feat: compression eval harness for agent/context_compressor.py

Ships a complete offline eval harness at scripts/compression_eval/. Runs
a real conversation fixture through ContextCompressor.compress(), asks
the compressor model to answer probe questions from the compressed
state, then has a judge model score each answer 0-5 on six dimensions
(accuracy, context_awareness, artifact_trail, completeness, continuity,
instruction_following). Methodology adapted from Factory's Dec 2025
write-up (https://factory.ai/news/evaluating-compression); the
scoreboard framing is not adopted.

Motivation: we edit context_compressor.py prompts and _template_sections
by hand and ship with no automated check that compression still
preserves file paths, error codes, or the active task. Until now there
has been no signal between 'test suite green' and 'a user hits a bad
summary in production.'

What's shipped
- DESIGN.md — full architecture, fixture/probe format, scrubber
  pipeline, grading rubric, open follow-ups
- README.md — usage, cost expectations, when to run it
- scrub_fixtures.py — reproducible pipeline that converts real sessions
  from ~/.hermes/sessions/*.jsonl into public-safe JSON fixtures. Applies
  agent.redact.redact_sensitive_text + username path normalisation +
  personal handle scrubbing + email/git-author normalisation + reasoning
  scratchpad stripping + platform-mention scrubbing + first-user
  paraphrase + system-prompt placeholder + orphan-message pruning + 2KB
  tool-output truncation
- fixtures/ — three scrubbed session snapshots covering three session
  shapes:
    feature-impl-context-priority  (75 msgs / ~17k tokens)
    debug-session-feishu-id-model  (59 msgs / ~13k tokens)
    config-build-competitive-scouts (61 msgs / ~23k tokens)
- probes/ — three probe banks (10-11 probes each) covering all four
  types (recall/artifact/continuation/decision) with expected_facts
  anchors (PR numbers, file paths, error codes, commands)
- rubric.py — six-dimension grading rubric, judge-prompt builder,
  JSON-with-fallback response parser
- compressor_driver.py — thin wrapper around ContextCompressor for
  forced single-shot compression (fixtures are below the default
  100k threshold so we force compress() to attribute score deltas
  to prompt changes, not threshold-fire variance)
- grader.py — two-phase continuation + grading calls via the OpenAI
  SDK directly against the resolved provider endpoint
- report.py — markdown report renderer (paste-ready for PR bodies),
  --compare-to delta mode, per-run JSON dumper
- run_eval.py — fire-style CLI (--fixtures, --runs, --judge-model,
  --compressor-model, --label, --focus-topic, --compare-to, --verbose)
- tests/scripts/test_compression_eval.py — 33 hermetic unit tests
  covering rubric parsing edge cases, judge-prompt building, report
  rendering, summariser medians, per-run JSON roundtrip, fixture and
  probe loading, and a PII smoke check on the checked-in fixtures

Non-LLM paths are covered by the 33-test suite that runs in CI. The
LLM paths (continuation + grading) require credentials and real API
calls, so they're exercised by running the eval itself — not by CI.

Validation
- 33/33 unit tests pass in 0.33s via scripts/run_tests.sh
- 50/50 adjacent tests (tests/agent/test_context_compressor.py) still
  pass — no regression introduced
- End-to-end dry run against debug-session-feishu-id-model with
  openai/gpt-5.4-mini via Nous Portal:
    Compression: 13081 -> 3055 tokens (76.6% ratio), 59 -> 10 messages
    Overall score: 3.25 (artifact_trail 1.50 is the weak spot,
    matching Factory's published observation)
    Specific probe misses surfaced with concrete judge notes

Noise floor (one empirical data point)
Same inputs re-run: overall 3.25 -> 3.17 (delta -0.08). Individual
dimensions varied up to ±0.5 between two single-run medians. Confirms
the DESIGN.md < 0.3 noise guidance is the right order of magnitude
for single-run comparisons. Tighter noise measurement (N=10) is
tracked as an open follow-up in DESIGN.md.

Why scripts/ and not tests/
Requires API credentials, costs ~$0.50-1.50 per run, minutes to
execute, LLM-graded (non-deterministic). Incompatible with
scripts/run_tests.sh which is hermetic, parallel, credential-free.
scripts/sample_and_compress.py is the existing precedent for offline
credentialed tooling.

Open follow-ups (tracked in DESIGN.md, not blocking this PR)
1. Iterative-merge fixture (two chained compressions on one session)
2. Precise noise-floor measurement at N=10
3. Scripted scrubber helpers to lower the cost of fixture #4+
4. Judge model selection policy (pin vs. per-user)

.dockerignore

@@ -14,3 +14,6 @@ node_modules
 .env
 
 *.md
+
+# Runtime data (bind-mounted at /opt/data; must not leak into build context)
+data/

.github/workflows/deploy-site.yml

@@ -53,6 +53,9 @@ jobs:
       - name: Extract skill metadata for dashboard
         run: python3 website/scripts/extract-skills.py
 
+      - name: Regenerate per-skill docs pages + catalogs
+        run: python3 website/scripts/generate-skill-docs.py
+
       - name: Build skills index (if not already present)
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/docs-site-checks.yml

@@ -36,6 +36,9 @@ jobs:
       - name: Extract skill metadata for dashboard
         run: python3 website/scripts/extract-skills.py
 
+      - name: Regenerate per-skill docs pages + catalogs
+        run: python3 website/scripts/generate-skill-docs.py
+
       - name: Lint docs diagrams
         run: npm run lint:diagrams
         working-directory: website

.gitignore

@@ -1,3 +1,4 @@
+.DS_Store
 /venv/
 /_pycache/
 *.pyc*
@@ -51,6 +52,10 @@ ignored/
 .worktrees/
 environments/benchmarks/evals/
 
+# Compression eval run outputs (harness lives in scripts/compression_eval/)
+scripts/compression_eval/results/*
+!scripts/compression_eval/results/.gitkeep
+
 # Web UI build output
 hermes_cli/web_dist/
 

AGENTS.md

@@ -5,78 +5,61 @@ Instructions for AI coding assistants and developers working on the hermes-agent
 ## Development Environment
 
 ```bash
-source venv/bin/activate  # ALWAYS activate before running Python
+# Prefer .venv; fall back to venv if that's what your checkout has.
+source .venv/bin/activate   # or: source venv/bin/activate
 ```
 
+`scripts/run_tests.sh` probes `.venv` first, then `venv`, then
+`$HOME/.hermes/hermes-agent/venv` (for worktrees that share a venv with the
+main checkout).
+
 ## Project Structure
 
+File counts shift constantly — don't treat the tree below as exhaustive.
+The canonical source is the filesystem. The notes call out the load-bearing
+entry points you'll actually edit.
+
 ```
 hermes-agent/
-├── run_agent.py          # AIAgent class — core conversation loop
+├── run_agent.py          # AIAgent class — core conversation loop (~12k LOC)
 ├── model_tools.py        # Tool orchestration, discover_builtin_tools(), handle_function_call()
 ├── toolsets.py           # Toolset definitions, _HERMES_CORE_TOOLS list
-├── cli.py                # HermesCLI class — interactive CLI orchestrator
+├── cli.py                # HermesCLI class — interactive CLI orchestrator (~11k LOC)
 ├── hermes_state.py       # SessionDB — SQLite session store (FTS5 search)
-├── agent/                # Agent internals
-│   ├── prompt_builder.py     # System prompt assembly
-│   ├── context_compressor.py # Auto context compression
-│   ├── prompt_caching.py     # Anthropic prompt caching
-│   ├── auxiliary_client.py   # Auxiliary LLM client (vision, summarization)
-│   ├── model_metadata.py     # Model context lengths, token estimation
-│   ├── models_dev.py         # models.dev registry integration (provider-aware context)
-│   ├── display.py            # KawaiiSpinner, tool preview formatting
-│   ├── skill_commands.py     # Skill slash commands (shared CLI/gateway)
-│   └── trajectory.py         # Trajectory saving helpers
-├── hermes_cli/           # CLI subcommands and setup
-│   ├── main.py           # Entry point — all `hermes` subcommands
-│   ├── config.py         # DEFAULT_CONFIG, OPTIONAL_ENV_VARS, migration
-│   ├── commands.py       # Slash command definitions + SlashCommandCompleter
-│   ├── callbacks.py      # Terminal callbacks (clarify, sudo, approval)
-│   ├── setup.py          # Interactive setup wizard
-│   ├── skin_engine.py    # Skin/theme engine — CLI visual customization
-│   ├── skills_config.py  # `hermes skills` — enable/disable skills per platform
-│   ├── tools_config.py   # `hermes tools` — enable/disable tools per platform
-│   ├── skills_hub.py     # `/skills` slash command (search, browse, install)
-│   ├── models.py         # Model catalog, provider model lists
-│   ├── model_switch.py   # Shared /model switch pipeline (CLI + gateway)
-│   └── auth.py           # Provider credential resolution
-├── tools/                # Tool implementations (one file per tool)
-│   ├── registry.py       # Central tool registry (schemas, handlers, dispatch)
-│   ├── approval.py       # Dangerous command detection
-│   ├── terminal_tool.py  # Terminal orchestration
-│   ├── process_registry.py # Background process management
-│   ├── file_tools.py     # File read/write/search/patch
-│   ├── web_tools.py      # Web search/extract (Parallel + Firecrawl)
-│   ├── browser_tool.py   # Browserbase browser automation
-│   ├── code_execution_tool.py # execute_code sandbox
-│   ├── delegate_tool.py  # Subagent delegation
-│   ├── mcp_tool.py       # MCP client (~1050 lines)
+├── hermes_constants.py   # get_hermes_home(), display_hermes_home() — profile-aware paths
+├── hermes_logging.py     # setup_logging() — agent.log / errors.log / gateway.log (profile-aware)
+├── batch_runner.py       # Parallel batch processing
+├── agent/                # Agent internals (provider adapters, memory, caching, compression, etc.)
+├── hermes_cli/           # CLI subcommands, setup wizard, plugins loader, skin engine
+├── tools/                # Tool implementations — auto-discovered via tools/registry.py
 │   └── environments/     # Terminal backends (local, docker, ssh, modal, daytona, singularity)
-├── gateway/              # Messaging platform gateway
-│   ├── run.py            # Main loop, slash commands, message dispatch
-│   ├── session.py        # SessionStore — conversation persistence
-│   └── platforms/        # Adapters: telegram, discord, slack, whatsapp, homeassistant, signal, qqbot
+├── gateway/              # Messaging gateway — run.py + session.py + platforms/
+│   ├── platforms/        # Adapter per platform (telegram, discord, slack, whatsapp,
+│   │                     #   homeassistant, signal, matrix, mattermost, email, sms,
+│   │                     #   dingtalk, wecom, weixin, feishu, qqbot, bluebubbles,
+│   │                     #   webhook, api_server, ...). See ADDING_A_PLATFORM.md.
+│   └── builtin_hooks/    # Always-registered gateway hooks (boot-md, ...)
+├── plugins/              # Plugin system (see "Plugins" section below)
+│   ├── memory/           # Memory-provider plugins (honcho, mem0, supermemory, ...)
+│   ├── context_engine/   # Context-engine plugins
+│   └── <others>/         # Dashboard, image-gen, disk-cleanup, examples, ...
+├── optional-skills/      # Heavier/niche skills shipped but NOT active by default
+├── skills/               # Built-in skills bundled with the repo
 ├── ui-tui/               # Ink (React) terminal UI — `hermes --tui`
-│   ├── src/entry.tsx        # TTY gate + render()
-│   ├── src/app.tsx          # Main state machine and UI
-│   ├── src/gatewayClient.ts # Child process + JSON-RPC bridge
-│   ├── src/app/             # Decomposed app logic (event handler, slash handler, stores, hooks)
-│   ├── src/components/      # Ink components (branding, markdown, prompts, pickers, etc.)
-│   ├── src/hooks/           # useCompletion, useInputHistory, useQueue, useVirtualHistory
-│   └── src/lib/             # Pure helpers (history, osc52, text, rpc, messages)
+│   └── src/              # entry.tsx, app.tsx, gatewayClient.ts + app/components/hooks/lib
 ├── tui_gateway/          # Python JSON-RPC backend for the TUI
-│   ├── entry.py             # stdio entrypoint
-│   ├── server.py            # RPC handlers and session logic
-│   ├── render.py            # Optional rich/ANSI bridge
-│   └── slash_worker.py      # Persistent HermesCLI subprocess for slash commands
 ├── acp_adapter/          # ACP server (VS Code / Zed / JetBrains integration)
-├── cron/                 # Scheduler (jobs.py, scheduler.py)
+├── cron/                 # Scheduler — jobs.py, scheduler.py
 ├── environments/         # RL training environments (Atropos)
-├── tests/                # Pytest suite (~3000 tests)
-└── batch_runner.py       # Parallel batch processing
+├── scripts/              # run_tests.sh, release.py, auxiliary scripts
+├── website/              # Docusaurus docs site
+└── tests/                # Pytest suite (~15k tests across ~700 files as of Apr 2026)
 ```
 
-**User config:** `~/.hermes/config.yaml` (settings), `~/.hermes/.env` (API keys)
+**User config:** `~/.hermes/config.yaml` (settings), `~/.hermes/.env` (API keys only).
+**Logs:** `~/.hermes/logs/` — `agent.log` (INFO+), `errors.log` (WARNING+),
+`gateway.log` when running the gateway. Profile-aware via `get_hermes_home()`.
+Browse with `hermes logs [--follow] [--level ...] [--session ...]`.
 
 ## File Dependency Chain
 
@@ -94,20 +77,30 @@ run_agent.py, cli.py, batch_runner.py, environments/
 
 ## AIAgent Class (run_agent.py)
 
+The real `AIAgent.__init__` takes ~60 parameters (credentials, routing, callbacks,
+session context, budget, credential pool, etc.). The signature below is the
+minimum subset you'll usually touch — read `run_agent.py` for the full list.
+
 ```python
 class AIAgent:
     def __init__(self,
-        model: str = "anthropic/claude-opus-4.6",
-        max_iterations: int = 90,
+        base_url: str = None,
+        api_key: str = None,
+        provider: str = None,
+        api_mode: str = None,              # "chat_completions" | "codex_responses" | ...
+        model: str = "",                   # empty → resolved from config/provider later
+        max_iterations: int = 90,          # tool-calling iterations (shared with subagents)
         enabled_toolsets: list = None,
         disabled_toolsets: list = None,
         quiet_mode: bool = False,
         save_trajectories: bool = False,
-        platform: str = None,           # "cli", "telegram", etc.
+        platform: str = None,              # "cli", "telegram", etc.
         session_id: str = None,
         skip_context_files: bool = False,
         skip_memory: bool = False,
-        # ... plus provider, api_mode, callbacks, routing params
+        credential_pool=None,
+        # ... plus callbacks, thread/user/chat IDs, iteration_budget, fallback_model,
+        # checkpoints config, prefill_messages, service_tier, reasoning_config, etc.
     ): ...
 
     def chat(self, message: str) -> str:
@@ -120,10 +113,13 @@ class AIAgent:
 
 ### Agent Loop
 
-The core loop is inside `run_conversation()` — entirely synchronous:
+The core loop is inside `run_conversation()` — entirely synchronous, with
+interrupt checks, budget tracking, and a one-turn grace call:
 
 ```python
-while api_call_count < self.max_iterations and self.iteration_budget.remaining > 0:
+while (api_call_count < self.max_iterations and self.iteration_budget.remaining > 0) \
+        or self._budget_grace_call:
+    if self._interrupt_requested: break
     response = client.chat.completions.create(model=model, messages=messages, tools=tool_schemas)
     if response.tool_calls:
         for tool_call in response.tool_calls:
@@ -134,7 +130,8 @@ while api_call_count < self.max_iterations and self.iteration_budget.remaining >
         return response.content
 ```
 
-Messages follow OpenAI format: `{"role": "system/user/assistant/tool", ...}`. Reasoning content is stored in `assistant_msg["reasoning"]`.
+Messages follow OpenAI format: `{"role": "system/user/assistant/tool", ...}`.
+Reasoning content is stored in `assistant_msg["reasoning"]`.
 
 ---
 
@@ -243,17 +240,6 @@ npm run fmt       # prettier
 npm test          # vitest
 ```
 
-### TUI in the Dashboard (`hermes dashboard` → `/chat`)
-
-The dashboard embeds the real `hermes --tui` — **not** a rewrite.  See `hermes_cli/pty_bridge.py` + the `@app.websocket("/api/pty")` endpoint in `hermes_cli/web_server.py`.
-
-- Browser loads `web/src/pages/ChatPage.tsx`, which mounts xterm.js's `Terminal` with the WebGL renderer, `@xterm/addon-fit` for container-driven resize, and `@xterm/addon-unicode11` for modern wide-character widths.
-- `/api/pty?token=…` upgrades to a WebSocket; auth uses the same ephemeral `_SESSION_TOKEN` as REST, via query param (browsers can't set `Authorization` on WS upgrade).
-- The server spawns whatever `hermes --tui` would spawn, through `ptyprocess` (POSIX PTY — WSL works, native Windows does not).
-- Frames: raw PTY bytes each direction; resize via `\x1b[RESIZE:<cols>;<rows>]` intercepted on the server and applied with `TIOCSWINSZ`.
-
-**Never add a parallel chat surface in React.** If you catch yourself re-implementing slash popover / model picker / tool cards for the dashboard, stop — the TUI already does those, and anything new you add to Ink will appear in the dashboard automatically.
-
 ---
 
 ## Adding New Tools
@@ -291,7 +277,7 @@ The registry handles schema collection, dispatch, availability checking, and err
 
 **State files**: If a tool stores persistent state (caches, logs, checkpoints), use `get_hermes_home()` for the base directory — never `Path.home() / ".hermes"`. This ensures each profile gets its own state.
 
-**Agent-level tools** (todo, memory): intercepted by `run_agent.py` before `handle_function_call()`. See `todo_tool.py` for the pattern.
+**Agent-level tools** (todo, memory): intercepted by `run_agent.py` before `handle_function_call()`. See `tools/todo_tool.py` for the pattern.
 
 ---
 
@@ -299,9 +285,13 @@ The registry handles schema collection, dispatch, availability checking, and err
 
 ### config.yaml options:
 1. Add to `DEFAULT_CONFIG` in `hermes_cli/config.py`
-2. Bump `_config_version` (currently 5) to trigger migration for existing users
+2. Bump `_config_version` (check the current value at the top of `DEFAULT_CONFIG`)
+   ONLY if you need to actively migrate/transform existing user config
+   (renaming keys, changing structure). Adding a new key to an existing
+   section is handled automatically by the deep-merge and does NOT require
+   a version bump.
 
-### .env variables:
+### .env variables (SECRETS ONLY — API keys, tokens, passwords):
 1. Add to `OPTIONAL_ENV_VARS` in `hermes_cli/config.py` with metadata:
 ```python
 "NEW_API_KEY": {
@@ -313,13 +303,29 @@ The registry handles schema collection, dispatch, availability checking, and err
 },
 ```
 
-### Config loaders (two separate systems):
+Non-secret settings (timeouts, thresholds, feature flags, paths, display
+preferences) belong in `config.yaml`, not `.env`. If internal code needs an
+env var mirror for backward compatibility, bridge it from `config.yaml` to
+the env var in code (see `gateway_timeout`, `terminal.cwd` → `TERMINAL_CWD`).
+
+### Config loaders (three paths — know which one you're in):
 
 | Loader | Used by | Location |
 |--------|---------|----------|
-| `load_cli_config()` | CLI mode | `cli.py` |
-| `load_config()` | `hermes tools`, `hermes setup` | `hermes_cli/config.py` |
-| Direct YAML load | Gateway | `gateway/run.py` |
+| `load_cli_config()` | CLI mode | `cli.py` — merges CLI-specific defaults + user YAML |
+| `load_config()` | `hermes tools`, `hermes setup`, most CLI subcommands | `hermes_cli/config.py` — merges `DEFAULT_CONFIG` + user YAML |
+| Direct YAML load | Gateway runtime | `gateway/run.py` + `gateway/config.py` — reads user YAML raw |
+
+If you add a new key and the CLI sees it but the gateway doesn't (or vice
+versa), you're on the wrong loader. Check `DEFAULT_CONFIG` coverage.
+
+### Working directory:
+- **CLI** — uses the process's current directory (`os.getcwd()`).
+- **Messaging** — uses `terminal.cwd` from `config.yaml`. The gateway bridges this
+  to the `TERMINAL_CWD` env var for child tools. **`MESSAGING_CWD` has been
+  removed** — the config loader prints a deprecation warning if it's set in
+  `.env`. Same for `TERMINAL_CWD` in `.env`; the canonical setting is
+  `terminal.cwd` in `config.yaml`.
 
 ---
 
@@ -412,7 +418,95 @@ Activate with `/skin cyberpunk` or `display.skin: cyberpunk` in config.yaml.
 
 ---
 
+## Plugins
+
+Hermes has two plugin surfaces. Both live under `plugins/` in the repo so
+repo-shipped plugins can be discovered alongside user-installed ones in
+`~/.hermes/plugins/` and pip-installed entry points.
+
+### General plugins (`hermes_cli/plugins.py` + `plugins/<name>/`)
+
+`PluginManager` discovers plugins from `~/.hermes/plugins/`, `./.hermes/plugins/`,
+and pip entry points. Each plugin exposes a `register(ctx)` function that
+can:
+
+- Register Python-callback lifecycle hooks:
+  `pre_tool_call`, `post_tool_call`, `pre_llm_call`, `post_llm_call`,
+  `on_session_start`, `on_session_end`
+- Register new tools via `ctx.register_tool(...)`
+- Register CLI subcommands via `ctx.register_cli_command(...)` — the
+  plugin's argparse tree is wired into `hermes` at startup so
+  `hermes <pluginname> <subcmd>` works with no change to `main.py`
+
+Hooks are invoked from `model_tools.py` (pre/post tool) and `run_agent.py`
+(lifecycle). **Discovery timing pitfall:** `discover_plugins()` only runs
+as a side effect of importing `model_tools.py`. Code paths that read plugin
+state without importing `model_tools.py` first must call `discover_plugins()`
+explicitly (it's idempotent).
+
+### Memory-provider plugins (`plugins/memory/<name>/`)
+
+Separate discovery system for pluggable memory backends. Current built-in
+providers include **honcho, mem0, supermemory, byterover, hindsight,
+holographic, openviking, retaindb**.
+
+Each provider implements the `MemoryProvider` ABC (see `agent/memory_provider.py`)
+and is orchestrated by `agent/memory_manager.py`. Lifecycle hooks include
+`sync_turn(turn_messages)`, `prefetch(query)`, `shutdown()`, and optional
+`post_setup(hermes_home, config)` for setup-wizard integration.
+
+**CLI commands via `plugins/memory/<name>/cli.py`:** if a memory plugin
+defines `register_cli(subparser)`, `discover_plugin_cli_commands()` finds
+it at argparse setup time and wires it into `hermes <plugin>`. The
+framework only exposes CLI commands for the **currently active** memory
+provider (read from `memory.provider` in config.yaml), so disabled
+providers don't clutter `hermes --help`.
+
+**Rule (Teknium, May 2026):** plugins MUST NOT modify core files
+(`run_agent.py`, `cli.py`, `gateway/run.py`, `hermes_cli/main.py`, etc.).
+If a plugin needs a capability the framework doesn't expose, expand the
+generic plugin surface (new hook, new ctx method) — never hardcode
+plugin-specific logic into core. PR #5295 removed 95 lines of hardcoded
+honcho argparse from `main.py` for exactly this reason.
+
+### Dashboard / context-engine / image-gen plugin directories
+
+`plugins/context_engine/`, `plugins/image_gen/`, `plugins/example-dashboard/`,
+etc. follow the same pattern (ABC + orchestrator + per-plugin directory).
+Context engines plug into `agent/context_engine.py`; image-gen providers
+into `agent/image_gen_provider.py`.
+
+---
+
+## Skills
+
+Two parallel surfaces:
+
+- **`skills/`** — built-in skills shipped and loadable by default.
+  Organized by category directories (e.g. `skills/github/`, `skills/mlops/`).
+- **`optional-skills/`** — heavier or niche skills shipped with the repo but
+  NOT active by default. Installed explicitly via
+  `hermes skills install official/<category>/<skill>`. Adapter lives in
+  `tools/skills_hub.py` (`OptionalSkillSource`). Categories include
+  `autonomous-ai-agents`, `blockchain`, `communication`, `creative`,
+  `devops`, `email`, `health`, `mcp`, `migration`, `mlops`, `productivity`,
+  `research`, `security`, `web-development`.
+
+When reviewing skill PRs, check which directory they target — heavy-dep or
+niche skills belong in `optional-skills/`.
+
+### SKILL.md frontmatter
+
+Standard fields: `name`, `description`, `version`, `platforms`
+(OS-gating list: `[macos]`, `[linux, macos]`, ...),
+`metadata.hermes.tags`, `metadata.hermes.category`,
+`metadata.hermes.config` (config.yaml settings the skill needs — stored
+under `skills.config.<key>`, prompted during setup, injected at load time).
+
+---
+
 ## Important Policies
+
 ### Prompt Caching Must Not Break
 
 Hermes-Agent ensures caching remains valid throughout a conversation. **Do NOT implement changes that would:**
@@ -422,9 +516,10 @@ Hermes-Agent ensures caching remains valid throughout a conversation. **Do NOT i
 
 Cache-breaking forces dramatically higher costs. The ONLY time we alter context is during context compression.
 
-### Working Directory Behavior
-- **CLI**: Uses current directory (`.` → `os.getcwd()`)
-- **Messaging**: Uses `MESSAGING_CWD` env var (default: home directory)
+Slash commands that mutate system-prompt state (skills, tools, memory, etc.)
+must be **cache-aware**: default to deferred invalidation (change takes
+effect next session), with an opt-in `--now` flag for immediate
+invalidation. See `/skills install --now` for the canonical pattern.
 
 ### Background Process Notifications (Gateway)
 
@@ -446,7 +541,7 @@ Hermes supports **profiles** — multiple fully isolated instances, each with it
 `HERMES_HOME` directory (config, API keys, memory, sessions, skills, gateway, etc.).
 
 The core mechanism: `_apply_profile_override()` in `hermes_cli/main.py` sets
-`HERMES_HOME` before any module imports. All 119+ references to `get_hermes_home()`
+`HERMES_HOME` before any module imports. All `get_hermes_home()` references
 automatically scope to the active profile.
 
 ### Rules for profile-safe code
@@ -503,8 +598,12 @@ Use `get_hermes_home()` from `hermes_constants` for code paths. Use `display_her
 for user-facing print/log messages. Hardcoding `~/.hermes` breaks profiles — each profile
 has its own `HERMES_HOME` directory. This was the source of 5 bugs fixed in PR #3575.
 
-### DO NOT use `simple_term_menu` for interactive menus
-Rendering bugs in tmux/iTerm2 — ghosting on scroll. Use `curses` (stdlib) instead. See `hermes_cli/tools_config.py` for the pattern.
+### DO NOT introduce new `simple_term_menu` usage
+Existing call sites in `hermes_cli/main.py` remain for legacy fallback only;
+the preferred UI is curses (stdlib) because `simple_term_menu` has
+ghost-duplication rendering bugs in tmux/iTerm2 with arrow keys. New
+interactive menus must use `hermes_cli/curses_ui.py` — see
+`hermes_cli/tools_config.py` for the canonical pattern.
 
 ### DO NOT use `\033[K` (ANSI erase-to-EOL) in spinner/display code
 Leaks as literal `?[K` text under `prompt_toolkit`'s `patch_stdout`. Use space-padding: `f"\r{line}{' ' * pad}"`.
@@ -515,6 +614,30 @@ Leaks as literal `?[K` text under `prompt_toolkit`'s `patch_stdout`. Use space-p
 ### DO NOT hardcode cross-tool references in schema descriptions
 Tool schema descriptions must not mention tools from other toolsets by name (e.g., `browser_navigate` saying "prefer web_search"). Those tools may be unavailable (missing API keys, disabled toolset), causing the model to hallucinate calls to non-existent tools. If a cross-reference is needed, add it dynamically in `get_tool_definitions()` in `model_tools.py` — see the `browser_navigate` / `execute_code` post-processing blocks for the pattern.
 
+### The gateway has TWO message guards — both must bypass approval/control commands
+When an agent is running, messages pass through two sequential guards:
+(1) **base adapter** (`gateway/platforms/base.py`) queues messages in
+`_pending_messages` when `session_key in self._active_sessions`, and
+(2) **gateway runner** (`gateway/run.py`) intercepts `/stop`, `/new`,
+`/queue`, `/status`, `/approve`, `/deny` before they reach
+`running_agent.interrupt()`. Any new command that must reach the runner
+while the agent is blocked (e.g. approval prompts) MUST bypass BOTH
+guards and be dispatched inline, not via `_process_message_background()`
+(which races session lifecycle).
+
+### Squash merges from stale branches silently revert recent fixes
+Before squash-merging a PR, ensure the branch is up to date with `main`
+(`git fetch origin main && git reset --hard origin/main` in the worktree,
+then re-apply the PR's commits). A stale branch's version of an unrelated
+file will silently overwrite recent fixes on main when squashed. Verify
+with `git diff HEAD~1..HEAD` after merging — unexpected deletions are a
+red flag.
+
+### Don't wire in dead code without E2E validation
+Unused code that was never shipped was dead for a reason. Before wiring an
+unused module into a live code path, E2E test the real resolution chain
+with actual imports (not mocks) against a temp `HERMES_HOME`.
+
 ### Tests must not write to `~/.hermes/`
 The `_isolate_hermes_home` autouse fixture in `tests/conftest.py` redirects `HERMES_HOME` to a temp dir. Never hardcode `~/.hermes/` paths in tests.
 
@@ -570,7 +693,7 @@ If you can't use the wrapper (e.g. on Windows or inside an IDE that shells
 pytest directly), at minimum activate the venv and pass `-n 4`:
 
 ```bash
-source venv/bin/activate
+source .venv/bin/activate   # or: source venv/bin/activate
 python -m pytest tests/ -q -n 4
 ```
 

CONTRIBUTING.md

@@ -9,7 +9,7 @@ Thank you for contributing to Hermes Agent! This guide covers everything you nee
 We value contributions in this order:
 
 1. **Bug fixes** — crashes, incorrect behavior, data loss. Always top priority.
-2. **Cross-platform compatibility** — Windows, macOS, different Linux distros, different terminal emulators. We want Hermes to work everywhere.
+2. **Cross-platform compatibility** — macOS, different Linux distros, and WSL2 on Windows. We want Hermes to work everywhere.
 3. **Security hardening** — shell injection, prompt injection, path traversal, privilege escalation. See [Security](#security-considerations).
 4. **Performance and robustness** — retry logic, error handling, graceful degradation.
 5. **New skills** — but only broadly useful ones. See [Should it be a Skill or a Tool?](#should-it-be-a-skill-or-a-tool)
@@ -55,10 +55,10 @@ If your skill is specialized, community-contributed, or niche, it's better suite
 
 | Requirement | Notes |
 |-------------|-------|
-| **Git** | With `--recurse-submodules` support |
+| **Git** | With `--recurse-submodules` support, and the `git-lfs` extension installed |
 | **Python 3.11+** | uv will install it if missing |
 | **uv** | Fast Python package manager ([install](https://docs.astral.sh/uv/)) |
-| **Node.js 18+** | Optional — needed for browser tools and WhatsApp bridge |
+| **Node.js 20+** | Optional — needed for browser tools and WhatsApp bridge (matches root `package.json` engines) |
 
 ### Clone and install
 
@@ -515,7 +515,7 @@ See `hermes_cli/skin_engine.py` for the full schema and existing skins as exampl
 
 ## Cross-Platform Compatibility
 
-Hermes runs on Linux, macOS, and Windows. When writing code that touches the OS:
+Hermes runs on Linux, macOS, and WSL2 on Windows. When writing code that touches the OS:
 
 ### Critical rules
 
@@ -597,7 +597,7 @@ refactor/description   # Code restructuring
 
 1. **Run tests**: `pytest tests/ -v`
 2. **Test manually**: Run `hermes` and exercise the code path you changed
-3. **Check cross-platform impact**: If you touch file I/O, process management, or terminal handling, consider Windows and macOS
+3. **Check cross-platform impact**: If you touch file I/O, process management, or terminal handling, consider macOS, Linux, and WSL2
 4. **Keep PRs focused**: One logical change per PR. Don't mix a bug fix with a refactor with a new feature.
 
 ### PR description

Dockerfile

@@ -10,9 +10,11 @@ ENV PYTHONUNBUFFERED=1
 ENV PLAYWRIGHT_BROWSERS_PATH=/opt/hermes/.playwright
 
 # Install system dependencies in one layer, clear APT cache
+# tini reaps orphaned zombie processes (MCP stdio subprocesses, git, bun, etc.)
+# that would otherwise accumulate when hermes runs as PID 1. See #15012.
 RUN apt-get update && \
     apt-get install -y --no-install-recommends \
-        build-essential nodejs npm python3 ripgrep ffmpeg gcc python3-dev libffi-dev procps git openssh-client docker-cli && \
+        build-essential nodejs npm python3 ripgrep ffmpeg gcc python3-dev libffi-dev procps git openssh-client docker-cli tini && \
     rm -rf /var/lib/apt/lists/*
 
 # Non-root user for runtime; UID can be overridden via HERMES_UID at runtime
@@ -41,9 +43,15 @@ COPY --chown=hermes:hermes . .
 # Build web dashboard (Vite outputs to hermes_cli/web_dist/)
 RUN cd web && npm run build
 
+# ---------- Permissions ----------
+# Make install dir world-readable so any HERMES_UID can read it at runtime.
+# The venv needs to be traversable too.
+USER root
+RUN chmod -R a+rX /opt/hermes
+# Start as root so the entrypoint can usermod/groupmod + gosu.
+# If HERMES_UID is unset, the entrypoint drops to the default hermes user (10000).
+
 # ---------- Python virtualenv ----------
-RUN chown hermes:hermes /opt/hermes
-USER hermes
 RUN uv venv && \
     uv pip install --no-cache-dir -e ".[all]"
 
@@ -52,4 +60,4 @@ ENV HERMES_WEB_DIST=/opt/hermes/hermes_cli/web_dist
 ENV HERMES_HOME=/opt/data
 ENV PATH="/opt/data/.local/bin:${PATH}"
 VOLUME [ "/opt/data" ]
-ENTRYPOINT [ "/opt/hermes/docker/entrypoint.sh" ]
+ENTRYPOINT [ "/usr/bin/tini", "-g", "--", "/opt/hermes/docker/entrypoint.sh" ]

README.md

@@ -76,7 +76,7 @@ Hermes has two entry points: start the terminal UI with `hermes`, or run the gat
 | Set a personality | `/personality [name]` | `/personality [name]` |
 | Retry or undo the last turn | `/retry`, `/undo` | `/retry`, `/undo` |
 | Compress context / check usage | `/compress`, `/usage`, `/insights [--days N]` | `/compress`, `/usage`, `/insights [days]` |
-| Browse skills | `/skills` or `/<skill-name>` | `/skills` or `/<skill-name>` |
+| Browse skills | `/skills` or `/<skill-name>` | `/<skill-name>` |
 | Interrupt current work | `Ctrl+C` or send a new message | `/stop` or send a new message |
 | Platform-specific status | `/platforms` | `/status`, `/sethome` |
 
@@ -157,14 +157,10 @@ curl -LsSf https://astral.sh/uv/install.sh | sh
 uv venv venv --python 3.11
 source venv/bin/activate
 uv pip install -e ".[all,dev]"
-python -m pytest tests/ -q
+scripts/run_tests.sh
 ```
 
-> **RL Training (optional):** To work on the RL/Tinker-Atropos integration:
-> ```bash
-> git submodule update --init tinker-atropos
-> uv pip install -e "./tinker-atropos"
-> ```
+> **RL Training (optional):** The RL/Atropos integration (`environments/`) ships via the `atroposlib` and `tinker` dependencies pulled in by `.[all,dev]` — no submodule setup required.
 
 ---
 

RELEASE_v0.11.0.md

@@ -0,0 +1,453 @@
+# Hermes Agent v0.11.0 (v2026.4.23)
+
+**Release Date:** April 23, 2026
+**Since v0.9.0:** 1,556 commits · 761 merged PRs · 1,314 files changed · 224,174 insertions · 29 community contributors (290 including co-authors)
+
+> The Interface release — a full React/Ink rewrite of the interactive CLI, a pluggable transport architecture underneath every provider, native AWS Bedrock support, five new inference paths, a 17th messaging platform (QQBot), a dramatically expanded plugin surface, and GPT-5.5 via Codex OAuth.
+
+This release also folds in all the highlights deferred from v0.10.0 (which shipped only the Nous Tool Gateway) — so it covers roughly two weeks of work across the whole stack.
+
+---
+
+## ✨ Highlights
+
+- **New Ink-based TUI** — `hermes --tui` is now a full React/Ink rewrite of the interactive CLI, with a Python JSON-RPC backend (`tui_gateway`). Sticky composer, live streaming with OSC-52 clipboard support, stable picker keys, status bar with per-turn stopwatch and git branch, `/clear` confirm, light-theme preset, and a subagent spawn observability overlay. ~310 commits to `ui-tui/` + `tui_gateway/`. (@OutThisLife + Teknium)
+
+- **Transport ABC + Native AWS Bedrock** — Format conversion and HTTP transport were extracted from `run_agent.py` into a pluggable `agent/transports/` layer. `AnthropicTransport`, `ChatCompletionsTransport`, `ResponsesApiTransport`, and `BedrockTransport` each own their own format conversion and API shape. Native AWS Bedrock support via the Converse API ships on top of the new abstraction. ([#10549](https://github.com/NousResearch/hermes-agent/pull/10549), [#13347](https://github.com/NousResearch/hermes-agent/pull/13347), [#13366](https://github.com/NousResearch/hermes-agent/pull/13366), [#13430](https://github.com/NousResearch/hermes-agent/pull/13430), [#13805](https://github.com/NousResearch/hermes-agent/pull/13805), [#13814](https://github.com/NousResearch/hermes-agent/pull/13814) — @kshitijk4poor + Teknium)
+
+- **Five new inference paths** — Native NVIDIA NIM ([#11774](https://github.com/NousResearch/hermes-agent/pull/11774)), Arcee AI ([#9276](https://github.com/NousResearch/hermes-agent/pull/9276)), Step Plan ([#13893](https://github.com/NousResearch/hermes-agent/pull/13893)), Google Gemini CLI OAuth ([#11270](https://github.com/NousResearch/hermes-agent/pull/11270)), and Vercel ai-gateway with pricing + dynamic discovery ([#13223](https://github.com/NousResearch/hermes-agent/pull/13223) — @jerilynzheng). Plus Gemini routed through the native AI Studio API for better performance ([#12674](https://github.com/NousResearch/hermes-agent/pull/12674)).
+
+- **GPT-5.5 over Codex OAuth** — OpenAI's new GPT-5.5 reasoning model is now available through your ChatGPT Codex OAuth, with live model discovery wired into the model picker so new OpenAI releases show up without catalog updates. ([#14720](https://github.com/NousResearch/hermes-agent/pull/14720))
+
+- **QQBot — 17th supported platform** — Native QQBot adapter via QQ Official API v2, with QR scan-to-configure setup wizard, streaming cursor, emoji reactions, and DM/group policy gating that matches WeCom/Weixin parity. ([#9364](https://github.com/NousResearch/hermes-agent/pull/9364), [#11831](https://github.com/NousResearch/hermes-agent/pull/11831))
+
+- **Plugin surface expanded** — Plugins can now register slash commands (`register_command`), dispatch tools directly (`dispatch_tool`), block tool execution from hooks (`pre_tool_call` can veto), rewrite tool results (`transform_tool_result`), transform terminal output (`transform_terminal_output`), ship image_gen backends, and add custom dashboard tabs. The bundled disk-cleanup plugin is opt-in by default as a reference implementation. ([#9377](https://github.com/NousResearch/hermes-agent/pull/9377), [#10626](https://github.com/NousResearch/hermes-agent/pull/10626), [#10763](https://github.com/NousResearch/hermes-agent/pull/10763), [#10951](https://github.com/NousResearch/hermes-agent/pull/10951), [#12929](https://github.com/NousResearch/hermes-agent/pull/12929), [#12944](https://github.com/NousResearch/hermes-agent/pull/12944), [#12972](https://github.com/NousResearch/hermes-agent/pull/12972), [#13799](https://github.com/NousResearch/hermes-agent/pull/13799), [#14175](https://github.com/NousResearch/hermes-agent/pull/14175))
+
+- **`/steer` — mid-run agent nudges** — `/steer <prompt>` injects a note that the running agent sees after its next tool call, without interrupting the turn or breaking prompt cache. For when you want to course-correct an agent in-flight. ([#12116](https://github.com/NousResearch/hermes-agent/pull/12116))
+
+- **Shell hooks** — Wire any shell script as a Hermes lifecycle hook (pre_tool_call, post_tool_call, on_session_start, etc.) without writing a Python plugin. ([#13296](https://github.com/NousResearch/hermes-agent/pull/13296))
+
+- **Webhook direct-delivery mode** — Webhook subscriptions can now forward payloads straight to a platform chat without going through the agent — zero-LLM push notifications for alerting, uptime checks, and event streams. ([#12473](https://github.com/NousResearch/hermes-agent/pull/12473))
+
+- **Smarter delegation** — Subagents now have an explicit `orchestrator` role that can spawn their own workers, with configurable `max_spawn_depth` (default flat). Concurrent sibling subagents share filesystem state through a file-coordination layer so they don't clobber each other's edits. ([#13691](https://github.com/NousResearch/hermes-agent/pull/13691), [#13718](https://github.com/NousResearch/hermes-agent/pull/13718))
+
+- **Auxiliary models — configurable UI + main-model-first** — `hermes model` has a dedicated "Configure auxiliary models" screen for per-task overrides (compression, vision, session_search, title_generation). `auto` routing now defaults to the main model for side tasks across all users (previously aggregator users were silently routed to a cheap provider-side default). ([#11891](https://github.com/NousResearch/hermes-agent/pull/11891), [#11900](https://github.com/NousResearch/hermes-agent/pull/11900))
+
+- **Dashboard plugin system + live theme switching** — The web dashboard is now extensible. Third-party plugins can add custom tabs, widgets, and views without forking. Paired with a live-switching theme system — themes now control colors, fonts, layout, and density — so users can hot-swap the dashboard look without a reload. Same theming discipline the CLI has, now on the web. ([#10951](https://github.com/NousResearch/hermes-agent/pull/10951), [#10687](https://github.com/NousResearch/hermes-agent/pull/10687), [#14725](https://github.com/NousResearch/hermes-agent/pull/14725))
+
+- **Dashboard polish** — i18n (English + Chinese), react-router sidebar layout, mobile-responsive, Vercel deployment, real per-session API call tracking, and one-click update + gateway restart buttons. ([#9228](https://github.com/NousResearch/hermes-agent/pull/9228), [#9370](https://github.com/NousResearch/hermes-agent/pull/9370), [#9453](https://github.com/NousResearch/hermes-agent/pull/9453), [#10686](https://github.com/NousResearch/hermes-agent/pull/10686), [#13526](https://github.com/NousResearch/hermes-agent/pull/13526), [#14004](https://github.com/NousResearch/hermes-agent/pull/14004) — @austinpickett + @DeployFaith + Teknium)
+
+---
+
+## 🏗️ Core Agent & Architecture
+
+### Transport Layer (NEW)
+- **Transport ABC** abstracts format conversion and HTTP transport from `run_agent.py` into `agent/transports/` ([#13347](https://github.com/NousResearch/hermes-agent/pull/13347))
+- **AnthropicTransport** — Anthropic Messages API path ([#13366](https://github.com/NousResearch/hermes-agent/pull/13366), @kshitijk4poor)
+- **ChatCompletionsTransport** — default path for OpenAI-compatible providers ([#13805](https://github.com/NousResearch/hermes-agent/pull/13805))
+- **ResponsesApiTransport** — OpenAI Responses API + Codex build_kwargs wiring ([#13430](https://github.com/NousResearch/hermes-agent/pull/13430), @kshitijk4poor)
+- **BedrockTransport** — AWS Bedrock Converse API transport ([#13814](https://github.com/NousResearch/hermes-agent/pull/13814))
+
+### Provider & Model Support
+- **Native AWS Bedrock provider** via Converse API ([#10549](https://github.com/NousResearch/hermes-agent/pull/10549))
+- **NVIDIA NIM native provider** (salvage of #11703) ([#11774](https://github.com/NousResearch/hermes-agent/pull/11774))
+- **Arcee AI direct provider** ([#9276](https://github.com/NousResearch/hermes-agent/pull/9276))
+- **Step Plan provider** (salvage #6005) ([#13893](https://github.com/NousResearch/hermes-agent/pull/13893), @kshitijk4poor)
+- **Google Gemini CLI OAuth** inference provider ([#11270](https://github.com/NousResearch/hermes-agent/pull/11270))
+- **Vercel ai-gateway** with pricing, attribution, and dynamic discovery ([#13223](https://github.com/NousResearch/hermes-agent/pull/13223), @jerilynzheng)
+- **GPT-5.5 over Codex OAuth** with live model discovery in the picker ([#14720](https://github.com/NousResearch/hermes-agent/pull/14720))
+- **Gemini routed through native AI Studio API** ([#12674](https://github.com/NousResearch/hermes-agent/pull/12674))
+- **xAI Grok upgraded to Responses API** ([#10783](https://github.com/NousResearch/hermes-agent/pull/10783))
+- **Ollama improvements** — Cloud provider support, GLM continuation, `think=false` control, surrogate sanitization, `/v1` hint ([#10782](https://github.com/NousResearch/hermes-agent/pull/10782))
+- **Kimi K2.6** across OpenRouter, Nous Portal, native Kimi, and HuggingFace ([#13148](https://github.com/NousResearch/hermes-agent/pull/13148), [#13152](https://github.com/NousResearch/hermes-agent/pull/13152), [#13169](https://github.com/NousResearch/hermes-agent/pull/13169))
+- **Kimi K2.5** promoted to first position in all model suggestion lists ([#11745](https://github.com/NousResearch/hermes-agent/pull/11745), @kshitijk4poor)
+- **Xiaomi MiMo v2.5-pro + v2.5** on OpenRouter, Nous Portal, and native ([#14184](https://github.com/NousResearch/hermes-agent/pull/14184), [#14635](https://github.com/NousResearch/hermes-agent/pull/14635), @kshitijk4poor)
+- **GLM-5V-Turbo** for coding plan ([#9907](https://github.com/NousResearch/hermes-agent/pull/9907))
+- **Claude Opus 4.7** in Nous Portal catalog ([#11398](https://github.com/NousResearch/hermes-agent/pull/11398))
+- **OpenRouter elephant-alpha** in curated lists ([#9378](https://github.com/NousResearch/hermes-agent/pull/9378))
+- **OpenCode-Go** — Kimi K2.6 and Qwen3.5/3.6 Plus in curated catalog ([#13429](https://github.com/NousResearch/hermes-agent/pull/13429))
+- **minimax/minimax-m2.5:free** in OpenRouter catalog ([#13836](https://github.com/NousResearch/hermes-agent/pull/13836))
+- **`/model` merges models.dev entries** for lesser-loved providers ([#14221](https://github.com/NousResearch/hermes-agent/pull/14221))
+- **Per-provider + per-model `request_timeout_seconds`** config ([#12652](https://github.com/NousResearch/hermes-agent/pull/12652))
+- **Configurable API retry count** via `agent.api_max_retries` ([#14730](https://github.com/NousResearch/hermes-agent/pull/14730))
+- **ctx_size context length key** for Lemonade server (salvage #8536) ([#14215](https://github.com/NousResearch/hermes-agent/pull/14215))
+- **Custom provider display name prompt** ([#9420](https://github.com/NousResearch/hermes-agent/pull/9420))
+- **Recommendation badges** on tool provider selection ([#9929](https://github.com/NousResearch/hermes-agent/pull/9929))
+- Fix: correct GPT-5 family context lengths in fallback defaults ([#9309](https://github.com/NousResearch/hermes-agent/pull/9309))
+- Fix: clamp `minimal` reasoning effort to `low` on Responses API ([#9429](https://github.com/NousResearch/hermes-agent/pull/9429))
+- Fix: strip reasoning item IDs from Responses API input when `store=False` ([#10217](https://github.com/NousResearch/hermes-agent/pull/10217))
+- Fix: OpenViking correct account default + commit session on `/new` and compress ([#10463](https://github.com/NousResearch/hermes-agent/pull/10463))
+- Fix: Kimi `/coding` thinking block survival + empty reasoning_content + block ordering (multiple PRs)
+- Fix: don't send Anthropic thinking to api.kimi.com/coding ([#13826](https://github.com/NousResearch/hermes-agent/pull/13826))
+- Fix: send `max_tokens`, `reasoning_effort`, and `thinking` for Kimi/Moonshot
+- Fix: stream reasoning content through OpenAI-compatible providers that emit it
+
+### Agent Loop & Conversation
+- **`/steer <prompt>`** — mid-run agent nudges after next tool call ([#12116](https://github.com/NousResearch/hermes-agent/pull/12116))
+- **Orchestrator role + configurable spawn depth** for `delegate_task` (default flat) ([#13691](https://github.com/NousResearch/hermes-agent/pull/13691))
+- **Cross-agent file state coordination** for concurrent subagents ([#13718](https://github.com/NousResearch/hermes-agent/pull/13718))
+- **Compressor smart collapse, dedup, anti-thrashing**, template upgrade, hardening ([#10088](https://github.com/NousResearch/hermes-agent/pull/10088))
+- **Compression summaries respect the conversation's language** ([#12556](https://github.com/NousResearch/hermes-agent/pull/12556))
+- **Compression model falls back to main model** on permanent 503/404 ([#10093](https://github.com/NousResearch/hermes-agent/pull/10093))
+- **Auto-continue interrupted agent work** after gateway restart ([#9934](https://github.com/NousResearch/hermes-agent/pull/9934))
+- **Activity heartbeats** prevent false gateway inactivity timeouts ([#10501](https://github.com/NousResearch/hermes-agent/pull/10501))
+- **Auxiliary models UI** — dedicated screen for per-task overrides ([#11891](https://github.com/NousResearch/hermes-agent/pull/11891))
+- **Auxiliary auto routing defaults to main model** for all users ([#11900](https://github.com/NousResearch/hermes-agent/pull/11900))
+- **PLATFORM_HINTS for Matrix, Mattermost, Feishu** ([#14428](https://github.com/NousResearch/hermes-agent/pull/14428), @alt-glitch)
+- Fix: reset retry counters after compression; stop poisoning conversation history ([#10055](https://github.com/NousResearch/hermes-agent/pull/10055))
+- Fix: break compression-exhaustion infinite loop and auto-reset session ([#10063](https://github.com/NousResearch/hermes-agent/pull/10063))
+- Fix: stale agent timeout, uv venv detection, empty response after tools ([#10065](https://github.com/NousResearch/hermes-agent/pull/10065))
+- Fix: prevent premature loop exit when weak models return empty after substantive tool calls ([#10472](https://github.com/NousResearch/hermes-agent/pull/10472))
+- Fix: preserve pre-start terminal interrupts ([#10504](https://github.com/NousResearch/hermes-agent/pull/10504))
+- Fix: improve interrupt responsiveness during concurrent tool execution ([#10935](https://github.com/NousResearch/hermes-agent/pull/10935))
+- Fix: word-wrap spinner, interruptable agent join, and delegate_task interrupt ([#10940](https://github.com/NousResearch/hermes-agent/pull/10940))
+- Fix: `/stop` no longer resets the session ([#9224](https://github.com/NousResearch/hermes-agent/pull/9224))
+- Fix: honor interrupts during MCP tool waits ([#9382](https://github.com/NousResearch/hermes-agent/pull/9382), @helix4u)
+- Fix: break stuck session resume loops after repeated restarts ([#9941](https://github.com/NousResearch/hermes-agent/pull/9941))
+- Fix: empty response nudge crash + placeholder leak to cron targets ([#11021](https://github.com/NousResearch/hermes-agent/pull/11021))
+- Fix: streaming cursor sanitization to prevent message truncation (multiple PRs)
+- Fix: resolve `context_length` for plugin context engines ([#9238](https://github.com/NousResearch/hermes-agent/pull/9238))
+
+### Session & Memory
+- **Auto-prune old sessions + VACUUM state.db** at startup ([#13861](https://github.com/NousResearch/hermes-agent/pull/13861))
+- **Honcho overhaul** — context injection, 5-tool surface, cost safety, session isolation ([#10619](https://github.com/NousResearch/hermes-agent/pull/10619))
+- **Hindsight richer session-scoped retain metadata** (salvage of #6290) ([#13987](https://github.com/NousResearch/hermes-agent/pull/13987))
+- Fix: deduplicate memory provider tools to prevent 400 on strict providers ([#10511](https://github.com/NousResearch/hermes-agent/pull/10511))
+- Fix: discover user-installed memory providers from `$HERMES_HOME/plugins/` ([#10529](https://github.com/NousResearch/hermes-agent/pull/10529))
+- Fix: add `on_memory_write` bridge to sequential tool execution path ([#10507](https://github.com/NousResearch/hermes-agent/pull/10507))
+- Fix: preserve `session_id` across `previous_response_id` chains in `/v1/responses` ([#10059](https://github.com/NousResearch/hermes-agent/pull/10059))
+
+---
+
+## 🖥️ New Ink-based TUI
+
+A full React/Ink rewrite of the interactive CLI — invoked via `hermes --tui` or `HERMES_TUI=1`. Shipped across ~310 commits to `ui-tui/` and `tui_gateway/`.
+
+### TUI Foundations
+- New TUI based on Ink + Python JSON-RPC backend
+- Prettier + ESLint + vitest tooling for `ui-tui/`
+- Entry split between `src/entry.tsx` (TTY gate) and `src/app.tsx` (state machine)
+- Persistent `_SlashWorker` subprocess for slash command dispatch
+
+### UX & Features
+- **Stable picker keys, /clear confirm, light-theme preset** ([#12312](https://github.com/NousResearch/hermes-agent/pull/12312), @OutThisLife)
+- **Git branch in status bar** cwd label ([#12305](https://github.com/NousResearch/hermes-agent/pull/12305), @OutThisLife)
+- **Per-turn elapsed stopwatch in FaceTicker + done-in sys line** ([#13105](https://github.com/NousResearch/hermes-agent/pull/13105), @OutThisLife)
+- **Subagent spawn observability overlay** ([#14045](https://github.com/NousResearch/hermes-agent/pull/14045), @OutThisLife)
+- **Per-prompt elapsed stopwatch in status bar** ([#12948](https://github.com/NousResearch/hermes-agent/pull/12948))
+- Sticky composer that freezes during scroll
+- OSC-52 clipboard support for copy across SSH sessions
+- Virtualized history rendering for performance
+- Slash command autocomplete via `complete.slash` RPC
+- Path autocomplete via `complete.path` RPC
+- Dozens of resize/ghosting/sticky-prompt fixes landed through the week
+
+### Structural Refactors
+- Decomposed `app.tsx` into `app/event-handler`, `app/slash-handler`, `app/stores`, `app/hooks` ([#14640](https://github.com/NousResearch/hermes-agent/pull/14640) and surrounding)
+- Component split: `branding.tsx`, `markdown.tsx`, `prompts.tsx`, `sessionPicker.tsx`, `messageLine.tsx`, `thinking.tsx`, `maskedPrompt.tsx`
+- Hook split: `useCompletion`, `useInputHistory`, `useQueue`, `useVirtualHistory`
+
+---
+
+## 📱 Messaging Platforms (Gateway)
+
+### New Platforms
+- **QQBot (17th platform)** — QQ Official API v2 adapter with QR setup, streaming, package split ([#9364](https://github.com/NousResearch/hermes-agent/pull/9364), [#11831](https://github.com/NousResearch/hermes-agent/pull/11831))
+
+### Telegram
+- **Dedicated `TELEGRAM_PROXY` env var + config.yaml proxy support** (closes #9414, #6530, #9074, #7786) ([#10681](https://github.com/NousResearch/hermes-agent/pull/10681))
+- **`ignored_threads` config** for Telegram groups ([#9530](https://github.com/NousResearch/hermes-agent/pull/9530))
+- **Config option to disable link previews** (closes #8728) ([#10610](https://github.com/NousResearch/hermes-agent/pull/10610))
+- **Auto-wrap markdown tables** in code blocks ([#11794](https://github.com/NousResearch/hermes-agent/pull/11794))
+- Fix: prevent duplicate replies when stream task is cancelled ([#9319](https://github.com/NousResearch/hermes-agent/pull/9319))
+- Fix: prevent streaming cursor (▉) from appearing as standalone messages ([#9538](https://github.com/NousResearch/hermes-agent/pull/9538))
+- Fix: retry transient tool sends + cold-boot budget ([#10947](https://github.com/NousResearch/hermes-agent/pull/10947))
+- Fix: Markdown special char escaping in `send_exec_approval`
+- Fix: parentheses in URLs during MarkdownV2 link conversion
+- Fix: Unicode dash normalization in model switch (closes iOS smart-punctuation issue)
+- Many platform hint / streaming / session-key fixes
+
+### Discord
+- **Forum channel support** (salvage of #10145 + media + polish) ([#11920](https://github.com/NousResearch/hermes-agent/pull/11920))
+- **`DISCORD_ALLOWED_ROLES`** for role-based access control ([#11608](https://github.com/NousResearch/hermes-agent/pull/11608))
+- **Config option to disable slash commands** (salvage #13130) ([#14315](https://github.com/NousResearch/hermes-agent/pull/14315))
+- **Native `send_animation`** for inline GIF playback ([#10283](https://github.com/NousResearch/hermes-agent/pull/10283))
+- **`send_message` Discord media attachments** ([#10246](https://github.com/NousResearch/hermes-agent/pull/10246))
+- **`/skill` command group** with category subcommands ([#9909](https://github.com/NousResearch/hermes-agent/pull/9909))
+- **Extract reply text from message references** ([#9781](https://github.com/NousResearch/hermes-agent/pull/9781))
+
+### Feishu
+- **Intelligent reply on document comments** with 3-tier access control ([#11898](https://github.com/NousResearch/hermes-agent/pull/11898))
+- **Show processing state via reactions** on user messages ([#12927](https://github.com/NousResearch/hermes-agent/pull/12927))
+- **Preserve @mention context for agent consumption** (salvage #13874) ([#14167](https://github.com/NousResearch/hermes-agent/pull/14167))
+
+### DingTalk
+- **`require_mention` + `allowed_users` gating** (parity with Slack/Telegram/Discord) ([#11564](https://github.com/NousResearch/hermes-agent/pull/11564))
+- **QR-code device-flow authorization** for setup wizard ([#11574](https://github.com/NousResearch/hermes-agent/pull/11574))
+- **AI Cards streaming, emoji reactions, and media handling** (salvage of #10985) ([#11910](https://github.com/NousResearch/hermes-agent/pull/11910))
+
+### WhatsApp
+- **`send_voice`** — native audio message delivery ([#13002](https://github.com/NousResearch/hermes-agent/pull/13002))
+- **`dm_policy` and `group_policy`** parity with WeCom/Weixin/QQ adapters ([#13151](https://github.com/NousResearch/hermes-agent/pull/13151))
+
+### WeCom / Weixin
+- **WeCom QR-scan bot creation + interactive setup wizard** (salvage #13923) ([#13961](https://github.com/NousResearch/hermes-agent/pull/13961))
+
+### Signal
+- **Media delivery support** via `send_message` ([#13178](https://github.com/NousResearch/hermes-agent/pull/13178))
+
+### Slack
+- **Per-thread sessions for DMs by default** ([#10987](https://github.com/NousResearch/hermes-agent/pull/10987))
+
+### BlueBubbles (iMessage)
+- Group chat session separation, webhook registration & auth fixes ([#9806](https://github.com/NousResearch/hermes-agent/pull/9806))
+
+### Gateway Core
+- **Gateway proxy mode** — forward messages to a remote API server ([#9787](https://github.com/NousResearch/hermes-agent/pull/9787))
+- **Per-channel ephemeral prompts** (Discord, Telegram, Slack, Mattermost) ([#10564](https://github.com/NousResearch/hermes-agent/pull/10564))
+- **Surface plugin slash commands** natively on all platforms + decision-capable command hook ([#14175](https://github.com/NousResearch/hermes-agent/pull/14175))
+- **Support document/archive extensions in MEDIA: tag extraction** (salvage #8255) ([#14307](https://github.com/NousResearch/hermes-agent/pull/14307))
+- **Recognize `.pdf` in MEDIA: tag extraction** ([#13683](https://github.com/NousResearch/hermes-agent/pull/13683))
+- **`--all` flag for `gateway start` and `restart`** ([#10043](https://github.com/NousResearch/hermes-agent/pull/10043))
+- **Notify active sessions on gateway shutdown** + update health check ([#9850](https://github.com/NousResearch/hermes-agent/pull/9850))
+- **Block agent from self-destructing the gateway** via terminal (closes #6666) ([#9895](https://github.com/NousResearch/hermes-agent/pull/9895))
+- Fix: suppress duplicate replies on interrupt and streaming flood control ([#10235](https://github.com/NousResearch/hermes-agent/pull/10235))
+- Fix: close temporary agents after one-off tasks ([#11028](https://github.com/NousResearch/hermes-agent/pull/11028), @kshitijk4poor)
+- Fix: busy-session ack when user messages during active agent run ([#10068](https://github.com/NousResearch/hermes-agent/pull/10068))
+- Fix: route watch-pattern notifications to the originating session ([#10460](https://github.com/NousResearch/hermes-agent/pull/10460))
+- Fix: preserve notify context in executor threads ([#10921](https://github.com/NousResearch/hermes-agent/pull/10921), @kshitijk4poor)
+- Fix: avoid duplicate replies after interrupted long tasks ([#11018](https://github.com/NousResearch/hermes-agent/pull/11018))
+- Fix: unlink stale PID + lock files on cleanup
+- Fix: force-unlink stale PID file after `--replace` takeover
+
+---
+
+## 🔧 Tool System
+
+### Plugin Surface (major expansion)
+- **`register_command()`** — plugins can now add slash commands ([#10626](https://github.com/NousResearch/hermes-agent/pull/10626))
+- **`dispatch_tool()`** — plugins can invoke tools from their code ([#10763](https://github.com/NousResearch/hermes-agent/pull/10763))
+- **`pre_tool_call` blocking** — plugins can veto tool execution ([#9377](https://github.com/NousResearch/hermes-agent/pull/9377))
+- **`transform_tool_result`** — plugins rewrite tool results generically ([#12972](https://github.com/NousResearch/hermes-agent/pull/12972))
+- **`transform_terminal_output`** — plugins rewrite terminal tool output ([#12929](https://github.com/NousResearch/hermes-agent/pull/12929))
+- **Namespaced skill registration** for plugin skill bundles ([#9786](https://github.com/NousResearch/hermes-agent/pull/9786))
+- **Opt-in-by-default + bundled disk-cleanup plugin** (salvage #12212) ([#12944](https://github.com/NousResearch/hermes-agent/pull/12944))
+- **Pluggable `image_gen` backends + OpenAI provider** ([#13799](https://github.com/NousResearch/hermes-agent/pull/13799))
+- **`openai-codex` image_gen plugin** (gpt-image-2 via Codex OAuth) ([#14317](https://github.com/NousResearch/hermes-agent/pull/14317))
+- **Shell hooks** — wire shell scripts as hook callbacks ([#13296](https://github.com/NousResearch/hermes-agent/pull/13296))
+
+### Browser
+- **`browser_cdp` raw DevTools Protocol passthrough** ([#12369](https://github.com/NousResearch/hermes-agent/pull/12369))
+- Camofox hardening + connection stability across the window
+
+### Execute Code
+- **Project/strict execution modes** (default: project) ([#11971](https://github.com/NousResearch/hermes-agent/pull/11971))
+
+### Image Generation
+- **Multi-model FAL support** with picker in `hermes tools` ([#11265](https://github.com/NousResearch/hermes-agent/pull/11265))
+- **Recraft V3 → V4 Pro, Nano Banana → Pro upgrades** ([#11406](https://github.com/NousResearch/hermes-agent/pull/11406))
+- **GPT Image 2** in FAL catalog ([#13677](https://github.com/NousResearch/hermes-agent/pull/13677))
+- **xAI image generation provider** (grok-imagine-image) ([#14765](https://github.com/NousResearch/hermes-agent/pull/14765))
+
+### TTS / STT / Voice
+- **Google Gemini TTS provider** ([#11229](https://github.com/NousResearch/hermes-agent/pull/11229))
+- **xAI Grok STT provider** ([#14473](https://github.com/NousResearch/hermes-agent/pull/14473))
+- **xAI TTS** (shipped with Responses API upgrade) ([#10783](https://github.com/NousResearch/hermes-agent/pull/10783))
+- **KittenTTS local provider** (salvage of #2109) ([#13395](https://github.com/NousResearch/hermes-agent/pull/13395))
+- **CLI record beep toggle** ([#13247](https://github.com/NousResearch/hermes-agent/pull/13247), @helix4u)
+
+### Webhook / Cron
+- **Webhook direct-delivery mode** — zero-LLM push notifications ([#12473](https://github.com/NousResearch/hermes-agent/pull/12473))
+- **Cron `wakeAgent` gate** — scripts can skip the agent entirely ([#12373](https://github.com/NousResearch/hermes-agent/pull/12373))
+- **Cron per-job `enabled_toolsets`** — cap token overhead + cost per job ([#14767](https://github.com/NousResearch/hermes-agent/pull/14767))
+
+### Delegate
+- **Orchestrator role** + configurable spawn depth (default flat) ([#13691](https://github.com/NousResearch/hermes-agent/pull/13691))
+- **Cross-agent file state coordination** ([#13718](https://github.com/NousResearch/hermes-agent/pull/13718))
+
+### File / Patch
+- **`patch` — "did you mean?" feedback** when patch fails to match ([#13435](https://github.com/NousResearch/hermes-agent/pull/13435))
+
+### API Server
+- **Stream `/v1/responses` SSE tool events** (salvage #9779) ([#10049](https://github.com/NousResearch/hermes-agent/pull/10049))
+- **Inline image inputs** on `/v1/chat/completions` and `/v1/responses` ([#12969](https://github.com/NousResearch/hermes-agent/pull/12969))
+
+### Docker / Podman
+- **Entry-level Podman support** — `find_docker()` + rootless entrypoint ([#10066](https://github.com/NousResearch/hermes-agent/pull/10066))
+- **Add docker-cli to Docker image** (salvage #10096) ([#14232](https://github.com/NousResearch/hermes-agent/pull/14232))
+- **File-sync back to host on teardown** (salvage of #8189 + hardening) ([#11291](https://github.com/NousResearch/hermes-agent/pull/11291))
+
+### MCP
+- 12 MCP improvements across the window (status, timeout handling, tool-call forwarding, etc.)
+
+---
+
+## 🧩 Skills Ecosystem
+
+### Skill System
+- **Namespaced skill registration** for plugin bundles ([#9786](https://github.com/NousResearch/hermes-agent/pull/9786))
+- **`hermes skills reset`** to un-stick bundled skills ([#11468](https://github.com/NousResearch/hermes-agent/pull/11468))
+- **Skills guard opt-in** — `config.skills.guard_agent_created` (default off) ([#14557](https://github.com/NousResearch/hermes-agent/pull/14557))
+- **Bundled skill scripts runnable out of the box** ([#13384](https://github.com/NousResearch/hermes-agent/pull/13384))
+- **`xitter` replaced with `xurl`** — the official X API CLI ([#12303](https://github.com/NousResearch/hermes-agent/pull/12303))
+- **MiniMax-AI/cli as default skill tap** (salvage #7501) ([#14493](https://github.com/NousResearch/hermes-agent/pull/14493))
+- **Fuzzy `@` file completions + mtime sorting** ([#9467](https://github.com/NousResearch/hermes-agent/pull/9467))
+
+### New Skills
+- **concept-diagrams** (salvage of #11045, @v1k22) ([#11363](https://github.com/NousResearch/hermes-agent/pull/11363))
+- **architecture-diagram** (Cocoon AI port) ([#9906](https://github.com/NousResearch/hermes-agent/pull/9906))
+- **pixel-art** with hardware palettes and video animation ([#12663](https://github.com/NousResearch/hermes-agent/pull/12663), [#12725](https://github.com/NousResearch/hermes-agent/pull/12725))
+- **baoyu-comic** ([#13257](https://github.com/NousResearch/hermes-agent/pull/13257), @JimLiu)
+- **baoyu-infographic** — 21 layouts × 21 styles (salvage #9901) ([#12254](https://github.com/NousResearch/hermes-agent/pull/12254))
+- **page-agent** — embed Alibaba's in-page GUI agent in your webapp ([#13976](https://github.com/NousResearch/hermes-agent/pull/13976))
+- **fitness-nutrition** optional skill + optional env var support ([#9355](https://github.com/NousResearch/hermes-agent/pull/9355))
+- **drug-discovery** — ChEMBL, PubChem, OpenFDA, ADMET ([#9443](https://github.com/NousResearch/hermes-agent/pull/9443))
+- **touchdesigner-mcp** (salvage of #10081) ([#12298](https://github.com/NousResearch/hermes-agent/pull/12298))
+- **adversarial-ux-test** optional skill (salvage of #2494, @omnissiah-comelse) ([#13425](https://github.com/NousResearch/hermes-agent/pull/13425))
+- **maps** — added `guest_house`, `camp_site`, and dual-key bakery lookup ([#13398](https://github.com/NousResearch/hermes-agent/pull/13398))
+- **llm-wiki** — port provenance markers, source hashing, and quality signals ([#13700](https://github.com/NousResearch/hermes-agent/pull/13700))
+
+---
+
+## 📊 Web Dashboard
+
+- **i18n (English + Chinese) language switcher** ([#9453](https://github.com/NousResearch/hermes-agent/pull/9453))
+- **Live-switching theme system** ([#10687](https://github.com/NousResearch/hermes-agent/pull/10687))
+- **Dashboard plugin system** — extend the web UI with custom tabs ([#10951](https://github.com/NousResearch/hermes-agent/pull/10951))
+- **react-router, sidebar layout, sticky header, dropdown component** ([#9370](https://github.com/NousResearch/hermes-agent/pull/9370), @austinpickett)
+- **Responsive for mobile** ([#9228](https://github.com/NousResearch/hermes-agent/pull/9228), @DeployFaith)
+- **Vercel deployment** ([#10686](https://github.com/NousResearch/hermes-agent/pull/10686), [#11061](https://github.com/NousResearch/hermes-agent/pull/11061), @austinpickett)
+- **Context window config support** ([#9357](https://github.com/NousResearch/hermes-agent/pull/9357))
+- **HTTP health probe for cross-container gateway detection** ([#9894](https://github.com/NousResearch/hermes-agent/pull/9894))
+- **Update + restart gateway buttons** ([#13526](https://github.com/NousResearch/hermes-agent/pull/13526), @austinpickett)
+- **Real API call count per session** (salvages #10140) ([#14004](https://github.com/NousResearch/hermes-agent/pull/14004))
+
+---
+
+## 🖱️ CLI & User Experience
+
+- **Dynamic shell completion for bash, zsh, and fish** ([#9785](https://github.com/NousResearch/hermes-agent/pull/9785))
+- **Light-mode skins + skin-aware completion menus** ([#9461](https://github.com/NousResearch/hermes-agent/pull/9461))
+- **Numbered keyboard shortcuts** on approval and clarify prompts ([#13416](https://github.com/NousResearch/hermes-agent/pull/13416))
+- **Markdown stripping, compact multiline previews, external editor** ([#12934](https://github.com/NousResearch/hermes-agent/pull/12934))
+- **`--ignore-user-config` and `--ignore-rules` flags** (port codex#18646) ([#14277](https://github.com/NousResearch/hermes-agent/pull/14277))
+- **Account limits section in `/usage`** ([#13428](https://github.com/NousResearch/hermes-agent/pull/13428))
+- **Doctor: Command Installation check** for `hermes` bin symlink ([#10112](https://github.com/NousResearch/hermes-agent/pull/10112))
+- **ESC cancels secret/sudo prompts**, clearer skip messaging ([#9902](https://github.com/NousResearch/hermes-agent/pull/9902))
+- Fix: agent-facing text uses `display_hermes_home()` instead of hardcoded `~/.hermes` ([#10285](https://github.com/NousResearch/hermes-agent/pull/10285))
+- Fix: enforce `config.yaml` as sole CWD source + deprecate `.env` CWD vars + add `hermes memory reset` ([#11029](https://github.com/NousResearch/hermes-agent/pull/11029))
+
+---
+
+## 🔒 Security & Reliability
+
+- **Global toggle to allow private/internal URL resolution** ([#14166](https://github.com/NousResearch/hermes-agent/pull/14166))
+- **Block agent from self-destructing the gateway** via terminal (closes #6666) ([#9895](https://github.com/NousResearch/hermes-agent/pull/9895))
+- **Telegram callback authorization** on update prompts ([#10536](https://github.com/NousResearch/hermes-agent/pull/10536))
+- **SECURITY.md** added ([#10532](https://github.com/NousResearch/hermes-agent/pull/10532), @I3eg1nner)
+- **Warn about legacy hermes.service units** during `hermes update` ([#11918](https://github.com/NousResearch/hermes-agent/pull/11918))
+- **Complete ASCII-locale UnicodeEncodeError recovery** for `api_messages`/`reasoning_content` (closes #6843) ([#10537](https://github.com/NousResearch/hermes-agent/pull/10537))
+- **Prevent stale `os.environ` leak** after `clear_session_vars` ([#10527](https://github.com/NousResearch/hermes-agent/pull/10527))
+- **Prevent agent hang when backgrounding processes** via terminal tool ([#10584](https://github.com/NousResearch/hermes-agent/pull/10584))
+- Many smaller session-resume, interrupt, streaming, and memory-race fixes throughout the window
+
+---
+
+## 🐛 Notable Bug Fixes
+
+The `fix:` category in this window covers 482 PRs. Highlights:
+
+- Streaming cursor artifacts filtered from Matrix, Telegram, WhatsApp, Discord (multiple PRs)
+- `<think>` and `<thought>` blocks filtered from gateway stream consumers ([#9408](https://github.com/NousResearch/hermes-agent/pull/9408))
+- Gateway display.streaming root-config override regression ([#9799](https://github.com/NousResearch/hermes-agent/pull/9799))
+- Context `session_search` coerces limit to int (prevents TypeError) ([#10522](https://github.com/NousResearch/hermes-agent/pull/10522))
+- Memory tool stays available when `fcntl` is unavailable (Windows) ([#9783](https://github.com/NousResearch/hermes-agent/pull/9783))
+- Trajectory compressor credentials load from `HERMES_HOME/.env` ([#9632](https://github.com/NousResearch/hermes-agent/pull/9632), @Dusk1e)
+- `@_context_completions` no longer crashes on `@` mention ([#9683](https://github.com/NousResearch/hermes-agent/pull/9683), @kshitijk4poor)
+- Group session `user_id` no longer treated as `thread_id` in shutdown notifications ([#10546](https://github.com/NousResearch/hermes-agent/pull/10546))
+- Telegram `platform_hint` — markdown is supported (closes #8261) ([#10612](https://github.com/NousResearch/hermes-agent/pull/10612))
+- Doctor checks for Kimi China credentials fixed
+- Streaming: don't suppress final response when commentary message is sent ([#10540](https://github.com/NousResearch/hermes-agent/pull/10540))
+- Rapid Telegram follow-ups no longer get cut off
+
+---
+
+## 🧪 Testing & CI
+
+- **Contributor attribution CI check** on PRs ([#9376](https://github.com/NousResearch/hermes-agent/pull/9376))
+- Hermetic test parity (`scripts/run_tests.sh`) held across this window
+- Test count stabilized post-Transport refactor; CI matrix held green through the transport rollout
+
+---
+
+## 📚 Documentation
+
+- Atropos + wandb links in user guide
+- ACP / VS Code / Zed / JetBrains integration docs refresh
+- Webhook subscription docs updated for direct-delivery mode
+- Plugin author guide expanded for new hooks (`register_command`, `dispatch_tool`, `transform_tool_result`)
+- Transport layer developer guide added
+- Website removed Discussions link from README
+
+---
+
+## 👥 Contributors
+
+### Core
+- **@teknium1** (Teknium)
+
+### Top Community Contributors (by merged PR count)
+- **@kshitijk4poor** — 49 PRs · Transport refactor (AnthropicTransport, ResponsesApiTransport), Step Plan provider, Xiaomi MiMo v2.5 support, numerous gateway fixes, promoted Kimi K2.5, @ mention crash fix
+- **@OutThisLife** (Brooklyn) — 31 PRs · TUI polish, git branch in status bar, per-turn stopwatch, stable picker keys, `/clear` confirm, light-theme preset, subagent spawn observability overlay
+- **@helix4u** — 11 PRs · Voice CLI record beep, MCP tool interrupt handling, assorted stability fixes
+- **@austinpickett** — 8 PRs · Dashboard react-router + sidebar + sticky header + dropdown, Vercel deployment, update + restart buttons
+- **@alt-glitch** — 8 PRs · PLATFORM_HINTS for Matrix/Mattermost/Feishu, Matrix fixes
+- **@ethernet8023** — 3 PRs
+- **@benbarclay** — 3 PRs
+- **@Aslaaen** — 2 PRs
+
+### Also contributing
+@jerilynzheng (ai-gateway pricing), @JimLiu (baoyu-comic skill), @Dusk1e (trajectory compressor credentials), @DeployFaith (mobile-responsive dashboard), @LeonSGP43, @v1k22 (concept-diagrams), @omnissiah-comelse (adversarial-ux-test), @coekfung (Telegram MarkdownV2 expandable blockquotes), @liftaris (TUI provider resolution), @arihantsethia (skill analytics dashboard), @topcheer + @xing8star (QQBot foundation), @kovyrin, @I3eg1nner (SECURITY.md), @PeterBerthelsen, @lengxii, @priveperfumes, @sjz-ks, @cuyua9, @Disaster-Terminator, @leozeli, @LehaoLin, @trevthefoolish, @loongfay, @MrNiceRicee, @WideLee, @bluefishs, @malaiwah, @bobashopcashier, @dsocolobsky, @iamagenius00, @IAvecilla, @aniruddhaadak80, @Es1la, @asheriif, @walli, @jquesnelle (original Tool Gateway work).
+
+### All Contributors (alphabetical)
+
+@0xyg3n, @10ishq, @A-afflatus, @Abnertheforeman, @admin28980, @adybag14-cyber, @akhater, @alexzhu0,
+@AllardQuek, @alt-glitch, @aniruddhaadak80, @anna-oake, @anniesurla, @anthhub, @areu01or00, @arihantsethia,
+@arthurbr11, @asheriif, @Aslaaen, @Asunfly, @austinpickett, @AviArora02-commits, @AxDSan, @azhengbot, @Bartok9,
+@benbarclay, @bennytimz, @bernylinville, @bingo906, @binhnt92, @bkadish, @bluefishs, @bobashopcashier,
+@brantzh6, @BrennerSpear, @brianclemens, @briandevans, @brooklynnicholson, @bugkill3r, @buray, @burtenshaw,
+@cdanis, @cgarwood82, @ChimingLiu, @chongweiliu, @christopherwoodall, @coekfung, @cola-runner, @corazzione,
+@counterposition, @cresslank, @cuyua9, @cypres0099, @danieldoderlein, @davetist, @davidvv, @DeployFaith,
+@Dev-Mriganka, @devorun, @dieutx, @Disaster-Terminator, @dodo-reach, @draix, @DrStrangerUJN, @dsocolobsky,
+@Dusk1e, @dyxushuai, @elkimek, @elmatadorgh, @emozilla, @entropidelic, @Erosika, @erosika, @Es1la, @etcircle,
+@etherman-os, @ethernet8023, @fancydirty, @farion1231, @fatinghenji, @Fatty911, @fengtianyu88, @Feranmi10,
+@flobo3, @francip, @fuleinist, @g-guthrie, @GenKoKo, @gianfrancopiana, @gnanam1990, @GuyCui, @haileymarshall,
+@haimu0x, @handsdiff, @hansnow, @hedgeho9X, @helix4u, @hengm3467, @HenkDz, @heykb, @hharry11, @HiddenPuppy,
+@honghua, @houko, @houziershi, @hsy5571616, @huangke19, @hxp-plus, @Hypn0sis, @I3eg1nner, @iacker,
+@iamagenius00, @IAvecilla, @iborazzi, @Ifkellx, @ifrederico, @imink, @isaachuangGMICLOUD, @ismell0992-afk,
+@j0sephz, @Jaaneek, @jackjin1997, @JackTheGit, @jaffarkeikei, @jerilynzheng, @JiaDe-Wu, @Jiawen-lee, @JimLiu,
+@jinzheng8115, @jneeee, @jplew, @jquesnelle, @Julientalbot, @Junass1, @jvcl, @kagura-agent, @keifergu,
+@kevinskysunny, @keyuyuan, @konsisumer, @kovyrin, @kshitijk4poor, @leeyang1990, @LehaoLin, @lengxii,
+@LeonSGP43, @leozeli, @li0near, @liftaris, @Lind3ey, @Linux2010, @liujinkun2025, @LLQWQ, @Llugaes, @lmoncany,
+@longsizhuo, @lrawnsley, @Lubrsy706, @lumenradley, @luyao618, @lvnilesh, @LVT382009, @m0n5t3r, @Magaav,
+@MagicRay1217, @malaiwah, @manuelschipper, @Marvae, @MassiveMassimo, @mavrickdeveloper, @maxchernin, @memosr,
+@meng93, @mengjian-github, @MestreY0d4-Uninter, @Mibayy, @MikeFac, @mikewaters, @milkoor, @minorgod,
+@MrNiceRicee, @ms-alan, @mvanhorn, @n-WN, @N0nb0at, @Nan93, @NIDNASSER-Abdelmajid, @nish3451, @niyoh120,
+@nocoo, @nosleepcassette, @NousResearch, @ogzerber, @omnissiah-comelse, @Only-Code-A, @opriz, @OwenYWT, @pedh,
+@pefontana, @PeterBerthelsen, @phpoh, @pinion05, @plgonzalezrx8, @pradeep7127, @priveperfumes,
+@projectadmin-dev, @PStarH, @rnijhara, @Roy-oss1, @roytian1217, @RucchiZ, @Ruzzgar, @RyanLee-Dev, @Salt-555,
+@Sanjays2402, @sgaofen, @sharziki, @shenuu, @shin4, @SHL0MS, @shushuzn, @sicnuyudidi, @simon-gtcl,
+@simon-marcus, @sirEven, @Sisyphus, @sjz-ks, @snreynolds, @Societus, @Somme4096, @sontianye, @sprmn24,
+@StefanIsMe, @stephenschoettler, @Swift42, @taeng0204, @taeuk178, @tannerfokkens-maker, @TaroballzChen,
+@ten-ltw, @teyrebaz33, @Tianworld, @topcheer, @Tranquil-Flow, @trevthefoolish, @TroyMitchell911, @UNLINEARITY,
+@v1k22, @vivganes, @vominh1919, @vrinek, @VTRiot, @WadydX, @walli, @wenhao7, @WhiteWorld, @WideLee, @wujhsu,
+@WuTianyi123, @Wysie, @xandersbell, @xiaoqiang243, @xiayh0107, @xinpengdr, @Xowiek, @ycbai, @yeyitech, @ygd58,
+@youngDoo, @yudaiyan, @Yukipukii1, @yule975, @yyq4193, @yzx9, @ZaynJarvis, @zhang9w0v5, @zhanggttry,
+@zhangxicen, @zhongyueming1121, @zhouxiaoya12, @zons-zhaozhy
+
+Also: @maelrx, @Marco Rutsch, @MaxsolcuCrypto, @Mind-Dragon, @Paul Bergeron, @say8hi, @whitehatjr1001.
+
+
+---
+
+**Full Changelog**: [v2026.4.13...v2026.4.23](https://github.com/NousResearch/hermes-agent/compare/v2026.4.13...v2026.4.23)

acp_adapter/server.py

@@ -60,7 +60,7 @@ from acp_adapter.events import (
     make_tool_progress_cb,
 )
 from acp_adapter.permissions import make_approval_callback
-from acp_adapter.session import SessionManager, SessionState
+from acp_adapter.session import SessionManager, SessionState, _expand_acp_enabled_toolsets
 
 logger = logging.getLogger(__name__)
 
@@ -287,7 +287,11 @@ class HermesACPAgent(acp.Agent):
         try:
             from model_tools import get_tool_definitions
 
-            enabled_toolsets = getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"]
+            enabled_toolsets = _expand_acp_enabled_toolsets(
+                getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"],
+                mcp_server_names=[server.name for server in mcp_servers],
+            )
+            state.agent.enabled_toolsets = enabled_toolsets
             disabled_toolsets = getattr(state.agent, "disabled_toolsets", None)
             state.agent.tools = get_tool_definitions(
                 enabled_toolsets=enabled_toolsets,
@@ -754,7 +758,9 @@ class HermesACPAgent(acp.Agent):
     def _cmd_tools(self, args: str, state: SessionState) -> str:
         try:
             from model_tools import get_tool_definitions
-            toolsets = getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"]
+            toolsets = _expand_acp_enabled_toolsets(
+                getattr(state.agent, "enabled_toolsets", None) or ["hermes-acp"]
+            )
             tools = get_tool_definitions(enabled_toolsets=toolsets, quiet_mode=True)
             if not tools:
                 return "No tools available."

acp_adapter/session.py

@@ -106,6 +106,24 @@ def _register_task_cwd(task_id: str, cwd: str) -> None:
         logger.debug("Failed to register ACP task cwd override", exc_info=True)
 
 
+def _expand_acp_enabled_toolsets(
+    toolsets: List[str] | None = None,
+    mcp_server_names: List[str] | None = None,
+) -> List[str]:
+    """Return ACP toolsets plus explicit MCP server toolsets for this session."""
+    expanded: List[str] = []
+    for name in list(toolsets or ["hermes-acp"]):
+        if name and name not in expanded:
+            expanded.append(name)
+
+    for server_name in list(mcp_server_names or []):
+        toolset_name = f"mcp-{server_name}"
+        if server_name and toolset_name not in expanded:
+            expanded.append(toolset_name)
+
+    return expanded
+
+
 def _clear_task_cwd(task_id: str) -> None:
     """Remove task-specific cwd overrides for an ACP session."""
     if not task_id:
@@ -537,9 +555,18 @@ class SessionManager:
         elif isinstance(model_cfg, str) and model_cfg.strip():
             default_model = model_cfg.strip()
 
+        configured_mcp_servers = [
+            name
+            for name, cfg in (config.get("mcp_servers") or {}).items()
+            if not isinstance(cfg, dict) or cfg.get("enabled", True) is not False
+        ]
+
         kwargs = {
             "platform": "acp",
-            "enabled_toolsets": ["hermes-acp"],
+            "enabled_toolsets": _expand_acp_enabled_toolsets(
+                ["hermes-acp"],
+                mcp_server_names=configured_mcp_servers,
+            ),
             "quiet_mode": True,
             "session_id": session_id,
             "model": model or default_model,

agent/anthropic_adapter.py

@@ -14,10 +14,11 @@ import copy
 import json
 import logging
 import os
+import platform
+import subprocess
 from pathlib import Path
 
 from hermes_constants import get_hermes_home
-from types import SimpleNamespace
 from typing import Any, Dict, List, Optional, Tuple
 from utils import normalize_proxy_env_vars
 
@@ -278,8 +279,9 @@ def _is_oauth_token(key: str) -> bool:
     Positively identifies Anthropic OAuth tokens by their key format:
     - ``sk-ant-`` prefix (but NOT ``sk-ant-api``) → setup tokens, managed keys
     - ``eyJ`` prefix → JWTs from the Anthropic OAuth flow
+    - ``cc-`` prefix → Claude Code OAuth access tokens (from CLAUDE_CODE_OAUTH_TOKEN)
 
-    Non-Anthropic keys (MiniMax, Alibaba, etc.) don't match either pattern
+    Non-Anthropic keys (MiniMax, Alibaba, etc.) don't match any pattern
     and correctly return False.
     """
     if not key:
@@ -293,6 +295,9 @@ def _is_oauth_token(key: str) -> bool:
     # JWTs from Anthropic OAuth flow
     if key.startswith("eyJ"):
         return True
+    # Claude Code OAuth access tokens (opaque, from CLAUDE_CODE_OAUTH_TOKEN)
+    if key.startswith("cc-"):
+        return True
     return False
 
 
@@ -462,8 +467,72 @@ def build_anthropic_bedrock_client(region: str):
     )
 
 
+def _read_claude_code_credentials_from_keychain() -> Optional[Dict[str, Any]]:
+    """Read Claude Code OAuth credentials from the macOS Keychain.
+
+    Claude Code >=2.1.114 stores credentials in the macOS Keychain under the
+    service name "Claude Code-credentials" rather than (or in addition to)
+    the JSON file at ~/.claude/.credentials.json.
+
+    The password field contains a JSON string with the same claudeAiOauth
+    structure as the JSON file.
+
+    Returns dict with {accessToken, refreshToken?, expiresAt?} or None.
+    """
+    import platform
+    import subprocess
+
+    if platform.system() != "Darwin":
+        return None
+
+    try:
+        # Read the "Claude Code-credentials" generic password entry
+        result = subprocess.run(
+            ["security", "find-generic-password",
+             "-s", "Claude Code-credentials",
+             "-w"],
+            capture_output=True,
+            text=True,
+            timeout=5,
+        )
+    except (OSError, subprocess.TimeoutExpired):
+        logger.debug("Keychain: security command not available or timed out")
+        return None
+
+    if result.returncode != 0:
+        logger.debug("Keychain: no entry found for 'Claude Code-credentials'")
+        return None
+
+    raw = result.stdout.strip()
+    if not raw:
+        return None
+
+    try:
+        data = json.loads(raw)
+    except json.JSONDecodeError:
+        logger.debug("Keychain: credentials payload is not valid JSON")
+        return None
+
+    oauth_data = data.get("claudeAiOauth")
+    if oauth_data and isinstance(oauth_data, dict):
+        access_token = oauth_data.get("accessToken", "")
+        if access_token:
+            return {
+                "accessToken": access_token,
+                "refreshToken": oauth_data.get("refreshToken", ""),
+                "expiresAt": oauth_data.get("expiresAt", 0),
+                "source": "macos_keychain",
+            }
+
+    return None
+
+
 def read_claude_code_credentials() -> Optional[Dict[str, Any]]:
-    """Read refreshable Claude Code OAuth credentials from ~/.claude/.credentials.json.
+    """Read refreshable Claude Code OAuth credentials.
+
+    Checks two sources in order:
+      1. macOS Keychain (Darwin only) — "Claude Code-credentials" entry
+      2. ~/.claude/.credentials.json file
 
     This intentionally excludes ~/.claude.json primaryApiKey. Opencode's
     subscription flow is OAuth/setup-token based with refreshable credentials,
@@ -472,6 +541,12 @@ def read_claude_code_credentials() -> Optional[Dict[str, Any]]:
 
     Returns dict with {accessToken, refreshToken?, expiresAt?} or None.
     """
+    # Try macOS Keychain first (covers Claude Code >=2.1.114)
+    kc_creds = _read_claude_code_credentials_from_keychain()
+    if kc_creds:
+        return kc_creds
+
+    # Fall back to JSON file
     cred_path = Path.home() / ".claude" / ".credentials.json"
     if cred_path.exists():
         try:
@@ -642,7 +717,9 @@ def _write_claude_code_credentials(
         existing["claudeAiOauth"] = oauth_data
 
         cred_path.parent.mkdir(parents=True, exist_ok=True)
-        cred_path.write_text(json.dumps(existing, indent=2), encoding="utf-8")
+        _tmp_cred = cred_path.with_suffix(".tmp")
+        _tmp_cred.write_text(json.dumps(existing, indent=2), encoding="utf-8")
+        _tmp_cred.replace(cred_path)
         # Restrict permissions (credentials file)
         cred_path.chmod(0o600)
     except (OSError, IOError) as e:
@@ -1599,70 +1676,3 @@ def build_anthropic_kwargs(
     return kwargs
 
 
-def normalize_anthropic_response(
-    response,
-    strip_tool_prefix: bool = False,
-) -> Tuple[SimpleNamespace, str]:
-    """Normalize Anthropic response to match the shape expected by AIAgent.
-
-    Returns (assistant_message, finish_reason) where assistant_message has
-    .content, .tool_calls, and .reasoning attributes.
-
-    When *strip_tool_prefix* is True, removes the ``mcp_`` prefix that was
-    added to tool names for OAuth Claude Code compatibility.
-    """
-    text_parts = []
-    reasoning_parts = []
-    reasoning_details = []
-    tool_calls = []
-
-    for block in response.content:
-        if block.type == "text":
-            text_parts.append(block.text)
-        elif block.type == "thinking":
-            reasoning_parts.append(block.thinking)
-            block_dict = _to_plain_data(block)
-            if isinstance(block_dict, dict):
-                reasoning_details.append(block_dict)
-        elif block.type == "tool_use":
-            name = block.name
-            if strip_tool_prefix and name.startswith(_MCP_TOOL_PREFIX):
-                name = name[len(_MCP_TOOL_PREFIX):]
-            tool_calls.append(
-                SimpleNamespace(
-                    id=block.id,
-                    type="function",
-                    function=SimpleNamespace(
-                        name=name,
-                        arguments=json.dumps(block.input),
-                    ),
-                )
-            )
-
-    # Map Anthropic stop_reason to OpenAI finish_reason.
-    # Newer stop reasons added in Claude 4.5+ / 4.7:
-    #   - refusal: the model declined to answer (cyber safeguards, CSAM, etc.)
-    #   - model_context_window_exceeded: hit context limit (not max_tokens)
-    # Both need distinct handling upstream — a refusal should surface to the
-    # user with a clear message, and a context-window overflow should trigger
-    # compression/truncation rather than be treated as normal end-of-turn.
-    stop_reason_map = {
-        "end_turn": "stop",
-        "tool_use": "tool_calls",
-        "max_tokens": "length",
-        "stop_sequence": "stop",
-        "refusal": "content_filter",
-        "model_context_window_exceeded": "length",
-    }
-    finish_reason = stop_reason_map.get(response.stop_reason, "stop")
-
-    return (
-        SimpleNamespace(
-            content="\n".join(text_parts) if text_parts else None,
-            tool_calls=tool_calls or None,
-            reasoning="\n\n".join(reasoning_parts) if reasoning_parts else None,
-            reasoning_content=None,
-            reasoning_details=reasoning_details or None,
-        ),
-        finish_reason,
-    )

agent/auxiliary_client.py

@@ -74,6 +74,12 @@ _PROVIDER_ALIASES = {
     "minimax_cn": "minimax-cn",
     "claude": "anthropic",
     "claude-code": "anthropic",
+    "github": "copilot",
+    "github-copilot": "copilot",
+    "github-model": "copilot",
+    "github-models": "copilot",
+    "github-copilot-acp": "copilot-acp",
+    "copilot-acp-agent": "copilot-acp",
 }
 
 
@@ -89,10 +95,11 @@ def _normalize_aux_provider(provider: Optional[str]) -> str:
     if normalized == "main":
         # Resolve to the user's actual main provider so named custom providers
         # and non-aggregator providers (DeepSeek, Alibaba, etc.) work correctly.
-        main_prov = _read_main_provider()
+        main_prov = (_read_main_provider() or "").strip().lower()
         if main_prov and main_prov not in ("auto", "main", ""):
-            return main_prov
-        return "custom"
+            normalized = main_prov
+        else:
+            return "custom"
     return _PROVIDER_ALIASES.get(normalized, normalized)
 
 
@@ -151,7 +158,7 @@ _API_KEY_PROVIDER_AUX_MODELS: Dict[str, str] = {
 # differs from their main chat model, map it here.  The vision auto-detect
 # "exotic provider" branch checks this before falling back to the main model.
 _PROVIDER_VISION_MODELS: Dict[str, str] = {
-    "xiaomi": "mimo-v2-omni",
+    "xiaomi": "mimo-v2.5",
     "zai": "glm-5v-turbo",
 }
 
@@ -573,7 +580,8 @@ class _AnthropicCompletionsAdapter:
         self._is_oauth = is_oauth
 
     def create(self, **kwargs) -> Any:
-        from agent.anthropic_adapter import build_anthropic_kwargs, normalize_anthropic_response
+        from agent.anthropic_adapter import build_anthropic_kwargs
+        from agent.transports import get_transport
 
         messages = kwargs.get("messages", [])
         model = kwargs.get("model", self._model)
@@ -610,7 +618,19 @@ class _AnthropicCompletionsAdapter:
                 anthropic_kwargs["temperature"] = temperature
 
         response = self._client.messages.create(**anthropic_kwargs)
-        assistant_message, finish_reason = normalize_anthropic_response(response)
+        _transport = get_transport("anthropic_messages")
+        _nr = _transport.normalize_response(
+            response, strip_tool_prefix=self._is_oauth
+        )
+
+        # ToolCall already duck-types as OpenAI shape (.type, .function.name,
+        # .function.arguments) via properties, so no wrapping needed.
+        assistant_message = SimpleNamespace(
+            content=_nr.content,
+            tool_calls=_nr.tool_calls,
+            reasoning=_nr.reasoning,
+        )
+        finish_reason = _nr.finish_reason
 
         usage = None
         if hasattr(response, "usage") and response.usage:
@@ -903,6 +923,19 @@ def _try_openrouter() -> Tuple[Optional[OpenAI], Optional[str]]:
                    default_headers=_OR_HEADERS), _OPENROUTER_MODEL
 
 
+def _describe_openrouter_unavailable() -> str:
+    """Return a more precise OpenRouter auth failure reason for logs."""
+    pool_present, entry = _select_pool_entry("openrouter")
+    if pool_present:
+        if entry is None:
+            return "OpenRouter credential pool has no usable entries (credentials may be exhausted)"
+        if not _pool_runtime_api_key(entry):
+            return "OpenRouter credential pool entry is missing a runtime API key"
+    if not str(os.getenv("OPENROUTER_API_KEY") or "").strip():
+        return "OPENROUTER_API_KEY not set"
+    return "no usable OpenRouter credentials found"
+
+
 def _try_nous(vision: bool = False) -> Tuple[Optional[OpenAI], Optional[str]]:
     # Check cross-session rate limit guard before attempting Nous —
     # if another session already recorded a 429, skip Nous entirely
@@ -1316,6 +1349,68 @@ def _is_auth_error(exc: Exception) -> bool:
     return "error code: 401" in err_lower or "authenticationerror" in type(exc).__name__.lower()
 
 
+def _evict_cached_clients(provider: str) -> None:
+    """Drop cached auxiliary clients for a provider so fresh creds are used."""
+    normalized = _normalize_aux_provider(provider)
+    with _client_cache_lock:
+        stale_keys = [
+            key for key in _client_cache
+            if _normalize_aux_provider(str(key[0])) == normalized
+        ]
+        for key in stale_keys:
+            client = _client_cache.get(key, (None, None, None))[0]
+            if client is not None:
+                _force_close_async_httpx(client)
+                try:
+                    close_fn = getattr(client, "close", None)
+                    if callable(close_fn):
+                        close_fn()
+                except Exception:
+                    pass
+            _client_cache.pop(key, None)
+
+
+def _refresh_provider_credentials(provider: str) -> bool:
+    """Refresh short-lived credentials for OAuth-backed auxiliary providers."""
+    normalized = _normalize_aux_provider(provider)
+    try:
+        if normalized == "openai-codex":
+            from hermes_cli.auth import resolve_codex_runtime_credentials
+
+            creds = resolve_codex_runtime_credentials(force_refresh=True)
+            if not str(creds.get("api_key", "") or "").strip():
+                return False
+            _evict_cached_clients(normalized)
+            return True
+        if normalized == "nous":
+            from hermes_cli.auth import resolve_nous_runtime_credentials
+
+            creds = resolve_nous_runtime_credentials(
+                min_key_ttl_seconds=max(60, int(os.getenv("HERMES_NOUS_MIN_KEY_TTL_SECONDS", "1800"))),
+                timeout_seconds=float(os.getenv("HERMES_NOUS_TIMEOUT_SECONDS", "15")),
+                force_mint=True,
+            )
+            if not str(creds.get("api_key", "") or "").strip():
+                return False
+            _evict_cached_clients(normalized)
+            return True
+        if normalized == "anthropic":
+            from agent.anthropic_adapter import read_claude_code_credentials, _refresh_oauth_token, resolve_anthropic_token
+
+            creds = read_claude_code_credentials()
+            token = _refresh_oauth_token(creds) if isinstance(creds, dict) and creds.get("refreshToken") else None
+            if not str(token or "").strip():
+                token = resolve_anthropic_token()
+            if not str(token or "").strip():
+                return False
+            _evict_cached_clients(normalized)
+            return True
+    except Exception as exc:
+        logger.debug("Auxiliary provider credential refresh failed for %s: %s", normalized, exc)
+        return False
+    return False
+
+
 def _try_payment_fallback(
     failed_provider: str,
     task: str = None,
@@ -1614,8 +1709,10 @@ def resolve_provider_client(
     if provider == "openrouter":
         client, default = _try_openrouter()
         if client is None:
-            logger.warning("resolve_provider_client: openrouter requested "
-                           "but OPENROUTER_API_KEY not set")
+            logger.warning(
+                "resolve_provider_client: openrouter requested but %s",
+                _describe_openrouter_unavailable(),
+            )
             return None, None
         final_model = _normalize_resolved_model(model or default, provider)
         return (_to_async_client(client, final_model) if async_mode
@@ -1708,7 +1805,7 @@ def resolve_provider_client(
                        "but no endpoint credentials found")
         return None, None
 
-    # ── Named custom providers (config.yaml custom_providers list) ───
+    # ── Named custom providers (config.yaml providers dict / custom_providers list) ───
     try:
         from hermes_cli.runtime_provider import _get_named_custom_provider
         custom_entry = _get_named_custom_provider(provider)
@@ -1719,16 +1816,51 @@ def resolve_provider_client(
             if not custom_key and custom_key_env:
                 custom_key = os.getenv(custom_key_env, "").strip()
             custom_key = custom_key or "no-key-required"
+            # An explicit per-task api_mode override (from _resolve_task_provider_model)
+            # wins; otherwise fall back to what the provider entry declared.
+            entry_api_mode = (api_mode or custom_entry.get("api_mode") or "").strip()
             if custom_base:
                 final_model = _normalize_resolved_model(
                     model or custom_entry.get("model") or _read_main_model() or "gpt-4o-mini",
                     provider,
                 )
-                client = OpenAI(api_key=custom_key, base_url=custom_base)
-                client = _wrap_if_needed(client, final_model, custom_base)
                 logger.debug(
-                    "resolve_provider_client: named custom provider %r (%s)",
-                    provider, final_model)
+                    "resolve_provider_client: named custom provider %r (%s, api_mode=%s)",
+                    provider, final_model, entry_api_mode or "chat_completions")
+                # anthropic_messages: route through the Anthropic Messages API
+                # via AnthropicAuxiliaryClient. Mirrors the anonymous-custom
+                # branch in _try_custom_endpoint(). See #15033.
+                if entry_api_mode == "anthropic_messages":
+                    try:
+                        from agent.anthropic_adapter import build_anthropic_client
+                        real_client = build_anthropic_client(custom_key, custom_base)
+                    except ImportError:
+                        logger.warning(
+                            "Named custom provider %r declares api_mode="
+                            "anthropic_messages but the anthropic SDK is not "
+                            "installed — falling back to OpenAI-wire.",
+                            provider,
+                        )
+                        client = OpenAI(api_key=custom_key, base_url=custom_base)
+                        return (_to_async_client(client, final_model) if async_mode
+                                else (client, final_model))
+                    sync_anthropic = AnthropicAuxiliaryClient(
+                        real_client, final_model, custom_key, custom_base, is_oauth=False,
+                    )
+                    if async_mode:
+                        return AsyncAnthropicAuxiliaryClient(sync_anthropic), final_model
+                    return sync_anthropic, final_model
+                client = OpenAI(api_key=custom_key, base_url=custom_base)
+                # codex_responses or inherited auto-detect (via _wrap_if_needed).
+                # _wrap_if_needed reads the closed-over `api_mode` (the task-level
+                # override). Named-provider entry api_mode=codex_responses also
+                # flows through here.
+                if entry_api_mode == "codex_responses" and not isinstance(
+                    client, CodexAuxiliaryClient
+                ):
+                    client = CodexAuxiliaryClient(client, final_model)
+                else:
+                    client = _wrap_if_needed(client, final_model, custom_base)
                 return (_to_async_client(client, final_model) if async_mode
                         else (client, final_model))
             logger.warning(
@@ -2829,6 +2961,49 @@ def call_llm(
                 return _validate_llm_response(
                     refreshed_client.chat.completions.create(**kwargs), task)
 
+        # ── Auth refresh retry ───────────────────────────────────────
+        if (_is_auth_error(first_err)
+                and resolved_provider not in ("auto", "", None)
+                and not client_is_nous):
+            if _refresh_provider_credentials(resolved_provider):
+                logger.info(
+                    "Auxiliary %s: refreshed %s credentials after auth error, retrying",
+                    task or "call", resolved_provider,
+                )
+                retry_client, retry_model = (
+                    resolve_vision_provider_client(
+                        provider=resolved_provider,
+                        model=final_model,
+                        async_mode=False,
+                    )[1:]
+                    if task == "vision"
+                    else _get_cached_client(
+                        resolved_provider,
+                        resolved_model,
+                        base_url=resolved_base_url,
+                        api_key=resolved_api_key,
+                        api_mode=resolved_api_mode,
+                        main_runtime=main_runtime,
+                    )
+                )
+                if retry_client is not None:
+                    retry_kwargs = _build_call_kwargs(
+                        resolved_provider,
+                        retry_model or final_model,
+                        messages,
+                        temperature=temperature,
+                        max_tokens=max_tokens,
+                        tools=tools,
+                        timeout=effective_timeout,
+                        extra_body=effective_extra_body,
+                        base_url=resolved_base_url,
+                    )
+                    _retry_base = str(getattr(retry_client, "base_url", "") or "")
+                    if _is_anthropic_compat_endpoint(resolved_provider, _retry_base):
+                        retry_kwargs["messages"] = _convert_openai_images_to_anthropic(retry_kwargs["messages"])
+                    return _validate_llm_response(
+                        retry_client.chat.completions.create(**retry_kwargs), task)
+
         # ── Payment / credit exhaustion fallback ──────────────────────
         # When the resolved provider returns 402 or a credit-related error,
         # try alternative providers instead of giving up.  This handles the
@@ -3049,6 +3224,48 @@ async def async_call_llm(
                 return _validate_llm_response(
                     await refreshed_client.chat.completions.create(**kwargs), task)
 
+        # ── Auth refresh retry (mirrors sync call_llm) ───────────────
+        if (_is_auth_error(first_err)
+                and resolved_provider not in ("auto", "", None)
+                and not client_is_nous):
+            if _refresh_provider_credentials(resolved_provider):
+                logger.info(
+                    "Auxiliary %s (async): refreshed %s credentials after auth error, retrying",
+                    task or "call", resolved_provider,
+                )
+                if task == "vision":
+                    _, retry_client, retry_model = resolve_vision_provider_client(
+                        provider=resolved_provider,
+                        model=final_model,
+                        async_mode=True,
+                    )
+                else:
+                    retry_client, retry_model = _get_cached_client(
+                        resolved_provider,
+                        resolved_model,
+                        async_mode=True,
+                        base_url=resolved_base_url,
+                        api_key=resolved_api_key,
+                        api_mode=resolved_api_mode,
+                    )
+                if retry_client is not None:
+                    retry_kwargs = _build_call_kwargs(
+                        resolved_provider,
+                        retry_model or final_model,
+                        messages,
+                        temperature=temperature,
+                        max_tokens=max_tokens,
+                        tools=tools,
+                        timeout=effective_timeout,
+                        extra_body=effective_extra_body,
+                        base_url=resolved_base_url,
+                    )
+                    _retry_base = str(getattr(retry_client, "base_url", "") or "")
+                    if _is_anthropic_compat_endpoint(resolved_provider, _retry_base):
+                        retry_kwargs["messages"] = _convert_openai_images_to_anthropic(retry_kwargs["messages"])
+                    return _validate_llm_response(
+                        await retry_client.chat.completions.create(**retry_kwargs), task)
+
         # ── Payment / connection fallback (mirrors sync call_llm) ─────
         should_fallback = _is_payment_error(first_err) or _is_connection_error(first_err)
         is_auto = resolved_provider in ("auto", "", None)

agent/context_compressor.py

@@ -1099,6 +1099,21 @@ The user has requested that this compaction PRIORITISE preserving all informatio
 
         return max(cut_idx, head_end + 1)
 
+    # ------------------------------------------------------------------
+    # ContextEngine: manual /compress preflight
+    # ------------------------------------------------------------------
+
+    def has_content_to_compress(self, messages: List[Dict[str, Any]]) -> bool:
+        """Return True if there is a non-empty middle region to compact.
+
+        Overrides the ABC default so the gateway ``/compress`` guard can
+        skip the LLM call when the transcript is still entirely inside
+        the protected head/tail.
+        """
+        compress_start = self._align_boundary_forward(messages, self.protect_first_n)
+        compress_end = self._find_tail_cut_by_tokens(messages, compress_start)
+        return compress_start < compress_end
+
     # ------------------------------------------------------------------
     # Main compression entry point
     # ------------------------------------------------------------------

agent/context_engine.py

@@ -78,6 +78,7 @@ class ContextEngine(ABC):
         self,
         messages: List[Dict[str, Any]],
         current_tokens: int = None,
+        focus_topic: str = None,
     ) -> List[Dict[str, Any]]:
         """Compact the message list and return the new message list.
 
@@ -86,6 +87,12 @@ class ContextEngine(ABC):
         context budget. The implementation is free to summarize, build a
         DAG, or do anything else — as long as the returned list is a valid
         OpenAI-format message sequence.
+
+        Args:
+            focus_topic: Optional topic string from manual ``/compress <focus>``.
+                Engines that support guided compression should prioritise
+                preserving information related to this topic.  Engines that
+                don't support it may simply ignore this argument.
         """
 
     # -- Optional: pre-flight check ----------------------------------------
@@ -98,6 +105,21 @@ class ContextEngine(ABC):
         """
         return False
 
+    # -- Optional: manual /compress preflight ------------------------------
+
+    def has_content_to_compress(self, messages: List[Dict[str, Any]]) -> bool:
+        """Quick check: is there anything in ``messages`` that can be compacted?
+
+        Used by the gateway ``/compress`` command as a preflight guard —
+        returning False lets the gateway report "nothing to compress yet"
+        without making an LLM call.
+
+        Default returns True (always attempt).  Engines with a cheap way
+        to introspect their own head/tail boundaries should override this
+        to return False when the transcript is still entirely protected.
+        """
+        return True
+
     # -- Optional: session lifecycle ---------------------------------------
 
     def on_session_start(self, session_id: str, **kwargs) -> None:

agent/copilot_acp_client.py

@@ -46,6 +46,47 @@ def _resolve_args() -> list[str]:
     return shlex.split(raw)
 
 
+def _resolve_home_dir() -> str:
+    """Return a stable HOME for child ACP processes."""
+
+    try:
+        from hermes_constants import get_subprocess_home
+
+        profile_home = get_subprocess_home()
+        if profile_home:
+            return profile_home
+    except Exception:
+        pass
+
+    home = os.environ.get("HOME", "").strip()
+    if home:
+        return home
+
+    expanded = os.path.expanduser("~")
+    if expanded and expanded != "~":
+        return expanded
+
+    try:
+        import pwd
+
+        resolved = pwd.getpwuid(os.getuid()).pw_dir.strip()
+        if resolved:
+            return resolved
+    except Exception:
+        pass
+
+    # Last resort: /tmp (writable on any POSIX system). Avoids crashing the
+    # subprocess with no HOME; callers can set HERMES_HOME explicitly if they
+    # need a different writable dir.
+    return "/tmp"
+
+
+def _build_subprocess_env() -> dict[str, str]:
+    env = os.environ.copy()
+    env["HOME"] = _resolve_home_dir()
+    return env
+
+
 def _jsonrpc_error(message_id: Any, code: int, message: str) -> dict[str, Any]:
     return {
         "jsonrpc": "2.0",
@@ -382,6 +423,7 @@ class CopilotACPClient:
                 text=True,
                 bufsize=1,
                 cwd=self._acp_cwd,
+                env=_build_subprocess_env(),
             )
         except FileNotFoundError as exc:
             raise RuntimeError(

agent/credential_pool.py

@@ -455,6 +455,61 @@ class CredentialPool:
             logger.debug("Failed to sync from credentials file: %s", exc)
         return entry
 
+    def _sync_nous_entry_from_auth_store(self, entry: PooledCredential) -> PooledCredential:
+        """Sync a Nous pool entry from auth.json if tokens differ.
+
+        Nous OAuth refresh tokens are single-use.  When another process
+        (e.g. a concurrent cron) refreshes the token via
+        ``resolve_nous_runtime_credentials``, it writes fresh tokens to
+        auth.json under ``_auth_store_lock``.  The pool entry's tokens
+        become stale.  This method detects that and adopts the newer pair,
+        avoiding a "refresh token reuse" revocation on the Nous Portal.
+        """
+        if self.provider != "nous" or entry.source != "device_code":
+            return entry
+        try:
+            with _auth_store_lock():
+                auth_store = _load_auth_store()
+                state = _load_provider_state(auth_store, "nous")
+            if not state:
+                return entry
+            store_refresh = state.get("refresh_token", "")
+            store_access = state.get("access_token", "")
+            if store_refresh and store_refresh != entry.refresh_token:
+                logger.debug(
+                    "Pool entry %s: syncing tokens from auth.json (Nous refresh token changed)",
+                    entry.id,
+                )
+                field_updates: Dict[str, Any] = {
+                    "access_token": store_access,
+                    "refresh_token": store_refresh,
+                    "last_status": None,
+                    "last_status_at": None,
+                    "last_error_code": None,
+                }
+                if state.get("expires_at"):
+                    field_updates["expires_at"] = state["expires_at"]
+                if state.get("agent_key"):
+                    field_updates["agent_key"] = state["agent_key"]
+                if state.get("agent_key_expires_at"):
+                    field_updates["agent_key_expires_at"] = state["agent_key_expires_at"]
+                if state.get("inference_base_url"):
+                    field_updates["inference_base_url"] = state["inference_base_url"]
+                extra_updates = dict(entry.extra)
+                for extra_key in ("obtained_at", "expires_in", "agent_key_id",
+                                  "agent_key_expires_in", "agent_key_reused",
+                                  "agent_key_obtained_at"):
+                    val = state.get(extra_key)
+                    if val is not None:
+                        extra_updates[extra_key] = val
+                updated = replace(entry, extra=extra_updates, **field_updates)
+                self._replace_entry(entry, updated)
+                self._persist()
+                return updated
+        except Exception as exc:
+            logger.debug("Failed to sync Nous entry from auth.json: %s", exc)
+        return entry
+
     def _sync_device_code_entry_to_auth_store(self, entry: PooledCredential) -> None:
         """Write refreshed pool entry tokens back to auth.json providers.
 
@@ -561,6 +616,9 @@ class CredentialPool:
                     last_refresh=refreshed.get("last_refresh"),
                 )
             elif self.provider == "nous":
+                synced = self._sync_nous_entry_from_auth_store(entry)
+                if synced is not entry:
+                    entry = synced
                 nous_state = {
                     "access_token": entry.access_token,
                     "refresh_token": entry.refresh_token,
@@ -635,6 +693,26 @@ class CredentialPool:
                     # Credentials file had a valid (non-expired) token — use it directly
                     logger.debug("Credentials file has valid token, using without refresh")
                     return synced
+            # For nous: another process may have consumed the refresh token
+            # between our proactive sync and the HTTP call.  Re-sync from
+            # auth.json and adopt the fresh tokens if available.
+            if self.provider == "nous":
+                synced = self._sync_nous_entry_from_auth_store(entry)
+                if synced.refresh_token != entry.refresh_token:
+                    logger.debug("Nous refresh failed but auth.json has newer tokens — adopting")
+                    updated = replace(
+                        synced,
+                        last_status=STATUS_OK,
+                        last_status_at=None,
+                        last_error_code=None,
+                        last_error_reason=None,
+                        last_error_message=None,
+                        last_error_reset_at=None,
+                    )
+                    self._replace_entry(synced, updated)
+                    self._persist()
+                    self._sync_device_code_entry_to_auth_store(updated)
+                    return updated
             self._mark_exhausted(entry, None)
             return None
 
@@ -698,6 +776,17 @@ class CredentialPool:
                 if synced is not entry:
                     entry = synced
                     cleared_any = True
+            # For nous entries, sync from auth.json before status checks.
+            # Another process may have successfully refreshed via
+            # resolve_nous_runtime_credentials(), making this entry's
+            # exhausted status stale.
+            if (self.provider == "nous"
+                    and entry.source == "device_code"
+                    and entry.last_status == STATUS_EXHAUSTED):
+                synced = self._sync_nous_entry_from_auth_store(entry)
+                if synced is not entry:
+                    entry = synced
+                    cleared_any = True
             if entry.last_status == STATUS_EXHAUSTED:
                 exhausted_until = _exhausted_until(entry)
                 if exhausted_until is not None and now < exhausted_until:
@@ -739,8 +828,11 @@ class CredentialPool:
 
         if self._strategy == STRATEGY_LEAST_USED and len(available) > 1:
             entry = min(available, key=lambda e: e.request_count)
+            # Increment usage counter so subsequent selections distribute load
+            updated = replace(entry, request_count=entry.request_count + 1)
+            self._replace_entry(entry, updated)
             self._current_id = entry.id
-            return entry
+            return updated
 
         if self._strategy == STRATEGY_ROUND_ROBIN and len(available) > 1:
             entry = available[0]
@@ -1056,6 +1148,18 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
                     "inference_base_url": state.get("inference_base_url"),
                     "agent_key": state.get("agent_key"),
                     "agent_key_expires_at": state.get("agent_key_expires_at"),
+                    # Carry the mint/refresh timestamps into the pool so
+                    # freshness-sensitive consumers (self-heal hooks, pool
+                    # pruning by age) can distinguish just-minted credentials
+                    # from stale ones.  Without these, fresh device_code
+                    # entries get obtained_at=None and look older than they
+                    # are (#15099).
+                    "obtained_at": state.get("obtained_at"),
+                    "expires_in": state.get("expires_in"),
+                    "agent_key_id": state.get("agent_key_id"),
+                    "agent_key_expires_in": state.get("agent_key_expires_in"),
+                    "agent_key_reused": state.get("agent_key_reused"),
+                    "agent_key_obtained_at": state.get("agent_key_obtained_at"),
                     "tls": state.get("tls") if isinstance(state.get("tls"), dict) else None,
                     "label": seeded_label,
                 },
@@ -1066,9 +1170,10 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
         # env vars (COPILOT_GITHUB_TOKEN / GH_TOKEN).  They don't live in
         # the auth store or credential pool, so we resolve them here.
         try:
-            from hermes_cli.copilot_auth import resolve_copilot_token
+            from hermes_cli.copilot_auth import resolve_copilot_token, get_copilot_api_token
             token, source = resolve_copilot_token()
             if token:
+                api_token = get_copilot_api_token(token)
                 source_name = "gh_cli" if "gh" in source.lower() else f"env:{source}"
                 if not _is_suppressed(provider, source_name):
                     active_sources.add(source_name)
@@ -1080,7 +1185,7 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
                         {
                             "source": source_name,
                             "auth_type": AUTH_TYPE_API_KEY,
-                            "access_token": token,
+                            "access_token": api_token,
                             "base_url": pconfig.inference_base_url if pconfig else "",
                             "label": source,
                         },

agent/error_classifier.py

@@ -45,6 +45,7 @@ class FailoverReason(enum.Enum):
 
     # Model
     model_not_found = "model_not_found"  # 404 or invalid model — fallback to different model
+    provider_policy_blocked = "provider_policy_blocked"  # Aggregator (e.g. OpenRouter) blocked the only endpoint due to account data/privacy policy
 
     # Request format
     format_error = "format_error"        # 400 bad request — abort or strip + retry
@@ -194,6 +195,29 @@ _MODEL_NOT_FOUND_PATTERNS = [
     "unsupported model",
 ]
 
+# OpenRouter aggregator policy-block patterns.
+#
+# When a user's OpenRouter account privacy setting (or a per-request
+# `provider.data_collection: deny` preference) excludes the only endpoint
+# serving a model, OpenRouter returns 404 with a *specific* message that is
+# distinct from "model not found":
+#
+#   "No endpoints available matching your guardrail restrictions and
+#    data policy. Configure: https://openrouter.ai/settings/privacy"
+#
+# We classify this as `provider_policy_blocked` rather than
+# `model_not_found` because:
+#   - The model *exists* — model_not_found is misleading in logs
+#   - Provider fallback won't help: the account-level setting applies to
+#     every call on the same OpenRouter account
+#   - The error body already contains the fix URL, so the user gets
+#     actionable guidance without us rewriting the message
+_PROVIDER_POLICY_BLOCKED_PATTERNS = [
+    "no endpoints available matching your guardrail",
+    "no endpoints available matching your data policy",
+    "no endpoints found matching your data policy",
+]
+
 # Auth patterns (non-status-code signals)
 _AUTH_PATTERNS = [
     "invalid api key",
@@ -319,6 +343,11 @@ def classify_api_error(
     """
     status_code = _extract_status_code(error)
     error_type = type(error).__name__
+    # Copilot/GitHub Models RateLimitError may not set .status_code; force 429
+    # so downstream rate-limit handling (classifier reason, pool rotation,
+    # fallback gating) fires correctly instead of misclassifying as generic.
+    if status_code is None and error_type == "RateLimitError":
+        status_code = 429
     body = _extract_error_body(error)
     error_code = _extract_error_code(body)
 
@@ -523,6 +552,17 @@ def _classify_by_status(
         return _classify_402(error_msg, result_fn)
 
     if status_code == 404:
+        # OpenRouter policy-block 404 — distinct from "model not found".
+        # The model exists; the user's account privacy setting excludes the
+        # only endpoint serving it. Falling back to another provider won't
+        # help (same account setting applies).  The error body already
+        # contains the fix URL, so just surface it.
+        if any(p in error_msg for p in _PROVIDER_POLICY_BLOCKED_PATTERNS):
+            return result_fn(
+                FailoverReason.provider_policy_blocked,
+                retryable=False,
+                should_fallback=False,
+            )
         if any(p in error_msg for p in _MODEL_NOT_FOUND_PATTERNS):
             return result_fn(
                 FailoverReason.model_not_found,
@@ -640,6 +680,12 @@ def _classify_400(
         )
 
     # Some providers return model-not-found as 400 instead of 404 (e.g. OpenRouter).
+    if any(p in error_msg for p in _PROVIDER_POLICY_BLOCKED_PATTERNS):
+        return result_fn(
+            FailoverReason.provider_policy_blocked,
+            retryable=False,
+            should_fallback=False,
+        )
     if any(p in error_msg for p in _MODEL_NOT_FOUND_PATTERNS):
         return result_fn(
             FailoverReason.model_not_found,
@@ -812,6 +858,15 @@ def _classify_by_message(
             should_fallback=True,
         )
 
+    # Provider policy-block (aggregator-side guardrail) — check before
+    # model_not_found so we don't mis-label as a missing model.
+    if any(p in error_msg for p in _PROVIDER_POLICY_BLOCKED_PATTERNS):
+        return result_fn(
+            FailoverReason.provider_policy_blocked,
+            retryable=False,
+            should_fallback=False,
+        )
+
     # Model not found patterns
     if any(p in error_msg for p in _MODEL_NOT_FOUND_PATTERNS):
         return result_fn(

agent/gemini_native_adapter.py

@@ -44,6 +44,97 @@ def is_native_gemini_base_url(base_url: str) -> bool:
     return not normalized.endswith("/openai")
 
 
+def probe_gemini_tier(
+    api_key: str,
+    base_url: str = DEFAULT_GEMINI_BASE_URL,
+    *,
+    model: str = "gemini-2.5-flash",
+    timeout: float = 10.0,
+) -> str:
+    """Probe a Google AI Studio API key and return its tier.
+
+    Returns one of:
+
+    - ``"free"``    -- key is on the free tier (unusable with Hermes)
+    - ``"paid"``    -- key is on a paid tier
+    - ``"unknown"`` -- probe failed; callers should proceed without blocking.
+    """
+    key = (api_key or "").strip()
+    if not key:
+        return "unknown"
+
+    normalized_base = str(base_url or DEFAULT_GEMINI_BASE_URL).strip().rstrip("/")
+    if not normalized_base:
+        normalized_base = DEFAULT_GEMINI_BASE_URL
+    if normalized_base.lower().endswith("/openai"):
+        normalized_base = normalized_base[: -len("/openai")]
+
+    url = f"{normalized_base}/models/{model}:generateContent"
+    payload = {
+        "contents": [{"role": "user", "parts": [{"text": "hi"}]}],
+        "generationConfig": {"maxOutputTokens": 1},
+    }
+
+    try:
+        with httpx.Client(timeout=timeout) as client:
+            resp = client.post(
+                url,
+                params={"key": key},
+                json=payload,
+                headers={"Content-Type": "application/json"},
+            )
+    except Exception as exc:
+        logger.debug("probe_gemini_tier: network error: %s", exc)
+        return "unknown"
+
+    headers_lower = {k.lower(): v for k, v in resp.headers.items()}
+    rpd_header = headers_lower.get("x-ratelimit-limit-requests-per-day")
+    if rpd_header:
+        try:
+            rpd_val = int(rpd_header)
+        except (TypeError, ValueError):
+            rpd_val = None
+        # Published free-tier daily caps (Dec 2025):
+        #   gemini-2.5-pro: 100, gemini-2.5-flash: 250, flash-lite: 1000
+        # Tier 1 starts at ~1500+ for Flash. We treat <= 1000 as free.
+        if rpd_val is not None and rpd_val <= 1000:
+            return "free"
+        if rpd_val is not None and rpd_val > 1000:
+            return "paid"
+
+    if resp.status_code == 429:
+        body_text = ""
+        try:
+            body_text = resp.text or ""
+        except Exception:
+            body_text = ""
+        if "free_tier" in body_text.lower():
+            return "free"
+        return "paid"
+
+    if 200 <= resp.status_code < 300:
+        return "paid"
+
+    return "unknown"
+
+
+def is_free_tier_quota_error(error_message: str) -> bool:
+    """Return True when a Gemini 429 message indicates free-tier exhaustion."""
+    if not error_message:
+        return False
+    return "free_tier" in error_message.lower()
+
+
+_FREE_TIER_GUIDANCE = (
+    "\n\nYour Google API key is on the free tier (<= 250 requests/day for "
+    "gemini-2.5-flash). Hermes typically makes 3-10 API calls per user turn, "
+    "so the free tier is exhausted in a handful of messages and cannot sustain "
+    "an agent session. Enable billing on your Google Cloud project and "
+    "regenerate the key in a billing-enabled project: "
+    "https://aistudio.google.com/apikey"
+)
+
+
 class GeminiAPIError(Exception):
     """Error shape compatible with Hermes retry/error classification."""
 
@@ -650,6 +741,12 @@ def gemini_http_error(response: httpx.Response) -> GeminiAPIError:
     else:
         message = f"Gemini returned HTTP {status}: {body_text[:500]}"
 
+    # Free-tier quota exhaustion -> append actionable guidance so users who
+    # bypassed the setup wizard (direct GOOGLE_API_KEY in .env) still learn
+    # that the free tier cannot sustain an agent session.
+    if status == 429 and is_free_tier_quota_error(err_message or body_text):
+        message = message + _FREE_TIER_GUIDANCE
+
     return GeminiAPIError(
         message,
         code=code,
@@ -704,6 +801,13 @@ class GeminiNativeClient:
         http_client: Optional[httpx.Client] = None,
         **_: Any,
     ) -> None:
+        if not (api_key or "").strip():
+            raise RuntimeError(
+                "Gemini native client requires an API key, but none was provided. "
+                "Set GOOGLE_API_KEY or GEMINI_API_KEY in your environment / ~/.hermes/.env "
+                "(get one at https://aistudio.google.com/app/apikey), or run `hermes setup` "
+                "to configure the Google provider."
+            )
         self.api_key = api_key
         normalized_base = (base_url or DEFAULT_GEMINI_BASE_URL).rstrip("/")
         if normalized_base.endswith("/openai"):

agent/gemini_schema.py

@@ -73,6 +73,20 @@ def sanitize_gemini_schema(schema: Any) -> Dict[str, Any]:
             ]
             continue
         cleaned[key] = value
+
+    # Gemini's Schema validator requires every ``enum`` entry to be a string,
+    # even when the parent ``type`` is ``integer`` / ``number`` / ``boolean``.
+    # OpenAI / OpenRouter / Anthropic accept typed enums (e.g. Discord's
+    # ``auto_archive_duration: {type: integer, enum: [60, 1440, 4320, 10080]}``),
+    # so we only drop the ``enum`` when it would collide with Gemini's rule.
+    # Keeping ``type: integer`` plus the human-readable description gives the
+    # model enough guidance; the tool handler still validates the value.
+    enum_val = cleaned.get("enum")
+    type_val = cleaned.get("type")
+    if isinstance(enum_val, list) and type_val in {"integer", "number", "boolean"}:
+        if any(not isinstance(item, str) for item in enum_val):
+            cleaned.pop("enum", None)
+
     return cleaned
 
 

agent/model_metadata.py

@@ -6,6 +6,7 @@ and run_agent.py for pre-flight context checks.
 
 import ipaddress
 import logging
+import os
 import re
 import time
 from pathlib import Path
@@ -21,6 +22,25 @@ from hermes_constants import OPENROUTER_MODELS_URL
 
 logger = logging.getLogger(__name__)
 
+
+def _resolve_requests_verify() -> bool | str:
+    """Resolve SSL verify setting for `requests` calls from env vars.
+
+    The `requests` library only honours REQUESTS_CA_BUNDLE / CURL_CA_BUNDLE
+    by default. Hermes also honours HERMES_CA_BUNDLE (its own convention)
+    and SSL_CERT_FILE (used by the stdlib `ssl` module and by httpx), so
+    that a single env var can cover both `requests` and `httpx` callsites
+    inside the same process.
+
+    Returns either a filesystem path to a CA bundle, or True to defer to
+    the requests default (certifi).
+    """
+    for env_var in ("HERMES_CA_BUNDLE", "REQUESTS_CA_BUNDLE", "SSL_CERT_FILE"):
+        val = os.getenv(env_var)
+        if val and os.path.isfile(val):
+            return val
+    return True
+
 # Provider names that can appear as a "provider:" prefix before a model ID.
 # Only these are stripped — Ollama-style "model:tag" colons (e.g. "qwen3.5:27b")
 # are preserved so the full model name reaches cache lookups and server queries.
@@ -123,6 +143,10 @@ DEFAULT_CONTEXT_LENGTHS = {
     "claude": 200000,
     # OpenAI — GPT-5 family (most have 400k; specific overrides first)
     # Source: https://developers.openai.com/api/docs/models
+    # GPT-5.5 (launched Apr 23 2026). 400k is the fallback for providers we
+    # can't probe live. ChatGPT Codex OAuth actually caps lower (272k as of
+    # Apr 2026) and is resolved via _resolve_codex_oauth_context_length().
+    "gpt-5.5": 400000,
     "gpt-5.4-nano": 400000,           # 400k (not 1.05M like full 5.4)
     "gpt-5.4-mini": 400000,           # 400k (not 1.05M like full 5.4)
     "gpt-5.4": 1050000,               # GPT-5.4, GPT-5.4 Pro (1.05M context)
@@ -183,12 +207,12 @@ DEFAULT_CONTEXT_LENGTHS = {
     "moonshotai/Kimi-K2.6": 262144,
     "moonshotai/Kimi-K2-Thinking": 262144,
     "MiniMaxAI/MiniMax-M2.5": 204800,
-    "XiaomiMiMo/MiMo-V2-Flash": 256000,
-    "mimo-v2-pro": 1000000,
-    "mimo-v2-omni": 256000,
-    "mimo-v2-flash": 256000,
-    "mimo-v2.5-pro": 1000000,
-    "mimo-v2.5": 1000000,
+    "XiaomiMiMo/MiMo-V2-Flash": 262144,
+    "mimo-v2-pro": 1048576,
+    "mimo-v2.5-pro": 1048576,
+    "mimo-v2.5": 1048576,
+    "mimo-v2-omni": 262144,
+    "mimo-v2-flash": 262144,
     "zai-org/GLM-5": 202752,
 }
 
@@ -491,7 +515,7 @@ def fetch_model_metadata(force_refresh: bool = False) -> Dict[str, Dict[str, Any
         return _model_metadata_cache
 
     try:
-        response = requests.get(OPENROUTER_MODELS_URL, timeout=10)
+        response = requests.get(OPENROUTER_MODELS_URL, timeout=10, verify=_resolve_requests_verify())
         response.raise_for_status()
         data = response.json()
 
@@ -558,6 +582,7 @@ def fetch_endpoint_model_metadata(
                     server_url.rstrip("/") + "/api/v1/models",
                     headers=headers,
                     timeout=10,
+                    verify=_resolve_requests_verify(),
                 )
                 response.raise_for_status()
                 payload = response.json()
@@ -606,7 +631,7 @@ def fetch_endpoint_model_metadata(
     for candidate in candidates:
         url = candidate.rstrip("/") + "/models"
         try:
-            response = requests.get(url, headers=headers, timeout=10)
+            response = requests.get(url, headers=headers, timeout=10, verify=_resolve_requests_verify())
             response.raise_for_status()
             payload = response.json()
             cache: Dict[str, Dict[str, Any]] = {}
@@ -637,9 +662,10 @@ def fetch_endpoint_model_metadata(
                 try:
                     # Try /v1/props first (current llama.cpp); fall back to /props for older builds
                     base = candidate.rstrip("/").replace("/v1", "")
-                    props_resp = requests.get(base + "/v1/props", headers=headers, timeout=5)
+                    _verify = _resolve_requests_verify()
+                    props_resp = requests.get(base + "/v1/props", headers=headers, timeout=5, verify=_verify)
                     if not props_resp.ok:
-                        props_resp = requests.get(base + "/props", headers=headers, timeout=5)
+                        props_resp = requests.get(base + "/props", headers=headers, timeout=5, verify=_verify)
                     if props_resp.ok:
                         props = props_resp.json()
                         gen_settings = props.get("default_generation_settings", {})
@@ -711,6 +737,22 @@ def get_cached_context_length(model: str, base_url: str) -> Optional[int]:
     return cache.get(key)
 
 
+def _invalidate_cached_context_length(model: str, base_url: str) -> None:
+    """Drop a stale cache entry so it gets re-resolved on the next lookup."""
+    key = f"{model}@{base_url}"
+    cache = _load_context_cache()
+    if key not in cache:
+        return
+    del cache[key]
+    path = _get_context_cache_path()
+    try:
+        path.parent.mkdir(parents=True, exist_ok=True)
+        with open(path, "w") as f:
+            yaml.dump({"context_lengths": cache}, f, default_flow_style=False)
+    except Exception as e:
+        logger.debug("Failed to invalidate context length cache entry %s: %s", key, e)
+
+
 def get_next_probe_tier(current_length: int) -> Optional[int]:
     """Return the next lower probe tier, or None if already at minimum."""
     for tier in CONTEXT_PROBE_TIERS:
@@ -988,7 +1030,7 @@ def _query_anthropic_context_length(model: str, base_url: str, api_key: str) ->
             "x-api-key": api_key,
             "anthropic-version": "2023-06-01",
         }
-        resp = requests.get(url, headers=headers, timeout=10)
+        resp = requests.get(url, headers=headers, timeout=10, verify=_resolve_requests_verify())
         if resp.status_code != 200:
             return None
         data = resp.json()
@@ -1002,6 +1044,116 @@ def _query_anthropic_context_length(model: str, base_url: str, api_key: str) ->
     return None
 
 
+# Known ChatGPT Codex OAuth context windows (observed via live
+# chatgpt.com/backend-api/codex/models probe, Apr 2026). These are the
+# `context_window` values, which are what Codex actually enforces — the
+# direct OpenAI API has larger limits for the same slugs, but Codex OAuth
+# caps lower (e.g. gpt-5.5 is 1.05M on the API, 272K on Codex).
+#
+# Used as a fallback when the live probe fails (no token, network error).
+# Longest keys first so substring match picks the most specific entry.
+_CODEX_OAUTH_CONTEXT_FALLBACK: Dict[str, int] = {
+    "gpt-5.1-codex-max": 272_000,
+    "gpt-5.1-codex-mini": 272_000,
+    "gpt-5.3-codex": 272_000,
+    "gpt-5.2-codex": 272_000,
+    "gpt-5.4-mini": 272_000,
+    "gpt-5.5": 272_000,
+    "gpt-5.4": 272_000,
+    "gpt-5.2": 272_000,
+    "gpt-5": 272_000,
+}
+
+
+_codex_oauth_context_cache: Dict[str, int] = {}
+_codex_oauth_context_cache_time: float = 0.0
+_CODEX_OAUTH_CONTEXT_CACHE_TTL = 3600  # 1 hour
+
+
+def _fetch_codex_oauth_context_lengths(access_token: str) -> Dict[str, int]:
+    """Probe the ChatGPT Codex /models endpoint for per-slug context windows.
+
+    Codex OAuth imposes its own context limits that differ from the direct
+    OpenAI API (e.g. gpt-5.5 is 1.05M on the API, 272K on Codex). The
+    `context_window` field in each model entry is the authoritative source.
+
+    Returns a ``{slug: context_window}`` dict. Empty on failure.
+    """
+    global _codex_oauth_context_cache, _codex_oauth_context_cache_time
+    now = time.time()
+    if (
+        _codex_oauth_context_cache
+        and now - _codex_oauth_context_cache_time < _CODEX_OAUTH_CONTEXT_CACHE_TTL
+    ):
+        return _codex_oauth_context_cache
+
+    try:
+        resp = requests.get(
+            "https://chatgpt.com/backend-api/codex/models?client_version=1.0.0",
+            headers={"Authorization": f"Bearer {access_token}"},
+            timeout=10,
+            verify=_resolve_requests_verify(),
+        )
+        if resp.status_code != 200:
+            logger.debug(
+                "Codex /models probe returned HTTP %s; falling back to hardcoded defaults",
+                resp.status_code,
+            )
+            return {}
+        data = resp.json()
+    except Exception as exc:
+        logger.debug("Codex /models probe failed: %s", exc)
+        return {}
+
+    entries = data.get("models", []) if isinstance(data, dict) else []
+    result: Dict[str, int] = {}
+    for item in entries:
+        if not isinstance(item, dict):
+            continue
+        slug = item.get("slug")
+        ctx = item.get("context_window")
+        if isinstance(slug, str) and isinstance(ctx, int) and ctx > 0:
+            result[slug.strip()] = ctx
+
+    if result:
+        _codex_oauth_context_cache = result
+        _codex_oauth_context_cache_time = now
+    return result
+
+
+def _resolve_codex_oauth_context_length(
+    model: str, access_token: str = ""
+) -> Optional[int]:
+    """Resolve a Codex OAuth model's real context window.
+
+    Prefers a live probe of chatgpt.com/backend-api/codex/models (when we
+    have a bearer token), then falls back to ``_CODEX_OAUTH_CONTEXT_FALLBACK``.
+    """
+    model_bare = _strip_provider_prefix(model).strip()
+    if not model_bare:
+        return None
+
+    if access_token:
+        live = _fetch_codex_oauth_context_lengths(access_token)
+        if model_bare in live:
+            return live[model_bare]
+        # Case-insensitive match in case casing drifts
+        model_lower = model_bare.lower()
+        for slug, ctx in live.items():
+            if slug.lower() == model_lower:
+                return ctx
+
+    # Fallback: longest-key-first substring match over hardcoded defaults.
+    model_lower = model_bare.lower()
+    for slug, ctx in sorted(
+        _CODEX_OAUTH_CONTEXT_FALLBACK.items(), key=lambda x: len(x[0]), reverse=True
+    ):
+        if slug in model_lower:
+            return ctx
+
+    return None
+
+
 def _resolve_nous_context_length(model: str) -> Optional[int]:
     """Resolve Nous Portal model context length via OpenRouter metadata.
 
@@ -1069,7 +1221,21 @@ def get_model_context_length(
     if base_url:
         cached = get_cached_context_length(model, base_url)
         if cached is not None:
-            return cached
+            # Invalidate stale Codex OAuth cache entries: pre-PR #14935 builds
+            # resolved gpt-5.x to the direct-API value (e.g. 1.05M) via
+            # models.dev and persisted it. Codex OAuth caps at 272K for every
+            # slug, so any cached Codex entry at or above 400K is a leftover
+            # from the old resolution path. Drop it and fall through to the
+            # live /models probe in step 5 below.
+            if provider == "openai-codex" and cached >= 400_000:
+                logger.info(
+                    "Dropping stale Codex cache entry %s@%s -> %s (pre-fix value); "
+                    "re-resolving via live /models probe",
+                    model, base_url, f"{cached:,}",
+                )
+                _invalidate_cached_context_length(model, base_url)
+            else:
+                return cached
 
     # 2. Active endpoint metadata for truly custom/unknown endpoints.
     # Known providers (Copilot, OpenAI, Anthropic, etc.) skip this — their
@@ -1142,10 +1308,32 @@ def get_model_context_length(
             if inferred:
                 effective_provider = inferred
 
+    # 5a. Copilot live /models API — max_prompt_tokens from the user's account.
+    # This catches account-specific models (e.g. claude-opus-4.6-1m) that
+    # don't exist in models.dev. For models that ARE in models.dev, this
+    # returns the provider-enforced limit which is what users can actually use.
+    if effective_provider in ("copilot", "copilot-acp", "github-copilot"):
+        try:
+            from hermes_cli.models import get_copilot_model_context
+            ctx = get_copilot_model_context(model, api_key=api_key)
+            if ctx:
+                return ctx
+        except Exception:
+            pass  # Fall through to models.dev
+
     if effective_provider == "nous":
         ctx = _resolve_nous_context_length(model)
         if ctx:
             return ctx
+    if effective_provider == "openai-codex":
+        # Codex OAuth enforces lower context limits than the direct OpenAI
+        # API for the same slug (e.g. gpt-5.5 is 1.05M on the API but 272K
+        # on Codex). Authoritative source is Codex's own /models endpoint.
+        codex_ctx = _resolve_codex_oauth_context_length(model, access_token=api_key or "")
+        if codex_ctx:
+            if base_url:
+                save_context_length(model, base_url, codex_ctx)
+            return codex_ctx
     if effective_provider:
         from agent.models_dev import lookup_models_dev_context
         ctx = lookup_models_dev_context(effective_provider, model)

agent/models_dev.py

@@ -418,6 +418,9 @@ def list_provider_models(provider: str) -> List[str]:
 
     Returns an empty list if the provider is unknown or has no data.
     """
+    from hermes_cli.models import normalize_provider
+    provider = normalize_provider(provider) or provider
+    
     models = _get_provider_models(provider)
     if models is None:
         return []

agent/moonshot_schema.py

@@ -0,0 +1,190 @@
+"""Helpers for translating OpenAI-style tool schemas to Moonshot's schema subset.
+
+Moonshot (Kimi) accepts a stricter subset of JSON Schema than standard OpenAI
+tool calling.  Requests that violate it fail with HTTP 400:
+
+    tools.function.parameters is not a valid moonshot flavored json schema,
+    details: <...>
+
+Known rejection modes documented at
+https://forum.moonshot.ai/t/tool-calling-specification-violation-on-moonshot-api/102
+and MoonshotAI/kimi-cli#1595:
+
+1. Every property schema must carry a ``type``.  Standard JSON Schema allows
+   type to be omitted (the value is then unconstrained); Moonshot refuses.
+2. When ``anyOf`` is used, ``type`` must be on the ``anyOf`` children, not
+   the parent.  Presence of both causes "type should be defined in anyOf
+   items instead of the parent schema".
+
+The ``#/definitions/...`` → ``#/$defs/...`` rewrite for draft-07 refs is
+handled separately in ``tools/mcp_tool._normalize_mcp_input_schema`` so it
+applies at MCP registration time for all providers.
+"""
+
+from __future__ import annotations
+
+import copy
+from typing import Any, Dict, List
+
+# Keys whose values are maps of name → schema (not schemas themselves).
+# When we recurse, we walk the values of these maps as schemas, but we do
+# NOT apply the missing-type repair to the map itself.
+_SCHEMA_MAP_KEYS = frozenset({"properties", "patternProperties", "$defs", "definitions"})
+
+# Keys whose values are lists of schemas.
+_SCHEMA_LIST_KEYS = frozenset({"anyOf", "oneOf", "allOf", "prefixItems"})
+
+# Keys whose values are a single nested schema.
+_SCHEMA_NODE_KEYS = frozenset({"items", "contains", "not", "additionalProperties", "propertyNames"})
+
+
+def _repair_schema(node: Any, is_schema: bool = True) -> Any:
+    """Recursively apply Moonshot repairs to a schema node.
+
+    ``is_schema=True`` means this dict is a JSON Schema node and gets the
+    missing-type + anyOf-parent repairs applied.  ``is_schema=False`` means
+    it's a container map (e.g. the value of ``properties``) and we only
+    recurse into its values.
+    """
+    if isinstance(node, list):
+        # Lists only show up under schema-list keys (anyOf/oneOf/allOf), so
+        # every element is itself a schema.
+        return [_repair_schema(item, is_schema=True) for item in node]
+    if not isinstance(node, dict):
+        return node
+
+    # Walk the dict, deciding per-key whether recursion is into a schema
+    # node, a container map, or a scalar.
+    repaired: Dict[str, Any] = {}
+    for key, value in node.items():
+        if key in _SCHEMA_MAP_KEYS and isinstance(value, dict):
+            # Map of name → schema.  Don't treat the map itself as a schema
+            # (it has no type / properties of its own), but each value is.
+            repaired[key] = {
+                sub_key: _repair_schema(sub_val, is_schema=True)
+                for sub_key, sub_val in value.items()
+            }
+        elif key in _SCHEMA_LIST_KEYS and isinstance(value, list):
+            repaired[key] = [_repair_schema(v, is_schema=True) for v in value]
+        elif key in _SCHEMA_NODE_KEYS:
+            # items / not / additionalProperties: single nested schema.
+            # additionalProperties can also be a bool — leave those alone.
+            if isinstance(value, dict):
+                repaired[key] = _repair_schema(value, is_schema=True)
+            else:
+                repaired[key] = value
+        else:
+            # Scalars (description, title, format, enum values, etc.) pass through.
+            repaired[key] = value
+
+    if not is_schema:
+        return repaired
+
+    # Rule 2: when anyOf is present, type belongs only on the children.
+    if "anyOf" in repaired and isinstance(repaired["anyOf"], list):
+        repaired.pop("type", None)
+        return repaired
+
+    # Rule 1: property schemas without type need one.  $ref nodes are exempt
+    # — their type comes from the referenced definition.
+    if "$ref" in repaired:
+        return repaired
+    return _fill_missing_type(repaired)
+
+
+def _fill_missing_type(node: Dict[str, Any]) -> Dict[str, Any]:
+    """Infer a reasonable ``type`` if this schema node has none."""
+    if "type" in node and node["type"] not in (None, ""):
+        return node
+
+    # Heuristic: presence of ``properties`` → object, ``items`` → array, ``enum``
+    # → type of first enum value, else fall back to ``string`` (safest scalar).
+    if "properties" in node or "required" in node or "additionalProperties" in node:
+        inferred = "object"
+    elif "items" in node or "prefixItems" in node:
+        inferred = "array"
+    elif "enum" in node and isinstance(node["enum"], list) and node["enum"]:
+        sample = node["enum"][0]
+        if isinstance(sample, bool):
+            inferred = "boolean"
+        elif isinstance(sample, int):
+            inferred = "integer"
+        elif isinstance(sample, float):
+            inferred = "number"
+        else:
+            inferred = "string"
+    else:
+        inferred = "string"
+
+    return {**node, "type": inferred}
+
+
+def sanitize_moonshot_tool_parameters(parameters: Any) -> Dict[str, Any]:
+    """Normalize tool parameters to a Moonshot-compatible object schema.
+
+    Returns a deep-copied schema with the two flavored-JSON-Schema repairs
+    applied.  Input is not mutated.
+    """
+    if not isinstance(parameters, dict):
+        return {"type": "object", "properties": {}}
+
+    repaired = _repair_schema(copy.deepcopy(parameters), is_schema=True)
+    if not isinstance(repaired, dict):
+        return {"type": "object", "properties": {}}
+
+    # Top-level must be an object schema
+    if repaired.get("type") != "object":
+        repaired["type"] = "object"
+    if "properties" not in repaired:
+        repaired["properties"] = {}
+
+    return repaired
+
+
+def sanitize_moonshot_tools(tools: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    """Apply ``sanitize_moonshot_tool_parameters`` to every tool's parameters."""
+    if not tools:
+        return tools
+
+    sanitized: List[Dict[str, Any]] = []
+    any_change = False
+    for tool in tools:
+        if not isinstance(tool, dict):
+            sanitized.append(tool)
+            continue
+        fn = tool.get("function")
+        if not isinstance(fn, dict):
+            sanitized.append(tool)
+            continue
+        params = fn.get("parameters")
+        repaired = sanitize_moonshot_tool_parameters(params)
+        if repaired is not params:
+            any_change = True
+            new_fn = {**fn, "parameters": repaired}
+            sanitized.append({**tool, "function": new_fn})
+        else:
+            sanitized.append(tool)
+
+    return sanitized if any_change else tools
+
+
+def is_moonshot_model(model: str | None) -> bool:
+    """True for any Kimi / Moonshot model slug, regardless of aggregator prefix.
+
+    Matches bare names (``kimi-k2.6``, ``moonshotai/Kimi-K2.6``) and aggregator-
+    prefixed slugs (``nous/moonshotai/kimi-k2.6``, ``openrouter/moonshotai/...``).
+    Detection by model name covers Nous / OpenRouter / other aggregators that
+    route to Moonshot's inference, where the base URL is the aggregator's, not
+    ``api.moonshot.ai``.
+    """
+    if not model:
+        return False
+    bare = model.strip().lower()
+    # Last path segment (covers aggregator-prefixed slugs)
+    tail = bare.rsplit("/", 1)[-1]
+    if tail.startswith("kimi-") or tail == "kimi":
+        return True
+    # Vendor-prefixed forms commonly used on aggregators
+    if "moonshot" in bare or "/kimi" in bare or bare.startswith("kimi"):
+        return True
+    return False

agent/prompt_builder.py

@@ -370,6 +370,32 @@ PLATFORM_HINTS = {
         "MEDIA:/absolute/path/to/file in your response. Images (.jpg, .png, "
         ".heic) appear as photos and other files arrive as attachments."
     ),
+    "mattermost": (
+        "You are in a Mattermost workspace communicating with your user. "
+        "Mattermost renders standard Markdown — headings, bold, italic, code "
+        "blocks, and tables all work. "
+        "You can send media files natively: include MEDIA:/absolute/path/to/file "
+        "in your response. Images (.jpg, .png, .webp) are uploaded as photo "
+        "attachments, audio and video as file attachments. "
+        "Image URLs in markdown format ![alt](url) are rendered as inline previews automatically."
+    ),
+    "matrix": (
+        "You are in a Matrix room communicating with your user. "
+        "Matrix renders Markdown — bold, italic, code blocks, and links work; "
+        "the adapter converts your Markdown to HTML for rich display. "
+        "You can send media files natively: include MEDIA:/absolute/path/to/file "
+        "in your response. Images (.jpg, .png, .webp) are sent as inline photos, "
+        "audio (.ogg, .mp3) as voice/audio messages, video (.mp4) inline, "
+        "and other files as downloadable attachments."
+    ),
+    "feishu": (
+        "You are in a Feishu (Lark) workspace communicating with your user. "
+        "Feishu renders Markdown in messages — bold, italic, code blocks, and "
+        "links are supported. "
+        "You can send media files natively: include MEDIA:/absolute/path/to/file "
+        "in your response. Images (.jpg, .png, .webp) are uploaded and displayed "
+        "inline, audio files as voice messages, and other files as attachments."
+    ),
     "weixin": (
         "You are on Weixin/WeChat. Markdown formatting is supported, so you may use it when "
         "it improves readability, but keep the message compact and chat-friendly. You can send media files natively: "

agent/skill_commands.py

@@ -1,15 +1,13 @@
-"""Shared slash command helpers for skills and built-in prompt-style modes.
+"""Shared slash command helpers for skills.
 
 Shared between CLI (cli.py) and gateway (gateway/run.py) so both surfaces
-can invoke skills via /skill-name commands and prompt-only built-ins like
-/plan.
+can invoke skills via /skill-name commands.
 """
 
 import json
 import logging
 import re
 import subprocess
-from datetime import datetime
 from pathlib import Path
 from typing import Any, Dict, Optional
 
@@ -18,7 +16,6 @@ from hermes_constants import display_hermes_home
 logger = logging.getLogger(__name__)
 
 _skill_commands: Dict[str, Dict[str, Any]] = {}
-_PLAN_SLUG_RE = re.compile(r"[^a-z0-9]+")
 # Patterns for sanitizing skill names into clean hyphen-separated slugs.
 _SKILL_INVALID_CHARS = re.compile(r"[^a-z0-9-]")
 _SKILL_MULTI_HYPHEN = re.compile(r"-{2,}")
@@ -128,27 +125,6 @@ def _expand_inline_shell(
     return _INLINE_SHELL_RE.sub(_replace, content)
 
 
-def build_plan_path(
-    user_instruction: str = "",
-    *,
-    now: datetime | None = None,
-) -> Path:
-    """Return the default workspace-relative markdown path for a /plan invocation.
-
-    Relative paths are intentional: file tools are task/backend-aware and resolve
-    them against the active working directory for local, docker, ssh, modal,
-    daytona, and similar terminal backends. That keeps the plan with the active
-    workspace instead of the Hermes host's global home directory.
-    """
-    slug_source = (user_instruction or "").strip().splitlines()[0] if user_instruction else ""
-    slug = _PLAN_SLUG_RE.sub("-", slug_source.lower()).strip("-")
-    if slug:
-        slug = "-".join(part for part in slug.split("-")[:8] if part)[:48].strip("-")
-    slug = slug or "conversation-plan"
-    timestamp = (now or datetime.now()).strftime("%Y-%m-%d_%H%M%S")
-    return Path(".hermes") / "plans" / f"{timestamp}-{slug}.md"
-
-
 def _load_skill_payload(skill_identifier: str, task_id: str | None = None) -> tuple[dict[str, Any], Path | None, str] | None:
     """Load a skill by name/path and return (loaded_payload, skill_dir, display_name)."""
     raw_identifier = (skill_identifier or "").strip()
@@ -345,7 +321,7 @@ def scan_skill_commands() -> Dict[str, Dict[str, Any]]:
     _skill_commands = {}
     try:
         from tools.skills_tool import SKILLS_DIR, _parse_frontmatter, skill_matches_platform, _get_disabled_skill_names
-        from agent.skill_utils import get_external_skills_dirs
+        from agent.skill_utils import get_external_skills_dirs, iter_skill_index_files
         disabled = _get_disabled_skill_names()
         seen_names: set = set()
 
@@ -356,7 +332,7 @@ def scan_skill_commands() -> Dict[str, Dict[str, Any]]:
         dirs_to_scan.extend(get_external_skills_dirs())
 
         for scan_dir in dirs_to_scan:
-            for skill_md in scan_dir.rglob("SKILL.md"):
+            for skill_md in iter_skill_index_files(scan_dir, "SKILL.md"):
                 if any(part in ('.git', '.github', '.hub') for part in skill_md.parts):
                     continue
                 try:

agent/title_generator.py

@@ -38,7 +38,7 @@ def generate_title(user_message: str, assistant_response: str, timeout: float =
         response = call_llm(
             task="title_generation",
             messages=messages,
-            max_tokens=30,
+            max_tokens=500,
             temperature=0.3,
             timeout=timeout,
         )

agent/transports/anthropic.py

@@ -78,31 +78,52 @@ class AnthropicTransport(ProviderTransport):
     def normalize_response(self, response: Any, **kwargs) -> NormalizedResponse:
         """Normalize Anthropic response to NormalizedResponse.
 
-        Calls the adapter's v1 normalize and maps the (SimpleNamespace, finish_reason)
-        tuple to the shared NormalizedResponse type.
+        Parses content blocks (text, thinking, tool_use), maps stop_reason
+        to OpenAI finish_reason, and collects reasoning_details in provider_data.
         """
-        from agent.anthropic_adapter import normalize_anthropic_response
-        from agent.transports.types import build_tool_call
+        import json
+        from agent.anthropic_adapter import _to_plain_data
+        from agent.transports.types import ToolCall
 
         strip_tool_prefix = kwargs.get("strip_tool_prefix", False)
-        assistant_msg, finish_reason = normalize_anthropic_response(response, strip_tool_prefix)
+        _MCP_PREFIX = "mcp_"
 
-        tool_calls = None
-        if assistant_msg.tool_calls:
-            tool_calls = [
-                build_tool_call(id=tc.id, name=tc.function.name, arguments=tc.function.arguments)
-                for tc in assistant_msg.tool_calls
-            ]
+        text_parts = []
+        reasoning_parts = []
+        reasoning_details = []
+        tool_calls = []
+
+        for block in response.content:
+            if block.type == "text":
+                text_parts.append(block.text)
+            elif block.type == "thinking":
+                reasoning_parts.append(block.thinking)
+                block_dict = _to_plain_data(block)
+                if isinstance(block_dict, dict):
+                    reasoning_details.append(block_dict)
+            elif block.type == "tool_use":
+                name = block.name
+                if strip_tool_prefix and name.startswith(_MCP_PREFIX):
+                    name = name[len(_MCP_PREFIX):]
+                tool_calls.append(
+                    ToolCall(
+                        id=block.id,
+                        name=name,
+                        arguments=json.dumps(block.input),
+                    )
+                )
+
+        finish_reason = self._STOP_REASON_MAP.get(response.stop_reason, "stop")
 
         provider_data = {}
-        if getattr(assistant_msg, "reasoning_details", None):
-            provider_data["reasoning_details"] = assistant_msg.reasoning_details
+        if reasoning_details:
+            provider_data["reasoning_details"] = reasoning_details
 
         return NormalizedResponse(
-            content=assistant_msg.content,
-            tool_calls=tool_calls,
+            content="\n".join(text_parts) if text_parts else None,
+            tool_calls=tool_calls or None,
             finish_reason=finish_reason,
-            reasoning=getattr(assistant_msg, "reasoning", None),
+            reasoning="\n\n".join(reasoning_parts) if reasoning_parts else None,
             usage=None,
             provider_data=provider_data or None,
         )

agent/transports/chat_completions.py

@@ -12,6 +12,7 @@ reasoning configuration, temperature handling, and extra_body assembly.
 import copy
 from typing import Any, Dict, List, Optional
 
+from agent.moonshot_schema import is_moonshot_model, sanitize_moonshot_tools
 from agent.prompt_builder import DEVELOPER_ROLE_MODELS
 from agent.transports.base import ProviderTransport
 from agent.transports.types import NormalizedResponse, ToolCall, Usage
@@ -172,6 +173,11 @@ class ChatCompletionsTransport(ProviderTransport):
 
         # Tools
         if tools:
+            # Moonshot/Kimi uses a stricter flavored JSON Schema.  Rewriting
+            # tool parameters here keeps aggregator routes (Nous, OpenRouter,
+            # etc.) compatible, in addition to direct moonshot.ai endpoints.
+            if is_moonshot_model(model):
+                tools = sanitize_moonshot_tools(tools)
             api_kwargs["tools"] = tools
 
         # max_tokens resolution — priority: ephemeral > user > provider default

agent/transports/types.py

@@ -37,6 +37,44 @@ class ToolCall:
     arguments: str  # JSON string
     provider_data: Optional[Dict[str, Any]] = field(default=None, repr=False)
 
+    # ── Backward compatibility ──────────────────────────────────
+    # The agent loop reads tc.function.name / tc.function.arguments
+    # throughout run_agent.py (45+ sites).  These properties let
+    # NormalizedResponse pass through without the _nr_to_assistant_message
+    # shim, while keeping ToolCall's canonical fields flat.
+    @property
+    def type(self) -> str:
+        return "function"
+
+    @property
+    def function(self) -> "ToolCall":
+        """Return self so tc.function.name / tc.function.arguments work."""
+        return self
+
+    @property
+    def call_id(self) -> Optional[str]:
+        """Codex call_id from provider_data, accessed via getattr by _build_assistant_message."""
+        return (self.provider_data or {}).get("call_id")
+
+    @property
+    def response_item_id(self) -> Optional[str]:
+        """Codex response_item_id from provider_data."""
+        return (self.provider_data or {}).get("response_item_id")
+
+    @property
+    def extra_content(self) -> Optional[Dict[str, Any]]:
+        """Gemini extra_content (thought_signature) from provider_data.
+
+        Gemini 3 thinking models attach ``extra_content`` with a
+        ``thought_signature`` to each tool call.  This signature must be
+        replayed on subsequent API calls — without it the API rejects the
+        request with HTTP 400.  The chat_completions transport stores this
+        in ``provider_data["extra_content"]``; this property exposes it so
+        ``_build_assistant_message`` can ``getattr(tc, "extra_content")``
+        uniformly.
+        """
+        return (self.provider_data or {}).get("extra_content")
+
 
 @dataclass
 class Usage:
@@ -70,6 +108,24 @@ class NormalizedResponse:
     usage: Optional[Usage] = None
     provider_data: Optional[Dict[str, Any]] = field(default=None, repr=False)
 
+    # ── Backward compatibility ──────────────────────────────────
+    # The shim _nr_to_assistant_message() mapped these from provider_data.
+    # These properties let NormalizedResponse pass through directly.
+    @property
+    def reasoning_content(self) -> Optional[str]:
+        pd = self.provider_data or {}
+        return pd.get("reasoning_content")
+
+    @property
+    def reasoning_details(self):
+        pd = self.provider_data or {}
+        return pd.get("reasoning_details")
+
+    @property
+    def codex_reasoning_items(self):
+        pd = self.provider_data or {}
+        return pd.get("codex_reasoning_items")
+
 
 # ---------------------------------------------------------------------------
 # Factory helpers

cli-config.yaml.example

@@ -326,6 +326,16 @@ compression:
   # To pin a specific model/provider for compression summaries, use the
   # auxiliary section below (auxiliary.compression.provider / model).
 
+# =============================================================================
+# Anthropic prompt caching TTL
+# =============================================================================
+# When prompt caching is active (Claude via OpenRouter or native Anthropic),
+# Anthropic supports two TTL tiers for cached prefixes: "5m" (default) and
+# "1h". Other values are ignored and "5m" is used.
+#
+prompt_caching:
+  cache_ttl: "5m" # use "1h" for long sessions with pauses between turns
+
 # =============================================================================
 # Auxiliary Models (Advanced — Experimental)
 # =============================================================================
@@ -507,6 +517,13 @@ agent:
   # finish, then interrupts anything still running after this timeout.
   # 0 = no drain, interrupt immediately.
   # restart_drain_timeout: 60
+
+  # Max app-level retry attempts for API errors (connection drops, provider
+  # timeouts, 5xx, etc.) before the agent surfaces the failure. Lower this
+  # to 1 if you use fallback providers and want fast failover on flaky
+  # primaries (default 3). The OpenAI SDK does its own low-level retries
+  # underneath this wrapper — this is the Hermes-level loop.
+  # api_max_retries: 3
   
   # Enable verbose logging
   verbose: false

cli.py

@@ -305,13 +305,23 @@ def load_cli_config() -> Dict[str, Any]:
     
     Environment variables take precedence over config file values.
     Returns default values if no config file exists.
+
+    If HERMES_IGNORE_USER_CONFIG=1 is set (via ``hermes chat --ignore-user-config``),
+    the user config at ``~/.hermes/config.yaml`` is skipped entirely and only the
+    built-in defaults plus the project-level ``cli-config.yaml`` (if any) are used.
+    Credentials in ``.env`` are still loaded — this flag only suppresses
+    behavioral/config settings.
     """
     # Check user config first ({HERMES_HOME}/config.yaml)
     user_config_path = _hermes_home / 'config.yaml'
     project_config_path = Path(__file__).parent / 'cli-config.yaml'
 
+    # --ignore-user-config: force-skip the user config.yaml (still honor project
+    # config as a fallback so defaults stay sensible).
+    ignore_user_config = os.environ.get("HERMES_IGNORE_USER_CONFIG") == "1"
+
     # Use user config if it exists, otherwise project config
-    if user_config_path.exists():
+    if user_config_path.exists() and not ignore_user_config:
         config_path = user_config_path
     else:
         config_path = project_config_path
@@ -1678,7 +1688,6 @@ def _looks_like_slash_command(text: str) -> bool:
 from agent.skill_commands import (
     scan_skill_commands,
     build_skill_invocation_message,
-    build_plan_path,
     build_preloaded_skills_prompt,
 )
 
@@ -1802,6 +1811,7 @@ class HermesCLI:
         resume: str = None,
         checkpoints: bool = False,
         pass_session_id: bool = False,
+        ignore_rules: bool = False,
     ):
         """
         Initialize the Hermes CLI.
@@ -1955,6 +1965,11 @@ class HermesCLI:
         self.checkpoints_enabled = checkpoints or cp_cfg.get("enabled", False)
         self.checkpoint_max_snapshots = cp_cfg.get("max_snapshots", 50)
         self.pass_session_id = pass_session_id
+        # --ignore-rules: honor either the constructor flag or the env var set
+        # by `hermes chat --ignore-rules` in hermes_cli/main.py. When true we
+        # pass skip_context_files=True and skip_memory=True to AIAgent so
+        # AGENTS.md/SOUL.md/.cursorrules and persistent memory are not loaded.
+        self.ignore_rules = ignore_rules or os.environ.get("HERMES_IGNORE_RULES") == "1"
         
         # Ephemeral system prompt: env var takes precedence, then config
         self.system_prompt = (
@@ -3068,6 +3083,8 @@ class HermesCLI:
             format_runtime_provider_error,
         )
 
+        _primary_exc = None
+        runtime = None
         try:
             runtime = resolve_runtime_provider(
                 requested=self.requested_provider,
@@ -3075,7 +3092,34 @@ class HermesCLI:
                 explicit_base_url=self._explicit_base_url,
             )
         except Exception as exc:
-            message = format_runtime_provider_error(exc)
+            _primary_exc = exc
+
+        # Primary provider auth failed — try fallback providers before giving up.
+        if runtime is None and _primary_exc is not None:
+            from hermes_cli.auth import AuthError
+            if isinstance(_primary_exc, AuthError):
+                _fb_chain = self._fallback_model if isinstance(self._fallback_model, list) else []
+                for _fb in _fb_chain:
+                    _fb_provider = (_fb.get("provider") or "").strip().lower()
+                    _fb_model = (_fb.get("model") or "").strip()
+                    if not _fb_provider or not _fb_model:
+                        continue
+                    try:
+                        runtime = resolve_runtime_provider(requested=_fb_provider)
+                        logger.warning(
+                            "Primary provider auth failed (%s). Falling through to fallback: %s/%s",
+                            _primary_exc, _fb_provider, _fb_model,
+                        )
+                        _cprint(f"⚠️  Primary auth failed — switching to fallback: {_fb_provider} / {_fb_model}")
+                        self.requested_provider = _fb_provider
+                        self.model = _fb_model
+                        _primary_exc = None
+                        break
+                    except Exception:
+                        continue
+
+        if runtime is None:
+            message = format_runtime_provider_error(_primary_exc) if _primary_exc else "Provider resolution failed."
             ChatConsole().print(f"[bold red]{message}[/]")
             return False
 
@@ -3238,6 +3282,23 @@ class HermesCLI:
                 _cprint(f"\033[1;31mSession not found: {self.session_id}{_RST}")
                 _cprint(f"{_DIM}Use a session ID from a previous CLI run (hermes sessions list).{_RST}")
                 return False
+            # If the requested session is the (empty) head of a compression
+            # chain, walk to the descendant that actually holds the messages.
+            # See #15000 and SessionDB.resolve_resume_session_id.
+            try:
+                resolved_id = self._session_db.resolve_resume_session_id(self.session_id)
+            except Exception:
+                resolved_id = self.session_id
+            if resolved_id and resolved_id != self.session_id:
+                ChatConsole().print(
+                    f"[{_DIM}]Session {_escape(self.session_id)} was compressed into "
+                    f"{_escape(resolved_id)}; resuming the descendant with your "
+                    f"transcript.[/]"
+                )
+                self.session_id = resolved_id
+                resolved_meta = self._session_db.get_session(self.session_id)
+                if resolved_meta:
+                    session_meta = resolved_meta
             restored = self._session_db.get_messages_as_conversation(self.session_id)
             if restored:
                 restored = [m for m in restored if m.get("role") != "session_meta"]
@@ -3312,6 +3373,8 @@ class HermesCLI:
                 checkpoints_enabled=self.checkpoints_enabled,
                 checkpoint_max_snapshots=self.checkpoint_max_snapshots,
                 pass_session_id=self.pass_session_id,
+                skip_context_files=self.ignore_rules,
+                skip_memory=self.ignore_rules,
                 tool_progress_callback=self._on_tool_progress,
                 tool_start_callback=self._on_tool_start if self._inline_diffs_enabled else None,
                 tool_complete_callback=self._on_tool_complete if self._inline_diffs_enabled else None,
@@ -3454,6 +3517,22 @@ class HermesCLI:
             )
             return False
 
+        # If the requested session is the (empty) head of a compression chain,
+        # walk to the descendant that actually holds the messages. See #15000.
+        try:
+            resolved_id = self._session_db.resolve_resume_session_id(self.session_id)
+        except Exception:
+            resolved_id = self.session_id
+        if resolved_id and resolved_id != self.session_id:
+            self._console_print(
+                f"[dim]Session {self.session_id} was compressed into "
+                f"{resolved_id}; resuming the descendant with your transcript.[/]"
+            )
+            self.session_id = resolved_id
+            resolved_meta = self._session_db.get_session(self.session_id)
+            if resolved_meta:
+                session_meta = resolved_meta
+
         restored = self._session_db.get_messages_as_conversation(self.session_id)
         if restored:
             restored = [m for m in restored if m.get("role") != "session_meta"]
@@ -4668,6 +4747,22 @@ class HermesCLI:
             _cprint("  Use /history or `hermes sessions list` to see available sessions.")
             return
 
+        # If the target is the empty head of a compression chain, redirect to
+        # the descendant that actually holds the transcript. See #15000.
+        try:
+            resolved_id = self._session_db.resolve_resume_session_id(target_id)
+        except Exception:
+            resolved_id = target_id
+        if resolved_id and resolved_id != target_id:
+            _cprint(
+                f"  Session {target_id} was compressed into {resolved_id}; "
+                f"resuming the descendant with your transcript."
+            )
+            target_id = resolved_id
+            resolved_meta = self._session_db.get_session(target_id)
+            if resolved_meta:
+                session_meta = resolved_meta
+
         if target_id == self.session_id:
             _cprint("  Already on that session.")
             return
@@ -5360,79 +5455,6 @@ class HermesCLI:
         except Exception:
             return False
 
-    def _show_model_and_providers(self):
-        """Show current model + provider and list all authenticated providers.
-
-        Shows current model + provider, then lists all authenticated
-        providers with their available models.
-        """
-        from hermes_cli.models import (
-            curated_models_for_provider, list_available_providers,
-            normalize_provider, _PROVIDER_LABELS,
-            get_pricing_for_provider, format_model_pricing_table,
-        )
-        from hermes_cli.auth import resolve_provider as _resolve_provider
-
-        # Resolve current provider
-        raw_provider = normalize_provider(self.provider)
-        if raw_provider == "auto":
-            try:
-                current = _resolve_provider(
-                    self.requested_provider,
-                    explicit_api_key=self._explicit_api_key,
-                    explicit_base_url=self._explicit_base_url,
-                )
-            except Exception:
-                current = "openrouter"
-        else:
-            current = raw_provider
-        current_label = _PROVIDER_LABELS.get(current, current)
-
-        print(f"\n  Current: {self.model} via {current_label}")
-        print()
-
-        # Show all authenticated providers with their models
-        providers = list_available_providers()
-        authed = [p for p in providers if p["authenticated"]]
-        unauthed = [p for p in providers if not p["authenticated"]]
-
-        if authed:
-            print("  Authenticated providers & models:")
-            for p in authed:
-                is_active = p["id"] == current
-                marker = " ← active" if is_active else ""
-                print(f"    [{p['id']}]{marker}")
-                curated = curated_models_for_provider(p["id"])
-                # Fetch pricing for providers that support it (openrouter, nous)
-                pricing_map = get_pricing_for_provider(p["id"]) if p["id"] in ("openrouter", "nous") else {}
-                if curated and pricing_map:
-                    cur_model = self.model if is_active else ""
-                    for line in format_model_pricing_table(curated, pricing_map, current_model=cur_model):
-                        print(line)
-                elif curated:
-                    for mid, desc in curated:
-                        current_marker = " ← current" if (is_active and mid == self.model) else ""
-                        print(f"      {mid}{current_marker}")
-                elif p["id"] == "custom":
-                    from hermes_cli.models import _get_custom_base_url
-                    custom_url = _get_custom_base_url()
-                    if custom_url:
-                        print(f"      endpoint: {custom_url}")
-                    if is_active:
-                        print(f"      model: {self.model} ← current")
-                    print("      (use hermes model to change)")
-                else:
-                    print("      (use hermes model to change)")
-                print()
-
-        if unauthed:
-            names = ", ".join(p["label"] for p in unauthed)
-            print(f"  Not configured: {names}")
-            print("  Run: hermes setup")
-            print()
-
-        print("  To change model or provider, use: hermes model")
-
     def _output_console(self):
         """Use prompt_toolkit-safe Rich rendering once the TUI is live."""
         if getattr(self, "_app", None):
@@ -6008,16 +6030,12 @@ class HermesCLI:
             self._handle_resume_command(cmd_original)
         elif canonical == "model":
             self._handle_model_switch(cmd_original)
-        elif canonical == "provider":
-            self._show_model_and_providers()
         elif canonical == "gquota":
             self._handle_gquota_command(cmd_original)
 
         elif canonical == "personality":
             # Use original case (handler lowercases the personality name itself)
             self._handle_personality_command(cmd_original)
-        elif canonical == "plan":
-            self._handle_plan_command(cmd_original)
         elif canonical == "retry":
             retry_msg = self.retry_last()
             if retry_msg and hasattr(self, '_pending_input'):
@@ -6252,32 +6270,6 @@ class HermesCLI:
         
         return True
     
-    def _handle_plan_command(self, cmd: str):
-        """Handle /plan [request] — load the bundled plan skill."""
-        parts = cmd.strip().split(maxsplit=1)
-        user_instruction = parts[1].strip() if len(parts) > 1 else ""
-
-        plan_path = build_plan_path(user_instruction)
-        msg = build_skill_invocation_message(
-            "/plan",
-            user_instruction,
-            task_id=self.session_id,
-            runtime_note=(
-                "Save the markdown plan with write_file to this exact relative path "
-                f"inside the active workspace/backend cwd: {plan_path}"
-            ),
-        )
-
-        if not msg:
-            ChatConsole().print("[bold red]Failed to load the bundled /plan skill[/]")
-            return
-
-        _cprint(f"  📝 Plan mode queued via skill. Markdown plan target: {plan_path}")
-        if hasattr(self, '_pending_input'):
-            self._pending_input.put(msg)
-        else:
-            ChatConsole().print("[bold red]Plan mode unavailable: input queue not initialized[/]")
-    
     def _handle_background_command(self, cmd: str):
         """Handle /background <prompt> — run a prompt in a separate background session.
 
@@ -6667,6 +6659,13 @@ class HermesCLI:
                 print(f"   ⚠ Port {_port} is not reachable at {cdp_url}")
 
             os.environ["BROWSER_CDP_URL"] = cdp_url
+            # Eagerly start the CDP supervisor so pending_dialogs + frame_tree
+            # show up in the next browser_snapshot.  No-op if already started.
+            try:
+                from tools.browser_tool import _ensure_cdp_supervisor  # type: ignore[import-not-found]
+                _ensure_cdp_supervisor("default")
+            except Exception:
+                pass
             print()
             print("🌐 Browser connected to live Chrome via CDP")
             print(f"   Endpoint: {cdp_url}")
@@ -6688,7 +6687,8 @@ class HermesCLI:
             if current:
                 os.environ.pop("BROWSER_CDP_URL", None)
                 try:
-                    from tools.browser_tool import cleanup_all_browsers
+                    from tools.browser_tool import cleanup_all_browsers, _stop_cdp_supervisor
+                    _stop_cdp_supervisor("default")
                     cleanup_all_browsers()
                 except Exception:
                     pass
@@ -10816,6 +10816,8 @@ def main(
     w: bool = False,
     checkpoints: bool = False,
     pass_session_id: bool = False,
+    ignore_user_config: bool = False,
+    ignore_rules: bool = False,
 ):
     """
     Hermes Agent CLI - Interactive AI Assistant
@@ -10925,6 +10927,7 @@ def main(
         resume=resume,
         checkpoints=checkpoints,
         pass_session_id=pass_session_id,
+        ignore_rules=ignore_rules,
     )
 
     if parsed_skills:

cron/jobs.py

@@ -371,6 +371,39 @@ def save_jobs(jobs: List[Dict[str, Any]]):
         raise
 
 
+def _normalize_workdir(workdir: Optional[str]) -> Optional[str]:
+    """Normalize and validate a cron job workdir.
+
+    Rules:
+      - Empty / None → None (feature off, preserves old behaviour).
+      - ``~`` is expanded.  Relative paths are rejected — cron jobs run detached
+        from any shell cwd, so relative paths have no stable meaning.
+      - The path must exist and be a directory at create/update time.  We do
+        NOT re-check at run time (a user might briefly unmount the dir; the
+        scheduler will just fall back to old behaviour with a logged warning).
+
+    Returns the absolute path string, or None when disabled.
+    Raises ValueError on invalid input.
+    """
+    if workdir is None:
+        return None
+    raw = str(workdir).strip()
+    if not raw:
+        return None
+    expanded = Path(raw).expanduser()
+    if not expanded.is_absolute():
+        raise ValueError(
+            f"Cron workdir must be an absolute path (got {raw!r}). "
+            f"Cron jobs run detached from any shell cwd, so relative paths are ambiguous."
+        )
+    resolved = expanded.resolve()
+    if not resolved.exists():
+        raise ValueError(f"Cron workdir does not exist: {resolved}")
+    if not resolved.is_dir():
+        raise ValueError(f"Cron workdir is not a directory: {resolved}")
+    return str(resolved)
+
+
 def create_job(
     prompt: str,
     schedule: str,
@@ -384,6 +417,8 @@ def create_job(
     provider: Optional[str] = None,
     base_url: Optional[str] = None,
     script: Optional[str] = None,
+    enabled_toolsets: Optional[List[str]] = None,
+    workdir: Optional[str] = None,
 ) -> Dict[str, Any]:
     """
     Create a new cron job.
@@ -403,6 +438,15 @@ def create_job(
         script: Optional path to a Python script whose stdout is injected into the
                 prompt each run.  The script runs before the agent turn, and its output
                 is prepended as context.  Useful for data collection / change detection.
+        enabled_toolsets: Optional list of toolset names to restrict the agent to.
+                          When set, only tools from these toolsets are loaded, reducing
+                          token overhead. When omitted, all default tools are loaded.
+        workdir: Optional absolute path.  When set, the job runs as if launched
+                from that directory: AGENTS.md / CLAUDE.md / .cursorrules from
+                that directory are injected into the system prompt, and the
+                terminal/file/code_exec tools use it as their working directory
+                (via TERMINAL_CWD).  When unset, the old behaviour is preserved
+                (no context files injected, tools use the scheduler's cwd).
 
     Returns:
         The created job dict
@@ -433,6 +477,9 @@ def create_job(
     normalized_base_url = normalized_base_url or None
     normalized_script = str(script).strip() if isinstance(script, str) else None
     normalized_script = normalized_script or None
+    normalized_toolsets = [str(t).strip() for t in enabled_toolsets if str(t).strip()] if enabled_toolsets else None
+    normalized_toolsets = normalized_toolsets or None
+    normalized_workdir = _normalize_workdir(workdir)
 
     label_source = (prompt or (normalized_skills[0] if normalized_skills else None)) or "cron job"
     job = {
@@ -464,6 +511,8 @@ def create_job(
         # Delivery configuration
         "deliver": deliver,
         "origin": origin,  # Tracks where job was created for "origin" delivery
+        "enabled_toolsets": normalized_toolsets,
+        "workdir": normalized_workdir,
     }
 
     jobs = load_jobs()
@@ -497,6 +546,15 @@ def update_job(job_id: str, updates: Dict[str, Any]) -> Optional[Dict[str, Any]]
         if job["id"] != job_id:
             continue
 
+        # Validate / normalize workdir if present in updates.  Empty string or
+        # None both mean "clear the field" (restore old behaviour).
+        if "workdir" in updates:
+            _wd = updates["workdir"]
+            if _wd in (None, "", False):
+                updates["workdir"] = None
+            else:
+                updates["workdir"] = _normalize_workdir(_wd)
+
         updated = _apply_skill_fields({**job, **updates})
         schedule_changed = "schedule" in updates
 

cron/scheduler.py

@@ -40,6 +40,37 @@ from hermes_time import now as _hermes_now
 
 logger = logging.getLogger(__name__)
 
+
+def _resolve_cron_enabled_toolsets(job: dict, cfg: dict) -> list[str] | None:
+    """Resolve the toolset list for a cron job.
+
+    Precedence:
+    1. Per-job ``enabled_toolsets`` (set via ``cronjob`` tool on create/update).
+       Keeps the agent's job-scoped toolset override intact — #6130.
+    2. Per-platform ``hermes tools`` config for the ``cron`` platform.
+       Mirrors gateway behavior (``_get_platform_tools(cfg, platform_key)``)
+       so users can gate cron toolsets globally without recreating every job.
+    3. ``None`` on any lookup failure — AIAgent loads the full default set
+       (legacy behavior before this change, preserved as the safety net).
+
+    _DEFAULT_OFF_TOOLSETS ({moa, homeassistant, rl}) are removed by
+    ``_get_platform_tools`` for unconfigured platforms, so fresh installs
+    get cron WITHOUT ``moa`` by default (issue reported by Norbert —
+    surprise $4.63 run).
+    """
+    per_job = job.get("enabled_toolsets")
+    if per_job:
+        return per_job
+    try:
+        from hermes_cli.tools_config import _get_platform_tools  # lazy: avoid heavy import at cron module load
+        return sorted(_get_platform_tools(cfg or {}, "cron"))
+    except Exception as exc:
+        logger.warning(
+            "Cron toolset resolution failed, falling back to full default toolset: %s",
+            exc,
+        )
+        return None
+
 # Valid delivery platforms — used to validate user-supplied platform names
 # in cron delivery targets, preventing env var enumeration via crafted names.
 _KNOWN_DELIVERY_PLATFORMS = frozenset({
@@ -764,6 +795,30 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
         chat_name=origin.get("chat_name", "") if origin else "",
     )
 
+    # Per-job working directory.  When set (and validated at create/update
+    # time), we point TERMINAL_CWD at it so:
+    #   - build_context_files_prompt() picks up AGENTS.md / CLAUDE.md /
+    #     .cursorrules from the job's project dir, AND
+    #   - the terminal, file, and code-exec tools run commands from there.
+    #
+    # tick() serializes workdir-jobs outside the parallel pool, so mutating
+    # os.environ["TERMINAL_CWD"] here is safe for those jobs.  For workdir-less
+    # jobs we leave TERMINAL_CWD untouched — preserves the original behaviour
+    # (skip_context_files=True, tools use whatever cwd the scheduler has).
+    _job_workdir = (job.get("workdir") or "").strip() or None
+    if _job_workdir and not Path(_job_workdir).is_dir():
+        # Directory was removed between create-time validation and now.  Log
+        # and drop back to old behaviour rather than crashing the job.
+        logger.warning(
+            "Job '%s': configured workdir %r no longer exists — running without it",
+            job_id, _job_workdir,
+        )
+        _job_workdir = None
+    _prior_terminal_cwd = os.environ.get("TERMINAL_CWD", "_UNSET_")
+    if _job_workdir:
+        os.environ["TERMINAL_CWD"] = _job_workdir
+        logger.info("Job '%s': using workdir %s", job_id, _job_workdir)
+
     try:
         # Re-read .env and config.yaml fresh every run so provider/key
         # changes take effect without a gateway restart.
@@ -840,6 +895,7 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             resolve_runtime_provider,
             format_runtime_provider_error,
         )
+        from hermes_cli.auth import AuthError
         try:
             runtime_kwargs = {
                 "requested": job.get("provider") or os.getenv("HERMES_INFERENCE_PROVIDER"),
@@ -847,6 +903,28 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             if job.get("base_url"):
                 runtime_kwargs["explicit_base_url"] = job.get("base_url")
             runtime = resolve_runtime_provider(**runtime_kwargs)
+        except AuthError as auth_exc:
+            # Primary provider auth failed — try fallback chain before giving up.
+            logger.warning("Job '%s': primary auth failed (%s), trying fallback", job_id, auth_exc)
+            fb = _cfg.get("fallback_providers") or _cfg.get("fallback_model")
+            fb_list = (fb if isinstance(fb, list) else [fb]) if fb else []
+            runtime = None
+            for entry in fb_list:
+                if not isinstance(entry, dict):
+                    continue
+                try:
+                    fb_kwargs = {"requested": entry.get("provider")}
+                    if entry.get("base_url"):
+                        fb_kwargs["explicit_base_url"] = entry["base_url"]
+                    if entry.get("api_key"):
+                        fb_kwargs["explicit_api_key"] = entry["api_key"]
+                    runtime = resolve_runtime_provider(**fb_kwargs)
+                    logger.info("Job '%s': fallback resolved to %s", job_id, runtime.get("provider"))
+                    break
+                except Exception as fb_exc:
+                    logger.debug("Job '%s': fallback %s failed: %s", job_id, entry.get("provider"), fb_exc)
+            if runtime is None:
+                raise RuntimeError(format_runtime_provider_error(auth_exc)) from auth_exc
         except Exception as exc:
             message = format_runtime_provider_error(exc)
             raise RuntimeError(message) from exc
@@ -886,9 +964,13 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             providers_ignored=pr.get("ignore"),
             providers_order=pr.get("order"),
             provider_sort=pr.get("sort"),
+            enabled_toolsets=_resolve_cron_enabled_toolsets(job, _cfg),
             disabled_toolsets=["cronjob", "messaging", "clarify"],
             quiet_mode=True,
-            skip_context_files=True,  # Don't inject SOUL.md/AGENTS.md from scheduler cwd
+            # When a workdir is configured, inject AGENTS.md / CLAUDE.md /
+            # .cursorrules from that directory; otherwise preserve the old
+            # behaviour (don't inject SOUL.md/AGENTS.md from the scheduler cwd).
+            skip_context_files=not bool(_job_workdir),
             skip_memory=True,  # Cron system prompts would corrupt user representations
             platform="cron",
             session_id=_cron_session_id,
@@ -972,6 +1054,12 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
                 f"— last activity: {_last_desc}"
             )
 
+        # Guard against non-dict returns from run_conversation under error conditions
+        if not isinstance(result, dict):
+            raise RuntimeError(
+                f"agent.run_conversation returned {type(result).__name__} instead of dict: {result!r}"
+            )
+
         final_response = result.get("final_response", "") or ""
         # Strip leaked placeholder text that upstream may inject on empty completions.
         if final_response.strip() == "(No response generated)":
@@ -1021,6 +1109,14 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
         return False, output, "", error_msg
 
     finally:
+        # Restore TERMINAL_CWD to whatever it was before this job ran.  We
+        # only ever mutate it when the job has a workdir; see the setup block
+        # at the top of run_job for the serialization guarantee.
+        if _job_workdir:
+            if _prior_terminal_cwd == "_UNSET_":
+                os.environ.pop("TERMINAL_CWD", None)
+            else:
+                os.environ["TERMINAL_CWD"] = _prior_terminal_cwd
         # Clean up ContextVar session/delivery state for this job.
         clear_session_vars(_ctx_tokens)
         if _session_db:
@@ -1148,14 +1244,28 @@ def tick(verbose: bool = True, adapters=None, loop=None) -> int:
                 mark_job_run(job["id"], False, str(e))
                 return False
 
-        # Run all due jobs concurrently, each in its own ContextVar copy
-        # so session/delivery state stays isolated per-thread.
-        with concurrent.futures.ThreadPoolExecutor(max_workers=_max_workers) as _tick_pool:
-            _futures = []
-            for job in due_jobs:
-                _ctx = contextvars.copy_context()
-                _futures.append(_tick_pool.submit(_ctx.run, _process_job, job))
-            _results = [f.result() for f in _futures]
+        # Partition due jobs: those with a per-job workdir mutate
+        # os.environ["TERMINAL_CWD"] inside run_job, which is process-global —
+        # so they MUST run sequentially to avoid corrupting each other.  Jobs
+        # without a workdir leave env untouched and stay parallel-safe.
+        workdir_jobs = [j for j in due_jobs if (j.get("workdir") or "").strip()]
+        parallel_jobs = [j for j in due_jobs if not (j.get("workdir") or "").strip()]
+
+        _results: list = []
+
+        # Sequential pass for workdir jobs.
+        for job in workdir_jobs:
+            _ctx = contextvars.copy_context()
+            _results.append(_ctx.run(_process_job, job))
+
+        # Parallel pass for the rest — same behaviour as before.
+        if parallel_jobs:
+            with concurrent.futures.ThreadPoolExecutor(max_workers=_max_workers) as _tick_pool:
+                _futures = []
+                for job in parallel_jobs:
+                    _ctx = contextvars.copy_context()
+                    _futures.append(_tick_pool.submit(_ctx.run, _process_job, job))
+                _results.extend(f.result() for f in _futures)
 
         return sum(_results)
     finally:

docker-compose.yml

@@ -0,0 +1,52 @@
+#
+# docker-compose.yml for Hermes Agent
+#
+# Usage:
+#   HERMES_UID=$(id -u) HERMES_GID=$(id -g) docker compose up -d
+#
+# Set HERMES_UID / HERMES_GID to the host user that owns ~/.hermes so
+# files created inside the container stay readable/writable on the host.
+# The entrypoint remaps the internal `hermes` user to these values via
+# usermod/groupmod + gosu.
+#
+# Security notes:
+#   - The dashboard service binds to 127.0.0.1 by default. It stores API
+#     keys; exposing it on LAN without auth is unsafe. If you want remote
+#     access, use an SSH tunnel or put it behind a reverse proxy that
+#     adds authentication — do NOT pass --insecure --host 0.0.0.0.
+#   - The gateway's API server is off unless you uncomment API_SERVER_KEY
+#     and API_SERVER_HOST. See docs/user-guide/api-server.md before doing
+#     this on an internet-facing host.
+#
+services:
+  gateway:
+    build: .
+    image: hermes-agent
+    container_name: hermes
+    restart: unless-stopped
+    network_mode: host
+    volumes:
+      - ~/.hermes:/opt/data
+    environment:
+      - HERMES_UID=${HERMES_UID:-10000}
+      - HERMES_GID=${HERMES_GID:-10000}
+      # To expose the OpenAI-compatible API server beyond localhost,
+      # uncomment BOTH lines (API_SERVER_KEY is mandatory for auth):
+      # - API_SERVER_HOST=0.0.0.0
+      # - API_SERVER_KEY=${API_SERVER_KEY}
+    command: ["gateway", "run"]
+
+  dashboard:
+    image: hermes-agent
+    container_name: hermes-dashboard
+    restart: unless-stopped
+    network_mode: host
+    depends_on:
+      - gateway
+    volumes:
+      - ~/.hermes:/opt/data
+    environment:
+      - HERMES_UID=${HERMES_UID:-10000}
+      - HERMES_GID=${HERMES_GID:-10000}
+    # Localhost-only. For remote access, tunnel via `ssh -L 9119:localhost:9119`.
+    command: ["dashboard", "--host", "127.0.0.1", "--no-open"]

docker/entrypoint.sh

@@ -22,9 +22,18 @@ if [ "$(id -u)" = "0" ]; then
         groupmod -o -g "$HERMES_GID" hermes 2>/dev/null || true
     fi
 
+    # Fix ownership of the data volume. When HERMES_UID remaps the hermes user,
+    # files created by previous runs (under the old UID) become inaccessible.
+    # Always chown -R when UID was remapped; otherwise only if top-level is wrong.
     actual_hermes_uid=$(id -u hermes)
-    if [ "$(stat -c %u "$HERMES_HOME" 2>/dev/null)" != "$actual_hermes_uid" ]; then
-        echo "$HERMES_HOME is not owned by $actual_hermes_uid, fixing"
+    needs_chown=false
+    if [ -n "$HERMES_UID" ] && [ "$HERMES_UID" != "10000" ]; then
+        needs_chown=true
+    elif [ "$(stat -c %u "$HERMES_HOME" 2>/dev/null)" != "$actual_hermes_uid" ]; then
+        needs_chown=true
+    fi
+    if [ "$needs_chown" = true ]; then
+        echo "Fixing ownership of $HERMES_HOME to hermes ($actual_hermes_uid)"
         # In rootless Podman the container's "root" is mapped to an unprivileged
         # host UID — chown will fail.  That's fine: the volume is already owned
         # by the mapped user on the host side.
@@ -58,6 +67,13 @@ if [ ! -f "$HERMES_HOME/config.yaml" ]; then
     cp "$INSTALL_DIR/cli-config.yaml.example" "$HERMES_HOME/config.yaml"
 fi
 
+# Ensure the main config file remains accessible to the hermes runtime user
+# even if it was edited on the host after initial ownership setup.
+if [ -f "$HERMES_HOME/config.yaml" ]; then
+    chown hermes:hermes "$HERMES_HOME/config.yaml"
+    chmod 640 "$HERMES_HOME/config.yaml"
+fi
+
 # SOUL.md
 if [ ! -f "$HERMES_HOME/SOUL.md" ]; then
     cp "$INSTALL_DIR/docker/SOUL.md" "$HERMES_HOME/SOUL.md"

gateway/platforms/base.py

@@ -752,7 +752,10 @@ class MessageEvent:
         if not self.is_command():
             return self.text
         parts = self.text.split(maxsplit=1)
-        return parts[1] if len(parts) > 1 else ""
+        args = parts[1] if len(parts) > 1 else ""
+        # iOS auto-corrects -- to — (em dash) and - to – (en dash)
+        args = args.replace("\u2014\u2014", "--").replace("\u2014", "--").replace("\u2013", "-")
+        return args
 
 
 @dataclass 
@@ -897,10 +900,16 @@ class BasePlatformAdapter(ABC):
         self._fatal_error_retryable = True
         self._fatal_error_handler: Optional[Callable[["BasePlatformAdapter"], Awaitable[None] | None]] = None
         
-        # Track active message handlers per session for interrupt support
-        # Key: session_key (e.g., chat_id), Value: (event, asyncio.Event for interrupt)
+        # Track active message handlers per session for interrupt support.
+        # _active_sessions stores the per-session interrupt Event; _session_tasks
+        # maps session → the specific Task currently processing it so that
+        # session-terminating commands (/stop, /new, /reset) can cancel the
+        # right task and release the adapter-level guard deterministically.
+        # Without the owner-task map, an old task's finally block could delete
+        # a newer task's guard, leaving stale busy state.
         self._active_sessions: Dict[str, asyncio.Event] = {}
         self._pending_messages: Dict[str, MessageEvent] = {}
+        self._session_tasks: Dict[str, asyncio.Task] = {}
         # Background message-processing tasks spawned by handle_message().
         # Gateway shutdown cancels these so an old gateway instance doesn't keep
         # working on a task after --replace or manual restarts.
@@ -1343,7 +1352,7 @@ class BasePlatformAdapter(ABC):
         # Extract MEDIA:<path> tags, allowing optional whitespace after the colon
         # and quoted/backticked paths for LLM-formatted outputs.
         media_pattern = re.compile(
-            r'''[`"']?MEDIA:\s*(?P<path>`[^`\n]+`|"[^"\n]+"|'[^'\n]+'|(?:~/|/)\S+(?:[^\S\n]+\S+)*?\.(?:png|jpe?g|gif|webp|mp4|mov|avi|mkv|webm|ogg|opus|mp3|wav|m4a|pdf)(?=[\s`"',;:)\]}]|$)|\S+)[`"']?'''
+            r'''[`"']?MEDIA:\s*(?P<path>`[^`\n]+`|"[^"\n]+"|'[^'\n]+'|(?:~/|/)\S+(?:[^\S\n]+\S+)*?\.(?:png|jpe?g|gif|webp|mp4|mov|avi|mkv|webm|ogg|opus|mp3|wav|m4a|epub|pdf|zip|rar|7z|docx?|xlsx?|pptx?|txt|csv|apk|ipa)(?=[\s`"',;:)\]}]|$)|\S+)[`"']?'''
         )
         for match in media_pattern.finditer(content):
             path = match.group("path").strip()
@@ -1677,6 +1686,222 @@ class BasePlatformAdapter(ABC):
             return f"{existing_text}\n\n{new_text}".strip()
         return existing_text
 
+    # ------------------------------------------------------------------
+    # Session task + guard ownership helpers
+    # ------------------------------------------------------------------
+    # These were introduced together with the _session_tasks owner map to
+    # make session lifecycle reconciliation deterministic across (a) the
+    # normal completion path, (b) /stop/ /new/ /reset bypass commands,
+    # and (c) stale-lock self-heal on the next inbound message.
+
+    def _release_session_guard(
+        self,
+        session_key: str,
+        *,
+        guard: Optional[asyncio.Event] = None,
+    ) -> None:
+        """Release the adapter-level guard for a session.
+
+        When ``guard`` is provided, only release the entry if it still points
+        at that exact Event.  This lets reset-like commands swap in a temporary
+        guard while the old processing task unwinds, without having the old
+        task's cleanup accidentally clear the replacement guard.
+        """
+        current_guard = self._active_sessions.get(session_key)
+        if current_guard is None:
+            return
+        if guard is not None and current_guard is not guard:
+            return
+        del self._active_sessions[session_key]
+
+    def _session_task_is_stale(self, session_key: str) -> bool:
+        """Return True if the owner task for ``session_key`` is done/cancelled.
+
+        A lock is "stale" when the adapter still has ``_active_sessions[key]``
+        AND a known owner task in ``_session_tasks`` that has already exited.
+        When there is no owner task at all, that usually means the guard was
+        installed by some path other than handle_message() (tests sometimes
+        install guards directly) — don't treat that as stale.  The on-entry
+        self-heal only needs to handle the production split-brain case where
+        an owner task was recorded, then exited without clearing its guard.
+        """
+        task = self._session_tasks.get(session_key)
+        if task is None:
+            return False
+        done = getattr(task, "done", None)
+        return bool(done and done())
+
+    def _heal_stale_session_lock(self, session_key: str) -> bool:
+        """Clear a stale session lock if the owner task is already gone.
+
+        Returns True if a stale lock was healed.  Returns False if there is
+        no lock, or the owner task is still alive (the normal busy case).
+
+        This is the on-entry safety net sidbin's issue #11016 analysis calls
+        for: without it, a split-brain — adapter still thinks the session is
+        active, but nothing is actually processing — traps the chat in
+        infinite "Interrupting current task..." until the gateway is
+        restarted.
+        """
+        if session_key not in self._active_sessions:
+            return False
+        if not self._session_task_is_stale(session_key):
+            return False
+        logger.warning(
+            "[%s] Healing stale session lock for %s (owner task is done/absent)",
+            self.name,
+            session_key,
+        )
+        self._active_sessions.pop(session_key, None)
+        self._pending_messages.pop(session_key, None)
+        self._session_tasks.pop(session_key, None)
+        return True
+
+    def _start_session_processing(
+        self,
+        event: MessageEvent,
+        session_key: str,
+        *,
+        interrupt_event: Optional[asyncio.Event] = None,
+    ) -> bool:
+        """Spawn a background processing task under the given session guard.
+
+        Returns True on success.  If the runtime stubs ``create_task`` with a
+        non-Task sentinel (some tests do this), the guard is rolled back and
+        False is returned so the caller isn't left holding a half-installed
+        session lock.
+        """
+        guard = interrupt_event or asyncio.Event()
+        self._active_sessions[session_key] = guard
+
+        task = asyncio.create_task(self._process_message_background(event, session_key))
+        self._session_tasks[session_key] = task
+        try:
+            self._background_tasks.add(task)
+        except TypeError:
+            # Tests stub create_task() with lightweight sentinels that are not
+            # hashable and do not support lifecycle callbacks.
+            self._session_tasks.pop(session_key, None)
+            self._release_session_guard(session_key, guard=guard)
+            return False
+        if hasattr(task, "add_done_callback"):
+            task.add_done_callback(self._background_tasks.discard)
+            task.add_done_callback(self._expected_cancelled_tasks.discard)
+        return True
+
+    async def cancel_session_processing(
+        self,
+        session_key: str,
+        *,
+        release_guard: bool = True,
+        discard_pending: bool = True,
+    ) -> None:
+        """Cancel in-flight processing for a single session.
+
+        ``release_guard=False`` keeps the adapter-level session guard in place
+        so reset-like commands can finish atomically before follow-up messages
+        are allowed to start a fresh background task.
+        """
+        task = self._session_tasks.pop(session_key, None)
+        if task is not None and not task.done():
+            logger.debug(
+                "[%s] Cancelling active processing for session %s",
+                self.name,
+                session_key,
+            )
+            self._expected_cancelled_tasks.add(task)
+            task.cancel()
+            try:
+                await task
+            except asyncio.CancelledError:
+                pass
+            except Exception:
+                logger.debug(
+                    "[%s] Session cancellation raised while unwinding %s",
+                    self.name,
+                    session_key,
+                    exc_info=True,
+                )
+        if discard_pending:
+            self._pending_messages.pop(session_key, None)
+        if release_guard:
+            self._release_session_guard(session_key)
+
+    async def _drain_pending_after_session_command(
+        self,
+        session_key: str,
+        command_guard: asyncio.Event,
+    ) -> None:
+        """Resume the latest queued follow-up once a session command completes.
+
+        Called at the tail of /stop, /new, and /reset dispatch.  Releases the
+        command-scoped guard, then — if a follow-up message landed while the
+        command was running — spawns a fresh processing task for it.
+        """
+        pending_event = self._pending_messages.pop(session_key, None)
+        self._release_session_guard(session_key, guard=command_guard)
+        if pending_event is None:
+            return
+        self._start_session_processing(pending_event, session_key)
+
+    async def _dispatch_active_session_command(
+        self,
+        event: MessageEvent,
+        session_key: str,
+        cmd: str,
+    ) -> None:
+        """Dispatch a reset-like bypass command while preserving guard ordering.
+
+        /stop, /new, and /reset must:
+          1. Keep the session guard installed while the runner processes the
+             command (so a racing follow-up message stays queued, not
+             dispatched as a second parallel run).
+          2. Cancel the old in-flight adapter task only AFTER the runner has
+             finished handling the command (so the runner sees consistent
+             state and its response is sent in order).
+          3. Release the command-scoped guard and drain the latest queued
+             follow-up exactly once, after 1 and 2 complete.
+        """
+        logger.debug(
+            "[%s] Command '/%s' bypassing active-session guard for %s",
+            self.name,
+            cmd,
+            session_key,
+        )
+
+        current_guard = self._active_sessions.get(session_key)
+        command_guard = asyncio.Event()
+        self._active_sessions[session_key] = command_guard
+        thread_meta = {"thread_id": event.source.thread_id} if event.source.thread_id else None
+
+        try:
+            response = await self._message_handler(event)
+            # Old adapter task (if any) is cancelled AFTER the runner has
+            # fully handled the command — keeps ordering deterministic.
+            await self.cancel_session_processing(
+                session_key,
+                release_guard=False,
+                discard_pending=False,
+            )
+            if response:
+                await self._send_with_retry(
+                    chat_id=event.source.chat_id,
+                    content=response,
+                    reply_to=event.message_id,
+                    metadata=thread_meta,
+                )
+        except Exception:
+            # On failure, restore the original guard if one still exists so
+            # we don't leave the session in a half-reset state.
+            if self._active_sessions.get(session_key) is command_guard:
+                if session_key in self._session_tasks and current_guard is not None:
+                    self._active_sessions[session_key] = current_guard
+                else:
+                    self._release_session_guard(session_key, guard=command_guard)
+            raise
+
+        await self._drain_pending_after_session_command(session_key, command_guard)
+
     async def handle_message(self, event: MessageEvent) -> None:
         """
         Process an incoming message.
@@ -1693,7 +1918,15 @@ class BasePlatformAdapter(ABC):
             group_sessions_per_user=self.config.extra.get("group_sessions_per_user", True),
             thread_sessions_per_user=self.config.extra.get("thread_sessions_per_user", False),
         )
-        
+
+        # On-entry self-heal: if the adapter still has an _active_sessions
+        # entry for this key but the owner task has already exited (done or
+        # cancelled), the lock is stale.  Clear it and fall through to
+        # normal dispatch so the user isn't trapped behind a dead guard —
+        # this is the split-brain tail described in issue #11016.
+        if session_key in self._active_sessions:
+            self._heal_stale_session_lock(session_key)
+
         # Check if there's already an active handler for this session
         if session_key in self._active_sessions:
             # Certain commands must bypass the active-session guard and be
@@ -1710,6 +1943,23 @@ class BasePlatformAdapter(ABC):
             from hermes_cli.commands import should_bypass_active_session
 
             if should_bypass_active_session(cmd):
+                # /stop, /new, /reset must cancel the in-flight adapter task
+                # and preserve ordering of queued follow-ups.  Route those
+                # through the dedicated handoff path that serializes
+                # cancellation + runner response + pending drain.
+                if cmd in ("stop", "new", "reset"):
+                    try:
+                        await self._dispatch_active_session_command(event, session_key, cmd)
+                    except Exception as e:
+                        logger.error(
+                            "[%s] Command '/%s' dispatch failed: %s",
+                            self.name, cmd, e, exc_info=True,
+                        )
+                    return
+
+                # Other bypass commands (/approve, /deny, /status,
+                # /background, /restart) just need direct dispatch — they
+                # don't cancel the running task.
                 logger.debug(
                     "[%s] Command '/%s' bypassing active-session guard for %s",
                     self.name, cmd, session_key,
@@ -1755,19 +2005,9 @@ class BasePlatformAdapter(ABC):
         # starts would also pass the _active_sessions check and spawn a
         # duplicate task.  (grammY sequentialize / aiogram EventIsolation
         # pattern — set the guard synchronously, not inside the task.)
-        self._active_sessions[session_key] = asyncio.Event()
-
-        # Spawn background task to process this message
-        task = asyncio.create_task(self._process_message_background(event, session_key))
-        try:
-            self._background_tasks.add(task)
-        except TypeError:
-            # Some tests stub create_task() with lightweight sentinels that are not
-            # hashable and do not support lifecycle callbacks.
-            return
-        if hasattr(task, "add_done_callback"):
-            task.add_done_callback(self._background_tasks.discard)
-            task.add_done_callback(self._expected_cancelled_tasks.discard)
+        # _start_session_processing installs the guard AND the owner-task
+        # mapping atomically so stale-lock detection works.
+        self._start_session_processing(event, session_key)
     
     @staticmethod
     def _get_human_delay() -> float:
@@ -2127,6 +2367,9 @@ class BasePlatformAdapter(ABC):
                 drain_task = asyncio.create_task(
                     self._process_message_background(late_pending, session_key)
                 )
+                # Hand ownership of the session to the drain task so stale-lock
+                # detection keeps working while it runs.
+                self._session_tasks[session_key] = drain_task
                 try:
                     self._background_tasks.add(drain_task)
                     drain_task.add_done_callback(self._background_tasks.discard)
@@ -2136,9 +2379,14 @@ class BasePlatformAdapter(ABC):
                 # Leave _active_sessions[session_key] populated — the drain
                 # task's own lifecycle will clean it up.
             else:
-                # Clean up session tracking
-                if session_key in self._active_sessions:
-                    del self._active_sessions[session_key]
+                # Clean up session tracking.  Guard-match both deletes so a
+                # reset-like command that already swapped in its own
+                # command_guard (and cancelled us) can't be accidentally
+                # cleared by our unwind.  The command owns the session now.
+                current_task = asyncio.current_task()
+                if current_task is not None and self._session_tasks.get(session_key) is current_task:
+                    del self._session_tasks[session_key]
+                self._release_session_guard(session_key, guard=interrupt_event)
     
     async def cancel_background_tasks(self) -> None:
         """Cancel any in-flight background message-processing tasks.
@@ -2168,6 +2416,7 @@ class BasePlatformAdapter(ABC):
             # will be in self._background_tasks now.  Re-check.
         self._background_tasks.clear()
         self._expected_cancelled_tasks.clear()
+        self._session_tasks.clear()
         self._pending_messages.clear()
         self._active_sessions.clear()
 

gateway/platforms/discord.py

@@ -23,6 +23,7 @@ from typing import Callable, Dict, Optional, Any
 logger = logging.getLogger(__name__)
 
 VALID_THREAD_AUTO_ARCHIVE_MINUTES = {60, 1440, 4320, 10080}
+_DISCORD_COMMAND_SYNC_POLICIES = {"safe", "bulk", "off"}
 
 try:
     import discord
@@ -527,6 +528,7 @@ class DiscordAdapter(BasePlatformAdapter):
         # Reply threading mode: "off" (no replies), "first" (reply on first
         # chunk only, default), "all" (reply-reference on every chunk).
         self._reply_to_mode: str = getattr(config, 'reply_to_mode', 'first') or 'first'
+        self._slash_commands: bool = self.config.extra.get("slash_commands", True)
 
     async def connect(self) -> bool:
         """Connect to Discord and start receiving events."""
@@ -744,7 +746,8 @@ class DiscordAdapter(BasePlatformAdapter):
                     )
 
             # Register slash commands
-            self._register_slash_commands()
+            if self._slash_commands:
+                self._register_slash_commands()
 
             # Start the bot in background
             self._bot_task = asyncio.create_task(self._client.start(self.config.token))
@@ -800,8 +803,27 @@ class DiscordAdapter(BasePlatformAdapter):
         if not self._client:
             return
         try:
-            synced = await asyncio.wait_for(self._client.tree.sync(), timeout=30)
-            logger.info("[%s] Synced %d slash command(s)", self.name, len(synced))
+            sync_policy = self._get_discord_command_sync_policy()
+            if sync_policy == "off":
+                logger.info("[%s] Skipping Discord slash command sync (policy=off)", self.name)
+                return
+
+            if sync_policy == "bulk":
+                synced = await asyncio.wait_for(self._client.tree.sync(), timeout=30)
+                logger.info("[%s] Synced %d slash command(s) via bulk tree sync", self.name, len(synced))
+                return
+
+            summary = await asyncio.wait_for(self._safe_sync_slash_commands(), timeout=30)
+            logger.info(
+                "[%s] Safely reconciled %d slash command(s): unchanged=%d updated=%d recreated=%d created=%d deleted=%d",
+                self.name,
+                summary["total"],
+                summary["unchanged"],
+                summary["updated"],
+                summary["recreated"],
+                summary["created"],
+                summary["deleted"],
+            )
         except asyncio.TimeoutError:
             logger.warning("[%s] Slash command sync timed out after 30s", self.name)
         except asyncio.CancelledError:
@@ -809,6 +831,183 @@ class DiscordAdapter(BasePlatformAdapter):
         except Exception as e:  # pragma: no cover - defensive logging
             logger.warning("[%s] Slash command sync failed: %s", self.name, e, exc_info=True)
 
+    def _get_discord_command_sync_policy(self) -> str:
+        raw = str(os.getenv("DISCORD_COMMAND_SYNC_POLICY", "safe") or "").strip().lower()
+        if raw in _DISCORD_COMMAND_SYNC_POLICIES:
+            return raw
+        if raw:
+            logger.warning(
+                "[%s] Invalid DISCORD_COMMAND_SYNC_POLICY=%r; falling back to 'safe'",
+                self.name,
+                raw,
+            )
+        return "safe"
+
+    def _canonicalize_app_command_payload(self, payload: Dict[str, Any]) -> Dict[str, Any]:
+        """Reduce command payloads to the semantic fields Hermes manages."""
+        contexts = payload.get("contexts")
+        integration_types = payload.get("integration_types")
+        return {
+            "type": int(payload.get("type", 1) or 1),
+            "name": str(payload.get("name", "") or ""),
+            "description": str(payload.get("description", "") or ""),
+            "default_member_permissions": self._normalize_permissions(
+                payload.get("default_member_permissions")
+            ),
+            "dm_permission": bool(payload.get("dm_permission", True)),
+            "nsfw": bool(payload.get("nsfw", False)),
+            "contexts": sorted(int(c) for c in contexts) if contexts else None,
+            "integration_types": (
+                sorted(int(i) for i in integration_types) if integration_types else None
+            ),
+            "options": [
+                self._canonicalize_app_command_option(item)
+                for item in payload.get("options", []) or []
+                if isinstance(item, dict)
+            ],
+        }
+
+    @staticmethod
+    def _normalize_permissions(value: Any) -> Optional[str]:
+        """Discord emits default_member_permissions as str server-side but discord.py
+        sets it as int locally. Normalize to str-or-None so the comparison is stable."""
+        if value is None:
+            return None
+        return str(value)
+
+    def _existing_command_to_payload(self, command: Any) -> Dict[str, Any]:
+        """Build a canonical-ready dict from an AppCommand.
+
+        discord.py's AppCommand.to_dict() does NOT include nsfw,
+        dm_permission, or default_member_permissions (they live only on the
+        attributes). Pull them from the attributes so the canonicalizer sees
+        the real server-side values instead of defaults — otherwise any
+        command using non-default permissions would diff on every startup.
+        """
+        payload = dict(command.to_dict())
+        nsfw = getattr(command, "nsfw", None)
+        if nsfw is not None:
+            payload["nsfw"] = bool(nsfw)
+        guild_only = getattr(command, "guild_only", None)
+        if guild_only is not None:
+            payload["dm_permission"] = not bool(guild_only)
+        default_permissions = getattr(command, "default_member_permissions", None)
+        if default_permissions is not None:
+            payload["default_member_permissions"] = getattr(
+                default_permissions, "value", default_permissions
+            )
+        return payload
+
+    def _canonicalize_app_command_option(self, payload: Dict[str, Any]) -> Dict[str, Any]:
+        return {
+            "type": int(payload.get("type", 0) or 0),
+            "name": str(payload.get("name", "") or ""),
+            "description": str(payload.get("description", "") or ""),
+            "required": bool(payload.get("required", False)),
+            "autocomplete": bool(payload.get("autocomplete", False)),
+            "choices": [
+                {
+                    "name": str(choice.get("name", "") or ""),
+                    "value": choice.get("value"),
+                }
+                for choice in payload.get("choices", []) or []
+                if isinstance(choice, dict)
+            ],
+            "channel_types": list(payload.get("channel_types", []) or []),
+            "min_value": payload.get("min_value"),
+            "max_value": payload.get("max_value"),
+            "min_length": payload.get("min_length"),
+            "max_length": payload.get("max_length"),
+            "options": [
+                self._canonicalize_app_command_option(item)
+                for item in payload.get("options", []) or []
+                if isinstance(item, dict)
+            ],
+        }
+
+    def _patchable_app_command_payload(self, payload: Dict[str, Any]) -> Dict[str, Any]:
+        """Fields supported by discord.py's edit_global_command route."""
+        canonical = self._canonicalize_app_command_payload(payload)
+        return {
+            "name": canonical["name"],
+            "description": canonical["description"],
+            "options": canonical["options"],
+        }
+
+    async def _safe_sync_slash_commands(self) -> Dict[str, int]:
+        """Diff existing global commands and only mutate the commands that changed."""
+        if not self._client:
+            return {
+                "total": 0,
+                "unchanged": 0,
+                "updated": 0,
+                "recreated": 0,
+                "created": 0,
+                "deleted": 0,
+            }
+
+        tree = self._client.tree
+        app_id = getattr(self._client, "application_id", None) or getattr(getattr(self._client, "user", None), "id", None)
+        if not app_id:
+            raise RuntimeError("Discord application ID is unavailable for slash command sync")
+
+        desired_payloads = [command.to_dict(tree) for command in tree.get_commands()]
+        desired_by_key = {
+            (int(payload.get("type", 1) or 1), str(payload.get("name", "") or "").lower()): payload
+            for payload in desired_payloads
+        }
+        existing_commands = await tree.fetch_commands()
+        existing_by_key = {
+            (
+                int(getattr(getattr(command, "type", None), "value", getattr(command, "type", 1)) or 1),
+                str(command.name or "").lower(),
+            ): command
+            for command in existing_commands
+        }
+
+        unchanged = 0
+        updated = 0
+        recreated = 0
+        created = 0
+        deleted = 0
+        http = self._client.http
+
+        for key, desired in desired_by_key.items():
+            current = existing_by_key.pop(key, None)
+            if current is None:
+                await http.upsert_global_command(app_id, desired)
+                created += 1
+                continue
+
+            current_existing_payload = self._existing_command_to_payload(current)
+            current_payload = self._canonicalize_app_command_payload(current_existing_payload)
+            desired_payload = self._canonicalize_app_command_payload(desired)
+            if current_payload == desired_payload:
+                unchanged += 1
+                continue
+
+            if self._patchable_app_command_payload(current_existing_payload) == self._patchable_app_command_payload(desired):
+                await http.delete_global_command(app_id, current.id)
+                await http.upsert_global_command(app_id, desired)
+                recreated += 1
+                continue
+
+            await http.edit_global_command(app_id, current.id, desired)
+            updated += 1
+
+        for current in existing_by_key.values():
+            await http.delete_global_command(app_id, current.id)
+            deleted += 1
+
+        return {
+            "total": len(desired_payloads),
+            "unchanged": unchanged,
+            "updated": updated,
+            "recreated": recreated,
+            "created": created,
+            "deleted": deleted,
+        }
+
     async def _add_reaction(self, message: Any, emoji: str) -> bool:
         """Add an emoji reaction to a Discord message."""
         if not message or not hasattr(message, "add_reaction"):
@@ -2047,10 +2246,6 @@ class DiscordAdapter(BasePlatformAdapter):
         async def slash_usage(interaction: discord.Interaction):
             await self._run_simple_slash(interaction, "/usage")
 
-        @tree.command(name="provider", description="Show available providers")
-        async def slash_provider(interaction: discord.Interaction):
-            await self._run_simple_slash(interaction, "/provider")
-
         @tree.command(name="help", description="Show available commands")
         async def slash_help(interaction: discord.Interaction):
             await self._run_simple_slash(interaction, "/help")
@@ -2520,7 +2715,12 @@ class DiscordAdapter(BasePlatformAdapter):
         return os.getenv("DISCORD_REQUIRE_MENTION", "true").lower() not in ("false", "0", "no", "off")
 
     def _discord_free_response_channels(self) -> set:
-        """Return Discord channel IDs where no bot mention is required."""
+        """Return Discord channel IDs where no bot mention is required.
+
+        A single ``"*"`` entry (either from a list or a comma-separated
+        string) is preserved in the returned set so callers can short-circuit
+        on wildcard membership, consistent with ``allowed_channels``.
+        """
         raw = self.config.extra.get("free_response_channels")
         if raw is None:
             raw = os.getenv("DISCORD_FREE_RESPONSE_CHANNELS", "")
@@ -3013,14 +3213,14 @@ class DiscordAdapter(BasePlatformAdapter):
             allowed_channels_raw = os.getenv("DISCORD_ALLOWED_CHANNELS", "")
             if allowed_channels_raw:
                 allowed_channels = {ch.strip() for ch in allowed_channels_raw.split(",") if ch.strip()}
-                if not (channel_ids & allowed_channels):
+                if "*" not in allowed_channels and not (channel_ids & allowed_channels):
                     logger.debug("[%s] Ignoring message in non-allowed channel: %s", self.name, channel_ids)
                     return
 
             # Check ignored channels - never respond even when mentioned
             ignored_channels_raw = os.getenv("DISCORD_IGNORED_CHANNELS", "")
             ignored_channels = {ch.strip() for ch in ignored_channels_raw.split(",") if ch.strip()}
-            if channel_ids & ignored_channels:
+            if "*" in ignored_channels or (channel_ids & ignored_channels):
                 logger.debug("[%s] Ignoring message in ignored channel: %s", self.name, channel_ids)
                 return
 
@@ -3034,7 +3234,11 @@ class DiscordAdapter(BasePlatformAdapter):
             voice_linked_ids = {str(ch_id) for ch_id in self._voice_text_channels.values()}
             current_channel_id = str(message.channel.id)
             is_voice_linked_channel = current_channel_id in voice_linked_ids
-            is_free_channel = bool(channel_ids & free_channels) or is_voice_linked_channel
+            is_free_channel = (
+                "*" in free_channels
+                or bool(channel_ids & free_channels)
+                or is_voice_linked_channel
+            )
 
             # Skip the mention check if the message is in a thread where
             # the bot has previously participated (auto-created or replied in).
@@ -3667,6 +3871,15 @@ if DISCORD_AVAILABLE:
 
             self.resolved = True
             model_id = interaction.data["values"][0]
+            self.clear_items()
+            await interaction.response.edit_message(
+                embed=discord.Embed(
+                    title="⚙ Switching Model",
+                    description=f"Switching to `{model_id}`...",
+                    color=discord.Color.blue(),
+                ),
+                view=None,
+            )
 
             try:
                 result_text = await self.on_model_selected(
@@ -3677,14 +3890,13 @@ if DISCORD_AVAILABLE:
             except Exception as exc:
                 result_text = f"Error switching model: {exc}"
 
-            self.clear_items()
-            await interaction.response.edit_message(
+            await interaction.edit_original_response(
                 embed=discord.Embed(
                     title="⚙ Model Switched",
                     description=result_text,
                     color=discord.Color.green(),
                 ),
-                view=self,
+                view=None,
             )
 
         async def _on_back(self, interaction: discord.Interaction):

gateway/platforms/feishu.py

@@ -1700,6 +1700,7 @@ class FeishuAdapter(BasePlatformAdapter):
         if not self._client:
             return SendResult(success=False, error="Not connected")
 
+        content = self.format_message(content)
         try:
             msg_type, payload = self._build_outbound_payload(content)
             body = self._build_update_message_body(msg_type=msg_type, content=payload)

gateway/platforms/qqbot/adapter.py

@@ -535,6 +535,9 @@ class QQAdapter(BasePlatformAdapter):
                     quick_disconnect_count = 0
                 else:
                     backoff_idx += 1
+                    if backoff_idx >= MAX_RECONNECT_ATTEMPTS:
+                        logger.error("[%s] Max reconnect attempts reached (QQCloseError)", self._log_tag)
+                        return
 
             except Exception as exc:
                 if not self._running:

gateway/platforms/wecom.py

@@ -508,6 +508,11 @@ class WeComAdapter(BasePlatformAdapter):
         self._remember_chat_req_id(chat_id, self._payload_req_id(payload))
 
         text, reply_text = self._extract_text(body)
+        # Strip leading @mention in group chats so slash commands like
+        # "@BotName /approve" are correctly recognized as "/approve".
+        # Mirrors what the Telegram adapter does (re.sub @botname).
+        if is_group and text:
+            text = re.sub(r"^@\S+\s*", "", text).strip()
         media_urls, media_types = await self._extract_media(body)
         message_type = self._derive_message_type(body, text, media_types)
         has_reply_context = bool(reply_text and (text or media_urls))

gateway/run.py

@@ -14,6 +14,7 @@ Usage:
 """
 
 import asyncio
+import dataclasses
 import json
 import logging
 import os
@@ -349,16 +350,30 @@ _AGENT_PENDING_SENTINEL = object()
 
 
 def _resolve_runtime_agent_kwargs() -> dict:
-    """Resolve provider credentials for gateway-created AIAgent instances."""
+    """Resolve provider credentials for gateway-created AIAgent instances.
+
+    If the primary provider fails with an authentication error, attempt to
+    resolve credentials using the fallback provider chain from config.yaml
+    before giving up.
+    """
     from hermes_cli.runtime_provider import (
         resolve_runtime_provider,
         format_runtime_provider_error,
     )
+    from hermes_cli.auth import AuthError
 
     try:
         runtime = resolve_runtime_provider(
             requested=os.getenv("HERMES_INFERENCE_PROVIDER"),
         )
+    except AuthError as auth_exc:
+        # Primary provider auth failed (expired token, revoked key, etc.).
+        # Try the fallback provider chain before raising.
+        logger.warning("Primary provider auth failed: %s — trying fallback", auth_exc)
+        fb_config = _try_resolve_fallback_provider()
+        if fb_config is not None:
+            return fb_config
+        raise RuntimeError(format_runtime_provider_error(auth_exc)) from auth_exc
     except Exception as exc:
         raise RuntimeError(format_runtime_provider_error(exc)) from exc
 
@@ -373,6 +388,48 @@ def _resolve_runtime_agent_kwargs() -> dict:
     }
 
 
+def _try_resolve_fallback_provider() -> dict | None:
+    """Attempt to resolve credentials from the fallback_model/fallback_providers config."""
+    from hermes_cli.runtime_provider import resolve_runtime_provider
+    try:
+        import yaml as _y
+        cfg_path = _hermes_home / "config.yaml"
+        if not cfg_path.exists():
+            return None
+        with open(cfg_path, encoding="utf-8") as _f:
+            cfg = _y.safe_load(_f) or {}
+        fb = cfg.get("fallback_providers") or cfg.get("fallback_model")
+        if not fb:
+            return None
+        # Normalize to list
+        fb_list = fb if isinstance(fb, list) else [fb]
+        for entry in fb_list:
+            if not isinstance(entry, dict):
+                continue
+            try:
+                runtime = resolve_runtime_provider(
+                    requested=entry.get("provider"),
+                    explicit_base_url=entry.get("base_url"),
+                    explicit_api_key=entry.get("api_key"),
+                )
+                logger.info("Fallback provider resolved: %s", runtime.get("provider"))
+                return {
+                    "api_key": runtime.get("api_key"),
+                    "base_url": runtime.get("base_url"),
+                    "provider": runtime.get("provider"),
+                    "api_mode": runtime.get("api_mode"),
+                    "command": runtime.get("command"),
+                    "args": list(runtime.get("args") or []),
+                    "credential_pool": runtime.get("credential_pool"),
+                }
+            except Exception as fb_exc:
+                logger.debug("Fallback entry %s failed: %s", entry.get("provider"), fb_exc)
+                continue
+    except Exception:
+        pass
+    return None
+
+
 def _build_media_placeholder(event) -> str:
     """Build a text placeholder for media-only events so they aren't dropped.
 
@@ -1551,27 +1608,23 @@ class GatewayRunner:
             )
             return True
 
-        # --- Normal busy case (agent actively running a task) ---
-        # The user sent a message while the agent is working.  Interrupt the
-        # agent immediately so it stops the current tool-calling loop and
-        # processes the new message.  The pending message is stored in the
-        # adapter so the base adapter picks it up once the interrupted run
-        # returns.  A brief ack tells the user what's happening (debounced
-        # to avoid spam when they fire multiple messages quickly).
-
+        # Normal busy case (agent actively running a task)
         adapter = self.adapters.get(event.source.platform)
         if not adapter:
             return False  # let default path handle it
 
         # Store the message so it's processed as the next turn after the
-        # interrupt causes the current run to exit.
+        # current run finishes (or is interrupted).
         from gateway.platforms.base import merge_pending_message_event
         merge_pending_message_event(adapter._pending_messages, session_key, event)
 
-        # Interrupt the running agent — this aborts in-flight tool calls and
-        # causes the agent loop to exit at the next check point.
+        is_queue_mode = self._busy_input_mode == "queue"
+
+        # If not in queue mode, interrupt the running agent immediately.
+        # This aborts in-flight tool calls and causes the agent loop to exit
+        # at the next check point.
         running_agent = self._running_agents.get(session_key)
-        if running_agent and running_agent is not _AGENT_PENDING_SENTINEL:
+        if not is_queue_mode and running_agent and running_agent is not _AGENT_PENDING_SENTINEL:
             try:
                 running_agent.interrupt(event.text)
             except Exception:
@@ -1583,7 +1636,7 @@ class GatewayRunner:
         now = time.time()
         last_ack = self._busy_ack_ts.get(session_key, 0)
         if now - last_ack < _BUSY_ACK_COOLDOWN:
-            return True  # interrupt sent, ack already delivered recently
+            return True  # interrupt sent (if not queue), ack already delivered recently
 
         self._busy_ack_ts[session_key] = now
 
@@ -1608,10 +1661,16 @@ class GatewayRunner:
                 pass
 
         status_detail = f" ({', '.join(status_parts)})" if status_parts else ""
-        message = (
-            f"⚡ Interrupting current task{status_detail}. "
-            f"I'll respond to your message shortly."
-        )
+        if is_queue_mode:
+            message = (
+                f"⏳ Queued for the next turn{status_detail}. "
+                f"I'll respond once the current task finishes."
+            )
+        else:
+            message = (
+                f"⚡ Interrupting current task{status_detail}. "
+                f"I'll respond to your message shortly."
+            )
 
         thread_meta = {"thread_id": event.source.thread_id} if event.source.thread_id else None
         try:
@@ -2307,6 +2366,17 @@ class GatewayRunner:
                 for key, entry in _expired_entries:
                     try:
                         await self._async_flush_memories(entry.session_id, key)
+                        try:
+                            from hermes_cli.plugins import invoke_hook as _invoke_hook
+                            _parts = key.split(":")
+                            _platform = _parts[2] if len(_parts) > 2 else ""
+                            _invoke_hook(
+                                "on_session_finalize",
+                                session_id=entry.session_id,
+                                platform=_platform,
+                            )
+                        except Exception:
+                            pass
                         # Shut down memory provider and close tool resources
                         # on the cached agent.  Idle agents live in
                         # _agent_cache (not _running_agents), so look there.
@@ -2560,6 +2630,40 @@ class GatewayRunner:
             return
 
         async def _stop_impl() -> None:
+            def _kill_tool_subprocesses(phase: str) -> None:
+                """Kill tool subprocesses + tear down terminal envs + browsers.
+
+                Called twice in the shutdown path: once eagerly after a
+                drain timeout forces agent interrupt (so we reclaim bash/
+                sleep children before systemd TimeoutStopSec escalates to
+                SIGKILL on the cgroup — #8202), and once as a final
+                catch-all at the end of _stop_impl() for the graceful
+                path or anything respawned mid-teardown.
+
+                All steps are best-effort; exceptions are swallowed so
+                one subsystem's failure doesn't block the rest.
+                """
+                try:
+                    from tools.process_registry import process_registry
+                    _killed = process_registry.kill_all()
+                    if _killed:
+                        logger.info(
+                            "Shutdown (%s): killed %d tool subprocess(es)",
+                            phase, _killed,
+                        )
+                except Exception as _e:
+                    logger.debug("process_registry.kill_all (%s) error: %s", phase, _e)
+                try:
+                    from tools.terminal_tool import cleanup_all_environments
+                    cleanup_all_environments()
+                except Exception as _e:
+                    logger.debug("cleanup_all_environments (%s) error: %s", phase, _e)
+                try:
+                    from tools.browser_tool import cleanup_all_browsers
+                    cleanup_all_browsers()
+                except Exception as _e:
+                    logger.debug("cleanup_all_browsers (%s) error: %s", phase, _e)
+
             logger.info(
                 "Stopping gateway%s...",
                 " for restart" if self._restart_requested else "",
@@ -2621,6 +2725,16 @@ class GatewayRunner:
                     self._update_runtime_status("draining")
                     await asyncio.sleep(0.1)
 
+                # Kill lingering tool subprocesses NOW, before we spend more
+                # budget on adapter disconnect / session DB close.  Under
+                # systemd (TimeoutStopSec bounded by drain_timeout+headroom),
+                # deferring this to the end of stop() risks systemd escalating
+                # to SIGKILL on the cgroup first — at which point bash/sleep
+                # children left behind by an interrupted terminal tool get
+                # killed by systemd instead of us (issue #8202).  The final
+                # catch-all cleanup below still runs for the graceful path.
+                _kill_tool_subprocesses("post-interrupt")
+
             if self._restart_requested and self._restart_detached:
                 try:
                     await self._launch_detached_restart_command()
@@ -2656,22 +2770,13 @@ class GatewayRunner:
             self._shutdown_event.set()
 
             # Global cleanup: kill any remaining tool subprocesses not tied
-            # to a specific agent (catch-all for zombie prevention).
-            try:
-                from tools.process_registry import process_registry
-                process_registry.kill_all()
-            except Exception:
-                pass
-            try:
-                from tools.terminal_tool import cleanup_all_environments
-                cleanup_all_environments()
-            except Exception:
-                pass
-            try:
-                from tools.browser_tool import cleanup_all_browsers
-                cleanup_all_browsers()
-            except Exception:
-                pass
+            # to a specific agent (catch-all for zombie prevention). On the
+            # drain-timeout path we already did this earlier after agent
+            # interrupt — this second call catches (a) the graceful path
+            # where drain succeeded without interrupt, and (b) anything
+            # that got respawned between the earlier call and adapter
+            # disconnect (defense in depth; safe to call repeatedly).
+            _kill_tool_subprocesses("final-cleanup")
 
             # Close SQLite session DBs so the WAL write lock is released.
             # Without this, --replace and similar restart flows leave the
@@ -3108,7 +3213,50 @@ class GatewayRunner:
 
         # Internal events (e.g. background-process completion notifications)
         # are system-generated and must skip user authorization.
-        if getattr(event, "internal", False):
+        is_internal = bool(getattr(event, "internal", False))
+
+        # Fire pre_gateway_dispatch plugin hook for user-originated messages.
+        # Plugins receive the MessageEvent and may return a dict influencing flow:
+        #   {"action": "skip",    "reason": ...}    -> drop (no reply, plugin handled)
+        #   {"action": "rewrite", "text":  ...}     -> replace event.text, continue
+        #   {"action": "allow"}   /   None          -> normal dispatch
+        # Hook runs BEFORE auth so plugins can handle unauthorized senders
+        # (e.g. customer handover ingest) without triggering the pairing flow.
+        if not is_internal:
+            try:
+                from hermes_cli.plugins import invoke_hook as _invoke_hook
+                _hook_results = _invoke_hook(
+                    "pre_gateway_dispatch",
+                    event=event,
+                    gateway=self,
+                    session_store=self.session_store,
+                )
+            except Exception as _hook_exc:
+                logger.warning("pre_gateway_dispatch invocation failed: %s", _hook_exc)
+                _hook_results = []
+
+            for _result in _hook_results:
+                if not isinstance(_result, dict):
+                    continue
+                _action = _result.get("action")
+                if _action == "skip":
+                    logger.info(
+                        "pre_gateway_dispatch skip: reason=%s platform=%s chat=%s",
+                        _result.get("reason"),
+                        source.platform.value if source.platform else "unknown",
+                        source.chat_id or "unknown",
+                    )
+                    return None
+                if _action == "rewrite":
+                    _new_text = _result.get("text")
+                    if isinstance(_new_text, str):
+                        event = dataclasses.replace(event, text=_new_text)
+                        source = event.source
+                    break
+                if _action == "allow":
+                    break
+
+        if is_internal:
             pass
         elif source.user_id is None:
             # Messages with no user identity (Telegram service messages,
@@ -3405,7 +3553,7 @@ class GatewayRunner:
             # running-agent guard. Reject gracefully rather than falling
             # through to interrupt + discard. Without this, commands
             # like /model, /reasoning, /voice, /insights, /title,
-            # /resume, /retry, /undo, /compress, /usage, /provider,
+            # /resume, /retry, /undo, /compress, /usage,
             # /reload-mcp, /sethome, /reset (all registered as Discord
             # slash commands) would interrupt the agent AND get
             # silently discarded by the slash-command safety net,
@@ -3592,34 +3740,9 @@ class GatewayRunner:
         if canonical == "model":
             return await self._handle_model_command(event)
 
-        if canonical == "provider":
-            return await self._handle_provider_command(event)
-        
         if canonical == "personality":
             return await self._handle_personality_command(event)
 
-        if canonical == "plan":
-            try:
-                from agent.skill_commands import build_plan_path, build_skill_invocation_message
-
-                user_instruction = event.get_command_args().strip()
-                plan_path = build_plan_path(user_instruction)
-                event.text = build_skill_invocation_message(
-                    "/plan",
-                    user_instruction,
-                    task_id=_quick_key,
-                    runtime_note=(
-                        "Save the markdown plan with write_file to this exact relative path "
-                        f"inside the active workspace/backend cwd: {plan_path}"
-                    ),
-                )
-                if not event.text:
-                    return "Failed to load the bundled /plan skill."
-                canonical = None
-            except Exception as e:
-                logger.exception("Failed to prepare /plan command")
-                return f"Failed to enter plan mode: {e}"
-        
         if canonical == "retry":
             return await self._handle_retry_command(event)
         
@@ -5484,6 +5607,7 @@ class GatewayRunner:
                 try:
                     providers = list_authenticated_providers(
                         current_provider=current_provider,
+                        current_base_url=current_base_url,
                         user_providers=user_provs,
                         custom_providers=custom_provs,
                         max_models=50,
@@ -5595,6 +5719,7 @@ class GatewayRunner:
             try:
                 providers = list_authenticated_providers(
                     current_provider=current_provider,
+                    current_base_url=current_base_url,
                     user_providers=user_provs,
                     custom_providers=custom_provs,
                     max_models=5,
@@ -5740,63 +5865,6 @@ class GatewayRunner:
 
         return "\n".join(lines)
 
-    async def _handle_provider_command(self, event: MessageEvent) -> str:
-        """Handle /provider command - show available providers."""
-        import yaml
-        from hermes_cli.models import (
-            list_available_providers,
-            normalize_provider,
-            _PROVIDER_LABELS,
-        )
-
-        # Resolve current provider from config
-        current_provider = "openrouter"
-        model_cfg = {}
-        config_path = _hermes_home / 'config.yaml'
-        try:
-            if config_path.exists():
-                with open(config_path, encoding="utf-8") as f:
-                    cfg = yaml.safe_load(f) or {}
-                model_cfg = cfg.get("model", {})
-                if isinstance(model_cfg, dict):
-                    current_provider = model_cfg.get("provider", current_provider)
-        except Exception:
-            pass
-
-        current_provider = normalize_provider(current_provider)
-        if current_provider == "auto":
-            try:
-                from hermes_cli.auth import resolve_provider as _resolve_provider
-                current_provider = _resolve_provider(current_provider)
-            except Exception:
-                current_provider = "openrouter"
-
-        # Detect custom endpoint from config base_url
-        if current_provider == "openrouter":
-            _cfg_base = model_cfg.get("base_url", "") if isinstance(model_cfg, dict) else ""
-            if _cfg_base and "openrouter.ai" not in _cfg_base:
-                current_provider = "custom"
-
-        current_label = _PROVIDER_LABELS.get(current_provider, current_provider)
-
-        lines = [
-            f"🔌 **Current provider:** {current_label} (`{current_provider}`)",
-            "",
-            "**Available providers:**",
-        ]
-
-        providers = list_available_providers()
-        for p in providers:
-            marker = " ← active" if p["id"] == current_provider else ""
-            auth = "✅" if p["authenticated"] else "❌"
-            aliases = f"  _(also: {', '.join(p['aliases'])})_" if p["aliases"] else ""
-            lines.append(f"{auth} `{p['id']}` — {p['label']}{aliases}{marker}")
-
-        lines.append("")
-        lines.append("Switch: `/model provider:model-name`")
-        lines.append("Setup: `hermes setup`")
-        return "\n".join(lines)
-    
     async def _handle_personality_command(self, event: MessageEvent) -> str:
         """Handle /personality command - list or set a personality."""
         import yaml
@@ -7063,10 +7131,7 @@ class GatewayRunner:
                 tmp_agent._print_fn = lambda *a, **kw: None
 
                 compressor = tmp_agent.context_compressor
-                compress_start = compressor.protect_first_n
-                compress_start = compressor._align_boundary_forward(msgs, compress_start)
-                compress_end = compressor._find_tail_cut_by_tokens(msgs, compress_start)
-                if compress_start >= compress_end:
+                if not compressor.has_content_to_compress(msgs):
                     return "Nothing to compress yet (the transcript is still all protected context)."
 
                 loop = asyncio.get_running_loop()
@@ -8663,7 +8728,12 @@ class GatewayRunner:
         override = self._session_model_overrides.get(session_key)
         return override is not None and override.get("model") == agent_model
 
-    def _release_running_agent_state(self, session_key: str) -> None:
+    def _release_running_agent_state(
+        self,
+        session_key: str,
+        *,
+        run_generation: Optional[int] = None,
+    ) -> bool:
         """Pop ALL per-running-agent state entries for ``session_key``.
 
         Replaces ad-hoc ``del self._running_agents[key]`` calls scattered
@@ -8679,13 +8749,25 @@ class GatewayRunner:
         across turns (``_session_model_overrides``, ``_voice_mode``,
         ``_pending_approvals``, ``_update_prompt_pending``) is NOT
         touched here — those have their own lifecycles.
+
+        When ``run_generation`` is provided, only clear the slot if that
+        generation is still current for the session.  This prevents an
+        older async run whose generation was bumped by /stop or /new from
+        clobbering a newer run's state during its own unwind.  Returns
+        True when the slot was cleared, False when an ownership guard
+        blocked it.
         """
         if not session_key:
-            return
+            return False
+        if run_generation is not None and not self._is_session_run_current(
+            session_key, run_generation
+        ):
+            return False
         self._running_agents.pop(session_key, None)
         self._running_agents_ts.pop(session_key, None)
         if hasattr(self, "_busy_ack_ts"):
             self._busy_ack_ts.pop(session_key, None)
+        return True
 
     def _clear_session_boundary_security_state(self, session_key: str) -> None:
         """Clear approval state that must not survive a real conversation switch."""
@@ -10247,10 +10329,24 @@ class GatewayRunner:
             # Wait for agent to be created
             while agent_holder[0] is None:
                 await asyncio.sleep(0.05)
-            if session_key:
-                self._running_agents[session_key] = agent_holder[0]
-                if self._draining:
-                    self._update_runtime_status("draining")
+            if not session_key:
+                return
+            # Only promote the sentinel to the real agent if this run is still
+            # current.  If /stop or /new bumped the generation while we were
+            # spinning up, leave the newer run's slot alone — we'll be
+            # discarded by the stale-result check in _handle_message_with_agent.
+            if run_generation is not None and not self._is_session_run_current(
+                session_key, run_generation
+            ):
+                logger.info(
+                    "Skipping stale agent promotion for %s — generation %s is no longer current",
+                    (session_key or "")[:20],
+                    run_generation,
+                )
+                return
+            self._running_agents[session_key] = agent_holder[0]
+            if self._draining:
+                self._update_runtime_status("draining")
         
         tracking_task = asyncio.create_task(track_agent())
         
@@ -10305,9 +10401,9 @@ class GatewayRunner:
         # Periodic "still working" notifications for long-running tasks.
         # Fires every N seconds so the user knows the agent hasn't died.
         # Config: agent.gateway_notify_interval in config.yaml, or
-        # HERMES_AGENT_NOTIFY_INTERVAL env var.  Default 600s (10 min).
+        # HERMES_AGENT_NOTIFY_INTERVAL env var.  Default 180s (3 min).
         # 0 = disable notifications.
-        _NOTIFY_INTERVAL_RAW = float(os.getenv("HERMES_AGENT_NOTIFY_INTERVAL", 600))
+        _NOTIFY_INTERVAL_RAW = float(os.getenv("HERMES_AGENT_NOTIFY_INTERVAL", 180))
         _NOTIFY_INTERVAL = _NOTIFY_INTERVAL_RAW if _NOTIFY_INTERVAL_RAW > 0 else None
         _notify_start = time.time()
 
@@ -10756,7 +10852,14 @@ class GatewayRunner:
             # Clean up tracking
             tracking_task.cancel()
             if session_key:
-                self._release_running_agent_state(session_key)
+                # Only release the slot if this run's generation still owns
+                # it.  A /stop or /new that bumped the generation while we
+                # were unwinding has already installed its own state; this
+                # guard prevents an old run from clobbering it on the way
+                # out.
+                self._release_running_agent_state(
+                    session_key, run_generation=run_generation
+                )
             if self._draining:
                 self._update_runtime_status("draining")
             
@@ -10879,6 +10982,7 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
     from gateway.status import (
         acquire_gateway_runtime_lock,
         get_running_pid,
+        get_process_start_time,
         release_gateway_runtime_lock,
         remove_pid_file,
         terminate_pid,
@@ -10886,6 +10990,7 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
     existing_pid = get_running_pid()
     if existing_pid is not None and existing_pid != os.getpid():
         if replace:
+            existing_start_time = get_process_start_time(existing_pid)
             logger.info(
                 "Replacing existing gateway instance (PID %d) with --replace.",
                 existing_pid,
@@ -10954,7 +11059,10 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
             # leaving stale lock files that block the new gateway from starting.
             try:
                 from gateway.status import release_all_scoped_locks
-                _released = release_all_scoped_locks()
+                _released = release_all_scoped_locks(
+                    owner_pid=existing_pid,
+                    owner_start_time=existing_start_time,
+                )
                 if _released:
                     logger.info("Released %d stale scoped lock(s) from old gateway.", _released)
             except Exception:

gateway/status.py

@@ -113,6 +113,11 @@ def _get_process_start_time(pid: int) -> Optional[int]:
         return None
 
 
+def get_process_start_time(pid: int) -> Optional[int]:
+    """Public wrapper for retrieving a process start time when available."""
+    return _get_process_start_time(pid)
+
+
 def _read_process_cmdline(pid: int) -> Optional[str]:
     """Return the process command line as a space-separated string."""
     cmdline_path = Path(f"/proc/{pid}/cmdline")
@@ -496,7 +501,8 @@ def acquire_scoped_lock(scope: str, identity: str, metadata: Optional[dict[str,
         if not stale:
             try:
                 os.kill(existing_pid, 0)
-            except (ProcessLookupError, PermissionError):
+            except (ProcessLookupError, PermissionError, OSError):
+                # Windows raises OSError with WinError 87 for invalid pid check
                 stale = True
             else:
                 current_start = _get_process_start_time(existing_pid)
@@ -561,17 +567,43 @@ def release_scoped_lock(scope: str, identity: str) -> None:
         pass
 
 
-def release_all_scoped_locks() -> int:
-    """Remove all scoped lock files in the lock directory.
+def release_all_scoped_locks(
+    *,
+    owner_pid: Optional[int] = None,
+    owner_start_time: Optional[int] = None,
+) -> int:
+    """Remove scoped lock files in the lock directory.
 
     Called during --replace to clean up stale locks left by stopped/killed
-    gateway processes that did not release their locks gracefully.
+    gateway processes that did not release their locks gracefully. When an
+    ``owner_pid`` is provided, only lock records belonging to that gateway
+    process are removed. ``owner_start_time`` further narrows the match to
+    protect against PID reuse.
+
+    When no owner is provided, preserves the legacy behavior and removes every
+    scoped lock file in the directory.
+
     Returns the number of lock files removed.
     """
     lock_dir = _get_lock_dir()
     removed = 0
     if lock_dir.exists():
         for lock_file in lock_dir.glob("*.lock"):
+            if owner_pid is not None:
+                record = _read_json_file(lock_file)
+                if not isinstance(record, dict):
+                    continue
+                try:
+                    record_pid = int(record.get("pid"))
+                except (TypeError, ValueError):
+                    continue
+                if record_pid != owner_pid:
+                    continue
+                if (
+                    owner_start_time is not None
+                    and record.get("start_time") != owner_start_time
+                ):
+                    continue
             try:
                 lock_file.unlink(missing_ok=True)
                 removed += 1
@@ -743,6 +775,10 @@ def get_running_pid(
             if _record_looks_like_gateway(record):
                 return pid
             continue
+        except OSError:
+            # Windows raises OSError with WinError 87 for an invalid pid
+            # (process is definitely gone). Treat as "process doesn't exist".
+            continue
 
         recorded_start = record.get("start_time")
         current_start = _get_process_start_time(pid)

hermes_cli/__init__.py

@@ -11,5 +11,5 @@ Provides subcommands for:
 - hermes cron          - Manage cron jobs
 """
 
-__version__ = "0.10.0"
-__release_date__ = "2026.4.16"
+__version__ = "0.11.0"
+__release_date__ = "2026.4.23"

hermes_cli/auth_commands.py

@@ -110,18 +110,40 @@ def _display_source(source: str) -> str:
     return source.split(":", 1)[1] if source.startswith("manual:") else source
 
 
+def _classify_exhausted_status(entry) -> tuple[str, bool]:
+    code = getattr(entry, "last_error_code", None)
+    reason = str(getattr(entry, "last_error_reason", "") or "").strip().lower()
+    message = str(getattr(entry, "last_error_message", "") or "").strip().lower()
+
+    if code == 429 or any(token in reason for token in ("rate_limit", "usage_limit", "quota", "exhausted")) or any(
+        token in message for token in ("rate limit", "usage limit", "quota", "too many requests")
+    ):
+        return "rate-limited", True
+
+    if code in {401, 403} or any(token in reason for token in ("invalid_token", "invalid_grant", "unauthorized", "forbidden", "auth")) or any(
+        token in message for token in ("unauthorized", "forbidden", "expired", "revoked", "invalid token", "authentication")
+    ):
+        return "auth failed", False
+
+    return "exhausted", True
+
+
+
 def _format_exhausted_status(entry) -> str:
     if entry.last_status != STATUS_EXHAUSTED:
         return ""
+    label, show_retry_window = _classify_exhausted_status(entry)
     reason = getattr(entry, "last_error_reason", None)
     reason_text = f" {reason}" if isinstance(reason, str) and reason.strip() else ""
     code = f" ({entry.last_error_code})" if entry.last_error_code else ""
+    if not show_retry_window:
+        return f" {label}{reason_text}{code} (re-auth may be required)"
     exhausted_until = _exhausted_until(entry)
     if exhausted_until is None:
-        return f" exhausted{reason_text}{code}"
+        return f" {label}{reason_text}{code}"
     remaining = max(0, int(math.ceil(exhausted_until - time.time())))
     if remaining <= 0:
-        return f" exhausted{reason_text}{code} (ready to retry)"
+        return f" {label}{reason_text}{code} (ready to retry)"
     minutes, seconds = divmod(remaining, 60)
     hours, minutes = divmod(minutes, 60)
     days, hours = divmod(hours, 24)
@@ -133,7 +155,7 @@ def _format_exhausted_status(entry) -> str:
         wait = f"{minutes}m {seconds}s"
     else:
         wait = f"{seconds}s"
-    return f" exhausted{reason_text}{code} ({wait} left)"
+    return f" {label}{reason_text}{code} ({wait} left)"
 
 
 def auth_add_command(args) -> None:
@@ -386,6 +408,44 @@ def auth_reset_command(args) -> None:
     print(f"Reset status on {count} {provider} credentials")
 
 
+def auth_status_command(args) -> None:
+    provider = _normalize_provider(getattr(args, "provider", "") or "")
+    if not provider:
+        raise SystemExit("Provider is required. Example: `hermes auth status spotify`.")
+    status = auth_mod.get_auth_status(provider)
+    if not status.get("logged_in"):
+        reason = status.get("error")
+        if reason:
+            print(f"{provider}: logged out ({reason})")
+        else:
+            print(f"{provider}: logged out")
+        return
+
+    print(f"{provider}: logged in")
+    for key in ("auth_type", "client_id", "redirect_uri", "scope", "expires_at", "api_base_url"):
+        value = status.get(key)
+        if value:
+            print(f"  {key}: {value}")
+
+
+def auth_logout_command(args) -> None:
+    auth_mod.logout_command(SimpleNamespace(provider=getattr(args, "provider", None)))
+
+
+def auth_spotify_command(args) -> None:
+    action = str(getattr(args, "spotify_action", "") or "login").strip().lower()
+    if action in {"", "login"}:
+        auth_mod.login_spotify_command(args)
+        return
+    if action == "status":
+        auth_status_command(SimpleNamespace(provider="spotify"))
+        return
+    if action == "logout":
+        auth_logout_command(SimpleNamespace(provider="spotify"))
+        return
+    raise SystemExit(f"Unknown Spotify auth action: {action}")
+
+
 def _interactive_auth() -> None:
     """Interactive credential pool management when `hermes auth` is called bare."""
     # Show current pool status first
@@ -583,5 +643,14 @@ def auth_command(args) -> None:
     if action == "reset":
         auth_reset_command(args)
         return
+    if action == "status":
+        auth_status_command(args)
+        return
+    if action == "logout":
+        auth_logout_command(args)
+        return
+    if action == "spotify":
+        auth_spotify_command(args)
+        return
     # No subcommand — launch interactive mode
     _interactive_auth()

hermes_cli/banner.py

@@ -238,6 +238,52 @@ def get_git_banner_state(repo_dir: Optional[Path] = None) -> Optional[dict]:
     return {"upstream": upstream, "local": local, "ahead": max(ahead, 0)}
 
 
+_RELEASE_URL_BASE = "https://github.com/NousResearch/hermes-agent/releases/tag"
+_latest_release_cache: Optional[tuple] = None  # (tag, url) once resolved
+
+
+def get_latest_release_tag(repo_dir: Optional[Path] = None) -> Optional[tuple]:
+    """Return ``(tag, release_url)`` for the latest git tag, or None.
+
+    Local-only — runs ``git describe --tags --abbrev=0`` against the
+    Hermes checkout. Cached per-process. Release URL always points at the
+    canonical NousResearch/hermes-agent repo (forks don't get a link).
+    """
+    global _latest_release_cache
+    if _latest_release_cache is not None:
+        return _latest_release_cache or None
+
+    repo_dir = repo_dir or _resolve_repo_dir()
+    if repo_dir is None:
+        _latest_release_cache = ()  # falsy sentinel — skip future lookups
+        return None
+
+    try:
+        result = subprocess.run(
+            ["git", "describe", "--tags", "--abbrev=0"],
+            capture_output=True,
+            text=True,
+            timeout=3,
+            cwd=str(repo_dir),
+        )
+    except Exception:
+        _latest_release_cache = ()
+        return None
+
+    if result.returncode != 0:
+        _latest_release_cache = ()
+        return None
+
+    tag = (result.stdout or "").strip()
+    if not tag:
+        _latest_release_cache = ()
+        return None
+
+    url = f"{_RELEASE_URL_BASE}/{tag}"
+    _latest_release_cache = (tag, url)
+    return _latest_release_cache
+
+
 def format_banner_version_label() -> str:
     """Return the version label shown in the startup banner title."""
     base = f"Hermes Agent v{VERSION} ({RELEASE_DATE})"
@@ -519,9 +565,16 @@ def build_welcome_banner(console: Console, model: str, cwd: str,
     agent_name = _skin_branding("agent_name", "Hermes Agent")
     title_color = _skin_color("banner_title", "#FFD700")
     border_color = _skin_color("banner_border", "#CD7F32")
+    version_label = format_banner_version_label()
+    release_info = get_latest_release_tag()
+    if release_info:
+        _tag, _url = release_info
+        title_markup = f"[bold {title_color}][link={_url}]{version_label}[/link][/]"
+    else:
+        title_markup = f"[bold {title_color}]{version_label}[/]"
     outer_panel = Panel(
         layout_table,
-        title=f"[bold {title_color}]{format_banner_version_label()}[/]",
+        title=title_markup,
         border_style=border_color,
         padding=(0, 2),
     )

hermes_cli/codex_models.py

@@ -12,6 +12,7 @@ import os
 logger = logging.getLogger(__name__)
 
 DEFAULT_CODEX_MODELS: List[str] = [
+    "gpt-5.5",
     "gpt-5.4-mini",
     "gpt-5.4",
     "gpt-5.3-codex",
@@ -21,6 +22,7 @@ DEFAULT_CODEX_MODELS: List[str] = [
 ]
 
 _FORWARD_COMPAT_TEMPLATE_MODELS: List[tuple[str, tuple[str, ...]]] = [
+    ("gpt-5.5", ("gpt-5.4", "gpt-5.4-mini", "gpt-5.3-codex")),
     ("gpt-5.4-mini", ("gpt-5.3-codex", "gpt-5.2-codex")),
     ("gpt-5.4", ("gpt-5.3-codex", "gpt-5.2-codex")),
     ("gpt-5.3-codex", ("gpt-5.2-codex",)),

hermes_cli/commands.py

@@ -77,7 +77,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("rollback", "List or restore filesystem checkpoints", "Session",
                args_hint="[number]"),
     CommandDef("snapshot", "Create or restore state snapshots of Hermes config/state", "Session",
-               aliases=("snap",), args_hint="[create|restore <id>|prune]"),
+               cli_only=True, aliases=("snap",), args_hint="[create|restore <id>|prune]"),
     CommandDef("stop", "Kill all running background processes", "Session"),
     CommandDef("approve", "Approve a pending dangerous command", "Session",
                gateway_only=True, args_hint="[session|always]"),
@@ -104,9 +104,8 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("config", "Show current configuration", "Configuration",
                cli_only=True),
     CommandDef("model", "Switch model for this session", "Configuration", args_hint="[model] [--provider name] [--global]"),
-    CommandDef("provider", "Show available providers and current provider",
-               "Configuration"),
-    CommandDef("gquota", "Show Google Gemini Code Assist quota usage", "Info"),
+    CommandDef("gquota", "Show Google Gemini Code Assist quota usage", "Info",
+               cli_only=True),
 
     CommandDef("personality", "Set a predefined personality", "Configuration",
                args_hint="[name]"),
@@ -124,7 +123,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
                args_hint="[normal|fast|status]",
                subcommands=("normal", "fast", "status", "on", "off")),
     CommandDef("skin", "Show or change the display skin/theme", "Configuration",
-               args_hint="[name]"),
+               cli_only=True, args_hint="[name]"),
     CommandDef("voice", "Toggle voice mode", "Configuration",
                args_hint="[on|off|tts|status]", subcommands=("on", "off", "tts", "status")),
 
@@ -139,7 +138,8 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("cron", "Manage scheduled tasks", "Tools & Skills",
                cli_only=True, args_hint="[subcommand]",
                subcommands=("list", "add", "create", "edit", "pause", "resume", "run", "remove")),
-    CommandDef("reload", "Reload .env variables into the running session", "Tools & Skills"),
+    CommandDef("reload", "Reload .env variables into the running session", "Tools & Skills",
+               cli_only=True),
     CommandDef("reload-mcp", "Reload MCP servers from config", "Tools & Skills",
                aliases=("reload_mcp",)),
     CommandDef("browser", "Connect browser tools to your live Chrome via CDP", "Tools & Skills",
@@ -317,7 +317,7 @@ def should_bypass_active_session(command_name: str | None) -> bool:
     safety net in gateway.run discards any command text that reaches
     the pending queue — which meant a mid-run /model (or /reasoning,
     /voice, /insights, /title, /resume, /retry, /undo, /compress,
-    /usage, /provider, /reload-mcp, /sethome, /reset) would silently
+    /usage, /reload-mcp, /sethome, /reset) would silently
     interrupt the agent AND get discarded, producing a zero-char
     response. See issue #5057 / PRs #6252, #10370, #4665.
 

hermes_cli/config.py

@@ -361,6 +361,15 @@ DEFAULT_CONFIG = {
         # to finish, then interrupts any remaining runs after the timeout.
         # 0 = no drain, interrupt immediately.
         "restart_drain_timeout": 60,
+        # Max app-level retry attempts for API errors (connection drops,
+        # provider timeouts, 5xx, etc.) before the agent surfaces the
+        # failure.  The OpenAI SDK already does its own low-level retries
+        # (max_retries=2 default) for transient network errors; this is
+        # the Hermes-level retry loop that wraps the whole call.  Lower
+        # this to 1 if you use fallback providers and want fast failover
+        # on flaky primaries; raise it if you prefer to tolerate longer
+        # provider hiccups on a single provider.
+        "api_max_retries": 3,
         "service_tier": "",
         # Tool-use enforcement: injects system prompt guidance that tells the
         # model to actually call tools instead of describing intended actions.
@@ -375,7 +384,11 @@ DEFAULT_CONFIG = {
         # Periodic "still working" notification interval (seconds).
         # Sends a status message every N seconds so the user knows the
         # agent hasn't died during long tasks.  0 = disable notifications.
-        "gateway_notify_interval": 600,
+        # Lower values mean faster feedback on slow tasks but more chat
+        # noise; 180s is a compromise that catches spinning weak-model runs
+        # (60+ tool iterations with tiny output) before users assume the
+        # bot is dead and /restart.
+        "gateway_notify_interval": 180,
     },
     
     "terminal": {
@@ -394,17 +407,23 @@ DEFAULT_CONFIG = {
         # (bash doesn't source bashrc in non-interactive login mode) or
         # zsh-specific files like ``~/.zshrc`` / ``~/.zprofile``.
         # Paths support ``~`` / ``${VAR}``. Missing files are silently
-        # skipped. When empty, Hermes auto-appends ``~/.bashrc`` if the
+        # skipped. When empty, Hermes auto-sources ``~/.profile``,
+        # ``~/.bash_profile``, and ``~/.bashrc`` (in that order) if the
         # snapshot shell is bash (this is the ``auto_source_bashrc``
         # behaviour — disable with that key if you want strict login-only
         # semantics).
         "shell_init_files": [],
-        # When true (default), Hermes sources ``~/.bashrc`` in the login
-        # shell used to build the environment snapshot.  This captures
-        # PATH additions, shell functions, and aliases defined in the
-        # user's bashrc — which a plain ``bash -l -c`` would otherwise
-        # miss because bash skips bashrc in non-interactive login mode.
-        # Turn this off if you have a bashrc that misbehaves when sourced
+        # When true (default), Hermes sources the user's shell rc files
+        # (``~/.profile``, ``~/.bash_profile``, ``~/.bashrc``) in the
+        # login shell used to build the environment snapshot. This
+        # captures PATH additions, shell functions, and aliases — which a
+        # plain ``bash -l -c`` would otherwise miss because bash skips
+        # bashrc in non-interactive login mode, and because a default
+        # Debian/Ubuntu ``~/.bashrc`` short-circuits on non-interactive
+        # sources. ``~/.profile`` and ``~/.bash_profile`` are tried first
+        # because ``n`` / ``nvm`` / ``asdf`` installers typically write
+        # their PATH exports there without an interactivity guard. Turn
+        # this off if your rc files misbehave when sourced
         # non-interactively (e.g. one that hard-exits on TTY checks).
         "auto_source_bashrc": True,
         "docker_image": "nikolaik/python-nodejs:python3.11-nodejs20",
@@ -447,6 +466,12 @@ DEFAULT_CONFIG = {
         "record_sessions": False,  # Auto-record browser sessions as WebM videos
         "allow_private_urls": False,  # Allow navigating to private/internal IPs (localhost, 192.168.x.x, etc.)
         "cdp_url": "",  # Optional persistent CDP endpoint for attaching to an existing Chromium/Chrome
+        # CDP supervisor — dialog + frame detection via a persistent WebSocket.
+        # Active only when a CDP-capable backend is attached (Browserbase or
+        # local Chrome via /browser connect). See
+        # website/docs/developer-guide/browser-supervisor.md.
+        "dialog_policy": "must_respond",  # must_respond | auto_dismiss | auto_accept
+        "dialog_timeout_s": 300,  # Safety auto-dismiss after N seconds under must_respond
         "camofox": {
             # When true, Hermes sends a stable profile-scoped userId to Camofox
             # so the server maps it to a persistent Firefox profile automatically.
@@ -467,7 +492,27 @@ DEFAULT_CONFIG = {
     # exceed this are rejected with guidance to use offset+limit.
     # 100K chars ≈ 25–35K tokens across typical tokenisers.
     "file_read_max_chars": 100_000,
-    
+
+    # Tool-output truncation thresholds. When terminal output or a
+    # single read_file page exceeds these limits, Hermes truncates the
+    # payload sent to the model (keeping head + tail for terminal,
+    # enforcing pagination for read_file). Tuning these trades context
+    # footprint against how much raw output the model can see in one
+    # shot. Ported from anomalyco/opencode PR #23770.
+    #
+    # - max_bytes:       terminal_tool output cap, in chars
+    #                    (default 50_000 ≈ 12-15K tokens).
+    # - max_lines:       read_file pagination cap — the maximum `limit`
+    #                    a single read_file call can request before
+    #                    being clamped (default 2000).
+    # - max_line_length: per-line cap applied when read_file emits a
+    #                    line-numbered view (default 2000 chars).
+    "tool_output": {
+        "max_bytes": 50_000,
+        "max_lines": 2000,
+        "max_line_length": 2000,
+    },
+
     "compression": {
         "enabled": True,
         "threshold": 0.50,            # compress when context usage exceeds this ratio
@@ -476,6 +521,12 @@ DEFAULT_CONFIG = {
 
     },
 
+    # Anthropic prompt caching (Claude via OpenRouter or native Anthropic API).
+    # cache_ttl must be "5m" or "1h" (Anthropic-supported tiers); other values are ignored.
+    "prompt_caching": {
+        "cache_ttl": "5m",
+    },
+
     # AWS Bedrock provider configuration.
     # Only used when model.provider is "bedrock".
     "bedrock": {
@@ -720,6 +771,10 @@ DEFAULT_CONFIG = {
         "inherit_mcp_toolsets": True,
         "max_iterations": 50,  # per-subagent iteration cap (each subagent gets its own budget,
                                # independent of the parent's max_iterations)
+        "child_timeout_seconds": 600,  # wall-clock timeout for each child agent (floor 30s,
+                                       # no ceiling). High-reasoning models on large tasks
+                                       # (e.g. gpt-5.5 xhigh, opus-4.6) need generous budgets;
+                                       # raise if children time out before producing output.
         "reasoning_effort": "",  # reasoning effort for subagents: "xhigh", "high", "medium",
                                  # "low", "minimal", "none" (empty = inherit parent's level)
         "max_concurrent_children": 3,  # max parallel children per batch; floor of 1 enforced, no ceiling
@@ -754,6 +809,17 @@ DEFAULT_CONFIG = {
         "inline_shell": False,
         # Timeout (seconds) for each !`cmd` snippet when inline_shell is on.
         "inline_shell_timeout": 10,
+        # Run the keyword/pattern security scanner on skills the agent
+        # writes via skill_manage (create/edit/patch).  Off by default
+        # because the agent can already execute the same code paths via
+        # terminal() with no gate, so the scan adds friction (blocks
+        # skills that mention risky keywords in prose) without meaningful
+        # security.  Turn on if you want the belt-and-suspenders — a
+        # dangerous verdict will then surface as a tool error to the
+        # agent, which can retry with the flagged content removed.
+        # External hub installs (trusted/community sources) are always
+        # scanned regardless of this setting.
+        "guard_agent_created": False,
     },
 
     # Honcho AI-native memory -- reads ~/.honcho/config.json as single source of truth.
@@ -1274,7 +1340,7 @@ OPTIONAL_ENV_VARS = {
         "advanced": True,
     },
     "XIAOMI_API_KEY": {
-        "description": "Xiaomi MiMo API key for MiMo models (mimo-v2-pro, mimo-v2-omni, mimo-v2-flash)",
+        "description": "Xiaomi MiMo API key for MiMo models (mimo-v2.5-pro, mimo-v2.5, mimo-v2-pro, mimo-v2-omni, mimo-v2-flash)",
         "prompt": "Xiaomi MiMo API Key",
         "url": "https://platform.xiaomimimo.com",
         "password": True,
@@ -2055,6 +2121,14 @@ def _normalize_custom_provider_entry(
     models = entry.get("models")
     if isinstance(models, dict) and models:
         normalized["models"] = models
+    elif isinstance(models, list) and models:
+        # Hand-edited configs (and older Hermes versions) write ``models`` as
+        # a plain list of model ids. Preserve them by converting to the dict
+        # shape downstream code expects; otherwise normalize silently drops
+        # the list and /model shows the provider with (0) models.
+        normalized["models"] = {
+            str(m): {} for m in models if isinstance(m, str) and m.strip()
+        }
 
     context_length = entry.get("context_length")
     if isinstance(context_length, int) and context_length > 0:

hermes_cli/copilot_auth.py

@@ -275,6 +275,99 @@ def copilot_device_code_login(
     return None
 
 
+# ─── Copilot Token Exchange ────────────────────────────────────────────────
+
+# Module-level cache for exchanged Copilot API tokens.
+# Maps raw_token_fingerprint -> (api_token, expires_at_epoch).
+_jwt_cache: dict[str, tuple[str, float]] = {}
+_JWT_REFRESH_MARGIN_SECONDS = 120  # refresh 2 min before expiry
+
+# Token exchange endpoint and headers (matching VS Code / Copilot CLI)
+_TOKEN_EXCHANGE_URL = "https://api.github.com/copilot_internal/v2/token"
+_EDITOR_VERSION = "vscode/1.104.1"
+_EXCHANGE_USER_AGENT = "GitHubCopilotChat/0.26.7"
+
+
+def _token_fingerprint(raw_token: str) -> str:
+    """Short fingerprint of a raw token for cache keying (avoids storing full token)."""
+    import hashlib
+    return hashlib.sha256(raw_token.encode()).hexdigest()[:16]
+
+
+def exchange_copilot_token(raw_token: str, *, timeout: float = 10.0) -> tuple[str, float]:
+    """Exchange a raw GitHub token for a short-lived Copilot API token.
+
+    Calls ``GET https://api.github.com/copilot_internal/v2/token`` with
+    the raw GitHub token and returns ``(api_token, expires_at)``.
+
+    The returned token is a semicolon-separated string (not a standard JWT)
+    used as ``Authorization: Bearer <token>`` for Copilot API requests.
+
+    Results are cached in-process and reused until close to expiry.
+    Raises ``ValueError`` on failure.
+    """
+    import urllib.request
+
+    fp = _token_fingerprint(raw_token)
+
+    # Check cache first
+    cached = _jwt_cache.get(fp)
+    if cached:
+        api_token, expires_at = cached
+        if time.time() < expires_at - _JWT_REFRESH_MARGIN_SECONDS:
+            return api_token, expires_at
+
+    req = urllib.request.Request(
+        _TOKEN_EXCHANGE_URL,
+        method="GET",
+        headers={
+            "Authorization": f"token {raw_token}",
+            "User-Agent": _EXCHANGE_USER_AGENT,
+            "Accept": "application/json",
+            "Editor-Version": _EDITOR_VERSION,
+        },
+    )
+
+    try:
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            data = json.loads(resp.read().decode())
+    except Exception as exc:
+        raise ValueError(f"Copilot token exchange failed: {exc}") from exc
+
+    api_token = data.get("token", "")
+    expires_at = data.get("expires_at", 0)
+    if not api_token:
+        raise ValueError("Copilot token exchange returned empty token")
+
+    # Convert expires_at to float if needed
+    expires_at = float(expires_at) if expires_at else time.time() + 1800
+
+    _jwt_cache[fp] = (api_token, expires_at)
+    logger.debug(
+        "Copilot token exchanged, expires_at=%s",
+        expires_at,
+    )
+    return api_token, expires_at
+
+
+def get_copilot_api_token(raw_token: str) -> str:
+    """Exchange a raw GitHub token for a Copilot API token, with fallback.
+
+    Convenience wrapper: returns the exchanged token on success, or the
+    raw token unchanged if the exchange fails (e.g. network error, unsupported
+    account type). This preserves existing behaviour for accounts that don't
+    need exchange while enabling access to internal-only models for those that do.
+    """
+    if not raw_token:
+        return raw_token
+    try:
+        api_token, _ = exchange_copilot_token(raw_token)
+        return api_token
+    except Exception as exc:
+        logger.debug("Copilot token exchange failed, using raw token: %s", exc)
+        return raw_token
+
+
 # ─── Copilot API Headers ───────────────────────────────────────────────────
 
 def copilot_request_headers(

hermes_cli/cron.py

@@ -93,6 +93,9 @@ def cron_list(show_all: bool = False):
         script = job.get("script")
         if script:
             print(f"    Script:    {script}")
+        workdir = job.get("workdir")
+        if workdir:
+            print(f"    Workdir:   {workdir}")
 
         # Execution history
         last_status = job.get("last_status")
@@ -168,6 +171,7 @@ def cron_create(args):
         skill=getattr(args, "skill", None),
         skills=_normalize_skills(getattr(args, "skill", None), getattr(args, "skills", None)),
         script=getattr(args, "script", None),
+        workdir=getattr(args, "workdir", None),
     )
     if not result.get("success"):
         print(color(f"Failed to create job: {result.get('error', 'unknown error')}", Colors.RED))
@@ -180,6 +184,8 @@ def cron_create(args):
     job_data = result.get("job", {})
     if job_data.get("script"):
         print(f"  Script: {job_data['script']}")
+    if job_data.get("workdir"):
+        print(f"  Workdir: {job_data['workdir']}")
     print(f"  Next run: {result['next_run_at']}")
     return 0
 
@@ -218,6 +224,7 @@ def cron_edit(args):
         repeat=getattr(args, "repeat", None),
         skills=final_skills,
         script=getattr(args, "script", None),
+        workdir=getattr(args, "workdir", None),
     )
     if not result.get("success"):
         print(color(f"Failed to update job: {result.get('error', 'unknown error')}", Colors.RED))
@@ -233,6 +240,8 @@ def cron_edit(args):
         print("  Skills: none")
     if updated.get("script"):
         print(f"  Script: {updated['script']}")
+    if updated.get("workdir"):
+        print(f"  Workdir: {updated['workdir']}")
     return 0
 
 

hermes_cli/doctor.py

@@ -29,6 +29,7 @@ if _env_path.exists():
 load_dotenv(PROJECT_ROOT / ".env", override=False, encoding="utf-8")
 
 from hermes_cli.colors import Colors, color
+from hermes_cli.models import _HERMES_USER_AGENT
 from hermes_constants import OPENROUTER_MODELS_URL
 from utils import base_url_host_matches
 
@@ -295,16 +296,33 @@ def run_doctor(args):
             except Exception:
                 pass
             try:
-                from hermes_cli.auth import resolve_provider as _resolve_provider
+                from hermes_cli.config import get_compatible_custom_providers as _compatible_custom_providers
+                from hermes_cli.providers import resolve_provider_full as _resolve_provider_full
             except Exception:
-                _resolve_provider = None
+                _compatible_custom_providers = None
+                _resolve_provider_full = None
+
+            custom_providers = []
+            if _compatible_custom_providers is not None:
+                try:
+                    custom_providers = _compatible_custom_providers(cfg)
+                except Exception:
+                    custom_providers = []
+
+            user_providers = cfg.get("providers")
+            if isinstance(user_providers, dict):
+                known_providers.update(str(name).strip().lower() for name in user_providers if str(name).strip())
+            for entry in custom_providers:
+                if not isinstance(entry, dict):
+                    continue
+                name = str(entry.get("name") or "").strip()
+                if name:
+                    known_providers.add("custom:" + name.lower().replace(" ", "-"))
 
             canonical_provider = provider
-            if provider and _resolve_provider is not None and provider != "auto":
-                try:
-                    canonical_provider = _resolve_provider(provider)
-                except Exception:
-                    canonical_provider = None
+            if provider and _resolve_provider_full is not None and provider != "auto":
+                provider_def = _resolve_provider_full(provider, user_providers, custom_providers)
+                canonical_provider = provider_def.id if provider_def is not None else None
 
             if provider and provider != "auto":
                 if canonical_provider is None or (known_providers and canonical_provider not in known_providers):
@@ -957,7 +975,10 @@ def run_doctor(args):
                 if base_url_host_matches(_base, "api.kimi.com") and _base.rstrip("/").endswith("/coding"):
                     _base = _base.rstrip("/") + "/v1"
                 _url = (_base.rstrip("/") + "/models") if _base else _default_url
-                _headers = {"Authorization": f"Bearer {_key}"}
+                _headers = {
+                    "Authorization": f"Bearer {_key}",
+                    "User-Agent": _HERMES_USER_AGENT,
+                }
                 if base_url_host_matches(_base, "api.kimi.com"):
                     _headers["User-Agent"] = "claude-code/0.1.0"
                 _resp = httpx.get(

hermes_cli/dump.py

@@ -267,6 +267,8 @@ def run_dump(args):
         ("ANTHROPIC_API_KEY", "anthropic"),
         ("ANTHROPIC_TOKEN", "anthropic_token"),
         ("NOUS_API_KEY", "nous"),
+        ("GOOGLE_API_KEY", "google/gemini"),
+        ("GEMINI_API_KEY", "gemini"),
         ("GLM_API_KEY", "glm/zai"),
         ("ZAI_API_KEY", "zai"),
         ("KIMI_API_KEY", "kimi"),

hermes_cli/gateway.py

@@ -175,6 +175,60 @@ def _request_gateway_self_restart(pid: int) -> bool:
     return True
 
 
+def _graceful_restart_via_sigusr1(pid: int, drain_timeout: float) -> bool:
+    """Send SIGUSR1 to a gateway PID and wait for it to exit gracefully.
+
+    SIGUSR1 is wired in gateway/run.py to ``request_restart(via_service=True)``
+    which drains in-flight agent runs (up to ``agent.restart_drain_timeout``
+    seconds), then exits with code 75.  Both systemd (``Restart=on-failure``
+    + ``RestartForceExitStatus=75``) and launchd (``KeepAlive.SuccessfulExit
+    = false``) relaunch the process after the graceful exit.
+
+    This is the drain-aware alternative to ``systemctl restart`` / ``SIGTERM``,
+    which SIGKILL in-flight agents after a short timeout.
+
+    Args:
+        pid: Gateway process PID (systemd MainPID, launchd PID, or bare
+            process PID).
+        drain_timeout: Seconds to wait for the process to exit after sending
+            SIGUSR1.  Should be slightly larger than the gateway's
+            ``agent.restart_drain_timeout`` to allow the drain loop to
+            finish cleanly.
+
+    Returns:
+        True if the PID was signalled and exited within the timeout.
+        False if SIGUSR1 couldn't be sent or the process didn't exit in
+        time (caller should fall back to a harder restart path).
+    """
+    if not hasattr(signal, "SIGUSR1"):
+        return False
+    if pid <= 0:
+        return False
+    try:
+        os.kill(pid, signal.SIGUSR1)
+    except ProcessLookupError:
+        # Already gone — nothing to drain.
+        return True
+    except (PermissionError, OSError):
+        return False
+
+    import time as _time
+
+    deadline = _time.monotonic() + max(drain_timeout, 1.0)
+    while _time.monotonic() < deadline:
+        try:
+            os.kill(pid, 0)  # signal 0 — probe liveness
+        except ProcessLookupError:
+            return True
+        except PermissionError:
+            # Process still exists but we can't signal it.  Treat as alive
+            # so the caller falls back.
+            pass
+        _time.sleep(0.5)
+    # Drain didn't finish in time.
+    return False
+
+
 def _append_unique_pid(pids: list[int], pid: int | None, exclude_pids: set[int]) -> None:
     if pid is None or pid <= 0:
         return
@@ -761,6 +815,21 @@ def get_systemd_unit_path(system: bool = False) -> Path:
     return Path.home() / ".config" / "systemd" / "user" / f"{name}.service"
 
 
+class UserSystemdUnavailableError(RuntimeError):
+    """Raised when ``systemctl --user`` cannot reach the user D-Bus session.
+
+    Typically hit on fresh RHEL/Debian SSH sessions where linger is disabled
+    and no user@.service is running, so ``/run/user/$UID/bus`` never exists.
+    Carries a user-facing remediation message in ``args[0]``.
+    """
+
+
+def _user_dbus_socket_path() -> Path:
+    """Return the expected per-user D-Bus socket path (regardless of existence)."""
+    xdg = os.environ.get("XDG_RUNTIME_DIR") or f"/run/user/{os.getuid()}"
+    return Path(xdg) / "bus"
+
+
 def _ensure_user_systemd_env() -> None:
     """Ensure DBUS_SESSION_BUS_ADDRESS and XDG_RUNTIME_DIR are set for systemctl --user.
 
@@ -783,6 +852,126 @@ def _ensure_user_systemd_env() -> None:
             os.environ["DBUS_SESSION_BUS_ADDRESS"] = f"unix:path={bus_path}"
 
 
+def _wait_for_user_dbus_socket(timeout: float = 3.0) -> bool:
+    """Poll for the user D-Bus socket to appear, up to ``timeout`` seconds.
+
+    Linger-enabled user@.service can take a second or two to spawn the socket
+    after ``loginctl enable-linger`` runs.  Returns True once the socket exists.
+    """
+    import time
+
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        if _user_dbus_socket_path().exists():
+            _ensure_user_systemd_env()
+            return True
+        time.sleep(0.2)
+    return _user_dbus_socket_path().exists()
+
+
+def _preflight_user_systemd(*, auto_enable_linger: bool = True) -> None:
+    """Ensure ``systemctl --user`` will reach the user D-Bus session bus.
+
+    No-op when the bus socket is already there (the common case on desktops
+    and linger-enabled servers).  On fresh SSH sessions where the socket is
+    missing:
+
+    * If linger is already enabled, wait briefly for user@.service to spawn
+      the socket.
+    * If linger is disabled and ``auto_enable_linger`` is True, try
+      ``loginctl enable-linger $USER`` (works as non-root when polkit permits
+      it, otherwise needs sudo).
+    * If the socket is still missing afterwards, raise
+      :class:`UserSystemdUnavailableError` with a precise remediation message.
+
+    Callers should treat the exception as a terminal condition for user-scope
+    systemd operations and surface the message to the user.
+    """
+    _ensure_user_systemd_env()
+    bus_path = _user_dbus_socket_path()
+    if bus_path.exists():
+        return
+
+    import getpass
+
+    username = getpass.getuser()
+    linger_enabled, linger_detail = get_systemd_linger_status()
+
+    if linger_enabled is True:
+        if _wait_for_user_dbus_socket(timeout=3.0):
+            return
+        # Linger is on but socket still missing — unusual; fall through to error.
+        _raise_user_systemd_unavailable(
+            username,
+            reason="User D-Bus socket is missing even though linger is enabled.",
+            fix_hint=(
+                f"  systemctl start user@{os.getuid()}.service\n"
+                "  (may require sudo; try again after the command succeeds)"
+            ),
+        )
+
+    if auto_enable_linger and shutil.which("loginctl"):
+        try:
+            result = subprocess.run(
+                ["loginctl", "enable-linger", username],
+                capture_output=True,
+                text=True,
+                check=False,
+                timeout=30,
+            )
+        except Exception as exc:
+            _raise_user_systemd_unavailable(
+                username,
+                reason=f"loginctl enable-linger failed ({exc}).",
+                fix_hint=f"  sudo loginctl enable-linger {username}",
+            )
+        else:
+            if result.returncode == 0:
+                if _wait_for_user_dbus_socket(timeout=5.0):
+                    print(f"✓ Enabled linger for {username} — user D-Bus now available")
+                    return
+                # enable-linger succeeded but the socket never appeared.
+                _raise_user_systemd_unavailable(
+                    username,
+                    reason="Linger was enabled, but the user D-Bus socket did not appear.",
+                    fix_hint=(
+                        "  Log out and log back in, then re-run the command.\n"
+                        f"  Or reboot and run: systemctl --user start {get_service_name()}"
+                    ),
+                )
+            detail = (result.stderr or result.stdout or f"exit {result.returncode}").strip()
+            _raise_user_systemd_unavailable(
+                username,
+                reason=f"loginctl enable-linger was denied: {detail}",
+                fix_hint=f"  sudo loginctl enable-linger {username}",
+            )
+
+    _raise_user_systemd_unavailable(
+        username,
+        reason=(
+            "User D-Bus session is not available "
+            f"({linger_detail or 'linger disabled'})."
+        ),
+        fix_hint=f"  sudo loginctl enable-linger {username}",
+    )
+
+
+def _raise_user_systemd_unavailable(username: str, *, reason: str, fix_hint: str) -> None:
+    """Build a user-facing error message and raise UserSystemdUnavailableError."""
+    msg = (
+        f"{reason}\n"
+        "  systemctl --user cannot reach the user D-Bus session in this shell.\n"
+        "\n"
+        "  To fix:\n"
+        f"{fix_hint}\n"
+        "\n"
+        "  Alternative: run the gateway in the foreground (stays up until\n"
+        "  you exit / close the terminal):\n"
+        "    hermes gateway run"
+    )
+    raise UserSystemdUnavailableError(msg)
+
+
 def _systemctl_cmd(system: bool = False) -> list[str]:
     if not system:
         _ensure_user_systemd_env()
@@ -1334,7 +1523,14 @@ def generate_systemd_unit(system: bool = False, run_as_user: str | None = None)
             path_entries.append(resolved_node_dir)
 
     common_bin_paths = ["/usr/local/sbin", "/usr/local/bin", "/usr/sbin", "/usr/bin", "/sbin", "/bin"]
-    restart_timeout = max(60, int(_get_restart_drain_timeout() or 0))
+    # systemd's TimeoutStopSec must exceed the gateway's drain_timeout so
+    # there's budget left for post-interrupt cleanup (tool subprocess kill,
+    # adapter disconnect, session DB close) before systemd escalates to
+    # SIGKILL on the cgroup — otherwise bash/sleep tool-call children left
+    # by a force-interrupted agent get reaped by systemd instead of us
+    # (#8202). 30s of headroom covers the worst case we've observed.
+    _drain_timeout = int(_get_restart_drain_timeout() or 0)
+    restart_timeout = max(60, _drain_timeout) + 30
 
     if system:
         username, group_name, home_dir = _system_service_identity(run_as_user)
@@ -1623,6 +1819,11 @@ def systemd_start(system: bool = False):
     system = _select_systemd_scope(system)
     if system:
         _require_root_for_system_service("start")
+    else:
+        # Fail fast with actionable guidance if the user D-Bus session is not
+        # reachable (common on fresh RHEL/Debian SSH sessions without linger).
+        # Raises UserSystemdUnavailableError with a remediation message.
+        _preflight_user_systemd()
     refresh_systemd_unit_if_needed(system=system)
     _run_systemctl(["start", get_service_name()], system=system, check=True, timeout=30)
     print(f"✓ {_service_scope_label(system).capitalize()} service started")
@@ -1642,6 +1843,8 @@ def systemd_restart(system: bool = False):
     system = _select_systemd_scope(system)
     if system:
         _require_root_for_system_service("restart")
+    else:
+        _preflight_user_systemd()
     refresh_systemd_unit_if_needed(system=system)
     from gateway.status import get_running_pid
 
@@ -3516,6 +3719,10 @@ def gateway_setup():
                     systemd_start()
                 elif is_macos():
                     launchd_start()
+            except UserSystemdUnavailableError as e:
+                print_error("  Failed to start — user systemd not reachable:")
+                for line in str(e).splitlines():
+                    print(f"  {line}")
             except subprocess.CalledProcessError as e:
                 print_error(f"  Failed to start: {e}")
     else:
@@ -3580,6 +3787,10 @@ def gateway_setup():
                     else:
                         stop_profile_gateway()
                         print_info("Start manually: hermes gateway")
+                except UserSystemdUnavailableError as e:
+                    print_error("  Restart failed — user systemd not reachable:")
+                    for line in str(e).splitlines():
+                        print(f"  {line}")
                 except subprocess.CalledProcessError as e:
                     print_error(f"  Restart failed: {e}")
         elif service_installed:
@@ -3589,6 +3800,10 @@ def gateway_setup():
                         systemd_start()
                     elif is_macos():
                         launchd_start()
+                except UserSystemdUnavailableError as e:
+                    print_error("  Start failed — user systemd not reachable:")
+                    for line in str(e).splitlines():
+                        print(f"  {line}")
                 except subprocess.CalledProcessError as e:
                     print_error(f"  Start failed: {e}")
         else:
@@ -3612,6 +3827,10 @@ def gateway_setup():
                                     systemd_start(system=installed_scope == "system")
                                 else:
                                     launchd_start()
+                            except UserSystemdUnavailableError as e:
+                                print_error("  Start failed — user systemd not reachable:")
+                                for line in str(e).splitlines():
+                                    print(f"  {line}")
                             except subprocess.CalledProcessError as e:
                                 print_error(f"  Start failed: {e}")
                     except subprocess.CalledProcessError as e:
@@ -3649,6 +3868,18 @@ def gateway_setup():
 
 def gateway_command(args):
     """Handle gateway subcommands."""
+    try:
+        return _gateway_command_inner(args)
+    except UserSystemdUnavailableError as e:
+        # Clean, actionable message instead of a traceback when the user D-Bus
+        # session is unreachable (fresh SSH shell, no linger, container, etc.).
+        print_error("User systemd not reachable:")
+        for line in str(e).splitlines():
+            print(f"  {line}")
+        sys.exit(1)
+
+
+def _gateway_command_inner(args):
     subcmd = getattr(args, 'gateway_command', None)
     
     # Default to run if no subcommand

hermes_cli/main.py

@@ -166,6 +166,27 @@ from hermes_cli.env_loader import load_hermes_dotenv
 
 load_hermes_dotenv(project_env=PROJECT_ROOT / ".env")
 
+# Bridge security.redact_secrets from config.yaml → HERMES_REDACT_SECRETS env
+# var BEFORE hermes_logging imports agent.redact (which snapshots the flag at
+# module-import time). Without this, config.yaml's toggle is ignored because
+# the setup_logging() call below imports agent.redact, which reads the env var
+# exactly once. Env var in .env still wins — this is config.yaml fallback only.
+try:
+    if "HERMES_REDACT_SECRETS" not in os.environ:
+        import yaml as _yaml_early
+        _cfg_path = get_hermes_home() / "config.yaml"
+        if _cfg_path.exists():
+            with open(_cfg_path, encoding="utf-8") as _f:
+                _early_sec_cfg = (_yaml_early.safe_load(_f) or {}).get("security", {})
+            if isinstance(_early_sec_cfg, dict):
+                _early_redact = _early_sec_cfg.get("redact_secrets")
+                if _early_redact is not None:
+                    os.environ["HERMES_REDACT_SECRETS"] = str(_early_redact).lower()
+            del _early_sec_cfg
+        del _cfg_path
+except Exception:
+    pass  # best-effort — redaction stays at default (enabled) on config errors
+
 # Initialize centralized file logging early — all `hermes` subcommands
 # (chat, setup, gateway, config, etc.) write to agent.log + errors.log.
 try:
@@ -1131,6 +1152,20 @@ def cmd_chat(args):
     if getattr(args, "yolo", False):
         os.environ["HERMES_YOLO_MODE"] = "1"
 
+    # --ignore-user-config: make load_cli_config() / load_config() skip the
+    # user's ~/.hermes/config.yaml and return built-in defaults. Set BEFORE
+    # importing cli (which runs `CLI_CONFIG = load_cli_config()` at module
+    # import time). Credentials in .env are still loaded — this flag only
+    # ignores behavioral/config settings.
+    if getattr(args, "ignore_user_config", False):
+        os.environ["HERMES_IGNORE_USER_CONFIG"] = "1"
+
+    # --ignore-rules: skip auto-injection of AGENTS.md/SOUL.md/.cursorrules
+    # (rules), memory entries, and any preloaded skills coming from user config.
+    # Maps to AIAgent(skip_context_files=True, skip_memory=True).
+    if getattr(args, "ignore_rules", False):
+        os.environ["HERMES_IGNORE_RULES"] = "1"
+
     # --source: tag session source for filtering (e.g. 'tool' for third-party integrations)
     if getattr(args, "source", None):
         os.environ["HERMES_SESSION_SOURCE"] = args.source
@@ -1159,6 +1194,8 @@ def cmd_chat(args):
         "checkpoints": getattr(args, "checkpoints", False),
         "pass_session_id": getattr(args, "pass_session_id", False),
         "max_turns": getattr(args, "max_turns", None),
+        "ignore_rules": getattr(args, "ignore_rules", False),
+        "ignore_user_config": getattr(args, "ignore_user_config", False),
     }
     # Filter out None values
     kwargs = {k: v for k, v in kwargs.items() if v is not None}
@@ -1413,6 +1450,7 @@ def select_provider_and_model(args=None):
         load_config,
         get_env_value,
     )
+    from hermes_cli.providers import resolve_provider_full
 
     config = load_config()
     current_model = config.get("model")
@@ -1430,14 +1468,30 @@ def select_provider_and_model(args=None):
     effective_provider = (
         config_provider or os.getenv("HERMES_INFERENCE_PROVIDER") or "auto"
     )
-    try:
-        active = resolve_provider(effective_provider)
-    except AuthError as exc:
-        warning = format_auth_error(exc)
-        print(f"Warning: {warning} Falling back to auto provider detection.")
+    compatible_custom_providers = get_compatible_custom_providers(config)
+    active = None
+    if effective_provider != "auto":
+        active_def = resolve_provider_full(
+            effective_provider,
+            config.get("providers"),
+            compatible_custom_providers,
+        )
+        if active_def is not None:
+            active = active_def.id
+        else:
+            warning = (
+                f"Unknown provider '{effective_provider}'. Check 'hermes model' for "
+                "available providers, or run 'hermes doctor' to diagnose config "
+                "issues."
+            )
+            print(f"Warning: {warning} Falling back to auto provider detection.")
+    if active is None:
         try:
             active = resolve_provider("auto")
-        except AuthError:
+        except AuthError as exc:
+            if effective_provider == "auto":
+                warning = format_auth_error(exc)
+                print(f"Warning: {warning} Falling back to auto provider detection.")
             active = None  # no provider yet; default to first in list
 
     # Detect custom endpoint
@@ -2295,7 +2349,41 @@ def _model_flow_openai_codex(config, current_model=""):
     from hermes_cli.codex_models import get_codex_model_ids
 
     status = get_codex_auth_status()
-    if not status.get("logged_in"):
+    if status.get("logged_in"):
+        print("  OpenAI Codex credentials: ✓")
+        print()
+        print("    1. Use existing credentials")
+        print("    2. Reauthenticate (new OAuth login)")
+        print("    3. Cancel")
+        print()
+        try:
+            choice = input("  Choice [1/2/3]: ").strip()
+        except (KeyboardInterrupt, EOFError):
+            choice = "1"
+
+        if choice == "2":
+            print("Starting a fresh OpenAI Codex login...")
+            print()
+            try:
+                mock_args = argparse.Namespace()
+                _login_openai_codex(
+                    mock_args,
+                    PROVIDER_REGISTRY["openai-codex"],
+                    force_new_login=True,
+                )
+            except SystemExit:
+                print("Login cancelled or failed.")
+                return
+            except Exception as exc:
+                print(f"Login failed: {exc}")
+                return
+            status = get_codex_auth_status()
+            if not status.get("logged_in"):
+                print("Login failed.")
+                return
+        elif choice == "3":
+            return
+    else:
         print("Not logged into OpenAI Codex. Starting login...")
         print()
         try:
@@ -2812,11 +2900,16 @@ def _model_flow_named_custom(config, provider_info):
 
     name = provider_info["name"]
     base_url = provider_info["base_url"]
+    api_mode = provider_info.get("api_mode", "")
     api_key = provider_info.get("api_key", "")
     key_env = provider_info.get("key_env", "")
     saved_model = provider_info.get("model", "")
     provider_key = (provider_info.get("provider_key") or "").strip()
 
+    # Resolve key from env var if api_key not set directly
+    if not api_key and key_env:
+        api_key = os.environ.get(key_env, "")
+
     print(f"  Provider: {name}")
     print(f"  URL:      {base_url}")
     if saved_model:
@@ -2824,7 +2917,10 @@ def _model_flow_named_custom(config, provider_info):
     print()
 
     print("Fetching available models...")
-    models = fetch_api_models(api_key, base_url, timeout=8.0)
+    models = fetch_api_models(
+        api_key, base_url, timeout=8.0,
+        api_mode=api_mode or None,
+    )
 
     if models:
         default_idx = 0
@@ -3914,12 +4010,71 @@ def _model_flow_api_key_provider(config, provider_id, current_model=""):
                 print("Cancelled.")
                 return
             save_env_value(key_env, new_key)
+            existing_key = new_key
             print("API key saved.")
             print()
     else:
         print(f"  {pconfig.name} API key: {existing_key[:8]}... ✓")
         print()
 
+    # Gemini free-tier gate: free-tier daily quotas (<= 250 RPD for Flash)
+    # are exhausted in a handful of agent turns, so refuse to wire up the
+    # provider with a free-tier key. Probe is best-effort; network or auth
+    # errors fall through without blocking.
+    if provider_id == "gemini" and existing_key:
+        try:
+            from agent.gemini_native_adapter import probe_gemini_tier
+        except Exception:
+            probe_gemini_tier = None
+        if probe_gemini_tier is not None:
+            print("  Checking Gemini API tier...")
+            probe_base = (
+                (get_env_value(base_url_env) if base_url_env else "")
+                or os.getenv(base_url_env or "", "")
+                or pconfig.inference_base_url
+            )
+            tier = probe_gemini_tier(existing_key, probe_base)
+            if tier == "free":
+                print()
+                print(
+                    "❌ This Google API key is on the free tier "
+                    "(<= 250 requests/day for gemini-2.5-flash)."
+                )
+                print(
+                    "   Hermes typically makes 3-10 API calls per user turn "
+                    "(tool iterations + auxiliary tasks),"
+                )
+                print(
+                    "   so the free tier is exhausted after a handful of "
+                    "messages and cannot sustain"
+                )
+                print("   an agent session.")
+                print()
+                print(
+                    "   To use Gemini with Hermes, enable billing on your "
+                    "Google Cloud project and regenerate"
+                )
+                print(
+                    "   the key in a billing-enabled project: "
+                    "https://aistudio.google.com/apikey"
+                )
+                print()
+                print(
+                    "   Alternatives with workable free usage: DeepSeek, "
+                    "OpenRouter (free models), Groq, Nous."
+                )
+                print()
+                print("Not saving Gemini as the default provider.")
+                return
+            if tier == "paid":
+                print("  Tier check: paid ✓")
+            else:
+                # "unknown" -- network issue, auth problem, unexpected response.
+                # Don't block; the runtime 429 handler will surface free-tier
+                # guidance if the key turns out to be free tier.
+                print("  Tier check: could not verify (proceeding anyway).")
+            print()
+
     # Optional base URL override
     current_base = ""
     if base_url_env:
@@ -3968,7 +4123,18 @@ def _model_flow_api_key_provider(config, provider_id, current_model=""):
             pass
 
         if mdev_models:
-            model_list = mdev_models
+            # Merge models.dev with curated list so newly added models
+            # (not yet in models.dev) still appear in the picker.
+            if curated:
+                seen = {m.lower() for m in mdev_models}
+                merged = list(mdev_models)
+                for m in curated:
+                    if m.lower() not in seen:
+                        merged.append(m)
+                        seen.add(m.lower())
+                model_list = merged
+            else:
+                model_list = mdev_models
             print(f"  Found {len(model_list)} model(s) from models.dev registry")
         elif curated and len(curated) >= 8:
             # Curated list is substantial — use it directly, skip live probe
@@ -4150,6 +4316,8 @@ def _model_flow_anthropic(config, current_model=""):
         from agent.anthropic_adapter import (
             read_claude_code_credentials,
             is_claude_code_token_valid,
+            _is_oauth_token,
+            _resolve_claude_code_token_from_credentials,
         )
 
         cc_creds = read_claude_code_credentials()
@@ -4158,7 +4326,14 @@ def _model_flow_anthropic(config, current_model=""):
     except Exception:
         pass
 
-    has_creds = bool(existing_key) or cc_available
+    # Stale-OAuth guard: if the only existing cred is an expired OAuth token
+    # (no valid cc_creds to fall back on), treat it as missing so the re-auth
+    # path is offered instead of silently accepting a broken token.
+    existing_is_stale_oauth = False
+    if existing_key and _is_oauth_token(existing_key) and not cc_available:
+        existing_is_stale_oauth = True
+
+    has_creds = (bool(existing_key) and not existing_is_stale_oauth) or cc_available
     needs_auth = not has_creds
 
     if has_creds:
@@ -5837,12 +6012,15 @@ def _cmd_update_impl(args, gateway_mode: bool):
         # Write exit code *before* the gateway restart attempt.
         # When running as ``hermes update --gateway`` (spawned by the gateway's
         # /update command), this process lives inside the gateway's systemd
-        # cgroup.  ``systemctl restart hermes-gateway`` kills everything in the
-        # cgroup (KillMode=mixed → SIGKILL to remaining processes), including
-        # us and the wrapping bash shell.  The shell never reaches its
-        # ``printf $status > .update_exit_code`` epilogue, so the exit-code
-        # marker file is never created.  The new gateway's update watcher then
-        # polls for 30 minutes and sends a spurious timeout message.
+        # cgroup.  A graceful SIGUSR1 restart keeps the drain loop alive long
+        # enough for the exit-code marker to be written below, but the
+        # fallback ``systemctl restart`` path (see below) kills everything in
+        # the cgroup (KillMode=mixed → SIGKILL to remaining processes),
+        # including us and the wrapping bash shell.  The shell never reaches
+        # its ``printf $status > .update_exit_code`` epilogue, so the
+        # exit-code marker file would never be created.  The new gateway's
+        # update watcher would then poll for 30 minutes and send a spurious
+        # timeout message.
         #
         # Writing the marker here — after git pull + pip install succeed but
         # before we attempt the restart — ensures the new gateway sees it
@@ -5864,9 +6042,37 @@ def _cmd_update_impl(args, gateway_mode: bool):
                 _ensure_user_systemd_env,
                 find_gateway_pids,
                 _get_service_pids,
+                _graceful_restart_via_sigusr1,
             )
             import signal as _signal
 
+            # Drain budget for graceful SIGUSR1 restarts.  The gateway drains
+            # for up to ``agent.restart_drain_timeout`` (default 60s) before
+            # exiting with code 75; we wait slightly longer so the drain
+            # completes before we fall back to a hard restart.  On older
+            # systemd units without SIGUSR1 wiring this wait just times out
+            # and we fall back to ``systemctl restart`` (the old behaviour).
+            try:
+                from hermes_constants import (
+                    DEFAULT_GATEWAY_RESTART_DRAIN_TIMEOUT as _DEFAULT_DRAIN,
+                )
+            except Exception:
+                _DEFAULT_DRAIN = 60.0
+            _cfg_drain = None
+            try:
+                from hermes_cli.config import load_config
+                _cfg_agent = (load_config().get("agent") or {})
+                _cfg_drain = _cfg_agent.get("restart_drain_timeout")
+            except Exception:
+                pass
+            try:
+                _drain_budget = float(_cfg_drain) if _cfg_drain is not None else float(_DEFAULT_DRAIN)
+            except (TypeError, ValueError):
+                _drain_budget = float(_DEFAULT_DRAIN)
+            # Add a 15s margin so the drain loop + final exit finish before
+            # we escalate to ``systemctl restart`` / SIGTERM.
+            _drain_budget = max(_drain_budget, 30.0) + 15.0
+
             restarted_services = []
             killed_pids = set()
 
@@ -5913,59 +6119,114 @@ def _cmd_update_impl(args, gateway_mode: bool):
                                 text=True,
                                 timeout=5,
                             )
-                            if check.stdout.strip() == "active":
-                                restart = subprocess.run(
-                                    scope_cmd + ["restart", svc_name],
+                            if check.stdout.strip() != "active":
+                                continue
+
+                            # Prefer a graceful SIGUSR1 restart so in-flight
+                            # agent runs drain instead of being SIGKILLed.
+                            # The gateway's SIGUSR1 handler calls
+                            # request_restart(via_service=True) → drain →
+                            # exit(75); systemd's Restart=on-failure (and
+                            # RestartForceExitStatus=75) respawns the unit.
+                            _main_pid = 0
+                            try:
+                                _show = subprocess.run(
+                                    scope_cmd + [
+                                        "show", svc_name,
+                                        "--property=MainPID", "--value",
+                                    ],
+                                    capture_output=True, text=True, timeout=5,
+                                )
+                                _main_pid = int((_show.stdout or "").strip() or 0)
+                            except (ValueError, subprocess.TimeoutExpired, FileNotFoundError):
+                                _main_pid = 0
+
+                            _graceful_ok = False
+                            if _main_pid > 0:
+                                print(
+                                    f"  → {svc_name}: draining (up to {int(_drain_budget)}s)..."
+                                )
+                                _graceful_ok = _graceful_restart_via_sigusr1(
+                                    _main_pid, drain_timeout=_drain_budget,
+                                )
+
+                            if _graceful_ok:
+                                # Gateway exited 75; systemd should relaunch
+                                # via Restart=on-failure.  Verify the new
+                                # process came up.
+                                _time.sleep(3)
+                                verify = subprocess.run(
+                                    scope_cmd + ["is-active", svc_name],
+                                    capture_output=True, text=True, timeout=5,
+                                )
+                                if verify.stdout.strip() == "active":
+                                    restarted_services.append(svc_name)
+                                    continue
+                                # Process exited but wasn't respawned (older
+                                # unit without Restart=on-failure or
+                                # RestartForceExitStatus=75).  Fall through
+                                # to systemctl start/restart.
+                                print(
+                                    f"  ⚠ {svc_name} drained but didn't relaunch — forcing restart"
+                                )
+
+                            # Fallback: blunt systemctl restart.  This is
+                            # what the old code always did; we get here only
+                            # when the graceful path failed (unit missing
+                            # SIGUSR1 wiring, drain exceeded the budget,
+                            # restart-policy mismatch).
+                            restart = subprocess.run(
+                                scope_cmd + ["restart", svc_name],
+                                capture_output=True,
+                                text=True,
+                                timeout=15,
+                            )
+                            if restart.returncode == 0:
+                                # Verify the service actually survived the
+                                # restart.  systemctl restart returns 0 even
+                                # if the new process crashes immediately.
+                                _time.sleep(3)
+                                verify = subprocess.run(
+                                    scope_cmd + ["is-active", svc_name],
                                     capture_output=True,
                                     text=True,
-                                    timeout=15,
+                                    timeout=5,
                                 )
-                                if restart.returncode == 0:
-                                    # Verify the service actually survived the
-                                    # restart.  systemctl restart returns 0 even
-                                    # if the new process crashes immediately.
+                                if verify.stdout.strip() == "active":
+                                    restarted_services.append(svc_name)
+                                else:
+                                    # Retry once — transient startup failures
+                                    # (stale module cache, import race) often
+                                    # resolve on the second attempt.
+                                    print(
+                                        f"  ⚠ {svc_name} died after restart, retrying..."
+                                    )
+                                    retry = subprocess.run(
+                                        scope_cmd + ["restart", svc_name],
+                                        capture_output=True,
+                                        text=True,
+                                        timeout=15,
+                                    )
                                     _time.sleep(3)
-                                    verify = subprocess.run(
+                                    verify2 = subprocess.run(
                                         scope_cmd + ["is-active", svc_name],
                                         capture_output=True,
                                         text=True,
                                         timeout=5,
                                     )
-                                    if verify.stdout.strip() == "active":
+                                    if verify2.stdout.strip() == "active":
                                         restarted_services.append(svc_name)
+                                        print(f"  ✓ {svc_name} recovered on retry")
                                     else:
-                                        # Retry once — transient startup failures
-                                        # (stale module cache, import race) often
-                                        # resolve on the second attempt.
                                         print(
-                                            f"  ⚠ {svc_name} died after restart, retrying..."
+                                            f"  ✗ {svc_name} failed to stay running after restart.\n"
+                                            f"    Check logs: journalctl --user -u {svc_name} --since '2 min ago'\n"
+                                            f"    Restart manually: systemctl {'--user ' if scope == 'user' else ''}restart {svc_name}"
                                         )
-                                        retry = subprocess.run(
-                                            scope_cmd + ["restart", svc_name],
-                                            capture_output=True,
-                                            text=True,
-                                            timeout=15,
-                                        )
-                                        _time.sleep(3)
-                                        verify2 = subprocess.run(
-                                            scope_cmd + ["is-active", svc_name],
-                                            capture_output=True,
-                                            text=True,
-                                            timeout=5,
-                                        )
-                                        if verify2.stdout.strip() == "active":
-                                            restarted_services.append(svc_name)
-                                            print(f"  ✓ {svc_name} recovered on retry")
-                                        else:
-                                            print(
-                                                f"  ✗ {svc_name} failed to stay running after restart.\n"
-                                                f"    Check logs: journalctl --user -u {svc_name} --since '2 min ago'\n"
-                                                f"    Restart manually: systemctl {'--user ' if scope == 'user' else ''}restart {svc_name}"
-                                            )
-                                else:
-                                    print(
-                                        f"  ⚠ Failed to restart {svc_name}: {restart.stderr.strip()}"
-                                    )
+                            else:
+                                print(
+                                    f"  ⚠ Failed to restart {svc_name}: {restart.stderr.strip()}"
+                                )
                     except (FileNotFoundError, subprocess.TimeoutExpired):
                         pass
 
@@ -6606,6 +6867,18 @@ For more help on a command:
         default=False,
         help="Include the session ID in the agent's system prompt",
     )
+    parser.add_argument(
+        "--ignore-user-config",
+        action="store_true",
+        default=False,
+        help="Ignore ~/.hermes/config.yaml and fall back to built-in defaults (credentials in .env are still loaded)",
+    )
+    parser.add_argument(
+        "--ignore-rules",
+        action="store_true",
+        default=False,
+        help="Skip auto-injection of AGENTS.md, SOUL.md, .cursorrules, memory, and preloaded skills",
+    )
     parser.add_argument(
         "--tui",
         action="store_true",
@@ -6745,6 +7018,18 @@ For more help on a command:
         default=argparse.SUPPRESS,
         help="Include the session ID in the agent's system prompt",
     )
+    chat_parser.add_argument(
+        "--ignore-user-config",
+        action="store_true",
+        default=argparse.SUPPRESS,
+        help="Ignore ~/.hermes/config.yaml and fall back to built-in defaults (credentials in .env are still loaded). Useful for isolated CI runs, reproduction, and third-party integrations.",
+    )
+    chat_parser.add_argument(
+        "--ignore-rules",
+        action="store_true",
+        default=argparse.SUPPRESS,
+        help="Skip auto-injection of AGENTS.md, SOUL.md, .cursorrules, memory, and preloaded skills. Combine with --ignore-user-config for a fully isolated run.",
+    )
     chat_parser.add_argument(
         "--source",
         default=None,
@@ -7048,7 +7333,7 @@ For more help on a command:
     )
     logout_parser.add_argument(
         "--provider",
-        choices=["nous", "openai-codex"],
+        choices=["nous", "openai-codex", "spotify"],
         default=None,
         help="Provider to log out from (default: active provider)",
     )
@@ -7105,6 +7390,17 @@ For more help on a command:
         "reset", help="Clear exhaustion status for all credentials for a provider"
     )
     auth_reset.add_argument("provider", help="Provider id")
+    auth_status = auth_subparsers.add_parser("status", help="Show auth status for a provider")
+    auth_status.add_argument("provider", help="Provider id")
+    auth_logout = auth_subparsers.add_parser("logout", help="Log out a provider and clear stored auth state")
+    auth_logout.add_argument("provider", help="Provider id")
+    auth_spotify = auth_subparsers.add_parser("spotify", help="Authenticate Hermes with Spotify via PKCE")
+    auth_spotify.add_argument("spotify_action", nargs="?", choices=["login", "status", "logout"], default="login")
+    auth_spotify.add_argument("--client-id", help="Spotify app client_id (or set HERMES_SPOTIFY_CLIENT_ID)")
+    auth_spotify.add_argument("--redirect-uri", help="Allow-listed localhost redirect URI for your Spotify app")
+    auth_spotify.add_argument("--scope", help="Override requested Spotify scopes")
+    auth_spotify.add_argument("--no-browser", action="store_true", help="Do not attempt to open the browser automatically")
+    auth_spotify.add_argument("--timeout", type=float, help="Callback/token exchange timeout in seconds")
     auth_parser.set_defaults(func=cmd_auth)
 
     # =========================================================================
@@ -7161,6 +7457,10 @@ For more help on a command:
         "--script",
         help="Path to a Python script whose stdout is injected into the prompt each run",
     )
+    cron_create.add_argument(
+        "--workdir",
+        help="Absolute path for the job to run from. Injects AGENTS.md / CLAUDE.md / .cursorrules from that directory and uses it as the cwd for terminal/file/code_exec tools. Omit to preserve old behaviour (no project context files).",
+    )
 
     # cron edit
     cron_edit = cron_subparsers.add_parser(
@@ -7199,6 +7499,10 @@ For more help on a command:
         "--script",
         help="Path to a Python script whose stdout is injected into the prompt each run. Pass empty string to clear.",
     )
+    cron_edit.add_argument(
+        "--workdir",
+        help="Absolute path for the job to run from (injects AGENTS.md etc. and sets terminal cwd). Pass empty string to clear.",
+    )
 
     # lifecycle actions
     cron_pause = cron_subparsers.add_parser("pause", help="Pause a scheduled job")

hermes_cli/model_normalize.py

@@ -12,8 +12,12 @@ Different LLM providers expect model identifiers in different formats:
   model IDs, but Claude still uses hyphenated native names like
   ``claude-sonnet-4-6``.
 - **OpenCode Go** preserves dots in model names: ``minimax-m2.7``.
-- **DeepSeek** only accepts two model identifiers:
-  ``deepseek-chat`` and ``deepseek-reasoner``.
+- **DeepSeek** accepts ``deepseek-chat`` (V3), ``deepseek-reasoner``
+  (R1-family), and the first-class V-series IDs (``deepseek-v4-pro``,
+  ``deepseek-v4-flash``, and any future ``deepseek-v<N>-*``).  Older
+  Hermes revisions folded every non-reasoner input into
+  ``deepseek-chat``, which on aggregators routes to V3 — so a user
+  picking V4 Pro was silently downgraded.
 - **Custom** and remaining providers pass the name through as-is.
 
 This module centralises that translation so callers can simply write::
@@ -25,6 +29,7 @@ Inspired by Clawdbot's ``normalizeAnthropicModelId`` pattern.
 
 from __future__ import annotations
 
+import re
 from typing import Optional
 
 # ---------------------------------------------------------------------------
@@ -100,6 +105,15 @@ _MATCHING_PREFIX_STRIP_PROVIDERS: frozenset[str] = frozenset({
     "custom",
 })
 
+# Providers whose APIs require lowercase model IDs.  Xiaomi's
+# ``api.xiaomimimo.com`` rejects mixed-case names like ``MiMo-V2.5-Pro``
+# that users might copy from marketing docs — it only accepts
+# ``mimo-v2.5-pro``.  After stripping a matching provider prefix, these
+# providers also get ``.lower()`` applied.
+_LOWERCASE_MODEL_PROVIDERS: frozenset[str] = frozenset({
+    "xiaomi",
+})
+
 # ---------------------------------------------------------------------------
 # DeepSeek special handling
 # ---------------------------------------------------------------------------
@@ -115,17 +129,30 @@ _DEEPSEEK_REASONER_KEYWORDS: frozenset[str] = frozenset({
 })
 
 _DEEPSEEK_CANONICAL_MODELS: frozenset[str] = frozenset({
-    "deepseek-chat",
-    "deepseek-reasoner",
+    "deepseek-chat",       # V3 on DeepSeek direct and most aggregators
+    "deepseek-reasoner",   # R1-family reasoning model
+    "deepseek-v4-pro",     # V4 Pro — first-class model ID
+    "deepseek-v4-flash",   # V4 Flash — first-class model ID
 })
 
+# First-class V-series IDs (``deepseek-v4-pro``, ``deepseek-v4-flash``,
+# future ``deepseek-v5-*``, dated variants like ``deepseek-v4-flash-20260423``).
+# Verified empirically 2026-04-24: DeepSeek's Chat Completions API returns
+# ``provider: DeepSeek`` / ``model: deepseek-v4-flash-20260423`` when called
+# with ``model=deepseek/deepseek-v4-flash``, so these names are not aliases
+# of ``deepseek-chat`` and must not be folded into it.
+_DEEPSEEK_V_SERIES_RE = re.compile(r"^deepseek-v\d+([-.].+)?$")
+
 
 def _normalize_for_deepseek(model_name: str) -> str:
-    """Map any model input to one of DeepSeek's two accepted identifiers.
+    """Map a model input to a DeepSeek-accepted identifier.
 
     Rules:
-    - Already ``deepseek-chat`` or ``deepseek-reasoner`` -> pass through.
-    - Contains any reasoner keyword (r1, think, reasoning, cot, reasoner)
+    - Already a known canonical (``deepseek-chat``/``deepseek-reasoner``/
+      ``deepseek-v4-pro``/``deepseek-v4-flash``) -> pass through.
+    - Matches the V-series pattern ``deepseek-v<digit>...`` -> pass through
+      (covers future ``deepseek-v5-*`` and dated variants without a release).
+    - Contains a reasoner keyword (r1, think, reasoning, cot, reasoner)
       -> ``deepseek-reasoner``.
     - Everything else -> ``deepseek-chat``.
 
@@ -133,13 +160,17 @@ def _normalize_for_deepseek(model_name: str) -> str:
         model_name: The bare model name (vendor prefix already stripped).
 
     Returns:
-        One of ``"deepseek-chat"`` or ``"deepseek-reasoner"``.
+        A DeepSeek-accepted model identifier.
     """
     bare = _strip_vendor_prefix(model_name).lower()
 
     if bare in _DEEPSEEK_CANONICAL_MODELS:
         return bare
 
+    # V-series first-class IDs (v4-pro, v4-flash, future v5-*, dated variants)
+    if _DEEPSEEK_V_SERIES_RE.match(bare):
+        return bare
+
     # Check for reasoner-like keywords anywhere in the name
     for keyword in _DEEPSEEK_REASONER_KEYWORDS:
         if keyword in bare:
@@ -347,6 +378,9 @@ def normalize_model_for_provider(model_input: str, target_provider: str) -> str:
 
         >>> normalize_model_for_provider("claude-sonnet-4.6", "zai")
         'claude-sonnet-4.6'
+
+        >>> normalize_model_for_provider("MiMo-V2.5-Pro", "xiaomi")
+        'mimo-v2.5-pro'
     """
     name = (model_input or "").strip()
     if not name:
@@ -410,7 +444,12 @@ def normalize_model_for_provider(model_input: str, target_provider: str) -> str:
 
     # --- Direct providers: repair matching provider prefixes only ---
     if provider in _MATCHING_PREFIX_STRIP_PROVIDERS:
-        return _strip_matching_provider_prefix(name, provider)
+        result = _strip_matching_provider_prefix(name, provider)
+        # Some providers require lowercase model IDs (e.g. Xiaomi's API
+        # rejects "MiMo-V2.5-Pro" but accepts "mimo-v2.5-pro").
+        if provider in _LOWERCASE_MODEL_PROVIDERS:
+            result = result.lower()
+        return result
 
     # --- Authoritative native providers: preserve user-facing slugs as-is ---
     if provider in _AUTHORITATIVE_NATIVE_PROVIDERS:

hermes_cli/model_switch.py

@@ -304,6 +304,113 @@ def parse_model_flags(raw_args: str) -> tuple[str, str, bool]:
 # Alias resolution
 # ---------------------------------------------------------------------------
 
+def _model_sort_key(model_id: str, prefix: str) -> tuple:
+    """Sort key for model version preference.
+
+    Extracts version numbers after the family prefix and returns a sort key
+    that prefers higher versions.  Suffix tokens (``pro``, ``omni``, etc.)
+    are used as tiebreakers, with common quality indicators ranked.
+
+    Examples (with prefix ``"mimo"``)::
+
+        mimo-v2.5-pro   → (-2.5, 0, 'pro')     # highest version wins
+        mimo-v2.5       → (-2.5, 1, '')          # no suffix = lower than pro
+        mimo-v2-pro     → (-2.0, 0, 'pro')
+        mimo-v2-omni    → (-2.0, 1, 'omni')
+        mimo-v2-flash   → (-2.0, 1, 'flash')
+    """
+    # Strip the prefix (and optional "/" separator for aggregator slugs)
+    rest = model_id[len(prefix):]
+    if rest.startswith("/"):
+        rest = rest[1:]
+    rest = rest.lstrip("-").strip()
+
+    # Parse version and suffix from the remainder.
+    # "v2.5-pro" → version [2.5], suffix "pro"
+    # "-omni"    → version [],    suffix "omni"
+    # State machine: start → in_version → between → in_suffix
+    nums: list[float] = []
+    suffix_buf = ""
+    state = "start"
+    num_buf = ""
+
+    for ch in rest:
+        if state == "start":
+            if ch in "vV":
+                state = "in_version"
+            elif ch.isdigit():
+                state = "in_version"
+                num_buf += ch
+            elif ch in "-_.":
+                pass  # skip separators before any content
+            else:
+                state = "in_suffix"
+                suffix_buf += ch
+        elif state == "in_version":
+            if ch.isdigit():
+                num_buf += ch
+            elif ch == ".":
+                if "." in num_buf:
+                    # Second dot — flush current number, start new component
+                    try:
+                        nums.append(float(num_buf.rstrip(".")))
+                    except ValueError:
+                        pass
+                    num_buf = ""
+                else:
+                    num_buf += ch
+            elif ch in "-_.":
+                if num_buf:
+                    try:
+                        nums.append(float(num_buf.rstrip(".")))
+                    except ValueError:
+                        pass
+                    num_buf = ""
+                state = "between"
+            else:
+                if num_buf:
+                    try:
+                        nums.append(float(num_buf.rstrip(".")))
+                    except ValueError:
+                        pass
+                    num_buf = ""
+                state = "in_suffix"
+                suffix_buf += ch
+        elif state == "between":
+            if ch.isdigit():
+                state = "in_version"
+                num_buf = ch
+            elif ch in "vV":
+                state = "in_version"
+            elif ch in "-_.":
+                pass
+            else:
+                state = "in_suffix"
+                suffix_buf += ch
+        elif state == "in_suffix":
+            suffix_buf += ch
+
+    # Flush remaining buffer (strip trailing dots — "5.4." → "5.4")
+    if num_buf and state == "in_version":
+        try:
+            nums.append(float(num_buf.rstrip(".")))
+        except ValueError:
+            pass
+
+    suffix = suffix_buf.lower().strip("-_.")
+    suffix = suffix.strip()
+
+    # Negate versions so higher → sorts first
+    version_key = tuple(-n for n in nums)
+
+    # Suffix quality ranking: pro/max > (no suffix) > omni/flash/mini/lite
+    # Lower number = preferred
+    _SUFFIX_RANK = {"pro": 0, "max": 0, "plus": 0, "turbo": 0}
+    suffix_rank = _SUFFIX_RANK.get(suffix, 1)
+
+    return version_key + (suffix_rank, suffix)
+
+
 def resolve_alias(
     raw_input: str,
     current_provider: str,
@@ -311,9 +418,9 @@ def resolve_alias(
     """Resolve a short alias against the current provider's catalog.
 
     Looks up *raw_input* in :data:`MODEL_ALIASES`, then searches the
-    current provider's models.dev catalog for the first model whose ID
-    starts with ``vendor/family`` (or just ``family`` for non-aggregator
-    providers).
+    current provider's models.dev catalog for the model whose ID starts
+    with ``vendor/family`` (or just ``family`` for non-aggregator
+    providers) and has the **highest version**.
 
     Returns:
         ``(provider, resolved_model_id, alias_name)`` if a match is
@@ -341,28 +448,44 @@ def resolve_alias(
 
     vendor, family = identity
 
-    # Search the provider's catalog from models.dev
+    # Build catalog from models.dev, then merge in static _PROVIDER_MODELS
+    # entries that models.dev may be missing (e.g. newly added models not
+    # yet synced to the registry).
     catalog = list_provider_models(current_provider)
-    if not catalog:
-        return None
+    try:
+        from hermes_cli.models import _PROVIDER_MODELS
+        static = _PROVIDER_MODELS.get(current_provider, [])
+        if static:
+            seen = {m.lower() for m in catalog}
+            for m in static:
+                if m.lower() not in seen:
+                    catalog.append(m)
+    except Exception:
+        pass
 
     # For aggregators, models are vendor/model-name format
     aggregator = is_aggregator(current_provider)
 
-    for model_id in catalog:
-        mid_lower = model_id.lower()
-        if aggregator:
-            # Match vendor/family prefix -- e.g. "anthropic/claude-sonnet"
-            prefix = f"{vendor}/{family}".lower()
-            if mid_lower.startswith(prefix):
-                return (current_provider, model_id, key)
-        else:
-            # Non-aggregator: bare names -- e.g. "claude-sonnet-4-6"
-            family_lower = family.lower()
-            if mid_lower.startswith(family_lower):
-                return (current_provider, model_id, key)
+    if aggregator:
+        prefix = f"{vendor}/{family}".lower()
+        matches = [
+            mid for mid in catalog
+            if mid.lower().startswith(prefix)
+        ]
+    else:
+        family_lower = family.lower()
+        matches = [
+            mid for mid in catalog
+            if mid.lower().startswith(family_lower)
+        ]
 
-    return None
+    if not matches:
+        return None
+
+    # Sort by version descending — prefer the latest/highest version
+    prefix_for_sort = f"{vendor}/{family}" if aggregator else family
+    matches.sort(key=lambda m: _model_sort_key(m, prefix_for_sort))
+    return (current_provider, matches[0], key)
 
 
 def get_authenticated_provider_slugs(
@@ -648,7 +771,10 @@ def switch_model(
 
     if provider_changed or explicit_provider:
         try:
-            runtime = resolve_runtime_provider(requested=target_provider)
+            runtime = resolve_runtime_provider(
+                requested=target_provider,
+                target_model=new_model,
+            )
             api_key = runtime.get("api_key", "")
             base_url = runtime.get("base_url", "")
             api_mode = runtime.get("api_mode", "")
@@ -665,7 +791,10 @@ def switch_model(
             )
     else:
         try:
-            runtime = resolve_runtime_provider(requested=current_provider)
+            runtime = resolve_runtime_provider(
+                requested=current_provider,
+                target_model=new_model,
+            )
             api_key = runtime.get("api_key", "")
             base_url = runtime.get("base_url", "")
             api_mode = runtime.get("api_mode", "")
@@ -692,6 +821,7 @@ def switch_model(
             target_provider,
             api_key=api_key,
             base_url=base_url,
+            api_mode=api_mode or None,
         )
     except Exception as e:
         validation = {
@@ -782,6 +912,7 @@ def switch_model(
 
 def list_authenticated_providers(
     current_provider: str = "",
+    current_base_url: str = "",
     user_providers: dict = None,
     custom_providers: list | None = None,
     max_models: int = 8,
@@ -812,7 +943,7 @@ def list_authenticated_providers(
     from hermes_cli.auth import PROVIDER_REGISTRY
     from hermes_cli.models import (
         OPENROUTER_MODELS, _PROVIDER_MODELS,
-        _MODELS_DEV_PREFERRED, _merge_with_models_dev,
+        _MODELS_DEV_PREFERRED, _merge_with_models_dev, provider_model_ids,
     )
 
     results: List[dict] = []
@@ -847,6 +978,10 @@ def list_authenticated_providers(
         # source of truth.  models.dev can have wrong mappings (e.g.
         # minimax-cn → MINIMAX_API_KEY instead of MINIMAX_CN_API_KEY).
         pconfig = PROVIDER_REGISTRY.get(hermes_id)
+        # Skip non-API-key auth providers here — they are handled in
+        # section 2 (HERMES_OVERLAYS) with proper auth store checking.
+        if pconfig and pconfig.auth_type != "api_key":
+            continue
         if pconfig and pconfig.api_key_env_vars:
             env_vars = list(pconfig.api_key_env_vars)
         else:
@@ -856,6 +991,14 @@ def list_authenticated_providers(
 
         # Check if any env var is set
         has_creds = any(os.environ.get(ev) for ev in env_vars)
+        if not has_creds:
+            try:
+                from hermes_cli.auth import _load_auth_store
+                store = _load_auth_store()
+                if store and hermes_id in store.get("credential_pool", {}):
+                    has_creds = True
+            except Exception:
+                pass
         if not has_creds:
             continue
 
@@ -967,11 +1110,14 @@ def list_authenticated_providers(
         if not has_creds:
             continue
 
-        # Use curated list — look up by Hermes slug, fall back to overlay key
-        model_ids = curated.get(hermes_slug, []) or curated.get(pid, [])
-        # Merge with models.dev for preferred providers (same rationale as above).
-        if hermes_slug in _MODELS_DEV_PREFERRED:
-            model_ids = _merge_with_models_dev(hermes_slug, model_ids)
+        if hermes_slug in {"copilot", "copilot-acp"}:
+            model_ids = provider_model_ids(hermes_slug)
+        else:
+            # Use curated list — look up by Hermes slug, fall back to overlay key
+            model_ids = curated.get(hermes_slug, []) or curated.get(pid, [])
+            # Merge with models.dev for preferred providers (same rationale as above).
+            if hermes_slug in _MODELS_DEV_PREFERRED:
+                model_ids = _merge_with_models_dev(hermes_slug, model_ids)
         total = len(model_ids)
         top = model_ids[:max_models]
 
@@ -1094,6 +1240,15 @@ def list_authenticated_providers(
                     if m and m not in models_list:
                         models_list.append(m)
 
+            # Official OpenAI API rows in providers: often have base_url but no
+            # explicit models: dict — avoid a misleading zero count in /model.
+            if not models_list:
+                url_lower = str(api_url).strip().lower()
+                if "api.openai.com" in url_lower:
+                    fb = curated.get("openai") or []
+                    if fb:
+                        models_list = list(fb)
+
             # Try to probe /v1/models if URL is set (but don't block on it)
             # For now just show what we know from config
             results.append({
@@ -1117,66 +1272,113 @@ def list_authenticated_providers(
 
     # --- 4. Saved custom providers from config ---
     # Each ``custom_providers`` entry represents one model under a named
-    # provider. Entries sharing the same provider name are grouped into a
-    # single picker row so that e.g. four Ollama Cloud entries
-    # (qwen3-coder, glm-5.1, kimi-k2, minimax-m2.7) appear as one
-    # "Ollama Cloud" row with four models inside instead of four
-    # duplicate "Ollama Cloud" rows. Entries with distinct provider names
-    # still produce separate rows (e.g. Ollama Cloud vs Moonshot).
+    # provider. Entries sharing the same endpoint (``base_url`` + ``api_key``)
+    # are grouped into a single picker row, so e.g. four Ollama entries
+    # pointing at ``http://localhost:11434/v1`` with per-model display names
+    # ("Ollama — GLM 5.1", "Ollama — Qwen3-coder", ...) appear as one
+    # "Ollama" row with four models inside instead of four near-duplicates
+    # that differ only by suffix. Entries with distinct endpoints still
+    # produce separate rows.
+    #
+    # When the grouped endpoint matches ``current_base_url`` the group's
+    # slug becomes ``current_provider`` so that selecting a model from the
+    # picker flows back through the runtime provider that already holds
+    # valid credentials — no re-resolution needed.
     if custom_providers and isinstance(custom_providers, list):
         from collections import OrderedDict
 
-        groups: "OrderedDict[str, dict]" = OrderedDict()
+        # Key by (base_url, api_key) instead of slug: names frequently
+        # differ per model ("Ollama — X") while the endpoint stays the
+        # same. Slug-based grouping left them as separate rows.
+        groups: "OrderedDict[tuple, dict]" = OrderedDict()
         for entry in custom_providers:
             if not isinstance(entry, dict):
                 continue
 
-            display_name = (entry.get("name") or "").strip()
+            raw_name = (entry.get("name") or "").strip()
             api_url = (
                 entry.get("base_url", "")
                 or entry.get("url", "")
                 or entry.get("api", "")
                 or ""
-            ).strip()
-            if not display_name or not api_url:
+            ).strip().rstrip("/")
+            if not raw_name or not api_url:
                 continue
+            api_key = (entry.get("api_key") or "").strip()
 
-            slug = custom_provider_slug(display_name)
-            if slug not in groups:
-                groups[slug] = {
+            group_key = (api_url, api_key)
+            if group_key not in groups:
+                # Strip per-model suffix so "Ollama — GLM 5.1" becomes
+                # "Ollama" for the grouped row. Em dash is the convention
+                # Hermes's own writer uses; a hyphen variant is accepted
+                # for hand-edited configs.
+                display_name = raw_name
+                for sep in ("—", " - "):
+                    if sep in display_name:
+                        display_name = display_name.split(sep)[0].strip()
+                        break
+                if not display_name:
+                    display_name = raw_name
+                # If this endpoint matches the currently active one, use
+                # ``current_provider`` as the slug so picker-driven switches
+                # route through the live credential pipeline.
+                if (
+                    current_base_url
+                    and api_url == current_base_url.strip().rstrip("/")
+                ):
+                    slug = current_provider or custom_provider_slug(display_name)
+                else:
+                    slug = custom_provider_slug(display_name)
+                groups[group_key] = {
+                    "slug": slug,
                     "name": display_name,
                     "api_url": api_url,
                     "models": [],
                 }
+
             # The singular ``model:`` field only holds the currently
             # active model. Hermes's own writer (main.py::_save_custom_provider)
             # stores every configured model as a dict under ``models:``;
             # downstream readers (agent/models_dev.py, gateway/run.py,
             # run_agent.py, hermes_cli/config.py) already consume that dict.
-            # The /model picker previously ignored it, so multi-model
-            # custom providers appeared to have only the active model.
             default_model = (entry.get("model") or "").strip()
-            if default_model and default_model not in groups[slug]["models"]:
-                groups[slug]["models"].append(default_model)
+            if default_model and default_model not in groups[group_key]["models"]:
+                groups[group_key]["models"].append(default_model)
 
             cfg_models = entry.get("models", {})
             if isinstance(cfg_models, dict):
                 for m in cfg_models:
-                    if m and m not in groups[slug]["models"]:
-                        groups[slug]["models"].append(m)
+                    if m and m not in groups[group_key]["models"]:
+                        groups[group_key]["models"].append(m)
             elif isinstance(cfg_models, list):
                 for m in cfg_models:
-                    if m and m not in groups[slug]["models"]:
-                        groups[slug]["models"].append(m)
+                    if m and m not in groups[group_key]["models"]:
+                        groups[group_key]["models"].append(m)
 
-        for slug, grp in groups.items():
-            if slug.lower() in seen_slugs:
+        _section4_emitted_slugs: set = set()
+        for grp in groups.values():
+            slug = grp["slug"]
+            # If the slug is already claimed by a built-in / overlay /
+            # user-provider row (sections 1-3), skip this custom group
+            # to avoid shadowing a real provider.
+            if slug.lower() in seen_slugs and slug.lower() not in _section4_emitted_slugs:
                 continue
+            # If a prior section-4 group already used this slug (two custom
+            # endpoints with the same cleaned name — e.g. two OpenAI-
+            # compatible gateways named identically with different keys),
+            # append a counter so both rows stay visible in the picker.
+            if slug.lower() in _section4_emitted_slugs:
+                base_slug = slug
+                n = 2
+                while f"{base_slug}-{n}".lower() in seen_slugs:
+                    n += 1
+                slug = f"{base_slug}-{n}"
+                grp["slug"] = slug
             # Skip if section 3 already emitted this endpoint under its
-            # ``providers:`` dict key — matches on (display_name, base_url),
-            # the tuple section 4 groups by.  Prevents two picker rows
-            # labelled identically when callers pass both ``user_providers``
-            # and a compatibility-merged ``custom_providers`` list.
+            # ``providers:`` dict key — matches on (display_name, base_url).
+            # Prevents two picker rows labelled identically when callers
+            # pass both ``user_providers`` and a compatibility-merged
+            # ``custom_providers`` list.
             _pair_key = (
                 str(grp["name"]).strip().lower(),
                 str(grp["api_url"]).strip().rstrip("/").lower(),
@@ -1194,6 +1396,7 @@ def list_authenticated_providers(
                 "api_url": grp["api_url"],
             })
             seen_slugs.add(slug.lower())
+            _section4_emitted_slugs.add(slug.lower())
 
     # Sort: current provider first, then by model count descending
     results.sort(key=lambda r: (not r["is_current"], -r["total_models"]))

hermes_cli/models.py

@@ -33,6 +33,8 @@ COPILOT_REASONING_EFFORTS_O_SERIES = ["low", "medium", "high"]
 # (model_id, display description shown in menus)
 OPENROUTER_MODELS: list[tuple[str, str]] = [
     ("moonshotai/kimi-k2.6",            "recommended"),
+    ("deepseek/deepseek-v4-pro",        ""),
+    ("deepseek/deepseek-v4-flash",      ""),
     ("anthropic/claude-opus-4.7",       ""),
     ("anthropic/claude-opus-4.6",       ""),
     ("anthropic/claude-sonnet-4.6",     ""),
@@ -109,6 +111,8 @@ def _codex_curated_models() -> list[str]:
 _PROVIDER_MODELS: dict[str, list[str]] = {
     "nous": [
         "moonshotai/kimi-k2.6",
+        "deepseek/deepseek-v4-pro",
+        "deepseek/deepseek-v4-flash",
         "xiaomi/mimo-v2.5-pro",
         "xiaomi/mimo-v2.5",
         "anthropic/claude-opus-4.7",
@@ -138,6 +142,18 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "openai/gpt-5.4-pro",
         "openai/gpt-5.4-nano",
     ],
+    # Native OpenAI Chat Completions (api.openai.com). Used by /model counts and
+    # provider_model_ids fallback when /v1/models is unavailable.
+    "openai": [
+        "gpt-5.4",
+        "gpt-5.4-mini",
+        "gpt-5-mini",
+        "gpt-5.3-codex",
+        "gpt-5.2-codex",
+        "gpt-4.1",
+        "gpt-4o",
+        "gpt-4o-mini",
+    ],
     "openai-codex": _codex_curated_models(),
     "copilot-acp": [
         "copilot-acp",
@@ -151,10 +167,13 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "gpt-4.1",
         "gpt-4o",
         "gpt-4o-mini",
-        "claude-opus-4.6",
         "claude-sonnet-4.6",
+        "claude-sonnet-4",
         "claude-sonnet-4.5",
         "claude-haiku-4.5",
+        "gemini-3.1-pro-preview",
+        "gemini-3-pro-preview",
+        "gemini-3-flash-preview",
         "gemini-2.5-pro",
         "grok-code-fast-1",
     ],
@@ -246,10 +265,14 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "claude-haiku-4-5-20251001",
     ],
     "deepseek": [
+        "deepseek-v4-pro",
+        "deepseek-v4-flash",
         "deepseek-chat",
         "deepseek-reasoner",
     ],
     "xiaomi": [
+        "mimo-v2.5-pro",
+        "mimo-v2.5",
         "mimo-v2-pro",
         "mimo-v2-omni",
         "mimo-v2-flash",
@@ -301,6 +324,8 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "kimi-k2.5",
         "glm-5.1",
         "glm-5",
+        "mimo-v2.5-pro",
+        "mimo-v2.5",
         "mimo-v2-pro",
         "mimo-v2-omni",
         "minimax-m2.7",
@@ -672,7 +697,7 @@ def get_nous_recommended_aux_model(
 # ---------------------------------------------------------------------------
 # Canonical provider list — single source of truth for provider identity.
 # Every code path that lists, displays, or iterates providers derives from
-# this list:  hermes model, /model, /provider, list_authenticated_providers.
+# this list:  hermes model, /model, list_authenticated_providers.
 #
 # Fields:
 #   slug        — internal provider ID (used in config.yaml, --provider flag)
@@ -692,7 +717,7 @@ CANONICAL_PROVIDERS: list[ProviderEntry] = [
     ProviderEntry("ai-gateway",     "Vercel AI Gateway",        "Vercel AI Gateway (200+ models, $5 free credit, no markup)"),
     ProviderEntry("anthropic",      "Anthropic",                "Anthropic (Claude models — API key or Claude Code)"),
     ProviderEntry("openai-codex",   "OpenAI Codex",             "OpenAI Codex"),
-    ProviderEntry("xiaomi",         "Xiaomi MiMo",              "Xiaomi MiMo (MiMo-V2 models — pro, omni, flash)"),
+    ProviderEntry("xiaomi",         "Xiaomi MiMo",              "Xiaomi MiMo (MiMo-V2.5 and V2 models — pro, omni, flash)"),
     ProviderEntry("nvidia",         "NVIDIA NIM",               "NVIDIA NIM (Nemotron models — build.nvidia.com or local NIM)"),
     ProviderEntry("qwen-oauth",     "Qwen OAuth (Portal)",      "Qwen OAuth (reuses local Qwen CLI login)"),
     ProviderEntry("copilot",        "GitHub Copilot",           "GitHub Copilot (uses GITHUB_TOKEN or gh auth token)"),
@@ -1100,7 +1125,10 @@ def fetch_models_with_pricing(
         return _pricing_cache[cache_key]
 
     url = cache_key.rstrip("/") + "/v1/models"
-    headers: dict[str, str] = {"Accept": "application/json"}
+    headers: dict[str, str] = {
+        "Accept": "application/json",
+        "User-Agent": _HERMES_USER_AGENT,
+    }
     if api_key:
         headers["Authorization"] = f"Bearer {api_key}"
 
@@ -1674,7 +1702,19 @@ def provider_model_ids(provider: Optional[str], *, force_refresh: bool = False)
     if normalized == "openai-codex":
         from hermes_cli.codex_models import get_codex_model_ids
 
-        return get_codex_model_ids()
+        # Pass the live OAuth access token so the picker matches whatever
+        # ChatGPT lists for this account right now (new models appear without
+        # a Hermes release). Falls back to the hardcoded catalog if no token
+        # or the endpoint is unreachable.
+        access_token = None
+        try:
+            from hermes_cli.auth import resolve_codex_runtime_credentials
+
+            creds = resolve_codex_runtime_credentials(refresh_if_expiring=True)
+            access_token = creds.get("api_key")
+        except Exception:
+            access_token = None
+        return get_codex_model_ids(access_token=access_token)
     if normalized in {"copilot", "copilot-acp"}:
         try:
             live = _fetch_github_models(_resolve_copilot_catalog_api_key())
@@ -1720,6 +1760,17 @@ def provider_model_ids(provider: Optional[str], *, force_refresh: bool = False)
         live = fetch_ollama_cloud_models(force_refresh=force_refresh)
         if live:
             return live
+    if normalized == "openai":
+        api_key = os.getenv("OPENAI_API_KEY", "").strip()
+        if api_key:
+            base_raw = os.getenv("OPENAI_BASE_URL", "").strip().rstrip("/")
+            base = base_raw or "https://api.openai.com/v1"
+            try:
+                live = fetch_api_models(api_key, base)
+                if live:
+                    return live
+            except Exception:
+                pass
     if normalized == "custom":
         base_url = _get_custom_base_url()
         if base_url:
@@ -1874,6 +1925,51 @@ def fetch_github_model_catalog(
     return None
 
 
+# ─── Copilot catalog context-window helpers ─────────────────────────────────
+
+# Module-level cache: {model_id: max_prompt_tokens}
+_copilot_context_cache: dict[str, int] = {}
+_copilot_context_cache_time: float = 0.0
+_COPILOT_CONTEXT_CACHE_TTL = 3600  # 1 hour
+
+
+def get_copilot_model_context(model_id: str, api_key: Optional[str] = None) -> Optional[int]:
+    """Look up max_prompt_tokens for a Copilot model from the live /models API.
+
+    Results are cached in-process for 1 hour to avoid repeated API calls.
+    Returns the token limit or None if not found.
+    """
+    global _copilot_context_cache, _copilot_context_cache_time
+
+    # Serve from cache if fresh
+    if _copilot_context_cache and (time.time() - _copilot_context_cache_time < _COPILOT_CONTEXT_CACHE_TTL):
+        if model_id in _copilot_context_cache:
+            return _copilot_context_cache[model_id]
+        # Cache is fresh but model not in it — don't re-fetch
+        return None
+
+    # Fetch and populate cache
+    catalog = fetch_github_model_catalog(api_key=api_key)
+    if not catalog:
+        return None
+
+    cache: dict[str, int] = {}
+    for item in catalog:
+        mid = str(item.get("id") or "").strip()
+        if not mid:
+            continue
+        caps = item.get("capabilities") or {}
+        limits = caps.get("limits") or {}
+        max_prompt = limits.get("max_prompt_tokens")
+        if isinstance(max_prompt, int) and max_prompt > 0:
+            cache[mid] = max_prompt
+
+    _copilot_context_cache = cache
+    _copilot_context_cache_time = time.time()
+
+    return cache.get(model_id)
+
+
 def _is_github_models_base_url(base_url: Optional[str]) -> bool:
     normalized = (base_url or "").strip().rstrip("/").lower()
     return (
@@ -1907,6 +2003,7 @@ _COPILOT_MODEL_ALIASES = {
     "openai/o4-mini": "gpt-5-mini",
     "anthropic/claude-opus-4.6": "claude-opus-4.6",
     "anthropic/claude-sonnet-4.6": "claude-sonnet-4.6",
+    "anthropic/claude-sonnet-4": "claude-sonnet-4",
     "anthropic/claude-sonnet-4.5": "claude-sonnet-4.5",
     "anthropic/claude-haiku-4.5": "claude-haiku-4.5",
     # Dash-notation fallbacks: Hermes' default Claude IDs elsewhere use
@@ -1916,10 +2013,12 @@ _COPILOT_MODEL_ALIASES = {
     # "model_not_supported".  See issue #6879.
     "claude-opus-4-6": "claude-opus-4.6",
     "claude-sonnet-4-6": "claude-sonnet-4.6",
+    "claude-sonnet-4-0": "claude-sonnet-4",
     "claude-sonnet-4-5": "claude-sonnet-4.5",
     "claude-haiku-4-5": "claude-haiku-4.5",
     "anthropic/claude-opus-4-6": "claude-opus-4.6",
     "anthropic/claude-sonnet-4-6": "claude-sonnet-4.6",
+    "anthropic/claude-sonnet-4-0": "claude-sonnet-4",
     "anthropic/claude-sonnet-4-5": "claude-sonnet-4.5",
     "anthropic/claude-haiku-4-5": "claude-haiku-4.5",
 }
@@ -2144,8 +2243,15 @@ def probe_api_models(
     api_key: Optional[str],
     base_url: Optional[str],
     timeout: float = 5.0,
+    api_mode: Optional[str] = None,
 ) -> dict[str, Any]:
-    """Probe an OpenAI-compatible ``/models`` endpoint with light URL heuristics."""
+    """Probe a ``/models`` endpoint with light URL heuristics.
+
+    For ``anthropic_messages`` mode, uses ``x-api-key`` and
+    ``anthropic-version`` headers (Anthropic's native auth) instead of
+    ``Authorization: Bearer``.  The response shape (``data[].id``) is
+    identical, so the same parser works for both.
+    """
     normalized = (base_url or "").strip().rstrip("/")
     if not normalized:
         return {
@@ -2177,7 +2283,10 @@ def probe_api_models(
 
     tried: list[str] = []
     headers: dict[str, str] = {"User-Agent": _HERMES_USER_AGENT}
-    if api_key:
+    if api_key and api_mode == "anthropic_messages":
+        headers["x-api-key"] = api_key
+        headers["anthropic-version"] = "2023-06-01"
+    elif api_key:
         headers["Authorization"] = f"Bearer {api_key}"
     if normalized.startswith(COPILOT_BASE_URL):
         headers.update(copilot_default_headers())
@@ -2219,7 +2328,10 @@ def _fetch_ai_gateway_models(timeout: float = 5.0) -> Optional[list[str]]:
         base_url = AI_GATEWAY_BASE_URL
 
     url = base_url.rstrip("/") + "/models"
-    headers: dict[str, str] = {"Authorization": f"Bearer {api_key}"}
+    headers: dict[str, str] = {
+        "Authorization": f"Bearer {api_key}",
+        "User-Agent": _HERMES_USER_AGENT,
+    }
     req = urllib.request.Request(url, headers=headers)
     try:
         with urllib.request.urlopen(req, timeout=timeout) as resp:
@@ -2239,13 +2351,14 @@ def fetch_api_models(
     api_key: Optional[str],
     base_url: Optional[str],
     timeout: float = 5.0,
+    api_mode: Optional[str] = None,
 ) -> Optional[list[str]]:
     """Fetch the list of available model IDs from the provider's ``/models`` endpoint.
 
     Returns a list of model ID strings, or ``None`` if the endpoint could not
     be reached (network error, timeout, auth failure, etc.).
     """
-    return probe_api_models(api_key, base_url, timeout=timeout).get("models")
+    return probe_api_models(api_key, base_url, timeout=timeout, api_mode=api_mode).get("models")
 
 
 # ---------------------------------------------------------------------------
@@ -2373,6 +2486,7 @@ def validate_requested_model(
     *,
     api_key: Optional[str] = None,
     base_url: Optional[str] = None,
+    api_mode: Optional[str] = None,
 ) -> dict[str, Any]:
     """
     Validate a ``/model`` value for the active provider.
@@ -2414,7 +2528,11 @@ def validate_requested_model(
         }
 
     if normalized == "custom":
-        probe = probe_api_models(api_key, base_url)
+        # Try probing with correct auth for the api_mode.
+        if api_mode == "anthropic_messages":
+            probe = probe_api_models(api_key, base_url, api_mode=api_mode)
+        else:
+            probe = probe_api_models(api_key, base_url)
         api_models = probe.get("models")
         if api_models is not None:
             if requested_for_lookup in set(api_models):
@@ -2463,12 +2581,17 @@ def validate_requested_model(
             f"Note: could not reach this custom endpoint's model listing at `{probe.get('probed_url')}`. "
             f"Hermes will still save `{requested}`, but the endpoint should expose `/models` for verification."
         )
+        if api_mode == "anthropic_messages":
+            message += (
+                "\n  Many Anthropic-compatible proxies do not implement the Models API "
+                "(GET /v1/models).  The model name has been accepted without verification."
+            )
         if probe.get("suggested_base_url"):
             message += f"\n  If this server expects `/v1`, try base URL: `{probe.get('suggested_base_url')}`"
 
         return {
-            "accepted": False,
-            "persist": False,
+            "accepted": api_mode == "anthropic_messages",
+            "persist": True,
             "recognized": False,
             "message": message,
         }
@@ -2556,10 +2679,100 @@ def validate_requested_model(
                 ),
             }
 
+    # Native Anthropic provider: /v1/models requires x-api-key (or Bearer for
+    # OAuth) plus anthropic-version headers.  The generic OpenAI-style probe
+    # below uses plain Bearer auth and 401s against Anthropic, so dispatch to
+    # the native fetcher which handles both API keys and Claude-Code OAuth
+    # tokens.  (The api_mode=="anthropic_messages" branch below handles the
+    # Messages-API transport case separately.)
+    if normalized == "anthropic":
+        anthropic_models = _fetch_anthropic_models()
+        if anthropic_models is not None:
+            if requested_for_lookup in set(anthropic_models):
+                return {
+                    "accepted": True,
+                    "persist": True,
+                    "recognized": True,
+                    "message": None,
+                }
+            auto = get_close_matches(requested_for_lookup, anthropic_models, n=1, cutoff=0.9)
+            if auto:
+                return {
+                    "accepted": True,
+                    "persist": True,
+                    "recognized": True,
+                    "corrected_model": auto[0],
+                    "message": f"Auto-corrected `{requested}` → `{auto[0]}`",
+                }
+            suggestions = get_close_matches(requested, anthropic_models, n=3, cutoff=0.5)
+            suggestion_text = ""
+            if suggestions:
+                suggestion_text = "\n  Similar models: " + ", ".join(f"`{s}`" for s in suggestions)
+            # Accept anyway — Anthropic sometimes gates newer/preview models
+            # (e.g. snapshot IDs, early-access releases) behind accounts
+            # even though they aren't listed on /v1/models.
+            return {
+                "accepted": True,
+                "persist": True,
+                "recognized": False,
+                "message": (
+                    f"Note: `{requested}` was not found in Anthropic's /v1/models listing. "
+                    f"It may still work if you have early-access or snapshot IDs."
+                    f"{suggestion_text}"
+                ),
+            }
+        # _fetch_anthropic_models returned None — no token resolvable or
+        # network failure.  Fall through to the generic warning below.
+
+    # Anthropic Messages API: many proxies don't implement /v1/models.
+    # Try probing with correct auth; if it fails, accept with a warning.
+    if api_mode == "anthropic_messages":
+        api_models = fetch_api_models(api_key, base_url, api_mode=api_mode)
+        if api_models is not None:
+            if requested_for_lookup in set(api_models):
+                return {
+                    "accepted": True,
+                    "persist": True,
+                    "recognized": True,
+                    "message": None,
+                }
+            auto = get_close_matches(requested_for_lookup, api_models, n=1, cutoff=0.9)
+            if auto:
+                return {
+                    "accepted": True,
+                    "persist": True,
+                    "recognized": True,
+                    "corrected_model": auto[0],
+                    "message": f"Auto-corrected `{requested}` → `{auto[0]}`",
+                }
+        # Probe failed or model not found — accept anyway (proxy likely
+        # doesn't implement the Anthropic Models API).
+        return {
+            "accepted": True,
+            "persist": True,
+            "recognized": False,
+            "message": (
+                f"Note: could not verify `{requested}` against this endpoint's "
+                f"model listing.  Many Anthropic-compatible proxies do not "
+                f"implement GET /v1/models.  The model name has been accepted "
+                f"without verification."
+            ),
+        }
+
     # Probe the live API to check if the model actually exists
     api_models = fetch_api_models(api_key, base_url)
 
     if api_models is not None:
+        # Gemini's OpenAI-compat /v1beta/openai/models endpoint returns IDs
+        # prefixed with "models/" (e.g. "models/gemini-2.5-flash") — native
+        # Gemini-API convention.  Our curated list and user input both use
+        # the bare ID, so a direct set-membership check drops every known
+        # Gemini model.  Strip the prefix before comparison.  See #12532.
+        if normalized == "gemini":
+            api_models = [
+                m[len("models/"):] if isinstance(m, str) and m.startswith("models/") else m
+                for m in api_models
+            ]
         if requested_for_lookup in set(api_models):
             # API confirmed the model exists
             return {

hermes_cli/pairing.py

@@ -44,7 +44,7 @@ def _cmd_list(store):
         for p in pending:
             print(
                 f"  {p['platform']:<12} {p['code']:<10} {p['user_id']:<20} "
-                f"{p.get('user_name', ''):<20} {p['age_minutes']}m ago"
+                f"{(p.get('user_name') or ''):<20} {p['age_minutes']}m ago"
             )
     else:
         print("\n  No pending pairing requests.")
@@ -54,7 +54,7 @@ def _cmd_list(store):
         print(f"  {'Platform':<12} {'User ID':<20} {'Name':<20}")
         print(f"  {'--------':<12} {'-------':<20} {'----':<20}")
         for a in approved:
-            print(f"  {a['platform']:<12} {a['user_id']:<20} {a.get('user_name', ''):<20}")
+            print(f"  {a['platform']:<12} {a['user_id']:<20} {(a.get('user_name') or ''):<20}")
     else:
         print("\n  No approved users.")
 
@@ -69,7 +69,7 @@ def _cmd_approve(store, platform: str, code: str):
     result = store.approve_code(platform, code)
     if result:
         uid = result["user_id"]
-        name = result.get("user_name", "")
+        name = result.get("user_name") or ""
         display = f"{name} ({uid})" if name else uid
         print(f"\n  Approved! User {display} on {platform} can now use the bot~")
         print("  They'll be recognized automatically on their next message.\n")

hermes_cli/platforms.py

@@ -38,6 +38,7 @@ PLATFORMS: OrderedDict[str, PlatformInfo] = OrderedDict([
     ("qqbot",          PlatformInfo(label="💬 QQBot",           default_toolset="hermes-qqbot")),
     ("webhook",        PlatformInfo(label="🔗 Webhook",         default_toolset="hermes-webhook")),
     ("api_server",     PlatformInfo(label="🌐 API Server",      default_toolset="hermes-api-server")),
+    ("cron",           PlatformInfo(label="⏰ Cron",            default_toolset="hermes-cron")),
 ])
 
 

hermes_cli/plugins.py

@@ -71,6 +71,14 @@ VALID_HOOKS: Set[str] = {
     "on_session_finalize",
     "on_session_reset",
     "subagent_stop",
+    # Gateway pre-dispatch hook. Fired once per incoming MessageEvent
+    # after the internal-event guard but BEFORE auth/pairing and agent
+    # dispatch. Plugins may return a dict to influence flow:
+    #   {"action": "skip",    "reason": "..."}  -> drop message (no reply)
+    #   {"action": "rewrite", "text": "..."}    -> replace event.text, continue
+    #   {"action": "allow"}  /  None             -> normal dispatch
+    # Kwargs: event: MessageEvent, gateway: GatewayRunner, session_store.
+    "pre_gateway_dispatch",
 }
 
 ENTRY_POINTS_GROUP = "hermes_agent.plugins"
@@ -512,10 +520,23 @@ class PluginManager:
     # Public
     # -----------------------------------------------------------------------
 
-    def discover_and_load(self) -> None:
-        """Scan all plugin sources and load each plugin found."""
-        if self._discovered:
+    def discover_and_load(self, force: bool = False) -> None:
+        """Scan all plugin sources and load each plugin found.
+
+        When ``force`` is true, clear cached discovery state first so config
+        changes or newly-added bundled backends become visible in long-lived
+        sessions without requiring a full agent restart.
+        """
+        if self._discovered and not force:
             return
+        if force:
+            self._plugins.clear()
+            self._hooks.clear()
+            self._plugin_tool_names.clear()
+            self._cli_commands.clear()
+            self._plugin_commands.clear()
+            self._plugin_skills.clear()
+            self._context_engine = None
         self._discovered = True
 
         manifests: List[PluginManifest] = []
@@ -1029,9 +1050,13 @@ def get_plugin_manager() -> PluginManager:
     return _plugin_manager
 
 
-def discover_plugins() -> None:
-    """Discover and load all plugins (idempotent)."""
-    get_plugin_manager().discover_and_load()
+def discover_plugins(force: bool = False) -> None:
+    """Discover and load all plugins.
+
+    Default behavior is idempotent. Pass ``force=True`` to rescan plugin
+    manifests and reload state in the current process.
+    """
+    get_plugin_manager().discover_and_load(force=force)
 
 
 def invoke_hook(hook_name: str, **kwargs: Any) -> List[Any]:
@@ -1082,10 +1107,13 @@ def get_pre_tool_call_block_message(
     return None
 
 
-def _ensure_plugins_discovered() -> PluginManager:
-    """Return the global manager after running idempotent plugin discovery."""
+def _ensure_plugins_discovered(force: bool = False) -> PluginManager:
+    """Return the global manager after ensuring plugin discovery has run.
+
+    Pass ``force=True`` to rescan in the current process.
+    """
     manager = get_plugin_manager()
-    manager.discover_and_load()
+    manager.discover_and_load(force=force)
     return manager
 
 

hermes_cli/profiles.py

@@ -863,19 +863,15 @@ def _safe_extract_profile_archive(archive: Path, destination: Path) -> None:
                 pass
 
 
-def import_profile(archive_path: str, name: Optional[str] = None) -> Path:
-    """Import a profile from a tar.gz archive.
+def _inspect_profile_archive_roots(archive: Path) -> set[str]:
+    """Return the archive's top-level directory names.
 
-    If *name* is not given, infers it from the archive's top-level directory.
-    Returns the imported profile directory.
+    Profile imports expect exactly one root directory. Inspecting the archive
+    before extraction lets us stage the import safely instead of mutating a
+    live profile tree first and reconciling names later.
     """
     import tarfile
 
-    archive = Path(archive_path)
-    if not archive.exists():
-        raise FileNotFoundError(f"Archive not found: {archive}")
-
-    # Peek at the archive to find the top-level directory name
     with tarfile.open(archive, "r:gz") as tf:
         top_dirs = {
             parts[0]
@@ -889,13 +885,33 @@ def import_profile(archive_path: str, name: Optional[str] = None) -> Path:
                 for member in tf.getmembers()
                 if member.isdir()
             }
+    return top_dirs
 
-    inferred_name = name or (top_dirs.pop() if len(top_dirs) == 1 else None)
+
+def import_profile(archive_path: str, name: Optional[str] = None) -> Path:
+    """Import a profile from a tar.gz archive.
+
+    If *name* is not given, infers it from the archive's top-level directory.
+    Returns the imported profile directory.
+    """
+    import tempfile
+
+    archive = Path(archive_path)
+    if not archive.exists():
+        raise FileNotFoundError(f"Archive not found: {archive}")
+
+    top_dirs = _inspect_profile_archive_roots(archive)
+    archive_root = top_dirs.pop() if len(top_dirs) == 1 else None
+    inferred_name = name or archive_root
     if not inferred_name:
         raise ValueError(
             "Cannot determine profile name from archive. "
             "Specify it explicitly: hermes profile import <archive> --name <name>"
         )
+    if archive_root is None:
+        raise ValueError(
+            "Profile archive must contain exactly one top-level directory."
+        )
 
     # Archives exported from the default profile have "default/" as top-level
     # dir.  Importing as "default" would target ~/.hermes itself — disallow
@@ -914,12 +930,22 @@ def import_profile(archive_path: str, name: Optional[str] = None) -> Path:
     profiles_root = _get_profiles_root()
     profiles_root.mkdir(parents=True, exist_ok=True)
 
-    _safe_extract_profile_archive(archive, profiles_root)
+    with tempfile.TemporaryDirectory(prefix="hermes_profile_import_") as tmpdir:
+        staging_root = Path(tmpdir)
+        _safe_extract_profile_archive(archive, staging_root)
 
-    # If the archive extracted under a different name, rename
-    extracted = profiles_root / (top_dirs.pop() if top_dirs else inferred_name)
-    if extracted != profile_dir and extracted.exists():
-        extracted.rename(profile_dir)
+        extracted = staging_root / archive_root
+        if not extracted.is_dir():
+            raise ValueError(
+                f"Profile archive root is missing or invalid: {archive_root}"
+            )
+
+        final_source = extracted
+        if archive_root != inferred_name:
+            final_source = staging_root / inferred_name
+            extracted.rename(final_source)
+
+        shutil.move(str(final_source), str(profile_dir))
 
     return profile_dir
 

hermes_cli/providers.py

@@ -116,6 +116,10 @@ HERMES_OVERLAYS: Dict[str, HermesOverlay] = {
         transport="openai_chat",
         base_url_env_var="DASHSCOPE_BASE_URL",
     ),
+    "alibaba-coding-plan": HermesOverlay(
+        transport="openai_chat",
+        base_url_env_var="ALIBABA_CODING_PLAN_BASE_URL",
+    ),
     "vercel": HermesOverlay(
         transport="openai_chat",
         is_aggregator=True,
@@ -259,6 +263,9 @@ ALIASES: Dict[str, str] = {
     "aliyun": "alibaba",
     "qwen": "alibaba",
     "alibaba-cloud": "alibaba",
+    "alibaba_coding": "alibaba-coding-plan",
+    "alibaba-coding": "alibaba-coding-plan",
+    "alibaba_coding_plan": "alibaba-coding-plan",
 
     # google-gemini-cli (OAuth + Code Assist)
     "gemini-cli": "google-gemini-cli",

hermes_cli/pty_bridge.py

@@ -1,221 +0,0 @@
-"""PTY bridge for `hermes dashboard` chat tab.
-
-Wraps a child process behind a pseudo-terminal so its ANSI output can be
-streamed to a browser-side terminal emulator (xterm.js) and typed
-keystrokes can be fed back in.  The only caller today is the
-``/api/pty`` WebSocket endpoint in ``hermes_cli.web_server``.
-
-Design constraints:
-
-* **POSIX-only.**  Hermes Agent supports Windows exclusively via WSL, which
-  exposes a native POSIX PTY via ``openpty(3)``.  Native Windows Python
-  has no PTY; :class:`PtyUnavailableError` is raised with a user-readable
-  install/platform message so the dashboard can render a banner instead of
-  crashing.
-* **Zero Node dependency on the server side.**  We use :mod:`ptyprocess`,
-  which is a pure-Python wrapper around the OS calls.  The browser talks
-  to the same ``hermes --tui`` binary it would launch from the CLI, so
-  every TUI feature (slash popover, model picker, tool rows, markdown,
-  skin engine, clarify/sudo/approval prompts) ships automatically.
-* **Byte-safe I/O.**  Reads and writes go through the PTY master fd
-  directly — we avoid :class:`ptyprocess.PtyProcessUnicode` because
-  streaming ANSI is inherently byte-oriented and UTF-8 boundaries may land
-  mid-read.
-"""
-
-from __future__ import annotations
-
-import errno
-import fcntl
-import os
-import select
-import signal
-import struct
-import sys
-import termios
-import time
-from typing import Optional, Sequence
-
-try:
-    import ptyprocess  # type: ignore
-    _PTY_AVAILABLE = not sys.platform.startswith("win")
-except ImportError:  # pragma: no cover - dev env without ptyprocess
-    ptyprocess = None  # type: ignore
-    _PTY_AVAILABLE = False
-
-
-__all__ = ["PtyBridge", "PtyUnavailableError"]
-
-
-class PtyUnavailableError(RuntimeError):
-    """Raised when a PTY cannot be created on this platform.
-
-    Today this means native Windows (no ConPTY bindings) or a dev
-    environment missing the ``ptyprocess`` dependency.  The dashboard
-    surfaces the message to the user as a chat-tab banner.
-    """
-
-
-class PtyBridge:
-    """Thin wrapper around ``ptyprocess.PtyProcess`` for byte streaming.
-
-    Not thread-safe.  A single bridge is owned by the WebSocket handler
-    that spawned it; the reader runs in an executor thread while writes
-    happen on the event-loop thread.  Both sides are OK because the
-    kernel PTY is the actual synchronization point — we never call
-    :mod:`ptyprocess` methods concurrently, we only call ``os.read`` and
-    ``os.write`` on the master fd, which is safe.
-    """
-
-    def __init__(self, proc: "ptyprocess.PtyProcess"):  # type: ignore[name-defined]
-        self._proc = proc
-        self._fd: int = proc.fd
-        self._closed = False
-
-    # -- lifecycle --------------------------------------------------------
-
-    @classmethod
-    def is_available(cls) -> bool:
-        """True if a PTY can be spawned on this platform."""
-        return bool(_PTY_AVAILABLE)
-
-    @classmethod
-    def spawn(
-        cls,
-        argv: Sequence[str],
-        *,
-        cwd: Optional[str] = None,
-        env: Optional[dict] = None,
-        cols: int = 80,
-        rows: int = 24,
-    ) -> "PtyBridge":
-        """Spawn ``argv`` behind a new PTY and return a bridge.
-
-        Raises :class:`PtyUnavailableError` if the platform can't host a
-        PTY.  Raises :class:`FileNotFoundError` or :class:`OSError` for
-        ordinary exec failures (missing binary, bad cwd, etc.).
-        """
-        if not _PTY_AVAILABLE:
-            raise PtyUnavailableError(
-                "Pseudo-terminals are unavailable on this platform. "
-                "Hermes Agent supports Windows only via WSL."
-            )
-        # Let caller-supplied env fully override inheritance; if they pass
-        # None we inherit the server's env (same semantics as subprocess).
-        spawn_env = os.environ.copy() if env is None else env
-        proc = ptyprocess.PtyProcess.spawn(  # type: ignore[union-attr]
-            list(argv),
-            cwd=cwd,
-            env=spawn_env,
-            dimensions=(rows, cols),
-        )
-        return cls(proc)
-
-    @property
-    def pid(self) -> int:
-        return int(self._proc.pid)
-
-    def is_alive(self) -> bool:
-        if self._closed:
-            return False
-        try:
-            return bool(self._proc.isalive())
-        except Exception:
-            return False
-
-    # -- I/O --------------------------------------------------------------
-
-    def read(self, timeout: float = 0.2) -> Optional[bytes]:
-        """Read up to 64 KiB of raw bytes from the PTY master.
-
-        Returns:
-            * bytes — zero or more bytes of child output
-            * empty bytes (``b""``) — no data available within ``timeout``
-            * None — child has exited and the master fd is at EOF
-
-        Never blocks longer than ``timeout`` seconds.  Safe to call after
-        :meth:`close`; returns ``None`` in that case.
-        """
-        if self._closed:
-            return None
-        try:
-            readable, _, _ = select.select([self._fd], [], [], timeout)
-        except (OSError, ValueError):
-            return None
-        if not readable:
-            return b""
-        try:
-            data = os.read(self._fd, 65536)
-        except OSError as exc:
-            # EIO on Linux = slave side closed.  EBADF = already closed.
-            if exc.errno in (errno.EIO, errno.EBADF):
-                return None
-            raise
-        if not data:
-            return None
-        return data
-
-    def write(self, data: bytes) -> None:
-        """Write raw bytes to the PTY master (i.e. the child's stdin)."""
-        if self._closed or not data:
-            return
-        # os.write can return a short write under load; loop until drained.
-        view = memoryview(data)
-        while view:
-            try:
-                n = os.write(self._fd, view)
-            except OSError as exc:
-                if exc.errno in (errno.EIO, errno.EBADF, errno.EPIPE):
-                    return
-                raise
-            if n <= 0:
-                return
-            view = view[n:]
-
-    def resize(self, cols: int, rows: int) -> None:
-        """Forward a terminal resize to the child via ``TIOCSWINSZ``."""
-        if self._closed:
-            return
-        # struct winsize: rows, cols, xpixel, ypixel (all unsigned short)
-        winsize = struct.pack("HHHH", max(1, rows), max(1, cols), 0, 0)
-        try:
-            fcntl.ioctl(self._fd, termios.TIOCSWINSZ, winsize)
-        except OSError:
-            pass
-
-    # -- teardown ---------------------------------------------------------
-
-    def close(self) -> None:
-        """Terminate the child (SIGTERM → 0.5s grace → SIGKILL) and close fds.
-
-        Idempotent.  Reaping the child is important so we don't leak
-        zombies across the lifetime of the dashboard process.
-        """
-        if self._closed:
-            return
-        self._closed = True
-
-        # SIGHUP is the conventional "your terminal went away" signal.
-        # We escalate if the child ignores it.
-        for sig in (signal.SIGHUP, signal.SIGTERM, signal.SIGKILL):
-            if not self._proc.isalive():
-                break
-            try:
-                self._proc.kill(sig)
-            except Exception:
-                pass
-            deadline = time.monotonic() + 0.5
-            while self._proc.isalive() and time.monotonic() < deadline:
-                time.sleep(0.02)
-
-        try:
-            self._proc.close(force=True)
-        except Exception:
-            pass
-
-    # Context-manager sugar — handy in tests and ad-hoc scripts.
-    def __enter__(self) -> "PtyBridge":
-        return self
-
-    def __exit__(self, *_exc) -> None:
-        self.close()

hermes_cli/runtime_provider.py

@@ -36,6 +36,29 @@ def _normalize_custom_provider_name(value: str) -> str:
     return value.strip().lower().replace(" ", "-")
 
 
+def _loopback_hostname(host: str) -> bool:
+    h = (host or "").lower().rstrip(".")
+    return h in {"localhost", "127.0.0.1", "::1", "0.0.0.0"}
+
+
+def _config_base_url_trustworthy_for_bare_custom(cfg_base_url: str, cfg_provider: str) -> bool:
+    """Decide whether ``model.base_url`` may back bare ``custom`` runtime resolution.
+
+    GitHub #14676: the model picker can select Custom while ``model.provider`` still reflects a
+    previous provider. Reject non-loopback URLs unless the YAML provider is already ``custom``,
+    so a stale OpenRouter/Z.ai base_url cannot hijack local ``custom`` sessions.
+    """
+    cfg_provider_norm = (cfg_provider or "").strip().lower()
+    bu = (cfg_base_url or "").strip()
+    if not bu:
+        return False
+    if cfg_provider_norm == "custom":
+        return True
+    if base_url_host_matches(bu, "openrouter.ai"):
+        return False
+    return _loopback_hostname(base_url_hostname(bu))
+
+
 def _detect_api_mode_for_url(base_url: str) -> Optional[str]:
     """Auto-detect api_mode from the resolved base URL.
 
@@ -160,8 +183,16 @@ def _resolve_runtime_from_pool_entry(
     requested_provider: str,
     model_cfg: Optional[Dict[str, Any]] = None,
     pool: Optional[CredentialPool] = None,
+    target_model: Optional[str] = None,
 ) -> Dict[str, Any]:
     model_cfg = model_cfg or _get_model_config()
+    # When the caller is resolving for a specific target model (e.g. a /model
+    # mid-session switch), prefer that over the persisted model.default. This
+    # prevents api_mode being computed from a stale config default that no
+    # longer matches the model actually being used — the bug that caused
+    # opencode-zen /v1 to be stripped for chat_completions requests when
+    # config.default was still a Claude model.
+    effective_model = (target_model or model_cfg.get("default") or "")
     base_url = (getattr(entry, "runtime_base_url", None) or getattr(entry, "base_url", None) or "").rstrip("/")
     api_key = getattr(entry, "runtime_api_key", None) or getattr(entry, "access_token", "")
     api_mode = "chat_completions"
@@ -207,7 +238,7 @@ def _resolve_runtime_from_pool_entry(
             api_mode = configured_mode
         elif provider in ("opencode-zen", "opencode-go"):
             from hermes_cli.models import opencode_model_api_mode
-            api_mode = opencode_model_api_mode(provider, model_cfg.get("default", ""))
+            api_mode = opencode_model_api_mode(provider, effective_model)
         else:
             # Auto-detect Anthropic-compatible endpoints (/anthropic suffix,
             # Kimi /coding, api.openai.com → codex_responses, api.x.ai →
@@ -323,12 +354,16 @@ def _get_named_custom_provider(requested_provider: str) -> Optional[Dict[str, An
                 # Found match by provider key
                 base_url = entry.get("api") or entry.get("url") or entry.get("base_url") or ""
                 if base_url:
-                    return {
+                    result = {
                         "name": entry.get("name", ep_name),
                         "base_url": base_url.strip(),
                         "api_key": resolved_api_key,
                         "model": entry.get("default_model", ""),
                     }
+                    api_mode = _parse_api_mode(entry.get("api_mode"))
+                    if api_mode:
+                        result["api_mode"] = api_mode
+                    return result
             # Also check the 'name' field if present
             display_name = entry.get("name", "")
             if display_name:
@@ -337,12 +372,16 @@ def _get_named_custom_provider(requested_provider: str) -> Optional[Dict[str, An
                     # Found match by display name
                     base_url = entry.get("api") or entry.get("url") or entry.get("base_url") or ""
                     if base_url:
-                        return {
+                        result = {
                             "name": display_name,
                             "base_url": base_url.strip(),
                             "api_key": resolved_api_key,
                             "model": entry.get("default_model", ""),
                         }
+                        api_mode = _parse_api_mode(entry.get("api_mode"))
+                        if api_mode:
+                            result["api_mode"] = api_mode
+                        return result
 
     # Fall back to custom_providers: list (legacy format)
     custom_providers = config.get("custom_providers")
@@ -464,6 +503,7 @@ def _resolve_openrouter_runtime(
     cfg_provider = cfg_provider.strip().lower()
 
     env_openrouter_base_url = os.getenv("OPENROUTER_BASE_URL", "").strip()
+    env_custom_base_url = os.getenv("CUSTOM_BASE_URL", "").strip()
 
     # Use config base_url when available and the provider context matches.
     # OPENAI_BASE_URL env var is no longer consulted — config.yaml is
@@ -473,11 +513,14 @@ def _resolve_openrouter_runtime(
         if requested_norm == "auto":
             if not cfg_provider or cfg_provider == "auto":
                 use_config_base_url = True
-        elif requested_norm == "custom" and cfg_provider == "custom":
+        elif requested_norm == "custom" and _config_base_url_trustworthy_for_bare_custom(
+            cfg_base_url, cfg_provider
+        ):
             use_config_base_url = True
 
     base_url = (
         (explicit_base_url or "").strip()
+        or env_custom_base_url
         or (cfg_base_url.strip() if use_config_base_url else "")
         or env_openrouter_base_url
         or OPENROUTER_BASE_URL
@@ -689,8 +732,18 @@ def resolve_runtime_provider(
     requested: Optional[str] = None,
     explicit_api_key: Optional[str] = None,
     explicit_base_url: Optional[str] = None,
+    target_model: Optional[str] = None,
 ) -> Dict[str, Any]:
-    """Resolve runtime provider credentials for agent execution."""
+    """Resolve runtime provider credentials for agent execution.
+
+    target_model: Optional override for model_cfg.get("default") when
+    computing provider-specific api_mode (e.g. OpenCode Zen/Go where different
+    models route through different API surfaces). Callers performing an
+    explicit mid-session model switch should pass the new model here so
+    api_mode is derived from the model they are switching TO, not the stale
+    persisted default. Other callers can leave it None to preserve existing
+    behavior (api_mode derived from config).
+    """
     requested_provider = resolve_requested_provider(requested)
 
     custom_runtime = _resolve_named_custom_runtime(
@@ -772,6 +825,7 @@ def resolve_runtime_provider(
                 requested_provider=requested_provider,
                 model_cfg=model_cfg,
                 pool=pool,
+                target_model=target_model,
             )
 
     if provider == "nous":
@@ -990,7 +1044,11 @@ def resolve_runtime_provider(
                 api_mode = configured_mode
             elif provider in ("opencode-zen", "opencode-go"):
                 from hermes_cli.models import opencode_model_api_mode
-                api_mode = opencode_model_api_mode(provider, model_cfg.get("default", ""))
+                # Prefer the target_model from the caller (explicit mid-session
+                # switch) over the stale model.default; see _resolve_runtime_from_pool_entry
+                # for the same rationale.
+                _effective = target_model or model_cfg.get("default", "")
+                api_mode = opencode_model_api_mode(provider, _effective)
             else:
                 # Auto-detect Anthropic-compatible endpoints by URL convention
                 # (e.g. https://api.minimax.io/anthropic, https://dashscope.../anthropic)

hermes_cli/setup.py

@@ -103,7 +103,7 @@ _DEFAULT_PROVIDER_MODELS = {
     "ai-gateway": ["anthropic/claude-opus-4.6", "anthropic/claude-sonnet-4.6", "openai/gpt-5", "google/gemini-3-flash"],
     "kilocode": ["anthropic/claude-opus-4.6", "anthropic/claude-sonnet-4.6", "openai/gpt-5.4", "google/gemini-3-pro-preview", "google/gemini-3-flash-preview"],
     "opencode-zen": ["gpt-5.4", "gpt-5.3-codex", "claude-sonnet-4-6", "gemini-3-flash", "glm-5", "kimi-k2.5", "minimax-m2.7"],
-    "opencode-go": ["kimi-k2.6", "kimi-k2.5", "glm-5.1", "glm-5", "mimo-v2-pro", "mimo-v2-omni", "minimax-m2.5", "minimax-m2.7", "qwen3.6-plus", "qwen3.5-plus"],
+    "opencode-go": ["kimi-k2.6", "kimi-k2.5", "glm-5.1", "glm-5", "mimo-v2.5-pro", "mimo-v2.5", "mimo-v2-pro", "mimo-v2-omni", "minimax-m2.7", "minimax-m2.5", "qwen3.6-plus", "qwen3.5-plus"],
     "huggingface": [
         "Qwen/Qwen3.5-397B-A17B", "Qwen/Qwen3-235B-A22B-Thinking-2507",
         "Qwen/Qwen3-Coder-480B-A35B-Instruct", "deepseek-ai/DeepSeek-R1-0528",
@@ -500,6 +500,15 @@ def _print_setup_summary(config: dict, hermes_home):
     if get_env_value("HASS_TOKEN"):
         tool_status.append(("Smart Home (Home Assistant)", True, None))
 
+    # Spotify (OAuth via hermes auth spotify — check auth.json, not env vars)
+    try:
+        from hermes_cli.auth import get_provider_auth_state
+        _spotify_state = get_provider_auth_state("spotify") or {}
+        if _spotify_state.get("access_token") or _spotify_state.get("refresh_token"):
+            tool_status.append(("Spotify (PKCE OAuth)", True, None))
+    except Exception:
+        pass
+
     # Skills Hub
     if get_env_value("GITHUB_TOKEN"):
         tool_status.append(("Skills Hub (GitHub)", True, None))
@@ -2334,6 +2343,7 @@ def setup_gateway(config: dict):
             launchd_install,
             launchd_start,
             launchd_restart,
+            UserSystemdUnavailableError,
         )
 
         service_installed = _is_service_installed()
@@ -2357,6 +2367,10 @@ def setup_gateway(config: dict):
                         systemd_restart()
                     elif _is_macos:
                         launchd_restart()
+                except UserSystemdUnavailableError as e:
+                    print_error("  Restart failed — user systemd not reachable:")
+                    for line in str(e).splitlines():
+                        print(f"  {line}")
                 except Exception as e:
                     print_error(f"  Restart failed: {e}")
         elif service_installed:
@@ -2366,6 +2380,10 @@ def setup_gateway(config: dict):
                         systemd_start()
                     elif _is_macos:
                         launchd_start()
+                except UserSystemdUnavailableError as e:
+                    print_error("  Start failed — user systemd not reachable:")
+                    for line in str(e).splitlines():
+                        print(f"  {line}")
                 except Exception as e:
                     print_error(f"  Start failed: {e}")
         elif supports_service_manager:
@@ -2389,6 +2407,10 @@ def setup_gateway(config: dict):
                                 systemd_start(system=installed_scope == "system")
                             elif _is_macos:
                                 launchd_start()
+                        except UserSystemdUnavailableError as e:
+                            print_error("  Start failed — user systemd not reachable:")
+                            for line in str(e).splitlines():
+                                print(f"  {line}")
                         except Exception as e:
                             print_error(f"  Start failed: {e}")
                 except Exception as e:

hermes_cli/status.py

@@ -164,19 +164,26 @@ def show_status(args):
         qwen_status = {}
 
     nous_logged_in = bool(nous_status.get("logged_in"))
+    nous_error = nous_status.get("error")
+    nous_label = "logged in" if nous_logged_in else "not logged in (run: hermes auth add nous --type oauth)"
     print(
         f"  {'Nous Portal':<12}  {check_mark(nous_logged_in)} "
-        f"{'logged in' if nous_logged_in else 'not logged in (run: hermes model)'}"
+        f"{nous_label}"
     )
-    if nous_logged_in:
-        portal_url = nous_status.get("portal_base_url") or "(unknown)"
-        access_exp = _format_iso_timestamp(nous_status.get("access_expires_at"))
-        key_exp = _format_iso_timestamp(nous_status.get("agent_key_expires_at"))
-        refresh_label = "yes" if nous_status.get("has_refresh_token") else "no"
+    portal_url = nous_status.get("portal_base_url") or "(unknown)"
+    access_exp = _format_iso_timestamp(nous_status.get("access_expires_at"))
+    key_exp = _format_iso_timestamp(nous_status.get("agent_key_expires_at"))
+    refresh_label = "yes" if nous_status.get("has_refresh_token") else "no"
+    if nous_logged_in or portal_url != "(unknown)" or nous_error:
         print(f"    Portal URL: {portal_url}")
+    if nous_logged_in or nous_status.get("access_expires_at"):
         print(f"    Access exp: {access_exp}")
+    if nous_logged_in or nous_status.get("agent_key_expires_at"):
         print(f"    Key exp:    {key_exp}")
+    if nous_logged_in or nous_status.get("has_refresh_token"):
         print(f"    Refresh:    {refresh_label}")
+    if nous_error and not nous_logged_in:
+        print(f"    Error:      {nous_error}")
 
     codex_logged_in = bool(codex_status.get("logged_in"))
     print(

hermes_cli/tips.py

@@ -127,7 +127,7 @@ TIPS = [
 
     # --- Tools & Capabilities ---
     "execute_code runs Python scripts that call Hermes tools programmatically — results stay out of context.",
-    "delegate_task spawns up to 3 concurrent sub-agents by default (configurable via delegation.max_concurrent_children) with isolated contexts for parallel work.",
+    "delegate_task spawns up to 3 concurrent sub-agents by default (delegation.max_concurrent_children) with isolated contexts for parallel work.",
     "web_extract works on PDF URLs — pass any PDF link and it converts to markdown.",
     "search_files is ripgrep-backed and faster than grep — use it instead of terminal grep.",
     "patch uses 9 fuzzy matching strategies so minor whitespace differences won't break edits.",
@@ -289,6 +289,7 @@ TIPS = [
     "When a provider returns HTTP 402 (payment required), the auxiliary client auto-falls back to the next one.",
     "agent.tool_use_enforcement steers models that describe actions instead of calling tools — auto for GPT/Codex.",
     "agent.restart_drain_timeout (default 60s) lets running agents finish before a gateway restart takes effect.",
+    "agent.api_max_retries (default 3) controls how many times the agent retries a failed API call before surfacing the error — lower it for fast fallback.",
     "The gateway caches AIAgent instances per session — destroying this cache breaks Anthropic prompt caching.",
     "Any website can expose skills via /.well-known/skills/index.json — the skills hub discovers them automatically.",
     "The skills audit log at ~/.hermes/skills/.hub/audit.log tracks every install and removal operation.",

hermes_cli/tools_config.py

@@ -67,12 +67,13 @@ CONFIGURABLE_TOOLSETS = [
     ("messaging",       "📨 Cross-Platform Messaging",  "send_message"),
     ("rl",              "🧪 RL Training",               "Tinker-Atropos training tools"),
     ("homeassistant",    "🏠 Home Assistant",           "smart home device control"),
+    ("spotify",          "🎵 Spotify",                  "playback, search, playlists, library"),
 ]
 
 # Toolsets that are OFF by default for new installs.
 # They're still in _HERMES_CORE_TOOLS (available at runtime if enabled),
 # but the setup checklist won't pre-select them for first-time users.
-_DEFAULT_OFF_TOOLSETS = {"moa", "homeassistant", "rl"}
+_DEFAULT_OFF_TOOLSETS = {"moa", "homeassistant", "rl", "spotify"}
 
 
 def _get_effective_configurable_toolsets():
@@ -361,6 +362,22 @@ TOOL_CATEGORIES = {
             },
         ],
     },
+    "spotify": {
+        "name": "Spotify",
+        "icon": "🎵",
+        "providers": [
+            {
+                "name": "Spotify Web API",
+                "tag": "PKCE OAuth — run `hermes auth spotify` after this",
+                "env_vars": [
+                    {"key": "HERMES_SPOTIFY_CLIENT_ID", "prompt": "Spotify app client_id",
+                     "url": "https://developer.spotify.com/dashboard"},
+                    {"key": "HERMES_SPOTIFY_REDIRECT_URI", "prompt": "Redirect URI (must be allow-listed in your Spotify app)",
+                     "default": "http://127.0.0.1:43827/spotify/callback"},
+                ],
+            },
+        ],
+    },
     "rl": {
         "name": "RL Training",
         "icon": "🧪",
@@ -590,7 +607,10 @@ def _get_platform_tools(
             default_off.remove(platform)
         enabled_toolsets -= default_off
 
-    # Plugin toolsets: enabled by default unless explicitly disabled.
+    # Plugin toolsets: enabled by default unless explicitly disabled, or
+    # unless the toolset is in _DEFAULT_OFF_TOOLSETS (e.g. spotify —
+    # shipped as a bundled plugin but user must opt in via `hermes tools`
+    # so we don't ship 7 Spotify tool schemas to users who don't use it).
     # A plugin toolset is "known" for a platform once `hermes tools`
     # has been saved for that platform (tracked via known_plugin_toolsets).
     # Unknown plugins default to enabled; known-but-absent = disabled.
@@ -602,6 +622,9 @@ def _get_platform_tools(
             if pts in toolset_names:
                 # Explicitly listed in config — enabled
                 enabled_toolsets.add(pts)
+            elif pts in _DEFAULT_OFF_TOOLSETS:
+                # Opt-in plugin toolset — stay off until user picks it
+                continue
             elif pts not in known_for_platform:
                 # New plugin not yet seen by hermes tools — default enabled
                 enabled_toolsets.add(pts)
@@ -1019,6 +1042,11 @@ def _configure_tool_category(ts_key: str, cat: dict, config: dict):
 
 def _is_provider_active(provider: dict, config: dict) -> bool:
     """Check if a provider entry matches the currently active config."""
+    plugin_name = provider.get("image_gen_plugin_name")
+    if plugin_name:
+        image_cfg = config.get("image_gen", {})
+        return isinstance(image_cfg, dict) and image_cfg.get("provider") == plugin_name
+
     managed_feature = provider.get("managed_nous_feature")
     if managed_feature:
         features = get_nous_subscription_features(config)
@@ -1026,6 +1054,13 @@ def _is_provider_active(provider: dict, config: dict) -> bool:
         if feature is None:
             return False
         if managed_feature == "image_gen":
+            image_cfg = config.get("image_gen", {})
+            if isinstance(image_cfg, dict):
+                configured_provider = image_cfg.get("provider")
+                if configured_provider not in (None, "", "fal"):
+                    return False
+                if image_cfg.get("use_gateway") is False:
+                    return False
             return feature.managed_by_nous
         if provider.get("tts_provider"):
             return (
@@ -1048,6 +1083,16 @@ def _is_provider_active(provider: dict, config: dict) -> bool:
     if provider.get("web_backend"):
         current = config.get("web", {}).get("backend")
         return current == provider["web_backend"]
+    if provider.get("imagegen_backend"):
+        image_cfg = config.get("image_gen", {})
+        if not isinstance(image_cfg, dict):
+            return False
+        configured_provider = image_cfg.get("provider")
+        return (
+            provider["imagegen_backend"] == "fal"
+            and configured_provider in (None, "", "fal")
+            and not image_cfg.get("use_gateway")
+        )
     return False
 
 
@@ -1245,6 +1290,18 @@ def _configure_imagegen_model_for_plugin(plugin_name: str, config: dict) -> None
     _print_success(f"  Model set to: {chosen}")
 
 
+def _select_plugin_image_gen_provider(plugin_name: str, config: dict) -> None:
+    """Persist a plugin-backed image generation provider selection."""
+    img_cfg = config.setdefault("image_gen", {})
+    if not isinstance(img_cfg, dict):
+        img_cfg = {}
+        config["image_gen"] = img_cfg
+    img_cfg["provider"] = plugin_name
+    img_cfg["use_gateway"] = False
+    _print_success(f"  image_gen.provider set to: {plugin_name}")
+    _configure_imagegen_model_for_plugin(plugin_name, config)
+
+
 def _configure_provider(provider: dict, config: dict):
     """Configure a single provider - prompt for API keys and set config."""
     env_vars = provider.get("env_vars", [])
@@ -1305,13 +1362,7 @@ def _configure_provider(provider: dict, config: dict):
         # and route model selection to the plugin's own catalog.
         plugin_name = provider.get("image_gen_plugin_name")
         if plugin_name:
-            img_cfg = config.setdefault("image_gen", {})
-            if not isinstance(img_cfg, dict):
-                img_cfg = {}
-                config["image_gen"] = img_cfg
-            img_cfg["provider"] = plugin_name
-            _print_success(f"  image_gen.provider set to: {plugin_name}")
-            _configure_imagegen_model_for_plugin(plugin_name, config)
+            _select_plugin_image_gen_provider(plugin_name, config)
             return
         # Imagegen backends prompt for model selection after backend pick.
         backend = provider.get("imagegen_backend")
@@ -1359,13 +1410,7 @@ def _configure_provider(provider: dict, config: dict):
         _print_success(f"  {provider['name']} configured!")
         plugin_name = provider.get("image_gen_plugin_name")
         if plugin_name:
-            img_cfg = config.setdefault("image_gen", {})
-            if not isinstance(img_cfg, dict):
-                img_cfg = {}
-                config["image_gen"] = img_cfg
-            img_cfg["provider"] = plugin_name
-            _print_success(f"  image_gen.provider set to: {plugin_name}")
-            _configure_imagegen_model_for_plugin(plugin_name, config)
+            _select_plugin_image_gen_provider(plugin_name, config)
             return
         # Imagegen backends prompt for model selection after env vars are in.
         backend = provider.get("imagegen_backend")
@@ -1539,16 +1584,39 @@ def _reconfigure_provider(provider: dict, config: dict):
         config.setdefault("web", {})["backend"] = provider["web_backend"]
         _print_success(f"  Web backend set to: {provider['web_backend']}")
 
+    if managed_feature and managed_feature not in ("web", "tts", "browser"):
+        section = config.setdefault(managed_feature, {})
+        if not isinstance(section, dict):
+            section = {}
+            config[managed_feature] = section
+        section["use_gateway"] = True
+    elif not managed_feature:
+        for cat_key, cat in TOOL_CATEGORIES.items():
+            if provider in cat.get("providers", []):
+                section = config.get(cat_key)
+                if isinstance(section, dict) and section.get("use_gateway"):
+                    section["use_gateway"] = False
+                break
+
     if not env_vars:
         if provider.get("post_setup"):
             _run_post_setup(provider["post_setup"])
         _print_success(f"  {provider['name']} - no configuration needed!")
         if managed_feature:
             _print_info("  Requests for this tool will be billed to your Nous subscription.")
+        plugin_name = provider.get("image_gen_plugin_name")
+        if plugin_name:
+            _select_plugin_image_gen_provider(plugin_name, config)
+            return
         # Imagegen backends prompt for model selection on reconfig too.
         backend = provider.get("imagegen_backend")
         if backend:
             _configure_imagegen_model(backend, config)
+            if backend == "fal":
+                img_cfg = config.setdefault("image_gen", {})
+                if isinstance(img_cfg, dict):
+                    img_cfg["provider"] = "fal"
+                    img_cfg["use_gateway"] = False
         return
 
     for var in env_vars:
@@ -1567,9 +1635,19 @@ def _reconfigure_provider(provider: dict, config: dict):
             _print_info("    Kept current")
 
     # Imagegen backends prompt for model selection on reconfig too.
+    plugin_name = provider.get("image_gen_plugin_name")
+    if plugin_name:
+        _select_plugin_image_gen_provider(plugin_name, config)
+        return
+
     backend = provider.get("imagegen_backend")
     if backend:
         _configure_imagegen_model(backend, config)
+        if backend == "fal":
+            img_cfg = config.setdefault("image_gen", {})
+            if isinstance(img_cfg, dict):
+                img_cfg["provider"] = "fal"
+                img_cfg["use_gateway"] = False
 
 
 def _reconfigure_simple_requirements(ts_key: str):

hermes_cli/voice.py

@@ -0,0 +1,548 @@
+"""Process-wide voice recording + TTS API for the TUI gateway.
+
+Wraps ``tools.voice_mode`` (recording/transcription) and ``tools.tts_tool``
+(text-to-speech) behind idempotent, stateful entry points that the gateway's
+``voice.record``, ``voice.toggle``, and ``voice.tts`` JSON-RPC handlers can
+call from a dedicated thread. The gateway imports this module lazily so that
+missing optional audio deps (sounddevice, faster-whisper, numpy) surface as
+an ``ImportError`` at call time, not at startup.
+
+Two usage modes are exposed:
+
+* **Push-to-talk** (``start_recording`` / ``stop_and_transcribe``) — single
+  manually-bounded capture used when the caller drives the start/stop pair
+  explicitly.
+* **Continuous (VAD)** (``start_continuous`` / ``stop_continuous``) — mirrors
+  the classic CLI voice mode: recording auto-stops on silence, transcribes,
+  hands the result to a callback, and then auto-restarts for the next turn.
+  Three consecutive no-speech cycles stop the loop and fire
+  ``on_silent_limit`` so the UI can turn the mode off.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import sys
+import threading
+from typing import Any, Callable, Optional
+
+from tools.voice_mode import (
+    create_audio_recorder,
+    is_whisper_hallucination,
+    play_audio_file,
+    transcribe_recording,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def _debug(msg: str) -> None:
+    """Emit a debug breadcrumb when HERMES_VOICE_DEBUG=1.
+
+    Goes to stderr so the TUI gateway wraps it as a gateway.stderr event,
+    which createGatewayEventHandler shows as an Activity line — exactly
+    what we need to diagnose "why didn't the loop auto-restart?" in the
+    user's real terminal without shipping a separate debug RPC.
+
+    Any OSError / BrokenPipeError is swallowed because this fires from
+    background threads (silence callback, TTS daemon, beep) where a
+    broken stderr pipe must not kill the whole gateway — the main
+    command pipe (stdin+stdout) is what actually matters.
+    """
+    if os.environ.get("HERMES_VOICE_DEBUG", "").strip() != "1":
+        return
+    try:
+        print(f"[voice] {msg}", file=sys.stderr, flush=True)
+    except (BrokenPipeError, OSError):
+        pass
+
+
+def _beeps_enabled() -> bool:
+    """CLI parity: voice.beep_enabled in config.yaml (default True)."""
+    try:
+        from hermes_cli.config import load_config
+
+        voice_cfg = load_config().get("voice", {})
+        if isinstance(voice_cfg, dict):
+            return bool(voice_cfg.get("beep_enabled", True))
+    except Exception:
+        pass
+    return True
+
+
+def _play_beep(frequency: int, count: int = 1) -> None:
+    """Audible cue matching cli.py's record/stop beeps.
+
+    880 Hz single-beep on start (cli.py:_voice_start_recording line 7532),
+    660 Hz double-beep on stop (cli.py:_voice_stop_and_transcribe line 7585).
+    Best-effort — sounddevice failures are silently swallowed so the
+    voice loop never breaks because a speaker was unavailable.
+    """
+    if not _beeps_enabled():
+        return
+    try:
+        from tools.voice_mode import play_beep
+
+        play_beep(frequency=frequency, count=count)
+    except Exception as e:
+        _debug(f"beep {frequency}Hz failed: {e}")
+
+# ── Push-to-talk state ───────────────────────────────────────────────
+_recorder = None
+_recorder_lock = threading.Lock()
+
+# ── Continuous (VAD) state ───────────────────────────────────────────
+_continuous_lock = threading.Lock()
+_continuous_active = False
+_continuous_recorder: Any = None
+
+# ── TTS-vs-STT feedback guard ────────────────────────────────────────
+# When TTS plays the agent reply over the speakers, the live microphone
+# picks it up and transcribes the agent's own voice as user input — an
+# infinite loop the agent happily joins ("Ha, looks like we're in a loop").
+# This Event mirrors cli.py:_voice_tts_done: cleared while speak_text is
+# playing, set while silent. _continuous_on_silence waits on it before
+# re-arming the recorder, and speak_text itself cancels any live capture
+# before starting playback so the tail of the previous utterance doesn't
+# leak into the mic.
+_tts_playing = threading.Event()
+_tts_playing.set()  # initially "not playing"
+_continuous_on_transcript: Optional[Callable[[str], None]] = None
+_continuous_on_status: Optional[Callable[[str], None]] = None
+_continuous_on_silent_limit: Optional[Callable[[], None]] = None
+_continuous_no_speech_count = 0
+_CONTINUOUS_NO_SPEECH_LIMIT = 3
+
+
+# ── Push-to-talk API ─────────────────────────────────────────────────
+
+
+def start_recording() -> None:
+    """Begin capturing from the default input device (push-to-talk).
+
+    Idempotent — calling again while a recording is in progress is a no-op.
+    """
+    global _recorder
+
+    with _recorder_lock:
+        if _recorder is not None and getattr(_recorder, "is_recording", False):
+            return
+        rec = create_audio_recorder()
+        rec.start()
+        _recorder = rec
+
+
+def stop_and_transcribe() -> Optional[str]:
+    """Stop the active push-to-talk recording, transcribe, return text.
+
+    Returns ``None`` when no recording is active, when the microphone
+    captured no speech, or when Whisper returned a known hallucination.
+    """
+    global _recorder
+
+    with _recorder_lock:
+        rec = _recorder
+        _recorder = None
+
+    if rec is None:
+        return None
+
+    wav_path = rec.stop()
+    if not wav_path:
+        return None
+
+    try:
+        result = transcribe_recording(wav_path)
+    except Exception as e:
+        logger.warning("voice transcription failed: %s", e)
+        return None
+    finally:
+        try:
+            if os.path.isfile(wav_path):
+                os.unlink(wav_path)
+        except Exception:
+            pass
+
+    # transcribe_recording returns {"success": bool, "transcript": str, ...}
+    # — matches cli.py:_voice_stop_and_transcribe's result.get("transcript").
+    if not result.get("success"):
+        return None
+    text = (result.get("transcript") or "").strip()
+    if not text or is_whisper_hallucination(text):
+        return None
+
+    return text
+
+
+# ── Continuous (VAD) API ─────────────────────────────────────────────
+
+
+def start_continuous(
+    on_transcript: Callable[[str], None],
+    on_status: Optional[Callable[[str], None]] = None,
+    on_silent_limit: Optional[Callable[[], None]] = None,
+    silence_threshold: int = 200,
+    silence_duration: float = 3.0,
+) -> None:
+    """Start a VAD-driven continuous recording loop.
+
+    The loop calls ``on_transcript(text)`` each time speech is detected and
+    transcribed successfully, then auto-restarts. After
+    ``_CONTINUOUS_NO_SPEECH_LIMIT`` consecutive silent cycles (no speech
+    picked up at all) the loop stops itself and calls ``on_silent_limit``
+    so the UI can reflect "voice off". Idempotent — calling while already
+    active is a no-op.
+
+    ``on_status`` is called with ``"listening"`` / ``"transcribing"`` /
+    ``"idle"`` so the UI can show a live indicator.
+    """
+    global _continuous_active, _continuous_recorder
+    global _continuous_on_transcript, _continuous_on_status, _continuous_on_silent_limit
+    global _continuous_no_speech_count
+
+    with _continuous_lock:
+        if _continuous_active:
+            _debug("start_continuous: already active — no-op")
+            return
+        _continuous_active = True
+        _continuous_on_transcript = on_transcript
+        _continuous_on_status = on_status
+        _continuous_on_silent_limit = on_silent_limit
+        _continuous_no_speech_count = 0
+
+        if _continuous_recorder is None:
+            _continuous_recorder = create_audio_recorder()
+
+        _continuous_recorder._silence_threshold = silence_threshold
+        _continuous_recorder._silence_duration = silence_duration
+        rec = _continuous_recorder
+
+    _debug(
+        f"start_continuous: begin (threshold={silence_threshold}, duration={silence_duration}s)"
+    )
+
+    # CLI parity: single 880 Hz beep *before* opening the stream — placing
+    # the beep after stream.start() on macOS triggers a CoreAudio conflict
+    # (cli.py:7528 comment).
+    _play_beep(frequency=880, count=1)
+
+    try:
+        rec.start(on_silence_stop=_continuous_on_silence)
+    except Exception as e:
+        logger.error("failed to start continuous recording: %s", e)
+        _debug(f"start_continuous: rec.start raised {type(e).__name__}: {e}")
+        with _continuous_lock:
+            _continuous_active = False
+        raise
+
+    if on_status:
+        try:
+            on_status("listening")
+        except Exception:
+            pass
+
+
+def stop_continuous() -> None:
+    """Stop the active continuous loop and release the microphone.
+
+    Idempotent — calling while not active is a no-op. Any in-flight
+    transcription completes but its result is discarded (the callback
+    checks ``_continuous_active`` before firing).
+    """
+    global _continuous_active, _continuous_on_transcript
+    global _continuous_on_status, _continuous_on_silent_limit
+    global _continuous_recorder, _continuous_no_speech_count
+
+    with _continuous_lock:
+        if not _continuous_active:
+            return
+        _continuous_active = False
+        rec = _continuous_recorder
+        on_status = _continuous_on_status
+        _continuous_on_transcript = None
+        _continuous_on_status = None
+        _continuous_on_silent_limit = None
+        _continuous_no_speech_count = 0
+
+    if rec is not None:
+        try:
+            # cancel() (not stop()) discards buffered frames — the loop
+            # is over, we don't want to transcribe a half-captured turn.
+            rec.cancel()
+        except Exception as e:
+            logger.warning("failed to cancel recorder: %s", e)
+
+    # Audible "recording stopped" cue (CLI parity: same 660 Hz × 2 the
+    # silence-auto-stop path plays).
+    _play_beep(frequency=660, count=2)
+
+    if on_status:
+        try:
+            on_status("idle")
+        except Exception:
+            pass
+
+
+def is_continuous_active() -> bool:
+    """Whether a continuous voice loop is currently running."""
+    with _continuous_lock:
+        return _continuous_active
+
+
+def _continuous_on_silence() -> None:
+    """AudioRecorder silence callback — runs in a daemon thread.
+
+    Stops the current capture, transcribes, delivers the text via
+    ``on_transcript``, and — if the loop is still active — starts the
+    next capture. Three consecutive silent cycles end the loop.
+    """
+    global _continuous_active, _continuous_no_speech_count
+
+    _debug("_continuous_on_silence: fired")
+
+    with _continuous_lock:
+        if not _continuous_active:
+            _debug("_continuous_on_silence: loop inactive — abort")
+            return
+        rec = _continuous_recorder
+        on_transcript = _continuous_on_transcript
+        on_status = _continuous_on_status
+        on_silent_limit = _continuous_on_silent_limit
+
+    if rec is None:
+        _debug("_continuous_on_silence: no recorder — abort")
+        return
+
+    if on_status:
+        try:
+            on_status("transcribing")
+        except Exception:
+            pass
+
+    wav_path = rec.stop()
+    # Peak RMS is the critical diagnostic when stop() returns None despite
+    # the VAD firing — tells us at a glance whether the mic was too quiet
+    # for SILENCE_RMS_THRESHOLD (200) or the VAD + peak checks disagree.
+    peak_rms = getattr(rec, "_peak_rms", -1)
+    _debug(
+        f"_continuous_on_silence: rec.stop -> {wav_path!r} (peak_rms={peak_rms})"
+    )
+
+    # CLI parity: double 660 Hz beep after the stream stops (safe from the
+    # CoreAudio conflict that blocks pre-start beeps).
+    _play_beep(frequency=660, count=2)
+
+    transcript: Optional[str] = None
+
+    if wav_path:
+        try:
+            result = transcribe_recording(wav_path)
+            # transcribe_recording returns {"success": bool, "transcript": str,
+            # "error": str?} — NOT {"text": str}.  Using the wrong key silently
+            # produced empty transcripts even when Groq/local STT returned fine,
+            # which masqueraded as "not hearing the user" to the caller.
+            success = bool(result.get("success"))
+            text = (result.get("transcript") or "").strip()
+            err = result.get("error")
+            _debug(
+                f"_continuous_on_silence: transcribe -> success={success} "
+                f"text={text!r} err={err!r}"
+            )
+            if success and text and not is_whisper_hallucination(text):
+                transcript = text
+        except Exception as e:
+            logger.warning("continuous transcription failed: %s", e)
+            _debug(f"_continuous_on_silence: transcribe raised {type(e).__name__}: {e}")
+        finally:
+            try:
+                if os.path.isfile(wav_path):
+                    os.unlink(wav_path)
+            except Exception:
+                pass
+
+    with _continuous_lock:
+        if not _continuous_active:
+            # User stopped us while we were transcribing — discard.
+            _debug("_continuous_on_silence: stopped during transcribe — no restart")
+            return
+        if transcript:
+            _continuous_no_speech_count = 0
+        else:
+            _continuous_no_speech_count += 1
+        should_halt = _continuous_no_speech_count >= _CONTINUOUS_NO_SPEECH_LIMIT
+        no_speech = _continuous_no_speech_count
+
+    if transcript and on_transcript:
+        try:
+            on_transcript(transcript)
+        except Exception as e:
+            logger.warning("on_transcript callback raised: %s", e)
+
+    if should_halt:
+        _debug(f"_continuous_on_silence: {no_speech} silent cycles — halting")
+        with _continuous_lock:
+            _continuous_active = False
+            _continuous_no_speech_count = 0
+        if on_silent_limit:
+            try:
+                on_silent_limit()
+            except Exception:
+                pass
+        try:
+            rec.cancel()
+        except Exception:
+            pass
+        if on_status:
+            try:
+                on_status("idle")
+            except Exception:
+                pass
+        return
+
+    # CLI parity (cli.py:10619-10621): wait for any in-flight TTS to
+    # finish before re-arming the mic, then leave a small gap to avoid
+    # catching the tail of the speaker output.  Without this the voice
+    # loop becomes a feedback loop — the agent's spoken reply lands
+    # back in the mic and gets re-submitted.
+    if not _tts_playing.is_set():
+        _debug("_continuous_on_silence: waiting for TTS to finish")
+        _tts_playing.wait(timeout=60)
+        import time as _time
+        _time.sleep(0.3)
+
+        # User may have stopped the loop during the wait.
+        with _continuous_lock:
+            if not _continuous_active:
+                _debug("_continuous_on_silence: stopped while waiting for TTS")
+                return
+
+    # Restart for the next turn.
+    _debug(f"_continuous_on_silence: restarting loop (no_speech={no_speech})")
+    _play_beep(frequency=880, count=1)
+    try:
+        rec.start(on_silence_stop=_continuous_on_silence)
+    except Exception as e:
+        logger.error("failed to restart continuous recording: %s", e)
+        _debug(f"_continuous_on_silence: restart raised {type(e).__name__}: {e}")
+        with _continuous_lock:
+            _continuous_active = False
+        return
+
+    if on_status:
+        try:
+            on_status("listening")
+        except Exception:
+            pass
+
+
+# ── TTS API ──────────────────────────────────────────────────────────
+
+
+def speak_text(text: str) -> None:
+    """Synthesize ``text`` with the configured TTS provider and play it.
+
+    Mirrors cli.py:_voice_speak_response exactly — same markdown strip
+    pipeline, same 4000-char cap, same explicit mp3 output path, same
+    MP3-over-OGG playback choice (afplay misbehaves on OGG), same cleanup
+    of both extensions. Keeping these in sync means a voice-mode TTS
+    session in the TUI sounds identical to one in the classic CLI.
+
+    While playback is in flight the module-level _tts_playing Event is
+    cleared so the continuous-recording loop knows to wait before
+    re-arming the mic (otherwise the agent's spoken reply feedback-loops
+    through the microphone and the agent ends up replying to itself).
+    """
+    if not text or not text.strip():
+        return
+
+    import re
+    import tempfile
+    import time
+
+    # Cancel any live capture before we open the speakers — otherwise the
+    # last ~200ms of the user's turn tail + the first syllables of our TTS
+    # both end up in the next recording window.  The continuous loop will
+    # re-arm itself after _tts_playing flips back (see _continuous_on_silence).
+    paused_recording = False
+    with _continuous_lock:
+        if (
+            _continuous_active
+            and _continuous_recorder is not None
+            and getattr(_continuous_recorder, "is_recording", False)
+        ):
+            try:
+                _continuous_recorder.cancel()
+                paused_recording = True
+            except Exception as e:
+                logger.warning("failed to pause recorder for TTS: %s", e)
+
+    _tts_playing.clear()
+    _debug(f"speak_text: TTS begin (paused_recording={paused_recording})")
+
+    try:
+        from tools.tts_tool import text_to_speech_tool
+
+        tts_text = text[:4000] if len(text) > 4000 else text
+        tts_text = re.sub(r'```[\s\S]*?```', ' ', tts_text)             # fenced code blocks
+        tts_text = re.sub(r'\[([^\]]+)\]\([^)]+\)', r'\1', tts_text)    # [text](url) → text
+        tts_text = re.sub(r'https?://\S+', '', tts_text)                # bare URLs
+        tts_text = re.sub(r'\*\*(.+?)\*\*', r'\1', tts_text)            # bold
+        tts_text = re.sub(r'\*(.+?)\*', r'\1', tts_text)                # italic
+        tts_text = re.sub(r'`(.+?)`', r'\1', tts_text)                  # inline code
+        tts_text = re.sub(r'^#+\s*', '', tts_text, flags=re.MULTILINE)  # headers
+        tts_text = re.sub(r'^\s*[-*]\s+', '', tts_text, flags=re.MULTILINE)  # list bullets
+        tts_text = re.sub(r'---+', '', tts_text)                        # horizontal rules
+        tts_text = re.sub(r'\n{3,}', '\n\n', tts_text)                  # excess newlines
+        tts_text = tts_text.strip()
+        if not tts_text:
+            return
+
+        # MP3 output path, pre-chosen so we can play the MP3 directly even
+        # when text_to_speech_tool auto-converts to OGG for messaging
+        # platforms.  afplay's OGG support is flaky, MP3 always works.
+        os.makedirs(os.path.join(tempfile.gettempdir(), "hermes_voice"), exist_ok=True)
+        mp3_path = os.path.join(
+            tempfile.gettempdir(),
+            "hermes_voice",
+            f"tts_{time.strftime('%Y%m%d_%H%M%S')}.mp3",
+        )
+
+        _debug(f"speak_text: synthesizing {len(tts_text)} chars -> {mp3_path}")
+        text_to_speech_tool(text=tts_text, output_path=mp3_path)
+
+        if os.path.isfile(mp3_path) and os.path.getsize(mp3_path) > 0:
+            _debug(f"speak_text: playing {mp3_path} ({os.path.getsize(mp3_path)} bytes)")
+            play_audio_file(mp3_path)
+            try:
+                os.unlink(mp3_path)
+                ogg_path = mp3_path.rsplit(".", 1)[0] + ".ogg"
+                if os.path.isfile(ogg_path):
+                    os.unlink(ogg_path)
+            except OSError:
+                pass
+        else:
+            _debug(f"speak_text: TTS tool produced no audio at {mp3_path}")
+    except Exception as e:
+        logger.warning("Voice TTS playback failed: %s", e)
+        _debug(f"speak_text raised {type(e).__name__}: {e}")
+    finally:
+        _tts_playing.set()
+        _debug("speak_text: TTS done")
+
+        # Re-arm the mic so the user can answer without pressing Ctrl+B.
+        # Small delay lets the OS flush speaker output and afplay fully
+        # release the audio device before sounddevice re-opens the input.
+        if paused_recording:
+            time.sleep(0.3)
+            with _continuous_lock:
+                if _continuous_active and _continuous_recorder is not None:
+                    try:
+                        _continuous_recorder.start(
+                            on_silence_stop=_continuous_on_silence
+                        )
+                        _debug("speak_text: recording resumed after TTS")
+                    except Exception as e:
+                        logger.warning(
+                            "failed to resume recorder after TTS: %s", e
+                        )

hermes_cli/web_server.py

@@ -49,7 +49,7 @@ from hermes_cli.config import (
 from gateway.status import get_running_pid, read_runtime_status
 
 try:
-    from fastapi import FastAPI, HTTPException, Request, WebSocket, WebSocketDisconnect
+    from fastapi import FastAPI, HTTPException, Request
     from fastapi.middleware.cors import CORSMiddleware
     from fastapi.responses import FileResponse, HTMLResponse, JSONResponse
     from fastapi.staticfiles import StaticFiles
@@ -71,6 +71,7 @@ app = FastAPI(title="Hermes Agent", version=__version__)
 # Injected into the SPA HTML so only the legitimate web UI can use it.
 # ---------------------------------------------------------------------------
 _SESSION_TOKEN = secrets.token_urlsafe(32)
+_SESSION_HEADER_NAME = "X-Hermes-Session-Token"
 
 # Simple rate limiter for the reveal endpoint
 _reveal_timestamps: List[float] = []
@@ -104,14 +105,29 @@ _PUBLIC_API_PATHS: frozenset = frozenset({
 })
 
 
-def _require_token(request: Request) -> None:
-    """Validate the ephemeral session token.  Raises 401 on mismatch.
+def _has_valid_session_token(request: Request) -> bool:
+    """True if the request carries a valid dashboard session token.
 
-    Uses ``hmac.compare_digest`` to prevent timing side-channels.
+    The dedicated session header avoids collisions with reverse proxies that
+    already use ``Authorization`` (for example Caddy ``basic_auth``). We still
+    accept the legacy Bearer path for backward compatibility with older
+    dashboard bundles.
     """
+    session_header = request.headers.get(_SESSION_HEADER_NAME, "")
+    if session_header and hmac.compare_digest(
+        session_header.encode(),
+        _SESSION_TOKEN.encode(),
+    ):
+        return True
+
     auth = request.headers.get("authorization", "")
     expected = f"Bearer {_SESSION_TOKEN}"
-    if not hmac.compare_digest(auth.encode(), expected.encode()):
+    return hmac.compare_digest(auth.encode(), expected.encode())
+
+
+def _require_token(request: Request) -> None:
+    """Validate the ephemeral session token.  Raises 401 on mismatch."""
+    if not _has_valid_session_token(request):
         raise HTTPException(status_code=401, detail="Unauthorized")
 
 
@@ -205,9 +221,7 @@ async def auth_middleware(request: Request, call_next):
     """Require the session token on all /api/ routes except the public list."""
     path = request.url.path
     if path.startswith("/api/") and path not in _PUBLIC_API_PATHS and not path.startswith("/api/plugins/"):
-        auth = request.headers.get("authorization", "")
-        expected = f"Bearer {_SESSION_TOKEN}"
-        if not hmac.compare_digest(auth.encode(), expected.encode()):
+        if not _has_valid_session_token(request):
             return JSONResponse(
                 status_code=401,
                 content={"detail": "Unauthorized"},
@@ -417,7 +431,14 @@ class EnvVarReveal(BaseModel):
 
 
 _GATEWAY_HEALTH_URL = os.getenv("GATEWAY_HEALTH_URL")
-_GATEWAY_HEALTH_TIMEOUT = float(os.getenv("GATEWAY_HEALTH_TIMEOUT", "3"))
+try:
+    _GATEWAY_HEALTH_TIMEOUT = float(os.getenv("GATEWAY_HEALTH_TIMEOUT", "3"))
+except (ValueError, TypeError):
+    _log.warning(
+        "Invalid GATEWAY_HEALTH_TIMEOUT value %r — using default 3.0s",
+        os.getenv("GATEWAY_HEALTH_TIMEOUT"),
+    )
+    _GATEWAY_HEALTH_TIMEOUT = 3.0
 
 
 def _probe_gateway_health() -> tuple[bool, dict | None]:
@@ -2242,148 +2263,6 @@ async def get_usage_analytics(days: int = 30):
         db.close()
 
 
-# ---------------------------------------------------------------------------
-# /api/pty — PTY-over-WebSocket bridge for the dashboard "Chat" tab.
-#
-# The endpoint spawns the same ``hermes --tui`` binary the CLI uses, behind
-# a POSIX pseudo-terminal, and forwards bytes + resize escapes across a
-# WebSocket.  The browser renders the ANSI through xterm.js (see
-# web/src/pages/ChatPage.tsx).
-#
-# Auth: ``?token=<session_token>`` query param (browsers can't set
-# Authorization on the WS upgrade).  Same ephemeral ``_SESSION_TOKEN`` as
-# REST.  Localhost-only — we defensively reject non-loopback clients even
-# though uvicorn binds to 127.0.0.1.
-# ---------------------------------------------------------------------------
-
-import re
-import asyncio
-
-from hermes_cli.pty_bridge import PtyBridge, PtyUnavailableError
-
-_RESIZE_RE = re.compile(rb"\x1b\[RESIZE:(\d+);(\d+)\]")
-_PTY_READ_CHUNK_TIMEOUT = 0.2
-# Starlette's TestClient reports the peer as "testclient"; treat it as
-# loopback so tests don't need to rewrite request scope.
-_LOOPBACK_HOSTS = frozenset({"127.0.0.1", "::1", "localhost", "testclient"})
-
-
-def _resolve_chat_argv(
-    resume: Optional[str] = None,
-) -> tuple[list[str], Optional[str], Optional[dict]]:
-    """Resolve the argv + cwd + env for the chat PTY.
-
-    Default: whatever ``hermes --tui`` would run.  Tests monkeypatch this
-    function to inject a tiny fake command (``cat``, ``sh -c 'printf …'``)
-    so nothing has to build Node or the TUI bundle.
-
-    Session resume is propagated via the ``HERMES_TUI_RESUME`` env var —
-    matching what ``hermes_cli.main._launch_tui`` does for the CLI path.
-    Appending ``--resume <id>`` to argv doesn't work because ``ui-tui`` does
-    not parse its argv.
-    """
-    from hermes_cli.main import PROJECT_ROOT, _make_tui_argv
-
-    argv, cwd = _make_tui_argv(PROJECT_ROOT / "ui-tui", tui_dev=False)
-    env: Optional[dict] = None
-    if resume:
-        env = os.environ.copy()
-        env["HERMES_TUI_RESUME"] = resume
-    return list(argv), str(cwd) if cwd else None, env
-
-
-@app.websocket("/api/pty")
-async def pty_ws(ws: WebSocket) -> None:
-    # --- auth + loopback check (before accept so we can close cleanly) ---
-    token = ws.query_params.get("token", "")
-    expected = _SESSION_TOKEN
-    if not hmac.compare_digest(token.encode(), expected.encode()):
-        await ws.close(code=4401)
-        return
-
-    client_host = ws.client.host if ws.client else ""
-    if client_host and client_host not in _LOOPBACK_HOSTS:
-        await ws.close(code=4403)
-        return
-
-    await ws.accept()
-
-    # --- spawn PTY ------------------------------------------------------
-    resume = ws.query_params.get("resume") or None
-    try:
-        argv, cwd, env = _resolve_chat_argv(resume=resume)
-    except SystemExit as exc:
-        # _make_tui_argv calls sys.exit(1) when node/npm is missing.
-        await ws.send_text(f"\r\n\x1b[31mChat unavailable: {exc}\x1b[0m\r\n")
-        await ws.close(code=1011)
-        return
-
-
-    try:
-        bridge = PtyBridge.spawn(argv, cwd=cwd, env=env)
-    except PtyUnavailableError as exc:
-        await ws.send_text(f"\r\n\x1b[31mChat unavailable: {exc}\x1b[0m\r\n")
-        await ws.close(code=1011)
-        return
-    except (FileNotFoundError, OSError) as exc:
-        await ws.send_text(f"\r\n\x1b[31mChat failed to start: {exc}\x1b[0m\r\n")
-        await ws.close(code=1011)
-        return
-
-    loop = asyncio.get_running_loop()
-
-    # --- reader task: PTY master → WebSocket ----------------------------
-    async def pump_pty_to_ws() -> None:
-        while True:
-            chunk = await loop.run_in_executor(
-                None, bridge.read, _PTY_READ_CHUNK_TIMEOUT
-            )
-            if chunk is None:  # EOF
-                return
-            if not chunk:  # no data this tick; yield control and retry
-                await asyncio.sleep(0)
-                continue
-            try:
-                await ws.send_bytes(chunk)
-            except Exception:
-                return
-
-    reader_task = asyncio.create_task(pump_pty_to_ws())
-
-    # --- writer loop: WebSocket → PTY master ----------------------------
-    try:
-        while True:
-            msg = await ws.receive()
-            msg_type = msg.get("type")
-            if msg_type == "websocket.disconnect":
-                break
-            raw = msg.get("bytes")
-            if raw is None:
-                text = msg.get("text")
-                raw = text.encode("utf-8") if isinstance(text, str) else b""
-            if not raw:
-                continue
-
-            # Resize escape is consumed locally, never written to the PTY.
-            match = _RESIZE_RE.match(raw)
-            if match and match.end() == len(raw):
-                cols = int(match.group(1))
-                rows = int(match.group(2))
-                bridge.resize(cols=cols, rows=rows)
-                continue
-
-            bridge.write(raw)
-    except WebSocketDisconnect:
-        pass
-    finally:
-        reader_task.cancel()
-        try:
-            await reader_task
-        except (asyncio.CancelledError, Exception):
-            pass
-        bridge.close()
-
-
 def mount_spa(application: FastAPI):
     """Mount the built SPA. Falls back to index.html for client-side routing.
 
@@ -2446,8 +2325,227 @@ _BUILTIN_DASHBOARD_THEMES = [
 ]
 
 
+def _parse_theme_layer(value: Any, default_hex: str, default_alpha: float = 1.0) -> Optional[Dict[str, Any]]:
+    """Normalise a theme layer spec from YAML into `{hex, alpha}` form.
+
+    Accepts shorthand (a bare hex string) or full dict form.  Returns
+    ``None`` on garbage input so the caller can fall back to a built-in
+    default rather than blowing up.
+    """
+    if value is None:
+        return {"hex": default_hex, "alpha": default_alpha}
+    if isinstance(value, str):
+        return {"hex": value, "alpha": default_alpha}
+    if isinstance(value, dict):
+        hex_val = value.get("hex", default_hex)
+        alpha_val = value.get("alpha", default_alpha)
+        if not isinstance(hex_val, str):
+            return None
+        try:
+            alpha_f = float(alpha_val)
+        except (TypeError, ValueError):
+            alpha_f = default_alpha
+        return {"hex": hex_val, "alpha": max(0.0, min(1.0, alpha_f))}
+    return None
+
+
+_THEME_DEFAULT_TYPOGRAPHY: Dict[str, str] = {
+    "fontSans": 'system-ui, -apple-system, "Segoe UI", Roboto, "Helvetica Neue", Arial, sans-serif',
+    "fontMono": 'ui-monospace, "SF Mono", "Cascadia Mono", Menlo, Consolas, monospace',
+    "baseSize": "15px",
+    "lineHeight": "1.55",
+    "letterSpacing": "0",
+}
+
+_THEME_DEFAULT_LAYOUT: Dict[str, str] = {
+    "radius": "0.5rem",
+    "density": "comfortable",
+}
+
+_THEME_OVERRIDE_KEYS = {
+    "card", "cardForeground", "popover", "popoverForeground",
+    "primary", "primaryForeground", "secondary", "secondaryForeground",
+    "muted", "mutedForeground", "accent", "accentForeground",
+    "destructive", "destructiveForeground", "success", "warning",
+    "border", "input", "ring",
+}
+
+# Well-known named asset slots themes can populate.  Any other keys under
+# ``assets.custom`` are exposed as ``--theme-asset-custom-<key>`` CSS vars
+# for plugin/shell use.
+_THEME_NAMED_ASSET_KEYS = {"bg", "hero", "logo", "crest", "sidebar", "header"}
+
+# Component-style buckets themes can override.  The value under each bucket
+# is a mapping from camelCase property name to CSS string; each pair emits
+# ``--component-<bucket>-<kebab-property>`` on :root.  The frontend's shell
+# components (Card, App header, Backdrop, etc.) consume these vars so themes
+# can restyle chrome (clip-path, border-image, segmented progress, etc.)
+# without shipping their own CSS.
+_THEME_COMPONENT_BUCKETS = {
+    "card", "header", "footer", "sidebar", "tab",
+    "progress", "badge", "backdrop", "page",
+}
+
+_THEME_LAYOUT_VARIANTS = {"standard", "cockpit", "tiled"}
+
+# Cap on customCSS length so a malformed/oversized theme YAML can't blow up
+# the response payload or the <style> tag.  32 KiB is plenty for every
+# practical reskin (the Strike Freedom demo is ~2 KiB).
+_THEME_CUSTOM_CSS_MAX = 32 * 1024
+
+
+def _normalise_theme_definition(data: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+    """Normalise a user theme YAML into the wire format `ThemeProvider`
+    expects.  Returns ``None`` if the theme is unusable.
+
+    Accepts both the full schema (palette/typography/layout) and a loose
+    form with bare hex strings, so hand-written YAMLs stay friendly.
+    """
+    if not isinstance(data, dict):
+        return None
+    name = data.get("name")
+    if not isinstance(name, str) or not name.strip():
+        return None
+
+    # Palette
+    palette_src = data.get("palette", {}) if isinstance(data.get("palette"), dict) else {}
+    # Allow top-level `colors.background` as a shorthand too.
+    colors_src = data.get("colors", {}) if isinstance(data.get("colors"), dict) else {}
+
+    def _layer(key: str, default_hex: str, default_alpha: float = 1.0) -> Dict[str, Any]:
+        spec = palette_src.get(key, colors_src.get(key))
+        parsed = _parse_theme_layer(spec, default_hex, default_alpha)
+        return parsed if parsed is not None else {"hex": default_hex, "alpha": default_alpha}
+
+    palette = {
+        "background": _layer("background", "#041c1c", 1.0),
+        "midground": _layer("midground", "#ffe6cb", 1.0),
+        "foreground": _layer("foreground", "#ffffff", 0.0),
+        "warmGlow": palette_src.get("warmGlow") or data.get("warmGlow") or "rgba(255, 189, 56, 0.35)",
+        "noiseOpacity": 1.0,
+    }
+    raw_noise = palette_src.get("noiseOpacity", data.get("noiseOpacity"))
+    try:
+        palette["noiseOpacity"] = float(raw_noise) if raw_noise is not None else 1.0
+    except (TypeError, ValueError):
+        palette["noiseOpacity"] = 1.0
+
+    # Typography
+    typo_src = data.get("typography", {}) if isinstance(data.get("typography"), dict) else {}
+    typography = dict(_THEME_DEFAULT_TYPOGRAPHY)
+    for key in ("fontSans", "fontMono", "fontDisplay", "fontUrl", "baseSize", "lineHeight", "letterSpacing"):
+        val = typo_src.get(key)
+        if isinstance(val, str) and val.strip():
+            typography[key] = val
+
+    # Layout
+    layout_src = data.get("layout", {}) if isinstance(data.get("layout"), dict) else {}
+    layout = dict(_THEME_DEFAULT_LAYOUT)
+    radius = layout_src.get("radius")
+    if isinstance(radius, str) and radius.strip():
+        layout["radius"] = radius
+    density = layout_src.get("density")
+    if isinstance(density, str) and density in ("compact", "comfortable", "spacious"):
+        layout["density"] = density
+
+    # Color overrides — keep only valid keys with string values.
+    overrides_src = data.get("colorOverrides", {})
+    color_overrides: Dict[str, str] = {}
+    if isinstance(overrides_src, dict):
+        for key, val in overrides_src.items():
+            if key in _THEME_OVERRIDE_KEYS and isinstance(val, str) and val.strip():
+                color_overrides[key] = val
+
+    # Assets — named slots + arbitrary user-defined keys.  Values must be
+    # strings (URLs or CSS ``url(...)``/``linear-gradient(...)`` expressions).
+    # We don't fetch remote assets here; the frontend just injects them as
+    # CSS vars.  Empty values are dropped so a theme can explicitly clear a
+    # slot by setting ``hero: ""``.
+    assets_out: Dict[str, Any] = {}
+    assets_src = data.get("assets", {}) if isinstance(data.get("assets"), dict) else {}
+    for key in _THEME_NAMED_ASSET_KEYS:
+        val = assets_src.get(key)
+        if isinstance(val, str) and val.strip():
+            assets_out[key] = val
+    custom_assets_src = assets_src.get("custom")
+    if isinstance(custom_assets_src, dict):
+        custom_assets: Dict[str, str] = {}
+        for key, val in custom_assets_src.items():
+            if (
+                isinstance(key, str)
+                and key.replace("-", "").replace("_", "").isalnum()
+                and isinstance(val, str)
+                and val.strip()
+            ):
+                custom_assets[key] = val
+        if custom_assets:
+            assets_out["custom"] = custom_assets
+
+    # Custom CSS — raw CSS text the frontend injects as a scoped <style>
+    # tag on theme apply.  Clipped to _THEME_CUSTOM_CSS_MAX to keep the
+    # payload bounded.  We intentionally do NOT parse/sanitise the CSS
+    # here — the dashboard is localhost-only and themes are user-authored
+    # YAML in ~/.hermes/, same trust level as the config file itself.
+    custom_css_val = data.get("customCSS")
+    custom_css: Optional[str] = None
+    if isinstance(custom_css_val, str) and custom_css_val.strip():
+        custom_css = custom_css_val[:_THEME_CUSTOM_CSS_MAX]
+
+    # Component style overrides — per-bucket dicts of camelCase CSS
+    # property -> CSS string.  The frontend converts these into CSS vars
+    # that shell components (Card, App header, Backdrop) consume.
+    component_styles_src = data.get("componentStyles", {})
+    component_styles: Dict[str, Dict[str, str]] = {}
+    if isinstance(component_styles_src, dict):
+        for bucket, props in component_styles_src.items():
+            if bucket not in _THEME_COMPONENT_BUCKETS or not isinstance(props, dict):
+                continue
+            clean: Dict[str, str] = {}
+            for prop, value in props.items():
+                if (
+                    isinstance(prop, str)
+                    and prop.replace("-", "").replace("_", "").isalnum()
+                    and isinstance(value, (str, int, float))
+                    and str(value).strip()
+                ):
+                    clean[prop] = str(value)
+            if clean:
+                component_styles[bucket] = clean
+
+    layout_variant_src = data.get("layoutVariant")
+    layout_variant = (
+        layout_variant_src
+        if isinstance(layout_variant_src, str) and layout_variant_src in _THEME_LAYOUT_VARIANTS
+        else "standard"
+    )
+
+    result: Dict[str, Any] = {
+        "name": name,
+        "label": data.get("label") or name,
+        "description": data.get("description", ""),
+        "palette": palette,
+        "typography": typography,
+        "layout": layout,
+        "layoutVariant": layout_variant,
+    }
+    if color_overrides:
+        result["colorOverrides"] = color_overrides
+    if assets_out:
+        result["assets"] = assets_out
+    if custom_css is not None:
+        result["customCSS"] = custom_css
+    if component_styles:
+        result["componentStyles"] = component_styles
+    return result
+
+
 def _discover_user_themes() -> list:
-    """Scan ~/.hermes/dashboard-themes/*.yaml for user-created themes."""
+    """Scan ~/.hermes/dashboard-themes/*.yaml for user-created themes.
+
+    Returns a list of fully-normalised theme definitions ready to ship
+    to the frontend, so the client can apply them without a secondary
+    round-trip or a built-in stub.
+    """
     themes_dir = get_hermes_home() / "dashboard-themes"
     if not themes_dir.is_dir():
         return []
@@ -2455,33 +2553,42 @@ def _discover_user_themes() -> list:
     for f in sorted(themes_dir.glob("*.yaml")):
         try:
             data = yaml.safe_load(f.read_text(encoding="utf-8"))
-            if isinstance(data, dict) and data.get("name"):
-                result.append({
-                    "name": data["name"],
-                    "label": data.get("label", data["name"]),
-                    "description": data.get("description", ""),
-                })
         except Exception:
             continue
+        normalised = _normalise_theme_definition(data)
+        if normalised is not None:
+            result.append(normalised)
     return result
 
 
 @app.get("/api/dashboard/themes")
 async def get_dashboard_themes():
-    """Return available themes and the currently active one."""
+    """Return available themes and the currently active one.
+
+    Built-in entries ship name/label/description only (the frontend owns
+    their full definitions in `web/src/themes/presets.ts`).  User themes
+    from `~/.hermes/dashboard-themes/*.yaml` ship with their full
+    normalised definition under `definition`, so the client can apply
+    them without a stub.
+    """
     config = load_config()
     active = config.get("dashboard", {}).get("theme", "default")
     user_themes = _discover_user_themes()
-    # Merge built-in + user, user themes override built-in by name.
     seen = set()
     themes = []
     for t in _BUILTIN_DASHBOARD_THEMES:
         seen.add(t["name"])
         themes.append(t)
     for t in user_themes:
-        if t["name"] not in seen:
-            themes.append(t)
-            seen.add(t["name"])
+        if t["name"] in seen:
+            continue
+        themes.append({
+            "name": t["name"],
+            "label": t["label"],
+            "description": t["description"],
+            "definition": t,
+        })
+        seen.add(t["name"])
     return {"themes": themes, "active": active}
 
 
@@ -2538,13 +2645,35 @@ def _discover_dashboard_plugins() -> list:
                 if name in seen_names:
                     continue
                 seen_names.add(name)
+                # Tab options: ``path`` + ``position`` for a new tab, optional
+                # ``override`` to replace a built-in route, and ``hidden`` to
+                # register the plugin component/slots without adding a tab
+                # (useful for slot-only plugins like a header-crest injector).
+                raw_tab = data.get("tab", {}) if isinstance(data.get("tab"), dict) else {}
+                tab_info = {
+                    "path": raw_tab.get("path", f"/{name}"),
+                    "position": raw_tab.get("position", "end"),
+                }
+                override_path = raw_tab.get("override")
+                if isinstance(override_path, str) and override_path.startswith("/"):
+                    tab_info["override"] = override_path
+                if bool(raw_tab.get("hidden")):
+                    tab_info["hidden"] = True
+                # Slots: list of named slot locations this plugin populates.
+                # The frontend exposes ``registerSlot(pluginName, slotName, Component)``
+                # on window; plugins with non-empty slots call it from their JS bundle.
+                slots_src = data.get("slots")
+                slots: List[str] = []
+                if isinstance(slots_src, list):
+                    slots = [s for s in slots_src if isinstance(s, str) and s]
                 plugins.append({
                     "name": name,
                     "label": data.get("label", name),
                     "description": data.get("description", ""),
                     "icon": data.get("icon", "Puzzle"),
                     "version": data.get("version", "0.0.0"),
-                    "tab": data.get("tab", {"path": f"/{name}", "position": "end"}),
+                    "tab": tab_info,
+                    "slots": slots,
                     "entry": data.get("entry", "dist/index.js"),
                     "css": data.get("css"),
                     "has_api": bool(data.get("api")),

hermes_state.py

@@ -1039,6 +1039,71 @@ class SessionDB:
             result.append(msg)
         return result
 
+    def resolve_resume_session_id(self, session_id: str) -> str:
+        """Redirect a resume target to the descendant session that holds the messages.
+
+        Context compression ends the current session and forks a new child session
+        (linked via ``parent_session_id``). The flush cursor is reset, so the
+        child is where new messages actually land — the parent ends up with
+        ``message_count = 0`` rows unless messages had already been flushed to
+        it before compression. See #15000.
+
+        This helper walks ``parent_session_id`` forward from ``session_id`` and
+        returns the first descendant in the chain that has at least one message
+        row. If the original session already has messages, or no descendant
+        has any, the original ``session_id`` is returned unchanged.
+
+        The chain is always walked via the child whose ``started_at`` is
+        latest; that matches the single-chain shape that compression creates.
+        A depth cap (32) guards against accidental loops in malformed data.
+        """
+        if not session_id:
+            return session_id
+
+        with self._lock:
+            # If this session already has messages, nothing to redirect.
+            try:
+                row = self._conn.execute(
+                    "SELECT 1 FROM messages WHERE session_id = ? LIMIT 1",
+                    (session_id,),
+                ).fetchone()
+            except Exception:
+                return session_id
+            if row is not None:
+                return session_id
+
+            # Walk descendants: at each step, pick the most-recently-started
+                # child session; stop once we find one with messages.
+            current = session_id
+            seen = {current}
+            for _ in range(32):
+                try:
+                    child_row = self._conn.execute(
+                        "SELECT id FROM sessions "
+                        "WHERE parent_session_id = ? "
+                        "ORDER BY started_at DESC, id DESC LIMIT 1",
+                        (current,),
+                    ).fetchone()
+                except Exception:
+                    return session_id
+                if child_row is None:
+                    return session_id
+                child_id = child_row["id"] if hasattr(child_row, "keys") else child_row[0]
+                if not child_id or child_id in seen:
+                    return session_id
+                seen.add(child_id)
+                try:
+                    msg_row = self._conn.execute(
+                        "SELECT 1 FROM messages WHERE session_id = ? LIMIT 1",
+                        (child_id,),
+                    ).fetchone()
+                except Exception:
+                    return session_id
+                if msg_row is not None:
+                    return child_id
+                current = child_id
+        return session_id
+
     def get_messages_as_conversation(self, session_id: str) -> List[Dict[str, Any]]:
         """
         Load messages in the OpenAI conversation format (role + content dicts).

model_tools.py

@@ -343,6 +343,18 @@ def get_tool_definitions(
     global _last_resolved_tool_names
     _last_resolved_tool_names = [t["function"]["name"] for t in filtered_tools]
 
+    # Sanitize schemas for broad backend compatibility. llama.cpp's
+    # json-schema-to-grammar converter (used by its OAI server to build
+    # GBNF tool-call parsers) rejects some shapes that cloud providers
+    # silently accept — bare "type": "object" with no properties,
+    # string-valued schema nodes from malformed MCP servers, etc. This
+    # is a no-op for schemas that are already well-formed.
+    try:
+        from tools.schema_sanitizer import sanitize_tool_schemas
+        filtered_tools = sanitize_tool_schemas(filtered_tools)
+    except Exception as e:  # pragma: no cover — defensive
+        logger.warning("Schema sanitization skipped: %s", e)
+
     return filtered_tools
 
 
@@ -418,6 +430,31 @@ def _coerce_value(value: str, expected_type):
         return _coerce_number(value, integer_only=(expected_type == "integer"))
     if expected_type == "boolean":
         return _coerce_boolean(value)
+    if expected_type == "array":
+        return _coerce_json(value, list)
+    if expected_type == "object":
+        return _coerce_json(value, dict)
+    return value
+
+
+def _coerce_json(value: str, expected_python_type: type):
+    """Parse *value* as JSON when the schema expects an array or object.
+
+    Handles model output drift where a complex oneOf/discriminated-union schema
+    causes the LLM to emit the array/object as a JSON string instead of a native
+    structure.  Returns the original string if parsing fails or yields the wrong
+    Python type.
+    """
+    try:
+        parsed = json.loads(value)
+    except (ValueError, TypeError):
+        return value
+    if isinstance(parsed, expected_python_type):
+        logger.debug(
+            "coerce_tool_args: coerced string to %s via json.loads",
+            expected_python_type.__name__,
+        )
+        return parsed
     return value
 
 

nix/nixosModules.nix

@@ -777,7 +777,10 @@ HERMES_NIX_ENV_EOF
             NoNewPrivileges = true;
             ProtectSystem = "strict";
             ProtectHome = false;
-            ReadWritePaths = [ cfg.stateDir ];
+            ReadWritePaths = [
+              cfg.stateDir
+              cfg.workingDirectory
+            ];
             PrivateTmp = true;
           };
 

plugins/image_gen/openai-codex/__init__.py

@@ -0,0 +1,378 @@
+"""OpenAI image generation backend — ChatGPT/Codex OAuth variant.
+
+Identical model catalog and tier semantics to the ``openai`` image-gen plugin
+(``gpt-image-2`` at low/medium/high quality), but routes the request through
+the Codex Responses API ``image_generation`` tool instead of the
+``images.generate`` REST endpoint. This lets users who are already
+authenticated with Codex/ChatGPT generate images without configuring a
+separate ``OPENAI_API_KEY``.
+
+Selection precedence for the tier (first hit wins):
+
+1. ``OPENAI_IMAGE_MODEL`` env var (escape hatch for scripts / tests)
+2. ``image_gen.openai-codex.model`` in ``config.yaml``
+3. ``image_gen.model`` in ``config.yaml`` (when it's one of our tier IDs)
+4. :data:`DEFAULT_MODEL` — ``gpt-image-2-medium``
+
+Output is saved as PNG under ``$HERMES_HOME/cache/images/``.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Any, Dict, List, Optional, Tuple
+
+from agent.image_gen_provider import (
+    DEFAULT_ASPECT_RATIO,
+    ImageGenProvider,
+    error_response,
+    resolve_aspect_ratio,
+    save_b64_image,
+    success_response,
+)
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Model catalog — mirrors the ``openai`` plugin so the picker UX is identical.
+# ---------------------------------------------------------------------------
+
+API_MODEL = "gpt-image-2"
+
+_MODELS: Dict[str, Dict[str, Any]] = {
+    "gpt-image-2-low": {
+        "display": "GPT Image 2 (Low)",
+        "speed": "~15s",
+        "strengths": "Fast iteration, lowest cost",
+        "quality": "low",
+    },
+    "gpt-image-2-medium": {
+        "display": "GPT Image 2 (Medium)",
+        "speed": "~40s",
+        "strengths": "Balanced — default",
+        "quality": "medium",
+    },
+    "gpt-image-2-high": {
+        "display": "GPT Image 2 (High)",
+        "speed": "~2min",
+        "strengths": "Highest fidelity, strongest prompt adherence",
+        "quality": "high",
+    },
+}
+
+DEFAULT_MODEL = "gpt-image-2-medium"
+
+_SIZES = {
+    "landscape": "1536x1024",
+    "square": "1024x1024",
+    "portrait": "1024x1536",
+}
+
+# Codex Responses surface used for the request. The chat model itself is only
+# the host that calls the ``image_generation`` tool; the actual image work is
+# done by ``API_MODEL``.
+_CODEX_CHAT_MODEL = "gpt-5.4"
+_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
+_CODEX_INSTRUCTIONS = (
+    "You are an assistant that must fulfill image generation requests by "
+    "using the image_generation tool when provided."
+)
+
+
+# ---------------------------------------------------------------------------
+# Config + auth helpers
+# ---------------------------------------------------------------------------
+
+
+def _load_image_gen_config() -> Dict[str, Any]:
+    """Read ``image_gen`` from config.yaml (returns {} on any failure)."""
+    try:
+        from hermes_cli.config import load_config
+
+        cfg = load_config()
+        section = cfg.get("image_gen") if isinstance(cfg, dict) else None
+        return section if isinstance(section, dict) else {}
+    except Exception as exc:
+        logger.debug("Could not load image_gen config: %s", exc)
+        return {}
+
+
+def _resolve_model() -> Tuple[str, Dict[str, Any]]:
+    """Decide which tier to use and return ``(model_id, meta)``."""
+    import os
+
+    env_override = os.environ.get("OPENAI_IMAGE_MODEL")
+    if env_override and env_override in _MODELS:
+        return env_override, _MODELS[env_override]
+
+    cfg = _load_image_gen_config()
+    sub = cfg.get("openai-codex") if isinstance(cfg.get("openai-codex"), dict) else {}
+    candidate: Optional[str] = None
+    if isinstance(sub, dict):
+        value = sub.get("model")
+        if isinstance(value, str) and value in _MODELS:
+            candidate = value
+    if candidate is None:
+        top = cfg.get("model")
+        if isinstance(top, str) and top in _MODELS:
+            candidate = top
+
+    if candidate is not None:
+        return candidate, _MODELS[candidate]
+
+    return DEFAULT_MODEL, _MODELS[DEFAULT_MODEL]
+
+
+def _read_codex_access_token() -> Optional[str]:
+    """Return a usable Codex OAuth token, or None.
+
+    Delegates to the canonical reader in ``agent.auxiliary_client`` so token
+    expiry, credential pool selection, and JWT decoding stay in one place.
+    """
+    try:
+        from agent.auxiliary_client import _read_codex_access_token as _reader
+
+        token = _reader()
+        if isinstance(token, str) and token.strip():
+            return token.strip()
+        return None
+    except Exception as exc:
+        logger.debug("Could not resolve Codex access token: %s", exc)
+        return None
+
+
+def _build_codex_client():
+    """Return an OpenAI client pointed at the ChatGPT/Codex backend, or None."""
+    token = _read_codex_access_token()
+    if not token:
+        return None
+    try:
+        import openai
+        from agent.auxiliary_client import _codex_cloudflare_headers
+
+        return openai.OpenAI(
+            api_key=token,
+            base_url=_CODEX_BASE_URL,
+            default_headers=_codex_cloudflare_headers(token),
+        )
+    except Exception as exc:
+        logger.debug("Could not build Codex image client: %s", exc)
+        return None
+
+
+def _collect_image_b64(client: Any, *, prompt: str, size: str, quality: str) -> Optional[str]:
+    """Stream a Codex Responses image_generation call and return the b64 image."""
+    image_b64: Optional[str] = None
+
+    with client.responses.stream(
+        model=_CODEX_CHAT_MODEL,
+        store=False,
+        instructions=_CODEX_INSTRUCTIONS,
+        input=[{
+            "type": "message",
+            "role": "user",
+            "content": [{"type": "input_text", "text": prompt}],
+        }],
+        tools=[{
+            "type": "image_generation",
+            "model": API_MODEL,
+            "size": size,
+            "quality": quality,
+            "output_format": "png",
+            "background": "opaque",
+            "partial_images": 1,
+        }],
+        tool_choice={
+            "type": "allowed_tools",
+            "mode": "required",
+            "tools": [{"type": "image_generation"}],
+        },
+    ) as stream:
+        for event in stream:
+            event_type = getattr(event, "type", "")
+            if event_type == "response.output_item.done":
+                item = getattr(event, "item", None)
+                if getattr(item, "type", None) == "image_generation_call":
+                    result = getattr(item, "result", None)
+                    if isinstance(result, str) and result:
+                        image_b64 = result
+            elif event_type == "response.image_generation_call.partial_image":
+                partial = getattr(event, "partial_image_b64", None)
+                if isinstance(partial, str) and partial:
+                    image_b64 = partial
+        final = stream.get_final_response()
+
+    # Final-response sweep covers the case where the stream finished before
+    # we observed the ``output_item.done`` event for the image call.
+    for item in getattr(final, "output", None) or []:
+        if getattr(item, "type", None) == "image_generation_call":
+            result = getattr(item, "result", None)
+            if isinstance(result, str) and result:
+                image_b64 = result
+
+    return image_b64
+
+
+# ---------------------------------------------------------------------------
+# Provider
+# ---------------------------------------------------------------------------
+
+
+class OpenAICodexImageGenProvider(ImageGenProvider):
+    """gpt-image-2 routed through ChatGPT/Codex OAuth instead of an API key."""
+
+    @property
+    def name(self) -> str:
+        return "openai-codex"
+
+    @property
+    def display_name(self) -> str:
+        return "OpenAI (Codex auth)"
+
+    def is_available(self) -> bool:
+        if not _read_codex_access_token():
+            return False
+        try:
+            import openai  # noqa: F401
+        except ImportError:
+            return False
+        return True
+
+    def list_models(self) -> List[Dict[str, Any]]:
+        return [
+            {
+                "id": model_id,
+                "display": meta["display"],
+                "speed": meta["speed"],
+                "strengths": meta["strengths"],
+                "price": "varies",
+            }
+            for model_id, meta in _MODELS.items()
+        ]
+
+    def default_model(self) -> Optional[str]:
+        return DEFAULT_MODEL
+
+    def get_setup_schema(self) -> Dict[str, Any]:
+        return {
+            "name": "OpenAI (Codex auth)",
+            "badge": "free",
+            "tag": "gpt-image-2 via ChatGPT/Codex OAuth — no API key required",
+            "env_vars": [],
+            "post_setup_hint": (
+                "Sign in with `hermes auth codex` (or `hermes setup` → Codex) "
+                "if you haven't already. No API key needed."
+            ),
+        }
+
+    def generate(
+        self,
+        prompt: str,
+        aspect_ratio: str = DEFAULT_ASPECT_RATIO,
+        **kwargs: Any,
+    ) -> Dict[str, Any]:
+        prompt = (prompt or "").strip()
+        aspect = resolve_aspect_ratio(aspect_ratio)
+
+        if not prompt:
+            return error_response(
+                error="Prompt is required and must be a non-empty string",
+                error_type="invalid_argument",
+                provider="openai-codex",
+                aspect_ratio=aspect,
+            )
+
+        if not _read_codex_access_token():
+            return error_response(
+                error=(
+                    "No Codex/ChatGPT OAuth credentials available. Run "
+                    "`hermes auth codex` (or `hermes setup` → Codex) to sign in."
+                ),
+                error_type="auth_required",
+                provider="openai-codex",
+                aspect_ratio=aspect,
+            )
+
+        try:
+            import openai  # noqa: F401
+        except ImportError:
+            return error_response(
+                error="openai Python package not installed (pip install openai)",
+                error_type="missing_dependency",
+                provider="openai-codex",
+                aspect_ratio=aspect,
+            )
+
+        tier_id, meta = _resolve_model()
+        size = _SIZES.get(aspect, _SIZES["square"])
+
+        client = _build_codex_client()
+        if client is None:
+            return error_response(
+                error="Could not initialize Codex image client",
+                error_type="auth_required",
+                provider="openai-codex",
+                model=tier_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+
+        try:
+            b64 = _collect_image_b64(
+                client,
+                prompt=prompt,
+                size=size,
+                quality=meta["quality"],
+            )
+        except Exception as exc:
+            logger.debug("Codex image generation failed", exc_info=True)
+            return error_response(
+                error=f"OpenAI image generation via Codex auth failed: {exc}",
+                error_type="api_error",
+                provider="openai-codex",
+                model=tier_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+
+        if not b64:
+            return error_response(
+                error="Codex response contained no image_generation_call result",
+                error_type="empty_response",
+                provider="openai-codex",
+                model=tier_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+
+        try:
+            saved_path = save_b64_image(b64, prefix=f"openai_codex_{tier_id}")
+        except Exception as exc:
+            return error_response(
+                error=f"Could not save image to cache: {exc}",
+                error_type="io_error",
+                provider="openai-codex",
+                model=tier_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+
+        return success_response(
+            image=str(saved_path),
+            model=tier_id,
+            prompt=prompt,
+            aspect_ratio=aspect,
+            provider="openai-codex",
+            extra={"size": size, "quality": meta["quality"]},
+        )
+
+
+# ---------------------------------------------------------------------------
+# Plugin entry point
+# ---------------------------------------------------------------------------
+
+
+def register(ctx) -> None:
+    """Plugin entry point — register the Codex-backed image-gen provider."""
+    ctx.register_image_gen_provider(OpenAICodexImageGenProvider())

plugins/image_gen/openai-codex/plugin.yaml

@@ -0,0 +1,5 @@
+name: openai-codex
+version: 1.0.0
+description: "OpenAI image generation backed by ChatGPT/Codex OAuth (gpt-image-2 via the Responses image_generation tool). Saves generated images to $HERMES_HOME/cache/images/."
+author: NousResearch
+kind: backend

plugins/image_gen/xai/__init__.py

@@ -0,0 +1,313 @@
+"""xAI image generation backend.
+
+Exposes xAI's ``grok-imagine-image`` model as an
+:class:`ImageGenProvider` implementation.
+
+Features:
+- Text-to-image generation
+- Multiple aspect ratios (1:1, 16:9, 9:16, etc.)
+- Multiple resolutions (1K, 2K)
+- Base64 output saved to cache
+
+Selection precedence (first hit wins):
+1. ``XAI_IMAGE_MODEL`` env var
+2. ``image_gen.xai.model`` in ``config.yaml``
+3. :data:`DEFAULT_MODEL`
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any, Dict, List, Optional, Tuple
+
+import requests
+
+from agent.image_gen_provider import (
+    DEFAULT_ASPECT_RATIO,
+    ImageGenProvider,
+    error_response,
+    resolve_aspect_ratio,
+    save_b64_image,
+    success_response,
+)
+from tools.xai_http import hermes_xai_user_agent
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Model catalog
+# ---------------------------------------------------------------------------
+
+API_MODEL = "grok-imagine-image"
+
+_MODELS: Dict[str, Dict[str, Any]] = {
+    "grok-imagine-image": {
+        "display": "Grok Imagine Image",
+        "speed": "~5-10s",
+        "strengths": "Fast, high-quality",
+    },
+}
+
+DEFAULT_MODEL = "grok-imagine-image"
+
+# xAI aspect ratios (more options than FAL/OpenAI)
+_XAI_ASPECT_RATIOS = {
+    "landscape": "16:9",
+    "square": "1:1",
+    "portrait": "9:16",
+    "4:3": "4:3",
+    "3:4": "3:4",
+    "3:2": "3:2",
+    "2:3": "2:3",
+}
+
+# xAI resolutions
+_XAI_RESOLUTIONS = {
+    "1k": "1024",
+    "2k": "2048",
+}
+
+DEFAULT_RESOLUTION = "1k"
+
+
+# ---------------------------------------------------------------------------
+# Config
+# ---------------------------------------------------------------------------
+
+
+def _load_xai_config() -> Dict[str, Any]:
+    """Read ``image_gen.xai`` from config.yaml."""
+    try:
+        from hermes_cli.config import load_config
+
+        cfg = load_config()
+        section = cfg.get("image_gen") if isinstance(cfg, dict) else None
+        xai_section = section.get("xai") if isinstance(section, dict) else None
+        return xai_section if isinstance(xai_section, dict) else {}
+    except Exception as exc:
+        logger.debug("Could not load image_gen.xai config: %s", exc)
+        return {}
+
+
+def _resolve_model() -> Tuple[str, Dict[str, Any]]:
+    """Decide which model to use and return ``(model_id, meta)``."""
+    env_override = os.environ.get("XAI_IMAGE_MODEL")
+    if env_override and env_override in _MODELS:
+        return env_override, _MODELS[env_override]
+
+    cfg = _load_xai_config()
+    candidate = cfg.get("model") if isinstance(cfg.get("model"), str) else None
+    if candidate and candidate in _MODELS:
+        return candidate, _MODELS[candidate]
+
+    return DEFAULT_MODEL, _MODELS[DEFAULT_MODEL]
+
+
+def _resolve_resolution() -> str:
+    """Get configured resolution."""
+    cfg = _load_xai_config()
+    res = cfg.get("resolution") if isinstance(cfg.get("resolution"), str) else None
+    if res and res in _XAI_RESOLUTIONS:
+        return res
+    return DEFAULT_RESOLUTION
+
+
+# ---------------------------------------------------------------------------
+# Provider
+# ---------------------------------------------------------------------------
+
+
+class XAIImageGenProvider(ImageGenProvider):
+    """xAI ``grok-imagine-image`` backend."""
+
+    @property
+    def name(self) -> str:
+        return "xai"
+
+    @property
+    def display_name(self) -> str:
+        return "xAI (Grok)"
+
+    def is_available(self) -> bool:
+        return bool(os.getenv("XAI_API_KEY"))
+
+    def list_models(self) -> List[Dict[str, Any]]:
+        return [
+            {
+                "id": model_id,
+                "display": meta.get("display", model_id),
+                "speed": meta.get("speed", ""),
+                "strengths": meta.get("strengths", ""),
+            }
+            for model_id, meta in _MODELS.items()
+        ]
+
+    def get_setup_schema(self) -> Dict[str, Any]:
+        return {
+            "name": "xAI (Grok)",
+            "badge": "paid",
+            "tag": "Native xAI image generation via grok-imagine-image",
+            "env_vars": [
+                {
+                    "key": "XAI_API_KEY",
+                    "prompt": "xAI API key",
+                    "url": "https://console.x.ai/",
+                },
+            ],
+        }
+
+    def generate(
+        self,
+        prompt: str,
+        aspect_ratio: str = DEFAULT_ASPECT_RATIO,
+        **kwargs: Any,
+    ) -> Dict[str, Any]:
+        """Generate an image using xAI's grok-imagine-image."""
+        api_key = os.getenv("XAI_API_KEY", "").strip()
+        if not api_key:
+            return error_response(
+                error="XAI_API_KEY not set. Get one at https://console.x.ai/",
+                error_type="missing_api_key",
+                provider="xai",
+                aspect_ratio=aspect_ratio,
+            )
+
+        model_id, meta = _resolve_model()
+        aspect = resolve_aspect_ratio(aspect_ratio)
+        xai_ar = _XAI_ASPECT_RATIOS.get(aspect, "1:1")
+        resolution = _resolve_resolution()
+        xai_res = _XAI_RESOLUTIONS.get(resolution, "1024")
+
+        payload: Dict[str, Any] = {
+            "model": API_MODEL,
+            "prompt": prompt,
+            "aspect_ratio": xai_ar,
+            "resolution": xai_res,
+        }
+
+        headers = {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json",
+            "User-Agent": hermes_xai_user_agent(),
+        }
+
+        base_url = (os.getenv("XAI_BASE_URL") or "https://api.x.ai/v1").strip().rstrip("/")
+
+        try:
+            response = requests.post(
+                f"{base_url}/images/generations",
+                headers=headers,
+                json=payload,
+                timeout=120,
+            )
+            response.raise_for_status()
+        except requests.HTTPError as exc:
+            status = exc.response.status_code if exc.response else 0
+            try:
+                err_msg = exc.response.json().get("error", {}).get("message", exc.response.text[:300])
+            except Exception:
+                err_msg = exc.response.text[:300] if exc.response else str(exc)
+            logger.error("xAI image gen failed (%d): %s", status, err_msg)
+            return error_response(
+                error=f"xAI image generation failed ({status}): {err_msg}",
+                error_type="api_error",
+                provider="xai",
+                model=model_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+        except requests.Timeout:
+            return error_response(
+                error="xAI image generation timed out (120s)",
+                error_type="timeout",
+                provider="xai",
+                model=model_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+        except requests.ConnectionError as exc:
+            return error_response(
+                error=f"xAI connection error: {exc}",
+                error_type="connection_error",
+                provider="xai",
+                model=model_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+
+        try:
+            result = response.json()
+        except Exception as exc:
+            return error_response(
+                error=f"xAI returned invalid JSON: {exc}",
+                error_type="invalid_response",
+                provider="xai",
+                model=model_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+
+        # Parse response — xAI returns data[0].b64_json or data[0].url
+        data = result.get("data", [])
+        if not data:
+            return error_response(
+                error="xAI returned no image data",
+                error_type="empty_response",
+                provider="xai",
+                model=model_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+
+        first = data[0]
+        b64 = first.get("b64_json")
+        url = first.get("url")
+
+        if b64:
+            try:
+                saved_path = save_b64_image(b64, prefix=f"xai_{model_id}")
+            except Exception as exc:
+                return error_response(
+                    error=f"Could not save image to cache: {exc}",
+                    error_type="io_error",
+                    provider="xai",
+                    model=model_id,
+                    prompt=prompt,
+                    aspect_ratio=aspect,
+                )
+            image_ref = str(saved_path)
+        elif url:
+            image_ref = url
+        else:
+            return error_response(
+                error="xAI response contained neither b64_json nor URL",
+                error_type="empty_response",
+                provider="xai",
+                model=model_id,
+                prompt=prompt,
+                aspect_ratio=aspect,
+            )
+
+        extra: Dict[str, Any] = {
+            "resolution": xai_res,
+        }
+
+        return success_response(
+            image=image_ref,
+            model=model_id,
+            prompt=prompt,
+            aspect_ratio=aspect,
+            provider="xai",
+            extra=extra,
+        )
+
+
+# ---------------------------------------------------------------------------
+# Plugin registration
+# ---------------------------------------------------------------------------
+
+
+def register(ctx: Any) -> None:
+    """Register this provider with the image gen registry."""
+    ctx.register_image_gen_provider(XAIImageGenProvider())

plugins/image_gen/xai/plugin.yaml

@@ -0,0 +1,7 @@
+name: xai
+version: 1.0.0
+description: "xAI image generation backend (grok-imagine-image). Text-to-image."
+author: Julien Talbot
+kind: backend
+requires_env:
+  - XAI_API_KEY

plugins/memory/hindsight/README.md

@@ -59,7 +59,8 @@ Config file: `~/.hermes/hindsight/config.json`
 
 | Key | Default | Description |
 |-----|---------|-------------|
-| `bank_id` | `hermes` | Memory bank name |
+| `bank_id` | `hermes` | Memory bank name (static fallback used when `bank_id_template` is unset or resolves empty) |
+| `bank_id_template` | — | Optional template to derive the bank name dynamically. Placeholders: `{profile}`, `{workspace}`, `{platform}`, `{user}`, `{session}`. Example: `hermes-{profile}` isolates memory per active Hermes profile. Empty placeholders collapse cleanly (e.g. `hermes-{user}` with no user becomes `hermes`). |
 | `bank_mission` | — | Reflect mission (identity/framing for reflect reasoning). Applied via Banks API. |
 | `bank_retain_mission` | — | Retain mission (steers what gets extracted). Applied via Banks API. |
 

plugins/memory/hindsight/__init__.py

@@ -3,6 +3,8 @@
 Long-term memory with knowledge graph, entity resolution, and multi-strategy
 retrieval. Supports cloud (API key) and local modes.
 
+Configurable timeout via HINDSIGHT_TIMEOUT env var or config.json.
+
 Original PR #1811 by benfrank241, adapted to MemoryProvider ABC.
 
 Config via environment variables:
@@ -11,6 +13,7 @@ Config via environment variables:
   HINDSIGHT_BUDGET                 — recall budget: low/mid/high (default: mid)
   HINDSIGHT_API_URL                — API endpoint
   HINDSIGHT_MODE                   — cloud or local (default: cloud)
+  HINDSIGHT_TIMEOUT                — API request timeout in seconds (default: 120)
   HINDSIGHT_RETAIN_TAGS            — comma-separated tags attached to retained memories
   HINDSIGHT_RETAIN_SOURCE          — metadata source value attached to retained memories
   HINDSIGHT_RETAIN_USER_PREFIX     — label used before user turns in retained transcripts
@@ -23,6 +26,7 @@ Or via $HERMES_HOME/hindsight/config.json (profile-scoped), falling back to
 from __future__ import annotations
 
 import asyncio
+import importlib
 import json
 import logging
 import os
@@ -40,6 +44,7 @@ logger = logging.getLogger(__name__)
 _DEFAULT_API_URL = "https://api.hindsight.vectorize.io"
 _DEFAULT_LOCAL_URL = "http://localhost:8888"
 _MIN_CLIENT_VERSION = "0.4.22"
+_DEFAULT_TIMEOUT = 120  # seconds — cloud API can take 30-40s per request
 _VALID_BUDGETS = {"low", "mid", "high"}
 _PROVIDER_DEFAULT_MODELS = {
     "openai": "gpt-4o-mini",
@@ -54,6 +59,22 @@ _PROVIDER_DEFAULT_MODELS = {
 }
 
 
+def _check_local_runtime() -> tuple[bool, str | None]:
+    """Return whether local embedded Hindsight imports cleanly.
+
+    On older CPUs, importing the local Hindsight stack can raise a runtime
+    error from NumPy before the daemon starts. Treat that as "unavailable"
+    so Hermes can degrade gracefully instead of repeatedly trying to start
+    a broken local memory backend.
+    """
+    try:
+        importlib.import_module("hindsight")
+        importlib.import_module("hindsight_embed.daemon_embed_manager")
+        return True, None
+    except Exception as exc:
+        return False, str(exc)
+
+
 # ---------------------------------------------------------------------------
 # Dedicated event loop for Hindsight async calls (one per process, reused).
 # Avoids creating ephemeral loops that leak aiohttp sessions.
@@ -81,13 +102,18 @@ def _get_loop() -> asyncio.AbstractEventLoop:
         return _loop
 
 
-def _run_sync(coro, timeout: float = 120.0):
+def _run_sync(coro, timeout: float = _DEFAULT_TIMEOUT):
     """Schedule *coro* on the shared loop and block until done."""
     loop = _get_loop()
     future = asyncio.run_coroutine_threadsafe(coro, loop)
     return future.result(timeout=timeout)
 
 
+# ---------------------------------------------------------------------------
+# Backward-compatible alias — instances use self._run_sync() instead.
+# ---------------------------------------------------------------------------
+
+
 # ---------------------------------------------------------------------------
 # Tool schemas
 # ---------------------------------------------------------------------------
@@ -233,6 +259,126 @@ def _utc_timestamp() -> str:
     return datetime.now(timezone.utc).isoformat(timespec="milliseconds").replace("+00:00", "Z")
 
 
+def _embedded_profile_name(config: dict[str, Any]) -> str:
+    """Return the Hindsight embedded profile name for this Hermes config."""
+    profile = config.get("profile", "hermes")
+    return str(profile or "hermes")
+
+
+def _load_simple_env(path) -> dict[str, str]:
+    """Parse a simple KEY=VALUE env file, ignoring comments and blank lines."""
+    if not path.exists():
+        return {}
+
+    values: dict[str, str] = {}
+    for line in path.read_text(encoding="utf-8").splitlines():
+        if not line or line.startswith("#") or "=" not in line:
+            continue
+        key, value = line.split("=", 1)
+        values[key.strip()] = value.strip()
+    return values
+
+
+def _build_embedded_profile_env(config: dict[str, Any], *, llm_api_key: str | None = None) -> dict[str, str]:
+    """Build the profile-scoped env file that standalone hindsight-embed consumes."""
+    current_key = llm_api_key
+    if current_key is None:
+        current_key = (
+            config.get("llmApiKey")
+            or config.get("llm_api_key")
+            or os.environ.get("HINDSIGHT_LLM_API_KEY", "")
+        )
+
+    current_provider = config.get("llm_provider", "")
+    current_model = config.get("llm_model", "")
+    current_base_url = config.get("llm_base_url") or os.environ.get("HINDSIGHT_API_LLM_BASE_URL", "")
+
+    # The embedded daemon expects OpenAI wire format for these providers.
+    daemon_provider = "openai" if current_provider in ("openai_compatible", "openrouter") else current_provider
+
+    env_values = {
+        "HINDSIGHT_API_LLM_PROVIDER": str(daemon_provider),
+        "HINDSIGHT_API_LLM_API_KEY": str(current_key or ""),
+        "HINDSIGHT_API_LLM_MODEL": str(current_model),
+        "HINDSIGHT_API_LOG_LEVEL": "info",
+    }
+    if current_base_url:
+        env_values["HINDSIGHT_API_LLM_BASE_URL"] = str(current_base_url)
+    return env_values
+
+
+def _embedded_profile_env_path(config: dict[str, Any]):
+    from pathlib import Path
+
+    return Path.home() / ".hindsight" / "profiles" / f"{_embedded_profile_name(config)}.env"
+
+
+def _materialize_embedded_profile_env(config: dict[str, Any], *, llm_api_key: str | None = None):
+    """Write the profile-scoped env file that standalone hindsight-embed uses."""
+    profile_env = _embedded_profile_env_path(config)
+    profile_env.parent.mkdir(parents=True, exist_ok=True)
+    env_values = _build_embedded_profile_env(config, llm_api_key=llm_api_key)
+    profile_env.write_text(
+        "".join(f"{key}={value}\n" for key, value in env_values.items()),
+        encoding="utf-8",
+    )
+    return profile_env
+
+def _sanitize_bank_segment(value: str) -> str:
+    """Sanitize a bank_id_template placeholder value.
+
+    Bank IDs should be safe for URL paths and filesystem use. Replaces any
+    character that isn't alphanumeric, dash, or underscore with a dash, and
+    collapses runs of dashes.
+    """
+    if not value:
+        return ""
+    out = []
+    prev_dash = False
+    for ch in str(value):
+        if ch.isalnum() or ch == "-" or ch == "_":
+            out.append(ch)
+            prev_dash = False
+        else:
+            if not prev_dash:
+                out.append("-")
+                prev_dash = True
+    return "".join(out).strip("-_")
+
+
+def _resolve_bank_id_template(template: str, fallback: str, **placeholders: str) -> str:
+    """Resolve a bank_id template string with the given placeholders.
+
+    Supported placeholders (each is sanitized before substitution):
+      {profile}   — active Hermes profile name (from agent_identity)
+      {workspace} — Hermes workspace name (from agent_workspace)
+      {platform}  — "cli", "telegram", "discord", etc.
+      {user}      — platform user id (gateway sessions)
+      {session}   — current session id
+
+    Missing/empty placeholders are rendered as the empty string and then
+    collapsed — e.g. ``hermes-{user}`` with no user becomes ``hermes``.
+
+    If the template is empty, resolution falls back to *fallback*.
+    Returns the sanitized bank id.
+    """
+    if not template:
+        return fallback
+    sanitized = {k: _sanitize_bank_segment(v) for k, v in placeholders.items()}
+    try:
+        rendered = template.format(**sanitized)
+    except (KeyError, IndexError) as exc:
+        logger.warning("Invalid bank_id_template %r: %s — using fallback %r",
+                       template, exc, fallback)
+        return fallback
+    while "--" in rendered:
+        rendered = rendered.replace("--", "-")
+    while "__" in rendered:
+        rendered = rendered.replace("__", "_")
+    rendered = rendered.strip("-_")
+    return rendered or fallback
+
+
 # ---------------------------------------------------------------------------
 # MemoryProvider implementation
 # ---------------------------------------------------------------------------
@@ -262,13 +408,17 @@ class HindsightMemoryProvider(MemoryProvider):
         self._chat_type = ""
         self._thread_id = ""
         self._agent_identity = ""
+        self._agent_workspace = ""
         self._turn_index = 0
         self._client = None
+        self._timeout = _DEFAULT_TIMEOUT
         self._prefetch_result = ""
         self._prefetch_lock = threading.Lock()
         self._prefetch_thread = None
         self._sync_thread = None
         self._session_id = ""
+        self._parent_session_id = ""
+        self._document_id = ""
 
         # Tags
         self._tags: list[str] | None = None
@@ -293,6 +443,7 @@ class HindsightMemoryProvider(MemoryProvider):
         # Bank
         self._bank_mission = ""
         self._bank_retain_mission: str | None = None
+        self._bank_id_template = ""
 
     @property
     def name(self) -> str:
@@ -302,9 +453,16 @@ class HindsightMemoryProvider(MemoryProvider):
         try:
             cfg = _load_config()
             mode = cfg.get("mode", "cloud")
-            if mode in ("local", "local_embedded", "local_external"):
+            if mode in ("local", "local_embedded"):
+                available, _ = _check_local_runtime()
+                return available
+            if mode == "local_external":
                 return True
-            has_key = bool(cfg.get("apiKey") or os.environ.get("HINDSIGHT_API_KEY", ""))
+            has_key = bool(
+                cfg.get("apiKey")
+                or cfg.get("api_key")
+                or os.environ.get("HINDSIGHT_API_KEY", "")
+            )
             has_url = bool(cfg.get("api_url") or os.environ.get("HINDSIGHT_API_URL", ""))
             return has_key or has_url
         except Exception:
@@ -363,7 +521,7 @@ class HindsightMemoryProvider(MemoryProvider):
         else:
             deps_to_install = [cloud_dep]
 
-        print(f"\n  Checking dependencies...")
+        print("\n  Checking dependencies...")
         uv_path = shutil.which("uv")
         if not uv_path:
             print("  ⚠ uv not found — install it: curl -LsSf https://astral.sh/uv/install.sh | sh")
@@ -374,14 +532,14 @@ class HindsightMemoryProvider(MemoryProvider):
                     [uv_path, "pip", "install", "--python", sys.executable, "--quiet", "--upgrade"] + deps_to_install,
                     check=True, timeout=120, capture_output=True,
                 )
-                print(f"  ✓ Dependencies up to date")
+                print("  ✓ Dependencies up to date")
             except Exception as e:
                 print(f"  ⚠ Install failed: {e}")
                 print(f"  Run manually: uv pip install --python {sys.executable} {' '.join(deps_to_install)}")
 
         # Step 3: Mode-specific config
         if mode == "cloud":
-            print(f"\n  Get your API key at https://ui.hindsight.vectorize.io\n")
+            print("\n  Get your API key at https://ui.hindsight.vectorize.io\n")
             existing_key = os.environ.get("HINDSIGHT_API_KEY", "")
             if existing_key:
                 masked = f"...{existing_key[-4:]}" if len(existing_key) > 4 else "set"
@@ -434,13 +592,19 @@ class HindsightMemoryProvider(MemoryProvider):
             sys.stdout.write("  LLM API key: ")
             sys.stdout.flush()
             llm_key = getpass.getpass(prompt="") if sys.stdin.isatty() else sys.stdin.readline().strip()
-            if llm_key:
-                env_writes["HINDSIGHT_LLM_API_KEY"] = llm_key
+            # Always write explicitly (including empty) so the provider sees ""
+            # rather than a missing variable.  The daemon reads from .env at
+            # startup and fails when HINDSIGHT_LLM_API_KEY is unset.
+            env_writes["HINDSIGHT_LLM_API_KEY"] = llm_key
 
         # Step 4: Save everything
         provider_config["bank_id"] = "hermes"
         provider_config["recall_budget"] = "mid"
-        bank_id = "hermes"
+        # Read existing timeout from config if present, otherwise use default
+        existing_timeout = self._config.get("timeout") if self._config else None
+        timeout_val = existing_timeout if existing_timeout else _DEFAULT_TIMEOUT
+        provider_config["timeout"] = timeout_val
+        env_writes["HINDSIGHT_TIMEOUT"] = str(timeout_val)
         config["memory"]["provider"] = "hindsight"
         save_config(config)
 
@@ -466,10 +630,32 @@ class HindsightMemoryProvider(MemoryProvider):
                     new_lines.append(f"{k}={v}")
             env_path.write_text("\n".join(new_lines) + "\n")
 
+        if mode == "local_embedded":
+            materialized_config = dict(provider_config)
+            config_path = Path(hermes_home) / "hindsight" / "config.json"
+            try:
+                materialized_config = json.loads(config_path.read_text(encoding="utf-8"))
+            except Exception:
+                pass
+
+            llm_api_key = env_writes.get("HINDSIGHT_LLM_API_KEY", "")
+            if not llm_api_key:
+                llm_api_key = _load_simple_env(Path(hermes_home) / ".env").get("HINDSIGHT_LLM_API_KEY", "")
+            if not llm_api_key:
+                llm_api_key = _load_simple_env(_embedded_profile_env_path(materialized_config)).get(
+                    "HINDSIGHT_API_LLM_API_KEY",
+                    "",
+                )
+
+            _materialize_embedded_profile_env(
+                materialized_config,
+                llm_api_key=llm_api_key or None,
+            )
+
         print(f"\n  ✓ Hindsight memory configured ({mode} mode)")
         if env_writes:
-            print(f"  API keys saved to .env")
-        print(f"\n  Start a new session to activate.\n")
+            print("  API keys saved to .env")
+        print("\n  Start a new session to activate.\n")
 
     def get_config_schema(self):
         return [
@@ -485,7 +671,8 @@ class HindsightMemoryProvider(MemoryProvider):
             {"key": "llm_base_url", "description": "Endpoint URL (e.g. http://192.168.1.10:8080/v1)", "default": "", "when": {"mode": "local_embedded", "llm_provider": "openai_compatible"}},
             {"key": "llm_api_key", "description": "LLM API key (optional for openai_compatible)", "secret": True, "env_var": "HINDSIGHT_LLM_API_KEY", "when": {"mode": "local_embedded"}},
             {"key": "llm_model", "description": "LLM model", "default": "gpt-4o-mini", "default_from": {"field": "llm_provider", "map": _PROVIDER_DEFAULT_MODELS}, "when": {"mode": "local_embedded"}},
-            {"key": "bank_id", "description": "Memory bank name", "default": "hermes"},
+            {"key": "bank_id", "description": "Memory bank name (static fallback when bank_id_template is unset)", "default": "hermes"},
+            {"key": "bank_id_template", "description": "Optional template to derive bank_id dynamically. Placeholders: {profile}, {workspace}, {platform}, {user}, {session}. Example: hermes-{profile}", "default": ""},
             {"key": "bank_mission", "description": "Mission/purpose description for the memory bank"},
             {"key": "bank_retain_mission", "description": "Custom extraction prompt for memory retention"},
             {"key": "recall_budget", "description": "Recall thoroughness", "default": "mid", "choices": ["low", "mid", "high"]},
@@ -505,12 +692,19 @@ class HindsightMemoryProvider(MemoryProvider):
             {"key": "recall_max_tokens", "description": "Maximum tokens for recall results", "default": 4096},
             {"key": "recall_max_input_chars", "description": "Maximum input query length for auto-recall", "default": 800},
             {"key": "recall_prompt_preamble", "description": "Custom preamble for recalled memories in context"},
+            {"key": "timeout", "description": "API request timeout in seconds", "default": _DEFAULT_TIMEOUT},
         ]
 
     def _get_client(self):
         """Return the cached Hindsight client (created once, reused)."""
         if self._client is None:
             if self._mode == "local_embedded":
+                available, reason = _check_local_runtime()
+                if not available:
+                    raise RuntimeError(
+                        "Hindsight local runtime is unavailable"
+                        + (f": {reason}" if reason else "")
+                    )
                 from hindsight import HindsightEmbedded
                 HindsightEmbedded.__del__ = lambda self: None
                 llm_provider = self._config.get("llm_provider", "")
@@ -529,16 +723,30 @@ class HindsightMemoryProvider(MemoryProvider):
                 self._client = HindsightEmbedded(**kwargs)
             else:
                 from hindsight_client import Hindsight
-                kwargs = {"base_url": self._api_url, "timeout": 30.0}
+                timeout = self._timeout or _DEFAULT_TIMEOUT
+                kwargs = {"base_url": self._api_url, "timeout": float(timeout)}
                 if self._api_key:
                     kwargs["api_key"] = self._api_key
-                logger.debug("Creating Hindsight cloud client (url=%s, has_key=%s)",
-                             self._api_url, bool(self._api_key))
+                logger.debug("Creating Hindsight cloud client (url=%s, has_key=%s, timeout=%s)",
+                             self._api_url, bool(self._api_key), kwargs["timeout"])
                 self._client = Hindsight(**kwargs)
         return self._client
 
+    def _run_sync(self, coro):
+        """Schedule *coro* on the shared loop using the configured timeout."""
+        return _run_sync(coro, timeout=self._timeout)
+
     def initialize(self, session_id: str, **kwargs) -> None:
         self._session_id = str(session_id or "").strip()
+        self._parent_session_id = str(kwargs.get("parent_session_id", "") or "").strip()
+
+        # Each process lifecycle gets its own document_id. Reusing session_id
+        # alone caused overwrites on /resume — the reloaded session starts
+        # with an empty _session_turns, so the next retain would replace the
+        # previously stored content. session_id stays in tags so processes
+        # for the same session remain filterable together.
+        start_ts = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
+        self._document_id = f"{self._session_id}-{start_ts}"
 
         # Check client version and auto-upgrade if needed
         try:
@@ -548,7 +756,9 @@ class HindsightMemoryProvider(MemoryProvider):
             if Version(installed) < Version(_MIN_CLIENT_VERSION):
                 logger.warning("hindsight-client %s is outdated (need >=%s), attempting upgrade...",
                                installed, _MIN_CLIENT_VERSION)
-                import shutil, subprocess, sys
+                import shutil
+                import subprocess
+                import sys
                 uv_path = shutil.which("uv")
                 if uv_path:
                     try:
@@ -575,19 +785,41 @@ class HindsightMemoryProvider(MemoryProvider):
         self._chat_type = str(kwargs.get("chat_type") or "").strip()
         self._thread_id = str(kwargs.get("thread_id") or "").strip()
         self._agent_identity = str(kwargs.get("agent_identity") or "").strip()
+        self._agent_workspace = str(kwargs.get("agent_workspace") or "").strip()
         self._turn_index = 0
         self._session_turns = []
         self._mode = self._config.get("mode", "cloud")
+        # Read timeout from config or env var, fall back to default
+        self._timeout = self._config.get("timeout") or int(os.environ.get("HINDSIGHT_TIMEOUT", str(_DEFAULT_TIMEOUT)))
         # "local" is a legacy alias for "local_embedded"
         if self._mode == "local":
             self._mode = "local_embedded"
+        if self._mode == "local_embedded":
+            available, reason = _check_local_runtime()
+            if not available:
+                logger.warning(
+                    "Hindsight local mode disabled because its runtime could not be imported: %s",
+                    reason,
+                )
+                self._mode = "disabled"
+                return
         self._api_key = self._config.get("apiKey") or self._config.get("api_key") or os.environ.get("HINDSIGHT_API_KEY", "")
         default_url = _DEFAULT_LOCAL_URL if self._mode in ("local_embedded", "local_external") else _DEFAULT_API_URL
         self._api_url = self._config.get("api_url") or os.environ.get("HINDSIGHT_API_URL", default_url)
         self._llm_base_url = self._config.get("llm_base_url", "")
 
         banks = self._config.get("banks", {}).get("hermes", {})
-        self._bank_id = self._config.get("bank_id") or banks.get("bankId", "hermes")
+        static_bank_id = self._config.get("bank_id") or banks.get("bankId", "hermes")
+        self._bank_id_template = self._config.get("bank_id_template", "") or ""
+        self._bank_id = _resolve_bank_id_template(
+            self._bank_id_template,
+            fallback=static_bank_id,
+            profile=self._agent_identity,
+            workspace=self._agent_workspace,
+            platform=self._platform,
+            user=self._user_id,
+            session=self._session_id,
+        )
         budget = self._config.get("recall_budget") or self._config.get("budget") or banks.get("budget", "mid")
         self._budget = budget if budget in _VALID_BUDGETS else "mid"
 
@@ -640,6 +872,10 @@ class HindsightMemoryProvider(MemoryProvider):
             pass
         logger.info("Hindsight initialized: mode=%s, api_url=%s, bank=%s, budget=%s, memory_mode=%s, prefetch_method=%s, client=%s",
                      self._mode, self._api_url, self._bank_id, self._budget, self._memory_mode, self._prefetch_method, _client_version)
+        if self._bank_id_template:
+            logger.debug("Hindsight bank resolved from template %r: profile=%s workspace=%s platform=%s user=%s -> bank=%s",
+                         self._bank_id_template, self._agent_identity, self._agent_workspace,
+                         self._platform, self._user_id, self._bank_id)
         logger.debug("Hindsight config: auto_retain=%s, auto_recall=%s, retain_every_n=%d, "
                      "retain_async=%s, retain_context=%s, recall_max_tokens=%d, recall_max_input_chars=%d, tags=%s, recall_tags=%s",
                      self._auto_retain, self._auto_recall, self._retain_every_n_turns,
@@ -669,42 +905,13 @@ class HindsightMemoryProvider(MemoryProvider):
                     # Update the profile .env to match our current config so
                     # the daemon always starts with the right settings.
                     # If the config changed and the daemon is running, stop it.
-                    from pathlib import Path as _Path
-                    profile_env = _Path.home() / ".hindsight" / "profiles" / f"{profile}.env"
-                    current_key = self._config.get("llm_api_key") or os.environ.get("HINDSIGHT_LLM_API_KEY", "")
-                    current_provider = self._config.get("llm_provider", "")
-                    current_model = self._config.get("llm_model", "")
-                    current_base_url = self._config.get("llm_base_url") or os.environ.get("HINDSIGHT_API_LLM_BASE_URL", "")
-                    # Map openai_compatible/openrouter → openai for the daemon (OpenAI wire format)
-                    daemon_provider = "openai" if current_provider in ("openai_compatible", "openrouter") else current_provider
-
-                    # Read saved profile config
-                    saved = {}
-                    if profile_env.exists():
-                        for line in profile_env.read_text().splitlines():
-                            if "=" in line and not line.startswith("#"):
-                                k, v = line.split("=", 1)
-                                saved[k.strip()] = v.strip()
-
-                    config_changed = (
-                        saved.get("HINDSIGHT_API_LLM_PROVIDER") != daemon_provider or
-                        saved.get("HINDSIGHT_API_LLM_MODEL") != current_model or
-                        saved.get("HINDSIGHT_API_LLM_API_KEY") != current_key or
-                        saved.get("HINDSIGHT_API_LLM_BASE_URL", "") != current_base_url
-                    )
+                    profile_env = _embedded_profile_env_path(self._config)
+                    expected_env = _build_embedded_profile_env(self._config)
+                    saved = _load_simple_env(profile_env)
+                    config_changed = saved != expected_env
 
                     if config_changed:
-                        # Write updated profile .env
-                        profile_env.parent.mkdir(parents=True, exist_ok=True)
-                        env_lines = (
-                            f"HINDSIGHT_API_LLM_PROVIDER={daemon_provider}\n"
-                            f"HINDSIGHT_API_LLM_API_KEY={current_key}\n"
-                            f"HINDSIGHT_API_LLM_MODEL={current_model}\n"
-                            f"HINDSIGHT_API_LOG_LEVEL=info\n"
-                        )
-                        if current_base_url:
-                            env_lines += f"HINDSIGHT_API_LLM_BASE_URL={current_base_url}\n"
-                        profile_env.write_text(env_lines)
+                        profile_env = _materialize_embedded_profile_env(self._config)
                         if client._manager.is_running(profile):
                             with open(log_path, "a") as f:
                                 f.write("\n=== Config changed, restarting daemon ===\n")
@@ -777,7 +984,7 @@ class HindsightMemoryProvider(MemoryProvider):
                 client = self._get_client()
                 if self._prefetch_method == "reflect":
                     logger.debug("Prefetch: calling reflect (bank=%s, query_len=%d)", self._bank_id, len(query))
-                    resp = _run_sync(client.areflect(bank_id=self._bank_id, query=query, budget=self._budget))
+                    resp = self._run_sync(client.areflect(bank_id=self._bank_id, query=query, budget=self._budget))
                     text = resp.text or ""
                 else:
                     recall_kwargs: dict = {
@@ -791,7 +998,7 @@ class HindsightMemoryProvider(MemoryProvider):
                         recall_kwargs["types"] = self._recall_types
                     logger.debug("Prefetch: calling recall (bank=%s, query_len=%d, budget=%s)",
                                  self._bank_id, len(query), self._budget)
-                    resp = _run_sync(client.arecall(**recall_kwargs))
+                    resp = self._run_sync(client.arecall(**recall_kwargs))
                     num_results = len(resp.results) if resp.results else 0
                     logger.debug("Prefetch: recall returned %d results", num_results)
                     text = "\n".join(f"- {r.text}" for r in resp.results if r.text) if resp.results else ""
@@ -888,7 +1095,7 @@ class HindsightMemoryProvider(MemoryProvider):
         if session_id:
             self._session_id = str(session_id).strip()
 
-        turn = json.dumps(self._build_turn_messages(user_content, assistant_content))
+        turn = json.dumps(self._build_turn_messages(user_content, assistant_content), ensure_ascii=False)
         self._session_turns.append(turn)
         self._turn_counter += 1
         self._turn_index = self._turn_counter
@@ -902,6 +1109,12 @@ class HindsightMemoryProvider(MemoryProvider):
                      len(self._session_turns), sum(len(t) for t in self._session_turns))
         content = "[" + ",".join(self._session_turns) + "]"
 
+        lineage_tags: list[str] = []
+        if self._session_id:
+            lineage_tags.append(f"session:{self._session_id}")
+        if self._parent_session_id:
+            lineage_tags.append(f"parent:{self._parent_session_id}")
+
         def _sync():
             try:
                 client = self._get_client()
@@ -912,15 +1125,16 @@ class HindsightMemoryProvider(MemoryProvider):
                         message_count=len(self._session_turns) * 2,
                         turn_index=self._turn_index,
                     ),
+                    tags=lineage_tags or None,
                 )
                 item.pop("bank_id", None)
                 item.pop("retain_async", None)
                 logger.debug("Hindsight retain: bank=%s, doc=%s, async=%s, content_len=%d, num_turns=%d",
-                             self._bank_id, self._session_id, self._retain_async, len(content), len(self._session_turns))
-                _run_sync(client.aretain_batch(
+                             self._bank_id, self._document_id, self._retain_async, len(content), len(self._session_turns))
+                self._run_sync(client.aretain_batch(
                     bank_id=self._bank_id,
                     items=[item],
-                    document_id=self._session_id,
+                    document_id=self._document_id,
                     retain_async=self._retain_async,
                 ))
                 logger.debug("Hindsight retain succeeded")
@@ -957,7 +1171,7 @@ class HindsightMemoryProvider(MemoryProvider):
                 )
                 logger.debug("Tool hindsight_retain: bank=%s, content_len=%d, context=%s",
                              self._bank_id, len(content), context)
-                _run_sync(client.aretain(**retain_kwargs))
+                self._run_sync(client.aretain(**retain_kwargs))
                 logger.debug("Tool hindsight_retain: success")
                 return json.dumps({"result": "Memory stored successfully."})
             except Exception as e:
@@ -980,7 +1194,7 @@ class HindsightMemoryProvider(MemoryProvider):
                     recall_kwargs["types"] = self._recall_types
                 logger.debug("Tool hindsight_recall: bank=%s, query_len=%d, budget=%s",
                              self._bank_id, len(query), self._budget)
-                resp = _run_sync(client.arecall(**recall_kwargs))
+                resp = self._run_sync(client.arecall(**recall_kwargs))
                 num_results = len(resp.results) if resp.results else 0
                 logger.debug("Tool hindsight_recall: %d results", num_results)
                 if not resp.results:
@@ -998,7 +1212,7 @@ class HindsightMemoryProvider(MemoryProvider):
             try:
                 logger.debug("Tool hindsight_reflect: bank=%s, query_len=%d, budget=%s",
                              self._bank_id, len(query), self._budget)
-                resp = _run_sync(client.areflect(
+                resp = self._run_sync(client.areflect(
                     bank_id=self._bank_id, query=query, budget=self._budget
                 ))
                 logger.debug("Tool hindsight_reflect: response_len=%d", len(resp.text or ""))
@@ -1011,7 +1225,6 @@ class HindsightMemoryProvider(MemoryProvider):
 
     def shutdown(self) -> None:
         logger.debug("Hindsight shutdown: waiting for background threads")
-        global _loop, _loop_thread
         for t in (self._prefetch_thread, self._sync_thread):
             if t and t.is_alive():
                 t.join(timeout=5.0)
@@ -1026,17 +1239,21 @@ class HindsightMemoryProvider(MemoryProvider):
                     except RuntimeError:
                         pass
                 else:
-                    _run_sync(self._client.aclose())
+                    self._run_sync(self._client.aclose())
             except Exception:
                 pass
             self._client = None
-        # Stop the background event loop so no tasks are pending at exit
-        if _loop is not None and _loop.is_running():
-            _loop.call_soon_threadsafe(_loop.stop)
-            if _loop_thread is not None:
-                _loop_thread.join(timeout=5.0)
-            _loop = None
-            _loop_thread = None
+        # The module-global background event loop (_loop / _loop_thread)
+        # is intentionally NOT stopped here. It is shared across every
+        # HindsightMemoryProvider instance in the process — the plugin
+        # loader creates a new provider per AIAgent, and the gateway
+        # creates one AIAgent per concurrent chat session. Stopping the
+        # loop from one provider's shutdown() strands the aiohttp
+        # ClientSession + TCPConnector owned by every sibling provider
+        # on a dead loop, which surfaces as the "Unclosed client session"
+        # / "Unclosed connector" warnings reported in #11923. The loop
+        # runs on a daemon thread and is reclaimed on process exit;
+        # per-session cleanup happens via self._client.aclose() above.
 
 
 def register(ctx) -> None:

plugins/spotify/__init__.py

@@ -0,0 +1,66 @@
+"""Spotify integration plugin — bundled, auto-loaded.
+
+Registers 7 tools (playback, devices, queue, search, playlists, albums,
+library) into the ``spotify`` toolset. Each tool's handler is gated by
+``_check_spotify_available()`` — when the user has not run ``hermes auth
+spotify``, the tools remain registered (so they appear in ``hermes
+tools``) but the runtime check prevents dispatch.
+
+Why a plugin instead of a top-level ``tools/`` file?
+
+- ``plugins/`` is where third-party service integrations live (see
+  ``plugins/image_gen/`` for the backend-provider pattern, ``plugins/
+  disk-cleanup/`` for the standalone pattern). ``tools/`` is reserved
+  for foundational capabilities (terminal, read_file, web_search, etc.).
+- Mirroring the image_gen plugin layout (``plugins/<category>/<backend>/``
+  for categories, flat ``plugins/<name>/`` for standalones) makes new
+  service integrations a pattern contributors can copy.
+- Bundled + ``kind: backend`` auto-loads on startup just like image_gen
+  backends — no user opt-in needed, no ``plugins.enabled`` config.
+
+The Spotify auth flow (``hermes auth spotify``), CLI plumbing, and docs
+are unchanged. This move is purely structural.
+"""
+
+from __future__ import annotations
+
+from plugins.spotify.tools import (
+    SPOTIFY_ALBUMS_SCHEMA,
+    SPOTIFY_DEVICES_SCHEMA,
+    SPOTIFY_LIBRARY_SCHEMA,
+    SPOTIFY_PLAYBACK_SCHEMA,
+    SPOTIFY_PLAYLISTS_SCHEMA,
+    SPOTIFY_QUEUE_SCHEMA,
+    SPOTIFY_SEARCH_SCHEMA,
+    _check_spotify_available,
+    _handle_spotify_albums,
+    _handle_spotify_devices,
+    _handle_spotify_library,
+    _handle_spotify_playback,
+    _handle_spotify_playlists,
+    _handle_spotify_queue,
+    _handle_spotify_search,
+)
+
+_TOOLS = (
+    ("spotify_playback",  SPOTIFY_PLAYBACK_SCHEMA,  _handle_spotify_playback,  "🎵"),
+    ("spotify_devices",   SPOTIFY_DEVICES_SCHEMA,   _handle_spotify_devices,   "🔈"),
+    ("spotify_queue",     SPOTIFY_QUEUE_SCHEMA,     _handle_spotify_queue,     "📻"),
+    ("spotify_search",    SPOTIFY_SEARCH_SCHEMA,    _handle_spotify_search,    "🔎"),
+    ("spotify_playlists", SPOTIFY_PLAYLISTS_SCHEMA, _handle_spotify_playlists, "📚"),
+    ("spotify_albums",    SPOTIFY_ALBUMS_SCHEMA,    _handle_spotify_albums,    "💿"),
+    ("spotify_library",   SPOTIFY_LIBRARY_SCHEMA,   _handle_spotify_library,   "❤️"),
+)
+
+
+def register(ctx) -> None:
+    """Register all Spotify tools. Called once by the plugin loader."""
+    for name, schema, handler, emoji in _TOOLS:
+        ctx.register_tool(
+            name=name,
+            toolset="spotify",
+            schema=schema,
+            handler=handler,
+            check_fn=_check_spotify_available,
+            emoji=emoji,
+        )

plugins/spotify/client.py

@@ -0,0 +1,435 @@
+"""Thin Spotify Web API helper used by Hermes native tools."""
+
+from __future__ import annotations
+
+import json
+from typing import Any, Dict, Iterable, Optional
+from urllib.parse import urlparse
+
+import httpx
+
+from hermes_cli.auth import (
+    AuthError,
+    resolve_spotify_runtime_credentials,
+)
+
+
+class SpotifyError(RuntimeError):
+    """Base Spotify tool error."""
+
+
+class SpotifyAuthRequiredError(SpotifyError):
+    """Raised when the user needs to authenticate with Spotify first."""
+
+
+class SpotifyAPIError(SpotifyError):
+    """Structured Spotify API failure."""
+
+    def __init__(
+        self,
+        message: str,
+        *,
+        status_code: Optional[int] = None,
+        response_body: Optional[str] = None,
+    ) -> None:
+        super().__init__(message)
+        self.status_code = status_code
+        self.response_body = response_body
+        self.path = None
+
+
+class SpotifyClient:
+    def __init__(self) -> None:
+        self._runtime = self._resolve_runtime(refresh_if_expiring=True)
+
+    def _resolve_runtime(self, *, force_refresh: bool = False, refresh_if_expiring: bool = True) -> Dict[str, Any]:
+        try:
+            return resolve_spotify_runtime_credentials(
+                force_refresh=force_refresh,
+                refresh_if_expiring=refresh_if_expiring,
+            )
+        except AuthError as exc:
+            raise SpotifyAuthRequiredError(str(exc)) from exc
+
+    @property
+    def base_url(self) -> str:
+        return str(self._runtime.get("base_url") or "").rstrip("/")
+
+    def _headers(self) -> Dict[str, str]:
+        return {
+            "Authorization": f"Bearer {self._runtime['access_token']}",
+            "Content-Type": "application/json",
+        }
+
+    def request(
+        self,
+        method: str,
+        path: str,
+        *,
+        params: Optional[Dict[str, Any]] = None,
+        json_body: Optional[Dict[str, Any]] = None,
+        allow_retry_on_401: bool = True,
+        empty_response: Optional[Dict[str, Any]] = None,
+    ) -> Any:
+        url = f"{self.base_url}{path}"
+        response = httpx.request(
+            method,
+            url,
+            headers=self._headers(),
+            params=_strip_none(params),
+            json=_strip_none(json_body) if json_body is not None else None,
+            timeout=30.0,
+        )
+        if response.status_code == 401 and allow_retry_on_401:
+            self._runtime = self._resolve_runtime(force_refresh=True, refresh_if_expiring=True)
+            return self.request(
+                method,
+                path,
+                params=params,
+                json_body=json_body,
+                allow_retry_on_401=False,
+            )
+        if response.status_code >= 400:
+            self._raise_api_error(response, method=method, path=path)
+        if response.status_code == 204 or not response.content:
+            return empty_response or {"success": True, "status_code": response.status_code, "empty": True}
+        if "application/json" in response.headers.get("content-type", ""):
+            return response.json()
+        return {"success": True, "text": response.text}
+
+    def _raise_api_error(self, response: httpx.Response, *, method: str, path: str) -> None:
+        detail = response.text.strip()
+        message = _friendly_spotify_error_message(
+            status_code=response.status_code,
+            detail=_extract_spotify_error_detail(response, fallback=detail),
+            method=method,
+            path=path,
+            retry_after=response.headers.get("Retry-After"),
+        )
+        error = SpotifyAPIError(message, status_code=response.status_code, response_body=detail)
+        error.path = path
+        raise error
+
+    def get_devices(self) -> Any:
+        return self.request("GET", "/me/player/devices")
+
+    def transfer_playback(self, *, device_id: str, play: bool = False) -> Any:
+        return self.request("PUT", "/me/player", json_body={
+            "device_ids": [device_id],
+            "play": play,
+        })
+
+    def get_playback_state(self, *, market: Optional[str] = None) -> Any:
+        return self.request(
+            "GET",
+            "/me/player",
+            params={"market": market},
+            empty_response={
+                "status_code": 204,
+                "empty": True,
+                "message": "No active Spotify playback session was found. Open Spotify on a device and start playback, or transfer playback to an available device.",
+            },
+        )
+
+    def get_currently_playing(self, *, market: Optional[str] = None) -> Any:
+        return self.request(
+            "GET",
+            "/me/player/currently-playing",
+            params={"market": market},
+            empty_response={
+                "status_code": 204,
+                "empty": True,
+                "message": "Spotify is not currently playing anything. Start playback in Spotify and try again.",
+            },
+        )
+
+    def start_playback(
+        self,
+        *,
+        device_id: Optional[str] = None,
+        context_uri: Optional[str] = None,
+        uris: Optional[list[str]] = None,
+        offset: Optional[Dict[str, Any]] = None,
+        position_ms: Optional[int] = None,
+    ) -> Any:
+        return self.request(
+            "PUT",
+            "/me/player/play",
+            params={"device_id": device_id},
+            json_body={
+                "context_uri": context_uri,
+                "uris": uris,
+                "offset": offset,
+                "position_ms": position_ms,
+            },
+        )
+
+    def pause_playback(self, *, device_id: Optional[str] = None) -> Any:
+        return self.request("PUT", "/me/player/pause", params={"device_id": device_id})
+
+    def skip_next(self, *, device_id: Optional[str] = None) -> Any:
+        return self.request("POST", "/me/player/next", params={"device_id": device_id})
+
+    def skip_previous(self, *, device_id: Optional[str] = None) -> Any:
+        return self.request("POST", "/me/player/previous", params={"device_id": device_id})
+
+    def seek(self, *, position_ms: int, device_id: Optional[str] = None) -> Any:
+        return self.request("PUT", "/me/player/seek", params={
+            "position_ms": position_ms,
+            "device_id": device_id,
+        })
+
+    def set_repeat(self, *, state: str, device_id: Optional[str] = None) -> Any:
+        return self.request("PUT", "/me/player/repeat", params={"state": state, "device_id": device_id})
+
+    def set_shuffle(self, *, state: bool, device_id: Optional[str] = None) -> Any:
+        return self.request("PUT", "/me/player/shuffle", params={"state": str(bool(state)).lower(), "device_id": device_id})
+
+    def set_volume(self, *, volume_percent: int, device_id: Optional[str] = None) -> Any:
+        return self.request("PUT", "/me/player/volume", params={
+            "volume_percent": volume_percent,
+            "device_id": device_id,
+        })
+
+    def get_queue(self) -> Any:
+        return self.request("GET", "/me/player/queue")
+
+    def add_to_queue(self, *, uri: str, device_id: Optional[str] = None) -> Any:
+        return self.request("POST", "/me/player/queue", params={"uri": uri, "device_id": device_id})
+
+    def search(
+        self,
+        *,
+        query: str,
+        search_types: list[str],
+        limit: int = 10,
+        offset: int = 0,
+        market: Optional[str] = None,
+        include_external: Optional[str] = None,
+    ) -> Any:
+        return self.request("GET", "/search", params={
+            "q": query,
+            "type": ",".join(search_types),
+            "limit": limit,
+            "offset": offset,
+            "market": market,
+            "include_external": include_external,
+        })
+
+    def get_my_playlists(self, *, limit: int = 20, offset: int = 0) -> Any:
+        return self.request("GET", "/me/playlists", params={"limit": limit, "offset": offset})
+
+    def get_playlist(self, *, playlist_id: str, market: Optional[str] = None) -> Any:
+        return self.request("GET", f"/playlists/{playlist_id}", params={"market": market})
+
+    def create_playlist(
+        self,
+        *,
+        name: str,
+        public: bool = False,
+        collaborative: bool = False,
+        description: Optional[str] = None,
+    ) -> Any:
+        return self.request("POST", "/me/playlists", json_body={
+            "name": name,
+            "public": public,
+            "collaborative": collaborative,
+            "description": description,
+        })
+
+    def add_playlist_items(
+        self,
+        *,
+        playlist_id: str,
+        uris: list[str],
+        position: Optional[int] = None,
+    ) -> Any:
+        return self.request("POST", f"/playlists/{playlist_id}/items", json_body={
+            "uris": uris,
+            "position": position,
+        })
+
+    def remove_playlist_items(
+        self,
+        *,
+        playlist_id: str,
+        uris: list[str],
+        snapshot_id: Optional[str] = None,
+    ) -> Any:
+        return self.request("DELETE", f"/playlists/{playlist_id}/items", json_body={
+            "items": [{"uri": uri} for uri in uris],
+            "snapshot_id": snapshot_id,
+        })
+
+    def update_playlist_details(
+        self,
+        *,
+        playlist_id: str,
+        name: Optional[str] = None,
+        public: Optional[bool] = None,
+        collaborative: Optional[bool] = None,
+        description: Optional[str] = None,
+    ) -> Any:
+        return self.request("PUT", f"/playlists/{playlist_id}", json_body={
+            "name": name,
+            "public": public,
+            "collaborative": collaborative,
+            "description": description,
+        })
+
+    def get_album(self, *, album_id: str, market: Optional[str] = None) -> Any:
+        return self.request("GET", f"/albums/{album_id}", params={"market": market})
+
+    def get_album_tracks(self, *, album_id: str, limit: int = 20, offset: int = 0, market: Optional[str] = None) -> Any:
+        return self.request("GET", f"/albums/{album_id}/tracks", params={
+            "limit": limit,
+            "offset": offset,
+            "market": market,
+        })
+
+    def get_saved_tracks(self, *, limit: int = 20, offset: int = 0, market: Optional[str] = None) -> Any:
+        return self.request("GET", "/me/tracks", params={"limit": limit, "offset": offset, "market": market})
+
+    def save_library_items(self, *, uris: list[str]) -> Any:
+        return self.request("PUT", "/me/library", params={"uris": ",".join(uris)})
+
+    def library_contains(self, *, uris: list[str]) -> Any:
+        return self.request("GET", "/me/library/contains", params={"uris": ",".join(uris)})
+
+    def get_saved_albums(self, *, limit: int = 20, offset: int = 0, market: Optional[str] = None) -> Any:
+        return self.request("GET", "/me/albums", params={"limit": limit, "offset": offset, "market": market})
+
+    def remove_saved_tracks(self, *, track_ids: list[str]) -> Any:
+        uris = [f"spotify:track:{track_id}" for track_id in track_ids]
+        return self.request("DELETE", "/me/library", params={"uris": ",".join(uris)})
+
+    def remove_saved_albums(self, *, album_ids: list[str]) -> Any:
+        uris = [f"spotify:album:{album_id}" for album_id in album_ids]
+        return self.request("DELETE", "/me/library", params={"uris": ",".join(uris)})
+
+    def get_recently_played(
+        self,
+        *,
+        limit: int = 20,
+        after: Optional[int] = None,
+        before: Optional[int] = None,
+    ) -> Any:
+        return self.request("GET", "/me/player/recently-played", params={
+            "limit": limit,
+            "after": after,
+            "before": before,
+        })
+
+
+def _extract_spotify_error_detail(response: httpx.Response, *, fallback: str) -> str:
+    detail = fallback
+    try:
+        payload = response.json()
+        if isinstance(payload, dict):
+            error_obj = payload.get("error")
+            if isinstance(error_obj, dict):
+                detail = str(error_obj.get("message") or detail)
+            elif isinstance(error_obj, str):
+                detail = error_obj
+    except Exception:
+        pass
+    return detail.strip()
+
+
+def _friendly_spotify_error_message(
+    *,
+    status_code: int,
+    detail: str,
+    method: str,
+    path: str,
+    retry_after: Optional[str],
+) -> str:
+    normalized_detail = detail.lower()
+    is_playback_path = path.startswith("/me/player")
+
+    if status_code == 401:
+        return "Spotify authentication failed or expired. Run `hermes auth spotify` again."
+
+    if status_code == 403:
+        if is_playback_path:
+            return (
+                "Spotify rejected this playback request. Playback control usually requires a Spotify Premium account "
+                "and an active Spotify Connect device."
+            )
+        if "scope" in normalized_detail or "permission" in normalized_detail:
+            return "Spotify rejected the request because the current auth scope is insufficient. Re-run `hermes auth spotify` to refresh permissions."
+        return "Spotify rejected the request. The account may not have permission for this action."
+
+    if status_code == 404:
+        if is_playback_path:
+            return "Spotify could not find an active playback device or player session for this request."
+        return "Spotify resource not found."
+
+    if status_code == 429:
+        message = "Spotify rate limit exceeded."
+        if retry_after:
+            message += f" Retry after {retry_after} seconds."
+        return message
+
+    if detail:
+        return detail
+    return f"Spotify API request failed with status {status_code}."
+
+
+def _strip_none(payload: Optional[Dict[str, Any]]) -> Dict[str, Any]:
+    if not payload:
+        return {}
+    return {key: value for key, value in payload.items() if value is not None}
+
+
+def normalize_spotify_id(value: str, expected_type: Optional[str] = None) -> str:
+    cleaned = (value or "").strip()
+    if not cleaned:
+        raise SpotifyError("Spotify id/uri/url is required.")
+    if cleaned.startswith("spotify:"):
+        parts = cleaned.split(":")
+        if len(parts) >= 3:
+            item_type = parts[1]
+            if expected_type and item_type != expected_type:
+                raise SpotifyError(f"Expected a Spotify {expected_type}, got {item_type}.")
+            return parts[2]
+    if "open.spotify.com" in cleaned:
+        parsed = urlparse(cleaned)
+        path_parts = [part for part in parsed.path.split("/") if part]
+        if len(path_parts) >= 2:
+            item_type, item_id = path_parts[0], path_parts[1]
+            if expected_type and item_type != expected_type:
+                raise SpotifyError(f"Expected a Spotify {expected_type}, got {item_type}.")
+            return item_id
+    return cleaned
+
+
+def normalize_spotify_uri(value: str, expected_type: Optional[str] = None) -> str:
+    cleaned = (value or "").strip()
+    if not cleaned:
+        raise SpotifyError("Spotify URI/url/id is required.")
+    if cleaned.startswith("spotify:"):
+        if expected_type:
+            parts = cleaned.split(":")
+            if len(parts) >= 3 and parts[1] != expected_type:
+                raise SpotifyError(f"Expected a Spotify {expected_type}, got {parts[1]}.")
+        return cleaned
+    item_id = normalize_spotify_id(cleaned, expected_type)
+    if expected_type:
+        return f"spotify:{expected_type}:{item_id}"
+    return cleaned
+
+
+def normalize_spotify_uris(values: Iterable[str], expected_type: Optional[str] = None) -> list[str]:
+    uris: list[str] = []
+    for value in values:
+        uri = normalize_spotify_uri(str(value), expected_type)
+        if uri not in uris:
+            uris.append(uri)
+    if not uris:
+        raise SpotifyError("At least one Spotify item is required.")
+    return uris
+
+
+def compact_json(data: Any) -> str:
+    return json.dumps(data, ensure_ascii=False)

plugins/spotify/plugin.yaml

@@ -0,0 +1,13 @@
+name: spotify
+version: 1.0.0
+description: "Native Spotify integration — 7 tools (playback, devices, queue, search, playlists, albums, library) using Spotify Web API + PKCE OAuth. Auth via `hermes auth spotify`. Tools gate on `providers.spotify` in ~/.hermes/auth.json."
+author: NousResearch
+kind: backend
+provides_tools:
+  - spotify_playback
+  - spotify_devices
+  - spotify_queue
+  - spotify_search
+  - spotify_playlists
+  - spotify_albums
+  - spotify_library

plugins/spotify/tools.py

@@ -0,0 +1,454 @@
+"""Native Spotify tools for Hermes (registered via plugins/spotify)."""
+
+from __future__ import annotations
+
+from typing import Any, Dict, List
+
+from hermes_cli.auth import get_auth_status
+from plugins.spotify.client import (
+    SpotifyAPIError,
+    SpotifyAuthRequiredError,
+    SpotifyClient,
+    SpotifyError,
+    normalize_spotify_id,
+    normalize_spotify_uri,
+    normalize_spotify_uris,
+)
+from tools.registry import tool_error, tool_result
+
+
+def _check_spotify_available() -> bool:
+    try:
+        return bool(get_auth_status("spotify").get("logged_in"))
+    except Exception:
+        return False
+
+
+def _spotify_client() -> SpotifyClient:
+    return SpotifyClient()
+
+
+def _spotify_tool_error(exc: Exception) -> str:
+    if isinstance(exc, (SpotifyError, SpotifyAuthRequiredError)):
+        return tool_error(str(exc))
+    if isinstance(exc, SpotifyAPIError):
+        return tool_error(str(exc), status_code=exc.status_code)
+    return tool_error(f"Spotify tool failed: {type(exc).__name__}: {exc}")
+
+
+def _coerce_limit(raw: Any, *, default: int = 20, minimum: int = 1, maximum: int = 50) -> int:
+    try:
+        value = int(raw)
+    except Exception:
+        value = default
+    return max(minimum, min(maximum, value))
+
+
+def _coerce_bool(raw: Any, default: bool = False) -> bool:
+    if isinstance(raw, bool):
+        return raw
+    if isinstance(raw, str):
+        cleaned = raw.strip().lower()
+        if cleaned in {"1", "true", "yes", "on"}:
+            return True
+        if cleaned in {"0", "false", "no", "off"}:
+            return False
+    return default
+
+
+def _as_list(raw: Any) -> List[str]:
+    if raw is None:
+        return []
+    if isinstance(raw, list):
+        return [str(item).strip() for item in raw if str(item).strip()]
+    return [str(raw).strip()] if str(raw).strip() else []
+
+
+def _describe_empty_playback(payload: Any, *, action: str) -> dict | None:
+    if not isinstance(payload, dict) or not payload.get("empty"):
+        return None
+    if action == "get_currently_playing":
+        return {
+            "success": True,
+            "action": action,
+            "is_playing": False,
+            "status_code": payload.get("status_code", 204),
+            "message": payload.get("message") or "Spotify is not currently playing anything.",
+        }
+    if action == "get_state":
+        return {
+            "success": True,
+            "action": action,
+            "has_active_device": False,
+            "status_code": payload.get("status_code", 204),
+            "message": payload.get("message") or "No active Spotify playback session was found.",
+        }
+    return None
+
+
+def _handle_spotify_playback(args: dict, **kw) -> str:
+    action = str(args.get("action") or "get_state").strip().lower()
+    client = _spotify_client()
+    try:
+        if action == "get_state":
+            payload = client.get_playback_state(market=args.get("market"))
+            empty_result = _describe_empty_playback(payload, action=action)
+            return tool_result(empty_result or payload)
+        if action == "get_currently_playing":
+            payload = client.get_currently_playing(market=args.get("market"))
+            empty_result = _describe_empty_playback(payload, action=action)
+            return tool_result(empty_result or payload)
+        if action == "play":
+            offset = args.get("offset")
+            if isinstance(offset, dict):
+                payload_offset = {k: v for k, v in offset.items() if v is not None}
+            else:
+                payload_offset = None
+            uris = normalize_spotify_uris(_as_list(args.get("uris")), "track") if args.get("uris") else None
+            context_uri = None
+            if args.get("context_uri"):
+                raw_context = str(args.get("context_uri"))
+                context_type = None
+                if raw_context.startswith("spotify:album:") or "/album/" in raw_context:
+                    context_type = "album"
+                elif raw_context.startswith("spotify:playlist:") or "/playlist/" in raw_context:
+                    context_type = "playlist"
+                elif raw_context.startswith("spotify:artist:") or "/artist/" in raw_context:
+                    context_type = "artist"
+                context_uri = normalize_spotify_uri(raw_context, context_type)
+            result = client.start_playback(
+                device_id=args.get("device_id"),
+                context_uri=context_uri,
+                uris=uris,
+                offset=payload_offset,
+                position_ms=args.get("position_ms"),
+            )
+            return tool_result({"success": True, "action": action, "result": result})
+        if action == "pause":
+            result = client.pause_playback(device_id=args.get("device_id"))
+            return tool_result({"success": True, "action": action, "result": result})
+        if action == "next":
+            result = client.skip_next(device_id=args.get("device_id"))
+            return tool_result({"success": True, "action": action, "result": result})
+        if action == "previous":
+            result = client.skip_previous(device_id=args.get("device_id"))
+            return tool_result({"success": True, "action": action, "result": result})
+        if action == "seek":
+            if args.get("position_ms") is None:
+                return tool_error("position_ms is required for action='seek'")
+            result = client.seek(position_ms=int(args["position_ms"]), device_id=args.get("device_id"))
+            return tool_result({"success": True, "action": action, "result": result})
+        if action == "set_repeat":
+            state = str(args.get("state") or "").strip().lower()
+            if state not in {"track", "context", "off"}:
+                return tool_error("state must be one of: track, context, off")
+            result = client.set_repeat(state=state, device_id=args.get("device_id"))
+            return tool_result({"success": True, "action": action, "result": result})
+        if action == "set_shuffle":
+            result = client.set_shuffle(state=_coerce_bool(args.get("state")), device_id=args.get("device_id"))
+            return tool_result({"success": True, "action": action, "result": result})
+        if action == "set_volume":
+            if args.get("volume_percent") is None:
+                return tool_error("volume_percent is required for action='set_volume'")
+            result = client.set_volume(volume_percent=max(0, min(100, int(args["volume_percent"]))), device_id=args.get("device_id"))
+            return tool_result({"success": True, "action": action, "result": result})
+        if action == "recently_played":
+            after = args.get("after")
+            before = args.get("before")
+            if after and before:
+                return tool_error("Provide only one of 'after' or 'before'")
+            return tool_result(client.get_recently_played(
+                limit=_coerce_limit(args.get("limit"), default=20),
+                after=int(after) if after is not None else None,
+                before=int(before) if before is not None else None,
+            ))
+        return tool_error(f"Unknown spotify_playback action: {action}")
+    except Exception as exc:
+        return _spotify_tool_error(exc)
+
+
+def _handle_spotify_devices(args: dict, **kw) -> str:
+    action = str(args.get("action") or "list").strip().lower()
+    client = _spotify_client()
+    try:
+        if action == "list":
+            return tool_result(client.get_devices())
+        if action == "transfer":
+            device_id = str(args.get("device_id") or "").strip()
+            if not device_id:
+                return tool_error("device_id is required for action='transfer'")
+            result = client.transfer_playback(device_id=device_id, play=_coerce_bool(args.get("play")))
+            return tool_result({"success": True, "action": action, "result": result})
+        return tool_error(f"Unknown spotify_devices action: {action}")
+    except Exception as exc:
+        return _spotify_tool_error(exc)
+
+
+def _handle_spotify_queue(args: dict, **kw) -> str:
+    action = str(args.get("action") or "get").strip().lower()
+    client = _spotify_client()
+    try:
+        if action == "get":
+            return tool_result(client.get_queue())
+        if action == "add":
+            uri = normalize_spotify_uri(str(args.get("uri") or ""), None)
+            result = client.add_to_queue(uri=uri, device_id=args.get("device_id"))
+            return tool_result({"success": True, "action": action, "uri": uri, "result": result})
+        return tool_error(f"Unknown spotify_queue action: {action}")
+    except Exception as exc:
+        return _spotify_tool_error(exc)
+
+
+def _handle_spotify_search(args: dict, **kw) -> str:
+    client = _spotify_client()
+    query = str(args.get("query") or "").strip()
+    if not query:
+        return tool_error("query is required")
+    raw_types = _as_list(args.get("types") or args.get("type") or ["track"])
+    search_types = [value.lower() for value in raw_types if value.lower() in {"album", "artist", "playlist", "track", "show", "episode", "audiobook"}]
+    if not search_types:
+        return tool_error("types must contain one or more of: album, artist, playlist, track, show, episode, audiobook")
+    try:
+        return tool_result(client.search(
+            query=query,
+            search_types=search_types,
+            limit=_coerce_limit(args.get("limit"), default=10),
+            offset=max(0, int(args.get("offset") or 0)),
+            market=args.get("market"),
+            include_external=args.get("include_external"),
+        ))
+    except Exception as exc:
+        return _spotify_tool_error(exc)
+
+
+def _handle_spotify_playlists(args: dict, **kw) -> str:
+    action = str(args.get("action") or "list").strip().lower()
+    client = _spotify_client()
+    try:
+        if action == "list":
+            return tool_result(client.get_my_playlists(
+                limit=_coerce_limit(args.get("limit"), default=20),
+                offset=max(0, int(args.get("offset") or 0)),
+            ))
+        if action == "get":
+            playlist_id = normalize_spotify_id(str(args.get("playlist_id") or ""), "playlist")
+            return tool_result(client.get_playlist(playlist_id=playlist_id, market=args.get("market")))
+        if action == "create":
+            name = str(args.get("name") or "").strip()
+            if not name:
+                return tool_error("name is required for action='create'")
+            return tool_result(client.create_playlist(
+                name=name,
+                public=_coerce_bool(args.get("public")),
+                collaborative=_coerce_bool(args.get("collaborative")),
+                description=args.get("description"),
+            ))
+        if action == "add_items":
+            playlist_id = normalize_spotify_id(str(args.get("playlist_id") or ""), "playlist")
+            uris = normalize_spotify_uris(_as_list(args.get("uris")))
+            return tool_result(client.add_playlist_items(
+                playlist_id=playlist_id,
+                uris=uris,
+                position=args.get("position"),
+            ))
+        if action == "remove_items":
+            playlist_id = normalize_spotify_id(str(args.get("playlist_id") or ""), "playlist")
+            uris = normalize_spotify_uris(_as_list(args.get("uris")))
+            return tool_result(client.remove_playlist_items(
+                playlist_id=playlist_id,
+                uris=uris,
+                snapshot_id=args.get("snapshot_id"),
+            ))
+        if action == "update_details":
+            playlist_id = normalize_spotify_id(str(args.get("playlist_id") or ""), "playlist")
+            return tool_result(client.update_playlist_details(
+                playlist_id=playlist_id,
+                name=args.get("name"),
+                public=args.get("public"),
+                collaborative=args.get("collaborative"),
+                description=args.get("description"),
+            ))
+        return tool_error(f"Unknown spotify_playlists action: {action}")
+    except Exception as exc:
+        return _spotify_tool_error(exc)
+
+
+def _handle_spotify_albums(args: dict, **kw) -> str:
+    action = str(args.get("action") or "get").strip().lower()
+    client = _spotify_client()
+    try:
+        album_id = normalize_spotify_id(str(args.get("album_id") or args.get("id") or ""), "album")
+        if action == "get":
+            return tool_result(client.get_album(album_id=album_id, market=args.get("market")))
+        if action == "tracks":
+            return tool_result(client.get_album_tracks(
+                album_id=album_id,
+                limit=_coerce_limit(args.get("limit"), default=20),
+                offset=max(0, int(args.get("offset") or 0)),
+                market=args.get("market"),
+            ))
+        return tool_error(f"Unknown spotify_albums action: {action}")
+    except Exception as exc:
+        return _spotify_tool_error(exc)
+
+
+def _handle_spotify_library(args: dict, **kw) -> str:
+    """Unified handler for saved tracks + saved albums (formerly two tools)."""
+    kind = str(args.get("kind") or "").strip().lower()
+    if kind not in {"tracks", "albums"}:
+        return tool_error("kind must be one of: tracks, albums")
+    action = str(args.get("action") or "list").strip().lower()
+    item_type = "track" if kind == "tracks" else "album"
+    client = _spotify_client()
+    try:
+        if action == "list":
+            limit = _coerce_limit(args.get("limit"), default=20)
+            offset = max(0, int(args.get("offset") or 0))
+            market = args.get("market")
+            if kind == "tracks":
+                return tool_result(client.get_saved_tracks(limit=limit, offset=offset, market=market))
+            return tool_result(client.get_saved_albums(limit=limit, offset=offset, market=market))
+        if action == "save":
+            uris = normalize_spotify_uris(_as_list(args.get("uris") or args.get("items")), item_type)
+            return tool_result(client.save_library_items(uris=uris))
+        if action == "remove":
+            ids = [normalize_spotify_id(item, item_type) for item in _as_list(args.get("ids") or args.get("items"))]
+            if not ids:
+                return tool_error("ids/items is required for action='remove'")
+            if kind == "tracks":
+                return tool_result(client.remove_saved_tracks(track_ids=ids))
+            return tool_result(client.remove_saved_albums(album_ids=ids))
+        return tool_error(f"Unknown spotify_library action: {action}")
+    except Exception as exc:
+        return _spotify_tool_error(exc)
+
+
+COMMON_STRING = {"type": "string"}
+
+SPOTIFY_PLAYBACK_SCHEMA = {
+    "name": "spotify_playback",
+    "description": "Control Spotify playback, inspect the active playback state, or fetch recently played tracks.",
+    "parameters": {
+        "type": "object",
+        "properties": {
+            "action": {"type": "string", "enum": ["get_state", "get_currently_playing", "play", "pause", "next", "previous", "seek", "set_repeat", "set_shuffle", "set_volume", "recently_played"]},
+            "device_id": COMMON_STRING,
+            "market": COMMON_STRING,
+            "context_uri": COMMON_STRING,
+            "uris": {"type": "array", "items": COMMON_STRING},
+            "offset": {"type": "object"},
+            "position_ms": {"type": "integer"},
+            "state": {"description": "For set_repeat use track/context/off. For set_shuffle use boolean-like true/false.", "oneOf": [{"type": "string"}, {"type": "boolean"}]},
+            "volume_percent": {"type": "integer"},
+            "limit": {"type": "integer", "description": "For recently_played: number of tracks (max 50)"},
+            "after": {"type": "integer", "description": "For recently_played: Unix ms cursor (after this timestamp)"},
+            "before": {"type": "integer", "description": "For recently_played: Unix ms cursor (before this timestamp)"},
+        },
+        "required": ["action"],
+    },
+}
+
+SPOTIFY_DEVICES_SCHEMA = {
+    "name": "spotify_devices",
+    "description": "List Spotify Connect devices or transfer playback to a different device.",
+    "parameters": {
+        "type": "object",
+        "properties": {
+            "action": {"type": "string", "enum": ["list", "transfer"]},
+            "device_id": COMMON_STRING,
+            "play": {"type": "boolean"},
+        },
+        "required": ["action"],
+    },
+}
+
+SPOTIFY_QUEUE_SCHEMA = {
+    "name": "spotify_queue",
+    "description": "Inspect the user's Spotify queue or add an item to it.",
+    "parameters": {
+        "type": "object",
+        "properties": {
+            "action": {"type": "string", "enum": ["get", "add"]},
+            "uri": COMMON_STRING,
+            "device_id": COMMON_STRING,
+        },
+        "required": ["action"],
+    },
+}
+
+SPOTIFY_SEARCH_SCHEMA = {
+    "name": "spotify_search",
+    "description": "Search the Spotify catalog for tracks, albums, artists, playlists, shows, or episodes.",
+    "parameters": {
+        "type": "object",
+        "properties": {
+            "query": COMMON_STRING,
+            "types": {"type": "array", "items": COMMON_STRING},
+            "type": COMMON_STRING,
+            "limit": {"type": "integer"},
+            "offset": {"type": "integer"},
+            "market": COMMON_STRING,
+            "include_external": COMMON_STRING,
+        },
+        "required": ["query"],
+    },
+}
+
+SPOTIFY_PLAYLISTS_SCHEMA = {
+    "name": "spotify_playlists",
+    "description": "List, inspect, create, update, and modify Spotify playlists.",
+    "parameters": {
+        "type": "object",
+        "properties": {
+            "action": {"type": "string", "enum": ["list", "get", "create", "add_items", "remove_items", "update_details"]},
+            "playlist_id": COMMON_STRING,
+            "market": COMMON_STRING,
+            "limit": {"type": "integer"},
+            "offset": {"type": "integer"},
+            "name": COMMON_STRING,
+            "description": COMMON_STRING,
+            "public": {"type": "boolean"},
+            "collaborative": {"type": "boolean"},
+            "uris": {"type": "array", "items": COMMON_STRING},
+            "position": {"type": "integer"},
+            "snapshot_id": COMMON_STRING,
+        },
+        "required": ["action"],
+    },
+}
+
+SPOTIFY_ALBUMS_SCHEMA = {
+    "name": "spotify_albums",
+    "description": "Fetch Spotify album metadata or album tracks.",
+    "parameters": {
+        "type": "object",
+        "properties": {
+            "action": {"type": "string", "enum": ["get", "tracks"]},
+            "album_id": COMMON_STRING,
+            "id": COMMON_STRING,
+            "market": COMMON_STRING,
+            "limit": {"type": "integer"},
+            "offset": {"type": "integer"},
+        },
+        "required": ["action"],
+    },
+}
+
+SPOTIFY_LIBRARY_SCHEMA = {
+    "name": "spotify_library",
+    "description": "List, save, or remove the user's saved Spotify tracks or albums. Use `kind` to select which.",
+    "parameters": {
+        "type": "object",
+        "properties": {
+            "kind": {"type": "string", "enum": ["tracks", "albums"], "description": "Which library to operate on"},
+            "action": {"type": "string", "enum": ["list", "save", "remove"]},
+            "limit": {"type": "integer"},
+            "offset": {"type": "integer"},
+            "market": COMMON_STRING,
+            "uris": {"type": "array", "items": COMMON_STRING},
+            "ids": {"type": "array", "items": COMMON_STRING},
+            "items": {"type": "array", "items": COMMON_STRING},
+        },
+        "required": ["kind", "action"],
+    },
+}

plugins/strike-freedom-cockpit/README.md

@@ -0,0 +1,70 @@
+# Strike Freedom Cockpit — dashboard skin demo
+
+Demonstrates how the dashboard skin+plugin system can be used to build a
+fully custom cockpit-style reskin without touching the core dashboard.
+
+Two pieces:
+
+- `theme/strike-freedom.yaml` — a dashboard theme YAML that paints the
+  palette, typography, layout variant (`cockpit`), component chrome
+  (notched card corners, scanlines, accent colors), and declares asset
+  slots (`hero`, `crest`, `bg`).
+- `dashboard/` — a plugin that populates the `sidebar`, `header-left`,
+  and `footer-right` slots reserved by the cockpit layout. The sidebar
+  renders an MS-STATUS panel with segmented telemetry bars driven by
+  real agent status; the header-left injects a COMPASS crest; the
+  footer-right replaces the default org tagline.
+
+## Install
+
+1. **Theme** — copy the theme YAML into your Hermes home:
+
+   ```
+   cp theme/strike-freedom.yaml ~/.hermes/dashboard-themes/
+   ```
+
+2. **Plugin** — the `dashboard/` directory gets auto-discovered because
+   it lives under `plugins/` in the repo. On a user install, copy the
+   whole plugin directory into `~/.hermes/plugins/`:
+
+   ```
+   cp -r . ~/.hermes/plugins/strike-freedom-cockpit
+   ```
+
+3. Restart the web UI (or `GET /api/dashboard/plugins/rescan`), open it,
+   pick **Strike Freedom** from the theme switcher.
+
+## Customising the artwork
+
+The sidebar plugin reads `--theme-asset-hero` and `--theme-asset-crest`
+from the active theme. Drop your own URLs into the theme YAML:
+
+```yaml
+assets:
+  hero: "/my-images/strike-freedom.png"
+  crest: "/my-images/compass-crest.svg"
+  bg: "/my-images/cosmic-era-bg.jpg"
+```
+
+The plugin reads those at render time — no plugin code changes needed
+to swap artwork across themes.
+
+## What this demo proves
+
+The dashboard skin+plugin system supports (ref: `web/src/themes/types.ts`,
+`web/src/plugins/slots.ts`):
+
+- Palette, typography, font URLs, density, radius — already present
+- **Asset URLs exposed as CSS vars** (bg / hero / crest / logo /
+  sidebar / header + arbitrary `custom.*`)
+- **Raw `customCSS` blocks** injected as scoped `<style>` tags
+- **Per-component style overrides** (card / header / sidebar / backdrop /
+  tab / progress / footer / badge / page) via CSS vars
+- **`layoutVariant`** — `standard`, `cockpit`, or `tiled`
+- **Plugin slots** — 10 named shell slots plugins can inject into
+  (`backdrop`, `header-left/right/banner`, `sidebar`, `pre-main`,
+  `post-main`, `footer-left/right`, `overlay`)
+- **Route overrides** — plugins can replace a built-in page entirely
+  (`tab.override: "/"`) instead of just adding a tab
+- **Hidden plugins** — slot-only plugins that never show in the nav
+  (`tab.hidden: true`) — as used here

plugins/strike-freedom-cockpit/dashboard/dist/index.js

@@ -0,0 +1,309 @@
+/**
+ * Strike Freedom Cockpit — dashboard plugin demo.
+ *
+ * A slot-only plugin (manifest sets tab.hidden: true) that populates
+ * three shell slots when the user has the ``strike-freedom`` theme
+ * selected (or any theme that picks layoutVariant: cockpit):
+ *
+ *   - sidebar       → MS-STATUS panel: ENERGY / SHIELD / POWER bars,
+ *                     ZGMF-X20A identity line, pilot block, hero
+ *                     render (from --theme-asset-hero when the theme
+ *                     provides one).
+ *   - header-left   → COMPASS faction crest (uses --theme-asset-crest
+ *                     if provided, falls back to a geometric SVG).
+ *   - footer-right  → COSMIC ERA tagline that replaces the default
+ *                     footer org line.
+ *
+ * The plugin demonstrates every extension point added alongside the
+ * slot system: registerSlot, tab.hidden, reading theme asset CSS vars
+ * from plugin code, and rendering above the built-in route content.
+ */
+(function () {
+  "use strict";
+
+  const SDK = window.__HERMES_PLUGIN_SDK__;
+  const PLUGINS = window.__HERMES_PLUGINS__;
+  if (!SDK || !PLUGINS || !PLUGINS.registerSlot) {
+    // Old dashboard bundle without slot support — bail silently rather
+    // than breaking the page.
+    return;
+  }
+
+  const { React } = SDK;
+  const { useState, useEffect } = SDK.hooks;
+  const { api } = SDK;
+
+  // ---------------------------------------------------------------------
+  // Helpers
+  // ---------------------------------------------------------------------
+
+  /** Read a CSS custom property from :root. Empty string when unset. */
+  function cssVar(name) {
+    if (typeof document === "undefined") return "";
+    return getComputedStyle(document.documentElement).getPropertyValue(name).trim();
+  }
+
+  /** Segmented chip progress bar — 10 cells filled proportionally to value. */
+  function TelemetryBar(props) {
+    const { label, value, color } = props;
+    const cells = [];
+    for (let i = 0; i < 10; i++) {
+      const filled = Math.round(value / 10) > i;
+      cells.push(
+        React.createElement("span", {
+          key: i,
+          style: {
+            flex: 1,
+            height: 8,
+            background: filled ? color : "rgba(255,255,255,0.06)",
+            transition: "background 200ms",
+            clipPath: "polygon(2px 0, 100% 0, calc(100% - 2px) 100%, 0 100%)",
+          },
+        }),
+      );
+    }
+    return React.createElement(
+      "div",
+      { style: { display: "flex", flexDirection: "column", gap: 4 } },
+      React.createElement(
+        "div",
+        {
+          style: {
+            display: "flex",
+            justifyContent: "space-between",
+            fontSize: "0.65rem",
+            letterSpacing: "0.12em",
+            opacity: 0.75,
+          },
+        },
+        React.createElement("span", null, label),
+        React.createElement("span", { style: { color, fontWeight: 700 } }, value + "%"),
+      ),
+      React.createElement(
+        "div",
+        { style: { display: "flex", gap: 2 } },
+        cells,
+      ),
+    );
+  }
+
+  // ---------------------------------------------------------------------
+  // Sidebar: MS-STATUS panel
+  // ---------------------------------------------------------------------
+
+  function SidebarSlot() {
+    // Pull live-ish numbers from the status API so the plugin isn't just
+    // a static decoration. Fall back to full bars if the API is slow /
+    // unavailable.
+    const [status, setStatus] = useState(null);
+    useEffect(function () {
+      let cancel = false;
+      api.getStatus()
+        .then(function (s) { if (!cancel) setStatus(s); })
+        .catch(function () {});
+      return function () { cancel = true; };
+    }, []);
+
+    // Map real status signals to HUD telemetry. Energy/shield/power
+    // aren't literal concepts on a software agent, so we read them from
+    // adjacent signals: active sessions, gateway connected-platforms,
+    // and agent-online health.
+    const energy = status && status.gateway_online ? 92 : 18;
+    const shield = status && status.connected_platforms
+      ? Math.min(100, 40 + (status.connected_platforms.length * 15))
+      : 70;
+    const power = status && status.active_sessions
+      ? Math.min(100, 55 + (status.active_sessions.length * 10))
+      : 87;
+
+    const hero = cssVar("--theme-asset-hero");
+
+    return React.createElement(
+      "div",
+      {
+        style: {
+          padding: "1rem 0.75rem",
+          display: "flex",
+          flexDirection: "column",
+          gap: "1rem",
+          fontFamily: "var(--theme-font-display, sans-serif)",
+          letterSpacing: "0.08em",
+          textTransform: "uppercase",
+          fontSize: "0.65rem",
+        },
+      },
+      // Header line
+      React.createElement(
+        "div",
+        {
+          style: {
+            borderBottom: "1px solid rgba(64,200,255,0.3)",
+            paddingBottom: 8,
+            display: "flex",
+            flexDirection: "column",
+            gap: 2,
+          },
+        },
+        React.createElement("span", { style: { opacity: 0.6 } }, "ms status"),
+        React.createElement("span", { style: { fontWeight: 700, fontSize: "0.85rem" } }, "zgmf-x20a"),
+        React.createElement("span", { style: { opacity: 0.6, fontSize: "0.6rem" } }, "strike freedom"),
+      ),
+      // Hero slot — only renders when the theme provides one.
+      hero
+        ? React.createElement("div", {
+            style: {
+              width: "100%",
+              aspectRatio: "3 / 4",
+              backgroundImage: hero,
+              backgroundSize: "contain",
+              backgroundPosition: "center",
+              backgroundRepeat: "no-repeat",
+              opacity: 0.85,
+            },
+            "aria-hidden": true,
+          })
+        : React.createElement("div", {
+            style: {
+              width: "100%",
+              aspectRatio: "3 / 4",
+              border: "1px dashed rgba(64,200,255,0.25)",
+              display: "flex",
+              alignItems: "center",
+              justifyContent: "center",
+              fontSize: "0.55rem",
+              opacity: 0.4,
+            },
+          }, "hero slot — set assets.hero in theme"),
+      // Pilot block
+      React.createElement(
+        "div",
+        {
+          style: {
+            borderTop: "1px solid rgba(64,200,255,0.18)",
+            borderBottom: "1px solid rgba(64,200,255,0.18)",
+            padding: "8px 0",
+            display: "flex",
+            flexDirection: "column",
+            gap: 2,
+          },
+        },
+        React.createElement("span", { style: { opacity: 0.5, fontSize: "0.55rem" } }, "pilot"),
+        React.createElement("span", { style: { fontWeight: 700 } }, "hermes agent"),
+        React.createElement("span", { style: { opacity: 0.5, fontSize: "0.55rem" } }, "compass"),
+      ),
+      // Telemetry bars
+      React.createElement(TelemetryBar, { label: "energy",  value: energy, color: "#ffce3a" }),
+      React.createElement(TelemetryBar, { label: "shield",  value: shield, color: "#3fd3ff" }),
+      React.createElement(TelemetryBar, { label: "power",   value: power,  color: "#ff3a5e" }),
+      // System online
+      React.createElement(
+        "div",
+        {
+          style: {
+            marginTop: 4,
+            padding: "6px 8px",
+            border: "1px solid rgba(74,222,128,0.4)",
+            color: "#4ade80",
+            textAlign: "center",
+            fontWeight: 700,
+            fontSize: "0.6rem",
+          },
+        },
+        status && status.gateway_online ? "system online" : "system offline",
+      ),
+    );
+  }
+
+  // ---------------------------------------------------------------------
+  // Header-left: COMPASS crest
+  // ---------------------------------------------------------------------
+
+  function HeaderCrestSlot() {
+    const crest = cssVar("--theme-asset-crest");
+    const inner = crest
+      ? React.createElement("div", {
+          style: {
+            width: 28,
+            height: 28,
+            backgroundImage: crest,
+            backgroundSize: "contain",
+            backgroundPosition: "center",
+            backgroundRepeat: "no-repeat",
+          },
+          "aria-hidden": true,
+        })
+      : React.createElement(
+          "svg",
+          {
+            width: 28,
+            height: 28,
+            viewBox: "0 0 28 28",
+            fill: "none",
+            stroke: "currentColor",
+            strokeWidth: 1.5,
+            "aria-hidden": true,
+          },
+          React.createElement("path", { d: "M14 2 L26 14 L14 26 L2 14 Z" }),
+          React.createElement("path", { d: "M14 8 L20 14 L14 20 L8 14 Z" }),
+          React.createElement("circle", { cx: 14, cy: 14, r: 2, fill: "currentColor" }),
+        );
+    return React.createElement(
+      "div",
+      {
+        style: {
+          display: "flex",
+          alignItems: "center",
+          paddingLeft: 12,
+          paddingRight: 8,
+          color: "var(--color-accent, #3fd3ff)",
+        },
+      },
+      inner,
+    );
+  }
+
+  // ---------------------------------------------------------------------
+  // Footer-right: COSMIC ERA tagline
+  // ---------------------------------------------------------------------
+
+  function FooterTaglineSlot() {
+    return React.createElement(
+      "span",
+      {
+        style: {
+          fontFamily: "var(--theme-font-display, sans-serif)",
+          fontSize: "0.6rem",
+          letterSpacing: "0.18em",
+          textTransform: "uppercase",
+          opacity: 0.75,
+          mixBlendMode: "plus-lighter",
+        },
+      },
+      "compass hermes systems / cosmic era 71",
+    );
+  }
+
+  // ---------------------------------------------------------------------
+  // Hidden tab placeholder — tab.hidden=true means this never renders in
+  // the nav, but we still register something sensible in case someone
+  // manually navigates to /strike-freedom-cockpit (e.g. via a bookmark).
+  // ---------------------------------------------------------------------
+
+  function HiddenPage() {
+    return React.createElement(
+      "div",
+      { style: { padding: "2rem", opacity: 0.6, fontSize: "0.8rem" } },
+      "Strike Freedom cockpit is a slot-only plugin — it populates the sidebar, header, and footer instead of showing a tab page.",
+    );
+  }
+
+  // ---------------------------------------------------------------------
+  // Registration
+  // ---------------------------------------------------------------------
+
+  const NAME = "strike-freedom-cockpit";
+  PLUGINS.register(NAME, HiddenPage);
+  PLUGINS.registerSlot(NAME, "sidebar", SidebarSlot);
+  PLUGINS.registerSlot(NAME, "header-left", HeaderCrestSlot);
+  PLUGINS.registerSlot(NAME, "footer-right", FooterTaglineSlot);
+})();

plugins/strike-freedom-cockpit/dashboard/manifest.json

@@ -0,0 +1,14 @@
+{
+  "name": "strike-freedom-cockpit",
+  "label": "Strike Freedom Cockpit",
+  "description": "MS-STATUS sidebar + header crest for the Strike Freedom theme",
+  "icon": "Shield",
+  "version": "1.0.0",
+  "tab": {
+    "path": "/strike-freedom-cockpit",
+    "position": "end",
+    "hidden": true
+  },
+  "slots": ["sidebar", "header-left", "footer-right"],
+  "entry": "dist/index.js"
+}

plugins/strike-freedom-cockpit/theme/strike-freedom.yaml

@@ -0,0 +1,126 @@
+# Strike Freedom — Hermes dashboard theme demo
+#
+# Copy this file to ~/.hermes/dashboard-themes/strike-freedom.yaml and
+# restart the web UI (or hit `/api/dashboard/plugins/rescan`). Pair with
+# the `strike-freedom-cockpit` plugin (plugins/strike-freedom-cockpit/)
+# for the full cockpit experience — this theme paints the palette,
+# chrome, and layout; the plugin supplies the MS-STATUS sidebar + header
+# crest that the cockpit layout variant reserves space for.
+#
+# Demonstrates every theme extension point added alongside the plugin
+# slot system: palette, typography, layoutVariant, assets, customCSS,
+# componentStyles, colorOverrides.
+name: strike-freedom
+label: "Strike Freedom"
+description: "Cockpit HUD — deep navy + cyan + gold accents"
+
+# ------- palette (3-layer) -------
+palette:
+  background: "#05091a"
+  midground: "#d8f0ff"
+  foreground:
+    hex: "#ffffff"
+    alpha: 0
+  warmGlow: "rgba(255, 199, 55, 0.24)"
+  noiseOpacity: 0.7
+
+# ------- typography -------
+typography:
+  fontSans: '"Orbitron", "Eurostile", "Bank Gothic", "Impact", sans-serif'
+  fontMono: '"Share Tech Mono", "JetBrains Mono", ui-monospace, monospace'
+  fontDisplay: '"Orbitron", "Eurostile", "Impact", sans-serif'
+  fontUrl: "https://fonts.googleapis.com/css2?family=Orbitron:wght@400;500;600;700;800&family=Share+Tech+Mono&display=swap"
+  baseSize: "14px"
+  lineHeight: "1.5"
+  letterSpacing: "0.04em"
+
+# ------- layout -------
+layout:
+  radius: "0"
+  density: "compact"
+
+# ``cockpit`` reserves a 260px left rail that the shell renders when the
+# user is on this theme. A paired plugin populates the rail via the
+# ``sidebar`` slot; with no plugin the rail shows a placeholder.
+layoutVariant: cockpit
+
+# ------- assets -------
+# Use any URL (https, data:, /dashboard-plugins/...) or a pre-wrapped
+# ``url(...)``/``linear-gradient(...)`` expression. The shell exposes
+# each as a CSS var so plugins can read the same imagery.
+assets:
+  bg: "linear-gradient(140deg, #05091a 0%, #0a1530 55%, #102048 100%)"
+  # Plugin reads --theme-asset-hero / --theme-asset-crest to populate
+  # its sidebar hero render + header crest. Replace these URLs with your
+  # own artwork (copy files into ~/.hermes/dashboard-themes/assets/ and
+  # reference them as /dashboard-themes-assets/strike-freedom/hero.png
+  # once that static route is wired up — for now use inline data URLs or
+  # remote URLs).
+  hero: ""
+  crest: ""
+
+# ------- component chrome -------
+# Each bucket's props become CSS vars (--component-<bucket>-<kebab>) that
+# built-in shell components (Card, header, sidebar, backdrop) consume.
+componentStyles:
+  card:
+    # Notched corners on the top-left + bottom-right — classic mecha UI.
+    clipPath: "polygon(12px 0, 100% 0, 100% calc(100% - 12px), calc(100% - 12px) 100%, 0 100%, 0 12px)"
+    background: "linear-gradient(180deg, rgba(10, 22, 52, 0.85) 0%, rgba(5, 9, 26, 0.92) 100%)"
+    boxShadow: "inset 0 0 0 1px rgba(64, 200, 255, 0.28), 0 0 18px -6px rgba(64, 200, 255, 0.4)"
+  header:
+    background: "linear-gradient(180deg, rgba(16, 32, 72, 0.95) 0%, rgba(5, 9, 26, 0.9) 100%)"
+  sidebar:
+    background: "linear-gradient(180deg, rgba(8, 18, 42, 0.88) 0%, rgba(5, 9, 26, 0.85) 100%)"
+  tab:
+    clipPath: "polygon(6px 0, 100% 0, calc(100% - 6px) 100%, 0 100%)"
+  backdrop:
+    backgroundSize: "cover"
+    backgroundPosition: "center"
+    fillerOpacity: "1"
+    fillerBlendMode: "normal"
+
+# ------- color overrides -------
+colorOverrides:
+  primary: "#ffce3a"
+  primaryForeground: "#05091a"
+  accent: "#3fd3ff"
+  accentForeground: "#05091a"
+  ring: "#3fd3ff"
+  success: "#4ade80"
+  warning: "#ffce3a"
+  destructive: "#ff3a5e"
+  border: "rgba(64, 200, 255, 0.28)"
+
+# ------- customCSS -------
+# Raw CSS injected as a scoped <style> tag on theme apply. Use this for
+# selector-level tweaks componentStyles can't express (pseudo-elements,
+# animations, media queries). Bounded to 32 KiB per theme.
+customCSS: |
+  /* Scanline overlay — subtle, only when theme is active. */
+  :root[data-layout-variant="cockpit"] body::before {
+    content: "";
+    position: fixed;
+    inset: 0;
+    pointer-events: none;
+    z-index: 100;
+    background: repeating-linear-gradient(
+      to bottom,
+      transparent 0px,
+      transparent 2px,
+      rgba(64, 200, 255, 0.035) 3px,
+      rgba(64, 200, 255, 0.035) 4px
+    );
+    mix-blend-mode: screen;
+  }
+
+  /* Chevron pips on card corners. */
+  [data-layout-variant="cockpit"] .border-border::before,
+  [data-layout-variant="cockpit"] .border-border::after {
+    content: "";
+    position: absolute;
+    width: 8px;
+    height: 8px;
+    border: 1px solid rgba(64, 200, 255, 0.55);
+    pointer-events: none;
+  }

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "hermes-agent"
-version = "0.10.0"
+version = "0.11.0"
 description = "The self-improving AI agent — creates skills from experience, improves them during use, and runs anywhere"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -39,7 +39,7 @@ dependencies = [
 [project.optional-dependencies]
 modal = ["modal>=1.0.0,<2"]
 daytona = ["daytona>=0.148.0,<1"]
-dev = ["debugpy>=1.8.0,<2", "pytest>=9.0.2,<10", "pytest-asyncio>=1.3.0,<2", "pytest-xdist>=3.0,<4", "mcp>=1.2.0,<2"]
+dev = ["debugpy>=1.8.0,<2", "pytest>=9.0.2,<10", "pytest-asyncio>=1.3.0,<2", "pytest-xdist>=3.0,<4", "mcp>=1.2.0,<2", "ty>=0.0.1a29,<0.0.22", "ruff"]
 messaging = ["python-telegram-bot[webhooks]>=22.6,<23", "discord.py[voice]>=2.7.1,<3", "aiohttp>=3.13.3,<4", "slack-bolt>=1.18.0,<2", "slack-sdk>=3.27.0,<4", "qrcode>=7.0,<8"]
 cron = ["croniter>=6.0.0,<7"]
 slack = ["slack-bolt>=1.18.0,<2", "slack-sdk>=3.27.0,<4"]
@@ -134,3 +134,28 @@ markers = [
     "integration: marks tests requiring external services (API keys, Modal, etc.)",
 ]
 addopts = "-m 'not integration' -n auto"
+
+[tool.ty.environment]
+python-version = "3.13"
+
+[tool.ty.rules]
+unknown-argument = "warn"
+redundant-cast = "ignore"
+
+[tool.ty.src]
+exclude = ["**"]
+
+[[tool.ty.overrides]]
+include = ["**"]
+
+[tool.ty.overrides.rules]
+unresolved-import = "ignore"
+invalid-method-override = "ignore"
+invalid-assignment = "ignore"
+not-iterable = "ignore"
+
+[tool.ruff]
+exclude = ["*"]
+
+[tool.uv]
+exclude-newer = "7 days"

run_agent.py

@@ -31,11 +31,13 @@ logger = logging.getLogger(__name__)
 import os
 import random
 import re
+import ssl
 import sys
 import tempfile
 import time
 import threading
 from types import SimpleNamespace
+import urllib.request
 import uuid
 from typing import List, Dict, Any, Optional
 from openai import OpenAI
@@ -181,6 +183,25 @@ def _get_proxy_from_env() -> Optional[str]:
     return None
 
 
+def _get_proxy_for_base_url(base_url: Optional[str]) -> Optional[str]:
+    """Return an env-configured proxy unless NO_PROXY excludes this base URL."""
+    proxy = _get_proxy_from_env()
+    if not proxy or not base_url:
+        return proxy
+
+    host = base_url_hostname(base_url)
+    if not host:
+        return proxy
+
+    try:
+        if urllib.request.proxy_bypass_environment(host):
+            return None
+    except Exception:
+        pass
+
+    return proxy
+
+
 def _install_safe_stdio() -> None:
     """Wrap stdout/stderr so best-effort console output cannot crash the agent."""
     for stream_name in ("stdout", "stderr"):
@@ -262,6 +283,7 @@ _MAX_TOOL_WORKERS = 8
 _DESTRUCTIVE_PATTERNS = re.compile(
     r"""(?:^|\s|&&|\|\||;|`)(?:
         rm\s|rmdir\s|
+        cp\s|install\s|
         mv\s|
         sed\s+-i|
         truncate\s|
@@ -663,6 +685,40 @@ def _sanitize_structure_non_ascii(payload: Any) -> bool:
 _QWEN_CODE_VERSION = "0.14.1"
 
 
+def _routermint_headers() -> dict:
+    """Return the User-Agent RouterMint needs to avoid Cloudflare 1010 blocks."""
+    from hermes_cli import __version__ as _HERMES_VERSION
+
+    return {
+        "User-Agent": f"HermesAgent/{_HERMES_VERSION}",
+    }
+
+
+def _pool_may_recover_from_rate_limit(pool) -> bool:
+    """Decide whether to wait for credential-pool rotation instead of falling back.
+
+    The existing pool-rotation path requires the pool to (1) exist and (2) have
+    at least one entry not currently in exhaustion cooldown.  But rotation is
+    only meaningful when the pool has more than one entry.
+
+    With a single-credential pool (common for Gemini OAuth, Vertex service
+    accounts, and any "one personal key" configuration), the primary entry
+    just 429'd and there is nothing to rotate to.  Waiting for the pool
+    cooldown to expire means retrying against the same exhausted quota — the
+    daily-quota 429 will recur immediately, and the retry budget is burned.
+
+    In that case we must fall back to the configured ``fallback_model``
+    instead.  Returns True only when rotation has somewhere to go.
+
+    See issue #11314.
+    """
+    if pool is None:
+        return False
+    if not pool.has_available():
+        return False
+    return len(pool.entries()) > 1
+
+
 def _qwen_portal_headers() -> dict:
     """Return default HTTP headers required by Qwen Portal API."""
     import platform as _plat
@@ -1006,8 +1062,21 @@ class AIAgent:
         self._use_prompt_caching, self._use_native_cache_layout = (
             self._anthropic_prompt_cache_policy()
         )
-        self._cache_ttl = "5m"  # Default 5-minute TTL (1.25x write cost)
-        
+        # Anthropic supports "5m" (default) and "1h" cache TTL tiers. Read from
+        # config.yaml under prompt_caching.cache_ttl; unknown values keep "5m".
+        # 1h tier costs 2x on write vs 1.25x for 5m, but amortizes across long
+        # sessions with >5-minute pauses between turns (#14971).
+        self._cache_ttl = "5m"
+        try:
+            from hermes_cli.config import load_config as _load_pc_cfg
+
+            _pc_cfg = _load_pc_cfg().get("prompt_caching", {}) or {}
+            _ttl = _pc_cfg.get("cache_ttl", "5m")
+            if _ttl in ("5m", "1h"):
+                self._cache_ttl = _ttl
+        except Exception:
+            pass
+
         # Iteration budget: the LLM is only notified when it actually exhausts
         # the iteration budget (api_call_count >= max_iterations).  At that
         # point we inject ONE message, allow one final API call, and if the
@@ -1179,6 +1248,8 @@ class AIAgent:
                         "X-OpenRouter-Title": "Hermes Agent",
                         "X-OpenRouter-Categories": "productivity,cli-agent",
                     }
+                elif base_url_host_matches(effective_base, "api.routermint.com"):
+                    client_kwargs["default_headers"] = _routermint_headers()
                 elif base_url_host_matches(effective_base, "api.githubcopilot.com"):
                     from hermes_cli.models import copilot_default_headers
 
@@ -1548,6 +1619,17 @@ class AIAgent:
             _agent_section = {}
         self._tool_use_enforcement = _agent_section.get("tool_use_enforcement", "auto")
 
+        # App-level API retry count (wraps each model API call).  Default 3,
+        # overridable via agent.api_max_retries in config.yaml.  See #11616.
+        try:
+            _raw_api_retries = _agent_section.get("api_max_retries", 3)
+            _api_retries = int(_raw_api_retries)
+            if _api_retries < 1:
+                _api_retries = 1  # 1 = no retry (single attempt)
+        except (TypeError, ValueError):
+            _api_retries = 3
+        self._api_max_retries = _api_retries
+
         # Initialize context compressor for automatic context management
         # Compresses conversation when approaching model's context limit
         # Configuration via config.yaml (compression section)
@@ -2037,12 +2119,14 @@ class AIAgent:
         # ("switched to anthropic, tui keeps trying openrouter").
         old_norm = (old_provider or "").strip().lower()
         new_norm = (new_provider or "").strip().lower()
+        fallback_chain = list(getattr(self, "_fallback_chain", []) or [])
         if old_norm and new_norm and old_norm != new_norm:
-            self._fallback_chain = [
-                entry for entry in self._fallback_chain
+            fallback_chain = [
+                entry for entry in fallback_chain
                 if (entry.get("provider") or "").strip().lower() not in {old_norm, new_norm}
             ]
-            self._fallback_model = self._fallback_chain[0] if self._fallback_chain else None
+        self._fallback_chain = fallback_chain
+        self._fallback_model = fallback_chain[0] if fallback_chain else None
 
         logging.info(
             "Model switched in-place: %s (%s) -> %s (%s)",
@@ -2864,6 +2948,69 @@ class AIAgent:
         "If nothing stands out, just say 'Nothing to save.' and stop."
     )
 
+    @staticmethod
+    def _summarize_background_review_actions(
+        review_messages: List[Dict],
+        prior_snapshot: List[Dict],
+    ) -> List[str]:
+        """Build the human-facing action summary for a background review pass.
+
+        Walks the review agent's session messages and collects "successful tool
+        action" descriptions to surface to the user (e.g. "Memory updated").
+        Tool messages already present in ``prior_snapshot`` are skipped so we
+        don't re-surface stale results from the prior conversation that the
+        review agent inherited via ``conversation_history`` (issue #14944).
+
+        Matching is by ``tool_call_id`` when available, with a content-equality
+        fallback for tool messages that lack one.
+        """
+        existing_tool_call_ids = set()
+        existing_tool_contents = set()
+        for prior in prior_snapshot or []:
+            if not isinstance(prior, dict) or prior.get("role") != "tool":
+                continue
+            tcid = prior.get("tool_call_id")
+            if tcid:
+                existing_tool_call_ids.add(tcid)
+            else:
+                content = prior.get("content")
+                if isinstance(content, str):
+                    existing_tool_contents.add(content)
+
+        actions: List[str] = []
+        for msg in review_messages or []:
+            if not isinstance(msg, dict) or msg.get("role") != "tool":
+                continue
+            tcid = msg.get("tool_call_id")
+            if tcid and tcid in existing_tool_call_ids:
+                continue
+            if not tcid:
+                content_str = msg.get("content")
+                if isinstance(content_str, str) and content_str in existing_tool_contents:
+                    continue
+            try:
+                data = json.loads(msg.get("content", "{}"))
+            except (json.JSONDecodeError, TypeError):
+                continue
+            if not isinstance(data, dict) or not data.get("success"):
+                continue
+            message = data.get("message", "")
+            target = data.get("target", "")
+            if "created" in message.lower():
+                actions.append(message)
+            elif "updated" in message.lower():
+                actions.append(message)
+            elif "added" in message.lower() or (target and "add" in message.lower()):
+                label = "Memory" if target == "memory" else "User profile" if target == "user" else target
+                actions.append(f"{label} updated")
+            elif "Entry added" in message:
+                label = "Memory" if target == "memory" else "User profile" if target == "user" else target
+                actions.append(f"{label} updated")
+            elif "removed" in message.lower() or "replaced" in message.lower():
+                label = "Memory" if target == "memory" else "User profile" if target == "user" else target
+                actions.append(f"{label} updated")
+        return actions
+
     def _spawn_background_review(
         self,
         messages_snapshot: List[Dict],
@@ -2913,32 +3060,15 @@ class AIAgent:
                     )
 
                 # Scan the review agent's messages for successful tool actions
-                # and surface a compact summary to the user.
-                actions = []
-                for msg in getattr(review_agent, "_session_messages", []):
-                    if not isinstance(msg, dict) or msg.get("role") != "tool":
-                        continue
-                    try:
-                        data = json.loads(msg.get("content", "{}"))
-                    except (json.JSONDecodeError, TypeError):
-                        continue
-                    if not data.get("success"):
-                        continue
-                    message = data.get("message", "")
-                    target = data.get("target", "")
-                    if "created" in message.lower():
-                        actions.append(message)
-                    elif "updated" in message.lower():
-                        actions.append(message)
-                    elif "added" in message.lower() or (target and "add" in message.lower()):
-                        label = "Memory" if target == "memory" else "User profile" if target == "user" else target
-                        actions.append(f"{label} updated")
-                    elif "Entry added" in message:
-                        label = "Memory" if target == "memory" else "User profile" if target == "user" else target
-                        actions.append(f"{label} updated")
-                    elif "removed" in message.lower() or "replaced" in message.lower():
-                        label = "Memory" if target == "memory" else "User profile" if target == "user" else target
-                        actions.append(f"{label} updated")
+                # and surface a compact summary to the user. Tool messages
+                # already present in messages_snapshot must be skipped, since
+                # the review agent inherits that history and would otherwise
+                # re-surface stale "created"/"updated" messages from the prior
+                # conversation as if they just happened (issue #14944).
+                actions = self._summarize_background_review_actions(
+                    getattr(review_agent, "_session_messages", []),
+                    messages_snapshot,
+                )
 
                 if actions:
                     summary = " · ".join(dict.fromkeys(actions))
@@ -4344,25 +4474,69 @@ class AIAgent:
     def _repair_tool_call(self, tool_name: str) -> str | None:
         """Attempt to repair a mismatched tool name before aborting.
 
-        1. Try lowercase
-        2. Try normalized (lowercase + hyphens/spaces -> underscores)
-        3. Try fuzzy match (difflib, cutoff=0.7)
+        Models sometimes emit variants of a tool name that differ only
+        in casing, separators, or class-like suffixes. Normalize
+        aggressively before falling back to fuzzy match:
+
+        1. Lowercase direct match.
+        2. Lowercase + hyphens/spaces -> underscores.
+        3. CamelCase -> snake_case (TodoTool -> todo_tool).
+        4. Strip trailing ``_tool`` / ``-tool`` / ``tool`` suffix that
+           Claude-style models sometimes tack on (TodoTool_tool ->
+           TodoTool -> Todo -> todo). Applied twice so double-tacked
+           suffixes like ``TodoTool_tool`` reduce all the way.
+        5. Fuzzy match (difflib, cutoff=0.7).
+
+        See #14784 for the original reports (TodoTool_tool, Patch_tool,
+        BrowserClick_tool were all returning "Unknown tool" before).
 
         Returns the repaired name if found in valid_tool_names, else None.
         """
+        import re
         from difflib import get_close_matches
 
-        # 1. Lowercase
+        if not tool_name:
+            return None
+
+        def _norm(s: str) -> str:
+            return s.lower().replace("-", "_").replace(" ", "_")
+
+        def _camel_snake(s: str) -> str:
+            return re.sub(r"(?<!^)(?=[A-Z])", "_", s).lower()
+
+        def _strip_tool_suffix(s: str) -> str | None:
+            lc = s.lower()
+            for suffix in ("_tool", "-tool", "tool"):
+                if lc.endswith(suffix):
+                    return s[: -len(suffix)].rstrip("_-")
+            return None
+
+        # Cheap fast-paths first — these cover the common case.
         lowered = tool_name.lower()
         if lowered in self.valid_tool_names:
             return lowered
-
-        # 2. Normalize
-        normalized = lowered.replace("-", "_").replace(" ", "_")
+        normalized = _norm(tool_name)
         if normalized in self.valid_tool_names:
             return normalized
 
-        # 3. Fuzzy match
+        # Build the full candidate set for class-like emissions.
+        cands: set[str] = {tool_name, lowered, normalized, _camel_snake(tool_name)}
+        # Strip trailing tool-suffix up to twice — TodoTool_tool needs it.
+        for _ in range(2):
+            extra: set[str] = set()
+            for c in cands:
+                stripped = _strip_tool_suffix(c)
+                if stripped:
+                    extra.add(stripped)
+                    extra.add(_norm(stripped))
+                    extra.add(_camel_snake(stripped))
+            cands |= extra
+
+        for c in cands:
+            if c and c in self.valid_tool_names:
+                return c
+
+        # Fuzzy match as last resort.
         matches = get_close_matches(lowered, self.valid_tool_names, n=1, cutoff=0.7)
         if matches:
             return matches[0]
@@ -4454,7 +4628,7 @@ class AIAgent:
         return False
 
     @staticmethod
-    def _build_keepalive_http_client() -> Any:
+    def _build_keepalive_http_client(base_url: str = "") -> Any:
         try:
             import httpx as _httpx
             import socket as _socket
@@ -4468,8 +4642,9 @@ class AIAgent:
                 _sock_opts.append((_socket.IPPROTO_TCP, _socket.TCP_KEEPALIVE, 30))
             # When a custom transport is provided, httpx won't auto-read proxy
             # from env vars (allow_env_proxies = trust_env and transport is None).
-            # Explicitly read proxy settings to ensure HTTP_PROXY/HTTPS_PROXY work.
-            _proxy = _get_proxy_from_env()
+            # Explicitly read proxy settings while still honoring NO_PROXY for
+            # loopback / local endpoints such as a locally hosted sub2api.
+            _proxy = _get_proxy_for_base_url(base_url)
             return _httpx.Client(
                 transport=_httpx.HTTPTransport(socket_options=_sock_opts),
                 proxy=_proxy,
@@ -4527,7 +4702,7 @@ class AIAgent:
                     if k in {"api_key", "base_url", "default_headers", "timeout", "http_client"}
                 }
                 if "http_client" not in safe_kwargs:
-                    keepalive_http = self._build_keepalive_http_client()
+                    keepalive_http = self._build_keepalive_http_client(base_url)
                     if keepalive_http is not None:
                         safe_kwargs["http_client"] = keepalive_http
                 client = GeminiNativeClient(**safe_kwargs)
@@ -4556,7 +4731,7 @@ class AIAgent:
         # Tests in ``tests/run_agent/test_create_openai_client_reuse.py`` and
         # ``tests/run_agent/test_sequential_chats_live.py`` pin this invariant.
         if "http_client" not in client_kwargs:
-            keepalive_http = self._build_keepalive_http_client()
+            keepalive_http = self._build_keepalive_http_client(client_kwargs.get("base_url", ""))
             if keepalive_http is not None:
                 client_kwargs["http_client"] = keepalive_http
         client = OpenAI(**client_kwargs)
@@ -5032,6 +5207,41 @@ class AIAgent:
 
         return True
 
+    def _try_refresh_copilot_client_credentials(self) -> bool:
+        """Refresh Copilot credentials and rebuild the shared OpenAI client.
+
+        Copilot tokens may remain the same string across refreshes (`gh auth token`
+        returns a stable OAuth token in many setups). We still rebuild the client
+        on 401 so retries recover from stale auth/client state without requiring
+        a session restart.
+        """
+        if self.provider != "copilot":
+            return False
+
+        try:
+            from hermes_cli.copilot_auth import resolve_copilot_token
+
+            new_token, token_source = resolve_copilot_token()
+        except Exception as exc:
+            logger.debug("Copilot credential refresh failed: %s", exc)
+            return False
+
+        if not isinstance(new_token, str) or not new_token.strip():
+            return False
+
+        new_token = new_token.strip()
+
+        self.api_key = new_token
+        self._client_kwargs["api_key"] = self.api_key
+        self._client_kwargs["base_url"] = self.base_url
+        self._apply_client_headers_for_base_url(str(self.base_url or ""))
+
+        if not self._replace_primary_openai_client(reason="copilot_credential_refresh"):
+            return False
+
+        logger.info("Copilot credentials refreshed from %s", token_source)
+        return True
+
     def _try_refresh_anthropic_client_credentials(self) -> bool:
         if self.api_mode != "anthropic_messages" or not hasattr(self, "_anthropic_api_key"):
             return False
@@ -5085,6 +5295,8 @@ class AIAgent:
             self._client_kwargs["default_headers"] = dict(_OR_HEADERS)
         elif base_url_host_matches(base_url, "ai-gateway.vercel.sh"):
             self._client_kwargs["default_headers"] = dict(_AI_GATEWAY_HEADERS)
+        elif base_url_host_matches(base_url, "api.routermint.com"):
+            self._client_kwargs["default_headers"] = _routermint_headers()
         elif base_url_host_matches(base_url, "api.githubcopilot.com"):
             from hermes_cli.models import copilot_default_headers
 
@@ -5163,7 +5375,7 @@ class AIAgent:
                 effective_reason = FailoverReason.billing
             elif status_code == 429:
                 effective_reason = FailoverReason.rate_limit
-            elif status_code == 401:
+            elif status_code in (401, 403):
                 effective_reason = FailoverReason.auth
 
         if effective_reason == FailoverReason.billing:
@@ -6282,7 +6494,7 @@ class AIAgent:
 
     # ── Provider fallback ──────────────────────────────────────────────────
 
-    def _try_activate_fallback(self) -> bool:
+    def _try_activate_fallback(self, reason: "FailoverReason | None" = None) -> bool:
         """Switch to the next fallback model/provider in the chain.
 
         Called when the current model is failing after retries.  Swaps the
@@ -6294,6 +6506,15 @@ class AIAgent:
         auth resolution and client construction — no duplicated provider→key
         mappings.
         """
+        if reason in (FailoverReason.rate_limit, FailoverReason.billing):
+            # Only start cooldown when leaving the primary provider.  If we're
+            # already on a fallback and chain-switching, the primary wasn't the
+            # source of the 429 so the cooldown should not be reset/extended.
+            fallback_already_active = bool(getattr(self, "_fallback_activated", False))
+            current_provider = (getattr(self, "provider", "") or "").strip().lower()
+            primary_provider = ((self._primary_runtime or {}).get("provider") or "").strip().lower()
+            if (not fallback_already_active) or (primary_provider and current_provider == primary_provider):
+                self._rate_limited_until = time.monotonic() + 60
         if self._fallback_index >= len(self._fallback_chain):
             return False
 
@@ -6430,11 +6651,15 @@ class AIAgent:
             # Without this, compression decisions use the primary model's
             # context window (e.g. 200K) instead of the fallback's (e.g. 32K),
             # causing oversized sessions to overflow the fallback.
+            # Also pass _config_context_length so the explicit config override
+            # (model.context_length in config.yaml) is respected — without this,
+            # the fallback activation drops to 128K even when config says 204800.
             if hasattr(self, 'context_compressor') and self.context_compressor:
                 from agent.model_metadata import get_model_context_length
                 fb_context_length = get_model_context_length(
                     self.model, base_url=self.base_url,
                     api_key=self.api_key, provider=self.provider,
+                    config_context_length=getattr(self, "_config_context_length", None),
                 )
                 self.context_compressor.update_model(
                     model=self.model,
@@ -6473,6 +6698,9 @@ class AIAgent:
         if not self._fallback_activated:
             return False
 
+        if getattr(self, "_rate_limited_until", 0) > time.monotonic():
+            return False  # primary still in rate-limit cooldown, stay on fallback
+
         rt = self._primary_runtime
         try:
             # ── Core runtime state ──
@@ -6766,42 +6994,6 @@ class AIAgent:
             cache[mode] = t
         return t
 
-    @staticmethod
-    def _nr_to_assistant_message(nr):
-        """Convert a NormalizedResponse to the SimpleNamespace shape downstream expects.
-
-        This is the single back-compat shim between the transport layer
-        (NormalizedResponse) and the agent loop (SimpleNamespace with
-        .content, .tool_calls, .reasoning, .reasoning_content,
-        .reasoning_details, .codex_reasoning_items, and per-tool-call
-        .call_id / .response_item_id).
-
-        TODO: Remove when downstream code reads NormalizedResponse directly.
-        """
-        tc_list = None
-        if nr.tool_calls:
-            tc_list = []
-            for tc in nr.tool_calls:
-                tc_ns = SimpleNamespace(
-                    id=tc.id,
-                    type="function",
-                    function=SimpleNamespace(name=tc.name, arguments=tc.arguments),
-                )
-                if tc.provider_data:
-                    for key in ("call_id", "response_item_id"):
-                        if tc.provider_data.get(key):
-                            setattr(tc_ns, key, tc.provider_data[key])
-                tc_list.append(tc_ns)
-        pd = nr.provider_data or {}
-        return SimpleNamespace(
-            content=nr.content,
-            tool_calls=tc_list or None,
-            reasoning=nr.reasoning,
-            reasoning_content=pd.get("reasoning_content"),
-            reasoning_details=pd.get("reasoning_details"),
-            codex_reasoning_items=pd.get("codex_reasoning_items"),
-        )
-
     def _prepare_anthropic_messages_for_api(self, api_messages: list) -> list:
         if not any(
             isinstance(msg, dict) and self._content_has_image_parts(msg.get("content"))
@@ -7503,20 +7695,25 @@ class AIAgent:
                     ]
             elif self.api_mode == "anthropic_messages" and not _aux_available:
                 _tfn = self._get_transport()
-                _flush_nr = _tfn.normalize_response(response, strip_tool_prefix=self._is_anthropic_oauth)
-                if _flush_nr and _flush_nr.tool_calls:
+                _flush_result = _tfn.normalize_response(response, strip_tool_prefix=self._is_anthropic_oauth)
+                if _flush_result and _flush_result.tool_calls:
                     tool_calls = [
                         SimpleNamespace(
                             id=tc.id, type="function",
                             function=SimpleNamespace(name=tc.name, arguments=tc.arguments),
-                        ) for tc in _flush_nr.tool_calls
+                        ) for tc in _flush_result.tool_calls
                     ]
-            elif hasattr(response, "choices") and response.choices:
+            elif self.api_mode in ("chat_completions", "bedrock_converse"):
                 # chat_completions / bedrock — normalize through transport
-                _flush_cc_nr = self._get_transport().normalize_response(response)
-                _flush_msg = self._nr_to_assistant_message(_flush_cc_nr)
-                if _flush_msg.tool_calls:
-                    tool_calls = _flush_msg.tool_calls
+                _flush_result = self._get_transport().normalize_response(response)
+                if _flush_result.tool_calls:
+                    tool_calls = _flush_result.tool_calls
+            elif _aux_available and hasattr(response, "choices") and response.choices:
+                # Auxiliary client returned OpenAI-shaped response while main
+                # api_mode is codex/anthropic — extract tool_calls from .choices
+                _aux_msg = response.choices[0].message
+                if hasattr(_aux_msg, "tool_calls") and _aux_msg.tool_calls:
+                    tool_calls = _aux_msg.tool_calls
 
             for tc in tool_calls:
                 if tc.function.name == "memory":
@@ -7575,7 +7772,12 @@ class AIAgent:
             except Exception:
                 pass
 
-        compressed = self.context_compressor.compress(messages, current_tokens=approx_tokens, focus_topic=focus_topic)
+        try:
+            compressed = self.context_compressor.compress(messages, current_tokens=approx_tokens, focus_topic=focus_topic)
+        except TypeError:
+            # Plugin context engine with strict signature that doesn't accept
+            # focus_topic — fall back to calling without it.
+            compressed = self.context_compressor.compress(messages, current_tokens=approx_tokens)
 
         todo_snapshot = self._todo_store.format_for_injection()
         if todo_snapshot:
@@ -8582,12 +8784,12 @@ class AIAgent:
                                    is_oauth=self._is_anthropic_oauth,
                                    preserve_dots=self._anthropic_preserve_dots())
                     summary_response = self._anthropic_messages_create(_ant_kw)
-                    _sum_nr = _tsum.normalize_response(summary_response, strip_tool_prefix=self._is_anthropic_oauth)
-                    final_response = (_sum_nr.content or "").strip()
+                    _summary_result = _tsum.normalize_response(summary_response, strip_tool_prefix=self._is_anthropic_oauth)
+                    final_response = (_summary_result.content or "").strip()
                 else:
                     summary_response = self._ensure_primary_openai_client(reason="iteration_limit_summary").chat.completions.create(**summary_kwargs)
-                    _sum_cc_nr = self._get_transport().normalize_response(summary_response)
-                    final_response = (_sum_cc_nr.content or "").strip()
+                    _summary_result = self._get_transport().normalize_response(summary_response)
+                    final_response = (_summary_result.content or "").strip()
 
             if final_response:
                 if "<think>" in final_response:
@@ -8612,8 +8814,8 @@ class AIAgent:
                                     max_tokens=self.max_tokens, reasoning_config=self.reasoning_config,
                                     preserve_dots=self._anthropic_preserve_dots())
                     retry_response = self._anthropic_messages_create(_ant_kw2)
-                    _retry_nr = _tretry.normalize_response(retry_response, strip_tool_prefix=self._is_anthropic_oauth)
-                    final_response = (_retry_nr.content or "").strip()
+                    _retry_result = _tretry.normalize_response(retry_response, strip_tool_prefix=self._is_anthropic_oauth)
+                    final_response = (_retry_result.content or "").strip()
                 else:
                     summary_kwargs = {
                         "model": self.model,
@@ -8627,8 +8829,8 @@ class AIAgent:
                         summary_kwargs["extra_body"] = summary_extra_body
 
                     summary_response = self._ensure_primary_openai_client(reason="iteration_limit_summary_retry").chat.completions.create(**summary_kwargs)
-                    _retry_cc_nr = self._get_transport().normalize_response(summary_response)
-                    final_response = (_retry_cc_nr.content or "").strip()
+                    _retry_result = self._get_transport().normalize_response(summary_response)
+                    final_response = (_retry_result.content or "").strip()
 
                 if final_response:
                     if "<think>" in final_response:
@@ -9290,12 +9492,13 @@ class AIAgent:
             
             api_start_time = time.time()
             retry_count = 0
-            max_retries = 3
+            max_retries = self._api_max_retries
             primary_recovery_attempted = False
             max_compression_attempts = 3
             codex_auth_retry_attempted=False
             anthropic_auth_retry_attempted=False
             nous_auth_retry_attempted=False
+            copilot_auth_retry_attempted=False
             thinking_sig_retry_attempted = False
             has_retried_429 = False
             restart_with_compressed_messages = False
@@ -9453,28 +9656,47 @@ class AIAgent:
                                 response_invalid = True
                                 error_details.append("response is None")
                             else:
-                                # output_text fallback: stream backfill may have failed
-                                # but normalize can still recover from output_text
-                                _out_text = getattr(response, "output_text", None)
-                                _out_text_stripped = _out_text.strip() if isinstance(_out_text, str) else ""
-                                if _out_text_stripped:
-                                    logger.debug(
-                                        "Codex response.output is empty but output_text is present "
-                                        "(%d chars); deferring to normalization.",
-                                        len(_out_text_stripped),
+                                # Provider returned a terminal failure (e.g. quota exhaustion).
+                                # Treat as invalid so the fallback chain is triggered instead of
+                                # letting the error bubble up outside the retry/fallback loop.
+                                _codex_resp_status = str(getattr(response, "status", "") or "").strip().lower()
+                                if _codex_resp_status in {"failed", "cancelled"}:
+                                    _codex_error_obj = getattr(response, "error", None)
+                                    _codex_error_msg = (
+                                        _codex_error_obj.get("message") if isinstance(_codex_error_obj, dict)
+                                        else str(_codex_error_obj) if _codex_error_obj
+                                        else f"Responses API returned status '{_codex_resp_status}'"
                                     )
-                                else:
-                                    _resp_status = getattr(response, "status", None)
-                                    _resp_incomplete = getattr(response, "incomplete_details", None)
-                                    logger.warning(
-                                        "Codex response.output is empty after stream backfill "
-                                        "(status=%s, incomplete_details=%s, model=%s). %s",
-                                        _resp_status, _resp_incomplete,
-                                        getattr(response, "model", None),
-                                        f"api_mode={self.api_mode} provider={self.provider}",
+                                    logging.warning(
+                                        "Codex response status='%s' (error=%s). Routing to fallback. %s",
+                                        _codex_resp_status, _codex_error_msg,
+                                        self._client_log_context(),
                                     )
                                     response_invalid = True
-                                    error_details.append("response.output is empty")
+                                    error_details.append(f"response.status={_codex_resp_status}: {_codex_error_msg}")
+                                else:
+                                    # output_text fallback: stream backfill may have failed
+                                    # but normalize can still recover from output_text
+                                    _out_text = getattr(response, "output_text", None)
+                                    _out_text_stripped = _out_text.strip() if isinstance(_out_text, str) else ""
+                                    if _out_text_stripped:
+                                        logger.debug(
+                                            "Codex response.output is empty but output_text is present "
+                                            "(%d chars); deferring to normalization.",
+                                            len(_out_text_stripped),
+                                        )
+                                    else:
+                                        _resp_status = getattr(response, "status", None)
+                                        _resp_incomplete = getattr(response, "incomplete_details", None)
+                                        logger.warning(
+                                            "Codex response.output is empty after stream backfill "
+                                            "(status=%s, incomplete_details=%s, model=%s). %s",
+                                            _resp_status, _resp_incomplete,
+                                            getattr(response, "model", None),
+                                            f"api_mode={self.api_mode} provider={self.provider}",
+                                        )
+                                        response_invalid = True
+                                        error_details.append("response.output is empty")
                     elif self.api_mode == "anthropic_messages":
                         _tv = self._get_transport()
                         if not _tv.validate_response(response):
@@ -9657,13 +9879,13 @@ class AIAgent:
                     elif self.api_mode == "bedrock_converse":
                         # Bedrock response already normalized at dispatch — use transport
                         _bt_fr = self._get_transport()
-                        _bt_fr_nr = _bt_fr.normalize_response(response)
-                        finish_reason = _bt_fr_nr.finish_reason
+                        _bedrock_result = _bt_fr.normalize_response(response)
+                        finish_reason = _bedrock_result.finish_reason
                     else:
                         _cc_fr = self._get_transport()
-                        _cc_fr_nr = _cc_fr.normalize_response(response)
-                        finish_reason = _cc_fr_nr.finish_reason
-                        assistant_message = self._nr_to_assistant_message(_cc_fr_nr)
+                        _finish_result = _cc_fr.normalize_response(response)
+                        finish_reason = _finish_result.finish_reason
+                        assistant_message = _finish_result
                         if self._should_treat_stop_as_truncated(
                             finish_reason,
                             assistant_message,
@@ -9688,12 +9910,12 @@ class AIAgent:
                         _trunc_msg = None
                         _trunc_transport = self._get_transport()
                         if self.api_mode == "anthropic_messages":
-                            _trunc_nr = _trunc_transport.normalize_response(
+                            _trunc_result = _trunc_transport.normalize_response(
                                 response, strip_tool_prefix=self._is_anthropic_oauth
                             )
                         else:
-                            _trunc_nr = _trunc_transport.normalize_response(response)
-                        _trunc_msg = self._nr_to_assistant_message(_trunc_nr)
+                            _trunc_result = _trunc_transport.normalize_response(response)
+                        _trunc_msg = _trunc_result
 
                         _trunc_content = getattr(_trunc_msg, "content", None) if _trunc_msg else None
                         _trunc_has_tool_calls = bool(getattr(_trunc_msg, "tool_calls", None)) if _trunc_msg else False
@@ -10240,6 +10462,15 @@ class AIAgent:
                         print(f"{self.log_prefix}     • Check credits / billing: https://portal.nousresearch.com")
                         print(f"{self.log_prefix}     • Verify stored credentials: {_dhh}/auth.json")
                         print(f"{self.log_prefix}     • Switch providers temporarily: /model <model> --provider openrouter")
+                    if (
+                        self.provider == "copilot"
+                        and status_code == 401
+                        and not copilot_auth_retry_attempted
+                    ):
+                        copilot_auth_retry_attempted = True
+                        if self._try_refresh_copilot_client_credentials():
+                            self._vprint(f"{self.log_prefix}🔐 Copilot credentials refreshed after 401. Retrying request...")
+                            continue
                     if (
                         self.api_mode == "anthropic_messages"
                         and status_code == 401
@@ -10439,14 +10670,14 @@ class AIAgent:
                     )
                     if is_rate_limited and self._fallback_index < len(self._fallback_chain):
                         # Don't eagerly fallback if credential pool rotation may
-                        # still recover.  The pool's retry-then-rotate cycle needs
-                        # at least one more attempt to fire — jumping to a fallback
-                        # provider here short-circuits it.
-                        pool = self._credential_pool
-                        pool_may_recover = pool is not None and pool.has_available()
+                        # still recover.  See _pool_may_recover_from_rate_limit
+                        # for the single-credential-pool exception.  Fixes #11314.
+                        pool_may_recover = _pool_may_recover_from_rate_limit(
+                            self._credential_pool
+                        )
                         if not pool_may_recover:
                             self._emit_status("⚠️ Rate limited — switching to fallback provider...")
-                            if self._try_activate_fallback():
+                            if self._try_activate_fallback(reason=classified.reason):
                                 retry_count = 0
                                 compression_attempts = 0
                                 primary_recovery_attempted = False
@@ -10594,9 +10825,30 @@ class AIAgent:
                         # Error is about the INPUT being too large — reduce context_length.
                         # Try to parse the actual limit from the error message
                         parsed_limit = parse_context_limit_from_error(error_msg)
+                        _provider_lower = (getattr(self, "provider", "") or "").lower()
+                        _base_lower = (getattr(self, "base_url", "") or "").rstrip("/").lower()
+                        is_minimax_provider = (
+                            _provider_lower in {"minimax", "minimax-cn"}
+                            or _base_lower.startswith((
+                                "https://api.minimax.io/anthropic",
+                                "https://api.minimaxi.com/anthropic",
+                            ))
+                        )
+                        minimax_delta_only_overflow = (
+                            is_minimax_provider
+                            and parsed_limit is None
+                            and "context window exceeds limit (" in error_msg
+                        )
                         if parsed_limit and parsed_limit < old_ctx:
                             new_ctx = parsed_limit
-                            self._vprint(f"{self.log_prefix}⚠️  Context limit detected from API: {new_ctx:,} tokens (was {old_ctx:,})", force=True)
+                            self._vprint(f"{self.log_prefix}Context limit detected from API: {new_ctx:,} tokens (was {old_ctx:,})", force=True)
+                        elif minimax_delta_only_overflow:
+                            new_ctx = old_ctx
+                            self._vprint(
+                                f"{self.log_prefix}Provider reported overflow amount only; "
+                                f"keeping context_length at {old_ctx:,} tokens and compressing.",
+                                force=True,
+                            )
                         else:
                             # Step down to the next probe tier
                             new_ctx = get_next_probe_tier(old_ctx)
@@ -10678,9 +10930,26 @@ class AIAgent:
                     # already accounts for 413, 429, 529 (transient), context
                     # overflow, and generic-400 heuristics.  Local validation
                     # errors (ValueError, TypeError) are programming bugs.
+                    # Exclude UnicodeEncodeError — it's a ValueError subclass
+                    # but is handled separately by the surrogate sanitization
+                    # path above.  Exclude json.JSONDecodeError — also a
+                    # ValueError subclass, but it indicates a transient
+                    # provider/network failure (malformed response body,
+                    # truncated stream, routing layer corruption), not a
+                    # local programming bug, and should be retried (#14782).
                     is_local_validation_error = (
                         isinstance(api_error, (ValueError, TypeError))
-                        and not isinstance(api_error, UnicodeEncodeError)
+                        and not isinstance(
+                            api_error, (UnicodeEncodeError, json.JSONDecodeError)
+                        )
+                        # ssl.SSLError (and its subclass SSLCertVerificationError)
+                        # inherits from OSError *and* ValueError via Python MRO,
+                        # so the isinstance(ValueError) check above would
+                        # misclassify a TLS transport failure as a local
+                        # programming bug and abort without retrying.  Exclude
+                        # ssl.SSLError explicitly so the error classifier's
+                        # retryable=True mapping takes effect instead.
+                        and not isinstance(api_error, ssl.SSLError)
                     )
                     is_client_error = (
                         is_local_validation_error
@@ -10928,9 +11197,9 @@ class AIAgent:
                 _normalize_kwargs = {}
                 if self.api_mode == "anthropic_messages":
                     _normalize_kwargs["strip_tool_prefix"] = self._is_anthropic_oauth
-                _nr = _transport.normalize_response(response, **_normalize_kwargs)
-                assistant_message = self._nr_to_assistant_message(_nr)
-                finish_reason = _nr.finish_reason
+                normalized = _transport.normalize_response(response, **_normalize_kwargs)
+                assistant_message = normalized
+                finish_reason = normalized.finish_reason
                 
                 # Normalize content to string — some OpenAI-compatible servers
                 # (llama-server, etc.) return content as a dict or list instead

scripts/compression_eval/DESIGN.md

@@ -0,0 +1,377 @@
+# Compression Eval — Design
+
+Status: proposal. Nothing under `scripts/compression_eval/` runs in CI.
+This is an offline tool authors run before merging prompt or algorithm
+changes to `agent/context_compressor.py`.
+
+## Why
+
+We tune the compressor prompt and the `_template_sections` checklist by
+hand, ship, and wait for the next real session to notice regressions.
+There is no automated check that a prompt edit still preserves file
+paths, error messages, or the active task across a compression.
+
+Factory.ai's December 2025 write-up
+(https://factory.ai/news/evaluating-compression) describes a
+probe-based eval that scores compressed state on six dimensions. The
+methodology is the valuable part — the benchmarks in the post are a
+marketing piece. We adopt the methodology and discard the scoreboard.
+
+## Goal
+
+Given a real session transcript and a bank of probe questions that
+exercise what the transcript contained, answer:
+
+1. After `ContextCompressor.compress()` runs, can the agent still
+   answer each probe correctly from the compressed state?
+2. Which of the six dimensions (accuracy, context awareness, artifact
+   trail, completeness, continuity, instruction following) is the
+   prompt weakest on?
+3. Does a prompt change improve or regress any dimension vs. the
+   previous run?
+
+That is the full scope. No "compare against OpenAI and Anthropic"
+benchmarking, no public scoreboard, no marketing claims.
+
+## Non-goals
+
+- Not a pytest. Requires API credentials, costs money, takes minutes
+  per fixture, and output is LLM-graded and non-deterministic.
+- Not part of `scripts/run_tests.sh`. Not invoked by CI.
+- Not a replacement for the existing compressor unit tests in
+  `tests/agent/test_context_compressor.py` — those stay as the
+  structural / boundary / tool-pair-sanitization guard.
+- Not a general trajectory eval. Scoped to context compaction only.
+
+## Where it lives
+
+```
+scripts/compression_eval/
+├── DESIGN.md                 # this file
+├── README.md                 # how to run, cost expectations, caveats
+├── run_eval.py               # entry point (fire CLI, like sample_and_compress.py)
+├── scrub_fixtures.py         # regenerate fixtures from ~/.hermes/sessions/*.jsonl
+├── fixtures/                 # checked-in scrubbed session snapshots
+│   ├── feature-impl-context-priority.json
+│   ├── debug-session-feishu-id-model.json
+│   └── config-build-competitive-scouts.json
+├── probes/                   # probe banks paired with fixtures
+│   └── <fixture>.probes.json
+├── rubric.py                 # grading prompt + dimension definitions
+├── grader.py                 # judge-model call + score parsing
+├── compressor_driver.py      # thin wrapper over ContextCompressor
+└── results/                  # gitignored; timestamped output per run
+    └── .gitkeep
+```
+
+`scripts/` is the right home: offline tooling, no CI involvement,
+precedent already set by `sample_and_compress.py`,
+`contributor_audit.py`, `discord-voice-doctor.py`.
+
+`environments/` is for Atropos RL training environments — wrong shape.
+`tests/` is hermetic and credential-free — incompatible with a
+probe-based eval that needs a judge model.
+
+## Fixture format
+
+A fixture is a single compressed-enough conversation captured from a
+real session. Stored as JSON (pretty-printed, reviewable in PRs):
+
+```json
+{
+  "name": "401-debug",
+  "description": "178-turn session debugging a 401 on /api/auth/login",
+  "model": "anthropic/claude-sonnet-4.6",
+  "context_length": 200000,
+  "messages": [
+    {"role": "system", "content": "..."},
+    {"role": "user", "content": "..."},
+    {"role": "assistant", "content": "...", "tool_calls": [...]},
+    {"role": "tool", "tool_call_id": "...", "content": "..."}
+  ],
+  "notes": "Captured 2026-04-24 from session 20260424_*.jsonl; \
+            PII scrubbed; secrets redacted via redact_sensitive_text."
+}
+```
+
+### Sourcing fixtures
+
+Fixtures are scrubbed snapshots of real sessions from the
+maintainer's `~/.hermes/sessions/*.jsonl` store, generated
+reproducibly by `scrub_fixtures.py` in this directory. Re-run the
+scrubber with `python3 scripts/compression_eval/scrub_fixtures.py`
+to regenerate them after a scrubber change.
+
+Three shipped fixtures cover three different session shapes:
+
+| Fixture | Source shape | Messages | Tokens (rough) | Tests |
+|---|---|---|---|---|
+| `feature-impl-context-priority` | investigate → patch → test → PR → merge | 75 | ~45k | continuation, artifact trail (2 files modified, 1 PR, ~16k skill_view in head) |
+| `debug-session-feishu-id-model` | PR triage + upstream docs + decision | 59 | ~28k | recall (PR #, error shape), decision (outcome + reason), large PR diff blocks |
+| `config-build-competitive-scouts` | iterative config: 11 cron jobs across 7 weekdays | 61 | ~26k | artifact trail (which jobs, which days), iterative-merge |
+
+The `~26k-45k` token range is below the default 50%-of-200k
+compression threshold, so the eval will always **force** a
+`compress()` call rather than wait for the natural trigger. That is
+the intended shape — we want a controlled single-shot compression so
+score deltas are attributable to the prompt change, not to whether
+the threshold happened to fire at the same boundary twice.
+
+### Scrubber pipeline
+
+`scrub_fixtures.py` applies, per message:
+
+1. `agent.redact.redact_sensitive_text` — API keys, tokens,
+   connection strings
+2. Username paths: `/home/teknium` → `/home/user`
+3. Personal handles: all case variants of the maintainer name → `user`
+4. Email addresses → `contributor@example.com`; git
+   `Author: Name <addr>` header lines normalised
+5. `<REASONING_SCRATCHPAD>...</REASONING_SCRATCHPAD>` and
+   `<think>...</think>` stripped from assistant content
+6. Messaging-platform user mentions (`<@123456>`, `<@***>`) →
+   `<@user>`
+7. First user message paraphrased to remove personal voice;
+   subsequent user turns kept verbatim after the redactions above
+8. System prompt replaced with a generic public-safe placeholder so
+   we don't check in the maintainer's tuned soul/skills/memory system
+   block
+9. Orphan empty-assistant messages (artifact of scratchpad-only
+   turns) and trailing tool messages with no matching assistant are
+   dropped
+10. Tool outputs preserved verbatim. An earlier iteration truncated
+    > 2KB tool bodies to keep fixture JSON small, but that defeats
+    the purpose: real sessions have 30KB `skill_view` dumps, 10KB
+    `read_file` outputs, 5KB `web_extract` bodies — compression has
+    to handle them. Truncation is now a no-op; the pipeline note
+    remains in `scrubbing_passes` for audit trail clarity.
+
+Before every fixture PR: grep the fixture for PII patterns. An
+audit is embedded at the bottom of the scrubber as comments.
+
+**Fixtures must stay small.** Target <200 KB per fixture, <500 KB
+total for the directory. Current total: ~410 KB across three
+fixtures.  Larger sessions are truncated with a
+`truncated_to: <index>` field in the fixture header so the cut is
+reviewable.
+
+## Probe format
+
+One probe file per fixture, so reviewers can see the question bank
+evolve alongside the fixture:
+
+```json
+{
+  "fixture": "401-debug",
+  "probes": [
+    {
+      "id": "recall-error-code",
+      "type": "recall",
+      "question": "What was the original error code and endpoint?",
+      "expected_facts": ["401", "/api/auth/login"]
+    },
+    {
+      "id": "artifact-files-modified",
+      "type": "artifact",
+      "question": "Which files have been modified in this session?",
+      "expected_facts": ["session_store.py", "redis_client.py"]
+    },
+    {
+      "id": "continuation-next-step",
+      "type": "continuation",
+      "question": "What should we do next?",
+      "expected_facts": ["re-run the integration tests", "restart the worker"]
+    },
+    {
+      "id": "decision-redis-approach",
+      "type": "decision",
+      "question": "What did we decide about the Redis issue?",
+      "expected_facts": ["switch to redis-py 5.x", "pooled connection"]
+    }
+  ]
+}
+```
+
+The four probe types come directly from Factory's methodology:
+**recall, artifact, continuation, decision**. `expected_facts` gives
+the grader concrete anchors instead of relying purely on LLM taste.
+
+Authoring a probe bank is a one-time cost per fixture. 8-12 probes per
+fixture is the target — enough to cover all four types, few enough to
+grade in under a minute at reasonable cost.
+
+## Grading
+
+Each probe gets scored 0-5 on **six dimensions** (Factory's six):
+
+| Dimension             | What it measures                                    |
+|-----------------------|-----------------------------------------------------|
+| accuracy              | File paths, function names, error codes are correct |
+| context_awareness     | Reflects current state, not a mid-session snapshot  |
+| artifact_trail        | Knows which files were read / modified / created    |
+| completeness          | Addresses all parts of the probe                    |
+| continuity            | Agent can continue without re-fetching              |
+| instruction_following | Probe answered in the requested form                |
+
+Grading is done by a single judge-model call per probe with a
+deterministic rubric prompt (see `rubric.py`). The rubric includes the
+`expected_facts` list so the judge has a concrete anchor. Default
+judge model: whatever the user has configured as their main model at
+run time (same resolution path as `auxiliary_client.call_llm`). A
+`--judge-model` flag allows overriding for consistency across runs.
+
+Non-determinism caveat: two runs of the same fixture will produce
+different scores. A single run means nothing. Report medians over
+N=3 runs by default, and require an improvement of >=0.3 on any
+dimension before claiming a prompt change is a win.
+
+## Run flow
+
+```
+python scripts/compression_eval/run_eval.py [OPTIONS]
+```
+
+Options (fire-style, mirroring `sample_and_compress.py`):
+
+| Flag                   | Default    | Purpose                                   |
+|------------------------|------------|-------------------------------------------|
+| `--fixtures`           | all        | Comma-separated fixture names             |
+| `--runs`               | 3          | Runs per fixture (for median)             |
+| `--judge-model`        | auto       | Override judge model                      |
+| `--compressor-model`   | auto       | Override model used *inside* the compressor |
+| `--label`              | timestamp  | Subdirectory under `results/`             |
+| `--focus-topic`        | none       | Pass-through to `compress(focus_topic=)`  |
+| `--compare-to`         | none       | Path to a previous run for diff output    |
+
+Steps per fixture per run:
+
+1. Load fixture JSON and probe bank.
+2. Construct a `ContextCompressor` against the fixture's model.
+3. Call `compressor.compress(messages)` — capture the compressed
+   message list.
+4. For each probe: ask the judge model to role-play as the continuing
+   agent with only the compressed state, then grade the answer on the
+   six dimensions using `rubric.py`.
+5. Write a per-run JSON to `results/<label>/<fixture>-run-N.json`.
+6. After all runs, emit a markdown summary to
+   `results/<label>/report.md`.
+
+## Report format
+
+Pasted verbatim into PR descriptions that touch the compressor:
+
+```
+## Compression eval — label 2026-04-25_13-40-02
+
+Main model: anthropic/claude-sonnet-4.6   Judge: same
+3 runs per fixture, medians reported.
+
+| Fixture        | Accuracy | Context | Artifact | Complete | Continuity | Instruction | Overall |
+|----------------|----------|---------|----------|----------|------------|-------------|---------|
+| 401-debug      | 4.1      | 4.0     | 2.5      | 4.3      | 3.8        | 5.0         | 3.95    |
+| pr-review      | 3.9      | 3.8     | 3.1      | 4.2      | 3.9        | 5.0         | 3.98    |
+| feature-impl   | 4.0      | 3.9     | 2.9      | 4.1      | 4.0        | 5.0         | 3.98    |
+
+Per-probe misses (score < 3.0):
+- 401-debug / artifact-files-modified: 1.7 — summary dropped redis_client.py
+- pr-review / decision-auth-rewrite: 2.3 — outcome captured, reasoning dropped
+```
+
+## Cost expectations
+
+Dominated by the judge calls. For 3 fixtures × 10 probes × 3 runs =
+90 judge calls per eval run. On Claude Sonnet 4.6 that is roughly
+$0.50-$1.50 per full eval depending on probe length. The compressor
+itself makes 1 call per fixture × 3 runs = 9 additional calls.
+
+**This is not a check to run after every commit.** It is a
+before-merge check for PRs that touch:
+
+- `agent/context_compressor.py` — any change to `_template_sections`,
+  `_generate_summary`, or `compress()`.
+- `agent/auxiliary_client.py` — when changing how compression tasks
+  are routed.
+- `agent/prompt_builder.py` — when the compression-note phrasing
+  changes.
+
+## Open questions (to resolve before implementing)
+
+1. **Fixture scrubbing: manual or scripted?** A scripted scrub that
+   also replaces project names / hostnames would lower the cost of
+   contributing a new fixture. Risk: over-aggressive replacement
+   destroys the signal the probe depends on. Propose: start manual,
+   add scripted helpers once we have 3 fixtures and know the common
+   PII shapes.
+
+2. **Judge model selection.** Factory uses GPT-5.2. We can't pin one
+   — user's main model changes. Options: (a) grade with main model
+   (cheap, inconsistent across users), (b) require a specific judge
+   model (e.g. `claude-sonnet-4.6`), inconsistent for users without
+   access. Propose (a) with a `--judge-model` override, and make the
+   model name prominent in the report so comparisons across machines
+   are legible.
+
+3. **Noise floor.** Before landing prompt changes, run the current
+   prompt N=10 times to measure per-dimension stddev. That tells us
+   the minimum delta to call a change significant. Suspect 0.2-0.3 on
+   a 0-5 scale. Decision deferred until after the first fixture is
+   landed.
+
+4. **Iterative-merge coverage.** The real Factory-vs-Anthropic
+   difference is incremental merge vs. regenerate. A fixture that
+   only compresses once doesn't exercise our iterative path. Add a
+   fourth fixture that forces two compressions (manually chained),
+   with probes that test whether information from the first
+   compression survives the second. Deferred to a follow-up PR.
+
+## Implementation status
+
+This PR ships the full eval end-to-end:
+
+- `scrub_fixtures.py` — reproducible scrubber
+- `fixtures/` — three scrubbed session fixtures
+- `probes/` — three probe banks (10-11 probes each, all four types)
+- `rubric.py` — six-dimension grading rubric + judge-prompt builder + response parser
+- `compressor_driver.py` — thin wrapper around `ContextCompressor` for forced single-shot compression
+- `grader.py` — two-phase continuation + grading calls via OpenAI SDK
+- `report.py` — markdown report renderer + `--compare-to` delta mode + per-run JSON dumper
+- `run_eval.py` — entry point (`fire`-style CLI)
+- `tests/scripts/test_compression_eval.py` — 33 unit tests covering rubric parsing, report rendering, fixture/probe loading, and a PII smoke test on the fixtures (LLM paths not tested — they require credentials and are exercised by the eval itself)
+
+### Noise floor — one empirical data point
+
+A single same-inputs re-run of `debug-session-feishu-id-model`
+(compressor + judge = `openai/gpt-5.4-mini` via Nous Portal,
+runs=1) produced:
+
+- Run A overall: 3.25
+- Run B overall: 3.17 (delta -0.08)
+
+Individual dimensions varied by up to ±0.5 between the two runs on
+single-run medians. This confirms DESIGN.md's "< 0.3 is noise"
+guidance is the right order of magnitude for a single-run
+comparison. With `runs=3` default, per-dimension variance should
+tighten; noise-floor measurement at N=10 is still a useful
+follow-up to calibrate precisely.
+
+## Open follow-ups (not blocking this PR)
+
+1. **Iterative-merge fixture** — our actual compression win over
+   "regenerate from scratch" approaches is only exercised when
+   `_previous_summary` is re-used on a second compression. None of
+   the three shipped fixtures force two compressions. The natural
+   basis is `config-build-competitive-scouts` (already iterative by
+   shape); splitting it at the Monday/Tuesday boundary would force
+   the second compression to merge rather than regenerate.
+2. **Noise-floor precision** — run the current prompt N=10 times
+   against one fixture to pin down per-dimension stddev and publish
+   the numbers in README.
+3. **Scripted scrubber helpers** — the current scrubber is manual
+   per-fixture. A helper that identifies candidate sessions to
+   scrub (by shape or by keyword) would lower the cost of adding
+   fixture #4+.
+4. **Judge model selection policy** — current code uses whatever
+   the user passes as `--judge-model` (default: same as compressor).
+   Pinning the judge across users would stabilise cross-machine
+   comparisons, at the cost of blocking users without access to
+   the pinned model.

scripts/compression_eval/README.md

@@ -0,0 +1,110 @@
+# compression_eval
+
+Offline eval harness for `agent/context_compressor.py`. Runs a real
+conversation transcript through the compressor, then probes the
+compressed state with targeted questions graded on six dimensions.
+
+## When to run
+
+Before merging changes to:
+
+- `agent/context_compressor.py` — any change to `_template_sections`,
+  `_generate_summary`, `compress()`, or its boundary logic
+- `agent/auxiliary_client.py` — when changing how compression tasks
+  are routed
+- `agent/prompt_builder.py` — when the compression-note phrasing
+  changes
+
+## Not for CI
+
+This harness makes real model calls (compressor + continuation +
+judge = ~3 calls per probe × probes per fixture × runs). Costs ~$0.50
+to ~$1.50 per full run depending on models, takes minutes, is
+LLM-graded (non-deterministic). It lives in `scripts/` and is
+invoked by hand. `tests/` and `scripts/run_tests.sh` do not touch it.
+
+`tests/scripts/test_compression_eval.py` covers the non-LLM code
+paths (rubric parsing, report rendering, fixture/probe loading, PII
+smoke check on the checked-in fixtures) and DOES run in CI.
+
+## Usage
+
+```bash
+# Run all three fixtures, 3 runs each, with your configured provider
+python3 scripts/compression_eval/run_eval.py
+
+# Faster iteration — one fixture, one run
+python3 scripts/compression_eval/run_eval.py \
+    --fixtures=debug-session-feishu-id-model --runs=1
+
+# Pin a cheap model for both compression + judge (recommended)
+python3 scripts/compression_eval/run_eval.py \
+    --compressor-provider=nous --compressor-model=openai/gpt-5.4-mini \
+    --judge-provider=nous      --judge-model=openai/gpt-5.4-mini \
+    --runs=3 --label=baseline
+
+# After editing context_compressor.py, rerun with a new label and diff
+python3 scripts/compression_eval/run_eval.py \
+    --compressor-provider=nous --compressor-model=openai/gpt-5.4-mini \
+    --judge-provider=nous      --judge-model=openai/gpt-5.4-mini \
+    --runs=3 --label=my-prompt-tweak \
+    --compare-to=results/baseline
+```
+
+Results land in `results/<label>/report.md` and are intended to be
+pasted verbatim into PR descriptions. `--compare-to` renders a delta
+column per dimension so reviewers can see "did this actually help?"
+at a glance.
+
+Rule of thumb: dimension deltas below ±0.3 are within run-to-run
+noise on `runs=3`. Publish a bigger N if you want tighter bounds.
+
+## Fixtures
+
+Three scrubbed session snapshots live under `fixtures/`:
+
+- `feature-impl-context-priority.json` — 75 msgs, investigate →
+  patch → test → PR → merge
+- `debug-session-feishu-id-model.json` — 59 msgs, PR triage +
+  upstream docs + decision
+- `config-build-competitive-scouts.json` — 61 msgs, iterative
+  config accumulation (11 cron jobs)
+
+Regenerate them from the maintainer's `~/.hermes/sessions/*.jsonl`
+with `python3 scripts/compression_eval/scrub_fixtures.py`. The
+scrubber pipeline and PII-audit checklist are documented in
+`DESIGN.md` under **Scrubber pipeline**.
+
+## Probes
+
+One probe bank per fixture under `probes/`, 10-11 probes each,
+covering all four types: **recall**, **artifact**, **continuation**,
+**decision**. Each probe carries an `expected_facts` list of concrete
+anchors (PR numbers, file paths, error codes, commands run) that the
+judge sees alongside the assistant's answer.
+
+## How it scores
+
+Six dimensions, 0-5 per probe:
+
+| Dimension             | What it measures                                     |
+|-----------------------|------------------------------------------------------|
+| accuracy              | File paths, function names, PR/issue numbers correct |
+| context_awareness     | Reflects current session state, not a snapshot       |
+| artifact_trail        | Correctly enumerates files / commands / PRs          |
+| completeness          | Addresses ALL parts of the probe                     |
+| continuity            | Next assistant could continue without re-fetching    |
+| instruction_following | Answer in the requested form                         |
+
+Report renders medians across N runs; probes scoring below 3.0
+overall surface in a separate section with the judge's specific
+complaint noted inline.
+
+## Related
+
+- `agent/context_compressor.py` — the thing under test
+- `tests/agent/test_context_compressor.py` — structural unit tests
+  that do run in CI
+- `scripts/sample_and_compress.py` — the closest existing script in
+  shape (offline, credential-requiring, not in CI)
+- `DESIGN.md` — full architecture + methodology + open follow-ups

scripts/compression_eval/compressor_driver.py

@@ -0,0 +1,114 @@
+"""Wraps ContextCompressor to run a single forced compression on a fixture.
+
+The real agent loop checks ``should_compress()`` before calling ``compress()``.
+Fixtures are intentionally sized below the 100k threshold so ``compress()``
+runs in a controlled, single-shot mode — score deltas attribute to the
+prompt change, not to whether the threshold happened to fire at the same
+boundary twice.
+
+Resolves the provider for the compression call via the same path the real
+agent uses (``hermes_cli.runtime_provider.resolve_runtime_provider``) so
+behaviour matches production aside from being a single call.
+"""
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+# Make sibling imports work whether invoked as a script or as a module.
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+from agent.context_compressor import (  # noqa: E402
+    ContextCompressor,
+    estimate_messages_tokens_rough,
+)
+
+
+def run_compression(
+    *,
+    messages: List[Dict[str, Any]],
+    compressor_model: str,
+    compressor_provider: str,
+    compressor_base_url: str,
+    compressor_api_key: str,
+    compressor_api_mode: str,
+    context_length: int,
+    focus_topic: Optional[str] = None,
+    summary_model_override: Optional[str] = None,
+) -> Dict[str, Any]:
+    """Run a single forced compression pass over the fixture messages.
+
+    Returns a dict with:
+      - compressed_messages: the post-compression message list
+      - summary_text: the summary produced (extracted from the compressed head)
+      - pre_tokens, post_tokens: rough token counts before/after
+      - compression_ratio: 1 - (post/pre)
+      - pre_message_count, post_message_count
+    """
+    compressor = ContextCompressor(
+        model=compressor_model,
+        threshold_percent=0.50,
+        protect_first_n=3,
+        protect_last_n=20,
+        summary_target_ratio=0.20,
+        quiet_mode=True,
+        summary_model_override=summary_model_override or "",
+        base_url=compressor_base_url,
+        api_key=compressor_api_key,
+        config_context_length=context_length,
+        provider=compressor_provider,
+        api_mode=compressor_api_mode,
+    )
+
+    pre_tokens = estimate_messages_tokens_rough(messages)
+    compressed = compressor.compress(
+        messages,
+        current_tokens=pre_tokens,
+        focus_topic=focus_topic,
+    )
+    post_tokens = estimate_messages_tokens_rough(compressed)
+
+    summary_text = _extract_summary_from_messages(compressed)
+
+    ratio = (1.0 - (post_tokens / pre_tokens)) if pre_tokens > 0 else 0.0
+
+    return {
+        "compressed_messages": compressed,
+        "summary_text": summary_text,
+        "pre_tokens": pre_tokens,
+        "post_tokens": post_tokens,
+        "compression_ratio": ratio,
+        "pre_message_count": len(messages),
+        "post_message_count": len(compressed),
+    }
+
+
+_SUMMARY_MARKERS = (
+    "## Active Task",
+    "## Goal",
+    "## Completed Actions",
+)
+
+
+def _extract_summary_from_messages(messages: List[Dict[str, Any]]) -> str:
+    """Find the structured summary block inside the compressed message list.
+
+    The compressor injects the summary as a user (or system-appended) message
+    near the head. We look for the section-header markers from
+    ``_template_sections`` in ``agent/context_compressor.py``.
+    """
+    for msg in messages:
+        content = msg.get("content")
+        if not isinstance(content, str):
+            if isinstance(content, list):
+                content = "\n".join(
+                    p.get("text", "") for p in content if isinstance(p, dict)
+                )
+            else:
+                continue
+        if any(marker in content for marker in _SUMMARY_MARKERS):
+            return content
+    return ""

scripts/compression_eval/grader.py

@@ -0,0 +1,181 @@
+"""Two-phase probe grading.
+
+Phase 1 — **Continuation**: simulate the next assistant turn. Feed the
+compressed message list plus the probe question and ask the continuing
+model to answer using only the compressed context. This is exactly what
+a real next-turn call would look like.
+
+Phase 2 — **Grading**: a separate judge-model call scores the answer on
+the six rubric dimensions using ``rubric.build_judge_prompt``.
+
+Both phases use the OpenAI SDK directly against the resolved provider
+endpoint, so the explicit api_key + base_url we pass always reaches the
+wire. (``agent.auxiliary_client.call_llm`` is designed for task-tagged
+auxiliary calls backed by config lookups; for eval we need the explicit
+credentials to win unconditionally.)
+"""
+from __future__ import annotations
+
+import logging
+import sys
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+from openai import OpenAI  # noqa: E402
+
+from rubric import build_judge_prompt, parse_judge_response  # noqa: E402
+
+logger = logging.getLogger(__name__)
+
+
+_CONTINUATION_SYSTEM = (
+    "You are the continuing assistant in a long session. Earlier turns have "
+    "been compacted into a handoff summary that is now part of the "
+    "conversation history. The user has just asked you a question. "
+    "Answer using ONLY what you can determine from the conversation history "
+    "you see (including the handoff summary). Do NOT invent details. If the "
+    "summary does not contain a specific fact, say so explicitly rather "
+    "than guessing. Be direct and concrete — cite file paths, PR numbers, "
+    "error codes, and exact values when they are present in the summary."
+)
+
+
+def answer_probe(
+    *,
+    compressed_messages: List[Dict[str, Any]],
+    probe_question: str,
+    model: str,
+    provider: str,
+    base_url: str,
+    api_key: str,
+    max_tokens: int = 1024,
+    timeout: Optional[float] = 120.0,
+) -> str:
+    """Run the continuation call: what does the next assistant answer?
+
+    Builds a messages list of [system_continuation, *compressed, probe_user]
+    and asks the configured model. Returns the answer content as a string.
+    """
+    # Strip any pre-existing system message from the compressed list and
+    # replace with our continuation system prompt. The fixture's generic
+    # system is not the right frame for the continuation simulation.
+    history = [m for m in compressed_messages if m.get("role") != "system"]
+    messages = (
+        [{"role": "system", "content": _CONTINUATION_SYSTEM}]
+        + _sanitize_for_chat_api(history)
+        + [{"role": "user", "content": probe_question}]
+    )
+
+    client = OpenAI(api_key=api_key, base_url=base_url, timeout=timeout)
+    response = client.chat.completions.create(
+        model=model,
+        messages=messages,
+        max_tokens=max_tokens,
+    )
+    content = response.choices[0].message.content
+    if not isinstance(content, str):
+        content = "" if content is None else str(content)
+    return content.strip()
+
+
+def grade_probe(
+    *,
+    probe_question: str,
+    probe_type: str,
+    expected_facts: List[str],
+    assistant_answer: str,
+    judge_model: str,
+    judge_provider: str,
+    judge_base_url: str,
+    judge_api_key: str,
+    max_tokens: int = 512,
+    timeout: Optional[float] = 120.0,
+) -> Dict[str, Any]:
+    """Run the judge call and parse the six dimension scores.
+
+    Returns dict {scores: {dim: int}, notes: str, overall: float,
+    raw: str, parse_error: str|None}. On parse failure, scores are zeros
+    and parse_error is populated — the caller decides whether to retry
+    or accept.
+    """
+    prompt = build_judge_prompt(
+        probe_question=probe_question,
+        probe_type=probe_type,
+        expected_facts=expected_facts,
+        assistant_answer=assistant_answer,
+    )
+    client = OpenAI(api_key=judge_api_key, base_url=judge_base_url, timeout=timeout)
+    response = client.chat.completions.create(
+        model=judge_model,
+        messages=[{"role": "user", "content": prompt}],
+        max_tokens=max_tokens,
+    )
+    raw = response.choices[0].message.content or ""
+    if not isinstance(raw, str):
+        raw = str(raw)
+
+    try:
+        parsed = parse_judge_response(raw)
+        parsed["raw"] = raw
+        parsed["parse_error"] = None
+        return parsed
+    except ValueError as exc:
+        logger.warning("Judge response parse failed: %s | raw=%r", exc, raw[:200])
+        from rubric import DIMENSIONS
+        return {
+            "scores": {d: 0 for d in DIMENSIONS},
+            "notes": "",
+            "overall": 0.0,
+            "raw": raw,
+            "parse_error": str(exc),
+        }
+
+
+def _sanitize_for_chat_api(
+    messages: List[Dict[str, Any]],
+) -> List[Dict[str, Any]]:
+    """Drop tool_calls/tool pairs that are incomplete.
+
+    A compressed message list may contain tool_call references whose matching
+    ``tool`` result was summarized away, which breaks strict-validator
+    providers (Anthropic, OpenAI). Easiest correct behaviour for the eval:
+    strip tool_calls entirely and drop ``tool`` role messages — the
+    continuation model only needs the summary + recent turns to answer the
+    probe, not the precise tool-call bookkeeping.
+    """
+    clean: List[Dict[str, Any]] = []
+    for m in messages:
+        role = m.get("role")
+        if role == "tool":
+            # Convert tool result to a plain user note so the continuation
+            # model still sees the content without needing the structured
+            # tool_call_id pairing.
+            content = m.get("content")
+            if isinstance(content, list):
+                content = "\n".join(
+                    p.get("text", "") for p in content if isinstance(p, dict)
+                )
+            clean.append({
+                "role": "user",
+                "content": f"[earlier tool result]\n{content or ''}",
+            })
+            continue
+        new = {"role": role, "content": m.get("content", "")}
+        # Drop tool_calls — the downstream assistant message's content
+        # still describes what the agent was doing.
+        clean.append(new)
+    # Collapse consecutive same-role turns into one (alternation rule)
+    merged: List[Dict[str, Any]] = []
+    for m in clean:
+        if merged and merged[-1]["role"] == m["role"]:
+            prev = merged[-1]
+            prev_c = prev.get("content") or ""
+            new_c = m.get("content") or ""
+            prev["content"] = f"{prev_c}\n\n{new_c}" if prev_c else new_c
+        else:
+            merged.append(m)
+    return merged

scripts/compression_eval/probes/config-build-competitive-scouts.probes.json

@@ -0,0 +1,96 @@
+{
+  "fixture": "config-build-competitive-scouts",
+  "description": "Probes for the competitive-scout cron-job setup session. Anchors are which agents were configured, which day of the week each runs, and the full final schedule. This fixture most directly tests artifact-trail and iterative-merge because the job list grows by one per user turn.",
+  "probes": [
+    {
+      "id": "recall-first-repo",
+      "type": "recall",
+      "question": "What was the first repository the user asked to create a scout cron for, and on what day of the week?",
+      "expected_facts": ["openclaw", "Sunday"]
+    },
+    {
+      "id": "recall-closed-source-target",
+      "type": "recall",
+      "question": "One of the scout targets does not have an open-source repository and had to be configured as a web scan instead. Which one, and on what day?",
+      "expected_facts": ["claude code", "Friday", "web scan"]
+    },
+    {
+      "id": "artifact-all-jobs",
+      "type": "artifact",
+      "question": "List every scout cron job created in this session.",
+      "expected_facts": [
+        "openclaw-pr-scout",
+        "nanoclaw-pr-scout",
+        "ironclaw-pr-scout",
+        "kilocode-pr-scout",
+        "codex-pr-scout",
+        "gemini-cli-pr-scout",
+        "cline-pr-scout",
+        "opencode-pr-scout",
+        "claude-code-scout",
+        "aider-pr-scout",
+        "roocode-pr-scout"
+      ]
+    },
+    {
+      "id": "artifact-final-schedule",
+      "type": "artifact",
+      "question": "What is the final weekly schedule? Give the day and the agents scanned on each day.",
+      "expected_facts": [
+        "Sun: openclaw, nanoclaw, ironclaw",
+        "Mon: kilo code",
+        "Tue: codex",
+        "Wed: gemini cli, cline",
+        "Thu: opencode",
+        "Fri: claude code",
+        "Sat: aider, roo"
+      ]
+    },
+    {
+      "id": "artifact-sunday-count",
+      "type": "artifact",
+      "question": "How many cron jobs run on Sunday?",
+      "expected_facts": ["3", "three", "openclaw, nanoclaw, ironclaw"]
+    },
+    {
+      "id": "artifact-total-count",
+      "type": "artifact",
+      "question": "How many scout cron jobs were created in total by the end of the session?",
+      "expected_facts": ["11", "eleven"]
+    },
+    {
+      "id": "decision-kilo-open-source",
+      "type": "decision",
+      "question": "The user asked whether Kilo Code is open source. What was the answer, and what did the user decide to do with it?",
+      "expected_facts": [
+        "yes, open source",
+        "Kilo-Org/kilocode",
+        "added as Monday scout"
+      ]
+    },
+    {
+      "id": "decision-saturday-fill",
+      "type": "decision",
+      "question": "Saturday was the last open day at one point. Which scout(s) were placed on Saturday, and why were those chosen?",
+      "expected_facts": ["aider", "roo", "filled in last based on openrouter popularity / cli comparison rankings"]
+    },
+    {
+      "id": "continuation-execution-time",
+      "type": "continuation",
+      "question": "At what local time of day do these scout cron jobs run?",
+      "expected_facts": ["10 AM Pacific", "17:00 UTC", "0 17 * * *"]
+    },
+    {
+      "id": "continuation-skill-used",
+      "type": "continuation",
+      "question": "Each scout job runs with a specific skill preloaded. Which one?",
+      "expected_facts": ["hermes-agent-dev"]
+    },
+    {
+      "id": "continuation-weekday-coverage",
+      "type": "continuation",
+      "question": "After the session ended, are there any weekdays still uncovered by a scout job?",
+      "expected_facts": ["no", "all 7 days covered", "full week loaded"]
+    }
+  ]
+}

scripts/compression_eval/probes/debug-session-feishu-id-model.probes.json

@@ -0,0 +1,72 @@
+{
+  "fixture": "debug-session-feishu-id-model",
+  "description": "Probes for the Feishu identity-model PR #8388 triage session. Anchors are the PR number, what the PR actually contained, what upstream docs confirmed, and the final decision + reasoning.",
+  "probes": [
+    {
+      "id": "recall-pr-number",
+      "type": "recall",
+      "question": "What is the PR number under review in this session, and what repository is it against?",
+      "expected_facts": ["PR #8388", "NousResearch/hermes-agent", "hermes-agent"]
+    },
+    {
+      "id": "recall-bug-claim",
+      "type": "recall",
+      "question": "What is the core bug the PR claims to fix? Be specific about the identifier involved.",
+      "expected_facts": ["open_id", "app-scoped", "not canonical", "Feishu identity model"]
+    },
+    {
+      "id": "recall-upstream-confirmation",
+      "type": "recall",
+      "question": "Do upstream Feishu/Lark docs confirm that open_id is app-scoped rather than a canonical cross-app identity?",
+      "expected_facts": ["yes", "confirmed", "open.feishu.cn", "same user has different Open IDs in different apps"]
+    },
+    {
+      "id": "artifact-pr-scope",
+      "type": "artifact",
+      "question": "Roughly how large is PR #8388, and which gateway subsystems does it touch beyond the Feishu adapter?",
+      "expected_facts": ["4647 lines", "gateway/run.py", "cron/scheduler.py", "gateway/config.py", "multi-account", "bind"]
+    },
+    {
+      "id": "artifact-new-tool",
+      "type": "artifact",
+      "question": "Does the PR add a new tool file? If so, what is its path?",
+      "expected_facts": ["tools/feishu_id_tool.py", "new file"]
+    },
+    {
+      "id": "decision-pr-assessment",
+      "type": "decision",
+      "question": "What is the reviewer's overall assessment of PR #8388 — approve, reject, or something more nuanced? Explain in one sentence.",
+      "expected_facts": [
+        "core claim is correct",
+        "scope is wrong",
+        "bait-and-switch",
+        "overbuilt",
+        "implement cleaner ourselves"
+      ]
+    },
+    {
+      "id": "decision-core-claim-validity",
+      "type": "decision",
+      "question": "Setting aside the PR's size, is the underlying identity-model concern technically valid or not?",
+      "expected_facts": ["technically valid", "correct", "open_id is app-scoped"]
+    },
+    {
+      "id": "continuation-next-action",
+      "type": "continuation",
+      "question": "Based on the review outcome, what is the next action the agent has been asked to take regarding this PR?",
+      "expected_facts": ["close the PR", "implement ourselves", "cleaner", "less complex"]
+    },
+    {
+      "id": "continuation-implementation-scope",
+      "type": "continuation",
+      "question": "If implementing the Feishu fix cleanly ourselves, which specific behaviour needs to change — what should replace the current use of open_id?",
+      "expected_facts": ["use union_id", "or user_id", "canonical identity", "cross-app stable ID"]
+    },
+    {
+      "id": "continuation-sources-to-reference",
+      "type": "continuation",
+      "question": "Which upstream documentation sources were fetched during review that should be referenced when writing the clean implementation?",
+      "expected_facts": ["open.feishu.cn", "open.larkoffice.com", "user-identity-introduction"]
+    }
+  ]
+}

scripts/compression_eval/probes/feature-impl-context-priority.probes.json

@@ -0,0 +1,74 @@
+{
+  "fixture": "feature-impl-context-priority",
+  "description": "Probes for the .hermes.md / AGENTS.md / CLAUDE.md / .cursorrules priority feature session. Anchors are the concrete facts the next assistant would need to continue: user's priority order, files modified, helper-function structure, live-test scenarios, and PR number.",
+  "probes": [
+    {
+      "id": "recall-priority-order",
+      "type": "recall",
+      "question": "What is the priority order the user asked for when multiple project-context files are present? List them from highest to lowest priority.",
+      "expected_facts": [".hermes.md", "AGENTS.md", "CLAUDE.md", ".cursorrules", "highest to lowest"]
+    },
+    {
+      "id": "recall-selection-mode",
+      "type": "recall",
+      "question": "When multiple context files exist in the same directory, does the agent now load all of them or pick only one?",
+      "expected_facts": ["only one", "priority-based selection", "highest-priority winner"]
+    },
+    {
+      "id": "artifact-files-modified",
+      "type": "artifact",
+      "question": "Which files in the hermes-agent repository were modified during this session? List them.",
+      "expected_facts": [
+        "agent/prompt_builder.py",
+        "tests/agent/test_prompt_builder.py"
+      ]
+    },
+    {
+      "id": "artifact-helper-functions",
+      "type": "artifact",
+      "question": "The session introduced separate helper functions for each context-file type. What are their names?",
+      "expected_facts": [
+        "_load_hermes_md",
+        "_load_agents_md",
+        "_load_claude_md",
+        "_load_cursorrules"
+      ]
+    },
+    {
+      "id": "artifact-test-scenarios",
+      "type": "artifact",
+      "question": "A scratch directory was created with scenario subdirectories to live-test the priority chain. Roughly how many scenarios, and what directory was it created under?",
+      "expected_facts": ["10 scenarios", "/tmp/context-priority-test"]
+    },
+    {
+      "id": "decision-claude-md-was-unsupported",
+      "type": "decision",
+      "question": "What was the finding about CLAUDE.md support in the existing loader before this session's changes?",
+      "expected_facts": ["CLAUDE.md was not handled", "not supported", "new handler added"]
+    },
+    {
+      "id": "decision-load-all-or-one",
+      "type": "decision",
+      "question": "Was the decision to load multiple context files when present, or to load only the highest-priority one? Explain the reasoning in one sentence.",
+      "expected_facts": ["load only one", "highest priority", "user preference", "do not want to load multiple"]
+    },
+    {
+      "id": "continuation-pr-number-and-status",
+      "type": "continuation",
+      "question": "A pull request was opened for this feature. What is the PR number and what is its merge status?",
+      "expected_facts": ["PR #2301", "merged", "squash"]
+    },
+    {
+      "id": "continuation-test-suite-result",
+      "type": "continuation",
+      "question": "What was the result of the full test suite run after the implementation changes?",
+      "expected_facts": ["5680 passed", "0 failures", "clean"]
+    },
+    {
+      "id": "continuation-next-step",
+      "type": "continuation",
+      "question": "If asked to pick up this session, what is the current state of main? Anything left to do?",
+      "expected_facts": ["merged to main", "main is current", "nothing outstanding", "pulled"]
+    }
+  ]
+}