docs(lsp): add feature page — setup, CLI, supported languages, troubleshooting

Covers: enable flow, server installation (detect-only default vs
hermes lsp install), how diagnostics reach the model, config knobs,
all 26 supported languages, and troubleshooting common issues.

.env.example

@@ -143,6 +143,18 @@
 # Also requires ~/.honcho/config.json with enabled=true (see README).
 # HONCHO_API_KEY=
 
+# =============================================================================
+# HYPERLIQUID OPTIONAL SKILL
+# =============================================================================
+# Optional defaults for the Hyperliquid skill in optional-skills/blockchain/hyperliquid
+#
+# Hyperliquid API base URL override
+# Default: https://api.hyperliquid.xyz
+# HYPERLIQUID_API_URL=https://api.hyperliquid-testnet.xyz
+#
+# Default address for account-level commands like state, fills, orders, and review
+# HYPERLIQUID_USER_ADDRESS=0x0000000000000000000000000000000000000000
+
 # =============================================================================
 # TERMINAL TOOL CONFIGURATION
 # =============================================================================

.github/workflows/lint.yml

@@ -122,7 +122,8 @@ jobs:
           retention-days: 14
 
       - name: Post / update PR comment
-        if: github.event_name == 'pull_request'
+        if: github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository
+        continue-on-error: true
         uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7
         with:
           script: |

AGENTS.md

@@ -540,10 +540,14 @@ Full authoring guide: `website/docs/developer-guide/model-provider-plugin.md`.
 
 ### 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`.
+`plugins/context_engine/`, `plugins/image_gen/`, 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`. Reference / docs-companion plugins
+(`example-dashboard`, `strike-freedom-cockpit`, `plugin-llm-example`,
+`plugin-llm-async-example`) live in the
+[`hermes-example-plugins`](https://github.com/NousResearch/hermes-example-plugins)
+companion repo, not in this tree.
 
 ---
 

SECURITY.md

@@ -1,84 +1,331 @@
 # Hermes Agent Security Policy
 
-This document outlines the security protocols, trust model, and deployment hardening guidelines for the **Hermes Agent** project.
+This document describes Hermes Agent's trust model, names the one
+security boundary the project treats as load-bearing, and defines the
+scope for vulnerability reports.
 
-## 1. Vulnerability Reporting
+## 1. Reporting a Vulnerability
 
-Hermes Agent does **not** operate a bug bounty program. Security issues should be reported via [GitHub Security Advisories (GHSA)](https://github.com/NousResearch/hermes-agent/security/advisories/new) or by emailing **security@nousresearch.com**. Do not open public issues for security vulnerabilities.
+Report privately via [GitHub Security Advisories](https://github.com/NousResearch/hermes-agent/security/advisories/new)
+or **security@nousresearch.com**. Do not open public issues for
+security vulnerabilities. **Hermes Agent does not operate a bug
+bounty program.**
 
-### Required Submission Details
-- **Title & Severity:** Concise description and CVSS score/rating.
-- **Affected Component:** Exact file path and line range (e.g., `tools/approval.py:120-145`).
-- **Environment:** Output of `hermes version`, commit SHA, OS, and Python version.
-- **Reproduction:** Step-by-step Proof-of-Concept (PoC) against `main` or the latest release.
-- **Impact:** Explanation of what trust boundary was crossed.
+A useful report includes:
+
+- A concise description and severity assessment.
+- The affected component, identified by file path and line range
+  (e.g. `path/to/file.py:120-145`).
+- Environment details (`hermes version`, commit SHA, OS, Python
+  version).
+- A reproduction against `main` or the latest release.
+- A statement of which trust boundary in §2 is crossed.
+
+Please read §2 and §3 before submitting. Reports that demonstrate
+limits of an in-process heuristic this policy does not treat as a
+boundary will be closed as out-of-scope under §3 — but see §3.2:
+they are still welcome as regular issues or pull requests, just not
+through the private security channel.
 
 ---
 
 ## 2. Trust Model
 
-The core assumption is that Hermes is a **personal agent** with one trusted operator.
+Hermes Agent is a single-tenant personal agent. Its posture is
+layered, and the layers are not equally load-bearing. Reporters and
+operators should reason about them in the same terms.
 
-### Operator & Session Trust
-- **Single Tenant:** The system protects the operator from LLM actions, not from malicious co-tenants. Multi-user isolation must happen at the OS/host level.
-- **Gateway Security:** Authorized callers (Telegram, Discord, Slack, etc.) receive equal trust. Session keys are used for routing, not as authorization boundaries.
-- **Execution:** Defaults to `terminal.backend: local` (direct host execution). Container isolation (Docker, Modal, Daytona) is opt-in for sandboxing.
+### 2.1 Definitions
 
-### Dangerous Command Approval
-The approval system (`tools/approval.py`) is a core security boundary. Terminal commands, file operations, and other potentially destructive actions are gated behind explicit user confirmation before execution. The approval mode is configurable via `approvals.mode` in `config.yaml`:
-- `"on"` (default) — prompts the user to approve dangerous commands.
-- `"auto"` — auto-approves after a configurable delay.
-- `"off"` — disables the gate entirely (break-glass; see Section 3).
+- **Agent process.** The Python interpreter running Hermes Agent,
+  including any Python modules it has loaded (skills, plugins,
+  hook handlers).
+- **Terminal backend.** A pluggable execution target for the
+  `terminal()` tool. The default runs commands directly on the host.
+  Other backends run commands inside a container, cloud sandbox, or
+  remote host.
+- **Input surface.** Any channel through which content enters the
+  agent's context: operator input, web fetches, email, gateway
+  messages, file reads, MCP server responses, tool results.
+- **Trust envelope.** The set of resources an operator has implicitly
+  granted Hermes Agent access to by running it — typically, whatever
+  the operator's own user account can reach on the host.
+- **Stance.** An explicit statement in Hermes Agent's documentation
+  or code about how a consuming layer (adapter, UI, file writer,
+  shell) should treat agent output — e.g. "the dashboard renders
+  agent output as inert HTML."
 
-### Output Redaction
-`agent/redact.py` strips secret-like patterns (API keys, tokens, credentials) from all display output before it reaches the terminal or gateway platform. This prevents accidental credential leakage in chat logs, tool previews, and response text. Redaction operates on the display layer only — underlying values remain intact for internal agent operations.
+### 2.2 The Boundary: OS-Level Isolation
 
-### Skills vs. MCP Servers
-- **Installed Skills:** High trust. Equivalent to local host code; skills can read environment variables and run arbitrary commands.
-- **MCP Servers:** Lower trust. MCP subprocesses receive a filtered environment (`_build_safe_env()` in `tools/mcp_tool.py`) — only safe baseline variables (`PATH`, `HOME`, `XDG_*`) plus variables explicitly declared in the server's `env` config block are passed through. Host credentials are stripped by default. Additionally, packages invoked via `npx`/`uvx` are checked against the OSV malware database before spawning.
+**The only security boundary against an adversarial LLM is the
+operating system.** Nothing inside the agent process constitutes
+containment — not the approval gate, not output redaction, not any
+pattern scanner, not any tool allowlist. Any in-process component
+that screens LLM output is a heuristic operating on an
+attacker-influenced string, and this policy treats it as such.
 
-### Code Execution Sandbox
-The `execute_code` tool (`tools/code_execution_tool.py`) runs LLM-generated Python scripts in a child process with API keys and tokens stripped from the environment to prevent credential exfiltration. Only environment variables explicitly declared by loaded skills (via `env_passthrough`) or by the user in `config.yaml` (`terminal.env_passthrough`) are passed through. The child accesses Hermes tools via RPC, not direct API calls.
+Hermes Agent supports two OS-level isolation postures. They address
+different threats and an operator should choose deliberately.
 
-### Subagents
-- **No recursive delegation:** The `delegate_task` tool is disabled for child agents.
-- **Depth limit:** `MAX_DEPTH = 2` — parent (depth 0) can spawn a child (depth 1); grandchildren are rejected.
-- **Memory isolation:** Subagents run with `skip_memory=True` and do not have access to the parent's persistent memory provider. The parent receives only the task prompt and final response as an observation.
+#### Terminal-backend isolation
+
+A non-default terminal backend runs LLM-emitted shell commands
+inside a container, remote host, or cloud sandbox. The file tools
+(`read_file`, `write_file`, `patch`) also run through this backend,
+since they are implemented on top of the shell contract — they
+cannot reach paths the backend doesn't expose.
+
+What this confines: anything the agent does by issuing shell or
+file operations. What this does **not** confine: everything the
+agent does in its own Python process. That includes the
+code-execution tool (spawned as a host subprocess), MCP subprocesses
+(spawned from the agent's environment), plugin loading, hook
+dispatch, and skill loading (all imported into the agent
+interpreter).
+
+Terminal-backend isolation is the right posture when the concern is
+LLM-emitted destructive shell or unwanted file-tool writes, and the
+operator is otherwise trusted.
+
+#### Whole-process wrapping
+
+Whole-process wrapping runs the entire agent process tree inside a
+sandbox. Every code path — shell, code-execution, MCP, file tools,
+plugins, hooks, skill loading — is subject to the same filesystem,
+network, process, and (where applicable) inference policy.
+
+Hermes Agent supports this in two ways:
+
+- **Hermes Agent's own Docker image and Compose setup.** Lighter-
+  weight; the agent runs in a standard container with operator-
+  configured mounts and network policy.
+- **[NVIDIA OpenShell](https://github.com/NVIDIA/OpenShell)**.
+  OpenShell provides per-session sandboxes with declarative policy
+  across filesystem, network (L7 egress), process/syscall, and
+  inference-routing layers. Network and inference policies are
+  hot-reloadable. Credentials are injected from a Provider store
+  and never touch the sandbox filesystem.
+
+Under a whole-process wrapper, Hermes Agent's in-process heuristics
+(§2.4) function as accident-prevention layered on top of a real
+boundary. This is the supported posture when the agent ingests
+content from surfaces the operator does not control — the open web,
+inbound email, multi-user channels, untrusted MCP servers — and for
+production or shared deployments.
+
+Operators running the default local backend with untrusted input
+surfaces, or running a terminal-backend sandbox and expecting it to
+contain code paths that don't go through the shell, are operating
+outside the supported security posture.
+
+### 2.3 Credential Scoping
+
+Hermes Agent filters the environment it passes to its lower-trust
+in-process components: shell subprocesses, MCP subprocesses, and
+the code-execution child. Credentials like provider API keys and
+gateway tokens are stripped by default; variables explicitly
+declared by the operator or by a loaded skill are passed through.
+
+This reduces casual exfiltration. It is not containment. Any
+component running inside the agent process (skills, plugins, hook
+handlers) can read whatever the agent itself can read, including
+in-memory credentials. The mitigation against a compromised
+in-process component is operator review before install (§2.4,
+§2.5), not environment scrubbing.
+
+### 2.4 In-Process Heuristics
+
+The following components screen or warn about LLM behavior. They
+are useful. They are not boundaries.
+
+- The **approval gate** detects common destructive shell patterns
+  and prompts the operator before execution. Shell is Turing-
+  complete; a denylist over shell strings is structurally
+  incomplete. The gate catches cooperative-mode mistakes, not
+  adversarial output.
+- **Output redaction** strips secret-like patterns from display.
+  A motivated output producer will defeat it.
+- **Skills Guard** scans installable skill content for injection
+  patterns. It is a review aid; the boundary for third-party skills
+  is operator review before install. Reviewing a skill means
+  reading its Python code and scripts, not just its SKILL.md
+  description — skills execute arbitrary Python at import time.
+
+### 2.5 Plugin Trust Model
+
+Plugins load into the agent process and run with full agent
+privileges: they can read the same credentials, call the same
+tools, register the same hooks, and import the same modules as
+anything shipped in-tree. The boundary for third-party plugins is
+operator review before install — the same rule as skills (§2.4),
+called out separately because plugins are architecturally heavier
+and often ship their own background services, network listeners,
+and dependencies.
+
+A malicious or buggy plugin is not a vulnerability in Hermes Agent
+itself. Bugs in Hermes Agent's plugin-install or plugin-discovery
+path that prevent the operator from seeing what they're installing
+are in scope under §3.1.
+
+### 2.6 External Surfaces
+
+An **external surface** is any channel outside the local agent
+process through which a caller can dispatch agent work, resolve
+approvals, or receive agent output. Each surface has its own
+authorization model, but the rules below apply uniformly.
+
+**Surfaces in Hermes Agent:**
+
+- **Gateway platform adapters.** Messaging integrations in
+  `gateway/platforms/` (Telegram, Discord, Slack, email, SMS, etc.)
+  and analogous adapters shipped as plugins.
+- **Network-exposed HTTP surfaces.** The API server adapter, the
+  dashboard plugin, the kanban plugin's HTTP endpoints, and any
+  other plugin that binds a listening socket.
+- **Editor / IDE adapters.** The ACP adapter (`acp_adapter/`) and
+  equivalent integrations that accept requests from a local client
+  process.
+- **The TUI gateway (`tui_gateway/`).** JSON-RPC backend for the
+  Ink terminal UI, reached over local IPC.
+
+**Uniform rules:**
+
+1. **Authorization is required at every surface that crosses a
+   trust boundary.** For messaging and network HTTP surfaces, the
+   boundary is the network: authorization means an operator-
+   configured caller allowlist. For editor and local-IPC surfaces
+   (ACP, TUI gateway), the boundary is the host's user account:
+   authorization means relying on OS-level access control (file
+   permissions, loopback-only binds) and not exposing the surface
+   beyond the local user without an explicit network auth layer.
+2. **An allowlist is required for every enabled network-exposed
+   adapter.** Adapters must refuse to dispatch agent work, resolve
+   approvals, or relay output until an allowlist is set. Code paths
+   that fail open when no allowlist is configured are code bugs in
+   scope under §3.1.
+3. **Session identifiers are routing handles, not authorization
+   boundaries.** Knowing another caller's session ID does not grant
+   access to their approvals or output; authorization is always
+   re-checked against the allowlist (or OS-level equivalent).
+4. **Within the authorized set, all callers are equally trusted.**
+   Hermes Agent does not model per-caller capabilities inside a
+   single adapter. Operators who need capability separation should
+   run separate agent instances with separate allowlists.
+5. **Binding a local-only surface to a non-loopback interface is a
+   break-glass operator decision (§3.2).** The dashboard and other
+   plugin HTTP servers default to loopback; exposing them via
+   `--host 0.0.0.0` or equivalent makes public-exposure hardening
+   (§4) the operator's responsibility.
 
 ---
 
-## 3. Out of Scope (Non-Vulnerabilities)
+## 3. Scope
 
-The following scenarios are **not** considered security breaches:
-- **Prompt Injection:** Unless it results in a concrete bypass of the approval system, toolset restrictions, or container sandbox.
-- **Public Exposure:** Deploying the gateway to the public internet without external authentication or network protection.
-- **Trusted State Access:** Reports that require pre-existing write access to `~/.hermes/`, `.env`, or `config.yaml` (these are operator-owned files).
-- **Default Behavior:** Host-level command execution when `terminal.backend` is set to `local` — this is the documented default, not a vulnerability.
-- **Configuration Trade-offs:** Intentional break-glass settings such as `approvals.mode: "off"` or `terminal.backend: local` in production.
-- **Tool-level read/access restrictions:** The agent has unrestricted shell access via the `terminal` tool by design. Reports that a specific tool (e.g., `read_file`) can access a resource are not vulnerabilities if the same access is available through `terminal`. Tool-level deny lists only constitute a meaningful security boundary when paired with equivalent restrictions on the terminal side (as with write operations, where `WRITE_DENIED_PATHS` is paired with the dangerous command approval system).
+### 3.1 In Scope
+
+- Escape from a declared OS-level isolation posture (§2.2): an
+  attacker-controlled code path reaching state that the posture
+  claimed to confine.
+- Unauthorized external-surface access: a caller outside the
+  configured authorization set (allowlist, or OS-level equivalent
+  for local-IPC surfaces) dispatching work, receiving output, or
+  resolving approvals (§2.6).
+- Credential exfiltration: leakage of operator credentials or
+  session authorization material to a destination outside the
+  trust envelope, via a mechanism that should have prevented it
+  (environment scrubbing bug, adapter logging, transport error
+  that flushes credentials to an upstream, etc.).
+- Trust-model documentation violations: code behaving contrary to
+  what this policy, Hermes Agent's own documentation, or reasonable
+  operator expectations would predict — including cases where
+  Hermes Agent has documented a stance about how its output should
+  be rendered by a consuming layer (dashboard, gateway adapter,
+  file writer, shell) and a code path breaks that stance.
+
+### 3.2 Out of Scope
+
+"Out of scope" here means "not a security vulnerability under this
+policy." It does not mean "not worth reporting." Improvements to the
+in-process heuristics, hardening ideas, and UX fixes are welcome as
+regular issues or pull requests — the approval gate can always catch
+more patterns, redaction can always get smarter, adapter behavior
+can always be tightened. These items just don't go through the
+private-disclosure channel and don't receive advisories.
+
+- **Bypasses of in-process heuristics (§2.4)** — approval-gate regex
+  bypasses, redaction bypasses, Skills Guard pattern bypasses, and
+  analogous reports against future heuristics. These components are
+  not boundaries; defeating them is not a vulnerability under this
+  policy.
+- **Prompt injection per se.** Getting the LLM to emit unusual
+  output — via injected content, hallucination, training artifacts,
+  or any other cause — is not itself a vulnerability. "I achieved
+  prompt injection" without a chained §3.1 outcome is not an
+  actionable report under this policy.
+- **Consequences of a chosen isolation posture.** Reports that a
+  code path operating within its posture's scope can do what that
+  posture permits are not vulnerabilities. Examples: shell or file
+  tools reaching host state under the local backend; code-execution
+  or MCP subprocesses reaching host state under terminal-backend
+  isolation that only sandboxes shell; reports whose preconditions
+  require pre-existing write access to operator-owned configuration
+  or credential files (those are already inside the trust envelope).
+- **Documented break-glass settings.** Operator-selected trade-offs
+  that explicitly disable protections: `--insecure` and equivalent
+  flags on the dashboard or other components, disabled approvals,
+  local backend in production, development profiles that bypass
+  hermes-home security, and similar. Reports against those
+  configurations are not vulnerabilities — that's the flag's job.
+- **Community-contributed skills and plugins.** Third-party skills
+  (including the community skills repository) and third-party
+  plugins are in the operator's review surface, not Hermes Agent's
+  trust surface (§2.4, §2.5). A skill or plugin doing something
+  malicious is the expected failure mode of one that wasn't
+  reviewed, not a vulnerability in Hermes Agent. Bugs in Hermes
+  Agent's skill-install or plugin-install path that prevent the
+  operator from seeing what they're installing are in scope under
+  §3.1.
+- **Public exposure without external controls.** Exposing the
+  gateway or API to the public internet without authentication,
+  VPN, or firewall.
+- **Tool-level read/write restrictions on a posture where shell is
+  permitted.** If a path is reachable via the terminal tool, reports
+  that other file tools can reach it add nothing.
 
 ---
 
-## 4. Deployment Hardening & Best Practices
+## 4. Deployment Hardening
 
-### Filesystem & Network
-- **Production sandboxing:** Use container backends (`docker`, `modal`, `daytona`) instead of `local` for untrusted workloads.
-- **File permissions:** Run as non-root (the Docker image uses UID 10000); protect credentials with `chmod 600 ~/.hermes/.env` on local installs.
-- **Network exposure:** Do not expose the gateway or API server to the public internet without VPN, Tailscale, or firewall protection. SSRF protection is enabled by default across all gateway platform adapters (Telegram, Discord, Slack, Matrix, Mattermost, etc.) with redirect validation. Note: the local terminal backend does not apply SSRF filtering, as it operates within the trusted operator's environment.
+The single most important hardening decision is matching isolation
+(§2.2) to the trust of the content the agent will ingest. Beyond
+that:
 
-### Skills & Supply Chain
-- **Skill installation:** Review Skills Guard reports (`tools/skills_guard.py`) before installing third-party skills. The audit log at `~/.hermes/skills/.hub/audit.log` tracks every install and removal.
-- **MCP safety:** OSV malware checking runs automatically for `npx`/`uvx` packages before MCP server processes are spawned.
-- **CI/CD:** GitHub Actions are pinned to full commit SHAs. The `supply-chain-audit.yml` workflow blocks PRs containing `.pth` files or suspicious `base64`+`exec` patterns.
-
-### Credential Storage
-- API keys and tokens belong exclusively in `~/.hermes/.env` — never in `config.yaml` or checked into version control.
-- The credential pool system (`agent/credential_pool.py`) handles key rotation and fallback. Credentials are resolved from environment variables, not stored in plaintext databases.
+- Run the agent as a non-root user. The supplied container image
+  does this by default.
+- Keep credentials in the operator credential file with tight
+  permissions, never in the main config, never in version control.
+  Under OpenShell, use the Provider store rather than an on-disk
+  credential file.
+- Do not expose the gateway or API to the public internet without
+  VPN, Tailscale, or firewall protection. Under OpenShell, use the
+  network policy layer to restrict egress.
+- Configure a caller allowlist for every network-exposed adapter
+  you enable (§2.6).
+- Review third-party skills and plugins before install (§2.4,
+  §2.5). For skills, this means reading the Python and scripts,
+  not just SKILL.md. Skills Guard reports and the install audit
+  log are the review surface.
+- Hermes Agent includes supply-chain guards for MCP server
+  launches and for dependency / bundled-package changes in CI; see
+  `CONTRIBUTING.md` for specifics.
 
 ---
 
-## 5. Disclosure Process
+## 5. Disclosure
 
-- **Coordinated Disclosure:** 90-day window or until a fix is released, whichever comes first.
-- **Communication:** All updates occur via the GHSA thread or email correspondence with security@nousresearch.com.
-- **Credits:** Reporters are credited in release notes unless anonymity is requested.
+- **Coordinated disclosure window:** 90 days from report, or until a
+  fix is released, whichever comes first.
+- **Channel:** the GHSA thread or email correspondence with
+  security@nousresearch.com.
+- **Credit:** reporters are credited in release notes unless
+  anonymity is requested.

acp_adapter/session.py

@@ -601,6 +601,7 @@ class SessionManager:
             ),
             "quiet_mode": True,
             "session_id": session_id,
+            "session_db": self._get_db(),
             "model": model or default_model,
         }
 

acp_adapter/tools.py

@@ -769,8 +769,8 @@ def _build_patch_mode_content(patch_text: str) -> List[Any]:
                 old_chunks: list[str] = []
                 new_chunks: list[str] = []
                 for hunk in op.hunks:
-                    old_lines = [line.content for line in hunk.lines if line.prefix in (" ", "-")]
-                    new_lines = [line.content for line in hunk.lines if line.prefix in (" ", "+")]
+                    old_lines = [line.content for line in hunk.lines if line.prefix in {" ", "-"}]
+                    new_lines = [line.content for line in hunk.lines if line.prefix in {" ", "+"}]
                     if old_lines or new_lines:
                         old_chunks.append("\n".join(old_lines))
                         new_chunks.append("\n".join(new_lines))

agent/account_usage.py

@@ -47,7 +47,7 @@ def _title_case_slug(value: Optional[str]) -> Optional[str]:
 
 
 def _parse_dt(value: Any) -> Optional[datetime]:
-    if value in (None, ""):
+    if value in {None, ""}:
         return None
     if isinstance(value, (int, float)):
         return datetime.fromtimestamp(float(value), tz=timezone.utc)

agent/anthropic_adapter.py

@@ -35,6 +35,14 @@ def _get_anthropic_sdk():
     """Return the ``anthropic`` SDK module, importing lazily. None if not installed."""
     global _anthropic_sdk
     if _anthropic_sdk is ...:
+        try:
+            from tools.lazy_deps import ensure as _lazy_ensure
+            _lazy_ensure("provider.anthropic", prompt=False)
+        except ImportError:
+            pass
+        except Exception:
+            # FeatureUnavailable — fall through to ImportError handling below
+            pass
         try:
             import anthropic as _sdk
             _anthropic_sdk = _sdk
@@ -1289,13 +1297,21 @@ def convert_tools_to_anthropic(tools: List[Dict]) -> List[Dict]:
             continue
         if name:
             seen_names.add(name)
-        result.append({
+        anthropic_tool: Dict[str, Any] = {
             "name": name,
             "description": fn.get("description", ""),
             "input_schema": _normalize_tool_input_schema(
                 fn.get("parameters", {"type": "object", "properties": {}})
             ),
-        })
+        }
+        # Forward cache_control marker when present on the OpenAI-format
+        # tool dict (set by ``mark_tools_for_long_lived_cache``). Anthropic's
+        # tools array supports cache_control on the last tool to cache the
+        # entire schema cross-session.
+        cache_control = t.get("cache_control")
+        if isinstance(cache_control, dict):
+            anthropic_tool["cache_control"] = dict(cache_control)
+        result.append(anthropic_tool)
     return result
 
 
@@ -1537,7 +1553,7 @@ def convert_messages_to_anthropic(
             # downgraded to a spurious text block on the last assistant message.
             reasoning_content = m.get("reasoning_content")
             _already_has_thinking = any(
-                isinstance(b, dict) and b.get("type") in ("thinking", "redacted_thinking")
+                isinstance(b, dict) and b.get("type") in {"thinking", "redacted_thinking"}
                 for b in blocks
             )
             if isinstance(reasoning_content, str) and not _already_has_thinking:
@@ -1688,7 +1704,7 @@ def convert_messages_to_anthropic(
                 if isinstance(m["content"], list):
                     m["content"] = [
                         b for b in m["content"]
-                        if not (isinstance(b, dict) and b.get("type") in ("thinking", "redacted_thinking"))
+                        if not (isinstance(b, dict) and b.get("type") in {"thinking", "redacted_thinking"})
                     ]
                 prev_blocks = fixed[-1]["content"]
                 curr_blocks = m["content"]

agent/auxiliary_client.py

@@ -175,7 +175,7 @@ def _normalize_aux_provider(provider: Optional[str]) -> str:
         # 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() or "").strip().lower()
-        if main_prov and main_prov not in ("auto", "main", ""):
+        if main_prov and main_prov not in {"auto", "main", ""}:
             normalized = main_prov
         else:
             return "custom"
@@ -382,7 +382,7 @@ _AI_GATEWAY_HEADERS = {
 # Nous Portal extra_body for product attribution.
 # Callers should pass this as extra_body in chat.completions.create()
 # when the auxiliary client is backed by Nous Portal.
-NOUS_EXTRA_BODY = {"tags": ["product=hermes-agent"]}
+NOUS_EXTRA_BODY = {"tags": ["product=hermes-agent", "client=aux"]}
 
 # Set at resolve time — True if the auxiliary client points to Nous Portal
 auxiliary_is_nous: bool = False
@@ -490,6 +490,29 @@ def _select_pool_entry(provider: str) -> Tuple[bool, Optional[Any]]:
         return True, None
 
 
+def _peek_pool_entry(provider: str) -> Optional[Any]:
+    """Best-effort current/next pool entry without mutating selection order."""
+    try:
+        pool = load_pool(provider)
+    except Exception as exc:
+        logger.debug("Auxiliary client: could not load pool for %s (peek): %s", provider, exc)
+        return None
+    if not pool or not pool.has_credentials():
+        return None
+    try:
+        current_fn = getattr(pool, "current", None)
+        if callable(current_fn):
+            current = current_fn()
+            if current is not None:
+                return current
+        peek_fn = getattr(pool, "peek", None)
+        if callable(peek_fn):
+            return peek_fn()
+    except Exception as exc:
+        logger.debug("Auxiliary client: could not peek pool entry for %s: %s", provider, exc)
+    return None
+
+
 def _pool_runtime_api_key(entry: Any) -> str:
     if entry is None:
         return ""
@@ -555,7 +578,7 @@ def _convert_content_for_responses(content: Any) -> Any:
             if detail:
                 entry["detail"] = detail
             converted.append(entry)
-        elif ptype in ("input_text", "input_image"):
+        elif ptype in {"input_text", "input_image"}:
             # Already in Responses format — pass through
             converted.append(part)
         else:
@@ -683,6 +706,16 @@ class _CodexCompletionsAdapter:
                     close()
                 except Exception:
                     logger.debug("Codex auxiliary: client close during timeout failed", exc_info=True)
+            # The cached auxiliary client wraps this same ``self._client``
+            # (or *is* a ``CodexAuxiliaryClient`` whose ``_real_client`` is
+            # this instance).  After we close the httpx transport above, the
+            # cache must drop that entry — otherwise the next auxiliary call
+            # (compression retry, memory flush, etc.) reuses the dead client
+            # and fails fast with a connection error.  See issue #23432.
+            try:
+                _evict_cached_client_instance(self._client)
+            except Exception:
+                logger.debug("Codex auxiliary: cache eviction on timeout failed", exc_info=True)
 
         def _check_cancelled() -> None:
             if deadline is not None and time.monotonic() >= deadline:
@@ -765,7 +798,7 @@ class _CodexCompletionsAdapter:
                 if item_type == "message":
                     for part in (_item_get(item, "content") or []):
                         ptype = _item_get(part, "type")
-                        if ptype in ("output_text", "text"):
+                        if ptype in {"output_text", "text"}:
                             text_parts.append(_item_get(part, "text", ""))
                 elif item_type == "function_call":
                     tool_calls_raw.append(SimpleNamespace(
@@ -867,6 +900,14 @@ class AsyncCodexAuxiliaryClient:
         self.chat = _AsyncCodexChatShim(async_adapter)
         self.api_key = sync_wrapper.api_key
         self.base_url = sync_wrapper.base_url
+        # Mirror the sync wrapper's _real_client so cache eviction by leaf
+        # OpenAI client (e.g. _close_client_on_timeout in #23482) drops
+        # this async entry too. Without this, sync and async cache entries
+        # diverge on poisoning: the sync entry is evicted but the async
+        # entry keeps reusing the closed transport, failing every
+        # subsequent async aux call with 'Connection error' until the
+        # gateway restarts.
+        self._real_client = sync_wrapper._real_client
 
 
 class _AnthropicCompletionsAdapter:
@@ -1002,6 +1043,9 @@ class AsyncAnthropicAuxiliaryClient:
         self.chat = _AsyncAnthropicChatShim(async_adapter)
         self.api_key = sync_wrapper.api_key
         self.base_url = sync_wrapper.base_url
+        # See AsyncCodexAuxiliaryClient: mirror _real_client so cache
+        # eviction on a poisoned underlying client also drops this entry.
+        self._real_client = sync_wrapper._real_client
 
 
 def _endpoint_speaks_anthropic_messages(base_url: str) -> bool:
@@ -1440,7 +1484,16 @@ def _read_main_model() -> str:
 
     config.yaml model.default is the single source of truth for the active
     model. Environment variables are no longer consulted.
+
+    Runtime override: when an AIAgent is active with a CLI/gateway-provided
+    model that differs from config.yaml, ``set_runtime_main()`` records the
+    override in a process-local global. This is consulted FIRST so tools
+    that gate on "the active main model" (e.g. ``vision_analyze``'s native
+    fast path) see the live runtime, not the persisted config default.
     """
+    override = _RUNTIME_MAIN_MODEL
+    if isinstance(override, str) and override.strip():
+        return override.strip()
     try:
         from hermes_cli.config import load_config
         cfg = load_config()
@@ -1461,7 +1514,13 @@ def _read_main_provider() -> str:
 
     Returns the lowercase provider id (e.g. "alibaba", "openrouter") or ""
     if not configured.
+
+    Runtime override: see ``_read_main_model`` — same mechanism for the
+    provider half of the runtime tuple.
     """
+    override = _RUNTIME_MAIN_PROVIDER
+    if isinstance(override, str) and override.strip():
+        return override.strip().lower()
     try:
         from hermes_cli.config import load_config
         cfg = load_config()
@@ -1475,6 +1534,32 @@ def _read_main_provider() -> str:
     return ""
 
 
+# Process-local override set by AIAgent at session/turn start. Single-threaded
+# per turn — no lock needed. Cleared by ``clear_runtime_main()``.
+_RUNTIME_MAIN_PROVIDER: str = ""
+_RUNTIME_MAIN_MODEL: str = ""
+
+
+def set_runtime_main(provider: str, model: str) -> None:
+    """Record the live runtime provider/model for the current AIAgent.
+
+    Called by ``run_agent.AIAgent._sync_runtime_main_for_aux_routing`` (or
+    equivalent setter) at the top of each turn so that
+    ``_read_main_provider`` / ``_read_main_model`` reflect CLI/gateway
+    overrides instead of the stale config.yaml default.
+    """
+    global _RUNTIME_MAIN_PROVIDER, _RUNTIME_MAIN_MODEL
+    _RUNTIME_MAIN_PROVIDER = (provider or "").strip().lower()
+    _RUNTIME_MAIN_MODEL = (model or "").strip()
+
+
+def clear_runtime_main() -> None:
+    """Clear the runtime override (e.g. on session end)."""
+    global _RUNTIME_MAIN_PROVIDER, _RUNTIME_MAIN_MODEL
+    _RUNTIME_MAIN_PROVIDER = ""
+    _RUNTIME_MAIN_MODEL = ""
+
+
 def _resolve_custom_runtime() -> Tuple[Optional[str], Optional[str], Optional[str]]:
     """Resolve the active custom/main endpoint the same way the main CLI does.
 
@@ -1756,6 +1841,113 @@ def _get_provider_chain() -> List[tuple]:
     ]
 
 
+# ── Auxiliary "recently 402'd" unhealthy-provider cache ────────────────────
+#
+# When an auxiliary provider returns HTTP 402 (Payment Required / credit
+# exhaustion), retrying it on every subsequent aux call is wasteful — the
+# provider stays depleted for hours or days, but the chain re-tries it as
+# the FIRST entry on every compression/title-gen/session-search call,
+# burns ~1 RTT, gets 402 again, then falls back. On a long Discord/LCM
+# session that adds up to dozens of doomed 402s.
+#
+# Solution: when ANY caller observes a payment error against a provider,
+# mark it unhealthy for ``_AUX_UNHEALTHY_TTL_SECONDS``. ``_resolve_auto``
+# Step-2 and ``_try_payment_fallback`` both consult this cache and skip
+# unhealthy entries (logging once per skip-reason so the user sees what
+# happened). Entries auto-expire so a topped-up account recovers without
+# manual intervention.
+#
+# Failure isolation: the cache is in-process only. A second hermes
+# process won't inherit the unhealthy mark — that's intentional, since
+# the user might be running two profiles with different OpenRouter keys.
+
+_AUX_UNHEALTHY_TTL_SECONDS = 600  # 10 minutes
+_aux_unhealthy_until: Dict[str, float] = {}
+_aux_unhealthy_logged_at: Dict[str, float] = {}
+
+# Map provider names that show up in resolved_provider / explicit-config
+# back to the chain labels used by _get_provider_chain(). Keep in sync
+# with the alias map in _try_payment_fallback below.
+_AUX_UNHEALTHY_LABEL_ALIASES = {
+    "openrouter": "openrouter",
+    "nous": "nous",
+    "custom": "local/custom",
+    "local/custom": "local/custom",
+    "openai-codex": "openai-codex",
+    "codex": "openai-codex",
+}
+
+
+def _normalize_chain_label(provider: str) -> str:
+    """Normalize a resolved_provider value to a chain label used by
+    ``_get_provider_chain()``. Falls back to the lowercased input for
+    direct API-key providers (deepseek, alibaba, minimax, etc.) which
+    each report their own provider name from the api-key chain.
+    """
+    if not provider:
+        return ""
+    p = str(provider).strip().lower()
+    return _AUX_UNHEALTHY_LABEL_ALIASES.get(p, p)
+
+
+def _mark_provider_unhealthy(provider: str, ttl: Optional[float] = None) -> None:
+    """Mark ``provider`` as recently-402'd, hidden from chain iteration
+    until the TTL expires. Called from the payment-fallback branches in
+    ``call_llm`` and ``acall_llm`` after a confirmed payment error.
+    """
+    label = _normalize_chain_label(provider)
+    if not label:
+        return
+    expires_at = time.time() + (ttl if ttl is not None else _AUX_UNHEALTHY_TTL_SECONDS)
+    _aux_unhealthy_until[label] = expires_at
+    logger.warning(
+        "Auxiliary: marking %s unhealthy for %ds (payment / credit error). "
+        "Subsequent auxiliary calls will skip it until %s.",
+        label,
+        int(ttl if ttl is not None else _AUX_UNHEALTHY_TTL_SECONDS),
+        time.strftime("%H:%M:%S", time.localtime(expires_at)),
+    )
+
+
+def _is_provider_unhealthy(label: str) -> bool:
+    """True iff ``label`` is in the unhealthy cache and the TTL hasn't expired.
+    Lazily evicts expired entries so the cache stays small.
+    """
+    if not label:
+        return False
+    expires_at = _aux_unhealthy_until.get(label)
+    if expires_at is None:
+        return False
+    if time.time() >= expires_at:
+        _aux_unhealthy_until.pop(label, None)
+        _aux_unhealthy_logged_at.pop(label, None)
+        return False
+    return True
+
+
+def _log_skip_unhealthy(label: str, task: Optional[str] = None) -> None:
+    """Emit a single info-level log per minute when we skip an unhealthy
+    provider. Avoids spamming the log on bursty sessions while still
+    giving the user a trail.
+    """
+    now = time.time()
+    last = _aux_unhealthy_logged_at.get(label, 0.0)
+    if now - last >= 60:
+        _aux_unhealthy_logged_at[label] = now
+        expires_at = _aux_unhealthy_until.get(label, now)
+        logger.info(
+            "Auxiliary %s: skipping %s (recently returned payment error, retry in %ds)",
+            task or "call", label, max(0, int(expires_at - now)),
+        )
+
+
+def _reset_aux_unhealthy_cache() -> None:
+    """Clear the unhealthy cache. Used by tests and by a future explicit
+    user trigger (e.g. ``hermes config aux reset``)."""
+    _aux_unhealthy_until.clear()
+    _aux_unhealthy_logged_at.clear()
+
+
 def _is_payment_error(exc: Exception) -> bool:
     """Detect payment/credit/quota exhaustion errors.
 
@@ -1768,7 +1960,7 @@ def _is_payment_error(exc: Exception) -> bool:
     err_lower = str(exc).lower()
     # OpenRouter and other providers include "credits" or "afford" in 402 bodies,
     # but sometimes wrap them in 429 or other codes.
-    if status in (402, 429, None):
+    if status in {402, 429, None}:
         if any(kw in err_lower for kw in ("credits", "insufficient funds",
                                            "can only afford", "billing",
                                            "payment required")):
@@ -1817,10 +2009,12 @@ def _is_connection_error(exc: Exception) -> bool:
     distinct from API errors (4xx/5xx) which indicate the provider IS
     reachable but returned an error.
     """
-    from openai import APIConnectionError, APITimeoutError
-
-    if isinstance(exc, (APIConnectionError, APITimeoutError)):
-        return True
+    try:
+        from openai import APIConnectionError, APITimeoutError
+        if isinstance(exc, (APIConnectionError, APITimeoutError)):
+            return True
+    except ImportError:
+        pass
     # urllib3 / httpx / httpcore connection errors
     err_type = type(exc).__name__
     if any(kw in err_type for kw in ("Connection", "Timeout", "DNS", "SSL")):
@@ -1830,6 +2024,16 @@ def _is_connection_error(exc: Exception) -> bool:
         "connection refused", "name or service not known",
         "no route to host", "network is unreachable",
         "timed out", "connection reset",
+        # httpcore / httpx streaming premature-close errors.  These surface
+        # when a proxy or provider drops the connection mid-stream and are
+        # transient by nature — the request should be retried or rerouted.
+        # See issue #18458.
+        "incomplete chunked read",
+        "peer closed connection",
+        "response ended prematurely",
+        "unexpected eof",
+        "remoteprotocolerror",
+        "localprotocolerror",
     )):
         return True
     return False
@@ -1908,6 +2112,246 @@ def _evict_cached_clients(provider: str) -> None:
             _client_cache.pop(key, None)
 
 
+def _evict_cached_client_instance(target: Any) -> bool:
+    """Drop the cache entry whose stored client is *target*.
+
+    Used when a specific cached client has been poisoned (closed httpx
+    transport after a timeout, broken streaming session, etc.) so the next
+    auxiliary call rebuilds rather than reusing the dead instance.
+
+    Walks both sync and async wrappers (``CodexAuxiliaryClient``,
+    ``AnthropicAuxiliaryClient``, ``AsyncCodexAuxiliaryClient``, etc.) via
+    their ``_real_client`` attribute so a timeout that closes the underlying
+    ``OpenAI`` (or native provider) client evicts every cached shim that
+    exposed it. Async wrappers must mirror their sync sibling's
+    ``_real_client`` for this to work — otherwise the sync entry is evicted
+    but the async entry survives and keeps reusing the dead transport.
+
+    Returns True when at least one entry was evicted.
+    """
+    if target is None:
+        return False
+    evicted = False
+    with _client_cache_lock:
+        for key in list(_client_cache.keys()):
+            entry = _client_cache.get(key)
+            if entry is None:
+                continue
+            cached = entry[0]
+            if cached is None:
+                continue
+            real = getattr(cached, "_real_client", None)
+            if cached is target or real is target:
+                del _client_cache[key]
+                evicted = True
+    return evicted
+
+
+def _pool_cache_hint(
+    provider: str,
+    *,
+    main_runtime: Optional[Dict[str, Any]] = None,
+) -> str:
+    """Return a stable cache discriminator for pooled providers."""
+    normalized = _normalize_aux_provider(provider)
+    if normalized == "auto":
+        runtime = _normalize_main_runtime(main_runtime)
+        normalized = _normalize_aux_provider(runtime.get("provider") or _read_main_provider())
+    if normalized in {"", "auto", "custom"}:
+        return ""
+    entry = _peek_pool_entry(normalized)
+    if entry is None:
+        return ""
+    entry_id = str(getattr(entry, "id", "") or "").strip()
+    if not entry_id:
+        return ""
+    return f"{normalized}:{entry_id}"
+
+
+def _pool_error_context(exc: Exception) -> Dict[str, Any]:
+    status = getattr(exc, "status_code", None)
+    payload: Dict[str, Any] = {"message": str(exc)}
+    if status is not None:
+        payload["status_code"] = status
+    return payload
+
+
+def _recoverable_pool_provider(resolved_provider: str, client: Any) -> Optional[str]:
+    """Infer which provider pool can recover the current auxiliary client."""
+    normalized = _normalize_aux_provider(resolved_provider)
+    if normalized not in {"", "auto", "custom"}:
+        return normalized
+    base = str(getattr(client, "base_url", "") or "")
+    if base_url_host_matches(base, "chatgpt.com"):
+        return "openai-codex"
+    if base_url_host_matches(base, "openrouter.ai"):
+        return "openrouter"
+    if base_url_host_matches(base, "inference-api.nousresearch.com"):
+        return "nous"
+    if base_url_host_matches(base, "api.anthropic.com"):
+        return "anthropic"
+    if base_url_host_matches(base, "api.githubcopilot.com"):
+        return "copilot"
+    if base_url_host_matches(base, "api.kimi.com"):
+        return "kimi-coding"
+    return None
+
+
+def _recover_provider_pool(provider: str, exc: Exception) -> bool:
+    """Try same-provider credential-pool recovery for auxiliary calls."""
+    normalized = _normalize_aux_provider(provider)
+    try:
+        pool = load_pool(normalized)
+    except Exception as load_exc:
+        logger.debug("Auxiliary client: could not load pool for %s recovery: %s", normalized, load_exc)
+        return False
+    if not pool or not pool.has_credentials():
+        return False
+
+    status_code = getattr(exc, "status_code", None)
+    error_context = _pool_error_context(exc)
+
+    if _is_auth_error(exc):
+        refreshed = pool.try_refresh_current()
+        if refreshed is not None:
+            _evict_cached_clients(normalized)
+            return True
+        next_entry = pool.mark_exhausted_and_rotate(
+            status_code=status_code if status_code is not None else 401,
+            error_context=error_context,
+        )
+        if next_entry is not None:
+            _evict_cached_clients(normalized)
+            return True
+        return False
+
+    if _is_payment_error(exc) or _is_rate_limit_error(exc):
+        fallback_status = 402 if _is_payment_error(exc) else 429
+        next_entry = pool.mark_exhausted_and_rotate(
+            status_code=status_code if status_code is not None else fallback_status,
+            error_context=error_context,
+        )
+        if next_entry is not None:
+            _evict_cached_clients(normalized)
+            return True
+    return False
+
+
+def _retry_same_provider_sync(
+    *,
+    task: Optional[str],
+    resolved_provider: str,
+    resolved_model: Optional[str],
+    resolved_base_url: Optional[str],
+    resolved_api_key: Optional[str],
+    resolved_api_mode: Optional[str],
+    main_runtime: Optional[Dict[str, Any]],
+    final_model: Optional[str],
+    messages: list,
+    temperature: Optional[float],
+    max_tokens: Optional[int],
+    tools: Optional[list],
+    effective_timeout: float,
+    effective_extra_body: dict,
+) -> Any:
+    if task == "vision":
+        _, retry_client, retry_model = resolve_vision_provider_client(
+            provider=resolved_provider,
+            model=final_model,
+            base_url=resolved_base_url,
+            api_key=resolved_api_key,
+            async_mode=False,
+        )
+    else:
+        retry_client, retry_model = _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 None:
+        raise RuntimeError(
+            f"Auxiliary {task or 'call'}: provider {resolved_provider} could not be rebuilt after recovery"
+        )
+
+    retry_base = str(getattr(retry_client, "base_url", "") or "")
+    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=retry_base or resolved_base_url,
+    )
+    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,
+    )
+
+
+async def _retry_same_provider_async(
+    *,
+    task: Optional[str],
+    resolved_provider: str,
+    resolved_model: Optional[str],
+    resolved_base_url: Optional[str],
+    resolved_api_key: Optional[str],
+    resolved_api_mode: Optional[str],
+    final_model: Optional[str],
+    messages: list,
+    temperature: Optional[float],
+    max_tokens: Optional[int],
+    tools: Optional[list],
+    effective_timeout: float,
+    effective_extra_body: dict,
+) -> Any:
+    if task == "vision":
+        _, retry_client, retry_model = resolve_vision_provider_client(
+            provider=resolved_provider,
+            model=final_model,
+            base_url=resolved_base_url,
+            api_key=resolved_api_key,
+            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 None:
+        raise RuntimeError(
+            f"Auxiliary {task or 'call'}: provider {resolved_provider} could not be rebuilt after recovery"
+        )
+
+    retry_base = str(getattr(retry_client, "base_url", "") or "")
+    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=retry_base or resolved_base_url,
+    )
+    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,
+    )
+
+
 def _refresh_provider_credentials(provider: str) -> bool:
     """Refresh short-lived credentials for OAuth-backed auxiliary providers."""
     normalized = _normalize_aux_provider(provider)
@@ -1980,6 +2424,10 @@ def _try_payment_fallback(
     for label, try_fn in _get_provider_chain():
         if label in skip_chain_labels:
             continue
+        if _is_provider_unhealthy(label):
+            _log_skip_unhealthy(label, task)
+            tried.append(f"{label} (unhealthy)")
+            continue
         client, model = try_fn()
         if client is not None:
             logger.info(
@@ -2048,7 +2496,7 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
     main_provider = runtime_provider or _read_main_provider()
     main_model = runtime_model or _read_main_model()
     if (main_provider and main_model
-            and main_provider not in ("auto", "")):
+            and main_provider not in {"auto", ""}):
         resolved_provider = main_provider
         explicit_base_url = None
         explicit_api_key = None
@@ -2056,21 +2504,34 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
             resolved_provider = "custom"
             explicit_base_url = runtime_base_url
             explicit_api_key = runtime_api_key or None
-        client, resolved = resolve_provider_client(
-            resolved_provider,
-            main_model,
-            explicit_base_url=explicit_base_url,
-            explicit_api_key=explicit_api_key,
-            api_mode=runtime_api_mode or None,
-        )
-        if client is not None:
-            logger.info("Auxiliary auto-detect: using main provider %s (%s)",
-                        main_provider, resolved or main_model)
-            return client, resolved or main_model
+        # Skip Step-1 if the main provider was recently 402'd. The unhealthy
+        # cache TTL bounds how long we bypass it, so a topped-up account
+        # recovers automatically. If we tried Step-1 anyway, every aux call
+        # on a depleted main provider would pay one doomed 402 RTT before
+        # falling to Step-2.
+        main_chain_label = _normalize_chain_label(resolved_provider)
+        if main_chain_label and _is_provider_unhealthy(main_chain_label):
+            _log_skip_unhealthy(main_chain_label)
+        else:
+            client, resolved = resolve_provider_client(
+                resolved_provider,
+                main_model,
+                explicit_base_url=explicit_base_url,
+                explicit_api_key=explicit_api_key,
+                api_mode=runtime_api_mode or None,
+            )
+            if client is not None:
+                logger.info("Auxiliary auto-detect: using main provider %s (%s)",
+                            main_provider, resolved or main_model)
+                return client, resolved or main_model
 
     # ── Step 2: aggregator / fallback chain ──────────────────────────────
     tried = []
     for label, try_fn in _get_provider_chain():
+        if _is_provider_unhealthy(label):
+            _log_skip_unhealthy(label)
+            tried.append(f"{label} (unhealthy)")
+            continue
         client, model = try_fn()
         if client is not None:
             if tried:
@@ -2696,7 +3157,7 @@ def resolve_provider_client(
         return (_to_async_client(client, final_model, is_vision=is_vision) if async_mode
                 else (client, final_model))
 
-    elif pconfig.auth_type in ("oauth_device_code", "oauth_external"):
+    elif pconfig.auth_type in {"oauth_device_code", "oauth_external"}:
         # OAuth providers — route through their specific try functions
         if provider == "nous":
             return resolve_provider_client("nous", model, async_mode)
@@ -2805,7 +3266,7 @@ def get_available_vision_backends() -> List[str]:
     available: List[str] = []
     # 1. Active provider — if the user configured a provider, try it first.
     main_provider = _read_main_provider()
-    if main_provider and main_provider not in ("auto", ""):
+    if main_provider and main_provider not in {"auto", ""}:
         if main_provider in _VISION_AUTO_PROVIDER_ORDER:
             if _strict_vision_backend_available(main_provider):
                 available.append(main_provider)
@@ -2851,7 +3312,7 @@ def resolve_vision_provider_client(
 
     if resolved_base_url:
         provider_for_base_override = (
-            requested if requested and requested not in ("", "auto") else "custom"
+            requested if requested and requested not in {"", "auto"} else "custom"
         )
         client, final_model = resolve_provider_client(
             provider_for_base_override,
@@ -2879,7 +3340,7 @@ def resolve_vision_provider_client(
         #   4. Stop
         main_provider = _read_main_provider()
         main_model = _read_main_model()
-        if main_provider and main_provider not in ("auto", ""):
+        if main_provider and main_provider not in {"auto", ""}:
             vision_model = _PROVIDER_VISION_MODELS.get(main_provider, main_model)
             if main_provider == "nous":
                 sync_client, default_model = _resolve_strict_vision_backend(
@@ -3033,7 +3494,8 @@ def _client_cache_key(
 ) -> tuple:
     runtime = _normalize_main_runtime(main_runtime)
     runtime_key = tuple(runtime.get(field, "") for field in _MAIN_RUNTIME_FIELDS) if provider == "auto" else ()
-    return (provider, async_mode, base_url or "", api_key or "", api_mode or "", runtime_key, is_vision)
+    pool_hint = _pool_cache_hint(provider, main_runtime=main_runtime)
+    return (provider, async_mode, base_url or "", api_key or "", api_mode or "", runtime_key, is_vision, pool_hint)
 
 
 def _store_cached_client(cache_key: tuple, client: Any, default_model: Optional[str], *, bound_loop: Any = None) -> None:
@@ -3564,7 +4026,7 @@ def _build_call_kwargs(
     # Provider-specific extra_body
     merged_extra = dict(extra_body or {})
     if provider == "nous" or auxiliary_is_nous:
-        merged_extra.setdefault("tags", []).extend(["product=hermes-agent"])
+        merged_extra.setdefault("tags", []).extend(NOUS_EXTRA_BODY["tags"])
     if merged_extra:
         kwargs["extra_body"] = merged_extra
 
@@ -3684,7 +4146,7 @@ def call_llm(
             # credentials were found, fail fast instead of silently routing
             # through OpenRouter (which causes confusing 404s).
             _explicit = (resolved_provider or "").strip().lower()
-            if _explicit and _explicit not in ("auto", "openrouter", "custom"):
+            if _explicit and _explicit not in {"auto", "openrouter", "custom"}:
                 raise RuntimeError(
                     f"Provider '{_explicit}' is set in config.yaml but no API key "
                     f"was found. Set the {_explicit.upper()}_API_KEY environment "
@@ -3814,46 +4276,63 @@ def call_llm(
 
         # ── Auth refresh retry ───────────────────────────────────────
         if (_is_auth_error(first_err)
-                and resolved_provider not in ("auto", "", None)
+                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,
-                    )
+                return _retry_same_provider_sync(
+                    task=task,
+                    resolved_provider=resolved_provider,
+                    resolved_model=resolved_model,
+                    resolved_base_url=resolved_base_url,
+                    resolved_api_key=resolved_api_key,
+                    resolved_api_mode=resolved_api_mode,
+                    main_runtime=main_runtime,
+                    final_model=final_model,
+                    messages=messages,
+                    temperature=temperature,
+                    max_tokens=max_tokens,
+                    tools=tools,
+                    effective_timeout=effective_timeout,
+                    effective_extra_body=effective_extra_body,
                 )
-                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"])
+
+        # ── Same-provider credential-pool recovery ─────────────────────
+        pool_provider = _recoverable_pool_provider(resolved_provider, client)
+        if pool_provider and (_is_auth_error(first_err) or _is_payment_error(first_err) or _is_rate_limit_error(first_err)):
+            recovery_err = first_err
+            if _is_rate_limit_error(first_err):
+                try:
                     return _validate_llm_response(
-                        retry_client.chat.completions.create(**retry_kwargs), task)
+                        client.chat.completions.create(**kwargs), task)
+                except Exception as retry_err:
+                    if not (_is_auth_error(retry_err) or _is_payment_error(retry_err) or _is_rate_limit_error(retry_err)):
+                        raise
+                    recovery_err = retry_err
+            if _recover_provider_pool(pool_provider, recovery_err):
+                logger.info(
+                    "Auxiliary %s: recovered %s via credential-pool rotation after %s",
+                    task or "call", pool_provider, type(recovery_err).__name__,
+                )
+                return _retry_same_provider_sync(
+                    task=task,
+                    resolved_provider=resolved_provider,
+                    resolved_model=resolved_model,
+                    resolved_base_url=resolved_base_url,
+                    resolved_api_key=resolved_api_key,
+                    resolved_api_mode=resolved_api_mode,
+                    main_runtime=main_runtime,
+                    final_model=final_model,
+                    messages=messages,
+                    temperature=temperature,
+                    max_tokens=max_tokens,
+                    tools=tools,
+                    effective_timeout=effective_timeout,
+                    effective_extra_body=effective_extra_body,
+                )
 
         # ── Payment / credit exhaustion fallback ──────────────────────
         # When the resolved provider returns 402 or a credit-related error,
@@ -3880,10 +4359,17 @@ def call_llm(
         # Only try alternative providers when the user didn't explicitly
         # configure this task's provider.  Explicit provider = hard constraint;
         # auto (the default) = best-effort fallback chain.  (#7559)
-        is_auto = resolved_provider in ("auto", "", None)
+        is_auto = resolved_provider in {"auto", "", None}
         if should_fallback and is_auto:
             if _is_payment_error(first_err):
                 reason = "payment error"
+                # Resolve the actual provider label (resolved_provider may be
+                # "auto"; the client's base_url tells us which backend got the
+                # 402). Mark THAT label unhealthy so subsequent aux calls
+                # skip it instead of paying another doomed RTT.
+                _mark_provider_unhealthy(
+                    _recoverable_pool_provider(resolved_provider, client) or resolved_provider
+                )
             elif _is_rate_limit_error(first_err):
                 reason = "rate limit"
             else:
@@ -3901,6 +4387,17 @@ def call_llm(
                     base_url=str(getattr(fb_client, "base_url", "") or ""))
                 return _validate_llm_response(
                     fb_client.chat.completions.create(**fb_kwargs), task)
+        # Connection/timeout errors leave the cached client poisoned (closed
+        # httpx transport, half-read stream, dead async loop).  Drop it from
+        # the cache regardless of whether we found a fallback above so the
+        # next auxiliary call rebuilds a fresh client instead of reusing the
+        # dead one.  See issue #23432.
+        if _is_connection_error(first_err):
+            try:
+                _evict_cached_client_instance(client)
+            except Exception:
+                logger.debug("Auxiliary: cache eviction after connection error failed",
+                             exc_info=True)
         raise
 
 
@@ -4018,7 +4515,7 @@ async def async_call_llm(
         )
         if client is None:
             _explicit = (resolved_provider or "").strip().lower()
-            if _explicit and _explicit not in ("auto", "openrouter", "custom"):
+            if _explicit and _explicit not in {"auto", "openrouter", "custom"}:
                 raise RuntimeError(
                     f"Provider '{_explicit}' is set in config.yaml but no API key "
                     f"was found. Set the {_explicit.upper()}_API_KEY environment "
@@ -4129,45 +4626,61 @@ async def async_call_llm(
 
         # ── Auth refresh retry (mirrors sync call_llm) ───────────────
         if (_is_auth_error(first_err)
-                and resolved_provider not in ("auto", "", None)
+                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 await _retry_same_provider_async(
+                    task=task,
+                    resolved_provider=resolved_provider,
+                    resolved_model=resolved_model,
+                    resolved_base_url=resolved_base_url,
+                    resolved_api_key=resolved_api_key,
+                    resolved_api_mode=resolved_api_mode,
+                    final_model=final_model,
+                    messages=messages,
+                    temperature=temperature,
+                    max_tokens=max_tokens,
+                    tools=tools,
+                    effective_timeout=effective_timeout,
+                    effective_extra_body=effective_extra_body,
+                )
+
+        # ── Same-provider credential-pool recovery (mirrors sync) ─────
+        pool_provider = _recoverable_pool_provider(resolved_provider, client)
+        if pool_provider and (_is_auth_error(first_err) or _is_payment_error(first_err) or _is_rate_limit_error(first_err)):
+            recovery_err = first_err
+            if _is_rate_limit_error(first_err):
+                try:
                     return _validate_llm_response(
-                        await retry_client.chat.completions.create(**retry_kwargs), task)
+                        await client.chat.completions.create(**kwargs), task)
+                except Exception as retry_err:
+                    if not (_is_auth_error(retry_err) or _is_payment_error(retry_err) or _is_rate_limit_error(retry_err)):
+                        raise
+                    recovery_err = retry_err
+            if _recover_provider_pool(pool_provider, recovery_err):
+                logger.info(
+                    "Auxiliary %s (async): recovered %s via credential-pool rotation after %s",
+                    task or "call", pool_provider, type(recovery_err).__name__,
+                )
+                return await _retry_same_provider_async(
+                    task=task,
+                    resolved_provider=resolved_provider,
+                    resolved_model=resolved_model,
+                    resolved_base_url=resolved_base_url,
+                    resolved_api_key=resolved_api_key,
+                    resolved_api_mode=resolved_api_mode,
+                    final_model=final_model,
+                    messages=messages,
+                    temperature=temperature,
+                    max_tokens=max_tokens,
+                    tools=tools,
+                    effective_timeout=effective_timeout,
+                    effective_extra_body=effective_extra_body,
+                )
 
         # ── Payment / connection / rate-limit fallback (mirrors sync call_llm) ──
         should_fallback = (
@@ -4175,10 +4688,13 @@ async def async_call_llm(
             or _is_connection_error(first_err)
             or _is_rate_limit_error(first_err)
         )
-        is_auto = resolved_provider in ("auto", "", None)
+        is_auto = resolved_provider in {"auto", "", None}
         if should_fallback and is_auto:
             if _is_payment_error(first_err):
                 reason = "payment error"
+                _mark_provider_unhealthy(
+                    _recoverable_pool_provider(resolved_provider, client) or resolved_provider
+                )
             elif _is_rate_limit_error(first_err):
                 reason = "rate limit"
             else:
@@ -4202,4 +4718,12 @@ async def async_call_llm(
                     fb_kwargs["model"] = async_fb_model
                 return _validate_llm_response(
                     await async_fb.chat.completions.create(**fb_kwargs), task)
+        # Mirror the sync path: drop poisoned clients on connection/timeout
+        # so the next aux call rebuilds.  See issue #23432.
+        if _is_connection_error(first_err):
+            try:
+                _evict_cached_client_instance(client)
+            except Exception:
+                logger.debug("Auxiliary (async): cache eviction after connection error failed",
+                             exc_info=True)
         raise

agent/codex_responses_adapter.py

@@ -410,10 +410,29 @@ def _chat_messages_to_responses_input(messages: List[Dict[str, Any]]) -> List[Di
                     call_id = raw_tool_call_id.strip()
             if not isinstance(call_id, str) or not call_id.strip():
                 continue
+
+            # Multimodal tool result: convert OpenAI-style content list into
+            # Responses ``function_call_output.output`` array. The Responses
+            # API accepts ``output`` as either a string or an array of
+            # ``input_text``/``input_image`` items. See
+            # https://developers.openai.com/api/reference/python/resources/responses/.
+            tool_content = msg.get("content")
+            output_value: Any
+            if isinstance(tool_content, list):
+                converted = _chat_content_to_responses_parts(
+                    tool_content, role="user",
+                )
+                if converted:
+                    output_value = converted
+                else:
+                    output_value = ""
+            else:
+                output_value = str(tool_content or "")
+
             items.append({
                 "type": "function_call_output",
                 "call_id": call_id,
-                "output": str(msg.get("content", "") or ""),
+                "output": output_value,
             })
 
     return items
@@ -466,6 +485,38 @@ def _preflight_codex_input_items(raw_items: Any) -> List[Dict[str, Any]]:
             output = item.get("output", "")
             if output is None:
                 output = ""
+            # Output may be a string OR an array of structured content
+            # items (input_text / input_image) for multimodal tool results.
+            # Both shapes are accepted by the Responses API. We preserve
+            # the array form when present.
+            if isinstance(output, list):
+                # Validate each item is a recognised content shape; drop
+                # anything else to avoid 4xx from the API.
+                cleaned: List[Dict[str, Any]] = []
+                for part in output:
+                    if not isinstance(part, dict):
+                        continue
+                    ptype = part.get("type")
+                    if ptype == "input_text":
+                        text = part.get("text")
+                        if isinstance(text, str) and text:
+                            cleaned.append({"type": "input_text", "text": text})
+                    elif ptype == "input_image":
+                        url = part.get("image_url")
+                        if isinstance(url, str) and url:
+                            entry: Dict[str, Any] = {"type": "input_image", "image_url": url}
+                            detail = part.get("detail")
+                            if isinstance(detail, str) and detail.strip():
+                                entry["detail"] = detail.strip()
+                            cleaned.append(entry)
+                normalized.append(
+                    {
+                        "type": "function_call_output",
+                        "call_id": call_id.strip(),
+                        "output": cleaned if cleaned else "",
+                    }
+                )
+                continue
             if not isinstance(output, str):
                 output = str(output)
 

agent/context_compressor.py

@@ -23,7 +23,7 @@ import re
 import time
 from typing import Any, Dict, List, Optional
 
-from agent.auxiliary_client import call_llm
+from agent.auxiliary_client import call_llm, _is_connection_error
 from agent.context_engine import ContextEngine
 from agent.model_metadata import (
     MINIMUM_CONTEXT_LENGTH,
@@ -167,7 +167,7 @@ def _strip_image_parts_from_parts(parts: Any) -> Any:
             out.append(part)
             continue
         ptype = part.get("type")
-        if ptype in ("image", "image_url", "input_image"):
+        if ptype in {"image", "image_url", "input_image"}:
             had_image = True
             out.append({"type": "text", "text": "[screenshot removed to save context]"})
         else:
@@ -274,8 +274,8 @@ def _summarize_tool_result(tool_name: str, tool_args: str, tool_content: str) ->
         mode = args.get("mode", "replace")
         return f"[patch] {mode} in {path} ({content_len:,} chars result)"
 
-    if tool_name in ("browser_navigate", "browser_click", "browser_snapshot",
-                     "browser_type", "browser_scroll", "browser_vision"):
+    if tool_name in {"browser_navigate", "browser_click", "browser_snapshot",
+                     "browser_type", "browser_scroll", "browser_vision"}:
         url = args.get("url", "")
         ref = args.get("ref", "")
         detail = f" {url}" if url else (f" ref={ref}" if ref else "")
@@ -304,7 +304,7 @@ def _summarize_tool_result(tool_name: str, tool_args: str, tool_content: str) ->
             code_preview += "..."
         return f"[execute_code] `{code_preview}` ({line_count} lines output)"
 
-    if tool_name in ("skill_view", "skills_list", "skill_manage"):
+    if tool_name in {"skill_view", "skills_list", "skill_manage"}:
         name = args.get("name", "?")
         return f"[{tool_name}] name={name} ({content_len:,} chars)"
 
@@ -979,13 +979,13 @@ The user has requested that this compaction PRIORITISE preserving all informatio
             _status = getattr(e, "status_code", None) or getattr(getattr(e, "response", None), "status_code", None)
             _err_str = str(e).lower()
             _is_model_not_found = (
-                _status in (404, 503)
+                _status in {404, 503}
                 or "model_not_found" in _err_str
                 or "does not exist" in _err_str
                 or "no available channel" in _err_str
             )
             _is_timeout = (
-                _status in (408, 429, 502, 504)
+                _status in {408, 429, 502, 504}
                 or "timeout" in _err_str
             )
             # Non-JSON / malformed-body responses from misconfigured providers
@@ -1000,6 +1000,14 @@ The user has requested that this compaction PRIORITISE preserving all informatio
                 isinstance(e, json.JSONDecodeError)
                 or "expecting value" in _err_str
             )
+            # httpcore / httpx streaming premature-close errors surface as
+            # ConnectionError subclasses or plain Exception with characteristic
+            # substrings ("incomplete chunked read", "peer closed connection",
+            # "response ended prematurely", "unexpected eof").  These are
+            # transient network events; treat them like a timeout so we fall
+            # back to the main model instead of entering a 60-second cooldown.
+            # See issue #18458.
+            _is_streaming_closed = _is_connection_error(e)
             if _is_json_decode and not _is_model_not_found and not _is_timeout:
                 logger.error(
                     "Context compression failed: auxiliary LLM returned a "
@@ -1012,7 +1020,7 @@ The user has requested that this compaction PRIORITISE preserving all informatio
                     e,
                 )
             if (
-                (_is_model_not_found or _is_timeout or _is_json_decode)
+                (_is_model_not_found or _is_timeout or _is_json_decode or _is_streaming_closed)
                 and self.summary_model
                 and self.summary_model != self.model
                 and not getattr(self, "_summary_model_fallen_back", False)
@@ -1021,6 +1029,8 @@ The user has requested that this compaction PRIORITISE preserving all informatio
                     _reason = "returned invalid JSON"
                 elif _is_model_not_found:
                     _reason = "unavailable"
+                elif _is_streaming_closed:
+                    _reason = "closed stream prematurely"
                 else:
                     _reason = "timed out"
                 self._fallback_to_main_for_compression(e, _reason)
@@ -1043,10 +1053,10 @@ The user has requested that this compaction PRIORITISE preserving all informatio
                 self._fallback_to_main_for_compression(e, "failed")
                 return self._generate_summary(turns_to_summarize, focus_topic=focus_topic)
 
-            # Transient errors (timeout, rate limit, network, JSON decode) —
-            # shorter cooldown for JSON decode since the body shape can flip
-            # back to valid quickly when an upstream proxy recovers.
-            _transient_cooldown = 30 if _is_json_decode else 60
+            # Transient errors (timeout, rate limit, network, JSON decode,
+            # streaming premature-close) — shorter cooldown for JSON decode and
+            # streaming-closed since those conditions can self-resolve quickly.
+            _transient_cooldown = 30 if (_is_json_decode or _is_streaming_closed) else 60
             self._summary_failure_cooldown_until = time.monotonic() + _transient_cooldown
             err_text = str(e).strip() or e.__class__.__name__
             if len(err_text) > 220:
@@ -1306,8 +1316,7 @@ The user has requested that this compaction PRIORITISE preserving all informatio
 
         # Ensure we protect at least min_tail messages
         fallback_cut = n - min_tail
-        if cut_idx > fallback_cut:
-            cut_idx = fallback_cut
+        cut_idx = min(cut_idx, fallback_cut)
 
         # If the token budget would protect everything (small conversations),
         # force a cut after the head so compression can still remove middle turns.
@@ -1470,7 +1479,7 @@ The user has requested that this compaction PRIORITISE preserving all informatio
         first_tail_role = messages[compress_end].get("role", "user") if compress_end < n_messages else "user"
         # Pick a role that avoids consecutive same-role with both neighbors.
         # Priority: avoid colliding with head (already committed), then tail.
-        if last_head_role in ("assistant", "tool"):
+        if last_head_role in {"assistant", "tool"}:
             summary_role = "user"
         else:
             summary_role = "assistant"

agent/credential_pool.py

@@ -149,7 +149,7 @@ class PooledCredential:
         }
         result: Dict[str, Any] = {}
         for field_def in fields(self):
-            if field_def.name in ("provider", "extra"):
+            if field_def.name in {"provider", "extra"}:
                 continue
             value = getattr(self, field_def.name)
             if value is not None or field_def.name in _ALWAYS_EMIT:

agent/curator.py

@@ -72,6 +72,7 @@ def _default_state() -> Dict[str, Any]:
         "last_run_at": None,
         "last_run_duration_seconds": None,
         "last_run_summary": None,
+        "last_run_summary_shown_at": None,
         "last_report_path": None,
         "paused": False,
         "run_count": 0,
@@ -876,6 +877,96 @@ def _reconcile_classification(
     return {"consolidated": consolidated, "pruned": pruned}
 
 
+def _build_rename_summary(
+    *,
+    before_names: Set[str],
+    after_report: List[Dict[str, Any]],
+    tool_calls: List[Dict[str, Any]],
+    model_final: str,
+) -> str:
+    """Format the user-visible rename map for a curator run.
+
+    Renders the "where did my skills go?" lines that get appended to the
+    `final_summary` string fed to gateway/CLI receivers. Empty string when
+    nothing was archived this run — most ticks are no-op and shouldn't add
+    extra log noise.
+
+    Format::
+
+        archived 4 skill(s):
+          • pdf-extraction → document-tools
+          • docx-extraction → document-tools
+          • flaky-thing — pruned (stale)
+          • old-utility → spreadsheet-ops
+        full report: hermes curator status
+        keep an umbrella stable: hermes curator pin document-tools
+
+    Cap is 10 entries so a 50-skill consolidation doesn't blow up
+    agent.log; the full list is always in REPORT.md. The pin hint only
+    appears when at least one consolidation produced an umbrella worth
+    pinning (pruned-only runs skip it).
+    """
+    after_by_name = {r.get("name"): r for r in after_report if isinstance(r, dict)}
+    after_names = set(after_by_name.keys())
+    removed = sorted(before_names - after_names)
+    added = sorted(after_names - before_names)
+    if not removed:
+        return ""
+
+    heuristic = _classify_removed_skills(
+        removed=removed,
+        added=added,
+        after_names=after_names,
+        tool_calls=tool_calls,
+    )
+    model_block = _parse_structured_summary(model_final)
+    destinations = set(after_names) | set(added)
+    absorbed_declarations = _extract_absorbed_into_declarations(tool_calls)
+    classification = _reconcile_classification(
+        removed=removed,
+        heuristic=heuristic,
+        model_block=model_block,
+        destinations=destinations,
+        absorbed_declarations=absorbed_declarations,
+    )
+    consolidated = classification["consolidated"]
+    pruned = classification["pruned"]
+
+    SHOW = 10
+    lines: List[str] = []
+    total = len(consolidated) + len(pruned)
+    lines.append(f"archived {total} skill(s):")
+    shown = 0
+    for entry in consolidated:
+        if shown >= SHOW:
+            break
+        name = entry.get("name", "?")
+        into = entry.get("into", "?")
+        lines.append(f"  • {name} → {into}")
+        shown += 1
+    for entry in pruned:
+        if shown >= SHOW:
+            break
+        name = entry.get("name", "?") if isinstance(entry, dict) else str(entry)
+        lines.append(f"  • {name} — pruned (stale)")
+        shown += 1
+    if total > SHOW:
+        lines.append(f"  … and {total - SHOW} more")
+    lines.append("full report: hermes curator status")
+    # Pin hint — only surface it when there's actually a destination skill
+    # worth pinning. The umbrella skills that absorbed content are the natural
+    # candidates: pinning one tells future curator runs to leave it alone.
+    # Pruned-only runs don't get this hint (nothing surviving to pin).
+    if consolidated:
+        umbrellas = sorted({e.get("into") for e in consolidated if e.get("into")})
+        if umbrellas:
+            example = umbrellas[0]
+            lines.append(
+                f"keep an umbrella stable: hermes curator pin {example}"
+            )
+    return "\n".join(lines)
+
+
 def _write_run_report(
     *,
     started_at: datetime,
@@ -1398,6 +1489,22 @@ def run_curator_review(
                 "error": str(e),
             }
 
+        # Append the rename map (`old-name → umbrella`) to the user-visible
+        # summary so people don't have to dig into REPORT.md to find out where
+        # their skills went. Best-effort: classification is pure but never
+        # block the run on a formatting issue.
+        try:
+            rename_lines = _build_rename_summary(
+                before_names=before_names,
+                after_report=skill_usage.agent_created_report(),
+                tool_calls=llm_meta.get("tool_calls", []) or [],
+                model_final=llm_meta.get("final", "") or "",
+            )
+            if rename_lines:
+                final_summary = f"{final_summary}\n{rename_lines}"
+        except Exception as e:
+            logger.debug("Curator rename summary build failed: %s", e, exc_info=True)
+
         elapsed = (datetime.now(timezone.utc) - start).total_seconds()
         state2 = load_state()
         state2["last_run_duration_seconds"] = elapsed

agent/error_classifier.py

@@ -83,7 +83,7 @@ class ClassifiedError:
 
     @property
     def is_auth(self) -> bool:
-        return self.reason in (FailoverReason.auth, FailoverReason.auth_permanent)
+        return self.reason in {FailoverReason.auth, FailoverReason.auth_permanent}
 
 
 
@@ -254,6 +254,20 @@ _THINKING_SIG_PATTERNS = [
     "signature",  # Combined with "thinking" check
 ]
 
+# Message-string patterns that indicate a provider-side timeout even when
+# the exception type is generic (e.g. RuntimeError from a local shim that
+# wraps a subprocess timeout).  Checked before the type-based transport
+# heuristics so custom-provider "timed out" errors don't fall through to
+# the unknown bucket and get misreported as empty responses.
+_TIMEOUT_MESSAGE_PATTERNS = [
+    "timed out",
+    "turn timed out",
+    "request timed out",
+    "deadline exceeded",
+    "operation timed out",
+    "upstream timed out",
+]
+
 # Transport error type names
 _TRANSPORT_ERROR_TYPES = frozenset({
     "ReadTimeout", "ConnectTimeout", "PoolTimeout",
@@ -674,10 +688,10 @@ def _classify_by_status(
             result_fn=result_fn,
         )
 
-    if status_code in (500, 502):
+    if status_code in {500, 502}:
         return result_fn(FailoverReason.server_error, retryable=True)
 
-    if status_code in (503, 529):
+    if status_code in {503, 529}:
         return result_fn(FailoverReason.overloaded, retryable=True)
 
     # Other 4xx — non-retryable
@@ -796,7 +810,7 @@ def _classify_400(
         # Responses API (and some providers) use flat body: {"message": "..."}
         if not err_body_msg:
             err_body_msg = str(body.get("message") or "").strip().lower()
-    is_generic = len(err_body_msg) < 30 or err_body_msg in ("error", "")
+    is_generic = len(err_body_msg) < 30 or err_body_msg in {"error", ""}
     # Absolute token/message-count thresholds are only a proxy for smaller
     # context windows.  Large-context sessions can have many messages while
     # still being far below their actual token budget.
@@ -827,14 +841,14 @@ def _classify_by_error_code(
     """Classify by structured error codes from the response body."""
     code_lower = error_code.lower()
 
-    if code_lower in ("resource_exhausted", "throttled", "rate_limit_exceeded"):
+    if code_lower in {"resource_exhausted", "throttled", "rate_limit_exceeded"}:
         return result_fn(
             FailoverReason.rate_limit,
             retryable=True,
             should_rotate_credential=True,
         )
 
-    if code_lower in ("insufficient_quota", "billing_not_active", "payment_required"):
+    if code_lower in {"insufficient_quota", "billing_not_active", "payment_required"}:
         return result_fn(
             FailoverReason.billing,
             retryable=False,
@@ -842,14 +856,14 @@ def _classify_by_error_code(
             should_fallback=True,
         )
 
-    if code_lower in ("model_not_found", "model_not_available", "invalid_model"):
+    if code_lower in {"model_not_found", "model_not_available", "invalid_model"}:
         return result_fn(
             FailoverReason.model_not_found,
             retryable=False,
             should_fallback=True,
         )
 
-    if code_lower in ("context_length_exceeded", "max_tokens_exceeded"):
+    if code_lower in {"context_length_exceeded", "max_tokens_exceeded"}:
         return result_fn(
             FailoverReason.context_overflow,
             retryable=True,
@@ -963,6 +977,14 @@ def _classify_by_message(
             should_fallback=True,
         )
 
+    # Timeout message patterns — generic exception types (e.g. RuntimeError)
+    # raised by local shims or custom providers that internally wrap a
+    # subprocess/HTTP timeout.  Classified as transport timeout so the retry
+    # loop rebuilds the client instead of treating the turn as an empty
+    # model response.
+    if any(p in error_msg for p in _TIMEOUT_MESSAGE_PATTERNS):
+        return result_fn(FailoverReason.timeout, retryable=True)
+
     return None
 
 

agent/gemini_cloudcode_adapter.py

@@ -77,7 +77,7 @@ def _coerce_content_to_text(content: Any) -> str:
                 if p.get("type") == "text" and isinstance(p.get("text"), str):
                     pieces.append(p["text"])
                 # Multimodal (image_url, etc.) — stub for now; log and skip
-                elif p.get("type") in ("image_url", "input_audio"):
+                elif p.get("type") in {"image_url", "input_audio"}:
                     logger.debug("Dropping multimodal part (not yet supported): %s", p.get("type"))
         return "\n".join(pieces)
     return str(content)

agent/gemini_native_adapter.py

@@ -945,6 +945,12 @@ class AsyncGeminiNativeClient:
         self.api_key = sync_client.api_key
         self.base_url = sync_client.base_url
         self.chat = _AsyncGeminiChatNamespace(self)
+        # Expose the underlying sync client as _real_client so the auxiliary
+        # cache's eviction-by-leaf-client helper (#23482) can find and drop
+        # this async entry when the sync GeminiNativeClient is poisoned.
+        # GeminiNativeClient is itself the leaf (no OpenAI client beneath
+        # it), so we point at the sync_client directly.
+        self._real_client = sync_client
 
     async def _create_chat_completion(self, **kwargs: Any) -> Any:
         stream = bool(kwargs.get("stream"))

agent/i18n.py

@@ -39,20 +39,45 @@ from typing import Any
 
 logger = logging.getLogger(__name__)
 
-SUPPORTED_LANGUAGES: tuple[str, ...] = ("en", "zh", "ja", "de", "es", "fr", "tr", "uk")
+SUPPORTED_LANGUAGES: tuple[str, ...] = (
+    "en", "zh", "zh-hant", "ja", "de", "es", "fr", "tr", "uk",
+    "af", "ko", "it", "ga", "pt", "ru", "hu",
+)
 DEFAULT_LANGUAGE = "en"
 
 # Accept a few natural aliases so users who type "chinese" / "zh-CN" / "jp"
 # get the right catalog instead of silently falling back to English.
 _LANGUAGE_ALIASES: dict[str, str] = {
     "english": "en", "en-us": "en", "en-gb": "en",
-    "chinese": "zh", "mandarin": "zh", "zh-cn": "zh", "zh-tw": "zh", "zh-hans": "zh", "zh-hant": "zh",
+    # Simplified Chinese — explicit codes route here; bare "chinese" / "mandarin"
+    # also default to Simplified since that's the larger user base.
+    "chinese": "zh", "mandarin": "zh", "zh-cn": "zh", "zh-hans": "zh", "zh-sg": "zh",
+    # Traditional Chinese — distinct catalog.  Cover Taiwan / Hong Kong / Macau
+    # locale tags plus the common "traditional" alias.
+    "traditional-chinese": "zh-hant", "traditional_chinese": "zh-hant",
+    "zh-tw": "zh-hant", "zh-hk": "zh-hant", "zh-mo": "zh-hant",
     "japanese": "ja", "jp": "ja", "ja-jp": "ja",
-    "german": "de", "deutsch": "de", "de-de": "de",
-    "spanish": "es", "español": "es", "espanol": "es", "es-es": "es", "es-mx": "es",
+    "german": "de", "deutsch": "de", "de-de": "de", "de-at": "de", "de-ch": "de",
+    "spanish": "es", "español": "es", "espanol": "es", "es-es": "es", "es-mx": "es", "es-ar": "es",
     "french": "fr", "français": "fr", "france": "fr", "fr-fr": "fr", "fr-be": "fr", "fr-ca": "fr", "fr-ch": "fr",
     "ukrainian": "uk", "ukrainisch": "uk", "українська": "uk", "uk-ua": "uk", "ua": "uk",
     "turkish": "tr", "türkçe": "tr", "tr-tr": "tr",
+    # Afrikaans — South African Dutch-derived language; "af-ZA" is the common BCP-47 tag.
+    "afrikaans": "af", "af-za": "af",
+    # Korean
+    "korean": "ko", "한국어": "ko", "ko-kr": "ko",
+    # Italian
+    "italian": "it", "italiano": "it", "it-it": "it", "it-ch": "it",
+    # Irish (Gaeilge) — ga is the BCP-47 code
+    "irish": "ga", "gaeilge": "ga", "ga-ie": "ga",
+    # Portuguese — bare "portuguese" routes to European Portuguese; pt-br
+    # is in the same family but rendered identically here (no separate br catalog).
+    "portuguese": "pt", "português": "pt", "portugues": "pt",
+    "pt-pt": "pt", "pt-br": "pt", "brazilian": "pt", "brasileiro": "pt",
+    # Russian
+    "russian": "ru", "русский": "ru", "ru-ru": "ru",
+    # Hungarian
+    "hungarian": "hu", "magyar": "hu", "hu-hu": "hu",
 }
 
 _catalog_cache: dict[str, dict[str, str]] = {}

agent/image_routing.py

@@ -76,7 +76,7 @@ def _explicit_aux_vision_override(cfg: Optional[Dict[str, Any]]) -> bool:
     base_url = str(vision.get("base_url") or "").strip()
 
     # "auto" / "" / blank = not explicit
-    if provider in ("", "auto") and not model and not base_url:
+    if provider in {"", "auto"} and not model and not base_url:
         return False
     return True
 
@@ -163,7 +163,7 @@ def _sniff_mime_from_bytes(raw: bytes) -> Optional[str]:
     if raw.startswith(b"\xff\xd8\xff"):
         return "image/jpeg"
     # GIF87a / GIF89a
-    if raw[:6] in (b"GIF87a", b"GIF89a"):
+    if raw[:6] in {b"GIF87a", b"GIF89a"}:
         return "image/gif"
     # WEBP: "RIFF" .... "WEBP"
     if len(raw) >= 12 and raw[:4] == b"RIFF" and raw[8:12] == b"WEBP":
@@ -172,9 +172,9 @@ def _sniff_mime_from_bytes(raw: bytes) -> Optional[str]:
     if raw.startswith(b"BM"):
         return "image/bmp"
     # HEIC/HEIF: ftypheic / ftypheix / ftypmif1 / ftypmsf1 etc.
-    if len(raw) >= 12 and raw[4:8] == b"ftyp" and raw[8:12] in (
+    if len(raw) >= 12 and raw[4:8] == b"ftyp" and raw[8:12] in {
         b"heic", b"heix", b"hevc", b"hevx", b"mif1", b"msf1", b"heim", b"heis",
-    ):
+    }:
         return "image/heic"
     return None
 

agent/markdown_tables.py

@@ -0,0 +1,309 @@
+"""CJK/wide-character-aware re-alignment of model-emitted markdown tables.
+
+Models pad markdown tables assuming each character occupies one terminal
+cell. CJK glyphs and most emoji render as two cells, so the model's
+spacing collapses into drift the moment a table reaches a real terminal —
+header pipes line up, every body row drifts right by N cells per CJK
+char.
+
+This module rebuilds row padding using ``wcwidth.wcswidth`` (display
+columns), preserving the table's pipes and dashes so it still reads as a
+plain-text table in ``strip`` / unrendered display modes. Standard Rich
+markdown rendering already aligns CJK correctly inside a wide enough
+panel; this helper is for the paths that print the model's text more or
+less verbatim.
+
+The helper is deliberately conservative:
+
+* Only contiguous ``| ... |`` blocks with a divider line are rewritten.
+* Anything that does not look like a table is passed through unchanged.
+* Single-line / mid-stream fragments are left alone — callers buffer
+  table rows and flush them once the block is complete.
+
+There is a small, intentional caveat: ``wcwidth`` returns ``-1`` for some
+emoji-with-variation-selector sequences (e.g. ``⚠️``); we clamp those to
+0 so they do not corrupt the column width math. The 1-cell drift on
+those specific glyphs is preferable to silently widening every table
+that contains one.
+"""
+
+from __future__ import annotations
+
+import re
+from typing import List
+
+from wcwidth import wcswidth
+
+__all__ = [
+    "is_table_divider",
+    "looks_like_table_row",
+    "realign_markdown_tables",
+    "split_table_row",
+]
+
+
+_DIVIDER_CELL_RE = re.compile(r"^\s*:?-{3,}:?\s*$")
+_MIN_COL_WIDTH = 3  # matches the divider's minimum dash run.
+
+
+def _disp_width(s: str) -> int:
+    """``wcswidth`` clamped to a non-negative integer.
+
+    ``wcswidth`` returns ``-1`` when it encounters a control char or an
+    unknown sequence; treat those as zero-width rather than letting a
+    negative number flow into ``max`` and break the column-width math.
+    """
+
+    w = wcswidth(s)
+    return w if w > 0 else 0
+
+
+def _pad_to_width(s: str, target: int) -> str:
+    return s + " " * max(0, target - _disp_width(s))
+
+
+def split_table_row(row: str) -> List[str]:
+    """Split ``| a | b | c |`` into ``["a", "b", "c"]`` with trims."""
+
+    s = row.strip()
+    if s.startswith("|"):
+        s = s[1:]
+    if s.endswith("|"):
+        s = s[:-1]
+    return [c.strip() for c in s.split("|")]
+
+
+def is_table_divider(row: str) -> bool:
+    """True when ``row`` is a markdown table separator line."""
+
+    cells = split_table_row(row)
+    return len(cells) > 1 and all(_DIVIDER_CELL_RE.match(c) for c in cells)
+
+
+def looks_like_table_row(row: str) -> bool:
+    """True when ``row`` could plausibly be a markdown table row.
+
+    Used by streaming callers to decide whether to buffer an in-flight
+    line. We are intentionally permissive here — the realigner itself
+    only rewrites blocks that are accompanied by a divider, so a false
+    positive here at most delays the print of one line.
+    """
+
+    if "|" not in row:
+        return False
+    stripped = row.strip()
+    if not stripped:
+        return False
+    # A leading pipe is the strongest signal; without it we still allow
+    # rows with at least two pipes so models that omit the leading pipe
+    # don't slip past us.
+    if stripped.startswith("|"):
+        return True
+    return stripped.count("|") >= 2
+
+
+def _render_block(rows: List[List[str]], available_width: int | None = None) -> List[str]:
+    """Render ``rows`` (header + body, divider implied) at uniform widths.
+
+    If ``available_width`` is given and the rebuilt horizontal table
+    would exceed it, fall back to a vertical key-value rendering so
+    rows do not soft-wrap mid-cell — terminal soft-wrap destroys
+    column alignment visually even when the underlying bytes are
+    perfectly padded, which is exactly the "tables look broken"
+    user report this code path is meant to address.
+    """
+
+    ncols = max(len(r) for r in rows)
+    rows = [r + [""] * (ncols - len(r)) for r in rows]
+
+    widths = [
+        max(_MIN_COL_WIDTH, *(_disp_width(r[c]) for r in rows))
+        for c in range(ncols)
+    ]
+
+    # Total horizontal width for the rendered row:
+    #   `| ` + cell + ` ` for each column, plus the final closing `|`.
+    horizontal_width = sum(widths) + 3 * ncols + 1
+
+    if available_width is not None and horizontal_width > max(available_width, 20):
+        return _render_vertical(rows, ncols, available_width)
+
+    def _row(cells: List[str]) -> str:
+        return (
+            "| "
+            + " | ".join(_pad_to_width(c, widths[k]) for k, c in enumerate(cells))
+            + " |"
+        )
+
+    out = [_row(rows[0])]
+    out.append("|" + "|".join("-" * (w + 2) for w in widths) + "|")
+    for r in rows[1:]:
+        out.append(_row(r))
+    return out
+
+
+def _wrap_to_width(text: str, width: int) -> List[str]:
+    """Soft-wrap ``text`` at word boundaries to fit ``width`` display cells.
+
+    Falls back to hard-breaking the longest word if a single token is
+    wider than ``width``.  Empty input yields a single empty string so
+    the caller's row count stays predictable.
+    """
+
+    if width <= 0 or not text:
+        return [text]
+
+    words = text.split()
+    if not words:
+        return [""]
+
+    lines: List[str] = []
+    current = ""
+    current_w = 0
+
+    def _hard_break(word: str, w: int) -> List[str]:
+        out: List[str] = []
+        buf = ""
+        bw = 0
+        for ch in word:
+            cw = _disp_width(ch) or 1
+            if bw + cw > w and buf:
+                out.append(buf)
+                buf = ch
+                bw = cw
+            else:
+                buf += ch
+                bw += cw
+        if buf:
+            out.append(buf)
+        return out
+
+    for word in words:
+        ww = _disp_width(word)
+        if not current:
+            if ww <= width:
+                current = word
+                current_w = ww
+            else:
+                pieces = _hard_break(word, width)
+                lines.extend(pieces[:-1])
+                current = pieces[-1] if pieces else ""
+                current_w = _disp_width(current)
+            continue
+        if current_w + 1 + ww <= width:
+            current += " " + word
+            current_w += 1 + ww
+        else:
+            lines.append(current)
+            if ww <= width:
+                current = word
+                current_w = ww
+            else:
+                pieces = _hard_break(word, width)
+                lines.extend(pieces[:-1])
+                current = pieces[-1] if pieces else ""
+                current_w = _disp_width(current)
+    if current:
+        lines.append(current)
+    return lines or [""]
+
+
+def _render_vertical(
+    rows: List[List[str]], ncols: int, available_width: int
+) -> List[str]:
+    """Render a too-wide table as vertical ``Header: value`` rows.
+
+    Mirrors Claude Code's narrow-terminal fallback in
+    ``MarkdownTable.tsx``: each body row becomes a small block of
+    ``Header: cell-value`` lines (continuation lines indented two
+    spaces) separated by a thin ``─`` divider between rows.  Keeps
+    every line narrower than ``available_width`` so the terminal does
+    not soft-wrap mid-cell.
+    """
+
+    if not rows:
+        return []
+
+    headers = rows[0] + [""] * (ncols - len(rows[0]))
+    body = rows[1:]
+
+    labels = [h or f"Column {i + 1}" for i, h in enumerate(headers)]
+
+    sep_width = max(20, min(40, available_width - 2)) if available_width else 30
+    separator = "─" * sep_width
+    indent = "  "
+    indent_w = _disp_width(indent)
+
+    out: List[str] = []
+    for ri, row in enumerate(body):
+        if ri > 0:
+            out.append(separator)
+        for ci in range(ncols):
+            label = labels[ci]
+            value = row[ci] if ci < len(row) else ""
+            label_w = _disp_width(label)
+            first_budget = max(10, available_width - label_w - 2)
+            cont_budget = max(10, available_width - indent_w)
+            if not value:
+                out.append(f"{label}:")
+                continue
+            wrapped = _wrap_to_width(value, first_budget)
+            out.append(f"{label}: {wrapped[0]}")
+            if len(wrapped) > 1:
+                # Re-flow continuation text at the wider continuation
+                # budget — words split across the narrower first-line
+                # budget should re-pack greedily for the rest.
+                cont_text = " ".join(wrapped[1:])
+                for cl in _wrap_to_width(cont_text, cont_budget):
+                    if cl.strip():
+                        out.append(f"{indent}{cl}")
+    return out
+
+
+def realign_markdown_tables(text: str, available_width: int | None = None) -> str:
+    """Rewrite every ``| ... |`` + divider block with wcwidth-aware padding.
+
+    Lines that are not part of a recognised table are returned verbatim,
+    so this is safe to apply to arbitrary assistant prose.
+
+    If ``available_width`` is given (terminal cells available for the
+    rendered table), tables wider than that are rendered as vertical
+    key-value pairs instead of a horizontal pipe-bordered grid.  This
+    avoids the terminal soft-wrapping mid-cell, which destroys column
+    alignment visually even when the bytes are perfectly padded.
+    """
+
+    if "|" not in text:
+        return text
+
+    lines = text.split("\n")
+    out: List[str] = []
+    i = 0
+    n = len(lines)
+
+    while i < n:
+        line = lines[i]
+        # A table starts with a header row whose next line is a divider.
+        if (
+            "|" in line
+            and i + 1 < n
+            and is_table_divider(lines[i + 1])
+        ):
+            header = split_table_row(line)
+            body: List[List[str]] = []
+            j = i + 2
+            while j < n and "|" in lines[j] and lines[j].strip():
+                if is_table_divider(lines[j]):
+                    j += 1
+                    continue
+                body.append(split_table_row(lines[j]))
+                j += 1
+
+            if any(c for c in header) or body:
+                out.extend(_render_block([header] + body, available_width))
+                i = j
+                continue
+        out.append(line)
+        i += 1
+
+    return "\n".join(out)

agent/memory_manager.py

@@ -470,11 +470,11 @@ class MemoryManager:
 
         accepted = [
             p for p in params
-            if p.kind in (
+            if p.kind in {
                 inspect.Parameter.POSITIONAL_ONLY,
                 inspect.Parameter.POSITIONAL_OR_KEYWORD,
                 inspect.Parameter.KEYWORD_ONLY,
-            )
+            }
         ]
         if len(accepted) >= 4:
             return "positional"

agent/model_metadata.py

@@ -157,6 +157,13 @@ DEFAULT_CONTEXT_LENGTHS = {
     "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)
+    # gpt-5.3-codex-spark is Codex-OAuth-only (ChatGPT Pro entitlement) and
+    # uses a smaller 128k window than other gpt-5.x slugs. Listed here as
+    # a defensive override so the longest-substring fallback doesn't match
+    # the generic "gpt-5" entry below (400k) and report the wrong limit if
+    # Spark's context ever needs to be resolved through this path. Real
+    # usage flows through _CODEX_OAUTH_CONTEXT_FALLBACK at line ~1113.
+    "gpt-5.3-codex-spark": 128000,
     "gpt-5.1-chat": 128000,           # Chat variant has 128k context
     "gpt-5": 400000,                  # GPT-5.x base, mini, codex variants (400k)
     "gpt-4.1": 1047576,
@@ -210,8 +217,10 @@ DEFAULT_CONTEXT_LENGTHS = {
     "grok": 131072,             # catch-all (grok-beta, unknown grok-*)
     # Kimi
     "kimi": 262144,
-    # Tencent — Hy3 Preview (Hunyuan) with 256K context window
-    "hy3-preview": 256000,
+    # Tencent — Hy3 Preview (Hunyuan) with 256K context window.
+    # OpenRouter live metadata reports 262144 (256 × 1024); align the
+    # static fallback so cache and offline both agree (issue #22268).
+    "hy3-preview": 262144,
     # Nemotron — NVIDIA's open-weights series (128K context across all sizes)
     "nemotron": 131072,
     # Arcee
@@ -235,6 +244,44 @@ DEFAULT_CONTEXT_LENGTHS = {
     "zai-org/GLM-5": 202752,
 }
 
+# xAI Grok models that ACCEPT the `reasoning.effort` parameter on
+# api.x.ai. Verified live against /v1/responses 2026-05-10:
+#
+#   ACCEPTS effort:  grok-3-mini, grok-3-mini-fast, grok-4.20-multi-agent-0309,
+#                    grok-4.3
+#   REJECTS effort:  grok-3, grok-4, grok-4-0709, grok-4-fast-(non-)reasoning,
+#                    grok-4-1-fast-(non-)reasoning, grok-4.20-0309-(non-)reasoning,
+#                    grok-code-fast-1
+#
+# REJECTS-side models still reason natively — they just don't expose an
+# effort dial — so callers should send no `reasoning` key at all rather
+# than a default `medium` (which 400s with "Model X does not support
+# parameter reasoningEffort").
+_GROK_EFFORT_CAPABLE_PREFIXES = (
+    "grok-3-mini",
+    "grok-4.20-multi-agent",
+    "grok-4.3",
+)
+
+
+def grok_supports_reasoning_effort(model: str) -> bool:
+    """Return True when an xAI Grok model accepts ``reasoning.effort``.
+
+    Allowlist by substring (matches both bare ``grok-3-mini`` and
+    aggregator-prefixed ``x-ai/grok-3-mini``). Conservative by design:
+    if a future Grok model isn't listed, we send no effort dial rather
+    than 400.
+    """
+    name = (model or "").strip().lower()
+    if not name:
+        return False
+    # Strip common aggregator prefixes (x-ai/, openrouter/x-ai/, xai/, ...)
+    for sep in ("/",):
+        if sep in name:
+            name = name.rsplit(sep, 1)[-1]
+    return any(name.startswith(prefix) for prefix in _GROK_EFFORT_CAPABLE_PREFIXES)
+
+
 _CONTEXT_LENGTH_KEYS = (
     "context_length",
     "context_window",
@@ -524,7 +571,7 @@ def _extract_pricing(payload: Dict[str, Any]) -> Dict[str, Any]:
         pricing: Dict[str, Any] = {}
         for target, aliases in alias_map.items():
             for alias in aliases:
-                if alias in normalized and normalized[alias] not in (None, ""):
+                if alias in normalized and normalized[alias] not in {None, ""}:
                     pricing[target] = normalized[alias]
                     break
         if pricing:
@@ -959,6 +1006,79 @@ def query_ollama_num_ctx(model: str, base_url: str, api_key: str = "") -> Option
     return None
 
 
+def _query_ollama_api_show(model: str, base_url: str, api_key: str = "") -> Optional[int]:
+    """Query an Ollama server's native ``/api/show`` for context length.
+
+    Provider-agnostic: works against ANY Ollama-compatible server regardless
+    of hostname — local Ollama, Ollama Cloud (``ollama.com``), custom Ollama
+    hosting behind a reverse proxy, etc.  For non-Ollama servers the POST
+    returns 404/405 quickly; the function handles errors gracefully.
+
+    For hosted servers the GGUF ``model_info.*.context_length`` is the
+    authoritative source: the user can't set their own ``num_ctx``, and the
+    OpenAI-compat ``/v1/models`` endpoint correctly omits ``context_length``
+    per the OpenAI schema.
+
+    Resolution order for hosted Ollama:
+      1. ``model_info.*.context_length`` — GGUF training max (authoritative)
+      2. ``parameters`` → ``num_ctx`` — server-side Modelfile override
+    The order is flipped vs ``query_ollama_num_ctx()`` because local users
+    control ``num_ctx`` themselves; hosted users can't.
+    """
+    import httpx
+
+    server_url = base_url.rstrip("/")
+    if server_url.endswith("/v1"):
+        server_url = server_url[:-3]
+
+    headers = _auth_headers(api_key)
+
+    try:
+        with httpx.Client(timeout=5.0, headers=headers) as client:
+            resp = client.post(f"{server_url}/api/show", json={"name": model})
+            if resp.status_code != 200:
+                return None
+            data = resp.json()
+
+            # Hosted Ollama: GGUF model_info is the real max — prefer it over
+            # num_ctx which the Cloud operator may have capped arbitrarily.
+            model_info = data.get("model_info", {})
+            for key, value in model_info.items():
+                if "context_length" in key and isinstance(value, (int, float)):
+                    ctx = int(value)
+                    if ctx >= 1024:
+                        return ctx
+
+            # Fall back to num_ctx from Modelfile parameters (rare on Cloud)
+            params = data.get("parameters", "")
+            if "num_ctx" in params:
+                for line in params.split("\n"):
+                    if "num_ctx" in line:
+                        parts = line.strip().split()
+                        if len(parts) >= 2:
+                            try:
+                                ctx = int(parts[-1])
+                                if ctx >= 1024:
+                                    return ctx
+                            except ValueError:
+                                pass
+    except Exception:
+        pass
+    return None
+
+
+def _model_name_suggests_kimi(model: str) -> bool:
+    """Return True if the model name looks like a Kimi-family model.
+
+    Catches ``kimi-k2.6``, ``kimi-k2.5``, ``kimi-k2-thinking``,
+    ``moonshotai/Kimi-K2.6``, and similar variants.  Used as a guard
+    against stale OpenRouter metadata that underreports these models
+    as 32K context when they actually support 262K+.
+    """
+    lower = model.lower()
+    return lower.startswith("kimi") or "moonshot" in lower
+
+
 def _query_local_context_length(model: str, base_url: str, api_key: str = "") -> Optional[int]:
     """Query a local server for the model's context length."""
     import httpx
@@ -1106,6 +1226,12 @@ _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,
+    # Spark runs on specialised low-latency hardware and exposes a smaller
+    # 128k window than other Codex OAuth slugs. Listed explicitly so the
+    # longest-key-first fallback resolves it correctly — substring match
+    # on "gpt-5.3-codex" otherwise wins and reports 272k. Availability is
+    # gated by ChatGPT Pro entitlement on the Codex backend.
+    "gpt-5.3-codex-spark": 128_000,
     "gpt-5.2-codex": 272_000,
     "gpt-5.4-mini": 272_000,
     "gpt-5.5": 272_000,
@@ -1212,16 +1338,35 @@ def _resolve_nous_context_length(model: str) -> Optional[int]:
     with version normalization (dot↔dash).
     """
     metadata = fetch_model_metadata()  # OpenRouter cache
+
+    def _safe_ctx(or_id: str, entry: dict) -> Optional[int]:
+        """Return context length, but reject stale 32k values for Kimi models.
+
+        Apply the same guard used for the generic OpenRouter path (step 6 in 
+        resolve_context_length) so the Nous portal path does not short-circuit it.
+        """
+        ctx = entry.get("context_length")
+        if ctx is None:
+            return None
+        if ctx <= 32768 and _model_name_suggests_kimi(or_id):
+            logger.info(
+                "Rejecting OpenRouter metadata context=%s for %r "
+                "(Kimi-family underreport, Nous path); falling through to hardcoded defaults",
+                ctx, or_id,
+            )
+            return None
+        return ctx
+
     # Exact match first
     if model in metadata:
-        return metadata[model].get("context_length")
+        return _safe_ctx(model, metadata[model])
 
     normalized = _normalize_model_version(model).lower()
 
     for or_id, entry in metadata.items():
         bare = or_id.split("/", 1)[1] if "/" in or_id else or_id
         if bare.lower() == model.lower() or _normalize_model_version(bare).lower() == normalized:
-            return entry.get("context_length")
+            return _safe_ctx(or_id, entry)
 
     # Partial prefix match for cases like gemini-3-flash → gemini-3-flash-preview
     # Require match to be at a word boundary (followed by -, :, or end of string)
@@ -1232,7 +1377,7 @@ def _resolve_nous_context_length(model: str) -> Optional[int]:
             if candidate.startswith(query) and (
                 len(candidate) == len(query) or candidate[len(query)] in "-:."
             ):
-                return entry.get("context_length")
+                return _safe_ctx(or_id, entry)
 
     return None
 
@@ -1254,12 +1399,17 @@ def get_model_context_length(
     2. Active endpoint metadata (/models for explicit custom endpoints)
     3. Local server query (for local endpoints)
     4. Anthropic /v1/models API (API-key users only, not OAuth)
-    5. OpenRouter live API metadata
-    6. Nous suffix-match via OpenRouter cache
-    7. models.dev registry lookup (provider-aware)
-    8. Thin hardcoded defaults (broad family patterns)
-    9. Default fallback (256K)
-    """
+    5. Provider-aware lookups (before generic OpenRouter cache):
+       a. Copilot live /models API
+       b. Nous suffix-match via OpenRouter cache
+       c. Codex OAuth /models probe
+       d. GMI /models endpoint
+       e. Ollama native /api/show probe (any base_url, provider-agnostic)
+       f. models.dev registry lookup (with :cloud/-cloud suffix fallback)
+    6. OpenRouter live API metadata (Kimi-family 32k guard)
+    7. Hardcoded defaults (broad family patterns, longest-key-first)
+    8. Local server query (last resort)
+    9. Default fallback (256K)"""
     # 0. Explicit config override — user knows best
     if config_context_length is not None and isinstance(config_context_length, int) and config_context_length > 0:
         return config_context_length
@@ -1306,6 +1456,14 @@ def get_model_context_length(
                     model, base_url, f"{cached:,}",
                 )
                 _invalidate_cached_context_length(model, base_url)
+            # Invalidate stale 32k cache entries for Kimi-family models.
+            elif cached <= 32768 and _model_name_suggests_kimi(model):
+                logger.info(
+                    "Dropping stale Kimi cache entry %s@%s -> %s (OpenRouter underreport); "
+                    "re-resolving via hardcoded defaults",
+                    model, base_url, f"{cached:,}",
+                )
+                _invalidate_cached_context_length(model, base_url)
             else:
                 return cached
 
@@ -1339,6 +1497,13 @@ def get_model_context_length(
         if context_length is not None:
             return context_length
         if not _is_known_provider_base_url(base_url):
+            # 2b. Ollama native /api/show — any URL might be an Ollama server
+            # (local, cloud, or custom hosting).  Non-Ollama servers return
+            # 404/405 quickly.  Fall through on failure.
+            ctx = _query_ollama_api_show(model, base_url, api_key=api_key)
+            if ctx is not None:
+                save_context_length(model, base_url, ctx)
+                return ctx
             # 3. Try querying local server directly
             if is_local_endpoint(base_url):
                 local_ctx = _query_local_context_length(model, base_url, api_key=api_key)
@@ -1370,7 +1535,7 @@ def get_model_context_length(
     # (e.g. claude-opus-4.6 is 1M on Anthropic but 128K on GitHub Copilot).
     # If provider is generic (openrouter/custom/empty), try to infer from URL.
     effective_provider = provider
-    if not effective_provider or effective_provider in ("openrouter", "custom"):
+    if not effective_provider or effective_provider in {"openrouter", "custom"}:
         if base_url:
             inferred = _infer_provider_from_url(base_url)
             if inferred:
@@ -1380,7 +1545,7 @@ def get_model_context_length(
     # 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"):
+    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)
@@ -1408,16 +1573,45 @@ def get_model_context_length(
         ctx = _resolve_endpoint_context_length(model, base_url, api_key=api_key)
         if ctx is not None:
             return ctx
+    # 5e. Ollama native /api/show probe — runs for ANY provider with a
+    # base_url, not just ollama-cloud.  Ollama-compatible servers expose
+    # this endpoint regardless of hostname (local Ollama, Ollama Cloud,
+    # custom Ollama hosting).  The OpenAI-compat /v1/models endpoint
+    # correctly omits context_length per the OpenAI schema, but /api/show
+    # returns the authoritative GGUF model_info.context_length.
+    # For non-Ollama servers (OpenAI, Anthropic, etc.), the POST returns
+    # 404/405 quickly.  Results are cached, so the hit is per-model+URL,
+    # once per hour.
+    if base_url:
+        ctx = _query_ollama_api_show(model, base_url, api_key=api_key)
+        if ctx is not None:
+            save_context_length(model, base_url, ctx)
+            return ctx
     if effective_provider:
         from agent.models_dev import lookup_models_dev_context
         ctx = lookup_models_dev_context(effective_provider, model)
         if ctx:
             return ctx
 
-    # 6. OpenRouter live API metadata (provider-unaware fallback)
-    metadata = fetch_model_metadata()
-    if model in metadata:
-        return metadata[model].get("context_length", DEFAULT_FALLBACK_CONTEXT)
+    # 6. OpenRouter live API metadata — provider-unaware fallback.
+    # Only consulted when the provider is unknown (no effective_provider),
+    # because OpenRouter data is community-maintained and can be incorrect
+    # for models that belong to known providers with curated defaults.
+    if not effective_provider:
+        metadata = fetch_model_metadata()
+        if model in metadata:
+            or_ctx = metadata[model].get("context_length", DEFAULT_FALLBACK_CONTEXT)
+            # Guard against stale OpenRouter metadata for Kimi-family models.
+            if or_ctx == 32768 and _model_name_suggests_kimi(model):
+                logger.info(
+                    "Rejecting OpenRouter metadata context=%s for %r "
+                    "(Kimi-family underreport); falling through to hardcoded defaults",
+                    or_ctx, model,
+                )
+            else:
+                return or_ctx
+
+    # 7. (reserved)
 
     # 8. Hardcoded defaults (fuzzy match — longest key first for specificity)
     # Only check `default_model in model` (is the key a substring of the input).
@@ -1480,7 +1674,7 @@ def _count_image_tokens(msg: Dict[str, Any], cost_per_image: int) -> int:
             if not isinstance(part, dict):
                 continue
             ptype = part.get("type")
-            if ptype in ("image", "image_url", "input_image"):
+            if ptype in {"image", "image_url", "input_image"}:
                 count += 1
     stashed = msg.get("_anthropic_content_blocks") if isinstance(msg, dict) else None
     if isinstance(stashed, list):
@@ -1492,7 +1686,7 @@ def _count_image_tokens(msg: Dict[str, Any], cost_per_image: int) -> int:
         inner = content.get("content")
         if isinstance(inner, list):
             for part in inner:
-                if isinstance(part, dict) and part.get("type") in ("image", "image_url"):
+                if isinstance(part, dict) and part.get("type") in {"image", "image_url"}:
                     count += 1
     return count * cost_per_image
 
@@ -1514,7 +1708,7 @@ def _estimate_message_chars(msg: Dict[str, Any]) -> int:
                 cleaned = []
                 for part in v:
                     if isinstance(part, dict):
-                        if part.get("type") in ("image", "image_url", "input_image"):
+                        if part.get("type") in {"image", "image_url", "input_image"}:
                             cleaned.append({"type": part.get("type"), "image": "[stripped]"})
                         else:
                             cleaned.append(part)

agent/models_dev.py

@@ -145,7 +145,9 @@ PROVIDER_TO_MODELS_DEV: Dict[str, str] = {
     "openai": "openai",
     "openai-codex": "openai",
     "zai": "zai",
+    "kimi": "kimi-for-coding",
     "kimi-coding": "kimi-for-coding",
+    "moonshot": "kimi-for-coding",
     "stepfun": "stepfun",
     "kimi-coding-cn": "kimi-for-coding",
     "minimax": "minimax",
@@ -197,6 +199,32 @@ def _load_disk_cache() -> Dict[str, Any]:
     return {}
 
 
+def _disk_cache_age_seconds() -> Optional[float]:
+    """Return age (in seconds) of the disk cache file, or None if missing.
+
+    Used by ``fetch_models_dev`` to short-circuit the network probe when
+    a recent on-disk cache exists. Errors (missing file, permission
+    denied, weird filesystem) all return None — callers fall through
+    to the network fetch path.
+    """
+    try:
+        cache_path = _get_cache_path()
+        if not cache_path.exists():
+            return None
+        mtime = cache_path.stat().st_mtime
+        age = time.time() - mtime
+        # Negative age means the file's mtime is in the future (clock skew
+        # or system clock reset). Treat as "unknown freshness" → fall
+        # through to network so we don't serve potentially-bad data
+        # forever.
+        if age < 0:
+            return None
+        return age
+    except Exception as e:
+        logger.debug("Failed to stat models.dev disk cache: %s", e)
+        return None
+
+
 def _save_disk_cache(data: Dict[str, Any]) -> None:
     """Save models.dev data to disk cache atomically."""
     try:
@@ -207,13 +235,29 @@ def _save_disk_cache(data: Dict[str, Any]) -> None:
 
 
 def fetch_models_dev(force_refresh: bool = False) -> Dict[str, Any]:
-    """Fetch models.dev registry. In-memory cache (1hr) + disk fallback.
+    """Fetch models.dev registry. Cache hierarchy: in-mem → disk → network.
 
     Returns the full registry dict keyed by provider ID, or empty dict on failure.
+
+    Cache hierarchy (when ``force_refresh=False``):
+      1. In-memory cache, populated and < TTL old → return immediately.
+      2. **Disk cache file < TTL old by mtime → load, populate in-mem, return.**
+         No network call. Saves ~500 ms per cold-start agent construction;
+         ``models.dev`` only changes when providers add new models, so a
+         1 hour staleness window is acceptable (same TTL as in-mem cache).
+      3. Network fetch → on success, save to disk + in-mem and return.
+      4. Network fails → fall back to ANY available disk cache (even stale)
+         with a short 5 min in-mem grace period before retrying network.
+
+    When ``force_refresh=True`` (used by ``hermes config refresh``, the
+    \"refresh model catalog\" code path), stages 1 and 2 are skipped. The
+    function always hits the network and only falls back to disk if the
+    network call fails.
     """
     global _models_dev_cache, _models_dev_cache_time
 
-    # Check in-memory cache
+    # Stage 1: fresh in-memory cache wins. This is the hot path on
+    # long-lived processes — no I/O, no system calls.
     if (
         not force_refresh
         and _models_dev_cache
@@ -221,7 +265,27 @@ def fetch_models_dev(force_refresh: bool = False) -> Dict[str, Any]:
     ):
         return _models_dev_cache
 
-    # Try network fetch
+    # Stage 2: fresh-by-mtime disk cache short-circuits the network call.
+    # Only kicks in on cold-start processes (in-mem cache is empty or
+    # expired) and only when the user hasn't asked for a forced refresh.
+    # Skipped if the disk cache file is missing, unreadable, or older
+    # than _MODELS_DEV_CACHE_TTL.
+    if not force_refresh:
+        disk_age = _disk_cache_age_seconds()
+        if disk_age is not None and disk_age < _MODELS_DEV_CACHE_TTL:
+            disk_data = _load_disk_cache()
+            if disk_data:
+                _models_dev_cache = disk_data
+                # Anchor in-mem TTL to the disk file's age so we don't
+                # extend an already-aging cache by another full hour.
+                _models_dev_cache_time = time.time() - disk_age
+                logger.debug(
+                    "Loaded models.dev from fresh disk cache "
+                    "(%d providers, age=%.0fs)", len(disk_data), disk_age,
+                )
+                return _models_dev_cache
+
+    # Stage 3: network fetch.
     try:
         response = requests.get(MODELS_DEV_URL, timeout=15)
         response.raise_for_status()
@@ -239,8 +303,9 @@ def fetch_models_dev(force_refresh: bool = False) -> Dict[str, Any]:
     except Exception as e:
         logger.debug("Failed to fetch models.dev: %s", e)
 
-    # Fall back to disk cache — use a short TTL (5 min) so we retry
-    # the network fetch soon instead of serving stale data for a full hour.
+    # Stage 4: network failed — fall back to whatever disk cache exists,
+    # even if it's stale. Give it a short 5 min in-mem TTL so we retry
+    # the network soon instead of serving stale data for a full hour.
     if not _models_dev_cache:
         _models_dev_cache = _load_disk_cache()
         if _models_dev_cache:
@@ -284,6 +349,28 @@ def lookup_models_dev_context(provider: str, model: str) -> Optional[int]:
             if ctx:
                 return ctx
 
+    # Suffix-aware fallback: some providers (e.g. ollama-cloud) store
+    # model IDs with :cloud / -cloud suffixes in models.dev while the
+    # live API returns bare names.  Without this, kimi-k2.6 misses the
+    # kimi-k2.6:cloud entry and falls through to stale OpenRouter metadata
+    # reporting 32768 — tripping the 64k minimum-context guard.
+    # The suffix-stripping in fetch_ollama_cloud_models() handles the
+    # model-picker UX; this handles the context-length lookup path.
+    for suffix in (":cloud", "-cloud"):
+        suffixed_key = model + suffix
+        entry = models.get(suffixed_key)
+        if entry:
+            ctx = _extract_context(entry)
+            if ctx:
+                return ctx
+        # Also try case-insensitive
+        suffixed_lower = model_lower + suffix
+        for mid, mdata in models.items():
+            if mid.lower() == suffixed_lower:
+                ctx = _extract_context(mdata)
+                if ctx:
+                    return ctx
+
     return None
 
 

agent/moonshot_schema.py

@@ -122,7 +122,7 @@ def _repair_schema(node: Any, is_schema: bool = True) -> Any:
     # empty, drop it entirely.
     if "enum" in repaired and isinstance(repaired["enum"], list):
         node_type = repaired.get("type")
-        if node_type in ("string", "integer", "number", "boolean"):
+        if node_type in {"string", "integer", "number", "boolean"}:
             cleaned = [v for v in repaired["enum"]
                        if v is not None and v != ""]
             if cleaned:
@@ -135,7 +135,7 @@ def _repair_schema(node: Any, is_schema: bool = True) -> Any:
 
 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, ""):
+    if "type" in node and node["type"] not in {None, ""}:
         return node
 
     # Heuristic: presence of ``properties`` → object, ``items`` → array, ``enum``

agent/prompt_builder.py

@@ -157,6 +157,9 @@ MEMORY_GUIDANCE = (
     "User preferences and recurring corrections matter more than procedural task details.\n"
     "Do NOT save task progress, session outcomes, completed-work logs, or temporary TODO "
     "state to memory; use session_search to recall those from past transcripts. "
+    "Specifically: do not record PR numbers, issue numbers, commit SHAs, 'fixed bug X', "
+    "'submitted PR Y', 'Phase N done', file counts, or any artifact that will be stale "
+    "in 7 days. If a fact will be stale in a week, it does not belong in memory. "
     "If you've discovered a new way to do something, solved a problem that could be "
     "necessary later, save it as a skill with the skill tool.\n"
     "Write memories as declarative facts, not instructions to yourself. "
@@ -213,7 +216,15 @@ KANBAN_GUIDANCE = (
     "artifacts. `metadata` is machine-readable facts "
     "(`{changed_files: [...], tests_run: N, decisions: [...]}`). Downstream "
     "workers read both via their own `kanban_show`. Never put secrets / "
-    "tokens / raw PII in either field — run rows are durable forever.\n"
+    "tokens / raw PII in either field — run rows are durable forever. "
+    "Exception: if your output is a code change that needs human review "
+    "before counting as merged/done (most coding tasks), drop the "
+    "structured metadata (changed_files / tests_run / diff_path) into a "
+    "`kanban_comment` first, then end with "
+    "`kanban_block(reason=\"review-required: <one-line summary>\")` so a "
+    "reviewer can approve+unblock or request changes. Reviewing-then-"
+    "completing is more honest than auto-completing work that still needs "
+    "eyes on it.\n"
     "6. **If follow-up work appears, create it; don't do it.** Use "
     "`kanban_create(title=..., assignee=<right-profile>, parents=[your-task-id])` "
     "to spawn a child task for the appropriate specialist profile instead of "

agent/prompt_caching.py

@@ -1,15 +1,25 @@
-"""Anthropic prompt caching (system_and_3 strategy).
+"""Anthropic prompt caching strategies.
 
-Reduces input token costs by ~75% on multi-turn conversations by caching
-the conversation prefix. Uses 4 cache_control breakpoints (Anthropic max):
-  1. System prompt (stable across all turns)
-  2-4. Last 3 non-system messages (rolling window)
+Two layouts:
+
+* ``system_and_3`` (default, used everywhere except the long-lived path):
+  4 cache_control breakpoints — system prompt + last 3 non-system messages.
+  All at the same TTL (5m or 1h). Reduces input token costs by ~75% on
+  multi-turn conversations within a single session.
+
+* ``prefix_and_2`` (Claude on Anthropic / OpenRouter / Nous Portal):
+  4 breakpoints split across two TTL tiers — tools[-1] (1h) +
+  stable system prefix (1h) + last 2 non-system messages (5m). The
+  long-lived prefix is byte-stable across sessions for a given user
+  config, so every fresh session reads the cached system+tools instead
+  of re-paying for them. Within-session rolling window shrinks from 3
+  messages to 2 to free the breakpoint budget.
 
 Pure functions -- no class state, no AIAgent dependency.
 """
 
 import copy
-from typing import Any, Dict, List
+from typing import Any, Dict, List, Optional
 
 
 def _apply_cache_marker(msg: dict, cache_marker: dict, native_anthropic: bool = False) -> None:
@@ -38,6 +48,14 @@ def _apply_cache_marker(msg: dict, cache_marker: dict, native_anthropic: bool =
             last["cache_control"] = cache_marker
 
 
+def _build_marker(ttl: str) -> Dict[str, str]:
+    """Build a cache_control marker dict for the given TTL ('5m' or '1h')."""
+    marker: Dict[str, str] = {"type": "ephemeral"}
+    if ttl == "1h":
+        marker["ttl"] = "1h"
+    return marker
+
+
 def apply_anthropic_cache_control(
     api_messages: List[Dict[str, Any]],
     cache_ttl: str = "5m",
@@ -45,7 +63,8 @@ def apply_anthropic_cache_control(
 ) -> List[Dict[str, Any]]:
     """Apply system_and_3 caching strategy to messages for Anthropic models.
 
-    Places up to 4 cache_control breakpoints: system prompt + last 3 non-system messages.
+    Places up to 4 cache_control breakpoints: system prompt + last 3 non-system
+    messages, all at the same TTL.
 
     Returns:
         Deep copy of messages with cache_control breakpoints injected.
@@ -54,9 +73,7 @@ def apply_anthropic_cache_control(
     if not messages:
         return messages
 
-    marker = {"type": "ephemeral"}
-    if cache_ttl == "1h":
-        marker["ttl"] = "1h"
+    marker = _build_marker(cache_ttl)
 
     breakpoints_used = 0
 
@@ -70,3 +87,115 @@ def apply_anthropic_cache_control(
         _apply_cache_marker(messages[idx], marker, native_anthropic=native_anthropic)
 
     return messages
+
+
+def _mark_system_stable_block(
+    messages: List[Dict[str, Any]],
+    long_lived_marker: Dict[str, str],
+) -> bool:
+    """Mark the *first* content block of the system message with the 1h marker.
+
+    The system message is expected to have been split into multiple content
+    blocks beforehand by the caller — block[0] is the cross-session-stable
+    prefix, subsequent blocks carry context files + volatile suffix.
+    Falls back to marking the whole system message as a single block when
+    the message hasn't been split (preserves correctness on the fallback path).
+
+    Returns True when a marker was placed.
+    """
+    if not messages or messages[0].get("role") != "system":
+        return False
+
+    sys_msg = messages[0]
+    content = sys_msg.get("content")
+
+    # Already a list of blocks → mark the first block.
+    if isinstance(content, list) and content:
+        first = content[0]
+        if isinstance(first, dict):
+            first["cache_control"] = long_lived_marker
+            return True
+        return False
+
+    # String content (no split) → cannot place a stable-prefix breakpoint
+    # without changing the byte content.  Caller is responsible for
+    # splitting; if they didn't, fall through to envelope marker so we still
+    # cache *something* for this turn.
+    if isinstance(content, str) and content:
+        sys_msg["content"] = [
+            {"type": "text", "text": content, "cache_control": long_lived_marker}
+        ]
+        return True
+
+    return False
+
+
+def apply_anthropic_cache_control_long_lived(
+    api_messages: List[Dict[str, Any]],
+    long_lived_ttl: str = "1h",
+    rolling_ttl: str = "5m",
+    native_anthropic: bool = False,
+) -> List[Dict[str, Any]]:
+    """Apply prefix_and_2 caching: long-lived stable prefix + rolling window.
+
+    Layout (4 breakpoints total):
+      * Stable system prefix (block[0]) → ``long_lived_ttl`` TTL
+      * Last 2 non-system messages → ``rolling_ttl`` TTL each
+
+    NOTE: this function does NOT mark the tools array. Tools cache_control
+    is attached separately (see ``mark_tools_for_long_lived_cache``) because
+    tools live outside the messages list in the API payload.
+
+    The caller MUST have split the system message into ordered content
+    blocks where block[0] is the cross-session-stable portion. If the system
+    message is still a single string, it is wrapped into a single block and
+    marked — this is correct, just less effective (the volatile suffix is
+    not isolated, so the prefix invalidates per-session).
+
+    Returns:
+        Deep copy of messages with cache_control breakpoints injected.
+    """
+    messages = copy.deepcopy(api_messages)
+    if not messages:
+        return messages
+
+    long_marker = _build_marker(long_lived_ttl)
+    rolling_marker = _build_marker(rolling_ttl)
+
+    placed_prefix = _mark_system_stable_block(messages, long_marker)
+
+    # Reserve 1 breakpoint for the system prefix (when placed); spend the
+    # remaining 3 on the rolling tail.  Anthropic max is 4 total —
+    # tools[-1] (when marked) consumes the 4th, so we cap rolling at 2 here.
+    rolling_budget = 2 if placed_prefix else 3
+    non_sys = [i for i in range(len(messages)) if messages[i].get("role") != "system"]
+    for idx in non_sys[-rolling_budget:]:
+        _apply_cache_marker(messages[idx], rolling_marker, native_anthropic=native_anthropic)
+
+    return messages
+
+
+def mark_tools_for_long_lived_cache(
+    tools: Optional[List[Dict[str, Any]]],
+    long_lived_ttl: str = "1h",
+) -> Optional[List[Dict[str, Any]]]:
+    """Attach cache_control to the last tool in the OpenAI-format tools list.
+
+    Anthropic prefix-cache order is ``tools → system → messages``.  Marking
+    the last tool dict caches the entire tools array (Anthropic's docs:
+    "the marker is placed on the last block you want included in the cached
+    prefix").  Marker is preserved across the OpenAI-wire boundary on
+    OpenRouter and Nous Portal (which proxies to OpenRouter); on native
+    Anthropic the marker is forwarded by ``convert_tools_to_anthropic``.
+
+    Returns a deep copy of the tools list with the marker attached, or the
+    input unchanged when tools is empty/None.  Pure function — does not
+    mutate the input.
+    """
+    if not tools:
+        return tools
+    out = copy.deepcopy(tools)
+    last = out[-1]
+    if isinstance(last, dict):
+        last["cache_control"] = _build_marker(long_lived_ttl)
+    return out

agent/redact.py

@@ -64,7 +64,7 @@ _SENSITIVE_BODY_KEYS = frozenset({
 # cli.py) or `HERMES_REDACT_SECRETS=false` in ~/.hermes/.env. An opt-out
 # warning is logged at gateway and CLI startup so operators see the
 # downgrade — see `_log_redaction_status()` in gateway/run.py and cli.py.
-_REDACT_ENABLED = os.getenv("HERMES_REDACT_SECRETS", "true").lower() in ("1", "true", "yes", "on")
+_REDACT_ENABLED = os.getenv("HERMES_REDACT_SECRETS", "true").lower() in {"1", "true", "yes", "on"}
 
 # Known API key prefixes -- match the prefix + contiguous token chars
 _PREFIX_PATTERNS = [

agent/shell_hooks.py

@@ -312,7 +312,7 @@ def _parse_single_entry(
         )
         matcher = None
 
-    if matcher is not None and event not in ("pre_tool_call", "post_tool_call"):
+    if matcher is not None and event not in {"pre_tool_call", "post_tool_call"}:
         logger.warning(
             "hooks.%s[%d].matcher=%r will be ignored at runtime — the "
             "matcher field is only honored for pre_tool_call / "
@@ -423,7 +423,7 @@ def _make_callback(spec: ShellHookSpec) -> Callable[..., Optional[Dict[str, Any]
 
     def _callback(**kwargs: Any) -> Optional[Dict[str, Any]]:
         # Matcher gate — only meaningful for tool-scoped events.
-        if spec.event in ("pre_tool_call", "post_tool_call"):
+        if spec.event in {"pre_tool_call", "post_tool_call"}:
             if not spec.matches_tool(kwargs.get("tool_name")):
                 return None
 
@@ -658,7 +658,7 @@ def _prompt_and_record(
         print()  # keep the terminal tidy after ^C
         return False
 
-    if answer in ("y", "yes"):
+    if answer in {"y", "yes"}:
         _record_approval(event, command)
         return True
 
@@ -752,13 +752,13 @@ def _resolve_effective_accept(
     if accept_hooks_arg:
         return True
     env = os.environ.get("HERMES_ACCEPT_HOOKS", "").strip().lower()
-    if env in ("1", "true", "yes", "on"):
+    if env in {"1", "true", "yes", "on"}:
         return True
     cfg_val = cfg.get("hooks_auto_accept", False)
     if isinstance(cfg_val, bool):
         return cfg_val
     if isinstance(cfg_val, str):
-        return cfg_val.strip().lower() in ("1", "true", "yes", "on")
+        return cfg_val.strip().lower() in {"1", "true", "yes", "on"}
     return False
 
 

agent/skill_commands.py

@@ -261,7 +261,7 @@ def scan_skill_commands() -> Dict[str, Dict[str, Any]]:
 
         for scan_dir in dirs_to_scan:
             for skill_md in iter_skill_index_files(scan_dir, "SKILL.md"):
-                if any(part in ('.git', '.github', '.hub', '.archive') for part in skill_md.parts):
+                if any(part in {'.git', '.github', '.hub', '.archive'} for part in skill_md.parts):
                     continue
                 try:
                     content = skill_md.read_text(encoding='utf-8')

agent/transports/chat_completions.py

@@ -279,7 +279,7 @@ class ChatCompletionsTransport(ProviderTransport):
                 _kimi_effort = "medium"
                 if reasoning_config and isinstance(reasoning_config, dict):
                     _e = (reasoning_config.get("effort") or "").strip().lower()
-                    if _e in ("low", "medium", "high"):
+                    if _e in {"low", "medium", "high"}:
                         _kimi_effort = _e
                 api_kwargs["reasoning_effort"] = _kimi_effort
 
@@ -294,7 +294,7 @@ class ChatCompletionsTransport(ProviderTransport):
                 _tokenhub_effort = "high"
                 if reasoning_config and isinstance(reasoning_config, dict):
                     _e = (reasoning_config.get("effort") or "").strip().lower()
-                    if _e in ("low", "medium", "high"):
+                    if _e in {"low", "medium", "high"}:
                         _tokenhub_effort = _e
                 api_kwargs["reasoning_effort"] = _tokenhub_effort
 
@@ -323,6 +323,21 @@ class ChatCompletionsTransport(ProviderTransport):
         if provider_prefs and is_openrouter:
             extra_body["provider"] = provider_prefs
 
+        # Pareto Code router plugin — model-gated. Same shape as the
+        # profile path in plugins/model-providers/openrouter/__init__.py;
+        # this branch only runs when the OpenRouter profile isn't loaded.
+        if is_openrouter and model == "openrouter/pareto-code":
+            _pareto_score = params.get("openrouter_min_coding_score")
+            if _pareto_score is not None and _pareto_score != "":
+                try:
+                    _pareto_score_f = float(_pareto_score)
+                except (TypeError, ValueError):
+                    _pareto_score_f = None
+                if _pareto_score_f is not None and 0.0 <= _pareto_score_f <= 1.0:
+                    extra_body["plugins"] = [
+                        {"id": "pareto-router", "min_coding_score": _pareto_score_f}
+                    ]
+
         # Kimi extra_body.thinking
         if is_kimi:
             _kimi_thinking_enabled = True
@@ -448,6 +463,7 @@ class ChatCompletionsTransport(ProviderTransport):
                 qwen_session_metadata=params.get("qwen_session_metadata"),
                 model=model,
                 ollama_num_ctx=params.get("ollama_num_ctx"),
+                session_id=params.get("session_id"),
             )
         )
         api_kwargs.update(top_level_from_profile)
@@ -462,6 +478,7 @@ class ChatCompletionsTransport(ProviderTransport):
             model=model,
             base_url=params.get("base_url"),
             reasoning_config=reasoning_config,
+            openrouter_min_coding_score=params.get("openrouter_min_coding_score"),
         )
         if profile_body:
             extra_body.update(profile_body)

agent/transports/codex.py

@@ -104,7 +104,16 @@ class ResponsesApiTransport(ProviderTransport):
             kwargs["prompt_cache_key"] = session_id
 
         if reasoning_enabled and is_xai_responses:
+            from agent.model_metadata import grok_supports_reasoning_effort
+
             kwargs["include"] = ["reasoning.encrypted_content"]
+            # xAI rejects `reasoning.effort` on grok-4 / grok-4-fast / grok-3
+            # / grok-code-fast / grok-4.20-0309-* with HTTP 400 even though
+            # those models reason natively. Only send the effort dial when
+            # the target model is on the allowlist; otherwise send no
+            # `reasoning` key at all and let the model reason on its own.
+            if grok_supports_reasoning_effort(model):
+                kwargs["reasoning"] = {"effort": reasoning_effort}
         elif reasoning_enabled:
             if is_github_responses:
                 github_reasoning = params.get("github_reasoning_extra")

batch_runner.py

@@ -337,6 +337,7 @@ def _process_single_prompt(
             providers_ignored=config.get("providers_ignored"),
             providers_order=config.get("providers_order"),
             provider_sort=config.get("provider_sort"),
+            openrouter_min_coding_score=config.get("openrouter_min_coding_score"),
             max_tokens=config.get("max_tokens"),
             reasoning_config=config.get("reasoning_config"),
             prefill_messages=config.get("prefill_messages"),
@@ -546,6 +547,7 @@ class BatchRunner:
         providers_ignored: List[str] = None,
         providers_order: List[str] = None,
         provider_sort: str = None,
+        openrouter_min_coding_score: Optional[float] = None,
         max_tokens: int = None,
         reasoning_config: Dict[str, Any] = None,
         prefill_messages: List[Dict[str, Any]] = None,
@@ -595,6 +597,7 @@ class BatchRunner:
         self.providers_ignored = providers_ignored
         self.providers_order = providers_order
         self.provider_sort = provider_sort
+        self.openrouter_min_coding_score = openrouter_min_coding_score
         self.max_tokens = max_tokens
         self.reasoning_config = reasoning_config
         self.prefill_messages = prefill_messages
@@ -792,7 +795,7 @@ class BatchRunner:
                 conversations = entry.get("conversations", [])
                 for msg in conversations:
                     role = msg.get("role") or msg.get("from")
-                    if role in ("user", "human"):
+                    if role in {"user", "human"}:
                         prompt_text = (msg.get("content") or msg.get("value", "")).strip()
                         break
             
@@ -873,6 +876,7 @@ class BatchRunner:
             "providers_ignored": self.providers_ignored,
             "providers_order": self.providers_order,
             "provider_sort": self.provider_sort,
+            "openrouter_min_coding_score": self.openrouter_min_coding_score,
             "max_tokens": self.max_tokens,
             "reasoning_config": self.reasoning_config,
             "prefill_messages": self.prefill_messages,

cli-config.yaml.example

@@ -203,6 +203,12 @@ terminal:
 #   docker_forward_env:
 #     - "GITHUB_TOKEN"
 #     - "NPM_TOKEN"
+#   # Optional: extra flags passed verbatim to docker run (appended after security defaults).
+#   # Useful for adding capabilities (e.g. apt installs needing SETUID) or custom options.
+#   # Example: add a Linux capability not included by default
+#   # docker_extra_args:
+#   #   - "--cap-add"
+#   #   - "SETUID"
 
 # -----------------------------------------------------------------------------
 # OPTION 4: Singularity/Apptainer container
@@ -657,6 +663,10 @@ platform_toolsets:
 # platforms:
 #   telegram:
 #     reply_to_mode: "first"  # off | first | all
+#     # guest_mode lets explicit @mentions from non-allowlisted groups through.
+#     # Default false; ordinary messages, replies, and regex wake words stay blocked.
+#     guest_mode: false
+#     # allowed_chats: ["-1001234567890"]
 #     extra:
 #       disable_link_previews: false  # Set true to suppress Telegram URL previews in bot messages
 
@@ -943,6 +953,9 @@ display:
   #   false: Wait for the full response before rendering
   streaming: true
 
+  # Show [HH:MM] timestamps on user input and assistant response labels.
+  # timestamps: false
+
   # ───────────────────────────────────────────────────────────────────────────
   # Skin / Theme
   # ───────────────────────────────────────────────────────────────────────────

cron/jobs.py

@@ -664,7 +664,7 @@ def update_job(job_id: str, updates: Dict[str, Any]) -> Optional[Dict[str, Any]]
         # None both mean "clear the field" (restore old behaviour).
         if "workdir" in updates:
             _wd = updates["workdir"]
-            if _wd in (None, "", False):
+            if _wd in {None, "", False}:
                 updates["workdir"] = None
             else:
                 updates["workdir"] = _normalize_workdir(_wd)
@@ -811,7 +811,7 @@ def mark_job_run(job_id: str, success: bool, error: Optional[str] = None,
                 # schedule quietly goes off. See issue #16265.
                 if job["next_run_at"] is None:
                     kind = job.get("schedule", {}).get("kind")
-                    if kind in ("cron", "interval"):
+                    if kind in {"cron", "interval"}:
                         job["state"] = "error"
                         if not job.get("last_error"):
                             job["last_error"] = (
@@ -855,7 +855,7 @@ def advance_next_run(job_id: str) -> bool:
         for job in jobs:
             if job["id"] == job_id:
                 kind = job.get("schedule", {}).get("kind")
-                if kind not in ("cron", "interval"):
+                if kind not in {"cron", "interval"}:
                     return False
                 now = _hermes_now().isoformat()
                 new_next = compute_next_run(job["schedule"], now)
@@ -909,7 +909,7 @@ def _get_due_jobs_locked() -> List[Dict[str, Any]]:
             # next_run_at unset.  Without this branch, such jobs are
             # silently skipped forever; recompute next_run_at from the
             # schedule so they pick up at their next scheduled tick.
-            if not recovered_next and kind in ("cron", "interval"):
+            if not recovered_next and kind in {"cron", "interval"}:
                 recovered_next = compute_next_run(schedule, now.isoformat())
                 if recovered_next:
                     recovery_kind = kind
@@ -940,7 +940,7 @@ def _get_due_jobs_locked() -> List[Dict[str, Any]]:
             # (gateway was down and missed the window). Fast-forward to
             # the next future occurrence instead of firing a stale run.
             grace = _compute_grace_seconds(schedule)
-            if kind in ("cron", "interval") and (now - next_run_dt).total_seconds() > grace:
+            if kind in {"cron", "interval"} and (now - next_run_dt).total_seconds() > grace:
                 # Job is past its catch-up grace window — this is a stale missed run.
                 # Grace scales with schedule period: daily=2h, hourly=30m, 10min=5m.
                 new_next = compute_next_run(schedule, now.isoformat())
@@ -1082,9 +1082,8 @@ def rewrite_skill_refs(
                         new_skills.append(target)
                 elif name in pruned_set:
                     dropped.append(name)
-                else:
-                    if name not in new_skills:
-                        new_skills.append(name)
+                elif name not in new_skills:
+                    new_skills.append(name)
 
             if not mapped and not dropped:
                 continue

cron/scheduler.py

@@ -754,7 +754,7 @@ def _run_job_script(script_path: str) -> tuple[bool, str]:
     # shebang: the scripts dir is trusted, but keeping the interpreter
     # choice explicit here keeps the allowed surface small and auditable.
     suffix = path.suffix.lower()
-    if suffix in (".sh", ".bash"):
+    if suffix in {".sh", ".bash"}:
         # Resolve bash dynamically so Windows (Git Bash) and Linux/macOS
         # all work.  On native Windows without Git for Windows installed
         # shutil.which returns None — fall back to a clear error rather
@@ -1439,6 +1439,7 @@ 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"),
+            openrouter_min_coding_score=(_cfg.get("openrouter") or {}).get("min_coding_score"),
             enabled_toolsets=_resolve_cron_enabled_toolsets(job, _cfg),
             disabled_toolsets=["cronjob", "messaging", "clarify"],
             quiet_mode=True,

environments/agentic_opd_env.py

@@ -264,7 +264,7 @@ def _parse_hint_result(text: str) -> tuple[int | None, str]:
     """Parse the judge's boxed decision and hint text."""
     boxed = _BOXED_RE.findall(text)
     score = int(boxed[-1]) if boxed else None
-    if score not in (1, -1):
+    if score not in {1, -1}:
         score = None
     hint_matches = _HINT_RE.findall(text)
     hint = hint_matches[-1].strip() if hint_matches else ""

environments/benchmarks/terminalbench_2/terminalbench2_env.py

@@ -162,7 +162,7 @@ def _normalize_tar_member_parts(member_name: str) -> list:
     ):
         raise ValueError(f"Unsafe archive member path: {member_name}")
 
-    parts = [part for part in posix_path.parts if part not in ("", ".")]
+    parts = [part for part in posix_path.parts if part not in {"", "."}]
     if not parts or any(part == ".." for part in parts):
         raise ValueError(f"Unsafe archive member path: {member_name}")
     return parts
@@ -561,7 +561,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
             # --- 5. Verify -- run test suite in the agent's sandbox ---
             # Skip verification if the agent produced no meaningful output
             only_system_and_user = all(
-                msg.get("role") in ("system", "user") for msg in result.messages
+                msg.get("role") in {"system", "user"} for msg in result.messages
             )
             if result.turns_used == 0 or only_system_and_user:
                 logger.warning(
@@ -919,7 +919,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
             eval_metrics[f"eval/pass_rate_{cat_key}"] = cat_pass_rate
 
         # Store metrics for wandb_log
-        self.eval_metrics = [(k, v) for k, v in eval_metrics.items()]
+        self.eval_metrics = list(eval_metrics.items())
 
         # ---- Print summary ----
         print(f"\n{'='*60}")

environments/benchmarks/yc_bench/yc_bench_env.py

@@ -759,7 +759,7 @@ class YCBenchEvalEnv(HermesAgentBaseEnv):
             eval_metrics[f"eval/survival_rate_{key}"] = ps / pt if pt else 0
             eval_metrics[f"eval/avg_score_{key}"] = pa
 
-        self.eval_metrics = [(k, v) for k, v in eval_metrics.items()]
+        self.eval_metrics = list(eval_metrics.items())
 
         # --- Print summary ---
         print(f"\n{'='*60}")

environments/hermes_base_env.py

@@ -571,7 +571,7 @@ class HermesAgentBaseEnv(BaseEnv):
         # (e.g., API call failed on turn 1). No point spinning up a Modal sandbox
         # just to verify files that were never created.
         only_system_and_user = all(
-            msg.get("role") in ("system", "user") for msg in result.messages
+            msg.get("role") in {"system", "user"} for msg in result.messages
         )
         if result.turns_used == 0 or only_system_and_user:
             logger.warning(

environments/tool_context.py

@@ -179,7 +179,7 @@ class ToolContext:
 
         # Ensure parent directory exists in the sandbox
         parent = str(_Path(remote_path).parent)
-        if parent not in (".", "/"):
+        if parent not in {".", "/"}:
             self.terminal(f"mkdir -p {parent}", timeout=10)
 
         # For small files, single command is fine

gateway/config.py

@@ -28,9 +28,9 @@ def _coerce_bool(value: Any, default: bool = True) -> bool:
         return default
     if isinstance(value, str):
         lowered = value.strip().lower()
-        if lowered in ("true", "1", "yes", "on"):
+        if lowered in {"true", "1", "yes", "on"}:
             return True
-        if lowered in ("false", "0", "no", "off"):
+        if lowered in {"false", "0", "no", "off"}:
             return False
         return default
     return is_truthy_value(value, default=default)
@@ -317,14 +317,32 @@ class PlatformConfig:
         )
 
 
+# Streaming defaults — single source of truth so both StreamingConfig and
+# StreamConsumerConfig agree on the out-of-the-box edit rhythm.  Tuned for
+# Telegram's ~1 edit/s flood envelope: a touch under 1s lets the cadence
+# breathe without bumping into rate limits, and a smaller buffer threshold
+# makes short replies feel near-instant in DMs.
+DEFAULT_STREAMING_EDIT_INTERVAL: float = 0.8
+DEFAULT_STREAMING_BUFFER_THRESHOLD: int = 24
+DEFAULT_STREAMING_CURSOR: str = " ▉"
+
+
 @dataclass
 class StreamingConfig:
     """Configuration for real-time token streaming to messaging platforms."""
     enabled: bool = False
-    transport: str = "edit"       # "edit" (progressive editMessageText) or "off"
-    edit_interval: float = 1.0    # Seconds between message edits (Telegram rate-limits at ~1/s)
-    buffer_threshold: int = 40    # Chars before forcing an edit
-    cursor: str = " ▉"           # Cursor shown during streaming
+    # Transport selection:
+    #   "auto"  — prefer native streaming-draft updates when the platform
+    #             supports them (Telegram sendMessageDraft, Bot API 9.5+);
+    #             fall back to edit-based when not.  Recommended.
+    #   "draft" — explicitly request native drafts; falls back to edit when
+    #             the platform/chat doesn't support them.
+    #   "edit"  — progressive editMessageText only (legacy behaviour).
+    #   "off"   — disable streaming entirely.
+    transport: str = "auto"
+    edit_interval: float = DEFAULT_STREAMING_EDIT_INTERVAL
+    buffer_threshold: int = DEFAULT_STREAMING_BUFFER_THRESHOLD
+    cursor: str = DEFAULT_STREAMING_CURSOR
     # Ported from openclaw/openclaw#72038.  When >0, the final edit for
     # a long-running streamed response is delivered as a fresh message
     # if the original preview has been visible for at least this many
@@ -350,10 +368,14 @@ class StreamingConfig:
             return cls()
         return cls(
             enabled=_coerce_bool(data.get("enabled"), False),
-            transport=data.get("transport", "edit"),
-            edit_interval=_coerce_float(data.get("edit_interval"), 1.0),
-            buffer_threshold=_coerce_int(data.get("buffer_threshold"), 40),
-            cursor=data.get("cursor", " ▉"),
+            transport=data.get("transport", "auto"),
+            edit_interval=_coerce_float(
+                data.get("edit_interval"), DEFAULT_STREAMING_EDIT_INTERVAL,
+            ),
+            buffer_threshold=_coerce_int(
+                data.get("buffer_threshold"), DEFAULT_STREAMING_BUFFER_THRESHOLD,
+            ),
+            cursor=data.get("cursor", DEFAULT_STREAMING_CURSOR),
             fresh_final_after_seconds=_coerce_float(
                 data.get("fresh_final_after_seconds"), 60.0
             ),
@@ -588,8 +610,7 @@ class GatewayConfig:
 
         try:
             session_store_max_age_days = int(data.get("session_store_max_age_days", 90))
-            if session_store_max_age_days < 0:
-                session_store_max_age_days = 0
+            session_store_max_age_days = max(session_store_max_age_days, 0)
         except (TypeError, ValueError):
             session_store_max_age_days = 90
 
@@ -766,11 +787,19 @@ def load_gateway_config() -> GatewayConfig:
                     bridged["dm_policy"] = platform_cfg["dm_policy"]
                 if "allow_from" in platform_cfg:
                     bridged["allow_from"] = platform_cfg["allow_from"]
+                if "allow_admin_from" in platform_cfg:
+                    bridged["allow_admin_from"] = platform_cfg["allow_admin_from"]
+                if "user_allowed_commands" in platform_cfg:
+                    bridged["user_allowed_commands"] = platform_cfg["user_allowed_commands"]
                 if "group_policy" in platform_cfg:
                     bridged["group_policy"] = platform_cfg["group_policy"]
                 if "group_allow_from" in platform_cfg:
                     bridged["group_allow_from"] = platform_cfg["group_allow_from"]
-                if plat in (Platform.DISCORD, Platform.SLACK) and "channel_skill_bindings" in platform_cfg:
+                if "group_allow_admin_from" in platform_cfg:
+                    bridged["group_allow_admin_from"] = platform_cfg["group_allow_admin_from"]
+                if "group_user_allowed_commands" in platform_cfg:
+                    bridged["group_user_allowed_commands"] = platform_cfg["group_user_allowed_commands"]
+                if plat in {Platform.DISCORD, Platform.SLACK} and "channel_skill_bindings" in platform_cfg:
                     bridged["channel_skill_bindings"] = platform_cfg["channel_skill_bindings"]
                 if "channel_prompts" in platform_cfg:
                     channel_prompts = platform_cfg["channel_prompts"]
@@ -896,6 +925,8 @@ def load_gateway_config() -> GatewayConfig:
                     os.environ["TELEGRAM_REQUIRE_MENTION"] = str(_effective_rm).lower()
                 if "mention_patterns" in telegram_cfg and not os.getenv("TELEGRAM_MENTION_PATTERNS"):
                     os.environ["TELEGRAM_MENTION_PATTERNS"] = json.dumps(telegram_cfg["mention_patterns"])
+                if "guest_mode" in telegram_cfg and not os.getenv("TELEGRAM_GUEST_MODE"):
+                    os.environ["TELEGRAM_GUEST_MODE"] = str(telegram_cfg["guest_mode"]).lower()
                 frc = telegram_cfg.get("free_response_chats")
                 if frc is not None and not os.getenv("TELEGRAM_FREE_RESPONSE_CHATS"):
                     if isinstance(frc, list):
@@ -941,16 +972,17 @@ def load_gateway_config() -> GatewayConfig:
                     if isinstance(group_allowed_chats, list):
                         group_allowed_chats = ",".join(str(v) for v in group_allowed_chats)
                     os.environ["TELEGRAM_GROUP_ALLOWED_CHATS"] = str(group_allowed_chats)
-                if "disable_link_previews" in telegram_cfg:
-                    plat_data = platforms_data.setdefault(Platform.TELEGRAM.value, {})
-                    if not isinstance(plat_data, dict):
-                        plat_data = {}
-                        platforms_data[Platform.TELEGRAM.value] = plat_data
-                    extra = plat_data.setdefault("extra", {})
-                    if not isinstance(extra, dict):
-                        extra = {}
-                        plat_data["extra"] = extra
-                    extra["disable_link_previews"] = telegram_cfg["disable_link_previews"]
+                for _telegram_extra_key in ("guest_mode", "disable_link_previews"):
+                    if _telegram_extra_key in telegram_cfg:
+                        plat_data = platforms_data.setdefault(Platform.TELEGRAM.value, {})
+                        if not isinstance(plat_data, dict):
+                            plat_data = {}
+                            platforms_data[Platform.TELEGRAM.value] = plat_data
+                        extra = plat_data.setdefault("extra", {})
+                        if not isinstance(extra, dict):
+                            extra = {}
+                            plat_data["extra"] = extra
+                        extra[_telegram_extra_key] = telegram_cfg[_telegram_extra_key]
 
             whatsapp_cfg = yaml_cfg.get("whatsapp", {})
             if isinstance(whatsapp_cfg, dict):
@@ -1147,7 +1179,7 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
     
     # Reply threading mode for Telegram (off/first/all)
     telegram_reply_mode = os.getenv("TELEGRAM_REPLY_TO_MODE", "").lower()
-    if telegram_reply_mode in ("off", "first", "all"):
+    if telegram_reply_mode in {"off", "first", "all"}:
         if Platform.TELEGRAM not in config.platforms:
             config.platforms[Platform.TELEGRAM] = PlatformConfig()
         config.platforms[Platform.TELEGRAM].reply_to_mode = telegram_reply_mode
@@ -1188,14 +1220,14 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
     
     # Reply threading mode for Discord (off/first/all)
     discord_reply_mode = os.getenv("DISCORD_REPLY_TO_MODE", "").lower()
-    if discord_reply_mode in ("off", "first", "all"):
+    if discord_reply_mode in {"off", "first", "all"}:
         if Platform.DISCORD not in config.platforms:
             config.platforms[Platform.DISCORD] = PlatformConfig()
         config.platforms[Platform.DISCORD].reply_to_mode = discord_reply_mode
     
     # WhatsApp (typically uses different auth mechanism)
-    whatsapp_enabled = os.getenv("WHATSAPP_ENABLED", "").lower() in ("true", "1", "yes")
-    whatsapp_disabled_explicitly = os.getenv("WHATSAPP_ENABLED", "").lower() in ("false", "0", "no")
+    whatsapp_enabled = os.getenv("WHATSAPP_ENABLED", "").lower() in {"true", "1", "yes"}
+    whatsapp_disabled_explicitly = os.getenv("WHATSAPP_ENABLED", "").lower() in {"false", "0", "no"}
     if Platform.WHATSAPP in config.platforms:
         # YAML config exists — respect explicit disable
         wa_cfg = config.platforms[Platform.WHATSAPP]
@@ -1253,7 +1285,7 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         config.platforms[Platform.SIGNAL].extra.update({
             "http_url": signal_url,
             "account": signal_account,
-            "ignore_stories": os.getenv("SIGNAL_IGNORE_STORIES", "true").lower() in ("true", "1", "yes"),
+            "ignore_stories": os.getenv("SIGNAL_IGNORE_STORIES", "true").lower() in {"true", "1", "yes"},
         })
     signal_home = os.getenv("SIGNAL_HOME_CHANNEL")
     if signal_home and Platform.SIGNAL in config.platforms:
@@ -1302,7 +1334,7 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         matrix_password = os.getenv("MATRIX_PASSWORD", "")
         if matrix_password:
             config.platforms[Platform.MATRIX].extra["password"] = matrix_password
-        matrix_e2ee = os.getenv("MATRIX_ENCRYPTION", "").lower() in ("true", "1", "yes")
+        matrix_e2ee = os.getenv("MATRIX_ENCRYPTION", "").lower() in {"true", "1", "yes"}
         config.platforms[Platform.MATRIX].extra["encryption"] = matrix_e2ee
         matrix_device_id = os.getenv("MATRIX_DEVICE_ID", "")
         if matrix_device_id:
@@ -1367,7 +1399,7 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         )
 
     # API Server
-    api_server_enabled = os.getenv("API_SERVER_ENABLED", "").lower() in ("true", "1", "yes")
+    api_server_enabled = os.getenv("API_SERVER_ENABLED", "").lower() in {"true", "1", "yes"}
     api_server_key = os.getenv("API_SERVER_KEY", "")
     api_server_cors_origins = os.getenv("API_SERVER_CORS_ORIGINS", "")
     api_server_port = os.getenv("API_SERVER_PORT")
@@ -1394,7 +1426,7 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             config.platforms[Platform.API_SERVER].extra["model_name"] = api_server_model_name
 
     # Webhook platform
-    webhook_enabled = os.getenv("WEBHOOK_ENABLED", "").lower() in ("true", "1", "yes")
+    webhook_enabled = os.getenv("WEBHOOK_ENABLED", "").lower() in {"true", "1", "yes"}
     webhook_port = os.getenv("WEBHOOK_PORT")
     webhook_secret = os.getenv("WEBHOOK_SECRET", "")
     if webhook_enabled:
@@ -1410,11 +1442,11 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             config.platforms[Platform.WEBHOOK].extra["secret"] = webhook_secret
 
     # Microsoft Graph webhook platform
-    msgraph_webhook_enabled = os.getenv("MSGRAPH_WEBHOOK_ENABLED", "").lower() in (
+    msgraph_webhook_enabled = os.getenv("MSGRAPH_WEBHOOK_ENABLED", "").lower() in {
         "true",
         "1",
         "yes",
-    )
+    }
     msgraph_webhook_port = os.getenv("MSGRAPH_WEBHOOK_PORT")
     msgraph_webhook_client_state = os.getenv("MSGRAPH_WEBHOOK_CLIENT_STATE", "")
     msgraph_webhook_resources = os.getenv("MSGRAPH_WEBHOOK_ACCEPTED_RESOURCES", "")
@@ -1608,7 +1640,7 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
             "webhook_host": os.getenv("BLUEBUBBLES_WEBHOOK_HOST", "127.0.0.1"),
             "webhook_port": int(os.getenv("BLUEBUBBLES_WEBHOOK_PORT", "8645")),
             "webhook_path": os.getenv("BLUEBUBBLES_WEBHOOK_PATH", "/bluebubbles-webhook"),
-            "send_read_receipts": os.getenv("BLUEBUBBLES_SEND_READ_RECEIPTS", "true").lower() in ("true", "1", "yes"),
+            "send_read_receipts": os.getenv("BLUEBUBBLES_SEND_READ_RECEIPTS", "true").lower() in {"true", "1", "yes"},
         })
     bluebubbles_home = os.getenv("BLUEBUBBLES_HOME_CHANNEL")
     if bluebubbles_home and Platform.BLUEBUBBLES in config.platforms:

gateway/display_config.py

@@ -81,7 +81,7 @@ _TIER_MINIMAL = {
 
 _PLATFORM_DEFAULTS: dict[str, dict[str, Any]] = {
     # Tier 1 — full edit support, personal/team use
-    "telegram":    _TIER_HIGH,
+    "telegram":    {**_TIER_HIGH, "tool_progress": "new"},
     "discord":     _TIER_HIGH,
 
     # Tier 2 — edit support, often customer/workspace channels
@@ -190,13 +190,13 @@ def _normalise(setting: str, value: Any) -> Any:
         if value is True:
             return "all"
         return str(value).lower()
-    if setting in ("show_reasoning", "streaming"):
+    if setting in {"show_reasoning", "streaming"}:
         if isinstance(value, str):
-            return value.lower() in ("true", "1", "yes", "on")
+            return value.lower() in {"true", "1", "yes", "on"}
         return bool(value)
     if setting == "cleanup_progress":
         if isinstance(value, str):
-            return value.lower() in ("true", "1", "yes", "on")
+            return value.lower() in {"true", "1", "yes", "on"}
         return bool(value)
     if setting == "tool_preview_length":
         try:

gateway/platforms/ADDING_A_PLATFORM.md

@@ -33,6 +33,17 @@ status display, gateway setup, and more.
   auto-populate `OPTIONAL_ENV_VARS` in `hermes_cli/config.py` so the setup
   wizard surfaces proper descriptions, prompts, password flags, and URLs.
 
+**Subclassing for platform-specific UX.** When a platform has a hard
+time-window constraint that the base adapter can't anticipate (LINE's
+60s single-use reply token, WhatsApp's 24h session window, etc.), an
+adapter can override `_keep_typing` to layer a mid-flight bubble at a
+threshold without expanding the kwarg surface. Always
+`await super()._keep_typing(...)` so the typing heartbeat keeps running,
+and tear down your side task in `finally`. See `plugins/platforms/line/`
+for the full pattern (Template Buttons postback at 45s, `RequestCache`
+state machine, `interrupt_session_activity` override for `/stop`
+orphans) and the developer-guide page for the prose walkthrough.
+
 See `plugins/platforms/irc/`, `plugins/platforms/teams/`, and
 `plugins/platforms/google_chat/` for complete working examples, and
 `website/docs/developer-guide/adding-platform-adapters.md` for the full

gateway/platforms/__init__.py

@@ -9,9 +9,19 @@ Each adapter handles:
 """
 
 from .base import BasePlatformAdapter, MessageEvent, SendResult
-from .qqbot import QQAdapter
-from .yuanbao import YuanbaoAdapter
 
+# QQAdapter and YuanbaoAdapter were previously imported eagerly here, but
+# nothing in the codebase consumes ``from gateway.platforms import
+# QQAdapter`` (every real call site uses the long-form path
+# ``from gateway.platforms.qqbot import QQAdapter``). The eager imports
+# pulled in qqbot's chunked-upload + keyboards + onboard machinery and
+# yuanbao's websocket stack — about 48 ms wall and ~8 MB RSS on every
+# CLI invocation, even ones that never touch a gateway adapter.
+#
+# Use PEP 562 module ``__getattr__`` to keep the public re-export working
+# while deferring the actual import to first attribute access. This is
+# 100% backward-compatible for any external code that still imports the
+# adapters from the package root.
 __all__ = [
     "BasePlatformAdapter",
     "MessageEvent",
@@ -19,3 +29,17 @@ __all__ = [
     "QQAdapter",
     "YuanbaoAdapter",
 ]
+
+
+def __getattr__(name):
+    if name == "QQAdapter":
+        from .qqbot import QQAdapter  # noqa: F401
+        return QQAdapter
+    if name == "YuanbaoAdapter":
+        from .yuanbao import YuanbaoAdapter  # noqa: F401
+        return YuanbaoAdapter
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
+
+def __dir__():
+    return sorted(__all__)

gateway/platforms/api_server.py

@@ -449,7 +449,7 @@ if AIOHTTP_AVAILABLE:
     @web.middleware
     async def body_limit_middleware(request, handler):
         """Reject overly large request bodies early based on Content-Length."""
-        if request.method in ("POST", "PUT", "PATCH"):
+        if request.method in {"POST", "PUT", "PATCH"}:
             cl = request.headers.get("Content-Length")
             if cl is not None:
                 try:
@@ -646,7 +646,7 @@ class APIServerAdapter(BasePlatformAdapter):
         try:
             from hermes_cli.profiles import get_active_profile_name
             profile = get_active_profile_name()
-            if profile and profile not in ("default", "custom"):
+            if profile and profile not in {"default", "custom"}:
                 return profile
         except Exception:
             pass
@@ -1003,7 +1003,7 @@ class APIServerAdapter(BasePlatformAdapter):
                     system_prompt = content
                 else:
                     system_prompt = system_prompt + "\n" + content
-            elif role in ("user", "assistant"):
+            elif role in {"user", "assistant"}:
                 try:
                     content = _normalize_multimodal_content(raw_content)
                 except ValueError as exc:
@@ -1206,10 +1206,49 @@ class APIServerAdapter(BasePlatformAdapter):
                     status=500,
                 )
 
-        final_response = result.get("final_response", "")
-        if not final_response:
-            final_response = result.get("error", "(No response generated)")
+        final_response = result.get("final_response") or ""
+        is_partial = bool(result.get("partial"))
+        is_failed = bool(result.get("failed"))
+        completed = bool(result.get("completed", True))
+        err_msg = result.get("error")
 
+        # Decide finish_reason. OpenAI uses "length" for truncation, "stop"
+        # for normal completion, and downstream SDKs accept "error" / custom
+        # codes. See issue #22496.
+        if is_partial and err_msg and "truncat" in err_msg.lower():
+            finish_reason = "length"
+        elif is_failed or (not completed and err_msg):
+            finish_reason = "error"
+        else:
+            finish_reason = "stop"
+
+        response_headers = {
+            "X-Hermes-Session-Id": result.get("session_id", session_id),
+        }
+        if gateway_session_key:
+            response_headers["X-Hermes-Session-Key"] = gateway_session_key
+
+        # Hard-fail path: no usable assistant text AND a real failure → 5xx
+        # with OpenAI-style error envelope so SDK clients raise instead of
+        # silently rendering the internal failure string as message.content.
+        if not final_response and (is_failed or is_partial):
+            err_body = _openai_error(
+                err_msg or "Agent run did not produce a response.",
+                err_type="server_error",
+                code="agent_incomplete",
+            )
+            err_body["error"]["hermes"] = {
+                "completed": completed,
+                "partial": is_partial,
+                "failed": is_failed,
+            }
+            response_headers["X-Hermes-Completed"] = "false"
+            response_headers["X-Hermes-Partial"] = "true" if is_partial else "false"
+            return web.json_response(err_body, status=502, headers=response_headers)
+
+        # Soft-partial path: we have *some* text but the run did not complete
+        # (e.g. truncation with partial buffered output). Still 200 but signal
+        # truncation via finish_reason="length" + Hermes-specific extras.
         response_data = {
             "id": completion_id,
             "object": "chat.completion",
@@ -1222,7 +1261,7 @@ class APIServerAdapter(BasePlatformAdapter):
                         "role": "assistant",
                         "content": final_response,
                     },
-                    "finish_reason": "stop",
+                    "finish_reason": finish_reason,
                 }
             ],
             "usage": {
@@ -1231,12 +1270,19 @@ class APIServerAdapter(BasePlatformAdapter):
                 "total_tokens": usage.get("total_tokens", 0),
             },
         }
+        if is_partial or is_failed or not completed:
+            response_data["hermes"] = {
+                "completed": completed,
+                "partial": is_partial,
+                "failed": is_failed,
+                "error": err_msg,
+                "error_code": "output_truncated" if finish_reason == "length" else "agent_error",
+            }
+            response_headers["X-Hermes-Completed"] = "false"
+            response_headers["X-Hermes-Partial"] = "true" if is_partial else "false"
+            if err_msg:
+                response_headers["X-Hermes-Error"] = err_msg[:200]
 
-        response_headers = {
-            "X-Hermes-Session-Id": result.get("session_id", session_id),
-        }
-        if gateway_session_key:
-            response_headers["X-Hermes-Session-Key"] = gateway_session_key
         return web.json_response(response_data, headers=response_headers)
 
     async def _write_sse_chat_completion(
@@ -2335,7 +2381,7 @@ class APIServerAdapter(BasePlatformAdapter):
         if cron_err:
             return cron_err
         try:
-            include_disabled = request.query.get("include_disabled", "").lower() in ("true", "1")
+            include_disabled = request.query.get("include_disabled", "").lower() in {"true", "1"}
             jobs = _cron_list(include_disabled=include_disabled)
             return web.json_response({"jobs": jobs})
         except Exception as e:

gateway/platforms/base.py

@@ -560,7 +560,7 @@ def _looks_like_image(data: bytes) -> bool:
         return True
     if data[:3] == b"\xff\xd8\xff":
         return True
-    if data[:6] in (b"GIF87a", b"GIF89a"):
+    if data[:6] in {b"GIF87a", b"GIF89a"}:
         return True
     if data[:2] == b"BM":
         return True
@@ -859,7 +859,7 @@ def cache_document_from_bytes(data: bytes, filename: str) -> str:
     # Sanitize: strip directory components, null bytes, and control characters
     safe_name = Path(filename).name if filename else "document"
     safe_name = safe_name.replace("\x00", "").strip()
-    if not safe_name or safe_name in (".", ".."):
+    if not safe_name or safe_name in {".", ".."}:
         safe_name = "document"
     cached_name = f"doc_{uuid.uuid4().hex[:12]}_{safe_name}"
     filepath = cache_dir / cached_name
@@ -1035,6 +1035,13 @@ class SendResult:
     error: Optional[str] = None
     raw_response: Any = None
     retryable: bool = False  # True for transient connection errors — base will retry automatically
+    # When the adapter had to split an oversized payload across multiple
+    # platform messages (e.g. Telegram edit_message overflow split-and-deliver),
+    # ``message_id`` is the LAST visible message id (so subsequent edits target
+    # the most recent chunk) and these are the additional message ids that
+    # made up the full payload, in send order.  Empty tuple for the common
+    # single-message case.
+    continuation_message_ids: tuple = ()
 
 
 class EphemeralReply(str):
@@ -1311,6 +1318,61 @@ class BasePlatformAdapter(ABC):
         # _keep_typing skips send_typing when the chat_id is in this set.
         self._typing_paused: set = set()
 
+    @property
+    def message_len_fn(self) -> Callable[[str], int]:
+        """Return the length function for measuring message size on this platform.
+
+        Override in adapters whose platform counts characters differently from
+        Python ``len`` (e.g. Telegram counts UTF-16 code units).
+        """
+        return len
+
+    def supports_draft_streaming(
+        self,
+        chat_type: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> bool:
+        """Whether this adapter supports native streaming-draft updates.
+
+        Telegram Bot API 9.5 introduced ``sendMessageDraft``, which renders an
+        animated streaming preview as the bot calls it repeatedly with the
+        same ``draft_id`` and growing text.  Adapters that implement
+        ``send_draft`` should return True here for the chat types where the
+        platform supports it (Telegram restricts drafts to private DMs).
+
+        Default implementation returns False.  Stream consumers fall back to
+        the edit-based path (``send`` + ``edit_message``) when this returns
+        False or when ``send_draft`` raises.
+        """
+        return False
+
+    async def send_draft(
+        self,
+        chat_id: str,
+        draft_id: int,
+        content: str,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Send or update an animated streaming-draft preview.
+
+        Reuse the same ``draft_id`` (any non-zero int) across consecutive
+        calls within a single response so the platform animates the preview
+        rather than re-creating it.  Different responses must use different
+        ``draft_id`` values within the same chat to avoid animating over a
+        prior bubble.
+
+        Drafts have no message_id and cannot be edited, replied to, or
+        deleted via normal message APIs.  When the response finishes, the
+        caller delivers the final answer as a regular ``send`` and the
+        draft preview clears naturally on the client.
+
+        Default implementation raises NotImplementedError; adapters that
+        also return True from :meth:`supports_draft_streaming` must override.
+        """
+        raise NotImplementedError(
+            f"{type(self).__name__} does not implement send_draft"
+        )
+
     @property
     def has_fatal_error(self) -> bool:
         return self._fatal_error_message is not None
@@ -1511,6 +1573,33 @@ class BasePlatformAdapter(ABC):
     # property) so the stream consumer knows not to short-circuit.
     REQUIRES_EDIT_FINALIZE: bool = False
 
+    async def create_handoff_thread(
+        self,
+        parent_chat_id: str,
+        name: str,
+    ) -> Optional[str]:
+        """Create a fresh thread under ``parent_chat_id`` for a session handoff.
+
+        Used by the gateway's handoff watcher when transferring a CLI
+        session to a thread-capable platform — the new thread isolates the
+        handed-off conversation from any pre-existing chat in the home
+        channel and gives users a clean per-handoff scrollback.
+
+        Returns the new thread/topic id (as a string) on success, or
+        ``None`` if the platform doesn't support threading or the
+        attempt failed (permissions, topics-mode off, etc.). When ``None``
+        is returned the watcher falls back to using ``parent_chat_id``
+        directly.
+
+        Default implementation returns ``None`` — adapters that support
+        threads override this. See:
+          - Telegram: forum topics in groups, DM topics with bot API 9.4+
+          - Discord:  text-channel threads (1440-min auto-archive)
+          - Slack:    seed-message thread anchoring
+        """
+        return None
+
+
     async def edit_message(
         self,
         chat_id: str,
@@ -2704,7 +2793,7 @@ class BasePlatformAdapter(ABC):
                 # 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"):
+                if cmd in {"stop", "new", "reset"}:
                     try:
                         await self._dispatch_active_session_command(event, session_key, cmd)
                     except Exception as e:
@@ -2950,6 +3039,18 @@ class BasePlatformAdapter(ABC):
                 if text_content:
                     logger.info("[%s] Sending response (%d chars) to %s", self.name, len(text_content), event.source.chat_id)
                     _reply_anchor = _reply_anchor_for_event(event)
+                    # Mark final response messages for notification delivery.
+                    # Platform adapters that support per-message notification
+                    # control (e.g. Telegram's disable_notification) use this
+                    # flag to override silent-mode and ensure the final
+                    # response triggers a push notification.
+                    # Clone to avoid mutating the metadata shared with the
+                    # typing-indicator task (which must remain unmarked).
+                    if _thread_metadata is not None:
+                        _thread_metadata = dict(_thread_metadata)
+                        _thread_metadata["notify"] = True
+                    else:
+                        _thread_metadata = {"notify": True}
                     result = await self._send_with_retry(
                         chat_id=event.source.chat_id,
                         content=text_content,

gateway/platforms/bluebubbles.py

@@ -223,7 +223,7 @@ class BlueBubblesAdapter(BasePlatformAdapter):
     def _webhook_url(self) -> str:
         """Compute the external webhook URL for BlueBubbles registration."""
         host = self.webhook_host
-        if host in ("0.0.0.0", "127.0.0.1", "localhost", "::"):
+        if host in {"0.0.0.0", "127.0.0.1", "localhost", "::"}:
             host = "localhost"
         return f"http://{host}:{self.webhook_port}{self.webhook_path}"
 

gateway/platforms/dingtalk.py

@@ -353,9 +353,9 @@ class DingTalkAdapter(BasePlatformAdapter):
         configured = self.config.extra.get("require_mention")
         if configured is not None:
             if isinstance(configured, str):
-                return configured.lower() in ("true", "1", "yes", "on")
+                return configured.lower() in {"true", "1", "yes", "on"}
             return bool(configured)
-        return os.getenv("DINGTALK_REQUIRE_MENTION", "false").lower() in ("true", "1", "yes", "on")
+        return os.getenv("DINGTALK_REQUIRE_MENTION", "false").lower() in {"true", "1", "yes", "on"}
 
     def _dingtalk_free_response_chats(self) -> Set[str]:
         raw = self.config.extra.get("free_response_chats")
@@ -886,6 +886,67 @@ class DingTalkAdapter(BasePlatformAdapter):
         """DingTalk does not support typing indicators."""
         pass
 
+    async def send_image(
+        self,
+        chat_id: str,
+        image_url: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Send an image via DingTalk markdown.
+
+        DingTalk's session webhook only supports text/markdown payloads, not
+        native image/file attachments. For remote image URLs, render the image
+        inline with markdown so the user still sees the image. Local files need
+        OpenAPI media upload and are handled separately.
+        """
+        image_block = f"![image]({image_url})"
+        content = f"{caption}\n\n{image_block}" if caption else image_block
+        return await self.send(
+            chat_id=chat_id,
+            content=content,
+            reply_to=reply_to,
+            metadata=metadata,
+        )
+
+    async def send_image_file(
+        self,
+        chat_id: str,
+        image_path: str,
+        caption: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        **kwargs,
+    ) -> SendResult:
+        """DingTalk webhook replies cannot send local image files directly."""
+        return SendResult(
+            success=False,
+            error=(
+                "DingTalk session webhook replies do not support local image uploads. "
+                "Only markdown/text replies are supported without OpenAPI media upload."
+            ),
+        )
+
+    async def send_document(
+        self,
+        chat_id: str,
+        file_path: str,
+        caption: Optional[str] = None,
+        file_name: Optional[str] = None,
+        reply_to: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        **kwargs,
+    ) -> SendResult:
+        """DingTalk webhook replies cannot send local file attachments directly."""
+        return SendResult(
+            success=False,
+            error=(
+                "DingTalk session webhook replies do not support local file attachments. "
+                "Only markdown/text replies are supported without OpenAPI message send."
+            ),
+        )
+
     async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
         """Return basic info about a DingTalk conversation."""
         return {

gateway/platforms/discord.py

@@ -86,8 +86,32 @@ def _clean_discord_id(entry: str) -> str:
 
 
 def check_discord_requirements() -> bool:
-    """Check if Discord dependencies are available."""
-    return DISCORD_AVAILABLE
+    """Check if Discord dependencies are available.
+
+    Lazy-installs discord.py via ``tools.lazy_deps.ensure("platform.discord")``
+    on first call if not present. After successful install, re-binds module
+    globals so ``DISCORD_AVAILABLE`` becomes True.
+    """
+    global DISCORD_AVAILABLE, discord, DiscordMessage, Intents, commands
+    if DISCORD_AVAILABLE:
+        return True
+    try:
+        from tools.lazy_deps import ensure as _lazy_ensure
+        _lazy_ensure("platform.discord", prompt=False)
+    except Exception:
+        return False
+    try:
+        import discord as _discord
+        from discord import Message as _DM, Intents as _Intents
+        from discord.ext import commands as _commands
+    except ImportError:
+        return False
+    discord = _discord
+    DiscordMessage = _DM
+    Intents = _Intents
+    commands = _commands
+    DISCORD_AVAILABLE = True
+    return True
 
 
 def _build_allowed_mentions():
@@ -115,7 +139,7 @@ def _build_allowed_mentions():
         raw = os.getenv(name, "").strip().lower()
         if not raw:
             return default
-        return raw in ("true", "1", "yes", "on")
+        return raw in {"true", "1", "yes", "on"}
 
     return discord.AllowedMentions(
         everyone=_b("DISCORD_ALLOW_MENTION_EVERYONE", False),
@@ -708,7 +732,7 @@ class DiscordAdapter(BasePlatformAdapter):
 
                 # Ignore Discord system messages (thread renames, pins, member joins, etc.)
                 # Allow both default and reply types — replies have a distinct MessageType.
-                if message.type not in (discord.MessageType.default, discord.MessageType.reply):
+                if message.type not in {discord.MessageType.default, discord.MessageType.reply}:
                     return
 
                 # Bot message filtering (DISCORD_ALLOW_BOTS):
@@ -769,7 +793,7 @@ class DiscordAdapter(BasePlatformAdapter):
                     # answer regardless of who is mentioned.
                     _ignore_no_mention = os.getenv(
                         "DISCORD_IGNORE_NO_MENTION", "true"
-                    ).lower() in ("true", "1", "yes")
+                    ).lower() in {"true", "1", "yes"}
                     if _ignore_no_mention and not _self_mentioned and not _other_bots_mentioned:
                         _channel_id = str(message.channel.id)
                         _parent_id = None
@@ -1317,7 +1341,7 @@ class DiscordAdapter(BasePlatformAdapter):
 
     def _reactions_enabled(self) -> bool:
         """Check if message reactions are enabled via config/env."""
-        return os.getenv("DISCORD_REACTIONS", "true").lower() not in ("false", "0", "no")
+        return os.getenv("DISCORD_REACTIONS", "true").lower() not in {"false", "0", "no"}
 
     async def on_processing_start(self, event: MessageEvent) -> None:
         """Add an in-progress reaction for normal Discord message events."""
@@ -2697,6 +2721,8 @@ class DiscordAdapter(BasePlatformAdapter):
                     await asyncio.sleep(8)
             except asyncio.CancelledError:
                 pass
+            finally:
+                self._typing_tasks.pop(chat_id, None)
 
         self._typing_tasks[chat_id] = asyncio.create_task(_typing_loop())
 
@@ -3135,9 +3161,9 @@ class DiscordAdapter(BasePlatformAdapter):
         # UX so users don't see commands they can't invoke. Off by default
         # to preserve the slash UX for deployments that intentionally allow
         # everyone in the guild.
-        if os.getenv("DISCORD_HIDE_SLASH_COMMANDS", "false").strip().lower() in (
+        if os.getenv("DISCORD_HIDE_SLASH_COMMANDS", "false").strip().lower() in {
             "true", "1", "yes", "on",
-        ):
+        }:
             self._apply_owner_only_visibility(tree)
 
     def _apply_owner_only_visibility(self, tree) -> None:
@@ -3524,9 +3550,9 @@ class DiscordAdapter(BasePlatformAdapter):
         configured = self.config.extra.get("require_mention")
         if configured is not None:
             if isinstance(configured, str):
-                return configured.lower() not in ("false", "0", "no", "off")
+                return configured.lower() not in {"false", "0", "no", "off"}
             return bool(configured)
-        return os.getenv("DISCORD_REQUIRE_MENTION", "true").lower() not in ("false", "0", "no", "off")
+        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.
@@ -3689,6 +3715,84 @@ class DiscordAdapter(BasePlatformAdapter):
                 )
                 return None
 
+    async def create_handoff_thread(
+        self,
+        parent_chat_id: str,
+        name: str,
+    ) -> Optional[str]:
+        """Create a Discord thread under a text channel for a handoff.
+
+        Falls back to a seed-message + ``message.create_thread`` path if
+        ``parent.create_thread`` is rejected (some channel types or
+        permission setups). Returns the new thread id as a string, or
+        ``None`` on failure or when the parent isn't a text channel
+        (DMs, voice channels, threads themselves can't host threads).
+        """
+        if not self._client or not DISCORD_AVAILABLE:
+            return None
+
+        try:
+            parent_id = int(parent_chat_id)
+        except (TypeError, ValueError):
+            return None
+
+        try:
+            parent = self._client.get_channel(parent_id)
+            if parent is None:
+                parent = await self._client.fetch_channel(parent_id)
+        except Exception as exc:
+            logger.warning(
+                "[%s] Handoff thread: cannot resolve parent %s: %s",
+                self.name, parent_chat_id, exc,
+            )
+            return None
+
+        # DMs, voice channels, and existing threads can't host child threads.
+        if isinstance(parent, getattr(discord, "DMChannel", ())):
+            logger.info(
+                "[%s] Handoff thread: parent %s is a DM; threads not supported here",
+                self.name, parent_chat_id,
+            )
+            return None
+
+        thread_name = (name or "handoff").strip()[:80] or "handoff"
+        reason = "Hermes session handoff"
+
+        # First try: create a thread directly on the channel.
+        try:
+            create = getattr(parent, "create_thread", None)
+            if create is not None:
+                thread = await create(
+                    name=thread_name,
+                    auto_archive_duration=1440,
+                    reason=reason,
+                )
+                return str(thread.id)
+        except Exception as direct_error:
+            logger.debug(
+                "[%s] Handoff thread: direct create failed (%s); trying seed-message fallback",
+                self.name, direct_error,
+            )
+
+        # Fallback: post a seed message and create the thread from it.
+        try:
+            send = getattr(parent, "send", None)
+            if send is None:
+                return None
+            seed_msg = await send(f"\U0001f9f5 Hermes handoff: **{thread_name}**")
+            thread = await seed_msg.create_thread(
+                name=thread_name,
+                auto_archive_duration=1440,
+                reason=reason,
+            )
+            return str(thread.id)
+        except Exception as fallback_error:
+            logger.warning(
+                "[%s] Handoff thread: both create paths failed for parent %s: %s",
+                self.name, parent_chat_id, fallback_error,
+            )
+            return None
+
     async def send_exec_approval(
         self, chat_id: str, command: str, session_key: str,
         description: str = "dangerous command",
@@ -4120,7 +4224,7 @@ class DiscordAdapter(BasePlatformAdapter):
             no_thread_channels_raw = os.getenv("DISCORD_NO_THREAD_CHANNELS", "")
             no_thread_channels = {ch.strip() for ch in no_thread_channels_raw.split(",") if ch.strip()}
             skip_thread = bool(channel_ids & no_thread_channels)
-            auto_thread = os.getenv("DISCORD_AUTO_THREAD", "true").lower() in ("true", "1", "yes")
+            auto_thread = os.getenv("DISCORD_AUTO_THREAD", "true").lower() in {"true", "1", "yes"}
             is_reply_message = getattr(message, "type", None) == discord.MessageType.reply
             if auto_thread and not skip_thread and not is_voice_linked_channel and not is_reply_message:
                 thread = await self._auto_create_thread(message)
@@ -4202,7 +4306,7 @@ class DiscordAdapter(BasePlatformAdapter):
                 try:
                     # Determine extension from content type (image/png -> .png)
                     ext = "." + content_type.split("/")[-1].split(";")[0]
-                    if ext not in (".jpg", ".jpeg", ".png", ".gif", ".webp"):
+                    if ext not in {".jpg", ".jpeg", ".png", ".gif", ".webp"}:
                         ext = ".jpg"
                     cached_path = await self._cache_discord_image(att, ext)
                     media_urls.append(cached_path)
@@ -4216,7 +4320,7 @@ class DiscordAdapter(BasePlatformAdapter):
             elif content_type.startswith("audio/"):
                 try:
                     ext = "." + content_type.split("/")[-1].split(";")[0]
-                    if ext not in (".ogg", ".mp3", ".wav", ".webm", ".m4a"):
+                    if ext not in {".ogg", ".mp3", ".wav", ".webm", ".m4a"}:
                         ext = ".ogg"
                     cached_path = await self._cache_discord_audio(att, ext)
                     media_urls.append(cached_path)
@@ -4259,7 +4363,7 @@ class DiscordAdapter(BasePlatformAdapter):
                             logger.info("[Discord] Cached user document: %s", cached_path)
                             # Inject text content for plain-text documents (capped at 100 KB)
                             MAX_TEXT_INJECT_BYTES = 100 * 1024
-                            if ext in (".md", ".txt", ".log") and len(raw_bytes) <= MAX_TEXT_INJECT_BYTES:
+                            if ext in {".md", ".txt", ".log"} and len(raw_bytes) <= MAX_TEXT_INJECT_BYTES:
                                 try:
                                     text_content = raw_bytes.decode("utf-8")
                                     display_name = att.filename or f"document{ext}"

gateway/platforms/email.py

@@ -54,7 +54,7 @@ _NOREPLY_PATTERNS = (
 # RFC headers that indicate bulk/automated mail
 _AUTOMATED_HEADERS = {
     "Auto-Submitted": lambda v: v.lower() != "no",
-    "Precedence": lambda v: v.lower() in ("bulk", "list", "junk"),
+    "Precedence": lambda v: v.lower() in {"bulk", "list", "junk"},
     "X-Auto-Response-Suppress": lambda v: bool(v),
     "List-Unsubscribe": lambda v: bool(v),
 }
@@ -65,6 +65,29 @@ MAX_MESSAGE_LENGTH = 50_000
 # Supported image extensions for inline detection
 _IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".gif", ".webp"}
 
+def _send_imap_id(imap: "imaplib.IMAP4") -> None:
+    """Send RFC 2971 IMAP ID command identifying this client.
+
+    Required by 163/NetEase mailbox after LOGIN: without it, every UID
+    SEARCH/FETCH returns ``BYE Unsafe Login`` and disconnects.  Other
+    IMAP servers either honor it silently or reject the unknown command;
+    we swallow failures so non-supporting servers keep working.
+    """
+    try:
+        try:
+            from hermes_cli import __version__ as _hermes_version
+        except Exception:  # noqa: BLE001 — keep ID best-effort if import fails
+            _hermes_version = "0"
+        imap.xatom(
+            "ID",
+            f'("name" "hermes-agent" "version" "{_hermes_version}" '
+            '"vendor" "NousResearch" '
+            '"support-email" "noreply@nousresearch.com")',
+        )
+    except Exception as e:  # noqa: BLE001 — best-effort, never fatal
+        logger.debug("[Email] IMAP ID command not accepted: %s", e)
+
+
 def _is_automated_sender(address: str, headers: dict) -> bool:
     """Return True if this email is from an automated/noreply source."""
     addr = address.lower()
@@ -180,7 +203,7 @@ def _extract_attachments(
             continue
         # Skip text/plain and text/html body parts
         content_type = part.get_content_type()
-        if content_type in ("text/plain", "text/html") and "attachment" not in disposition:
+        if content_type in {"text/plain", "text/html"} and "attachment" not in disposition:
             continue
 
         filename = part.get_filename()
@@ -276,6 +299,7 @@ class EmailAdapter(BasePlatformAdapter):
             # Test IMAP connection
             imap = imaplib.IMAP4_SSL(self._imap_host, self._imap_port, timeout=30)
             imap.login(self._address, self._password)
+            _send_imap_id(imap)
             # Mark all existing messages as seen so we only process new ones
             imap.select("INBOX")
             status, data = imap.uid("search", None, "ALL")
@@ -344,6 +368,7 @@ class EmailAdapter(BasePlatformAdapter):
             imap = imaplib.IMAP4_SSL(self._imap_host, self._imap_port, timeout=30)
             try:
                 imap.login(self._address, self._password)
+                _send_imap_id(imap)
                 imap.select("INBOX")
 
                 status, data = imap.uid("search", None, "UNSEEN")

gateway/platforms/feishu.py

@@ -428,7 +428,7 @@ RejectReason = Literal[
 
 def _is_bot_sender(sender: Any) -> bool:
     # receive_v1 docs say {user, bot}; accept "app" defensively.
-    return getattr(sender, "sender_type", "") in ("bot", "app")
+    return getattr(sender, "sender_type", "") in {"bot", "app"}
 
 
 def _sender_identity(sender: Any) -> frozenset:
@@ -1428,8 +1428,8 @@ class FeishuAdapter(BasePlatformAdapter):
                     per_chat_require_mention = _to_boolean(rule_cfg.get("require_mention"))
                 group_rules[str(chat_id)] = FeishuGroupRule(
                     policy=str(rule_cfg.get("policy", "open")).strip().lower(),
-                    allowlist=set(str(u).strip() for u in rule_cfg.get("allowlist", []) if str(u).strip()),
-                    blacklist=set(str(u).strip() for u in rule_cfg.get("blacklist", []) if str(u).strip()),
+                    allowlist={str(u).strip() for u in rule_cfg.get("allowlist", []) if str(u).strip()},
+                    blacklist={str(u).strip() for u in rule_cfg.get("blacklist", []) if str(u).strip()},
                     require_mention=per_chat_require_mention,
                 )
 
@@ -1443,7 +1443,7 @@ class FeishuAdapter(BasePlatformAdapter):
         # Env-only so adapter and gateway auth bypass share one source; yaml
         # feishu.allow_bots is bridged to this env var at config load.
         allow_bots = os.getenv("FEISHU_ALLOW_BOTS", "none").strip().lower()
-        if allow_bots not in ("none", "mentions", "all"):
+        if allow_bots not in {"none", "mentions", "all"}:
             logger.warning(
                 "[Feishu] Unknown allow_bots=%r, falling back to 'none'. Valid: none, mentions, all.",
                 allow_bots,
@@ -2752,7 +2752,7 @@ class FeishuAdapter(BasePlatformAdapter):
     # =========================================================================
 
     def _reactions_enabled(self) -> bool:
-        return os.getenv("FEISHU_REACTIONS", "true").strip().lower() not in ("false", "0", "no")
+        return os.getenv("FEISHU_REACTIONS", "true").strip().lower() not in {"false", "0", "no"}
 
     async def _add_reaction(self, message_id: str, emoji_type: str) -> Optional[str]:
         """Return the reaction_id on success, else None. The id is needed later for deletion."""
@@ -3219,7 +3219,7 @@ class FeishuAdapter(BasePlatformAdapter):
             self._on_bot_added_to_chat(data)
         elif event_type == "im.chat.member.bot.deleted_v1":
             self._on_bot_removed_from_chat(data)
-        elif event_type in ("im.message.reaction.created_v1", "im.message.reaction.deleted_v1"):
+        elif event_type in {"im.message.reaction.created_v1", "im.message.reaction.deleted_v1"}:
             self._on_reaction_event(event_type, data)
         elif event_type == "card.action.trigger":
             self._on_card_action_trigger(data)
@@ -4273,21 +4273,31 @@ class FeishuAdapter(BasePlatformAdapter):
             request = self._build_reply_message_request(effective_reply_to, body)
             return await asyncio.to_thread(self._client.im.v1.message.reply, request)
 
-        body = self._build_create_message_body(
-            receive_id=chat_id,
-            msg_type=msg_type,
-            content=payload,
-            uuid_value=str(uuid.uuid4()),
-        )
-        # Detect whether chat_id is a user open_id (DM) or a chat_id (group).
-        # Feishu API expects receive_id_type="open_id" for user DMs (ou_ prefix)
-        # and receive_id_type="chat_id" for group chats (oc_ prefix, which IS
-        # the chat_id format — see https://open.feishu.cn/document/).
-        if chat_id.startswith("ou_"):
-            receive_id_type = "open_id"
+        # For topic/thread messages that fell back from reply→create, use
+        # thread_id as receive_id so the message lands in the topic instead of
+        # the main chat.
+        _thread_id = (metadata or {}).get("thread_id")
+        if _thread_id:
+            body = self._build_create_message_body(
+                receive_id=_thread_id,
+                msg_type=msg_type,
+                content=payload,
+                uuid_value=str(uuid.uuid4()),
+            )
+            request = self._build_create_message_request("thread_id", body)
         else:
-            receive_id_type = "chat_id"
-        request = self._build_create_message_request(receive_id_type, body)
+            body = self._build_create_message_body(
+                receive_id=chat_id,
+                msg_type=msg_type,
+                content=payload,
+                uuid_value=str(uuid.uuid4()),
+            )
+            # Detect whether chat_id is a user open_id (DM) or a chat_id (group).
+            if chat_id.startswith("ou_"):
+                receive_id_type = "open_id"
+            else:
+                receive_id_type = "chat_id"
+            request = self._build_create_message_request(receive_id_type, body)
         return await asyncio.to_thread(self._client.im.v1.message.create, request)
 
     @staticmethod
@@ -4805,7 +4815,7 @@ def _poll_registration(
 
         # Terminal errors
         error = res.get("error", "")
-        if error in ("access_denied", "expired_token"):
+        if error in {"access_denied", "expired_token"}:
             if poll_count > 0:
                 print()
             logger.warning("[Feishu onboard] Registration %s", error)

gateway/platforms/feishu_comment.py

@@ -690,7 +690,7 @@ def _extract_docs_links(replies: List[Dict[str, Any]]) -> List[Dict[str, str]]:
             except (json.JSONDecodeError, TypeError):
                 continue
         for elem in content.get("elements", []):
-            if elem.get("type") not in ("docs_link", "link"):
+            if elem.get("type") not in {"docs_link", "link"}:
                 continue
             link_data = elem.get("docs_link") or elem.get("link") or {}
             url = link_data.get("url", "")
@@ -1031,7 +1031,7 @@ def _save_session_history(key: str, messages: List[Dict[str, Any]]) -> None:
     # Only keep user/assistant messages (strip system messages and tool internals)
     cleaned = [
         m for m in messages
-        if m.get("role") in ("user", "assistant") and m.get("content")
+        if m.get("role") in {"user", "assistant"} and m.get("content")
     ]
     # Keep last N
     if len(cleaned) > _SESSION_MAX_MESSAGES:
@@ -1170,7 +1170,7 @@ async def handle_drive_comment_event(
     rule = resolve_rule(comments_cfg, file_type, file_token)
 
     # If no exact match and config has wiki keys, try reverse-lookup
-    if rule.match_source in ("wildcard", "top") and has_wiki_keys(comments_cfg):
+    if rule.match_source in {"wildcard", "top"} and has_wiki_keys(comments_cfg):
         wiki_token = await _reverse_lookup_wiki_token(client, file_type, file_token)
         if wiki_token:
             rule = resolve_rule(comments_cfg, file_type, file_token, wiki_token=wiki_token)

gateway/platforms/feishu_comment_rules.py

@@ -228,7 +228,7 @@ def _load_pairing_approved() -> set:
     if isinstance(approved, dict):
         return set(approved.keys())
     if isinstance(approved, list):
-        return set(str(u) for u in approved if u)
+        return {str(u) for u in approved if u}
     return set()
 
 

gateway/platforms/helpers.py

@@ -246,7 +246,7 @@ class ThreadParticipationTracker:
         thread_list = list(self._threads)
         if len(thread_list) > self._max_tracked:
             thread_list = thread_list[-self._max_tracked:]
-            self._threads = {thread_id: None for thread_id in thread_list}
+            self._threads = dict.fromkeys(thread_list)
         atomic_json_write(path, thread_list, indent=None)
 
     def mark(self, thread_id: str) -> None:

gateway/platforms/homeassistant.py

@@ -256,7 +256,7 @@ class HomeAssistantAdapter(BasePlatformAdapter):
                         await self._handle_ha_event(data.get("event", {}))
                 except json.JSONDecodeError:
                     logger.debug("Invalid JSON from HA WS: %s", ws_msg.data[:200])
-            elif ws_msg.type in (aiohttp.WSMsgType.CLOSED, aiohttp.WSMsgType.ERROR):
+            elif ws_msg.type in {aiohttp.WSMsgType.CLOSED, aiohttp.WSMsgType.ERROR}:
                 break
 
     async def _handle_ha_event(self, event: Dict[str, Any]) -> None:
@@ -361,7 +361,7 @@ class HomeAssistantAdapter(BasePlatformAdapter):
                 f"(was {'triggered' if old_val == 'on' else 'cleared'})"
             )
 
-        if domain in ("light", "switch", "fan"):
+        if domain in {"light", "switch", "fan"}:
             return (
                 f"[Home Assistant] {friendly_name}: turned "
                 f"{'on' if new_val == 'on' else 'off'}"

gateway/platforms/matrix.py

@@ -245,11 +245,11 @@ def check_matrix_requirements() -> bool:
 
     # If encryption is requested, verify E2EE deps are available at startup
     # rather than silently degrading to plaintext-only at connect time.
-    encryption_requested = os.getenv("MATRIX_ENCRYPTION", "").lower() in (
+    encryption_requested = os.getenv("MATRIX_ENCRYPTION", "").lower() in {
         "true",
         "1",
         "yes",
-    )
+    }
     if encryption_requested and not _check_e2ee_deps():
         logger.error(
             "Matrix: MATRIX_ENCRYPTION=true but E2EE dependencies are missing. %s. "
@@ -312,7 +312,7 @@ class MatrixAdapter(BasePlatformAdapter):
         )
         self._encryption: bool = config.extra.get(
             "encryption",
-            os.getenv("MATRIX_ENCRYPTION", "").lower() in ("true", "1", "yes"),
+            os.getenv("MATRIX_ENCRYPTION", "").lower() in {"true", "1", "yes"},
         )
         self._device_id: str = config.extra.get("device_id", "") or os.getenv(
             "MATRIX_DEVICE_ID", ""
@@ -343,7 +343,7 @@ class MatrixAdapter(BasePlatformAdapter):
         # Mention/thread gating — parsed once from env vars.
         self._require_mention: bool = os.getenv(
             "MATRIX_REQUIRE_MENTION", "true"
-        ).lower() not in ("false", "0", "no")
+        ).lower() not in {"false", "0", "no"}
         free_rooms_raw = config.extra.get("free_response_rooms")
         if free_rooms_raw is None:
             free_rooms_raw = os.getenv("MATRIX_FREE_RESPONSE_ROOMS", "")
@@ -367,22 +367,22 @@ class MatrixAdapter(BasePlatformAdapter):
             self._allowed_rooms: Set[str] = {
                 r.strip() for r in str(allowed_rooms_raw).split(",") if r.strip()
             }
-        self._auto_thread: bool = os.getenv("MATRIX_AUTO_THREAD", "true").lower() in (
+        self._auto_thread: bool = os.getenv("MATRIX_AUTO_THREAD", "true").lower() in {
             "true",
             "1",
             "yes",
-        )
+        }
         self._dm_auto_thread: bool = os.getenv(
             "MATRIX_DM_AUTO_THREAD", "false"
-        ).lower() in ("true", "1", "yes")
+        ).lower() in {"true", "1", "yes"}
         self._dm_mention_threads: bool = os.getenv(
             "MATRIX_DM_MENTION_THREADS", "false"
-        ).lower() in ("true", "1", "yes")
+        ).lower() in {"true", "1", "yes"}
 
         # Reactions: configurable via MATRIX_REACTIONS (default: true).
         self._reactions_enabled: bool = os.getenv(
             "MATRIX_REACTIONS", "true"
-        ).lower() not in ("false", "0", "no")
+        ).lower() not in {"false", "0", "no"}
         self._pending_reactions: dict[tuple[str, str], str] = {}
         # Delay before redacting reactions so Matrix homeservers have time to
         # deliver the final message event without tripping "missing event"
@@ -1771,9 +1771,9 @@ class MatrixAdapter(BasePlatformAdapter):
 
         # Cache media locally when downstream tools need a real file path.
         cached_path = None
-        should_cache_locally = msg_type in (
+        should_cache_locally = msg_type in {
             MessageType.PHOTO, MessageType.AUDIO, MessageType.VIDEO, MessageType.DOCUMENT,
-        ) or is_voice_message or is_encrypted_media
+        } or is_voice_message or is_encrypted_media
         if should_cache_locally and url:
             try:
                 file_bytes = await self._client.download_media(ContentURI(url))
@@ -1834,7 +1834,7 @@ class MatrixAdapter(BasePlatformAdapter):
                             ext = ext_map.get(media_type, ".jpg")
                             cached_path = cache_image_from_bytes(file_bytes, ext=ext)
                             logger.info("[Matrix] Cached user image at %s", cached_path)
-                        elif msg_type in (MessageType.AUDIO, MessageType.VOICE):
+                        elif msg_type in {MessageType.AUDIO, MessageType.VOICE}:
                             ext = (
                                 Path(
                                     body
@@ -2602,7 +2602,7 @@ class MatrixAdapter(BasePlatformAdapter):
         """Sanitize a URL for use in an href attribute."""
         stripped = url.strip()
         scheme = stripped.split(":", 1)[0].lower().strip() if ":" in stripped else ""
-        if scheme in ("javascript", "data", "vbscript"):
+        if scheme in {"javascript", "data", "vbscript"}:
             return ""
         return stripped.replace('"', """)
 

gateway/platforms/mattermost.py

@@ -611,7 +611,7 @@ class MattermostAdapter(BasePlatformAdapter):
                 # succeed on retry — stop reconnecting instead of looping forever.
                 import aiohttp
                 err_str = str(exc).lower()
-                if isinstance(exc, aiohttp.WSServerHandshakeError) and exc.status in (401, 403):
+                if isinstance(exc, aiohttp.WSServerHandshakeError) and exc.status in {401, 403}:
                     logger.error("Mattermost WS auth failed (HTTP %d) — stopping reconnect", exc.status)
                     return
                 if "401" in err_str or "403" in err_str or "unauthorized" in err_str:
@@ -649,21 +649,21 @@ class MattermostAdapter(BasePlatformAdapter):
             if self._closing:
                 return
 
-            if raw_msg.type in (
+            if raw_msg.type in {
                 raw_msg.type.TEXT,
                 raw_msg.type.BINARY,
-            ):
+            }:
                 try:
                     event = json.loads(raw_msg.data)
                 except (json.JSONDecodeError, TypeError):
                     continue
                 await self._handle_ws_event(event)
-            elif raw_msg.type in (
+            elif raw_msg.type in {
                 raw_msg.type.ERROR,
                 raw_msg.type.CLOSE,
                 raw_msg.type.CLOSING,
                 raw_msg.type.CLOSED,
-            ):
+            }:
                 logger.info("Mattermost: WebSocket closed (%s)", raw_msg.type)
                 break
 
@@ -732,7 +732,7 @@ class MattermostAdapter(BasePlatformAdapter):
 
             require_mention = os.getenv(
                 "MATTERMOST_REQUIRE_MENTION", "true"
-            ).lower() not in ("false", "0", "no")
+            ).lower() not in {"false", "0", "no"}
 
             free_channels_raw = os.getenv("MATTERMOST_FREE_RESPONSE_CHANNELS", "")
             free_channels = {ch.strip() for ch in free_channels_raw.split(",") if ch.strip()}

gateway/platforms/qqbot/adapter.py

@@ -513,7 +513,7 @@ class QQAdapter(BasePlatformAdapter):
                 self._fail_pending("Connection closed")
 
                 # Stop reconnecting for fatal codes
-                if code in (4914, 4915):
+                if code in {4914, 4915}:
                     desc = "offline/sandbox-only" if code == 4914 else "banned"
                     logger.error(
                         "[%s] Bot is %s. Check QQ Open Platform.", self._log_tag, desc
@@ -550,7 +550,7 @@ class QQAdapter(BasePlatformAdapter):
                     self._token_expires_at = 0.0
 
                 # Session invalid → clear session, will re-identify on next Hello
-                if code in (
+                if code in {
                         4006,
                         4007,
                         4009,
@@ -568,7 +568,7 @@ class QQAdapter(BasePlatformAdapter):
                         4911,
                         4912,
                         4913,
-                ):
+                }:
                     logger.info(
                         "[%s] Session error (%d), clearing session for re-identify",
                         self._log_tag,
@@ -637,12 +637,12 @@ class QQAdapter(BasePlatformAdapter):
                 payload = self._parse_json(msg.data)
                 if payload:
                     self._dispatch_payload(payload)
-            elif msg.type in (aiohttp.WSMsgType.PING,):
+            elif msg.type in {aiohttp.WSMsgType.PING,}:
                 # aiohttp auto-replies with PONG
                 pass
             elif msg.type == aiohttp.WSMsgType.CLOSE:
                 raise QQCloseError(msg.data, msg.extra)
-            elif msg.type in (aiohttp.WSMsgType.CLOSED, aiohttp.WSMsgType.ERROR):
+            elif msg.type in {aiohttp.WSMsgType.CLOSED, aiohttp.WSMsgType.ERROR}:
                 raise RuntimeError("WebSocket closed")
 
     async def _heartbeat_loop(self) -> None:
@@ -783,13 +783,13 @@ class QQAdapter(BasePlatformAdapter):
                 self._handle_ready(d)
             elif t == "RESUMED":
                 logger.info("[%s] Session resumed", self._log_tag)
-            elif t in (
+            elif t in {
                     "C2C_MESSAGE_CREATE",
                     "GROUP_AT_MESSAGE_CREATE",
                     "DIRECT_MESSAGE_CREATE",
                     "GUILD_MESSAGE_CREATE",
                     "GUILD_AT_MESSAGE_CREATE",
-            ):
+            }:
                 asyncio.create_task(self._on_message(t, d))
             elif t == "INTERACTION_CREATE":
                 self._create_task(self._on_interaction(d))
@@ -859,9 +859,9 @@ class QQAdapter(BasePlatformAdapter):
         # Route by event type
         if event_type == "C2C_MESSAGE_CREATE":
             await self._handle_c2c_message(d, msg_id, content, author, timestamp)
-        elif event_type in ("GROUP_AT_MESSAGE_CREATE",):
+        elif event_type in {"GROUP_AT_MESSAGE_CREATE",}:
             await self._handle_group_message(d, msg_id, content, author, timestamp)
-        elif event_type in ("GUILD_MESSAGE_CREATE", "GUILD_AT_MESSAGE_CREATE"):
+        elif event_type in {"GUILD_MESSAGE_CREATE", "GUILD_AT_MESSAGE_CREATE"}:
             await self._handle_guild_message(d, msg_id, content, author, timestamp)
         elif event_type == "DIRECT_MESSAGE_CREATE":
             await self._handle_dm_message(d, msg_id, content, author, timestamp)
@@ -1864,7 +1864,7 @@ class QQAdapter(BasePlatformAdapter):
             return ".wav"
         if data[:4] == b"fLaC":
             return ".flac"
-        if data[:2] in (b"\xff\xfb", b"\xff\xf3", b"\xff\xf2"):
+        if data[:2] in {b"\xff\xfb", b"\xff\xf3", b"\xff\xf2"}:
             return ".mp3"
         if data[:4] == b"\x30\x26\xb2\x75" or data[:4] == b"\x4f\x67\x67\x53":
             return ".ogg"
@@ -2033,7 +2033,7 @@ class QQAdapter(BasePlatformAdapter):
                         "base_url": base_url,
                         "api_key": api_key,
                         "model": model
-                                 or ("glm-asr" if provider in ("zai", "glm") else "whisper-1"),
+                                 or ("glm-asr" if provider in {"zai", "glm"} else "whisper-1"),
                     }
 
         # 2. QQ-specific env vars (set by `hermes setup gateway` / `hermes gateway`)
@@ -2115,7 +2115,7 @@ class QQAdapter(BasePlatformAdapter):
             if urlparse(source_url).path
             else ""
         )
-        if not ext or ext not in (
+        if not ext or ext not in {
                 ".silk",
                 ".amr",
                 ".mp3",
@@ -2124,7 +2124,7 @@ class QQAdapter(BasePlatformAdapter):
                 ".m4a",
                 ".aac",
                 ".flac",
-        ):
+        }:
             ext = self._guess_ext_from_data(audio_data)
 
         with tempfile.NamedTemporaryFile(suffix=ext, delete=False) as tmp_src:
@@ -2870,7 +2870,7 @@ class QQAdapter(BasePlatformAdapter):
             raise ValueError("Media source is required")
 
         parsed = urlparse(source)
-        if parsed.scheme in ("http", "https"):
+        if parsed.scheme in {"http", "https"}:
             # For URLs, pass through directly to the upload API
             content_type = mimetypes.guess_type(source)[0] or "application/octet-stream"
             resolved_name = file_name or Path(parsed.path).name or "media"
@@ -2966,7 +2966,7 @@ class QQAdapter(BasePlatformAdapter):
         chat_type = self._guess_chat_type(chat_id)
         return {
             "name": chat_id,
-            "type": "group" if chat_type in ("group", "guild") else "dm",
+            "type": "group" if chat_type in {"group", "guild"} else "dm",
         }
 
     # ------------------------------------------------------------------
@@ -2975,7 +2975,7 @@ class QQAdapter(BasePlatformAdapter):
 
     @staticmethod
     def _is_url(source: str) -> bool:
-        return urlparse(str(source)).scheme in ("http", "https")
+        return urlparse(str(source)).scheme in {"http", "https"}
 
     def _guess_chat_type(self, chat_id: str) -> str:
         """Determine chat type from stored inbound metadata, fallback to 'c2c'."""

gateway/platforms/qqbot/chunked_upload.py

@@ -239,7 +239,7 @@ class ChunkedUploader:
         :raises UploadFileTooLargeError: When the file exceeds the platform limit.
         :raises RuntimeError: On other API or I/O failures.
         """
-        if chat_type not in ("c2c", "group"):
+        if chat_type not in {"c2c", "group"}:
             raise ValueError(
                 f"ChunkedUploader: unsupported chat_type {chat_type!r}"
             )
@@ -592,8 +592,7 @@ async def _run_with_concurrency(
     concurrency: int,
 ) -> None:
     """Run a list of thunks with a bounded number in flight at once."""
-    if concurrency < 1:
-        concurrency = 1
+    concurrency = max(concurrency, 1)
     sem = asyncio.Semaphore(concurrency)
 
     async def _wrap(thunk: Callable[[], Awaitable[None]]) -> None:

gateway/platforms/signal.py

@@ -99,11 +99,11 @@ def _guess_extension(data: bytes) -> str:
 
 
 def _is_image_ext(ext: str) -> bool:
-    return ext.lower() in (".jpg", ".jpeg", ".png", ".gif", ".webp")
+    return ext.lower() in {".jpg", ".jpeg", ".png", ".gif", ".webp"}
 
 
 def _is_audio_ext(ext: str) -> bool:
-    return ext.lower() in (".mp3", ".wav", ".ogg", ".m4a", ".aac")
+    return ext.lower() in {".mp3", ".wav", ".ogg", ".m4a", ".aac"}
 
 
 _EXT_TO_MIME = {
@@ -1449,7 +1449,7 @@ class SignalAdapter(BasePlatformAdapter):
            contacts from seeing the 👀 reaction (which fires before run.py's
            auth gate and would otherwise reveal that a bot is listening).
         """
-        if os.getenv("SIGNAL_REACTIONS", "true").lower() in ("false", "0", "no"):
+        if os.getenv("SIGNAL_REACTIONS", "true").lower() in {"false", "0", "no"}:
             return False
         if event is not None:
             sender = getattr(getattr(event, "source", None), "user_id", None)

gateway/platforms/slack.py

@@ -679,6 +679,41 @@ class SlackAdapter(BasePlatformAdapter):
             if lock_acquired and not self._running:
                 self._release_platform_lock()
 
+    async def create_handoff_thread(
+        self,
+        parent_chat_id: str,
+        name: str,
+    ) -> Optional[str]:
+        """Create a Slack thread anchor for a session handoff.
+
+        Slack threads are anchored to a parent message (``thread_ts``), not
+        a channel-level construct. So we post a seed message into the home
+        channel and return its ``ts`` — the watcher uses that as the
+        ``thread_id`` for subsequent sends.
+
+        Returns the seed message ts as a string, or ``None`` on failure.
+        """
+        if not self._app:
+            return None
+        try:
+            client = self._get_client(parent_chat_id)
+            if client is None:
+                return None
+            seed_text = f":thread: Hermes handoff — *{(name or 'session').strip()[:80]}*"
+            result = await client.chat_postMessage(
+                channel=parent_chat_id,
+                text=seed_text,
+            )
+            ts = result.get("ts") if isinstance(result, dict) else getattr(result, "get", lambda _k, _d=None: None)("ts")
+            if ts:
+                return str(ts)
+        except Exception as exc:
+            logger.warning(
+                "[%s] Handoff thread: seed-post failed for channel %s: %s",
+                self.name, parent_chat_id, exc,
+            )
+        return None
+
     async def disconnect(self) -> None:
         """Disconnect from Slack."""
         if self._handler:
@@ -900,7 +935,7 @@ class SlackAdapter(BasePlatformAdapter):
         raw = self.config.extra.get("dm_top_level_threads_as_sessions")
         if raw is None:
             return True  # default: each DM thread is its own session
-        return str(raw).strip().lower() in ("1", "true", "yes", "on")
+        return str(raw).strip().lower() in {"1", "true", "yes", "on"}
 
     def _resolve_thread_ts(
         self,
@@ -1265,7 +1300,7 @@ class SlackAdapter(BasePlatformAdapter):
 
     def _reactions_enabled(self) -> bool:
         """Check if message reactions are enabled via config/env."""
-        return os.getenv("SLACK_REACTIONS", "true").lower() not in ("false", "0", "no")
+        return os.getenv("SLACK_REACTIONS", "true").lower() not in {"false", "0", "no"}
 
     async def on_processing_start(self, event: MessageEvent) -> None:
         """Add an in-progress reaction when message processing begins."""
@@ -1738,7 +1773,7 @@ class SlackAdapter(BasePlatformAdapter):
 
         # Ignore message edits and deletions
         subtype = event.get("subtype")
-        if subtype in ("message_changed", "message_deleted"):
+        if subtype in {"message_changed", "message_deleted"}:
             return
 
         original_text = event.get("text", "")
@@ -1857,7 +1892,7 @@ class SlackAdapter(BasePlatformAdapter):
         channel_type = event.get("channel_type", "")
         if not channel_type and channel_id.startswith("D"):
             channel_type = "im"
-        is_dm = channel_type in ("im", "mpim")  # Both 1:1 and group DMs
+        is_dm = channel_type in {"im", "mpim"}  # Both 1:1 and group DMs
 
         # Build thread_ts for session keying.
         # In channels: fall back to ts so each top-level @mention starts a
@@ -1998,7 +2033,7 @@ class SlackAdapter(BasePlatformAdapter):
             if mimetype.startswith("image/") and url:
                 try:
                     ext = "." + mimetype.split("/")[-1].split(";")[0]
-                    if ext not in (".jpg", ".jpeg", ".png", ".gif", ".webp"):
+                    if ext not in {".jpg", ".jpeg", ".png", ".gif", ".webp"}:
                         ext = ".jpg"
                     # Slack private URLs require the bot token as auth header
                     cached = await self._download_slack_file(url, ext, team_id=team_id)
@@ -2014,7 +2049,7 @@ class SlackAdapter(BasePlatformAdapter):
             elif mimetype.startswith("audio/") and url:
                 try:
                     ext = "." + mimetype.split("/")[-1].split(";")[0]
-                    if ext not in (".ogg", ".mp3", ".wav", ".webm", ".m4a"):
+                    if ext not in {".ogg", ".mp3", ".wav", ".webm", ".m4a"}:
                         ext = ".ogg"
                     cached = await self._download_slack_file(url, ext, audio=True, team_id=team_id)
                     media_urls.append(cached)
@@ -2702,7 +2737,7 @@ class SlackAdapter(BasePlatformAdapter):
         if team_id and channel_id:
             self._channel_team[channel_id] = team_id
 
-        if slash_name in ("hermes", ""):
+        if slash_name in {"hermes", ""}:
             # Legacy /hermes <subcommand> [args] routing + free-form questions.
             # Empty slash_name falls into this branch for backward compat
             # with any caller that didn't populate command["command"].
@@ -2897,9 +2932,9 @@ class SlackAdapter(BasePlatformAdapter):
         configured = self.config.extra.get("require_mention")
         if configured is not None:
             if isinstance(configured, str):
-                return configured.lower() not in ("false", "0", "no", "off")
+                return configured.lower() not in {"false", "0", "no", "off"}
             return bool(configured)
-        return os.getenv("SLACK_REQUIRE_MENTION", "true").lower() not in ("false", "0", "no", "off")
+        return os.getenv("SLACK_REQUIRE_MENTION", "true").lower() not in {"false", "0", "no", "off"}
 
     def _slack_strict_mention(self) -> bool:
         """When true, channel threads require an explicit @-mention on every
@@ -2909,9 +2944,9 @@ class SlackAdapter(BasePlatformAdapter):
         configured = self.config.extra.get("strict_mention")
         if configured is not None:
             if isinstance(configured, str):
-                return configured.lower() in ("true", "1", "yes", "on")
+                return configured.lower() in {"true", "1", "yes", "on"}
             return bool(configured)
-        return os.getenv("SLACK_STRICT_MENTION", "false").lower() in ("true", "1", "yes", "on")
+        return os.getenv("SLACK_STRICT_MENTION", "false").lower() in {"true", "1", "yes", "on"}
 
     def _slack_free_response_channels(self) -> set:
         """Return channel IDs where no @mention is required."""

gateway/platforms/telegram.py

@@ -77,7 +77,6 @@ from gateway.platforms.base import (
     SUPPORTED_VIDEO_TYPES,
     SUPPORTED_DOCUMENT_TYPES,
     utf16_len,
-    _prefix_within_utf16_limit,
 )
 from gateway.platforms.telegram_network import (
     TelegramFallbackTransport,
@@ -104,8 +103,58 @@ _TELEGRAM_IMAGE_EXT_TO_MIME = {
 
 
 def check_telegram_requirements() -> bool:
-    """Check if Telegram dependencies are available."""
-    return TELEGRAM_AVAILABLE
+    """Check if Telegram dependencies are available.
+
+    If python-telegram-bot is missing, attempts to lazy-install it via
+    ``tools.lazy_deps.ensure("platform.telegram")``. After a successful
+    install, re-imports the SDK and flips ``TELEGRAM_AVAILABLE`` to True
+    so the adapter's class-level type aliases get rebound.
+    """
+    global TELEGRAM_AVAILABLE, Update, Bot, Message, InlineKeyboardButton
+    global InlineKeyboardMarkup, LinkPreviewOptions, Application
+    global CommandHandler, CallbackQueryHandler, TelegramMessageHandler
+    global ContextTypes, filters, ParseMode, ChatType, HTTPXRequest
+    if TELEGRAM_AVAILABLE:
+        return True
+    try:
+        from tools.lazy_deps import ensure as _lazy_ensure
+        _lazy_ensure("platform.telegram", prompt=False)
+    except Exception:
+        return False
+    try:
+        from telegram import Update as _Update, Bot as _Bot, Message as _Message
+        from telegram import InlineKeyboardButton as _IKB, InlineKeyboardMarkup as _IKM
+        try:
+            from telegram import LinkPreviewOptions as _LPO
+        except ImportError:
+            _LPO = None
+        from telegram.ext import (
+            Application as _App, CommandHandler as _CH,
+            CallbackQueryHandler as _CQH,
+            MessageHandler as _MH,
+            ContextTypes as _CT, filters as _filters,
+        )
+        from telegram.constants import ParseMode as _PM, ChatType as _CtT
+        from telegram.request import HTTPXRequest as _HR
+    except ImportError:
+        return False
+    Update = _Update
+    Bot = _Bot
+    Message = _Message
+    InlineKeyboardButton = _IKB
+    InlineKeyboardMarkup = _IKM
+    LinkPreviewOptions = _LPO
+    Application = _App
+    CommandHandler = _CH
+    CallbackQueryHandler = _CQH
+    TelegramMessageHandler = _MH
+    ContextTypes = _CT
+    filters = _filters
+    ParseMode = _PM
+    ChatType = _CtT
+    HTTPXRequest = _HR
+    TELEGRAM_AVAILABLE = True
+    return True
 
 
 # Matches every character that MarkdownV2 requires to be backslash-escaped
@@ -283,6 +332,50 @@ class TelegramAdapter(BasePlatformAdapter):
     MEDIA_GROUP_WAIT_SECONDS = 0.8
     _GENERAL_TOPIC_THREAD_ID = "1"
 
+    # Adaptive text-batch ingress: short messages need a tighter delay so the
+    # first token reaches the agent fast.  Numbers tuned for "feels instant":
+    # ≤320 codepoints (one short paragraph) settles in ~180ms; ≤1024
+    # (a normal paragraph) in ~240ms; longer waits the configured cap.
+    # Always clamped to ``_text_batch_delay_seconds`` so an operator can lower
+    # the cap further via env var.
+    _TEXT_BATCH_FAST_LEN = 320
+    _TEXT_BATCH_FAST_DELAY_S = 0.18
+    _TEXT_BATCH_SHORT_LEN = 1024
+    _TEXT_BATCH_SHORT_DELAY_S = 0.24
+
+    @staticmethod
+    def _env_float_clamped(
+        name: str,
+        default: float,
+        *,
+        min_value: Optional[float] = None,
+        max_value: Optional[float] = None,
+    ) -> float:
+        """Read a float env var, reject non-finite values, and clamp to bounds.
+
+        Guarantees the returned value is a finite number usable directly in
+        ``asyncio.sleep()`` and similar APIs that reject NaN / Inf.
+        """
+        import math
+
+        raw = os.getenv(name)
+        try:
+            value = float(raw) if raw is not None else float(default)
+        except (TypeError, ValueError):
+            value = float(default)
+        if not math.isfinite(value):
+            value = float(default)
+        if min_value is not None:
+            value = max(value, min_value)
+        if max_value is not None:
+            value = min(value, max_value)
+        return value
+
+    @property
+    def message_len_fn(self):
+        """Telegram measures message length in UTF-16 code units."""
+        return utf16_len
+
     def __init__(self, config: PlatformConfig):
         super().__init__(config, Platform.TELEGRAM)
         self._app: Optional[Application] = None
@@ -299,9 +392,24 @@ class TelegramAdapter(BasePlatformAdapter):
         self._media_group_events: Dict[str, MessageEvent] = {}
         self._media_group_tasks: Dict[str, asyncio.Task] = {}
         # Buffer rapid text messages so Telegram client-side splits of long
-        # messages are aggregated into a single MessageEvent.
-        self._text_batch_delay_seconds = float(os.getenv("HERMES_TELEGRAM_TEXT_BATCH_DELAY_SECONDS", "0.6"))
-        self._text_batch_split_delay_seconds = float(os.getenv("HERMES_TELEGRAM_TEXT_BATCH_SPLIT_DELAY_SECONDS", "2.0"))
+        # messages are aggregated into a single MessageEvent.  Lower defaults
+        # (0.3s / 1.0s instead of 0.6s / 2.0s) let short replies stream
+        # without a noticeable wait — combined with the adaptive fast-path
+        # in ``_calc_text_batch_delay`` below, ≤320-codepoint replies settle
+        # in ~180ms.  All bounds are conservative for Telegram's
+        # ~1 edit/s flood envelope.
+        self._text_batch_delay_seconds = self._env_float_clamped(
+            "HERMES_TELEGRAM_TEXT_BATCH_DELAY_SECONDS",
+            0.3,
+            min_value=0.08,
+            max_value=2.0,
+        )
+        self._text_batch_split_delay_seconds = self._env_float_clamped(
+            "HERMES_TELEGRAM_TEXT_BATCH_SPLIT_DELAY_SECONDS",
+            1.0,
+            min_value=self._text_batch_delay_seconds,
+            max_value=4.0,
+        )
         self._pending_text_batches: Dict[str, MessageEvent] = {}
         self._pending_text_batch_tasks: Dict[str, asyncio.Task] = {}
         self._polling_error_task: Optional[asyncio.Task] = None
@@ -319,6 +427,30 @@ class TelegramAdapter(BasePlatformAdapter):
         # Slash-confirm button state: confirm_id → session_key (for /reload-mcp
         # and any other slash-confirm prompts; see GatewayRunner._request_slash_confirm).
         self._slash_confirm_state: Dict[str, str] = {}
+        # Notification mode for message sends.
+        # "important" — only final responses, approvals, and slash confirmations
+        #               trigger notifications; tool progress, streaming, status
+        #               messages are delivered silently via disable_notification.
+        #               This is the default — Telegram users found per-tool-call
+        #               push notifications too noisy.
+        # "all"       — every message triggers a push notification (legacy
+        #               behavior; opt-in via display.platforms.telegram.notifications).
+        self._notifications_mode: str = "important"
+
+    def _notification_kwargs(
+        self, metadata: Optional[Dict[str, Any]]
+    ) -> Dict[str, Any]:
+        """Return disable_notification kwargs when the adapter is in silent mode.
+
+        In "important" mode, all message sends are silently delivered
+        (disable_notification=True) unless the caller explicitly requests a
+        notification by setting ``metadata["notify"] = True``.
+        """
+        if getattr(self, "_notifications_mode", "important") != "important":
+            return {}
+        if (metadata or {}).get("notify"):
+            return {}
+        return {"disable_notification": True}
 
     def _is_callback_user_authorized(
         self,
@@ -534,7 +666,7 @@ class TelegramAdapter(BasePlatformAdapter):
     def _looks_like_network_error(error: Exception) -> bool:
         """Return True for transient network errors that warrant a reconnect attempt."""
         name = error.__class__.__name__.lower()
-        if name in ("networkerror", "timedout", "connectionerror"):
+        if name in {"networkerror", "timedout", "connectionerror"}:
             return True
         try:
             from telegram.error import NetworkError, TimedOut
@@ -550,9 +682,9 @@ class TelegramAdapter(BasePlatformAdapter):
             return default
         if isinstance(value, str):
             lowered = value.strip().lower()
-            if lowered in ("true", "1", "yes", "on"):
+            if lowered in {"true", "1", "yes", "on"}:
                 return True
-            if lowered in ("false", "0", "no", "off"):
+            if lowered in {"false", "0", "no", "off"}:
                 return False
             return default
         return bool(value)
@@ -841,6 +973,24 @@ class TelegramAdapter(BasePlatformAdapter):
                 )
             return None
 
+    async def create_handoff_thread(
+        self,
+        parent_chat_id: str,
+        name: str,
+    ) -> Optional[str]:
+        """Create a forum topic for a session handoff.
+
+        Works for DM topics (Bot API 9.4+, requires user to enable Topics
+        in their chat with the bot) and forum supergroups. Returns the
+        ``message_thread_id`` as a string, or ``None`` on failure.
+        """
+        try:
+            chat_id_int = int(parent_chat_id)
+        except (TypeError, ValueError):
+            return None
+        thread_id = await self._create_dm_topic(chat_id_int, name=name)
+        return str(thread_id) if thread_id else None
+
     async def rename_dm_topic(
         self,
         chat_id: int,
@@ -1071,7 +1221,7 @@ class TelegramAdapter(BasePlatformAdapter):
                 "write_timeout": _env_float("HERMES_TELEGRAM_HTTP_WRITE_TIMEOUT", 20.0),
             }
 
-            disable_fallback = (os.getenv("HERMES_TELEGRAM_DISABLE_FALLBACK_IPS", "").strip().lower() in ("1", "true", "yes", "on"))
+            disable_fallback = (os.getenv("HERMES_TELEGRAM_DISABLE_FALLBACK_IPS", "").strip().lower() in {"1", "true", "yes", "on"})
             fallback_ips = self._fallback_ips()
             if not fallback_ips:
                 fallback_ips = await discover_fallback_ips()
@@ -1414,6 +1564,7 @@ class TelegramAdapter(BasePlatformAdapter):
                                 reply_to_message_id=reply_to_id,
                                 **thread_kwargs,
                                 **self._link_preview_kwargs(),
+                                **self._notification_kwargs(metadata),
                             )
                         except Exception as md_error:
                             # Markdown parsing failed, try plain text
@@ -1427,6 +1578,7 @@ class TelegramAdapter(BasePlatformAdapter):
                                     reply_to_message_id=reply_to_id,
                                     **thread_kwargs,
                                     **self._link_preview_kwargs(),
+                                    **self._notification_kwargs(metadata),
                                 )
                             else:
                                 raise
@@ -1510,10 +1662,18 @@ class TelegramAdapter(BasePlatformAdapter):
             
         except Exception as e:
             logger.error("[%s] Failed to send Telegram message: %s", self.name, e, exc_info=True)
+            err_str = str(e).lower()
+            # Message too long — content exceeded 4096 chars. Return failure so
+            # stream consumer enters fallback mode and sends the remainder.
+            if "message_too_long" in err_str or "too long" in err_str:
+                logger.debug(
+                    "[%s] send() content too long, falling back to new-message continuation",
+                    self.name,
+                )
+                return SendResult(success=False, error="message_too_long")
             # TimedOut means the request may have reached Telegram —
             # mark as non-retryable so _send_with_retry() doesn't re-send.
             _to = locals().get("_TimedOut")
-            err_str = str(e).lower()
             is_timeout = (_to and isinstance(e, _to)) or "timed out" in err_str
             return SendResult(success=False, error=str(e), retryable=not is_timeout)
 
@@ -1525,9 +1685,26 @@ class TelegramAdapter(BasePlatformAdapter):
         *,
         finalize: bool = False,
     ) -> SendResult:
-        """Edit a previously sent Telegram message."""
+        """Edit a previously sent Telegram message.
+
+        Telegram caps single-message text at 4096 UTF-16 codeunits.  Streaming
+        replies that grow past this limit must NOT be silently truncated and
+        must NOT return failure (the consumer would re-send and create a
+        duplicate).  Instead this method split-and-delivers: edit the
+        existing message with the first chunk and send the rest as
+        continuation messages, returning the final chunk's id so subsequent
+        edits target the most recent visible message.
+        """
         if not self._bot:
             return SendResult(success=False, error="Not connected")
+
+        # Pre-flight: if content already exceeds the limit, split-and-deliver
+        # without round-tripping a doomed edit.
+        if utf16_len(content) > self.MAX_MESSAGE_LENGTH:
+            return await self._edit_overflow_split(
+                chat_id, message_id, content, finalize=finalize,
+            )
+
         try:
             if not finalize:
                 await self._bot.edit_message_text(
@@ -1561,22 +1738,17 @@ class TelegramAdapter(BasePlatformAdapter):
             # "Message is not modified" — content identical, treat as success
             if "not modified" in err_str:
                 return SendResult(success=True, message_id=message_id)
-            # Message too long — content exceeded 4096 chars (e.g. during
-            # streaming).  Truncate and succeed so the stream consumer can
-            # split the overflow into a new message instead of dying.
+            # Reactive split-and-deliver: parse_mode formatting can inflate
+            # the payload past the limit even when the raw text was under
+            # (e.g. MarkdownV2 escapes).  Same fix as the pre-flight path.
             if "message_too_long" in err_str or "too long" in err_str:
-                truncated = _prefix_within_utf16_limit(
-                    content, self.MAX_MESSAGE_LENGTH - 20
-                ) + "…"
-                try:
-                    await self._bot.edit_message_text(
-                        chat_id=int(chat_id),
-                        message_id=int(message_id),
-                        text=truncated,
-                    )
-                except Exception:
-                    pass  # best-effort truncation
-                return SendResult(success=True, message_id=message_id)
+                logger.debug(
+                    "[%s] edit_message overflow (%d UTF-16 > %d), splitting",
+                    self.name, utf16_len(content), self.MAX_MESSAGE_LENGTH,
+                )
+                return await self._edit_overflow_split(
+                    chat_id, message_id, content, finalize=finalize,
+                )
             # Flood control / RetryAfter — short waits are retried inline,
             # long waits return a failure immediately so streaming can fall back
             # to a normal final send instead of leaving a truncated partial.
@@ -1612,6 +1784,147 @@ class TelegramAdapter(BasePlatformAdapter):
             )
             return SendResult(success=False, error=str(e))
 
+    async def _edit_overflow_split(
+        self,
+        chat_id: str,
+        message_id: str,
+        content: str,
+        *,
+        finalize: bool,
+    ) -> SendResult:
+        """Split an oversized edit across the existing message + continuations.
+
+        Edit the original ``message_id`` with chunk 1 (with the platform's
+        usual ``(1/N)`` suffix preserved), then send the remaining chunks as
+        new messages threaded as replies to the previous chunk so the user
+        sees them grouped.  Returns ``SendResult(success=True,
+        message_id=<last-chunk-id>, continuation_message_ids=(...))`` so the
+        stream consumer can keep editing the most recent visible message
+        and the gateway has full visibility into every message id we put on
+        screen.
+
+        Falls back to ``SendResult(success=False)`` only if even the first-
+        chunk edit fails — that's a real adapter problem, not an overflow.
+        """
+        chunks = self.truncate_message(
+            content, self.MAX_MESSAGE_LENGTH, len_fn=utf16_len,
+        )
+        if len(chunks) <= 1:
+            # Defensive: shouldn't happen given the caller's pre-flight, but
+            # if truncate_message returned a single chunk just edit normally.
+            chunks = [content]
+
+        # Step 1 — edit the existing message with the first chunk.
+        first_chunk = chunks[0]
+        try:
+            if finalize:
+                # Use format_message + parse_mode for the final chunk;
+                # mirror edit_message's main happy-path.
+                formatted = self.format_message(first_chunk)
+                try:
+                    await self._bot.edit_message_text(
+                        chat_id=int(chat_id),
+                        message_id=int(message_id),
+                        text=formatted,
+                        parse_mode=ParseMode.MARKDOWN_V2,
+                    )
+                except Exception as fmt_err:
+                    if "not modified" not in str(fmt_err).lower():
+                        await self._bot.edit_message_text(
+                            chat_id=int(chat_id),
+                            message_id=int(message_id),
+                            text=first_chunk,
+                        )
+            else:
+                await self._bot.edit_message_text(
+                    chat_id=int(chat_id),
+                    message_id=int(message_id),
+                    text=first_chunk,
+                )
+        except Exception as e:
+            err_str = str(e).lower()
+            if "not modified" in err_str:
+                # First chunk identical to current text — fall through to
+                # send continuations.
+                pass
+            else:
+                logger.error(
+                    "[%s] Overflow split: first-chunk edit failed: %s",
+                    self.name, e, exc_info=True,
+                )
+                return SendResult(success=False, error=str(e))
+
+        # Step 2 — send each remaining chunk as a continuation message,
+        # threaded as a reply to the previous so the user sees them as a
+        # contiguous block.  We call self._bot.send_message directly so the
+        # continuation skips ``self.send``'s own pre-chunking pass (chunks
+        # are already correctly sized).  Best-effort MarkdownV2 with plain
+        # fallback, mirroring send().
+        continuation_ids: list[str] = []
+        prev_id = message_id
+        for chunk in chunks[1:]:
+            sent_msg = None
+            for use_markdown in (True, False) if finalize else (False,):
+                try:
+                    text = self.format_message(chunk) if use_markdown else chunk
+                    sent_msg = await self._bot.send_message(
+                        chat_id=int(chat_id),
+                        text=text,
+                        parse_mode=ParseMode.MARKDOWN_V2 if use_markdown else None,
+                        reply_to_message_id=int(prev_id) if prev_id else None,
+                    )
+                    break
+                except Exception as send_err:
+                    if "reply message not found" in str(send_err).lower():
+                        # Drop the reply anchor and try again.
+                        try:
+                            sent_msg = await self._bot.send_message(
+                                chat_id=int(chat_id),
+                                text=chunk,
+                            )
+                            break
+                        except Exception as _retry_err:
+                            logger.warning(
+                                "[%s] Overflow continuation no-reply retry failed: %s",
+                                self.name, _retry_err,
+                            )
+                            sent_msg = None
+                            break
+                    if use_markdown:
+                        # try plain text on next loop iteration
+                        continue
+                    logger.warning(
+                        "[%s] Overflow continuation send failed: %s",
+                        self.name, send_err,
+                    )
+                    sent_msg = None
+                    break
+            if sent_msg is None:
+                # Continuation failed — the user has chunk 1 + however many
+                # continuations succeeded.  Report success with what we got
+                # so the stream consumer knows the edit landed; the
+                # remaining tail is lost on this attempt and the next
+                # streaming tick may retry.
+                logger.warning(
+                    "[%s] Overflow split: stopped at %d/%d chunks delivered",
+                    self.name, 1 + len(continuation_ids), len(chunks),
+                )
+                break
+            new_id = str(getattr(sent_msg, "message_id", "")) or prev_id
+            continuation_ids.append(new_id)
+            prev_id = new_id
+
+        last_id = continuation_ids[-1] if continuation_ids else message_id
+        logger.debug(
+            "[%s] Overflow split delivered %d chunks; last_id=%s",
+            self.name, 1 + len(continuation_ids), last_id,
+        )
+        return SendResult(
+            success=True,
+            message_id=last_id,
+            continuation_message_ids=tuple(continuation_ids),
+        )
+
     async def delete_message(self, chat_id: str, message_id: str) -> bool:
         """Delete a previously sent Telegram message.
 
@@ -1637,6 +1950,109 @@ class TelegramAdapter(BasePlatformAdapter):
             )
             return False
 
+    def supports_draft_streaming(
+        self,
+        chat_type: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> bool:
+        """Telegram supports sendMessageDraft for private chats only.
+
+        Bot API 9.5 (March 2026) opened ``sendMessageDraft`` to all bots
+        unconditionally for private (DM) chats.  Groups, supergroups, and
+        channels still rely on the edit-based path.
+
+        We additionally require ``self._bot`` to expose ``send_message_draft``
+        (added to python-telegram-bot in 22.6); older PTB installs gracefully
+        fall back to the edit path even on DMs.
+        """
+        if not self._bot or not hasattr(self._bot, "send_message_draft"):
+            return False
+        return (chat_type or "").lower() in {"dm", "private"}
+
+    async def send_draft(
+        self,
+        chat_id: str,
+        draft_id: int,
+        content: str,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Stream a partial message via Telegram's native sendMessageDraft.
+
+        The Bot API animates the preview when the same ``draft_id`` is reused
+        across consecutive calls in the same chat.  When the response
+        finishes, the caller sends the final text via the normal ``send``
+        path; the draft preview clears naturally on the client (Telegram has
+        no Bot API to "promote" a draft to a real message — the final
+        ``sendMessage`` is what the user receives in their history).
+        """
+        if not self._bot:
+            return SendResult(success=False, error="not_connected")
+        if not hasattr(self._bot, "send_message_draft"):
+            return SendResult(success=False, error="api_unavailable")
+
+        # Trim to the same UTF-16 budget the platform enforces on regular
+        # sends.  Drafts have the same length contract as messages.
+        text = content if len(content) <= self.MAX_MESSAGE_LENGTH else \
+            self.truncate_message(content, self.MAX_MESSAGE_LENGTH, len_fn=utf16_len)[0]
+
+        kwargs: Dict[str, Any] = {
+            "chat_id": int(chat_id),
+            "draft_id": int(draft_id),
+            "text": text,
+        }
+        thread_id = self._metadata_thread_id(metadata)
+        if thread_id is not None:
+            kwargs["message_thread_id"] = thread_id
+
+        try:
+            ok = await self._bot.send_message_draft(**kwargs)
+            if ok:
+                # Drafts have no message_id; we report success without one
+                # so the caller knows the animation frame landed.
+                return SendResult(success=True, message_id=None)
+            return SendResult(success=False, error="draft_rejected")
+        except Exception as e:
+            # Most likely: BadRequest because this bot/chat doesn't allow
+            # drafts, or a transient server hiccup.  The caller treats any
+            # failure as "fall back to edit-based for this response".
+            logger.debug(
+                "[%s] sendMessageDraft failed (chat=%s draft_id=%s): %s",
+                self.name, chat_id, draft_id, e,
+            )
+            return SendResult(success=False, error=str(e))
+
+    async def _send_message_with_thread_fallback(self, **kwargs):
+        """Send a Telegram message, retrying once without message_thread_id
+        if Telegram returns 'Message thread not found'.
+
+        Used for control-style sends (approval prompts, model picker,
+        update prompts) that can carry a stale thread_id from a DM
+        reply chain.  The streaming send loop has its own equivalent
+        (PR #3390) at the body of ``send``; this helper applies the
+        same retry pattern to the non-streaming control paths.
+        """
+        if not self._bot:
+            raise RuntimeError("Not connected")
+
+        message_thread_id = kwargs.get("message_thread_id")
+        try:
+            return await self._bot.send_message(**kwargs)
+        except Exception as send_err:
+            if (
+                message_thread_id is not None
+                and self._is_bad_request_error(send_err)
+                and self._is_thread_not_found_error(send_err)
+            ):
+                logger.warning(
+                    "[%s] Thread %s not found for control message, retrying without message_thread_id",
+                    self.name,
+                    message_thread_id,
+                )
+                retry_kwargs = dict(kwargs)
+                retry_kwargs.pop("message_thread_id", None)
+                return await self._bot.send_message(**retry_kwargs)
+            raise
+
     async def send_update_prompt(
         self, chat_id: str, prompt: str, default: str = "",
         session_key: str = "",
@@ -1660,7 +2076,7 @@ class TelegramAdapter(BasePlatformAdapter):
             ])
             thread_id = self._metadata_thread_id(metadata)
             reply_to_id = self._reply_to_message_id_for_send(None, metadata)
-            msg = await self._bot.send_message(
+            msg = await self._send_message_with_thread_fallback(
                 chat_id=int(chat_id),
                 text=text,
                 parse_mode=ParseMode.MARKDOWN,
@@ -1740,7 +2156,7 @@ class TelegramAdapter(BasePlatformAdapter):
                 )
             )
 
-            msg = await self._bot.send_message(**kwargs)
+            msg = await self._send_message_with_thread_fallback(**kwargs)
 
             # Store session_key keyed by approval_id for the callback handler
             self._approval_state[approval_id] = session_key
@@ -1792,7 +2208,7 @@ class TelegramAdapter(BasePlatformAdapter):
                 )
             )
 
-            msg = await self._bot.send_message(**kwargs)
+            msg = await self._send_message_with_thread_fallback(**kwargs)
             self._slash_confirm_state[confirm_id] = session_key
             return SendResult(success=True, message_id=str(msg.message_id))
         except Exception as e:
@@ -1850,7 +2266,7 @@ class TelegramAdapter(BasePlatformAdapter):
 
             thread_id = metadata.get("thread_id") if metadata else None
             reply_to_id = self._reply_to_message_id_for_send(None, metadata)
-            msg = await self._bot.send_message(
+            msg = await self._send_message_with_thread_fallback(
                 chat_id=int(chat_id),
                 text=text,
                 parse_mode=ParseMode.MARKDOWN,
@@ -2357,7 +2773,7 @@ class TelegramAdapter(BasePlatformAdapter):
             with open(audio_path, "rb") as audio_file:
                 ext = os.path.splitext(audio_path)[1].lower()
                 # .ogg / .opus files -> send as voice (round playable bubble)
-                if ext in (".ogg", ".opus"):
+                if ext in {".ogg", ".opus"}:
                     _voice_thread = self._metadata_thread_id(metadata)
                     reply_to_id = self._reply_to_message_id_for_send(reply_to, metadata)
                     voice_thread_kwargs = self._thread_kwargs_for_send(
@@ -2374,13 +2790,14 @@ class TelegramAdapter(BasePlatformAdapter):
                             "caption": caption[:1024] if caption else None,
                             "reply_to_message_id": reply_to_id,
                             **voice_thread_kwargs,
+                            **self._notification_kwargs(metadata),
                         },
                         metadata,
                         reply_to_id,
                         "voice",
                         reset_media=lambda: audio_file.seek(0),
                     )
-                elif ext in (".mp3", ".m4a"):
+                elif ext in {".mp3", ".m4a"}:
                     # Telegram's Bot API sendAudio only accepts MP3 / M4A.
                     _audio_thread = self._metadata_thread_id(metadata)
                     reply_to_id = self._reply_to_message_id_for_send(reply_to, metadata)
@@ -2398,6 +2815,7 @@ class TelegramAdapter(BasePlatformAdapter):
                             "caption": caption[:1024] if caption else None,
                             "reply_to_message_id": reply_to_id,
                             **audio_thread_kwargs,
+                            **self._notification_kwargs(metadata),
                         },
                         metadata,
                         reply_to_id,
@@ -2534,6 +2952,7 @@ class TelegramAdapter(BasePlatformAdapter):
                         "media": media,
                         "reply_to_message_id": reply_to_id,
                         **thread_kwargs,
+                        **self._notification_kwargs(metadata),
                     },
                     metadata,
                     reply_to_id,
@@ -2591,6 +3010,7 @@ class TelegramAdapter(BasePlatformAdapter):
                         "caption": caption[:1024] if caption else None,
                         "reply_to_message_id": reply_to_id,
                         **thread_kwargs,
+                        **self._notification_kwargs(metadata),
                     },
                     metadata,
                     reply_to_id,
@@ -2686,6 +3106,7 @@ class TelegramAdapter(BasePlatformAdapter):
                         "caption": caption[:1024] if caption else None,
                         "reply_to_message_id": reply_to_id,
                         **thread_kwargs,
+                        **self._notification_kwargs(metadata),
                     },
                     metadata,
                     reply_to_id,
@@ -2731,6 +3152,7 @@ class TelegramAdapter(BasePlatformAdapter):
                         "caption": caption[:1024] if caption else None,
                         "reply_to_message_id": reply_to_id,
                         **thread_kwargs,
+                        **self._notification_kwargs(metadata),
                     },
                     metadata,
                     reply_to_id,
@@ -2781,6 +3203,7 @@ class TelegramAdapter(BasePlatformAdapter):
                     "caption": caption[:1024] if caption else None,
                     "reply_to_message_id": reply_to_id,
                     **photo_thread_kwargs,
+                    **self._notification_kwargs(metadata),
                 },
                 metadata,
                 reply_to_id,
@@ -2816,6 +3239,7 @@ class TelegramAdapter(BasePlatformAdapter):
                         "caption": caption[:1024] if caption else None,
                         "reply_to_message_id": reply_to_id,
                         **upload_thread_kwargs,
+                        **self._notification_kwargs(metadata),
                     },
                     metadata,
                     reply_to_id,
@@ -2861,6 +3285,7 @@ class TelegramAdapter(BasePlatformAdapter):
                     "caption": caption[:1024] if caption else None,
                     "reply_to_message_id": reply_to_id,
                     **animation_thread_kwargs,
+                    **self._notification_kwargs(metadata),
                 },
                 metadata,
                 reply_to_id,
@@ -3123,9 +3548,18 @@ class TelegramAdapter(BasePlatformAdapter):
         configured = self.config.extra.get("require_mention")
         if configured is not None:
             if isinstance(configured, str):
-                return configured.lower() in ("true", "1", "yes", "on")
+                return configured.lower() in {"true", "1", "yes", "on"}
             return bool(configured)
-        return os.getenv("TELEGRAM_REQUIRE_MENTION", "false").lower() in ("true", "1", "yes", "on")
+        return os.getenv("TELEGRAM_REQUIRE_MENTION", "false").lower() in {"true", "1", "yes", "on"}
+
+    def _telegram_guest_mode(self) -> bool:
+        """Return whether non-allowlisted groups may trigger via direct @mention."""
+        configured = self.config.extra.get("guest_mode")
+        if configured is not None:
+            if isinstance(configured, str):
+                return configured.lower() in {"true", "1", "yes", "on"}
+            return bool(configured)
+        return os.getenv("TELEGRAM_GUEST_MODE", "false").lower() in {"true", "1", "yes", "on"}
 
     def _telegram_free_response_chats(self) -> set[str]:
         raw = self.config.extra.get("free_response_chats")
@@ -3138,8 +3572,9 @@ class TelegramAdapter(BasePlatformAdapter):
     def _telegram_allowed_chats(self) -> set[str]:
         """Return the whitelist of group/supergroup chat IDs the bot will respond in.
 
-        When non-empty, group messages from chats NOT in this set are silently
-        ignored — even if the bot is @mentioned.  DMs are never filtered.
+        When non-empty, group messages from chats NOT in this set are
+        silently ignored unless ``guest_mode`` is enabled and the bot is
+        explicitly @mentioned.  DMs are never filtered.
         Empty set means no restriction (fully backward compatible).
         """
         raw = self.config.extra.get("allowed_chats")
@@ -3213,7 +3648,7 @@ class TelegramAdapter(BasePlatformAdapter):
         if not chat:
             return False
         chat_type = str(getattr(chat, "type", "")).split(".")[-1].lower()
-        return chat_type in ("group", "supergroup")
+        return chat_type in {"group", "supergroup"}
 
     def _is_reply_to_bot(self, message: Message) -> bool:
         if not self._bot or not getattr(message, "reply_to_message", None):
@@ -3286,6 +3721,14 @@ class TelegramAdapter(BasePlatformAdapter):
                     return True
         return False
 
+    def _is_guest_mention(self, message: Message) -> bool:
+        """Return True for the narrow guest-mode bypass: explicit bot mention.
+
+        The caller (:meth:`_should_process_message`) has already verified
+        the message is a group chat, so that check is not repeated here.
+        """
+        return self._telegram_guest_mode() and self._message_mentions_bot(message)
+
     def _clean_bot_trigger_text(self, text: Optional[str]) -> Optional[str]:
         if not text or not self._bot or not getattr(self._bot, "username", None):
             return text
@@ -3297,16 +3740,18 @@ class TelegramAdapter(BasePlatformAdapter):
         """Apply Telegram group trigger rules.
 
         DMs remain unrestricted. Group/supergroup messages are accepted when:
-        - the chat passes the ``allowed_chats`` whitelist (when set)
+        - the chat passes the ``allowed_chats`` whitelist (when set), or
+          ``guest_mode`` is enabled and the bot is explicitly mentioned
         - the chat is explicitly allowlisted in ``free_response_chats``
         - ``require_mention`` is disabled
         - the message replies to the bot
         - the bot is @mentioned
         - the text/caption matches a configured regex wake-word pattern
 
-        When ``allowed_chats`` is non-empty, it acts as a hard gate — messages
-        from any chat not in the list are ignored regardless of the other
-        rules.  When ``require_mention`` is enabled, slash commands are not given
+        When ``allowed_chats`` is non-empty, it remains a hard gate except for
+        the narrow ``guest_mode`` bypass: group/supergroup messages that
+        explicitly @mention this bot. Replies and regex wake words do not bypass
+        ``allowed_chats``. When ``require_mention`` is enabled, slash commands are not given
         special treatment — they must pass the same mention/reply checks
         as any other group message.  Users can still trigger commands via
         the Telegram bot menu (``/command@botname``) or by explicitly
@@ -3315,14 +3760,7 @@ class TelegramAdapter(BasePlatformAdapter):
         """
         if not self._is_group_chat(message):
             return True
-        # allowed_chats check (whitelist — must pass before other gating).
-        # When set, group messages from chats NOT in this whitelist are
-        # silently ignored, even if @mentioned.  DMs are already excluded above.
-        allowed = self._telegram_allowed_chats()
-        if allowed:
-            chat_id_str = str(getattr(getattr(message, "chat", None), "id", ""))
-            if chat_id_str not in allowed:
-                return False
+
         thread_id = getattr(message, "message_thread_id", None)
         if thread_id is not None:
             try:
@@ -3330,13 +3768,31 @@ class TelegramAdapter(BasePlatformAdapter):
                     return False
             except (TypeError, ValueError):
                 logger.warning("[%s] Ignoring non-numeric Telegram message_thread_id: %r", self.name, thread_id)
-        if str(getattr(getattr(message, "chat", None), "id", "")) in self._telegram_free_response_chats():
+
+        chat_id_str = str(getattr(getattr(message, "chat", None), "id", ""))
+
+        # Resolve guest-mode mention bypass once so _message_mentions_bot
+        # is not called redundantly in the normal flow below.
+        guest_mention = self._is_guest_mention(message)
+
+        # allowed_chats check (whitelist). When set, group messages from chats
+        # outside the whitelist are ignored unless guest_mode permits this
+        # exact message as an explicit direct mention. DMs are excluded above.
+        allowed = self._telegram_allowed_chats()
+        if allowed and chat_id_str not in allowed:
+            return guest_mention
+
+        if guest_mention:
+            return True
+        if chat_id_str in self._telegram_free_response_chats():
             return True
         if not self._telegram_require_mention():
             return True
         if self._is_reply_to_bot(message):
             return True
-        if self._message_mentions_bot(message):
+        # When guest_mode is True, _is_guest_mention already called
+        # _message_mentions_bot above — skip the redundant second call.
+        if not self._telegram_guest_mode() and self._message_mentions_bot(message):
             return True
         return self._message_matches_mention_patterns(message)
 
@@ -3456,12 +3912,27 @@ class TelegramAdapter(BasePlatformAdapter):
         """
         current_task = asyncio.current_task()
         try:
-            # Adaptive delay: if the latest chunk is near Telegram's 4096-char
-            # split point, a continuation is almost certain — wait longer.
+            # Adaptive delay tiers:
+            #  - last chunk ≥ _SPLIT_THRESHOLD: a continuation is almost
+            #    certain → wait the longer split delay.
+            #  - total accumulated text ≤ _TEXT_BATCH_FAST_LEN (~320 cp):
+            #    short message → cap delay at _TEXT_BATCH_FAST_DELAY_S
+            #    so the agent sees the text near-instantly.
+            #  - total ≤ _TEXT_BATCH_SHORT_LEN (~1024 cp):
+            #    medium → cap at _TEXT_BATCH_SHORT_DELAY_S.
+            #  - otherwise: use the configured cap.
+            # Tiers compose with operator overrides via the env-var-driven
+            # ``_text_batch_delay_seconds`` (e.g. an operator who sets the
+            # cap below 0.18s gets that lower number on every tier).
             pending = self._pending_text_batches.get(key)
             last_len = getattr(pending, "_last_chunk_len", 0) if pending else 0
+            total_len = len(getattr(pending, "text", "") or "") if pending else 0
             if last_len >= self._SPLIT_THRESHOLD:
                 delay = self._text_batch_split_delay_seconds
+            elif total_len <= self._TEXT_BATCH_FAST_LEN:
+                delay = min(self._text_batch_delay_seconds, self._TEXT_BATCH_FAST_DELAY_S)
+            elif total_len <= self._TEXT_BATCH_SHORT_LEN:
+                delay = min(self._text_batch_delay_seconds, self._TEXT_BATCH_SHORT_DELAY_S)
             else:
                 delay = self._text_batch_delay_seconds
             await asyncio.sleep(delay)
@@ -3736,7 +4207,7 @@ class TelegramAdapter(BasePlatformAdapter):
 
                 # For text files, inject content into event.text (capped at 100 KB)
                 MAX_TEXT_INJECT_BYTES = 100 * 1024
-                if ext in (".md", ".txt") and len(raw_bytes) <= MAX_TEXT_INJECT_BYTES:
+                if ext in {".md", ".txt"} and len(raw_bytes) <= MAX_TEXT_INJECT_BYTES:
                     try:
                         text_content = raw_bytes.decode("utf-8")
                         display_name = original_filename or f"document{ext}"
@@ -3975,14 +4446,29 @@ class TelegramAdapter(BasePlatformAdapter):
         
         # Determine chat type
         chat_type = "dm"
-        if chat.type in (ChatType.GROUP, ChatType.SUPERGROUP):
+        if chat.type in {ChatType.GROUP, ChatType.SUPERGROUP}:
             chat_type = "group"
         elif chat.type == ChatType.CHANNEL:
             chat_type = "channel"
 
-        # Resolve DM topic name and skill binding
+        # Resolve DM topic name and skill binding.
+        # In private chats, only preserve thread ids for real topic messages
+        # (is_topic_message=True).  Telegram puts message_thread_id on every
+        # DM that is a reply, even when the user is just replying to a
+        # previous message in the same DM — that bogus id then routes to a
+        # nonexistent thread and Telegram returns 'Message thread not found'
+        # on send (#3206).
         thread_id_raw = message.message_thread_id
-        thread_id_str = str(thread_id_raw) if thread_id_raw is not None else None
+        is_topic_message = bool(getattr(message, "is_topic_message", False))
+        thread_id_str = None
+        if thread_id_raw is not None:
+            if chat_type == "group":
+                thread_id_str = str(thread_id_raw)
+            elif chat_type == "dm" and is_topic_message:
+                thread_id_str = str(thread_id_raw)
+        # For forum groups without an explicit topic, default to the
+        # General-topic id so the gateway routes back to the General topic
+        # rather than dropping into the bot's main channel (#22423).
         if chat_type == "group" and thread_id_str is None and getattr(chat, "is_forum", False):
             thread_id_str = self._GENERAL_TOPIC_THREAD_ID
         chat_topic = None
@@ -4026,12 +4512,28 @@ class TelegramAdapter(BasePlatformAdapter):
             chat_topic=chat_topic,
         )
         
-        # Extract reply context if this message is a reply
+        # Extract reply context if this message is a reply.
+        # Prefer Telegram's native partial quote (message.quote, TextQuote)
+        # so a user replying to a single selected substring of a prior
+        # multi-section message doesn't get the whole replied-to message
+        # injected into the agent's context — which can cause the agent
+        # to act on unrelated actionable-looking text the user didn't
+        # quote (#22619). Fall back to the full replied-to message text
+        # / caption when no native quote is present.
         reply_to_id = None
         reply_to_text = None
         if message.reply_to_message:
             reply_to_id = str(message.reply_to_message.message_id)
-            reply_to_text = message.reply_to_message.text or message.reply_to_message.caption or None
+            quote = getattr(message, "quote", None)
+            quote_text = getattr(quote, "text", None) if quote is not None else None
+            if quote_text:
+                reply_to_text = quote_text
+            else:
+                reply_to_text = (
+                    message.reply_to_message.text
+                    or message.reply_to_message.caption
+                    or None
+                )
 
         # Per-channel/topic ephemeral prompt
         from gateway.platforms.base import resolve_channel_prompt
@@ -4060,7 +4562,7 @@ class TelegramAdapter(BasePlatformAdapter):
 
     def _reactions_enabled(self) -> bool:
         """Check if message reactions are enabled via config/env."""
-        return os.getenv("TELEGRAM_REACTIONS", "false").lower() not in ("false", "0", "no")
+        return os.getenv("TELEGRAM_REACTIONS", "false").lower() not in {"false", "0", "no"}
 
     async def _set_reaction(self, chat_id: str, message_id: str, emoji: str) -> bool:
         """Set a single emoji reaction on a Telegram message."""

gateway/platforms/telegram_network.py

@@ -59,7 +59,7 @@ class TelegramFallbackTransport(httpx.AsyncBaseTransport):
     """
 
     def __init__(self, fallback_ips: Iterable[str], **transport_kwargs):
-        self._fallback_ips = [ip for ip in dict.fromkeys(_normalize_fallback_ips(fallback_ips))]
+        self._fallback_ips = list(dict.fromkeys(_normalize_fallback_ips(fallback_ips)))
         proxy_url = _resolve_proxy_url(target_hosts=[_TELEGRAM_API_HOST, *self._fallback_ips])
         if proxy_url and "proxy" not in transport_kwargs:
             transport_kwargs["proxy"] = proxy_url

gateway/platforms/wecom.py

@@ -295,7 +295,7 @@ class WeComAdapter(BasePlatformAdapter):
 
         auth_payload = await self._wait_for_handshake(req_id)
         errcode = auth_payload.get("errcode", 0)
-        if errcode not in (0, None):
+        if errcode not in {0, None}:
             errmsg = auth_payload.get("errmsg", "authentication failed")
             raise RuntimeError(f"{errmsg} (errcode={errcode})")
 
@@ -320,7 +320,7 @@ class WeComAdapter(BasePlatformAdapter):
                 if self._payload_req_id(payload) == req_id:
                     return payload
                 logger.debug("[%s] Ignoring pre-auth payload: %s", self.name, payload.get("cmd"))
-            elif msg.type in (aiohttp.WSMsgType.CLOSED, aiohttp.WSMsgType.CLOSE, aiohttp.WSMsgType.ERROR):
+            elif msg.type in {aiohttp.WSMsgType.CLOSED, aiohttp.WSMsgType.CLOSE, aiohttp.WSMsgType.ERROR}:
                 raise RuntimeError("WeCom websocket closed during authentication")
 
     async def _listen_loop(self) -> None:
@@ -360,7 +360,7 @@ class WeComAdapter(BasePlatformAdapter):
                 payload = self._parse_json(msg.data)
                 if payload:
                     await self._dispatch_payload(payload)
-            elif msg.type in (aiohttp.WSMsgType.CLOSE, aiohttp.WSMsgType.CLOSED, aiohttp.WSMsgType.ERROR):
+            elif msg.type in {aiohttp.WSMsgType.CLOSE, aiohttp.WSMsgType.CLOSED, aiohttp.WSMsgType.ERROR}:
                 raise RuntimeError("WeCom websocket closed")
 
     async def _heartbeat_loop(self) -> None:
@@ -998,7 +998,7 @@ class WeComAdapter(BasePlatformAdapter):
     @staticmethod
     def _response_error(response: Dict[str, Any]) -> Optional[str]:
         errcode = response.get("errcode", 0)
-        if errcode in (0, None):
+        if errcode in {0, None}:
             return None
         errmsg = str(response.get("errmsg") or "unknown error")
         return f"WeCom errcode {errcode}: {errmsg}"

gateway/platforms/weixin.py

@@ -605,7 +605,7 @@ def _assert_weixin_cdn_url(url: str) -> None:
     except Exception as exc:  # noqa: BLE001
         raise ValueError(f"Unparseable media URL: {url!r}") from exc
 
-    if scheme not in ("http", "https"):
+    if scheme not in {"http", "https"}:
         raise ValueError(
             f"Media URL has disallowed scheme {scheme!r}; only http/https are permitted."
         )
@@ -983,7 +983,7 @@ def _extract_text(item_list: List[Dict[str, Any]]) -> str:
             ref = item.get("ref_msg") or {}
             ref_item = ref.get("message_item") or {}
             ref_type = ref_item.get("type")
-            if ref_type in (ITEM_IMAGE, ITEM_VIDEO, ITEM_FILE, ITEM_VOICE):
+            if ref_type in {ITEM_IMAGE, ITEM_VIDEO, ITEM_FILE, ITEM_VOICE}:
                 title = ref.get("title") or ""
                 prefix = f"[引用媒体: {title}]\n" if title else "[引用媒体]\n"
                 return f"{prefix}{text}".strip()
@@ -1331,7 +1331,7 @@ class WeixinAdapter(BasePlatformAdapter):
 
                 ret = response.get("ret", 0)
                 errcode = response.get("errcode", 0)
-                if ret not in (0, None) or errcode not in (0, None):
+                if ret not in {0, None} or errcode not in {0, None}:
                     if (ret == SESSION_EXPIRED_ERRCODE or errcode == SESSION_EXPIRED_ERRCODE
                             or _is_stale_session_ret(ret, errcode, response.get("errmsg"))):
                         logger.error("[%s] Session expired; pausing for 10 minutes", self.name)
@@ -1601,7 +1601,7 @@ class WeixinAdapter(BasePlatformAdapter):
                 if resp and isinstance(resp, dict):
                     ret = resp.get("ret")
                     errcode = resp.get("errcode")
-                    if (ret is not None and ret not in (0,)) or (errcode is not None and errcode not in (0,)):
+                    if (ret is not None and ret not in {0,}) or (errcode is not None and errcode not in {0,}):
                         is_session_expired = (
                             ret == SESSION_EXPIRED_ERRCODE
                             or errcode == SESSION_EXPIRED_ERRCODE

gateway/platforms/whatsapp.py

@@ -301,9 +301,9 @@ class WhatsAppAdapter(BasePlatformAdapter):
         configured = self.config.extra.get("require_mention")
         if configured is not None:
             if isinstance(configured, str):
-                return configured.lower() in ("true", "1", "yes", "on")
+                return configured.lower() in {"true", "1", "yes", "on"}
             return bool(configured)
-        return os.getenv("WHATSAPP_REQUIRE_MENTION", "false").lower() in ("true", "1", "yes", "on")
+        return os.getenv("WHATSAPP_REQUIRE_MENTION", "false").lower() in {"true", "1", "yes", "on"}
 
     def _whatsapp_free_response_chats(self) -> set[str]:
         raw = self.config.extra.get("free_response_chats")
@@ -679,7 +679,7 @@ class WhatsAppAdapter(BasePlatformAdapter):
         # getattr-with-default keeps tests that construct the adapter via
         # ``WhatsAppAdapter.__new__`` (bypassing __init__) working without
         # every _make_adapter() helper having to seed the attribute.
-        if getattr(self, "_shutting_down", False) and returncode in (0, -2, -15):
+        if getattr(self, "_shutting_down", False) and returncode in {0, -2, -15}:
             logger.info(
                 "[%s] Bridge exited during shutdown (code %d).",
                 self.name,
@@ -1183,7 +1183,7 @@ class WhatsAppAdapter(BasePlatformAdapter):
             if msg_type == MessageType.DOCUMENT and cached_urls:
                 for doc_path in cached_urls:
                     ext = Path(doc_path).suffix.lower()
-                    if ext in (".txt", ".md", ".csv", ".json", ".xml", ".yaml", ".yml", ".log", ".py", ".js", ".ts", ".html", ".css"):
+                    if ext in {".txt", ".md", ".csv", ".json", ".xml", ".yaml", ".yml", ".log", ".py", ".js", ".ts", ".html", ".css"}:
                         try:
                             file_size = Path(doc_path).stat().st_size
                             if file_size > MAX_TEXT_INJECT_BYTES:

gateway/platforms/yuanbao.py

@@ -2228,7 +2228,7 @@ class MediaResolveMiddleware(InboundMiddleware):
                 resp.raise_for_status()
                 payload = resp.json()
                 code = payload.get("code")
-                if code not in (None, 0):
+                if code not in {None, 0}:
                     raise RuntimeError(
                         f"resource/v1/download failed: code={code}, msg={payload.get('msg', '')}"
                     )
@@ -2391,7 +2391,7 @@ class MediaResolveMiddleware(InboundMiddleware):
                 rid = m.group(2)
                 kind, _, filename = head.partition(":")
                 kind = kind.strip()
-                if kind not in ("image", "file"):
+                if kind not in {"image", "file"}:
                     continue
                 if rid in seen:
                     continue
@@ -2993,10 +2993,10 @@ class ConnectionManager:
 
         # Fire-and-forget heartbeat ACKs — server always responds but callers don't
         # wait on these; silently discard to avoid "Unmatched Response" noise.
-        if cmd_type == CMD_TYPE["Response"] and cmd in (
+        if cmd_type == CMD_TYPE["Response"] and cmd in {
             "send_group_heartbeat",
             "send_private_heartbeat",
-        ):
+        }:
             logger.debug("[%s] Heartbeat ACK received: cmd=%s msg_id=%s", adapter.name, cmd, msg_id)
             return
 
@@ -3369,7 +3369,7 @@ class MediaSendHandler(ABC):
                 # Remove keys already passed explicitly to avoid "multiple values" TypeError
                 fwd_kwargs = {
                     k: v for k, v in kwargs.items()
-                    if k not in ("file_uuid", "filename", "content_type")
+                    if k not in {"file_uuid", "filename", "content_type"}
                 }
                 msg_body = self.build_msg_body(
                     upload_result,

gateway/platforms/yuanbao_media.py

@@ -150,7 +150,7 @@ def _parse_jpeg_size(buf: bytes) -> Optional[dict[str, int]]:
             i += 1
             continue
         marker = buf[i + 1]
-        if marker in (0xC0, 0xC2):
+        if marker in {0xC0, 0xC2}:
             h = struct.unpack(">H", buf[i + 5: i + 7])[0]
             w = struct.unpack(">H", buf[i + 7: i + 9])[0]
             return {"width": w, "height": h}
@@ -165,7 +165,7 @@ def _parse_gif_size(buf: bytes) -> Optional[dict[str, int]]:
     if len(buf) < 10:
         return None
     sig = buf[:6].decode("ascii", errors="replace")
-    if sig not in ("GIF87a", "GIF89a"):
+    if sig not in {"GIF87a", "GIF89a"}:
         return None
     w = struct.unpack("<H", buf[6:8])[0]
     h = struct.unpack("<H", buf[8:10])[0]

gateway/platforms/yuanbao_proto.py

@@ -702,7 +702,7 @@ def decode_inbound_push(data: bytes) -> Optional[dict]:
             "trace_id": trace_id,
         }
         # 过滤空值（保持 API 整洁）
-        return {k: v for k, v in result.items() if v or k in ("msg_body", "msg_seq")}
+        return {k: v for k, v in result.items() if v or k in {"msg_body", "msg_seq"}}
     except Exception as e:
         if DEBUG_MODE:
             logger.debug("[yuanbao_proto] decode_inbound_push failed: %s", e)

gateway/session.py

@@ -764,12 +764,12 @@ class SessionStore:
 
         now = _now()
 
-        if policy.mode in ("idle", "both"):
+        if policy.mode in {"idle", "both"}:
             idle_deadline = entry.updated_at + timedelta(minutes=policy.idle_minutes)
             if now > idle_deadline:
                 return True
 
-        if policy.mode in ("daily", "both"):
+        if policy.mode in {"daily", "both"}:
             today_reset = now.replace(
                 hour=policy.at_hour,
                 minute=0, second=0, microsecond=0,
@@ -805,12 +805,12 @@ class SessionStore:
         
         now = _now()
         
-        if policy.mode in ("idle", "both"):
+        if policy.mode in {"idle", "both"}:
             idle_deadline = entry.updated_at + timedelta(minutes=policy.idle_minutes)
             if now > idle_deadline:
                 return "idle"
         
-        if policy.mode in ("daily", "both"):
+        if policy.mode in {"daily", "both"}:
             today_reset = now.replace(
                 hour=policy.at_hour, 
                 minute=0, 
@@ -1276,9 +1276,14 @@ class SessionStore:
         
         # Also write legacy JSONL (keeps existing tooling working during transition)
         transcript_path = self.get_transcript_path(session_id)
-        with self._lock:
-            with open(transcript_path, "a", encoding="utf-8") as f:
-                f.write(json.dumps(message, ensure_ascii=False) + "\n")
+        try:
+            with self._lock:
+                with open(transcript_path, "a", encoding="utf-8") as f:
+                    f.write(json.dumps(message, ensure_ascii=False) + "\n")
+        except OSError as e:
+            # Disk full / read-only fs / permission errors must not crash the
+            # message handler — the SQLite write above is the primary store.
+            logger.debug("Failed to write JSONL transcript for %s: %s", session_id, e)
     
     def rewrite_transcript(self, session_id: str, messages: List[Dict[str, Any]]) -> None:
         """Replace the entire transcript for a session with new messages.

gateway/session_context.py

@@ -55,6 +55,7 @@ _SESSION_THREAD_ID: ContextVar = ContextVar("HERMES_SESSION_THREAD_ID", default=
 _SESSION_USER_ID: ContextVar = ContextVar("HERMES_SESSION_USER_ID", default=_UNSET)
 _SESSION_USER_NAME: ContextVar = ContextVar("HERMES_SESSION_USER_NAME", default=_UNSET)
 _SESSION_KEY: ContextVar = ContextVar("HERMES_SESSION_KEY", default=_UNSET)
+_SESSION_ID: ContextVar = ContextVar("HERMES_SESSION_ID", default=_UNSET)
 
 # Cron auto-delivery vars — set per-job in run_job() so concurrent jobs
 # don't clobber each other's delivery targets.
@@ -70,6 +71,7 @@ _VAR_MAP = {
     "HERMES_SESSION_USER_ID": _SESSION_USER_ID,
     "HERMES_SESSION_USER_NAME": _SESSION_USER_NAME,
     "HERMES_SESSION_KEY": _SESSION_KEY,
+    "HERMES_SESSION_ID": _SESSION_ID,
     "HERMES_CRON_AUTO_DELIVER_PLATFORM": _CRON_AUTO_DELIVER_PLATFORM,
     "HERMES_CRON_AUTO_DELIVER_CHAT_ID": _CRON_AUTO_DELIVER_CHAT_ID,
     "HERMES_CRON_AUTO_DELIVER_THREAD_ID": _CRON_AUTO_DELIVER_THREAD_ID,

gateway/shutdown_forensics.py

@@ -0,0 +1,462 @@
+"""Shutdown forensics — capture context when the gateway receives SIGTERM/SIGINT.
+
+The gateway's ``shutdown_signal_handler`` runs synchronously inside the
+asyncio event loop.  We can't safely block it for long, but we DO want a
+durable record of who/what triggered the shutdown so that "the gateway
+keeps dying" incidents can be diagnosed after the fact.
+
+This module exposes :func:`snapshot_shutdown_context`, a fast (<10ms),
+non-blocking probe that returns a structured dict the signal handler can
+log immediately, plus :func:`spawn_async_diagnostic`, a fire-and-forget
+``ps`` walk that runs as a detached subprocess so it can't block teardown
+even if /proc is wedged.
+
+Anything that needs to wait (e.g. shelling out to ``ps aux``) belongs in
+the async helper, never in the synchronous probe.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import signal
+import subprocess
+import sys
+import time
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+
+_SIGNAL_NAME_BY_NUM: Dict[int, str] = {}
+for _name in ("SIGTERM", "SIGINT", "SIGHUP", "SIGQUIT", "SIGUSR1", "SIGUSR2"):
+    _val = getattr(signal, _name, None)
+    if _val is not None:
+        _SIGNAL_NAME_BY_NUM[int(_val)] = _name
+
+
+def _signal_name(sig: Any) -> str:
+    """Return a human-readable signal name (or ``str(sig)`` as fallback)."""
+    if sig is None:
+        return "UNKNOWN"
+    try:
+        sig_int = int(sig)
+    except (TypeError, ValueError):
+        return str(sig)
+    return _SIGNAL_NAME_BY_NUM.get(sig_int, f"signal#{sig_int}")
+
+
+def _read_proc_field(pid: int, key: str) -> Optional[str]:
+    """Read a single field from /proc/<pid>/status.  Linux only; None elsewhere."""
+    try:
+        with open(f"/proc/{pid}/status", encoding="utf-8") as fh:
+            for line in fh:
+                if line.startswith(key + ":"):
+                    return line.split(":", 1)[1].strip()
+    except (FileNotFoundError, PermissionError, OSError):
+        pass
+    return None
+
+
+def _read_proc_cmdline(pid: int) -> Optional[str]:
+    """Read /proc/<pid>/cmdline as a printable string.  Linux only; None elsewhere."""
+    try:
+        with open(f"/proc/{pid}/cmdline", "rb") as fh:
+            data = fh.read()
+    except (FileNotFoundError, PermissionError, OSError):
+        return None
+    if not data:
+        return None
+    # cmdline uses NUL separators
+    return data.replace(b"\x00", b" ").decode("utf-8", errors="replace").strip()
+
+
+def _proc_summary(pid: int) -> Dict[str, Any]:
+    """Compact /proc/<pid> snapshot: pid, ppid, state, uid, cmdline.
+
+    Best-effort.  Missing fields are simply omitted rather than raising.
+    """
+    summary: Dict[str, Any] = {"pid": pid}
+    if pid <= 0:
+        return summary
+    name = _read_proc_field(pid, "Name")
+    if name is not None:
+        summary["name"] = name
+    state = _read_proc_field(pid, "State")
+    if state is not None:
+        summary["state"] = state
+    ppid = _read_proc_field(pid, "PPid")
+    if ppid is not None:
+        try:
+            summary["ppid"] = int(ppid)
+        except ValueError:
+            pass
+    uid = _read_proc_field(pid, "Uid")
+    if uid is not None:
+        # "real effective saved fs"
+        summary["uid"] = uid.split()[0] if uid else uid
+    cmdline = _read_proc_cmdline(pid)
+    if cmdline:
+        # Truncate aggressively — these can be 4KB
+        summary["cmdline"] = cmdline[:300]
+    return summary
+
+
+def snapshot_shutdown_context(received_signal: Any = None) -> Dict[str, Any]:
+    """Fast (<10ms) snapshot of who/what is asking us to shut down.
+
+    Captures:
+
+    * The signal number/name (so SIGINT vs SIGTERM is visible)
+    * Our own PID/ppid + parent process info from /proc (Linux)
+    * Whether systemd is our parent (``ppid==1`` or ``INVOCATION_ID`` set)
+    * Whether takeover/planned-stop markers exist (consumed lazily by the caller)
+    * /proc/self limits + load average (1-min)
+    * Wall-clock and monotonic timestamps for cross-correlating later phases
+
+    Pure stdlib, never raises, never blocks on subprocesses.
+    """
+    now = time.time()
+    monotonic = time.monotonic()
+    pid = os.getpid()
+    ppid = os.getppid()
+
+    ctx: Dict[str, Any] = {
+        "ts": now,
+        "ts_monotonic": monotonic,
+        "signal": _signal_name(received_signal),
+        "signal_num": int(received_signal) if received_signal is not None else None,
+        "pid": pid,
+        "ppid": ppid,
+        "parent": _proc_summary(ppid),
+        "self": _proc_summary(pid),
+    }
+
+    # systemd context.  If we were started by a systemd unit, INVOCATION_ID
+    # is set in our env.  ppid==1 (init) is also a strong signal that
+    # systemd reaped+forwarded the SIGTERM.
+    invocation_id = os.environ.get("INVOCATION_ID")
+    if invocation_id:
+        ctx["systemd_invocation_id"] = invocation_id
+    journal_stream = os.environ.get("JOURNAL_STREAM")
+    if journal_stream:
+        ctx["systemd_journal_stream"] = journal_stream
+    ctx["under_systemd"] = bool(invocation_id) or ppid == 1
+
+    # Load average — high load points the finger at "something else
+    # crushing the box" rather than "external killer".
+    try:
+        ctx["loadavg_1m"] = os.getloadavg()[0]
+    except (OSError, AttributeError):
+        pass
+
+    # /proc/self/status TracerPid: nonzero means a debugger / strace is
+    # attached.  Useful when "phantom SIGKILL" turns out to be a manual
+    # gdb session.
+    try:
+        tracer = _read_proc_field(pid, "TracerPid")
+        if tracer is not None and tracer != "0":
+            ctx["tracer_pid"] = int(tracer) if tracer.isdigit() else tracer
+            ctx["tracer"] = _proc_summary(int(tracer)) if tracer.isdigit() else None
+    except (TypeError, ValueError):
+        pass
+
+    # Race-detection hint: did somebody recently start a sibling gateway
+    # with --replace?  We can't see the new process directly here, but if
+    # there's a takeover marker on disk that DOESN'T name us, that's a
+    # smoking gun for "another --replace instance is killing us".
+    # Filenames mirror gateway.status (._TAKEOVER_MARKER_FILENAME /
+    # _PLANNED_STOP_MARKER_FILENAME); we use string literals here so the
+    # signal-handler path stays import-light.
+    try:
+        hermes_home_str = os.environ.get("HERMES_HOME")
+        if hermes_home_str:
+            takeover_path = Path(hermes_home_str) / ".gateway-takeover.json"
+            if takeover_path.exists():
+                try:
+                    raw = takeover_path.read_text(encoding="utf-8")
+                    ctx["takeover_marker"] = raw[:300]
+                    ctx["takeover_marker_for_self"] = (
+                        f'"target_pid": {pid}' in raw
+                        or f"'target_pid': {pid}" in raw
+                    )
+                except OSError:
+                    pass
+            planned_stop_path = Path(hermes_home_str) / ".gateway-planned-stop.json"
+            if planned_stop_path.exists():
+                try:
+                    raw = planned_stop_path.read_text(encoding="utf-8")
+                    ctx["planned_stop_marker"] = raw[:300]
+                except OSError:
+                    pass
+    except Exception:  # noqa: BLE001 — never raise from a signal handler
+        pass
+
+    return ctx
+
+
+def spawn_async_diagnostic(
+    log_path: Path,
+    signal_name: str,
+    *,
+    timeout_seconds: float = 5.0,
+) -> Optional[int]:
+    """Fire-and-forget ``ps``-style snapshot written to ``log_path``.
+
+    Runs as a detached subprocess so it can't block the asyncio event loop
+    or compete with platform teardown.  The subprocess uses its own
+    ``timeout`` so a wedged ``ps`` still self-cleans within
+    ``timeout_seconds``.
+
+    Returns the subprocess PID on success, ``None`` on failure.  Never
+    raises.
+
+    We deliberately avoid ``subprocess.run(["ps", "aux"])`` from inside the
+    signal handler (the pre-existing pattern): on a busy host with hundreds
+    of processes, ``ps aux`` can take >2s to walk /proc, during which the
+    asyncio loop is frozen and adapter teardown can't begin.
+    """
+    try:
+        log_path.parent.mkdir(parents=True, exist_ok=True)
+    except OSError:
+        return None
+
+    # Inline shell so we don't have to ship a helper script.  bash -c is
+    # available on every POSIX target we support; on Windows we just skip
+    # the snapshot (the platform doesn't ship ps anyway).
+    if sys.platform == "win32":
+        return None
+
+    script = (
+        f"echo '=== shutdown diagnostic @ {signal_name} ==='; "
+        "echo '--- date ---'; date -u +%Y-%m-%dT%H:%M:%SZ; "
+        "echo '--- ps auxf (top 60 by cpu) ---'; "
+        "ps auxf --sort=-pcpu 2>/dev/null | head -60; "
+        "echo '--- pstree of self ---'; "
+        f"pstree -plau {os.getpid()} 2>/dev/null | head -40 || true; "
+        "echo '--- /proc/loadavg ---'; "
+        "cat /proc/loadavg 2>/dev/null || true; "
+        "echo '--- recent dmesg (oom/killed) ---'; "
+        "dmesg -T 2>/dev/null | tail -20 || journalctl --user -n 20 --no-pager 2>/dev/null | tail -20 || true; "
+        "echo '=== end ==='"
+    )
+
+    try:
+        # Open the log file in append mode and let the subprocess inherit.
+        # We use os.O_APPEND so concurrent diagnostics from rapid signals
+        # don't trample each other.
+        fd = os.open(str(log_path), os.O_WRONLY | os.O_CREAT | os.O_APPEND, 0o644)
+    except OSError:
+        return None
+
+    try:
+        # Detach from our process group so the subprocess survives even
+        # if systemd kills our cgroup with KillMode=control-group (which
+        # would also reap us anyway, but defense in depth).  Without
+        # start_new_session, a SIGKILL on our cgroup takes the diag down
+        # before it can flush.
+        proc = subprocess.Popen(
+            ["timeout", f"{timeout_seconds:.0f}", "bash", "-c", script],
+            stdout=fd,
+            stderr=subprocess.STDOUT,
+            stdin=subprocess.DEVNULL,
+            start_new_session=True,
+            close_fds=True,
+        )
+    except (FileNotFoundError, OSError):
+        try:
+            os.close(fd)
+        except OSError:
+            pass
+        return None
+    finally:
+        # Subprocess inherited the fd; we can drop our handle.
+        try:
+            os.close(fd)
+        except OSError:
+            pass
+
+    return proc.pid
+
+
+def format_context_for_log(ctx: Dict[str, Any]) -> str:
+    """Render a shutdown context dict as a single, scannable log line."""
+    sig = ctx.get("signal", "?")
+    parent = ctx.get("parent") or {}
+    parent_cmd = parent.get("cmdline", "(unknown)")
+    parent_name = parent.get("name") or "?"
+    parent_pid = parent.get("pid") or "?"
+    under_systemd = "yes" if ctx.get("under_systemd") else "no"
+    load = ctx.get("loadavg_1m")
+    load_str = f"{load:.2f}" if isinstance(load, (int, float)) else "?"
+    extras: List[str] = []
+    if ctx.get("takeover_marker") is not None:
+        for_self = ctx.get("takeover_marker_for_self")
+        extras.append(
+            f"takeover_marker_present={'self' if for_self else 'other'}"
+        )
+    if ctx.get("planned_stop_marker") is not None:
+        extras.append("planned_stop_marker_present=yes")
+    if ctx.get("tracer_pid"):
+        extras.append(f"tracer_pid={ctx['tracer_pid']}")
+    extras_str = (" " + " ".join(extras)) if extras else ""
+    # Parent cmdline is the most useful single signal — log it prominently.
+    return (
+        f"signal={sig} "
+        f"under_systemd={under_systemd} "
+        f"parent_pid={parent_pid} "
+        f"parent_name={parent_name} "
+        f"loadavg_1m={load_str}"
+        f"{extras_str} "
+        f"parent_cmdline={parent_cmd!r}"
+    )
+
+
+def context_as_json(ctx: Dict[str, Any]) -> str:
+    """JSON-serialise a context dict for structured ingestion.  Never raises."""
+    try:
+        return json.dumps(ctx, default=str, sort_keys=True)
+    except (TypeError, ValueError):
+        return "{}"
+
+
+def check_systemd_timing_alignment(drain_timeout: float) -> Optional[Dict[str, Any]]:
+    """At startup, sanity-check that systemd's TimeoutStopSec >= drain_timeout.
+
+    When the gateway is run under a stale systemd unit file (e.g. the user
+    upgraded hermes-agent but never re-ran ``hermes setup`` to regenerate
+    the unit), ``TimeoutStopSec`` can be smaller than the configured
+    ``restart_drain_timeout``.  Result: SIGTERM arrives, the drain starts,
+    and systemd SIGKILLs the cgroup mid-drain — looks like a phantom kill
+    in the journal because the journal only logs ``code=killed status=9``.
+
+    Returns ``None`` when the alignment is fine OR we can't determine it
+    (not running under systemd, ``systemctl`` unavailable, etc.).  Returns
+    a dict with ``timeout_stop_sec`` + ``drain_timeout`` + ``mismatch``
+    bool when we have data to report.
+
+    Best-effort.  Never raises.
+    """
+    invocation_id = os.environ.get("INVOCATION_ID")
+    if not invocation_id:
+        return None  # Not running under systemd (or at least not directly)
+
+    # Try to identify our unit name and ask systemctl for its config.
+    unit_name: Optional[str] = None
+    try:
+        # /proc/self/cgroup gives us "0::/user.slice/.../hermes-gateway.service"
+        with open("/proc/self/cgroup", encoding="utf-8") as fh:
+            for line in fh:
+                # systemd cgroup line ends with the unit name
+                if ".service" in line:
+                    parts = line.strip().split("/")
+                    for p in reversed(parts):
+                        if p.endswith(".service"):
+                            unit_name = p
+                            break
+                    if unit_name:
+                        break
+    except (OSError, FileNotFoundError):
+        pass
+    if not unit_name:
+        return None
+
+    # Query systemctl for TimeoutStopUSec.  Use --user OR system depending
+    # on which manager actually owns the unit.  Try user first since
+    # that's the common case for hermes.
+    timeout_us: Optional[int] = None
+    for flag in (["--user"], []):
+        try:
+            result = subprocess.run(
+                ["systemctl", *flag, "show", unit_name, "--property=TimeoutStopUSec"],
+                capture_output=True, text=True, timeout=2.0,
+            )
+        except (FileNotFoundError, subprocess.TimeoutExpired, OSError):
+            continue
+        if result.returncode != 0:
+            continue
+        # Output: "TimeoutStopUSec=1min 30s" or "TimeoutStopUSec=90000000"
+        for line in result.stdout.splitlines():
+            if line.startswith("TimeoutStopUSec="):
+                value = line.split("=", 1)[1].strip()
+                # Try numeric microseconds first
+                if value.isdigit():
+                    timeout_us = int(value)
+                else:
+                    timeout_us = _parse_systemd_duration_to_us(value)
+                if timeout_us is not None:
+                    break
+        if timeout_us is not None:
+            break
+
+    if timeout_us is None:
+        return None
+
+    timeout_stop_sec = timeout_us / 1_000_000.0
+    # systemd needs headroom for: post-interrupt kill, adapter disconnect,
+    # SessionDB close, file unlinks, etc.  30s matches the unit-template
+    # constant in hermes_cli/gateway.py.
+    headroom = 30.0
+    expected = drain_timeout + headroom
+    return {
+        "unit": unit_name,
+        "timeout_stop_sec": timeout_stop_sec,
+        "drain_timeout": drain_timeout,
+        "expected_min": expected,
+        "mismatch": timeout_stop_sec < expected,
+    }
+
+
+def _parse_systemd_duration_to_us(raw: str) -> Optional[int]:
+    """Parse 'TimeoutStopUSec=1min 30s' / '90s' style values to microseconds.
+
+    systemd accepts a wide grammar; we cover the common cases (s, ms, min,
+    h) and return None on anything unexpected.  Never raises.
+    """
+    if not raw:
+        return None
+    units = {
+        "us": 1,
+        "ms": 1_000,
+        "s": 1_000_000,
+        "sec": 1_000_000,
+        "min": 60_000_000,
+        "h": 3_600_000_000,
+        "hr": 3_600_000_000,
+    }
+    total_us = 0
+    token = ""
+    digits = ""
+    for ch in raw + " ":
+        if ch.isdigit() or ch == ".":
+            if token:
+                # End previous unit, start new number
+                multiplier = units.get(token.lower())
+                if multiplier is None or not digits:
+                    return None
+                try:
+                    total_us += int(float(digits) * multiplier)
+                except ValueError:
+                    return None
+                digits = ""
+                token = ""
+            digits += ch
+        elif ch.isalpha():
+            token += ch
+        elif digits and token:
+            multiplier = units.get(token.lower())
+            if multiplier is None:
+                return None
+            try:
+                total_us += int(float(digits) * multiplier)
+            except ValueError:
+                return None
+            digits = ""
+            token = ""
+        elif digits and not token:
+            # Bare number = seconds (rare but valid)
+            try:
+                total_us += int(float(digits) * 1_000_000)
+            except ValueError:
+                return None
+            digits = ""
+    return total_us if total_us > 0 else None

gateway/slash_access.py

@@ -0,0 +1,229 @@
+"""Per-platform slash command access control.
+
+This module sits beside the existing per-platform allowlist (``allow_from``)
+and adds a second axis: of the users who are *allowed to talk to the
+gateway*, which ones can run *which slash commands*.
+
+Two lists per platform scope (DM vs group, mirroring ``allow_from`` vs
+``group_allow_from``):
+
+  - ``allow_admin_from``      — user IDs that get every registered slash
+                                command (built-in + plugin-registered).
+  - ``user_allowed_commands`` — slash command names non-admin users may
+                                run. Empty / unset → non-admins get no
+                                slash commands.
+
+Backward compatibility:
+
+  If ``allow_admin_from`` is not set for a scope, slash command gating
+  is disabled entirely for that scope. Every allowed user can run every
+  slash command, exactly like before. This means existing installs are
+  unaffected until an operator opts in by listing at least one admin.
+
+The gate is applied at the slash command dispatch site in
+``gateway/run.py`` so it covers BOTH built-in and plugin-registered
+commands via the live registry. Gating slash commands does not affect
+plain chat — non-admin users can still talk to the agent normally,
+they just can't trigger commands outside ``user_allowed_commands``.
+
+Authored as a slimmed-down salvage of PR #4443's permission tiers
+(co-authored by @ReqX). The full tier system, audit log, usage
+tracking, rate limiting, and tool filtering from that PR are not
+included here — only the slash-command access split.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, FrozenSet, Iterable, Optional, Tuple
+
+
+# Slash commands that MUST stay reachable for any allowed user, even when
+# slash gating is enabled and the user has no commands listed. Without this
+# carve-out, a non-admin user has no way to discover what they can or
+# can't do (``/help``, ``/whoami``) and no way to see what state the agent
+# is in (``/status``). These mirror the smallest set of read-only commands
+# we'd hand to a guest. Operators can still narrow this further by writing
+# their own ``user_allowed_commands`` (this set is only the implicit
+# fallback floor — anything in ``user_allowed_commands`` overrides it
+# additively, never restrictively).
+_ALWAYS_ALLOWED_FOR_USERS: FrozenSet[str] = frozenset({
+    "help",
+    "whoami",
+})
+
+
+@dataclass(frozen=True)
+class SlashAccessPolicy:
+    """Resolved access policy for a single (platform, scope) pair.
+
+    ``scope`` is ``"dm"`` for direct messages and ``"group"`` for groups,
+    channels, threads, and any other multi-user context. The mapping from
+    SessionSource.chat_type → scope happens in ``policy_for_source``.
+    """
+
+    enabled: bool                      # gating active for this scope?
+    admin_user_ids: FrozenSet[str]
+    user_allowed_commands: FrozenSet[str]
+
+    def is_admin(self, user_id: Optional[str]) -> bool:
+        if not self.enabled:
+            # Gating disabled → treat every allowed user as admin so
+            # downstream code can keep using ``is_admin`` / ``can_run``
+            # uniformly.
+            return True
+        if not user_id:
+            return False
+        return str(user_id) in self.admin_user_ids
+
+    def can_run(self, user_id: Optional[str], canonical_cmd: str) -> bool:
+        if not self.enabled:
+            return True
+        if self.is_admin(user_id):
+            return True
+        if not canonical_cmd:
+            return False
+        if canonical_cmd in _ALWAYS_ALLOWED_FOR_USERS:
+            return True
+        return canonical_cmd in self.user_allowed_commands
+
+
+_DM_CHAT_TYPES = frozenset({"dm", "direct", "private", ""})
+
+
+def _coerce_id_list(raw: Any) -> FrozenSet[str]:
+    """Normalize a YAML-loaded admin/user list into a frozenset of strings.
+
+    Accepts ``None``, list, tuple, or comma-separated string. Stringifies
+    each entry and strips whitespace; empty entries are dropped.
+    """
+    if raw is None:
+        return frozenset()
+    if isinstance(raw, (list, tuple, set, frozenset)):
+        items: Iterable[Any] = raw
+    elif isinstance(raw, str):
+        items = (s for s in raw.split(",") if s.strip())
+    else:
+        # single scalar (int user id, etc.)
+        items = (raw,)
+    out: list[str] = []
+    for it in items:
+        s = str(it).strip()
+        if s:
+            out.append(s)
+    return frozenset(out)
+
+
+def _coerce_command_list(raw: Any) -> FrozenSet[str]:
+    """Normalize a slash command allowlist.
+
+    Strips leading slashes so YAML can read either ``["help", "status"]``
+    or ``["/help", "/status"]``. Lowercase canonicalization matches how
+    ``resolve_command()`` stores names.
+    """
+    if raw is None:
+        return frozenset()
+    if isinstance(raw, (list, tuple, set, frozenset)):
+        items: Iterable[Any] = raw
+    elif isinstance(raw, str):
+        items = (s for s in raw.split(",") if s.strip())
+    else:
+        items = (raw,)
+    out: list[str] = []
+    for it in items:
+        s = str(it).strip().lstrip("/").lower()
+        if s:
+            out.append(s)
+    return frozenset(out)
+
+
+def _scope_for_chat_type(chat_type: Optional[str]) -> str:
+    if chat_type and chat_type.lower() in _DM_CHAT_TYPES:
+        return "dm"
+    return "group"
+
+
+def _platform_extra(platform_config: Any) -> dict:
+    """Return the ``extra`` dict from a PlatformConfig-like object.
+
+    Defensively handles None and non-PlatformConfig shapes so calling
+    code can stay simple.
+    """
+    if platform_config is None:
+        return {}
+    extra = getattr(platform_config, "extra", None)
+    if isinstance(extra, dict):
+        return extra
+    if isinstance(platform_config, dict):
+        # Some test harnesses pass dicts directly.
+        return platform_config
+    return {}
+
+
+def _keys_for_scope(scope: str) -> Tuple[str, str]:
+    """Return (admin_key, user_cmd_key) names for a scope."""
+    if scope == "group":
+        return ("group_allow_admin_from", "group_user_allowed_commands")
+    return ("allow_admin_from", "user_allowed_commands")
+
+
+def policy_from_extra(extra: dict, scope: str) -> SlashAccessPolicy:
+    """Build a policy from a platform's ``extra`` dict for one scope.
+
+    DM scope falls back to group scope keys ONLY for ``user_allowed_commands``
+    when the DM scope didn't specify its own. This keeps the common case
+    (operator wants the same command set DM and group) ergonomic without
+    forcing duplication. Admin lists are NOT cross-scope: an admin in
+    DMs is not implicitly an admin in a group.
+    """
+    admin_key, cmd_key = _keys_for_scope(scope)
+    admin_ids = _coerce_id_list(extra.get(admin_key))
+    cmds = _coerce_command_list(extra.get(cmd_key))
+
+    if scope == "dm" and not cmds:
+        # DM didn't specify — let group's user_allowed_commands fall through
+        # so operators only need to list it once if it's the same.
+        cmds = _coerce_command_list(extra.get("group_user_allowed_commands"))
+
+    enabled = bool(admin_ids)
+    return SlashAccessPolicy(
+        enabled=enabled,
+        admin_user_ids=admin_ids,
+        user_allowed_commands=cmds,
+    )
+
+
+def policy_for_source(gateway_config: Any, source: Any) -> SlashAccessPolicy:
+    """Resolve the access policy for a SessionSource.
+
+    Returns a "disabled" policy (gating off, allow everything) when:
+      - gateway_config is None
+      - the platform has no PlatformConfig
+      - the platform's PlatformConfig has no admin list set for the scope
+
+    Callers should treat the returned policy as authoritative for slash
+    command gating only. It does not gate plain chat messages.
+    """
+    if gateway_config is None or source is None:
+        return SlashAccessPolicy(
+            enabled=False,
+            admin_user_ids=frozenset(),
+            user_allowed_commands=frozenset(),
+        )
+    platforms = getattr(gateway_config, "platforms", None)
+    platform_config = None
+    if platforms is not None:
+        try:
+            platform_config = platforms.get(source.platform)
+        except Exception:
+            platform_config = None
+    extra = _platform_extra(platform_config)
+    scope = _scope_for_chat_type(getattr(source, "chat_type", None))
+    return policy_from_extra(extra, scope)
+
+
+__all__ = [
+    "SlashAccessPolicy",
+    "policy_from_extra",
+    "policy_for_source",
+]

gateway/status.py

@@ -218,7 +218,11 @@ def _read_pid_record(pid_path: Optional[Path] = None) -> Optional[dict]:
     if not pid_path.exists():
         return None
 
-    raw = pid_path.read_text().strip()
+    try:
+        raw = pid_path.read_text().strip()
+    except OSError:
+        # File was deleted between exists() and read_text(), or permission flipped.
+        return None
     if not raw:
         return None
 
@@ -482,10 +486,12 @@ def write_runtime_status(
     """Persist gateway runtime health information for diagnostics/status."""
     path = _get_runtime_status_path()
     payload = _read_json_file(path) or _build_runtime_status_record()
+    current_record = _build_pid_record()
     payload.setdefault("platforms", {})
-    payload.setdefault("kind", _GATEWAY_KIND)
-    payload["pid"] = os.getpid()
-    payload["start_time"] = _get_process_start_time(os.getpid())
+    payload["kind"] = current_record["kind"]
+    payload["pid"] = current_record["pid"]
+    payload["argv"] = current_record["argv"]
+    payload["start_time"] = current_record["start_time"]
     payload["updated_at"] = _utc_now_iso()
 
     if gateway_state is not _UNSET:
@@ -598,7 +604,7 @@ def acquire_scoped_lock(scope: str, identity: str, metadata: Optional[dict[str,
                             for _line in _proc_status.read_text(encoding="utf-8").splitlines():
                                 if _line.startswith("State:"):
                                     _state = _line.split()[1]
-                                    if _state in ("T", "t"):  # stopped or tracing stop
+                                    if _state in {"T", "t"}:  # stopped or tracing stop
                                         stale = True
                                     break
                     except (OSError, PermissionError):

gateway/stream_consumer.py

@@ -21,7 +21,15 @@ import queue
 import re
 import time
 from dataclasses import dataclass
-from typing import Any, Optional
+from typing import Any, Callable, Optional
+
+from gateway.platforms.base import BasePlatformAdapter as _BasePlatformAdapter
+from gateway.platforms.base import _custom_unit_to_cp
+from gateway.config import (
+    DEFAULT_STREAMING_EDIT_INTERVAL as _DEFAULT_STREAMING_EDIT_INTERVAL,
+    DEFAULT_STREAMING_BUFFER_THRESHOLD as _DEFAULT_STREAMING_BUFFER_THRESHOLD,
+    DEFAULT_STREAMING_CURSOR as _DEFAULT_STREAMING_CURSOR,
+)
 
 logger = logging.getLogger("gateway.stream_consumer")
 
@@ -40,9 +48,9 @@ _COMMENTARY = object()
 @dataclass
 class StreamConsumerConfig:
     """Runtime config for a single stream consumer instance."""
-    edit_interval: float = 1.0
-    buffer_threshold: int = 40
-    cursor: str = " ▉"
+    edit_interval: float = _DEFAULT_STREAMING_EDIT_INTERVAL
+    buffer_threshold: int = _DEFAULT_STREAMING_BUFFER_THRESHOLD
+    cursor: str = _DEFAULT_STREAMING_CURSOR
     buffer_only: bool = False
     # When >0, the final edit for a streamed response is delivered as a
     # fresh message if the original preview has been visible for at least
@@ -52,6 +60,18 @@ class StreamConsumerConfig:
     # openclaw/openclaw#72038.  Default 0 = always edit in place (legacy
     # behavior).  The gateway enables this selectively per-platform.
     fresh_final_after_seconds: float = 0.0
+    # Streaming transport selection:
+    #   "auto"  — prefer native draft streaming (e.g. Telegram sendMessageDraft)
+    #             when the adapter + chat supports it; fall back to edit.
+    #   "draft" — explicitly request native draft streaming; fall back to
+    #             edit when unsupported.
+    #   "edit"  — progressive editMessageText (legacy behavior).
+    #   "off"   — handled by the gateway before the consumer is even built.
+    transport: str = "auto"
+    # Hint for the consumer about the originating chat type (e.g. "dm",
+    # "group", "supergroup", "forum").  Used to gate native draft streaming,
+    # which is platform-specific (Telegram drafts are DM-only).
+    chat_type: str = ""
 
 
 class GatewayStreamConsumer:
@@ -85,6 +105,11 @@ class GatewayStreamConsumer:
         "</THINKING>", "</thinking>", "</thought>",
     )
 
+    # Class-wide monotonic counter for native-streaming draft ids.  Telegram
+    # animates a draft when the same draft_id is reused across consecutive
+    # calls in the same chat, so we need a fresh non-zero id per response.
+    _draft_id_counter: int = 0
+
     def __init__(
         self,
         adapter: Any,
@@ -92,6 +117,7 @@ class GatewayStreamConsumer:
         config: Optional[StreamConsumerConfig] = None,
         metadata: Optional[dict] = None,
         on_new_message: Optional[callable] = None,
+        initial_reply_to_id: Optional[str] = None,
     ):
         self.adapter = adapter
         self.chat_id = chat_id
@@ -105,6 +131,7 @@ class GatewayStreamConsumer:
         # the content, not edit the old bubble above it.
         # Called with no arguments. Exceptions are swallowed.
         self._on_new_message = on_new_message
+        self._initial_reply_to_id = initial_reply_to_id
         self._queue: queue.Queue = queue.Queue()
         self._accumulated = ""
         self._message_id: Optional[str] = None
@@ -136,6 +163,20 @@ class GatewayStreamConsumer:
         self._in_think_block = False
         self._think_buffer = ""
 
+        # Native draft-streaming state.  Resolved at the start of run() based
+        # on cfg.transport, cfg.chat_type, and the adapter's
+        # supports_draft_streaming() probe.  When True, the consumer emits
+        # animated draft frames via adapter.send_draft instead of progressive
+        # edits via adapter.edit_message.  The final answer still goes
+        # through the normal first-send path so the user gets a real message
+        # in their chat history (drafts have no message_id).
+        self._use_draft_streaming = False
+        self._draft_id: Optional[int] = None
+        # Cumulative draft-frame failure count for this consumer.  After the
+        # first failure we permanently disable drafts for the remainder of
+        # this response and route through edit-based for graceful degradation.
+        self._draft_failures = 0
+
     @property
     def already_sent(self) -> bool:
         """True if at least one message was sent or edited during the run."""
@@ -174,6 +215,16 @@ class GatewayStreamConsumer:
         self._last_sent_text = ""
         self._fallback_final_send = False
         self._fallback_prefix = ""
+        # Native draft streaming: bump the draft_id so the next text segment
+        # animates as a fresh preview below the tool-progress bubbles, not
+        # over the prior segment's already-finalized draft.  This is how
+        # we avoid the "inter-tool-call text leak" failure mode openclaw
+        # documented in their issue #32535 — each text block becomes its
+        # own visible message via the finalize, then a new draft animates
+        # for the next one.
+        if self._use_draft_streaming:
+            type(self)._draft_id_counter += 1
+            self._draft_id = type(self)._draft_id_counter
 
     def on_delta(self, text: str) -> None:
         """Thread-safe callback — called from the agent's worker thread.
@@ -299,9 +350,32 @@ class GatewayStreamConsumer:
 
     async def run(self) -> None:
         """Async task that drains the queue and edits the platform message."""
-        # Platform message length limit — leave room for cursor + formatting
+        # Platform message length limit — leave room for cursor + formatting.
+        # Use the adapter's length function (e.g. utf16_len for Telegram) so
+        # overflow detection matches what the platform actually enforces.
+        # Gate on isinstance(BasePlatformAdapter) so test MagicMocks (whose
+        # auto-attributes return mock objects, not callables) fall back to len.
+        _len_fn: "Callable[[str], int]" = (
+            self.adapter.message_len_fn
+            if isinstance(self.adapter, _BasePlatformAdapter)
+            else len
+        )
         _raw_limit = getattr(self.adapter, "MAX_MESSAGE_LENGTH", 4096)
-        _safe_limit = max(500, _raw_limit - len(self.cfg.cursor) - 100)
+        _safe_limit = max(500, _raw_limit - _len_fn(self.cfg.cursor) - 100)
+
+        # Resolve native draft streaming once per run.  When enabled the
+        # consumer routes mid-stream frames through adapter.send_draft and
+        # leaves _message_id=None so the existing got_done path delivers the
+        # final answer as a regular sendMessage (drafts have no message_id
+        # to edit).
+        self._use_draft_streaming = self._resolve_draft_streaming()
+        if self._use_draft_streaming:
+            type(self)._draft_id_counter += 1
+            self._draft_id = type(self)._draft_id_counter
+            logger.debug(
+                "Stream consumer using native-draft transport (chat=%s draft_id=%s)",
+                self.chat_id, self._draft_id,
+            )
 
         try:
             while True:
@@ -343,6 +417,10 @@ class GatewayStreamConsumer:
                     should_edit = should_edit or (
                         (elapsed >= self._current_edit_interval
                             and self._accumulated)
+                        # buffer_threshold is intentionally codepoint-based:
+                        # it's a debounce heuristic ("send updates roughly
+                        # every N visible characters"), not a platform-limit
+                        # check. _len_fn is reserved for overflow detection.
                         or len(self._accumulated) >= self.cfg.buffer_threshold
                     )
 
@@ -351,7 +429,7 @@ class GatewayStreamConsumer:
                     # Split overflow: if accumulated text exceeds the platform
                     # limit, split into properly sized chunks.
                     if (
-                        len(self._accumulated) > _safe_limit
+                        _len_fn(self._accumulated) > _safe_limit
                         and self._message_id is None
                     ):
                         # No existing message to edit (first message or after a
@@ -360,15 +438,23 @@ class GatewayStreamConsumer:
                         # proper word/code-fence boundaries and chunk
                         # indicators like "(1/2)".
                         chunks = self.adapter.truncate_message(
-                            self._accumulated, _safe_limit
+                            self._accumulated, _safe_limit, len_fn=_len_fn,
                         )
+                        chunks_delivered = False
+                        reply_to = self._message_id or self._initial_reply_to_id
                         for chunk in chunks:
-                            await self._send_new_chunk(chunk, self._message_id)
+                            new_id = await self._send_new_chunk(chunk, reply_to)
+                            if new_id is not None and new_id != reply_to:
+                                chunks_delivered = True
                         self._accumulated = ""
                         self._last_sent_text = ""
                         self._last_edit_time = time.monotonic()
                         if got_done:
-                            self._final_response_sent = self._already_sent
+                            # Only claim final delivery if THESE chunks actually
+                            # landed.  ``_already_sent`` may be True from prior
+                            # tool-progress edits or fallback-mode promotion (#10748)
+                            # — that doesn't mean the final answer reached the user.
+                            self._final_response_sent = chunks_delivered
                             return
                         if got_segment_break:
                             self._message_id = None
@@ -379,11 +465,14 @@ class GatewayStreamConsumer:
                     # Existing message: edit it with the first chunk, then
                     # start a new message for the overflow remainder.
                     while (
-                        len(self._accumulated) > _safe_limit
+                        _len_fn(self._accumulated) > _safe_limit
                         and self._message_id is not None
                         and self._edit_supported
                     ):
-                        split_at = self._accumulated.rfind("\n", 0, _safe_limit)
+                        _cp_budget = _custom_unit_to_cp(
+                            self._accumulated, _safe_limit, _len_fn,
+                        )
+                        split_at = self._accumulated.rfind("\n", 0, _cp_budget)
                         if split_at < _safe_limit // 2:
                             split_at = _safe_limit
                         chunk = self._accumulated[:split_at]
@@ -411,7 +500,7 @@ class GatewayStreamConsumer:
                     # path below so we don't finalize here for it.
                     current_update_visible = await self._send_or_edit(
                         display_text,
-                        finalize=got_segment_break,
+                        finalize=(got_done or got_segment_break),
                     )
                     self._last_edit_time = time.monotonic()
 
@@ -574,14 +663,18 @@ class GatewayStreamConsumer:
         return final_text
 
     @staticmethod
-    def _split_text_chunks(text: str, limit: int) -> list[str]:
+    def _split_text_chunks(
+        text: str, limit: int,
+        len_fn: "Callable[[str], int]" = len,
+    ) -> list[str]:
         """Split text into reasonably sized chunks for fallback sends."""
-        if len(text) <= limit:
+        if len_fn(text) <= limit:
             return [text]
         chunks: list[str] = []
         remaining = text
-        while len(remaining) > limit:
-            split_at = remaining.rfind("\n", 0, limit)
+        while len_fn(remaining) > limit:
+            _cp_budget = _custom_unit_to_cp(remaining, limit, len_fn)
+            split_at = remaining.rfind("\n", 0, _cp_budget)
             if split_at < limit // 2:
                 split_at = limit
             chunks.append(remaining[:split_at])
@@ -637,9 +730,15 @@ class GatewayStreamConsumer:
                 return
 
         raw_limit = getattr(self.adapter, "MAX_MESSAGE_LENGTH", 4096)
+        _len_fn: "Callable[[str], int]" = (
+            self.adapter.message_len_fn
+            if isinstance(self.adapter, _BasePlatformAdapter)
+            else len
+        )
         safe_limit = max(500, raw_limit - 100)
-        chunks = self._split_text_chunks(continuation, safe_limit)
+        chunks = self._split_text_chunks(continuation, safe_limit, len_fn=_len_fn)
 
+        stale_message_id = self._message_id  # partial message to clean up
         last_message_id: Optional[str] = None
         last_successful_chunk = ""
         sent_any_chunk = False
@@ -687,6 +786,22 @@ class GatewayStreamConsumer:
             # so any stale tool-progress bubble gets closed off.
             self._notify_new_message()
 
+        # Remove the frozen partial message so the user only sees the
+        # complete fallback response.  Best-effort — if the platform doesn't
+        # implement ``delete_message``, the delete fails (flood control still
+        # active, bot lacks permission, message too old to delete), the
+        # partial remains but at least the full answer was delivered.
+        if stale_message_id and stale_message_id != last_message_id:
+            delete_fn = getattr(self.adapter, "delete_message", None)
+            if delete_fn is not None:
+                try:
+                    await delete_fn(self.chat_id, stale_message_id)
+                except Exception as e:
+                    logger.debug(
+                        "Fallback partial cleanup failed (%s): %s",
+                        stale_message_id, e,
+                    )
+
         self._message_id = last_message_id
         self._already_sent = True
         self._final_response_sent = True
@@ -699,6 +814,89 @@ class GatewayStreamConsumer:
         err_lower = err.lower()
         return "flood" in err_lower or "retry after" in err_lower or "rate" in err_lower
 
+    def _resolve_draft_streaming(self) -> bool:
+        """Decide whether this run should use native draft streaming.
+
+        Honors ``cfg.transport``:
+          * ``"edit"``  → never use drafts (legacy progressive-edit path).
+          * ``"draft"`` → require draft support; gracefully fall back to edit
+            when the adapter declines.  Logs the downgrade at debug.
+          * ``"auto"``  → use drafts when the adapter supports them for this
+            chat type; otherwise edit.
+
+        Adapter eligibility is checked via
+        :meth:`BasePlatformAdapter.supports_draft_streaming`, which considers
+        the chat type (e.g. Telegram drafts are DM-only) and platform-version
+        gates (e.g. python-telegram-bot 22.6+).
+        """
+        transport = (self.cfg.transport or "auto").lower()
+        if transport == "edit":
+            return False
+        # "off" is filtered upstream by the gateway; treat as edit defensively.
+        if transport == "off":
+            return False
+        # Test adapters are MagicMocks that don't subclass BasePlatformAdapter;
+        # default them to edit so existing test behaviour is preserved.
+        if not isinstance(self.adapter, _BasePlatformAdapter):
+            return False
+        try:
+            supported = self.adapter.supports_draft_streaming(
+                chat_type=self.cfg.chat_type or None,
+                metadata=self.metadata,
+            )
+        except Exception:
+            logger.debug("supports_draft_streaming probe raised", exc_info=True)
+            supported = False
+        if not supported:
+            if transport == "draft":
+                logger.debug(
+                    "Draft streaming requested but unsupported (chat=%s, type=%r) — "
+                    "falling back to edit",
+                    self.chat_id, self.cfg.chat_type,
+                )
+            return False
+        return True
+
+    async def _send_draft_frame(self, text: str) -> bool:
+        """Emit a single animated draft frame for the current accumulated text.
+
+        Returns True when the frame landed.  On any failure, permanently
+        disables drafts for the remainder of this run so subsequent frames
+        flow through the edit-based path (which can adapt with flood-control
+        backoff, etc.).  Drafts have no message_id and clear naturally on
+        the client when the response finalizes via a regular sendMessage.
+        """
+        if self._draft_id is None:
+            # Defensive: should never happen — _use_draft_streaming gate is
+            # set in tandem with _draft_id in run().  Disable to be safe.
+            self._use_draft_streaming = False
+            return False
+        try:
+            result = await self.adapter.send_draft(
+                chat_id=self.chat_id,
+                draft_id=self._draft_id,
+                content=text,
+                metadata=self.metadata,
+            )
+        except Exception as e:
+            logger.debug(
+                "send_draft raised, disabling draft transport for this run: %s", e,
+            )
+            self._draft_failures += 1
+            self._use_draft_streaming = False
+            return False
+        if not getattr(result, "success", False):
+            logger.debug(
+                "send_draft returned success=False, disabling draft transport: %s",
+                getattr(result, "error", "unknown"),
+            )
+            self._draft_failures += 1
+            self._use_draft_streaming = False
+            return False
+        # Frame delivered.  Track text for parity with edit-based no-op skip.
+        self._last_sent_text = text
+        return True
+
     async def _flush_segment_tail_on_edit_failure(self) -> None:
         """Deliver un-sent tail content before a segment-break reset.
 
@@ -893,6 +1091,35 @@ class GatewayStreamConsumer:
                 and self.cfg.cursor in text
                 and len(_visible_stripped) < _MIN_NEW_MSG_CHARS):
             return True  # too short for a standalone message — accumulate more
+
+        # Native draft streaming: route mid-stream frames through send_draft.
+        # The final answer is delivered via the regular sendMessage path
+        # below — drafts have no message_id so we can't finalize them
+        # in-place; the regular sendMessage clears the draft naturally on
+        # the client and gives the user a real message in their history.
+        # Skip when:
+        #   * finalize=True (this is the final answer; needs to be a real message)
+        #   * an edit path is already established (message_id is set, e.g. after
+        #     a tool-boundary segment break where the prior text was finalized
+        #     as a real sendMessage and the next text segment continues editing
+        #     that one — staying on edit-based for that segment is correct).
+        if (
+            self._use_draft_streaming
+            and not finalize
+            and self._message_id is None
+        ):
+            # No-op skip: identical to the last frame we sent.
+            if text == self._last_sent_text:
+                return True
+            ok = await self._send_draft_frame(text)
+            if ok:
+                # Drafts mark "we put something on screen" but DO NOT set
+                # _already_sent — that flag gates the gateway's fallback
+                # final-send path and we still need that to fire so the
+                # user gets a real message (drafts have no message_id).
+                return True
+            # Failure already disabled drafts for this run; fall through to
+            # the regular edit/send path below.
         try:
             if self._message_id is not None:
                 if self._edit_supported:
@@ -931,7 +1158,29 @@ class GatewayStreamConsumer:
                     )
                     if result.success:
                         self._already_sent = True
-                        self._last_sent_text = text
+                        # Adapter may have split-and-delivered an oversized
+                        # edit across the original message + N continuations.
+                        # When that happens, ``message_id`` is the LAST visible
+                        # continuation and ``_last_sent_text`` no longer reflects
+                        # the on-screen content (the new message only holds the
+                        # final chunk's text), so subsequent edits must target
+                        # the new id and skip-if-same comparisons must reset.
+                        # Fire on_new_message so tool-progress bubbles linearize
+                        # below the new continuation, not the original.
+                        # ``getattr`` with default keeps backwards compat with
+                        # SimpleNamespace mocks in tests that pre-date the field.
+                        _continuation_ids = getattr(result, "continuation_message_ids", ()) or ()
+                        if (
+                            _continuation_ids
+                            and result.message_id
+                            and result.message_id != self._message_id
+                        ):
+                            self._message_id = str(result.message_id)
+                            self._message_created_ts = time.monotonic()
+                            self._last_sent_text = ""
+                            self._notify_new_message()
+                        else:
+                            self._last_sent_text = text
                         # Successful edit — reset flood strike counter
                         self._flood_strikes = 0
                         return True
@@ -979,10 +1228,12 @@ class GatewayStreamConsumer:
                     # The final response will be sent by the fallback path.
                     return False
             else:
-                # First message — send new
+                # First message — send new, threaded to the original user message
+                # so it lands in the correct topic/thread.
                 result = await self.adapter.send(
                     chat_id=self.chat_id,
                     content=text,
+                    reply_to=self._initial_reply_to_id,
                     metadata=self.metadata,
                 )
                 if result.success:

hermes_cli/auth.py

@@ -1450,7 +1450,7 @@ def resolve_provider(
         # whose availability isn't implied by LM_API_KEY presence (it may be
         # offline, and the no-auth setup uses a placeholder value), so it
         # also requires explicit selection.
-        if pid in ("copilot", "lmstudio"):
+        if pid in {"copilot", "lmstudio"}:
             continue
         for env_var in pconfig.api_key_env_vars:
             if has_usable_secret(os.getenv(env_var, "")):
@@ -2541,7 +2541,7 @@ def refresh_codex_oauth_pure(
         # A 401/403 from the token endpoint always means the refresh token
         # is invalid/expired — force relogin even if the body error code
         # wasn't one of the known strings above.
-        if response.status_code in (401, 403) and not relogin_required:
+        if response.status_code in {401, 403} and not relogin_required:
             relogin_required = True
         raise AuthError(
             message,
@@ -2947,7 +2947,7 @@ def _merge_shared_nous_oauth_state(state: Dict[str, Any]) -> bool:
         "expires_at",
     ):
         value = shared.get(key)
-        if value not in (None, ""):
+        if value not in {None, ""}:
             state[key] = value
     return True
 
@@ -3986,7 +3986,7 @@ def get_api_key_provider_status(provider_id: str) -> Dict[str, Any]:
     if pconfig.base_url_env_var:
         env_url = os.getenv(pconfig.base_url_env_var, "").strip()
 
-    if provider_id in ("kimi-coding", "kimi-coding-cn"):
+    if provider_id in {"kimi-coding", "kimi-coding-cn"}:
         base_url = _resolve_kimi_base_url(api_key, pconfig.inference_base_url, env_url)
     elif env_url:
         base_url = env_url
@@ -4046,6 +4046,8 @@ def get_auth_status(provider_id: Optional[str] = None) -> Dict[str, Any]:
         return get_qwen_auth_status()
     if target == "google-gemini-cli":
         return get_gemini_oauth_auth_status()
+    if target == "minimax-oauth":
+        return get_minimax_oauth_auth_status()
     if target == "copilot-acp":
         return get_external_process_provider_status(target)
     # API-key providers
@@ -4090,7 +4092,7 @@ def resolve_api_key_provider_credentials(provider_id: str) -> Dict[str, Any]:
     if pconfig.base_url_env_var:
         env_url = os.getenv(pconfig.base_url_env_var, "").strip()
 
-    if provider_id in ("kimi-coding", "kimi-coding-cn"):
+    if provider_id in {"kimi-coding", "kimi-coding-cn"}:
         base_url = _resolve_kimi_base_url(api_key, pconfig.inference_base_url, env_url)
     elif provider_id == "zai":
         base_url = _resolve_zai_base_url(api_key, pconfig.inference_base_url, env_url)
@@ -4510,7 +4512,7 @@ def _login_openai_codex(
                     reuse = input("Use existing credentials? [Y/n]: ").strip().lower()
                 except (EOFError, KeyboardInterrupt):
                     reuse = "y"
-                if reuse in ("", "y", "yes"):
+                if reuse in {"", "y", "yes"}:
                     config_path = _update_config_for_provider("openai-codex", existing.get("base_url", DEFAULT_CODEX_BASE_URL))
                     print()
                     print("Login successful!")
@@ -4531,7 +4533,7 @@ def _login_openai_codex(
                 do_import = input("Import these credentials? (a separate login is recommended) [y/N]: ").strip().lower()
             except (EOFError, KeyboardInterrupt):
                 do_import = "n"
-            if do_import in ("y", "yes"):
+            if do_import in {"y", "yes"}:
                 _save_codex_tokens(cli_tokens)
                 base_url = os.getenv("HERMES_CODEX_BASE_URL", "").strip().rstrip("/") or DEFAULT_CODEX_BASE_URL
                 config_path = _update_config_for_provider("openai-codex", base_url)
@@ -4623,7 +4625,7 @@ def _codex_device_code_login() -> Dict[str, Any]:
                 if poll_resp.status_code == 200:
                     code_resp = poll_resp.json()
                     break
-                elif poll_resp.status_code in (403, 404):
+                elif poll_resp.status_code in {403, 404}:
                     continue  # User hasn't completed login yet
                 else:
                     raise AuthError(
@@ -4757,6 +4759,20 @@ def _minimax_request_user_code(
     return payload
 
 
+def _minimax_expired_in_looks_like_unix_ms(expired_in: int, *, now_ms: int) -> bool:
+    """True if ``expired_in`` is plausibly a unix-ms absolute time (vs TTL seconds)."""
+    return int(expired_in) > (now_ms // 2)
+
+
+def _minimax_resolve_token_expiry_unix(expired_in: int, *, now: datetime) -> float:
+    """Return access-token expiry as unix seconds (MiniMax uses ms epoch or TTL seconds)."""
+    raw = int(expired_in)
+    now_ms = int(now.timestamp() * 1000)
+    if _minimax_expired_in_looks_like_unix_ms(raw, now_ms=now_ms):
+        return raw / 1000.0
+    return now.timestamp() + max(1, raw)
+
+
 def _minimax_poll_token(
     client: httpx.Client, *, portal_base_url: str, client_id: str,
     user_code: str, code_verifier: str, expired_in: int, interval_ms: Optional[int],
@@ -4765,12 +4781,11 @@ def _minimax_poll_token(
     # Defensive parsing: if it's small enough to be a duration, treat as seconds.
     import time as _time
     now_ms = int(_time.time() * 1000)
-    if expired_in > now_ms // 2:
-        # Looks like a unix-ms timestamp.
-        deadline = expired_in / 1000.0
+    raw = int(expired_in)
+    if _minimax_expired_in_looks_like_unix_ms(raw, now_ms=now_ms):
+        deadline = raw / 1000.0
     else:
-        # Treat as duration in seconds from now.
-        deadline = _time.time() + max(1, expired_in)
+        deadline = _time.time() + max(1, raw)
     interval = max(2.0, (interval_ms or 2000) / 1000.0)
 
     while _time.time() < deadline:
@@ -4884,8 +4899,10 @@ def _minimax_oauth_login(
         )
 
     now = datetime.now(timezone.utc)
-    expires_in_s = int(token_data["expired_in"])
-    expires_at = now.timestamp() + expires_in_s
+    expires_at_unix = _minimax_resolve_token_expiry_unix(
+        int(token_data["expired_in"]), now=now,
+    )
+    expires_in_s = max(0, int(expires_at_unix - now.timestamp()))
 
     auth_state = {
         "provider": "minimax-oauth",
@@ -4899,7 +4916,7 @@ def _minimax_oauth_login(
         "refresh_token": token_data["refresh_token"],
         "resource_url": token_data.get("resource_url"),
         "obtained_at": now.isoformat(),
-        "expires_at": datetime.fromtimestamp(expires_at, tz=timezone.utc).isoformat(),
+        "expires_at": datetime.fromtimestamp(expires_at_unix, tz=timezone.utc).isoformat(),
         "expires_in": expires_in_s,
     }
 
@@ -4960,14 +4977,16 @@ def _refresh_minimax_oauth_state(
             relogin_required=True,
         )
     now_dt = datetime.now(timezone.utc)
-    expires_in_s = int(payload["expired_in"])
+    expires_at_unix = _minimax_resolve_token_expiry_unix(
+        int(payload["expired_in"]), now=now_dt,
+    )
+    expires_in_s = max(0, int(expires_at_unix - now_dt.timestamp()))
     new_state = dict(state)
     new_state.update({
         "access_token": payload["access_token"],
         "refresh_token": payload.get("refresh_token", state["refresh_token"]),
         "obtained_at": now_dt.isoformat(),
-        "expires_at": datetime.fromtimestamp(now_dt.timestamp() + expires_in_s,
-                                             tz=timezone.utc).isoformat(),
+        "expires_at": datetime.fromtimestamp(expires_at_unix, tz=timezone.utc).isoformat(),
         "expires_in": expires_in_s,
     })
     _minimax_save_auth_state(new_state)
@@ -5188,7 +5207,7 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
                 do_import = input("Import these credentials? [Y/n]: ").strip().lower()
             except (EOFError, KeyboardInterrupt):
                 do_import = "y"
-            if do_import in ("", "y", "yes"):
+            if do_import in {"", "y", "yes"}:
                 print("Rehydrating Nous session from shared credentials...")
                 auth_state = _try_import_shared_nous_state(
                     timeout_seconds=timeout_seconds,
@@ -5251,6 +5270,7 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
             from hermes_cli.models import (
                 get_curated_nous_model_ids, get_pricing_for_provider,
                 check_nous_free_tier, partition_nous_models_by_tier,
+                union_with_portal_free_recommendations,
             )
             model_ids = get_curated_nous_model_ids()
 
@@ -5260,6 +5280,15 @@ def _login_nous(args, pconfig: ProviderConfig) -> None:
                 pricing = get_pricing_for_provider("nous")
                 free_tier = check_nous_free_tier()
                 if free_tier:
+                    # The Portal's freeRecommendedModels endpoint is the
+                    # source of truth for what's free *right now*. Augment
+                    # the curated list with anything new the Portal flags
+                    # as free so users on older Hermes builds still see
+                    # newly-launched free models without a CLI release.
+                    _portal_for_recs = auth_state.get("portal_base_url", "")
+                    model_ids, pricing = union_with_portal_free_recommendations(
+                        model_ids, pricing, _portal_for_recs,
+                    )
                     model_ids, unavailable_models = partition_nous_models_by_tier(
                         model_ids, pricing, free_tier=True,
                     )

hermes_cli/auth_commands.py

@@ -266,7 +266,7 @@ def auth_add_command(args) -> None:
                 do_import = input("Import these credentials? [Y/n]: ").strip().lower()
             except (EOFError, KeyboardInterrupt):
                 do_import = "y"
-            if do_import in ("", "y", "yes"):
+            if do_import in {"", "y", "yes"}:
                 print("Rehydrating Nous session from shared credentials...")
                 rehydrated = auth_mod._try_import_shared_nous_state(
                     timeout_seconds=getattr(args, "timeout", None) or 15.0,
@@ -375,10 +375,12 @@ def auth_add_command(args) -> None:
         return
 
     if provider == "minimax-oauth":
-        from hermes_cli.auth import resolve_minimax_oauth_runtime_credentials
-        creds = resolve_minimax_oauth_runtime_credentials()
+        creds = auth_mod._minimax_oauth_login(
+            open_browser=not getattr(args, "no_browser", False),
+            timeout_seconds=getattr(args, "timeout", None) or 15.0,
+        )
         label = (getattr(args, "label", None) or "").strip() or label_from_token(
-            creds["api_key"],
+            creds["access_token"],
             _oauth_default_label(provider, len(pool.entries()) + 1),
         )
         entry = PooledCredential(
@@ -388,8 +390,9 @@ def auth_add_command(args) -> None:
             auth_type=AUTH_TYPE_OAUTH,
             priority=0,
             source=f"{SOURCE_MANUAL}:minimax_oauth",
-            access_token=creds["api_key"],
-            base_url=creds.get("base_url"),
+            access_token=creds["access_token"],
+            refresh_token=creds.get("refresh_token"),
+            base_url=creds.get("inference_base_url"),
         )
         pool.add_entry(entry)
         print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')

hermes_cli/backup.py

@@ -298,7 +298,7 @@ def _detect_prefix(zf: zipfile.ZipFile) -> str:
     if len(first_parts) == 1:
         prefix = first_parts.pop()
         # Only strip if it looks like a hermes dir name
-        if prefix in (".hermes", "hermes"):
+        if prefix in {".hermes", "hermes"}:
             return prefix + "/"
 
     return ""
@@ -349,7 +349,7 @@ def run_import(args) -> None:
             except (EOFError, KeyboardInterrupt):
                 print("\nAborted.")
                 sys.exit(1)
-            if answer not in ("y", "yes"):
+            if answer not in {"y", "yes"}:
                 print("Aborted.")
                 return
 
@@ -802,8 +802,7 @@ def _prune_pre_update_backups(backup_dir: Path, keep: int) -> int:
     Operators who genuinely don't want a backup should set
     ``updates.pre_update_backup: false`` in config — that gates creation.
     """
-    if keep < 1:
-        keep = 1
+    keep = max(keep, 1)
     if not backup_dir.exists():
         return 0
 
@@ -875,8 +874,7 @@ def _prune_pre_migration_backups(backup_dir: Path, keep: int) -> int:
     Only touches files matching ``pre-migration-*.zip`` so other backups in
     the same directory are never touched.
     """
-    if keep < 0:
-        keep = 0
+    keep = max(keep, 0)
     if not backup_dir.exists():
         return 0
 

hermes_cli/checkpoints.py

@@ -139,7 +139,7 @@ def _confirm(prompt: str) -> bool:
     except (EOFError, KeyboardInterrupt):
         print()
         return False
-    return resp in ("y", "yes")
+    return resp in {"y", "yes"}
 
 
 def cmd_clear(args: argparse.Namespace) -> int:

hermes_cli/claw.py

@@ -298,7 +298,7 @@ def claw_command(args):
 
     if action == "migrate":
         _cmd_migrate(args)
-    elif action in ("cleanup", "clean"):
+    elif action in {"cleanup", "clean"}:
         _cmd_cleanup(args)
     else:
         print("Usage: hermes claw <command> [options]")
@@ -670,17 +670,16 @@ def _cmd_cleanup(args):
         elif not auto_yes and not sys.stdin.isatty():
             print_info(f"Non-interactive session — would archive: {source_dir}")
             print_info("To execute, re-run with: hermes claw cleanup --yes")
+        elif auto_yes or prompt_yes_no(f"Archive {source_dir}?", default=True):
+            try:
+                archive_path = _archive_directory(source_dir)
+                print_success(f"Archived: {source_dir} → {archive_path}")
+                total_archived += 1
+            except OSError as e:
+                print_error(f"Could not archive: {e}")
+                print_info(f"Try manually: mv {source_dir} {source_dir}.pre-migration")
         else:
-            if auto_yes or prompt_yes_no(f"Archive {source_dir}?", default=True):
-                try:
-                    archive_path = _archive_directory(source_dir)
-                    print_success(f"Archived: {source_dir} → {archive_path}")
-                    total_archived += 1
-                except OSError as e:
-                    print_error(f"Could not archive: {e}")
-                    print_info(f"Try manually: mv {source_dir} {source_dir}.pre-migration")
-            else:
-                print_info("Skipped.")
+            print_info("Skipped.")
 
     # Summary
     print()

hermes_cli/codex_models.py

@@ -16,6 +16,19 @@ DEFAULT_CODEX_MODELS: List[str] = [
     "gpt-5.4-mini",
     "gpt-5.4",
     "gpt-5.3-codex",
+    # gpt-5.3-codex-spark is in research preview and is exposed *only* via
+    # the Codex CLI / OAuth backend (chatgpt.com/backend-api/codex/models)
+    # for ChatGPT Pro subscribers. It is NOT available in the public OpenAI
+    # API, so it intentionally stays out of the "openai" provider catalog
+    # in hermes_cli/models.py — only the openai-codex (OAuth) provider
+    # surfaces it. The Codex backend reports ``supported_in_api: false`` for
+    # this slug; that flag describes API availability, not Codex backend
+    # availability, so the fetch/cache code paths below intentionally do
+    # not filter on it. PR #12994 removed this entry on the assumption it
+    # was unsupported — that was wrong; restored here. Keep it in the
+    # curated fallback so Pro users still see Spark in `/model` when live
+    # discovery is unavailable (offline first run, transient API failure).
+    "gpt-5.3-codex-spark",
     "gpt-5.2-codex",
     "gpt-5.1-codex-max",
     "gpt-5.1-codex-mini",
@@ -26,6 +39,11 @@ _FORWARD_COMPAT_TEMPLATE_MODELS: List[tuple[str, tuple[str, ...]]] = [
     ("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",)),
+    # Surface Spark whenever any compatible Codex template is present so
+    # accounts hitting the live endpoint with an older lineup still see
+    # Spark in the picker. Backend gates real availability by ChatGPT Pro
+    # entitlement; Hermes does not.
+    ("gpt-5.3-codex-spark", ("gpt-5.3-codex", "gpt-5.2-codex")),
 ]
 
 
@@ -78,10 +96,12 @@ def _fetch_models_from_api(access_token: str) -> List[str]:
         if not isinstance(slug, str) or not slug.strip():
             continue
         slug = slug.strip()
-        if item.get("supported_in_api") is False:
-            continue
+        # Codex CLI's catalog uses ``supported_in_api`` for the public OpenAI
+        # API, not for the OAuth-backed Codex backend that this provider uses.
+        # Some valid Codex CLI models (for example gpt-5.3-codex-spark) are
+        # marked false here but are still accepted by the Codex route.
         visibility = item.get("visibility", "")
-        if isinstance(visibility, str) and visibility.strip().lower() in ("hide", "hidden"):
+        if isinstance(visibility, str) and visibility.strip().lower() in {"hide", "hidden"}:
             continue
         priority = item.get("priority")
         rank = int(priority) if isinstance(priority, (int, float)) else 10_000
@@ -128,10 +148,11 @@ def _read_cache_models(codex_home: Path) -> List[str]:
             if not isinstance(slug, str) or not slug.strip():
                 continue
             slug = slug.strip()
-            if item.get("supported_in_api") is False:
-                continue
+            # Do not filter on ``supported_in_api`` here.  It describes the
+            # public OpenAI API, while Hermes openai-codex talks to the same
+            # OAuth-backed Codex backend as Codex CLI.
             visibility = item.get("visibility")
-            if isinstance(visibility, str) and visibility.strip().lower() in ("hide", "hidden"):
+            if isinstance(visibility, str) and visibility.strip().lower() in {"hide", "hidden"}:
                 continue
             priority = item.get("priority")
             rank = int(priority) if isinstance(priority, (int, float)) else 10_000

hermes_cli/commands.py

@@ -105,6 +105,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
     CommandDef("goal", "Set a standing goal Hermes works on across turns until achieved", "Session",
                args_hint="[text | pause | resume | clear | status]"),
     CommandDef("status", "Show session info", "Session"),
+    CommandDef("whoami", "Show your slash command access (admin / user)", "Info"),
     CommandDef("profile", "Show active profile name and home directory", "Info"),
     CommandDef("sethome", "Set this chat as the home channel", "Session",
                gateway_only=True, aliases=("set-home",)),
@@ -810,7 +811,7 @@ def discord_skill_commands_by_category(
     # names are marked with a sentinel so the warning distinguishes
     # "skill collided with a reserved command" from "two skills collided
     # on the 32-char clamp" — the latter is the rename-worthy case.
-    _names_used: dict[str, str] = {n: "<reserved>" for n in reserved_names}
+    _names_used: dict[str, str] = dict.fromkeys(reserved_names, "<reserved>")
     hidden = 0
 
     try:

hermes_cli/completion.py

@@ -216,9 +216,9 @@ _hermes() {{
     typeset -A opt_args
 
     _arguments -C \\
-        '(-h --help){{-h,--help}}[Show help and exit]' \\
-        '(-V --version){{-V,--version}}[Show version and exit]' \\
-        '(-p --profile){{-p,--profile}}[Profile name]:profile:_hermes_profiles' \\
+        '(-)'{{-h,--help}}'[Show help and exit]' \\
+        '(-)'{{-V,--version}}'[Show version and exit]' \\
+        '(-)'{{-p,--profile}}'[Profile name]:profile:_hermes_profiles' \\
         '1:command:->commands' \\
         '*::arg:->args'
 

hermes_cli/config.py

@@ -28,6 +28,48 @@ from typing import Dict, Any, Optional, List, Tuple
 
 logger = logging.getLogger(__name__)
 
+# Track which (config_path, mtime_ns, size) tuples we've already warned about
+# so concurrent CLI/gateway loads of a broken config.yaml don't spam stderr
+# every time. Cleared automatically when the file changes (different mtime).
+_CONFIG_PARSE_WARNED: set = set()
+
+
+def _warn_config_parse_failure(config_path: Path, exc: Exception) -> None:
+    """Surface a config.yaml parse failure to user, log, and stderr.
+
+    A YAML parse error in ``~/.hermes/config.yaml`` causes ``load_config()``
+    to silently fall back to ``DEFAULT_CONFIG``, which means every user
+    override (auxiliary providers, fallback chain, model overrides, etc.)
+    is dropped. Before this helper that was a one-line ``print(...)`` that
+    scrolled off-screen on the first invocation and was never seen again.
+
+    Now: warn once per (path, mtime_ns, size) on stderr **and** in
+    ``agent.log`` / ``errors.log`` at WARNING level so ``hermes logs``
+    surfaces it. Re-warns automatically if the file changes (different
+    mtime/size), so users editing the config see the next failure.
+    """
+    try:
+        st = config_path.stat()
+        key = (str(config_path), st.st_mtime_ns, st.st_size)
+    except OSError:
+        key = (str(config_path), 0, 0)
+    if key in _CONFIG_PARSE_WARNED:
+        return
+    _CONFIG_PARSE_WARNED.add(key)
+
+    msg = (
+        f"Failed to parse {config_path}: {exc}. "
+        f"Falling back to default config — every user override "
+        f"(auxiliary providers, fallback chain, model settings) is being IGNORED. "
+        f"Fix the YAML and restart."
+    )
+    logger.warning(msg)
+    try:
+        sys.stderr.write(f"⚠️  hermes config: {msg}\n")
+        sys.stderr.flush()
+    except Exception:
+        pass
+
 _IS_WINDOWS = platform.system() == "Windows"
 _ENV_VAR_NAME_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*$")
 _LAST_EXPANDED_CONFIG_BY_PATH: Dict[str, Any] = {}
@@ -537,6 +579,7 @@ DEFAULT_CONFIG = {
         # Explicit opt-in: mount the host cwd into /workspace for Docker sessions.
         # Default off because passing host directories into a sandbox weakens isolation.
         "docker_mount_cwd_to_workspace": False,
+        "docker_extra_args": [],        # Extra flags passed verbatim to docker run
         # Explicit opt-in: run the Docker container as the host user's uid:gid
         # (via `--user`).  When enabled, files written into bind-mounted dirs
         # (docker_volumes, the persistent workspace, or the auto-mounted cwd)
@@ -680,8 +723,15 @@ 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.
+    # long_lived_prefix: when true (default), Claude on Anthropic / OpenRouter / Nous
+    #   Portal uses a split layout: tools[-1] + stable system prefix at long_lived_ttl
+    #   (cross-session cache), last 2 messages at cache_ttl (within-session rolling).
+    #   Set false to keep the legacy "system + last 3 messages" single-tier layout.
+    # long_lived_ttl: TTL for the cross-session prefix tier ("5m" or "1h"; default "1h").
     "prompt_caching": {
         "cache_ttl": "5m",
+        "long_lived_prefix": True,
+        "long_lived_ttl": "1h",
     },
 
     # OpenRouter-specific settings.
@@ -691,9 +741,18 @@ DEFAULT_CONFIG = {
     #   See: https://openrouter.ai/docs/guides/features/response-caching
     # response_cache_ttl: how long cached responses remain valid, in seconds (1-86400).
     #   Default 300 (5 minutes). Only used when response_cache is enabled.
+    # min_coding_score: knob for the openrouter/pareto-code router (0.0-1.0).
+    #   Only applied when model.model is "openrouter/pareto-code". Higher
+    #   values route to stronger (more expensive) coders; lower values open
+    #   up cheaper, faster options. Default 0.65 lands on the mid-tier
+    #   coder on the current Pareto frontier. Empty string = let OpenRouter
+    #   pick the strongest available coder (router's documented default
+    #   when the plugins block is omitted).
+    #   See: https://openrouter.ai/docs/guides/routing/routers/pareto-router
     "openrouter": {
         "response_cache": True,
         "response_cache_ttl": 300,
+        "min_coding_score": 0.65,
     },
 
     # AWS Bedrock provider configuration.
@@ -722,6 +781,26 @@ DEFAULT_CONFIG = {
     # Empty model = use provider's default auxiliary model.
     # All tasks fall back to openrouter:google/gemini-3-flash-preview if
     # the configured provider is unavailable.
+    #
+    # extra_body: forwarded verbatim as request body fields on every aux call
+    # for that task. Use this to set provider-specific knobs (independent of
+    # main-agent settings). On OpenRouter you can set provider routing prefs
+    # and the Pareto Code coding-score floor here. Example:
+    #
+    #   auxiliary:
+    #     compression:
+    #       provider: openrouter
+    #       model: openrouter/pareto-code
+    #       extra_body:
+    #         provider:           # OpenRouter provider routing
+    #           order: [anthropic, google]
+    #           sort: throughput  # or price | latency
+    #         plugins:            # OpenRouter Pareto Code router
+    #           - id: pareto-router
+    #             min_coding_score: 0.5
+    #
+    # Each aux task is independent — main-agent provider_routing and
+    # openrouter.min_coding_score do NOT propagate to aux calls by design.
     "auxiliary": {
         "vision": {
             "provider": "auto",    # auto | openrouter | nous | codex | custom
@@ -830,6 +909,7 @@ DEFAULT_CONFIG = {
         "bell_on_complete": False,
         "show_reasoning": False,
         "streaming": False,
+        "timestamps": False,      # Show [HH:MM] on user and assistant labels
         "final_response_markdown": "strip",  # render | strip | raw
         # Preserve recent classic CLI output across Ctrl+L, /redraw, and
         # terminal resize full-screen clears. Disable if a terminal emulator
@@ -1204,6 +1284,15 @@ DEFAULT_CONFIG = {
         # "Always Approve" to silence the prompt permanently; that flips
         # this key to false.
         "mcp_reload_confirm": True,
+        # When true, destructive session slash commands (/clear, /new, /reset,
+        # /undo) ask the user to confirm before discarding conversation state.
+        # Three-option prompt (Approve Once / Always Approve / Cancel) routed
+        # through tools.slash_confirm — native yes/no buttons on Telegram,
+        # Discord, and Slack; text fallback elsewhere.  Users click "Always
+        # Approve" to silence the prompt permanently; that flips this key to
+        # false.  TUI has its own modal overlay (HERMES_TUI_NO_CONFIRM=1 to
+        # opt out there).
+        "destructive_slash_confirm": True,
     },
 
     # Permanently allowed dangerous command patterns (added via "always" approval)
@@ -1243,6 +1332,21 @@ DEFAULT_CONFIG = {
             "domains": [],
             "shared_files": [],
         },
+        # Acknowledged supply-chain security advisories. Each entry is the
+        # ID of an advisory the user has read and acted on (uninstalled the
+        # compromised package, rotated credentials). Acked advisories no
+        # longer trigger the startup banner. Add via `hermes doctor --ack
+        # <id>`; remove by editing the list directly. See
+        # ``hermes_cli/security_advisories.py`` for the catalog.
+        "acked_advisories": [],
+        # Allow Hermes to lazy-install opt-in backend packages from PyPI
+        # the first time the user enables a backend that needs them
+        # (e.g. installing ``elevenlabs`` when the user picks ElevenLabs as
+        # their TTS provider). Set to false to require explicit
+        # ``pip install`` for everything beyond the base set — appropriate
+        # for restricted networks, audited environments, or air-gapped
+        # systems where any runtime install is unacceptable.
+        "allow_lazy_installs": True,
     },
 
     "cron": {
@@ -3120,7 +3224,7 @@ def warn_deprecated_cwd_env_vars(config: Optional[Dict[str, Any]] = None) -> Non
     terminal_cfg = config.get("terminal", {})
     config_cwd = terminal_cfg.get("cwd", ".") if isinstance(terminal_cfg, dict) else "."
     # Only warn if config.yaml doesn't have an explicit path
-    config_has_explicit_cwd = config_cwd not in (".", "auto", "cwd", "")
+    config_has_explicit_cwd = config_cwd not in {".", "auto", "cwd", ""}
 
     lines: list[str] = []
     if messaging_cwd:
@@ -3180,10 +3284,10 @@ def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, A
         if "tool_progress" not in display:
             old_enabled = get_env_value("HERMES_TOOL_PROGRESS")
             old_mode = get_env_value("HERMES_TOOL_PROGRESS_MODE")
-            if old_enabled and old_enabled.lower() in ("false", "0", "no"):
+            if old_enabled and old_enabled.lower() in {"false", "0", "no"}:
                 display["tool_progress"] = "off"
                 results["config_added"].append("display.tool_progress=off (from HERMES_TOOL_PROGRESS=false)")
-            elif old_mode and old_mode.lower() in ("new", "all"):
+            elif old_mode and old_mode.lower() in {"new", "all"}:
                 display["tool_progress"] = old_mode.lower()
                 results["config_added"].append(f"display.tool_progress={old_mode.lower()} (from HERMES_TOOL_PROGRESS_MODE)")
             else:
@@ -3262,7 +3366,7 @@ def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, A
                 new_entry = {"api": old_url}
                 if old_name:
                     new_entry["name"] = old_name
-                if old_key and old_key not in ("no-key", "no-key-required", ""):
+                if old_key and old_key not in {"no-key", "no-key-required", ""}:
                     new_entry["api_key"] = old_key
 
                 # Carry over model and api_mode if present
@@ -3320,7 +3424,7 @@ def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, A
             stt.pop("model", None)
             # Place it in the appropriate provider section only if the
             # user didn't already set a model there
-            if provider in ("local", "local_command"):
+            if provider in {"local", "local_command"}:
                 # Don't migrate an OpenAI model name into the local section
                 _local_models = {
                     "tiny.en", "tiny", "base.en", "base", "small.en", "small",
@@ -3404,7 +3508,7 @@ def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, A
                 if not aux_comp.get("model"):
                     aux_comp["model"] = str(s_model).strip()
                     migrated_keys.append(f"model={s_model}")
-            if s_provider and str(s_provider).strip() not in ("", "auto"):
+            if s_provider and str(s_provider).strip() not in {"", "auto"}:
                 aux = config.setdefault("auxiliary", {})
                 aux_comp = aux.setdefault("compression", {})
                 if not aux_comp.get("provider") or aux_comp.get("provider") == "auto":
@@ -3635,7 +3739,7 @@ def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, A
             except (EOFError, KeyboardInterrupt):
                 answer = "n"
 
-            if answer in ("y", "yes"):
+            if answer in {"y", "yes"}:
                 print()
                 for name, info in new_and_unset:
                     if info.get("url"):
@@ -3696,7 +3800,7 @@ def migrate_config(interactive: bool = True, quiet: bool = False) -> Dict[str, A
         except (EOFError, KeyboardInterrupt):
             answer = "n"
 
-        if answer in ("y", "yes"):
+        if answer in {"y", "yes"}:
             print()
             config = load_config()
             try:
@@ -3966,7 +4070,8 @@ def read_raw_config() -> Dict[str, Any]:
         try:
             with open(config_path, encoding="utf-8") as f:
                 data = yaml.safe_load(f) or {}
-        except Exception:
+        except Exception as e:
+            _warn_config_parse_failure(config_path, e)
             return {}
 
         if not isinstance(data, dict):
@@ -4016,7 +4121,7 @@ def load_config() -> Dict[str, Any]:
 
                 config = _deep_merge(config, user_config)
             except Exception as e:
-                print(f"Warning: Failed to load config: {e}")
+                _warn_config_parse_failure(config_path, e)
 
         normalized = _normalize_root_model_keys(_normalize_max_turns_config(config))
         expanded = _expand_env_vars(normalized)
@@ -4777,9 +4882,9 @@ def set_config_value(key: str, value: str):
     # inline navigation here silently overwrote lists with dicts.
 
     # Convert value to appropriate type
-    if value.lower() in ('true', 'yes', 'on'):
+    if value.lower() in {'true', 'yes', 'on'}:
         value = True
-    elif value.lower() in ('false', 'no', 'off'):
+    elif value.lower() in {'false', 'no', 'off'}:
         value = False
     elif value.isdigit():
         value = int(value)
@@ -4805,6 +4910,7 @@ def set_config_value(key: str, value: str):
         "terminal.vercel_runtime": "TERMINAL_VERCEL_RUNTIME",
         "terminal.docker_mount_cwd_to_workspace": "TERMINAL_DOCKER_MOUNT_CWD_TO_WORKSPACE",
         "terminal.docker_run_as_host_user": "TERMINAL_DOCKER_RUN_AS_HOST_USER",
+        "terminal.docker_env": "TERMINAL_DOCKER_ENV",
         # terminal.cwd intentionally excluded — CLI resolves at runtime,
         # gateway bridges it in gateway/run.py. Persisting to .env causes
         # stale values to poison child processes.
@@ -4983,7 +5089,7 @@ def _inject_profile_env_vars() -> None:
     try:
         from providers import list_providers
         for _pp in list_providers():
-            if _pp.auth_type not in ("api_key",):
+            if _pp.auth_type not in {"api_key",}:
                 continue
             for _var in _pp.env_vars:
                 if _var in OPTIONAL_ENV_VARS:

hermes_cli/copilot_auth.py

@@ -128,7 +128,7 @@ def _try_gh_cli_token() -> Optional[str]:
 
     # Build a clean env so gh doesn't short-circuit on GITHUB_TOKEN / GH_TOKEN
     clean_env = {k: v for k, v in os.environ.items()
-                 if k not in ("GITHUB_TOKEN", "GH_TOKEN")}
+                 if k not in {"GITHUB_TOKEN", "GH_TOKEN"}}
 
     for gh_path in _gh_cli_candidates():
         cmd = [gh_path, "auth", "token"]

hermes_cli/curator.py

@@ -55,7 +55,16 @@ def _cmd_status(args) -> int:
     print(f"curator: {status_line}")
     print(f"  runs:           {runs}")
     print(f"  last run:       {_fmt_ts(last_run)}")
-    print(f"  last summary:   {summary}")
+    # Summary may be multi-line when the curator archived skills (the rename
+    # map gets appended as `name → umbrella` lines). Indent continuation
+    # lines so the block reads as one logical field.
+    if "\n" in summary:
+        first, *rest = summary.splitlines()
+        print(f"  last summary:   {first}")
+        for line in rest:
+            print(f"                  {line}")
+    else:
+        print(f"  last summary:   {summary}")
     _report = state.get("last_report_path")
     if _report:
         suffix = "" if Path(_report).exists() else " (missing)"
@@ -338,7 +347,7 @@ def _cmd_prune(args) -> int:
         except (EOFError, KeyboardInterrupt):
             print("\ncurator: aborted")
             return 1
-        if reply not in ("y", "yes"):
+        if reply not in {"y", "yes"}:
             print("curator: aborted")
             return 1
 
@@ -440,7 +449,7 @@ def _cmd_rollback(args) -> int:
         except (EOFError, KeyboardInterrupt):
             print("\ncancelled")
             return 1
-        if ans not in ("y", "yes"):
+        if ans not in {"y", "yes"}:
             print("cancelled")
             return 1
 

hermes_cli/curses_ui.py

@@ -139,16 +139,16 @@ def curses_checklist(
                 stdscr.refresh()
                 key = stdscr.getch()
 
-                if key in (curses.KEY_UP, ord("k")):
+                if key in {curses.KEY_UP, ord("k")}:
                     cursor = (cursor - 1) % len(items)
-                elif key in (curses.KEY_DOWN, ord("j")):
+                elif key in {curses.KEY_DOWN, ord("j")}:
                     cursor = (cursor + 1) % len(items)
                 elif key == ord(" "):
                     chosen.symmetric_difference_update({cursor})
-                elif key in (curses.KEY_ENTER, 10, 13):
+                elif key in {curses.KEY_ENTER, 10, 13}:
                     result_holder[0] = set(chosen)
                     return
-                elif key in (27, ord("q")):
+                elif key in {27, ord("q")}:
                     result_holder[0] = cancel_returns
                     return
 
@@ -265,14 +265,14 @@ def curses_radiolist(
                 stdscr.refresh()
                 key = stdscr.getch()
 
-                if key in (curses.KEY_UP, ord("k")):
+                if key in {curses.KEY_UP, ord("k")}:
                     cursor = (cursor - 1) % len(items)
-                elif key in (curses.KEY_DOWN, ord("j")):
+                elif key in {curses.KEY_DOWN, ord("j")}:
                     cursor = (cursor + 1) % len(items)
-                elif key in (ord(" "), curses.KEY_ENTER, 10, 13):
+                elif key in {ord(" "), curses.KEY_ENTER, 10, 13}:
                     result_holder[0] = cursor
                     return
-                elif key in (27, ord("q")):
+                elif key in {27, ord("q")}:
                     result_holder[0] = cancel_returns
                     return
 
@@ -388,14 +388,14 @@ def curses_single_select(
                 stdscr.refresh()
                 key = stdscr.getch()
 
-                if key in (curses.KEY_UP, ord("k")):
+                if key in {curses.KEY_UP, ord("k")}:
                     cursor = (cursor - 1) % len(all_items)
-                elif key in (curses.KEY_DOWN, ord("j")):
+                elif key in {curses.KEY_DOWN, ord("j")}:
                     cursor = (cursor + 1) % len(all_items)
-                elif key in (curses.KEY_ENTER, 10, 13):
+                elif key in {curses.KEY_ENTER, 10, 13}:
                     result_holder[0] = cursor
                     return
-                elif key in (27, ord("q")):
+                elif key in {27, ord("q")}:
                     result_holder[0] = None
                     return
 

hermes_cli/dingtalk_auth.py

@@ -93,7 +93,7 @@ def poll_registration(device_code: str) -> dict:
     """
     data = _api_post("/app/registration/poll", {"device_code": device_code})
     status_raw = str(data.get("status", "")).strip().upper()
-    if status_raw not in ("WAITING", "SUCCESS", "FAIL", "EXPIRED"):
+    if status_raw not in {"WAITING", "SUCCESS", "FAIL", "EXPIRED"}:
         status_raw = "UNKNOWN"
     return {
         "status": status_raw,

hermes_cli/doctor.py

@@ -245,15 +245,31 @@ def _build_apikey_providers_list() -> list:
     }
     for _label, _canonical in _name_to_canonical.items():
         _known_canonical.add(_canonical)
+    # Providers that already have a dedicated health check above the generic
+    # API-key loop (with custom headers/auth). Skip their pluggable profiles
+    # here so the generic Bearer-auth loop doesn't run a duplicate, broken
+    # check (e.g. Anthropic native API requires x-api-key, not Bearer).
+    _dedicated_canonical = {"anthropic", "openrouter", "bedrock"}
+    _known_canonical.update(_dedicated_canonical)
     try:
         from providers import list_providers
         from providers.base import ProviderProfile as _PP
+        try:
+            from hermes_cli.providers import normalize_provider as _normalize_provider
+        except Exception:  # pragma: no cover - normalization is best-effort
+            def _normalize_provider(_name: str) -> str:
+                return (_name or "").strip().lower()
         for _pp in list_providers():
             if not isinstance(_pp, _PP) or _pp.auth_type != "api_key" or not _pp.env_vars:
                 continue
             _label = _pp.display_name or _pp.name
             if _label in _known_names or _pp.name in _known_canonical:
                 continue
+            _candidates = {_normalize_provider(_pp.name)}
+            for _alias in (_pp.aliases or ()):
+                _candidates.add(_normalize_provider(_alias))
+            if _candidates & _dedicated_canonical:
+                continue
             # Separate API-key vars from base-URL override vars — the health-check
             # loop sends the first found value as Authorization: Bearer, so a URL
             # string must never be picked.
@@ -280,19 +296,101 @@ def _build_apikey_providers_list() -> list:
 def run_doctor(args):
     """Run diagnostic checks."""
     should_fix = getattr(args, 'fix', False)
+    ack_target = getattr(args, 'ack', None)
 
     # Doctor runs from the interactive CLI, so CLI-gated tool availability
     # checks (like cronjob management) should see the same context as `hermes`.
     os.environ.setdefault("HERMES_INTERACTIVE", "1")
-    
+
+    # Handle `hermes doctor --ack <id>` as a fast path. Persist the ack and
+    # return without running the rest of the diagnostics — the user has
+    # already seen the advisory and just wants to silence it.
+    if ack_target:
+        from hermes_cli.security_advisories import (
+            ADVISORIES,
+            ack_advisory,
+        )
+        valid_ids = {a.id for a in ADVISORIES}
+        if ack_target not in valid_ids:
+            print(color(
+                f"Unknown advisory ID: {ack_target!r}. Known IDs: "
+                f"{', '.join(sorted(valid_ids)) or '(none)'}",
+                Colors.RED,
+            ))
+            sys.exit(2)
+        if ack_advisory(ack_target):
+            print(color(
+                f"  ✓ Acknowledged advisory {ack_target}. "
+                f"It will no longer trigger startup banners.",
+                Colors.GREEN,
+            ))
+        else:
+            print(color(
+                f"  ✗ Failed to persist ack for {ack_target}. "
+                f"Check ~/.hermes/config.yaml is writable.",
+                Colors.RED,
+            ))
+            sys.exit(1)
+        return
+
     issues = []
     manual_issues = []  # issues that can't be auto-fixed
     fixed_count = 0
-    
+
     print()
     print(color("┌─────────────────────────────────────────────────────────┐", Colors.CYAN))
     print(color("│                 🩺 Hermes Doctor                        │", Colors.CYAN))
     print(color("└─────────────────────────────────────────────────────────┘", Colors.CYAN))
+
+    # =========================================================================
+    # Check: Security advisories  (RUNS FIRST — these are the most urgent)
+    # =========================================================================
+    print()
+    print(color("◆ Security Advisories", Colors.CYAN, Colors.BOLD))
+    try:
+        from hermes_cli.security_advisories import (
+            detect_compromised,
+            filter_unacked,
+            full_remediation_text,
+            get_acked_ids,
+        )
+        all_hits = detect_compromised()
+        fresh_hits = filter_unacked(all_hits)
+        if fresh_hits:
+            for hit in fresh_hits:
+                check_fail(
+                    f"{hit.advisory.title}",
+                    f"({hit.package}=={hit.installed_version})",
+                )
+                # Print the full remediation block, indented under the
+                # check_fail header so it reads as a single section.
+                for line in full_remediation_text(hit):
+                    if line:
+                        print(f"    {color(line, Colors.YELLOW)}")
+                    else:
+                        print()
+                # Funnel into the action list so the summary block surfaces it
+                # for users who scroll past the section.
+                manual_issues.append(
+                    f"Resolve security advisory {hit.advisory.id}: "
+                    f"uninstall {hit.package}=={hit.installed_version} and "
+                    f"rotate credentials, then run "
+                    f"`hermes doctor --ack {hit.advisory.id}`."
+                )
+            # Acked-but-still-installed: show as informational so the user
+            # knows the package is still on disk after the ack.
+            acked_ids = get_acked_ids()
+            for h in all_hits:
+                if h.advisory.id in acked_ids:
+                    check_warn(
+                        f"{h.package}=={h.installed_version} still installed "
+                        f"(advisory {h.advisory.id} acknowledged)",
+                    )
+        else:
+            check_ok("No active security advisories")
+    except Exception as e:
+        # Never let a bug in the advisory check block the rest of doctor.
+        check_warn(f"Security advisory check failed: {e}")
     
     # =========================================================================
     # Check: Python version
@@ -457,7 +555,7 @@ def run_doctor(args):
             if (
                 provider
                 and _resolve_auth_provider is not None
-                and provider not in ("auto", "custom")
+                and provider not in {"auto", "custom"}
             ):
                 try:
                     runtime_provider = _resolve_auth_provider(provider)
@@ -469,7 +567,7 @@ def run_doctor(args):
             if (
                 provider
                 and _resolve_provider_full is not None
-                and provider not in ("auto", "custom")
+                and provider not in {"auto", "custom"}
             ):
                 provider_def = _resolve_provider_full(provider, user_providers, custom_providers)
                 catalog_provider = provider_def.id if provider_def is not None else None
@@ -526,7 +624,7 @@ def run_doctor(args):
             # own env-var checks elsewhere in doctor, and get_auth_status()
             # returns a bare {logged_in: False} for anything it doesn't
             # explicitly dispatch, which would produce false positives.
-            if runtime_provider and runtime_provider not in ("auto", "custom", "openrouter"):
+            if runtime_provider and runtime_provider not in {"auto", "custom", "openrouter"}:
                 try:
                     from hermes_cli.auth import PROVIDER_REGISTRY, get_auth_status
                     pconfig = PROVIDER_REGISTRY.get(runtime_provider)
@@ -713,13 +811,12 @@ def run_doctor(args):
     hermes_home = HERMES_HOME
     if hermes_home.exists():
         check_ok(f"{_DHH} directory exists")
+    elif should_fix:
+        hermes_home.mkdir(parents=True, exist_ok=True)
+        check_ok(f"Created {_DHH} directory")
+        fixed_count += 1
     else:
-        if should_fix:
-            hermes_home.mkdir(parents=True, exist_ok=True)
-            check_ok(f"Created {_DHH} directory")
-            fixed_count += 1
-        else:
-            check_warn(f"{_DHH} not found", "(will be created on first use)")
+        check_warn(f"{_DHH} not found", "(will be created on first use)")
     
     # Check expected subdirectories
     expected_subdirs = ["cron", "sessions", "logs", "skills", "memories"]
@@ -727,13 +824,12 @@ def run_doctor(args):
         subdir_path = hermes_home / subdir_name
         if subdir_path.exists():
             check_ok(f"{_DHH}/{subdir_name}/ exists")
+        elif should_fix:
+            subdir_path.mkdir(parents=True, exist_ok=True)
+            check_ok(f"Created {_DHH}/{subdir_name}/")
+            fixed_count += 1
         else:
-            if should_fix:
-                subdir_path.mkdir(parents=True, exist_ok=True)
-                check_ok(f"Created {_DHH}/{subdir_name}/")
-                fixed_count += 1
-            else:
-                check_warn(f"{_DHH}/{subdir_name}/ not found", "(will be created on first use)")
+            check_warn(f"{_DHH}/{subdir_name}/ not found", "(will be created on first use)")
     
     # Check for SOUL.md persona file
     soul_path = hermes_home / "SOUL.md"
@@ -939,14 +1035,12 @@ def run_doctor(args):
         else:
             check_fail("docker not found", "(required for TERMINAL_ENV=docker)")
             issues.append("Install Docker or change TERMINAL_ENV")
+    elif _safe_which("docker"):
+        check_ok("docker", "(optional)")
+    elif _is_termux():
+        check_info("Docker backend is not available inside Termux (expected on Android)")
     else:
-        if _safe_which("docker"):
-            check_ok("docker", "(optional)")
-        else:
-            if _is_termux():
-                check_info("Docker backend is not available inside Termux (expected on Android)")
-            else:
-                check_warn("docker not found", "(optional)")
+        check_warn("docker not found", "(optional)")
     
     # SSH (if using ssh backend)
     if terminal_env == "ssh":
@@ -998,7 +1092,7 @@ def run_doctor(args):
             issues.append(f"Set TERMINAL_VERCEL_RUNTIME to one of: {supported}")
 
         disk = os.getenv("TERMINAL_CONTAINER_DISK", "51200").strip()
-        if disk in ("", "0", "51200"):
+        if disk in {"", "0", "51200"}:
             check_ok("Vercel disk setting", "(uses platform default)")
         else:
             check_fail("Vercel custom disk unsupported", "(reset terminal.container_disk to 51200)")
@@ -1024,7 +1118,7 @@ def run_doctor(args):
         for line in auth_status.detail_lines:
             check_info(f"Vercel auth {line}")
 
-        persistent = os.getenv("TERMINAL_CONTAINER_PERSISTENT", "true").lower() in ("1", "true", "yes", "on")
+        persistent = os.getenv("TERMINAL_CONTAINER_PERSISTENT", "true").lower() in {"1", "true", "yes", "on"}
         if persistent:
             check_info("Vercel persistence: snapshot filesystem only; live processes do not survive sandbox recreation")
         else:
@@ -1042,15 +1136,14 @@ def run_doctor(args):
         elif shutil.which("agent-browser"):
             check_ok("agent-browser", "(browser automation)")
             agent_browser_ok = True
+        elif _is_termux():
+            check_info("agent-browser is not installed (expected in the tested Termux path)")
+            check_info("Install it manually later with: npm install -g agent-browser && agent-browser install")
+            check_info("Termux browser setup:")
+            for step in _termux_browser_setup_steps(node_installed=True):
+                check_info(step)
         else:
-            if _is_termux():
-                check_info("agent-browser is not installed (expected in the tested Termux path)")
-                check_info("Install it manually later with: npm install -g agent-browser && agent-browser install")
-                check_info("Termux browser setup:")
-                for step in _termux_browser_setup_steps(node_installed=True):
-                    check_info(step)
-            else:
-                check_warn("agent-browser not installed", "(run: npm install)")
+            check_warn("agent-browser not installed", "(run: npm install)")
 
         # Chromium presence — the browser tools silently fail to register when
         # agent-browser is found but no Playwright-managed Chromium is on disk
@@ -1101,15 +1194,14 @@ def run_doctor(args):
                                 f"Install with: cd {PROJECT_ROOT} && "
                                 "npx playwright install --with-deps chromium"
                             )
+    elif _is_termux():
+        check_info("Node.js not found (browser tools are optional in the tested Termux path)")
+        check_info("Install Node.js on Termux with: pkg install nodejs")
+        check_info("Termux browser setup:")
+        for step in _termux_browser_setup_steps(node_installed=False):
+            check_info(step)
     else:
-        if _is_termux():
-            check_info("Node.js not found (browser tools are optional in the tested Termux path)")
-            check_info("Install Node.js on Termux with: pkg install nodejs")
-            check_info("Termux browser setup:")
-            for step in _termux_browser_setup_steps(node_installed=False):
-                check_info(step)
-        else:
-            check_warn("Node.js not found", "(optional, needed for browser tools)")
+        check_warn("Node.js not found", "(optional, needed for browser tools)")
     
     # npm audit for all Node.js packages
     _npm_bin = _safe_which("npm")
@@ -1166,44 +1258,92 @@ def run_doctor(args):
     # =========================================================================
     print()
     print(color("◆ API Connectivity", Colors.CYAN, Colors.BOLD))
-    
-    openrouter_key = os.getenv("OPENROUTER_API_KEY")
-    if openrouter_key:
-        print("  Checking OpenRouter API...", end="", flush=True)
+
+    # Refactor: every connectivity probe below is HTTP-bound and fully
+    # independent. Running them in series spent ~5s wall on a typical
+    # workstation (2s of that was boto3's IMDS lookup for AWS credentials,
+    # which times out unless you're actually on EC2). Threading them with
+    # a small executor pool collapses the section to roughly the slowest
+    # single probe — about 2s — without changing the output format.
+    #
+    # Each ``_probe_*`` helper is a pure function: takes its inputs,
+    # makes one HTTP/SDK call, returns a ``_ConnectivityResult`` carrying
+    # the line(s) to print and any issue strings to append. No globals,
+    # no shared mutable state, no printing inside the workers.
+    import concurrent.futures as _futures
+    from collections import namedtuple as _namedtuple
+
+    _ConnectivityResult = _namedtuple(
+        "_ConnectivityResult", ["label", "lines", "issues"]
+    )
+    _probes: list = []  # list of (label, callable) submitted in display order
+
+    def _probe_openrouter() -> _ConnectivityResult:
+        key = os.getenv("OPENROUTER_API_KEY")
+        if not key:
+            return _ConnectivityResult(
+                "OpenRouter API",
+                [(color("⚠", Colors.YELLOW), "OpenRouter API",
+                  color("(not configured)", Colors.DIM))],
+                [],
+            )
         try:
             import httpx
-            response = httpx.get(
+            r = httpx.get(
                 OPENROUTER_MODELS_URL,
-                headers={"Authorization": f"Bearer {openrouter_key}"},
-                timeout=10
+                headers={"Authorization": f"Bearer {key}"},
+                timeout=10,
             )
-            if response.status_code == 200:
-                print(f"\r  {color('✓', Colors.GREEN)} OpenRouter API                          ")
-            elif response.status_code == 401:
-                print(f"\r  {color('✗', Colors.RED)} OpenRouter API {color('(invalid API key)', Colors.DIM)}                ")
-                issues.append("Check OPENROUTER_API_KEY in .env")
-            elif response.status_code == 402:
-                print(f"\r  {color('✗', Colors.RED)} OpenRouter API {color('(out of credits — payment required)', Colors.DIM)}")
-                issues.append(
-                    "OpenRouter account has insufficient credits. "
-                    "Fix: run 'hermes config set model.provider <provider>' to switch providers, "
-                    "or fund your OpenRouter account at https://openrouter.ai/settings/credits"
+            if r.status_code == 200:
+                return _ConnectivityResult(
+                    "OpenRouter API",
+                    [(color("✓", Colors.GREEN), "OpenRouter API", "")],
+                    [],
                 )
-            elif response.status_code == 429:
-                print(f"\r  {color('✗', Colors.RED)} OpenRouter API {color('(rate limited)', Colors.DIM)}                ")
-                issues.append("OpenRouter rate limit hit — consider switching to a different provider or waiting")
-            else:
-                print(f"\r  {color('✗', Colors.RED)} OpenRouter API {color(f'(HTTP {response.status_code})', Colors.DIM)}                ")
+            if r.status_code == 401:
+                return _ConnectivityResult(
+                    "OpenRouter API",
+                    [(color("✗", Colors.RED), "OpenRouter API",
+                      color("(invalid API key)", Colors.DIM))],
+                    ["Check OPENROUTER_API_KEY in .env"],
+                )
+            if r.status_code == 402:
+                return _ConnectivityResult(
+                    "OpenRouter API",
+                    [(color("✗", Colors.RED), "OpenRouter API",
+                      color("(out of credits — payment required)", Colors.DIM))],
+                    ["OpenRouter account has insufficient credits. "
+                     "Fix: run 'hermes config set model.provider <provider>' "
+                     "to switch providers, or fund your OpenRouter account "
+                     "at https://openrouter.ai/settings/credits"],
+                )
+            if r.status_code == 429:
+                return _ConnectivityResult(
+                    "OpenRouter API",
+                    [(color("✗", Colors.RED), "OpenRouter API",
+                      color("(rate limited)", Colors.DIM))],
+                    ["OpenRouter rate limit hit — consider switching to "
+                     "a different provider or waiting"],
+                )
+            return _ConnectivityResult(
+                "OpenRouter API",
+                [(color("✗", Colors.RED), "OpenRouter API",
+                  color(f"(HTTP {r.status_code})", Colors.DIM))],
+                [],
+            )
         except Exception as e:
-            print(f"\r  {color('✗', Colors.RED)} OpenRouter API {color(f'({e})', Colors.DIM)}                ")
-            issues.append("Check network connectivity")
-    else:
-        check_warn("OpenRouter API", "(not configured)")
-    
-    from hermes_cli.auth import get_anthropic_key
-    anthropic_key = get_anthropic_key()
-    if anthropic_key:
-        print("  Checking Anthropic API...", end="", flush=True)
+            return _ConnectivityResult(
+                "OpenRouter API",
+                [(color("✗", Colors.RED), "OpenRouter API",
+                  color(f"({e})", Colors.DIM))],
+                ["Check network connectivity"],
+            )
+
+    def _probe_anthropic() -> _ConnectivityResult:
+        from hermes_cli.auth import get_anthropic_key
+        key = get_anthropic_key()
+        if not key:
+            return _ConnectivityResult("Anthropic API", [], [])
         try:
             import httpx
             from agent.anthropic_adapter import (
@@ -1212,140 +1352,247 @@ def run_doctor(args):
                 _OAUTH_ONLY_BETAS,
                 _CONTEXT_1M_BETA,
             )
-
             headers = {"anthropic-version": "2023-06-01"}
-            is_oauth = _is_oauth_token(anthropic_key)
+            is_oauth = _is_oauth_token(key)
             if is_oauth:
-                headers["Authorization"] = f"Bearer {anthropic_key}"
+                headers["Authorization"] = f"Bearer {key}"
                 headers["anthropic-beta"] = ",".join(_COMMON_BETAS + _OAUTH_ONLY_BETAS)
             else:
-                headers["x-api-key"] = anthropic_key
-            response = httpx.get(
+                headers["x-api-key"] = key
+            r = httpx.get(
                 "https://api.anthropic.com/v1/models",
-                headers=headers,
-                timeout=10
+                headers=headers, timeout=10,
             )
-            # Reactive recovery: OAuth subscriptions that don't include 1M
-            # context reject the request with 400 "long context beta is not
-            # yet available for this subscription". Retry once with that
-            # beta stripped so the doctor check doesn't falsely report the
-            # Anthropic API as unreachable for those users.
+            # Reactive recovery: OAuth subscriptions without 1M context reject the
+            # request with 400 "long context beta is not yet available for this
+            # subscription". Retry once with that beta stripped so the doctor
+            # check doesn't falsely report Anthropic as unreachable.
             if (
                 is_oauth
-                and response.status_code == 400
-                and "long context beta" in response.text.lower()
-                and "not yet available" in response.text.lower()
+                and r.status_code == 400
+                and "long context beta" in r.text.lower()
+                and "not yet available" in r.text.lower()
             ):
                 headers["anthropic-beta"] = ",".join(
-                    [b for b in _COMMON_BETAS if b != _CONTEXT_1M_BETA] + list(_OAUTH_ONLY_BETAS)
+                    [b for b in _COMMON_BETAS if b != _CONTEXT_1M_BETA]
+                    + list(_OAUTH_ONLY_BETAS)
                 )
-                response = httpx.get(
+                r = httpx.get(
                     "https://api.anthropic.com/v1/models",
-                    headers=headers,
-                    timeout=10,
+                    headers=headers, timeout=10,
                 )
-            if response.status_code == 200:
-                print(f"\r  {color('✓', Colors.GREEN)} Anthropic API                           ")
-            elif response.status_code == 401:
-                print(f"\r  {color('✗', Colors.RED)} Anthropic API {color('(invalid API key)', Colors.DIM)}                 ")
-            else:
-                msg = "(couldn't verify)"
-                print(f"\r  {color('⚠', Colors.YELLOW)} Anthropic API {color(msg, Colors.DIM)}                 ")
+            if r.status_code == 200:
+                return _ConnectivityResult(
+                    "Anthropic API",
+                    [(color("✓", Colors.GREEN), "Anthropic API", "")],
+                    [],
+                )
+            if r.status_code == 401:
+                return _ConnectivityResult(
+                    "Anthropic API",
+                    [(color("✗", Colors.RED), "Anthropic API",
+                      color("(invalid API key)", Colors.DIM))],
+                    [],
+                )
+            return _ConnectivityResult(
+                "Anthropic API",
+                [(color("⚠", Colors.YELLOW), "Anthropic API",
+                  color("(couldn't verify)", Colors.DIM))],
+                [],
+            )
         except Exception as e:
-            print(f"\r  {color('⚠', Colors.YELLOW)} Anthropic API {color(f'({e})', Colors.DIM)}                 ")
+            return _ConnectivityResult(
+                "Anthropic API",
+                [(color("⚠", Colors.YELLOW), "Anthropic API",
+                  color(f"({e})", Colors.DIM))],
+                [],
+            )
+
+    def _probe_apikey_provider(pname, env_vars, default_url, base_env,
+                               supports_health_check) -> _ConnectivityResult:
+        key = ""
+        for ev in env_vars:
+            key = os.getenv(ev, "")
+            if key:
+                break
+        if not key:
+            return _ConnectivityResult(pname, [], [])
+        label = pname.ljust(20)
+        if not supports_health_check:
+            return _ConnectivityResult(
+                pname,
+                [(color("✓", Colors.GREEN), label,
+                  color("(key configured)", Colors.DIM))],
+                [],
+            )
+        try:
+            import httpx
+            base = os.getenv(base_env, "") if base_env else ""
+            # Auto-detect Kimi Code keys (sk-kimi-) → api.kimi.com/coding/v1
+            # (OpenAI-compat surface, which exposes /models for health check).
+            if not base and key.startswith("sk-kimi-"):
+                base = "https://api.kimi.com/coding/v1"
+            # Anthropic-compat endpoints (/anthropic, api.kimi.com/coding
+            # with no /v1) don't support /models. Rewrite to OpenAI-compat
+            # /v1 surface for health checks.
+            if base and base.rstrip("/").endswith("/anthropic"):
+                from agent.auxiliary_client import _to_openai_base_url
+                base = _to_openai_base_url(base)
+            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}",
+                "User-Agent": _HERMES_USER_AGENT,
+            }
+            if base_url_host_matches(base, "api.kimi.com"):
+                headers["User-Agent"] = "claude-code/0.1.0"
+            r = httpx.get(url, headers=headers, timeout=10)
+            if (
+                pname == "Alibaba/DashScope"
+                and not base
+                and r.status_code == 401
+            ):
+                r = httpx.get(
+                    "https://dashscope.aliyuncs.com/compatible-mode/v1/models",
+                    headers=headers, timeout=10,
+                )
+            if r.status_code == 200:
+                return _ConnectivityResult(
+                    pname,
+                    [(color("✓", Colors.GREEN), label, "")],
+                    [],
+                )
+            if r.status_code == 401:
+                return _ConnectivityResult(
+                    pname,
+                    [(color("✗", Colors.RED), label,
+                      color("(invalid API key)", Colors.DIM))],
+                    [f"Check {env_vars[0]} in .env"],
+                )
+            return _ConnectivityResult(
+                pname,
+                [(color("⚠", Colors.YELLOW), label,
+                  color(f"(HTTP {r.status_code})", Colors.DIM))],
+                [],
+            )
+        except Exception as e:
+            return _ConnectivityResult(
+                pname,
+                [(color("⚠", Colors.YELLOW), label,
+                  color(f"({e})", Colors.DIM))],
+                [],
+            )
+
+    def _probe_bedrock() -> _ConnectivityResult:
+        try:
+            from agent.bedrock_adapter import (
+                has_aws_credentials,
+                resolve_aws_auth_env_var,
+                resolve_bedrock_region,
+            )
+        except ImportError:
+            return _ConnectivityResult("AWS Bedrock", [], [])
+        if not has_aws_credentials():
+            return _ConnectivityResult("AWS Bedrock", [], [])
+        auth_var = resolve_aws_auth_env_var()
+        region = resolve_bedrock_region()
+        label = "AWS Bedrock".ljust(20)
+        try:
+            import boto3
+            from botocore.config import Config as _BotoConfig
+            # Trim retries on the actual Bedrock API call so a transient
+            # failure doesn't pad the doctor run by 30+ seconds.
+            cfg = _BotoConfig(
+                connect_timeout=5,
+                read_timeout=10,
+                retries={"max_attempts": 1},
+            )
+            client = boto3.client("bedrock", region_name=region, config=cfg)
+            resp = client.list_foundation_models()
+            n = len(resp.get("modelSummaries", []))
+            return _ConnectivityResult(
+                "AWS Bedrock",
+                [(color("✓", Colors.GREEN), label,
+                  color(f"({auth_var}, {region}, {n} models)", Colors.DIM))],
+                [],
+            )
+        except ImportError:
+            return _ConnectivityResult(
+                "AWS Bedrock",
+                [(color("⚠", Colors.YELLOW), label,
+                  color(f"(boto3 not installed — {sys.executable} -m pip install boto3)",
+                        Colors.DIM))],
+                [f"Install boto3 for Bedrock: {sys.executable} -m pip install boto3"],
+            )
+        except Exception as e:
+            err_name = type(e).__name__
+            return _ConnectivityResult(
+                "AWS Bedrock",
+                [(color("⚠", Colors.YELLOW), label,
+                  color(f"({err_name}: {e})", Colors.DIM))],
+                [f"AWS Bedrock: {err_name} — check IAM permissions for "
+                 f"bedrock:ListFoundationModels"],
+            )
+
+    # Build the probe submission list in display order
+    _probes.append(("OpenRouter API", _probe_openrouter))
+    _probes.append(("Anthropic API", _probe_anthropic))
 
-    # -- API-key providers --
-    # Tuple: (name, env_vars, default_url, base_env, supports_models_endpoint)
-    # If supports_models_endpoint is False, we skip the health check and just show "configured"
-    # Cached at module level after first build — profiles auto-extend it.
     global _APIKEY_PROVIDERS_CACHE
     if _APIKEY_PROVIDERS_CACHE is None:
         _APIKEY_PROVIDERS_CACHE = _build_apikey_providers_list()
-    _apikey_providers = _APIKEY_PROVIDERS_CACHE
-    for _pname, _env_vars, _default_url, _base_env, _supports_health_check in _apikey_providers:
-        _key = ""
-        for _ev in _env_vars:
-            _key = os.getenv(_ev, "")
-            if _key:
-                break
-        if _key:
-            _label = _pname.ljust(20)
-            # Some providers (like MiniMax) don't support /models endpoint
-            if not _supports_health_check:
-                print(f"  {color('✓', Colors.GREEN)} {_label} {color('(key configured)', Colors.DIM)}")
-                continue
-            print(f"  Checking {_pname} API...", end="", flush=True)
-            try:
-                import httpx
-                _base = os.getenv(_base_env, "") if _base_env else ""
-                # Auto-detect Kimi Code keys (sk-kimi-) → api.kimi.com/coding/v1
-                # (OpenAI-compat surface, which exposes /models for health check).
-                if not _base and _key.startswith("sk-kimi-"):
-                    _base = "https://api.kimi.com/coding/v1"
-                # Anthropic-compat endpoints (/anthropic, api.kimi.com/coding
-                # with no /v1) don't support /models.  Rewrite to the OpenAI-compat
-                # /v1 surface for health checks.
-                if _base and _base.rstrip("/").endswith("/anthropic"):
-                    from agent.auxiliary_client import _to_openai_base_url
-                    _base = _to_openai_base_url(_base)
-                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}",
-                    "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(
-                    _url,
-                    headers=_headers,
-                    timeout=10,
-                )
-                if (
-                    _pname == "Alibaba/DashScope"
-                    and not _base
-                    and _resp.status_code == 401
-                ):
-                    _resp = httpx.get(
-                        "https://dashscope.aliyuncs.com/compatible-mode/v1/models",
-                        headers=_headers,
-                        timeout=10,
-                    )
-                if _resp.status_code == 200:
-                    print(f"\r  {color('✓', Colors.GREEN)} {_label}                          ")
-                elif _resp.status_code == 401:
-                    print(f"\r  {color('✗', Colors.RED)} {_label} {color('(invalid API key)', Colors.DIM)}           ")
-                    issues.append(f"Check {_env_vars[0]} in .env")
-                else:
-                    print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color(f'(HTTP {_resp.status_code})', Colors.DIM)}           ")
-            except Exception as _e:
-                print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color(f'({_e})', Colors.DIM)}           ")
+    for _entry in _APIKEY_PROVIDERS_CACHE:
+        _pname, _env_vars, _default_url, _base_env, _supports = _entry
+        # Capture loop vars by binding default args — without this, all closures
+        # would share the final iteration's values and every probe would hit
+        # the last provider's URL.
+        _probes.append((_pname, lambda p=_pname, e=_env_vars, u=_default_url,
+                                       b=_base_env, s=_supports:
+                                _probe_apikey_provider(p, e, u, b, s)))
 
-    # -- AWS Bedrock --
-    # Bedrock uses the AWS SDK credential chain, not API keys.
+    _probes.append(("AWS Bedrock", _probe_bedrock))
+
+    # Print a single status line so users see something happening, then
+    # fan out. ``\r`` clears it once the first real result line lands.
+    print(f"  {color(f'Running {len(_probes)} connectivity checks in parallel…', Colors.DIM)}",
+          end="", flush=True)
+
+    # Disable boto3's EC2 instance-metadata-service probe for the duration
+    # of the parallel block. boto's default credential chain tries
+    # 169.254.169.254 with a multi-second timeout when we're not on EC2,
+    # which dominated the section's wall time before this fix
+    # (~2s on a developer laptop, even with the rest parallelized).
+    # Set on the parent thread before submitting work so the env-var
+    # mutation never races with another worker. has_aws_credentials() in
+    # the bedrock probe already gates on real env-var creds, so IMDS is
+    # never the legitimate source for `hermes doctor`.
+    _imds_prev = os.environ.get("AWS_EC2_METADATA_DISABLED")
+    os.environ["AWS_EC2_METADATA_DISABLED"] = "true"
     try:
-        from agent.bedrock_adapter import has_aws_credentials, resolve_aws_auth_env_var, resolve_bedrock_region
-        if has_aws_credentials():
-            _auth_var = resolve_aws_auth_env_var()
-            _region = resolve_bedrock_region()
-            _label = "AWS Bedrock".ljust(20)
-            print(f"  Checking AWS Bedrock...", end="", flush=True)
-            try:
-                import boto3
-                _br_client = boto3.client("bedrock", region_name=_region)
-                _br_resp = _br_client.list_foundation_models()
-                _model_count = len(_br_resp.get("modelSummaries", []))
-                print(f"\r  {color('✓', Colors.GREEN)} {_label} {color(f'({_auth_var}, {_region}, {_model_count} models)', Colors.DIM)}           ")
-            except ImportError:
-                print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color(f'(boto3 not installed — {sys.executable} -m pip install boto3)', Colors.DIM)}           ")
-                issues.append(f"Install boto3 for Bedrock: {sys.executable} -m pip install boto3")
-            except Exception as _e:
-                _err_name = type(_e).__name__
-                print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color(f'({_err_name}: {_e})', Colors.DIM)}           ")
-                issues.append(f"AWS Bedrock: {_err_name} — check IAM permissions for bedrock:ListFoundationModels")
-    except ImportError:
-        pass  # bedrock_adapter not available — skip silently
+        # 8 workers is plenty — each probe is a single HTTP call plus a TLS
+        # handshake. More than that wastes thread-startup cost and risks
+        # noisy output if anything ever printed from inside a worker.
+        with _futures.ThreadPoolExecutor(max_workers=8,
+                                         thread_name_prefix="doctor-probe") as _ex:
+            _futures_in_order = [_ex.submit(_fn) for _, _fn in _probes]
+            _results = [_f.result() for _f in _futures_in_order]
+    finally:
+        if _imds_prev is None:
+            os.environ.pop("AWS_EC2_METADATA_DISABLED", None)
+        else:
+            os.environ["AWS_EC2_METADATA_DISABLED"] = _imds_prev
+
+    # Clear the "Running …" line and print all results in submission order.
+    print("\r" + " " * 70 + "\r", end="")
+    for _r in _results:
+        for _glyph, _label, _detail in _r.lines:
+            if _detail:
+                print(f"  {_glyph} {_label} {_detail}")
+            else:
+                print(f"  {_glyph} {_label}")
+        for _issue in _r.issues:
+            issues.append(_issue)
 
     # =========================================================================
     # Check: Submodules

hermes_cli/fallback_cmd.py

@@ -307,7 +307,7 @@ def cmd_fallback_clear(args) -> None:  # noqa: ARG001
         print()
         print("  Cancelled.")
         return
-    if resp not in ("y", "yes"):
+    if resp not in {"y", "yes"}:
         print("  Cancelled — no change.")
         return
 
@@ -347,11 +347,11 @@ def _numbered_pick(question: str, choices: List[str]) -> Optional[int]:
 def cmd_fallback(args) -> None:
     """Top-level dispatcher for ``hermes fallback [subcommand]``."""
     sub = getattr(args, "fallback_command", None)
-    if sub in (None, "", "list", "ls"):
+    if sub in {None, "", "list", "ls"}:
         cmd_fallback_list(args)
     elif sub == "add":
         cmd_fallback_add(args)
-    elif sub in ("remove", "rm"):
+    elif sub in {"remove", "rm"}:
         cmd_fallback_remove(args)
     elif sub == "clear":
         cmd_fallback_clear(args)

hermes_cli/gateway.py

@@ -394,42 +394,68 @@ def _scan_gateway_pids(exclude_pids: set[int], all_profiles: bool = False) -> li
                             pass
                     current_cmd = ""
         else:
-            result = subprocess.run(
-                ["ps", "-A", "eww", "-o", "pid=,command="],
-                capture_output=True,
-                text=True,
-                timeout=10,
-            )
-            if result.returncode != 0:
-                return []
-            for line in result.stdout.split("\n"):
-                stripped = line.strip()
-                if not stripped or "grep" in stripped:
-                    continue
+            # Try /proc first (works in Docker without procps installed),
+            # fall back to ps -A eww.
+            _found_via_proc = False
+            if os.path.isdir("/proc"):
+                try:
+                    my_pid = os.getpid()
+                    for entry in os.listdir("/proc"):
+                        if not entry.isdigit():
+                            continue
+                        pid = int(entry)
+                        if pid == my_pid or pid in exclude_pids:
+                            continue
+                        try:
+                            cmdline = open(f"/proc/{pid}/cmdline", "rb").read().decode("utf-8", errors="replace")
+                            cmdline = cmdline.replace("\x00", " ")
+                            if any(p in cmdline for p in patterns) and (
+                                all_profiles or _matches_current_profile(cmdline)
+                            ):
+                                _append_unique_pid(pids, pid, exclude_pids)
+                        except (OSError, PermissionError):
+                            continue
+                    _found_via_proc = True
+                except Exception:
+                    pass
 
-                pid = None
-                command = ""
+            if not _found_via_proc:
+                result = subprocess.run(
+                    ["ps", "-A", "eww", "-o", "pid=,command="],
+                    capture_output=True,
+                    text=True,
+                    timeout=10,
+                )
+                if result.returncode != 0:
+                    return []
+                for line in result.stdout.split("\n"):
+                    stripped = line.strip()
+                    if not stripped or "grep" in stripped:
+                        continue
 
-                parts = stripped.split(None, 1)
-                if len(parts) == 2:
-                    try:
-                        pid = int(parts[0])
-                        command = parts[1]
-                    except ValueError:
-                        pid = None
+                    pid = None
+                    command = ""
 
-                if pid is None:
-                    aux_parts = stripped.split()
-                    if len(aux_parts) > 10 and aux_parts[1].isdigit():
-                        pid = int(aux_parts[1])
-                        command = " ".join(aux_parts[10:])
+                    parts = stripped.split(None, 1)
+                    if len(parts) == 2:
+                        try:
+                            pid = int(parts[0])
+                            command = parts[1]
+                        except ValueError:
+                            pid = None
 
-                if pid is None:
-                    continue
-                if any(pattern in command for pattern in patterns) and (
-                    all_profiles or _matches_current_profile(command)
-                ):
-                    _append_unique_pid(pids, pid, exclude_pids)
+                    if pid is None:
+                        aux_parts = stripped.split()
+                        if len(aux_parts) > 10 and aux_parts[1].isdigit():
+                            pid = int(aux_parts[1])
+                            command = " ".join(aux_parts[10:])
+
+                    if pid is None:
+                        continue
+                    if any(pattern in command for pattern in patterns) and (
+                        all_profiles or _matches_current_profile(command)
+                    ):
+                        _append_unique_pid(pids, pid, exclude_pids)
     except (OSError, subprocess.TimeoutExpired):
         return []
 
@@ -635,6 +661,66 @@ def _probe_systemd_service_running(system: bool = False) -> tuple[bool, bool]:
     return selected_system, result.stdout.strip() == "active"
 
 
+def _read_systemd_unit_environment(system: bool = False) -> dict[str, str]:
+    """Parse the gateway unit's ``Environment=`` directives.
+
+    ``systemctl show -p Environment`` returns a single line of
+    space-separated ``KEY=VALUE`` pairs; values are not quoted in the output
+    even when the unit file quoted them. We split on whitespace and ``=``.
+    """
+    selected_system = _select_systemd_scope(system)
+    try:
+        result = _run_systemctl(
+            [
+                "show",
+                get_service_name(),
+                "--no-pager",
+                "--property",
+                "Environment",
+            ],
+            system=selected_system,
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+    except (RuntimeError, subprocess.TimeoutExpired, OSError):
+        return {}
+    if result.returncode != 0:
+        return {}
+    parsed: dict[str, str] = {}
+    for line in result.stdout.splitlines():
+        if not line.startswith("Environment="):
+            continue
+        body = line[len("Environment="):].strip()
+        for token in body.split():
+            if "=" not in token:
+                continue
+            key, value = token.split("=", 1)
+            parsed[key] = value
+    return parsed
+
+
+def _sync_hermes_home_from_systemd_unit(system: bool) -> None:
+    """When acting on a system-scope unit, adopt its ``HERMES_HOME``.
+
+    Under ``sudo``, ``HERMES_HOME`` is stripped and ``HOME=/root``, so
+    :func:`get_hermes_home` falls back to ``/root/.hermes`` — the wrong
+    profile. The unit file pins ``HERMES_HOME`` for the actual gateway
+    process, so we mirror that into our own environment to make
+    ``read_runtime_status`` / ``get_running_pid`` read the correct files.
+    """
+    if not system:
+        return
+    env = _read_systemd_unit_environment(system=True)
+    unit_home = env.get("HERMES_HOME", "").strip()
+    if not unit_home:
+        return
+    current = os.environ.get("HERMES_HOME", "").strip()
+    if current == unit_home:
+        return
+    os.environ["HERMES_HOME"] = unit_home
+
+
 def _read_systemd_unit_properties(
     system: bool = False,
     properties: tuple[str, ...] = (
@@ -1108,7 +1194,7 @@ def _systemd_operational(system: bool = False) -> bool:
         )
         # "running", "degraded", "starting" all mean systemd is PID 1
         status = result.stdout.strip().lower()
-        return status in ("running", "degraded", "starting", "initializing")
+        return status in {"running", "degraded", "starting", "initializing"}
     except (RuntimeError, subprocess.TimeoutExpired, OSError):
         return False
 
@@ -1141,6 +1227,27 @@ def is_windows() -> bool:
     return sys.platform == 'win32'
 
 
+def _windows_gateway_should_absorb_console_controls() -> bool:
+    """Return True for detached Windows gateway runs that should ignore Ctrl+C.
+
+    Foreground ``hermes gateway run`` must remain interruptible from
+    PowerShell/CMD. Detached service-style launches opt in via
+    ``HERMES_GATEWAY_DETACHED=1``; older wrappers without the env marker are
+    treated as detached when no interactive stdin is attached.
+    """
+    if not is_windows():
+        return False
+
+    detached = os.getenv("HERMES_GATEWAY_DETACHED", "").strip().lower()
+    if detached in {"1", "true", "yes", "on"}:
+        return True
+
+    try:
+        return not bool(sys.stdin and sys.stdin.isatty())
+    except (ValueError, OSError):
+        return True
+
+
 # =============================================================================
 # Service Configuration
 # =============================================================================
@@ -2149,7 +2256,30 @@ def refresh_systemd_unit_if_needed(system: bool = False) -> bool:
         return False
 
     expected_user = _read_systemd_user_from_unit(unit_path) if system else None
-    unit_path.write_text(generate_systemd_unit(system=system, run_as_user=expected_user), encoding="utf-8")
+    new_unit = generate_systemd_unit(system=system, run_as_user=expected_user)
+
+    # ── Test-environment safety belt ─────────────────────────────────────
+    # The user-scope unit path resolves under ``Path.home()``, which is NOT
+    # sandboxed by the test conftest (only HERMES_HOME is). If a test
+    # exercises ``run_gateway()`` with a pytest-tmp HERMES_HOME, the freshly
+    # generated unit bakes that ``/tmp/pytest-of-.../hermes_test`` path into
+    # ``Environment="HERMES_HOME=..."``. Writing that to the developer's
+    # real user systemd unit file silently breaks their gateway on the next
+    # reboot (systemd loads the polluted env, the gateway looks at an empty
+    # tmp dir, and Telegram/Discord/etc. all show as "not configured").
+    # Refuse to write when the generated unit references a pytest tmpdir.
+    # Detection sniffs the unit body — tests that legitimately exercise the
+    # refresh flow patch ``generate_systemd_unit`` to return synthetic
+    # content (``"new unit\n"``) which doesn't contain these markers and
+    # still works.
+    if not system and (
+        "/pytest-of-" in new_unit
+        or "/hermes_test\"" in new_unit
+        or "/hermes_test/" in new_unit
+    ):
+        return False
+
+    unit_path.write_text(new_unit, encoding="utf-8")
     _run_systemctl(["daemon-reload"], system=system, check=True, timeout=30)
     print(f"↻ Updated gateway {_service_scope_label(system)} service definition to match the current Hermes install")
     return True
@@ -2380,6 +2510,7 @@ def systemd_stop(system: bool = False):
     if system:
         _require_root_for_system_service("stop")
     _require_service_installed("stop", system=system)
+    _sync_hermes_home_from_systemd_unit(system=system)
     try:
         from gateway.status import get_running_pid, write_planned_stop_marker
         pid = get_running_pid(cleanup_stale=False)
@@ -2408,6 +2539,7 @@ def systemd_restart(system: bool = False):
         _preflight_user_systemd()
     _require_service_installed("restart", system=system)
     refresh_systemd_unit_if_needed(system=system)
+    _sync_hermes_home_from_systemd_unit(system=system)
     from gateway.status import get_running_pid
 
     pid = get_running_pid() or _systemd_main_pid(system=system)
@@ -2503,6 +2635,8 @@ def systemd_status(deep: bool = False, system: bool = False, full: bool = False)
         print(f"  Run: {'sudo ' if system else ''}hermes gateway install{scope_flag}")
         return
 
+    _sync_hermes_home_from_systemd_unit(system=system)
+
     if has_conflicting_systemd_units():
         print_systemd_scope_conflict_warning()
         print()
@@ -2781,7 +2915,7 @@ def launchd_start():
     try:
         subprocess.run(["launchctl", "kickstart", f"{_launchd_domain()}/{label}"], check=True, timeout=30)
     except subprocess.CalledProcessError as e:
-        if e.returncode not in (3, 113):
+        if e.returncode not in {3, 113}:
             raise
         print("↻ launchd job was unloaded; reloading service definition")
         subprocess.run(["launchctl", "bootstrap", _launchd_domain(), str(plist_path)], check=True, timeout=30)
@@ -2805,7 +2939,7 @@ def launchd_stop():
     try:
         subprocess.run(["launchctl", "bootout", target], check=True, timeout=90)
     except subprocess.CalledProcessError as e:
-        if e.returncode in (3, 113):
+        if e.returncode in {3, 113}:
             pass  # Already unloaded — nothing to stop.
         else:
             raise
@@ -2877,7 +3011,7 @@ def launchd_restart():
         subprocess.run(["launchctl", "kickstart", "-k", target], check=True, timeout=90)
         print("✓ Service restarted")
     except subprocess.CalledProcessError as e:
-        if e.returncode not in (3, 113):
+        if e.returncode not in {3, 113}:
             raise
         # Job not loaded — bootstrap and start fresh
         print("↻ launchd job was unloaded; reloading")
@@ -2978,34 +3112,17 @@ def run_gateway(verbose: int = 0, quiet: bool = False, replace: bool = False):
     _guard_official_docker_root_gateway()
     sys.path.insert(0, str(PROJECT_ROOT))
 
-    # On Windows, when the gateway is launched as a detached background
-    # process (via ``hermes gateway install`` → Scheduled Task / Startup
-    # folder / direct pythonw.exe spawn) there is no console attached. In
-    # that case Windows can still deliver CTRL_C_EVENT / CTRL_BREAK_EVENT
-    # to the process group under some circumstances (e.g. when *another*
-    # process in the same group sends one), which Python 3.11 translates
-    # into KeyboardInterrupt inside asyncio.run(). The outer handler below
-    # catches that and exits cleanly — silently killing the gateway. On
-    # detached boots we must absorb those spurious signals so the gateway
-    # stays alive; real user Ctrl+C still comes through prompt_toolkit /
-    # the asyncio signal handler when running in a real console.
-    #
-    # IMPORTANT lesson (May 2026): we originally gated this on "stdin is
-    # NOT a TTY" assuming only detached pythonw runs would be vulnerable.
-    # Wrong. When the user runs `hermes gateway start` from a PowerShell
-    # console, the gateway inherits that console and stdin IS a TTY —
-    # but it's STILL vulnerable to CTRL_C_EVENT broadcast by any sibling
-    # `hermes` invocation (like `hermes gateway status` 30 seconds later)
-    # because Windows routes console events to all processes sharing the
-    # console. Every hermes CLI process after that sibling fires is a
-    # potential drive-by killer. So on Windows, for `gateway run`
-    # specifically (never interactive by design), always install the
-    # SIGINT absorber regardless of TTY state.
+    # Detached Windows gateway runs must ignore console-control broadcasts
+    # from sibling CLI processes, but foreground `hermes gateway run` still
+    # needs to obey the banner's "Press Ctrl+C to stop" contract.
+    # Service-style launchers set HERMES_GATEWAY_DETACHED=1; older wrappers
+    # without the marker are handled by the non-TTY fallback.
     try:
         _stdin_is_tty = bool(sys.stdin and sys.stdin.isatty())
     except (ValueError, OSError):
         _stdin_is_tty = False
-    if is_windows():
+    _absorb_windows_console_controls = _windows_gateway_should_absorb_console_controls()
+    if _absorb_windows_console_controls:
         try:
             signal.signal(signal.SIGINT, signal.SIG_IGN)
             if hasattr(signal, "SIGBREAK"):
@@ -3103,6 +3220,7 @@ def run_gateway(verbose: int = 0, quiet: bool = False, replace: bool = False):
         replace=replace,
         argv=sys.argv,
         stdin_is_tty=_stdin_is_tty,
+        absorb_windows_console_controls=_absorb_windows_console_controls,
     )
 
     def _atexit_hook() -> None:
@@ -3631,7 +3749,7 @@ def _platform_status(platform: dict) -> str:
         password = get_env_value("MATRIX_PASSWORD")
         if (val or password) and homeserver:
             e2ee = get_env_value("MATRIX_ENCRYPTION")
-            suffix = " + E2EE" if e2ee and e2ee.lower() in ("true", "1", "yes") else ""
+            suffix = " + E2EE" if e2ee and e2ee.lower() in {"true", "1", "yes"} else ""
             return f"configured{suffix}"
         if val or password or homeserver:
             return "partially configured"
@@ -4829,15 +4947,14 @@ def gateway_setup():
                 print_info("  Run in foreground: hermes gateway run")
                 print_info("  For persistence:   tmux new -s hermes 'hermes gateway run'")
                 print_info("  To enable systemd: add systemd=true to /etc/wsl.conf, then 'wsl --shutdown'")
+            elif is_termux():
+                from hermes_constants import display_hermes_home as _dhh
+                print_info("  Termux does not use systemd/launchd services.")
+                print_info("  Run in foreground: hermes gateway run")
+                print_info(f"  Or start it manually in the background (best effort): nohup hermes gateway run >{_dhh()}/logs/gateway.log 2>&1 &")
             else:
-                if is_termux():
-                    from hermes_constants import display_hermes_home as _dhh
-                    print_info("  Termux does not use systemd/launchd services.")
-                    print_info("  Run in foreground: hermes gateway run")
-                    print_info(f"  Or start it manually in the background (best effort): nohup hermes gateway run >{_dhh()}/logs/gateway.log 2>&1 &")
-                else:
-                    print_info("  Service install not supported on this platform.")
-                    print_info("  Run in foreground: hermes gateway run")
+                print_info("  Service install not supported on this platform.")
+                print_info("  Run in foreground: hermes gateway run")
     else:
         print()
         print_info("No platforms configured. Run 'hermes gateway setup' when ready.")

hermes_cli/gateway_windows.py

@@ -216,6 +216,7 @@ def _build_gateway_cmd_script(
     lines.append(f"cd /d {_quote_cmd_script_arg(working_dir)}")
     lines.append(f'set "HERMES_HOME={hermes_home}"')
     lines.append('set "PYTHONIOENCODING=utf-8"')
+    lines.append('set "HERMES_GATEWAY_DETACHED=1"')
     # VIRTUAL_ENV lets the gateway's own python detection find the venv
     # if someone imports hermes_constants-based logic during startup.
     venv_dir = str(Path(python_path).resolve().parent.parent)
@@ -371,6 +372,7 @@ def _build_gateway_argv() -> tuple[list[str], str, dict[str, str]]:
     env_overlay = {
         "HERMES_HOME": hermes_home,
         "PYTHONIOENCODING": "utf-8",
+        "HERMES_GATEWAY_DETACHED": "1",
         "VIRTUAL_ENV": str(Path(python_exe).resolve().parent.parent),
     }
     return argv, working_dir, env_overlay

hermes_cli/goals.py

@@ -270,7 +270,7 @@ def _parse_judge_response(raw: str) -> Tuple[bool, str, bool]:
 
     done_val = data.get("done")
     if isinstance(done_val, str):
-        done = done_val.strip().lower() in ("true", "yes", "1", "done")
+        done = done_val.strip().lower() in {"true", "yes", "1", "done"}
     else:
         done = bool(done_val)
     reason = str(data.get("reason") or "").strip()
@@ -389,11 +389,11 @@ class GoalManager:
         return self._state is not None and self._state.status == "active"
 
     def has_goal(self) -> bool:
-        return self._state is not None and self._state.status in ("active", "paused")
+        return self._state is not None and self._state.status in {"active", "paused"}
 
     def status_line(self) -> str:
         s = self._state
-        if s is None or s.status in ("cleared",):
+        if s is None or s.status in {"cleared",}:
             return "No active goal. Set one with /goal <text>."
         turns = f"{s.turns_used}/{s.max_turns} turns"
         if s.status == "active":

hermes_cli/hooks.py

@@ -32,11 +32,11 @@ def hooks_command(args) -> None:
         print("Run 'hermes hooks --help' for details.")
         return
 
-    if sub in ("list", "ls"):
+    if sub in {"list", "ls"}:
         _cmd_list(args)
     elif sub == "test":
         _cmd_test(args)
-    elif sub in ("revoke", "remove", "rm"):
+    elif sub in {"revoke", "remove", "rm"}:
         _cmd_revoke(args)
     elif sub == "doctor":
         _cmd_doctor(args)
@@ -220,7 +220,7 @@ def _cmd_test(args) -> None:
     if getattr(args, "for_tool", None):
         specs = [
             s for s in specs
-            if s.event not in ("pre_tool_call", "post_tool_call")
+            if s.event not in {"pre_tool_call", "post_tool_call"}
             or s.matches_tool(args.for_tool)
         ]
 

hermes_cli/kanban.py

@@ -82,7 +82,7 @@ def _parse_workspace_flag(value: str) -> tuple[str, Optional[str]]:
     if not value:
         return ("scratch", None)
     v = value.strip()
-    if v in ("scratch", "worktree"):
+    if v in {"scratch", "worktree"}:
         return (v, None)
     if v.startswith("dir:"):
         path = v[len("dir:"):].strip()
@@ -510,6 +510,10 @@ def build_parser(parent_subparsers: argparse._SubParsersAction) -> argparse.Argu
     p_nsub.add_argument("--chat-id", required=True)
     p_nsub.add_argument("--thread-id", default=None)
     p_nsub.add_argument("--user-id", default=None)
+    p_nsub.add_argument(
+        "--notifier-profile", default=None,
+        help="Profile gateway that owns/delivers this subscription (default: active profile)",
+    )
 
     p_nlist = sub.add_parser(
         "notify-list",
@@ -648,6 +652,16 @@ def kanban_command(args: argparse.Namespace) -> int:
     # keeps the patch small and inherits the exact same resolution the
     # dispatcher uses for workers — consistency is a feature here.
     board_override = getattr(args, "board", None)
+    prev_board_env = os.environ.get("HERMES_KANBAN_BOARD")
+    restore_board_env = False
+
+    def _restore_board_env() -> None:
+        if not restore_board_env:
+            return
+        if prev_board_env is None:
+            os.environ.pop("HERMES_KANBAN_BOARD", None)
+        else:
+            os.environ["HERMES_KANBAN_BOARD"] = prev_board_env
     if board_override:
         try:
             normed = kb._normalize_board_slug(board_override)
@@ -667,12 +681,16 @@ def kanban_command(args: argparse.Namespace) -> int:
             )
             return 1
         os.environ["HERMES_KANBAN_BOARD"] = normed
+        restore_board_env = True
 
     # Boards management doesn't touch the DB at all — dispatch early so
     # fresh installs that haven't initialized any DB can still use
     # `hermes kanban boards create …`.
     if action == "boards":
-        return _dispatch_boards(args)
+        try:
+            return _dispatch_boards(args)
+        finally:
+            _restore_board_env()
 
     # Auto-initialize the DB before dispatching any subcommand. init_db
     # is idempotent, so running it every invocation is cheap (one
@@ -685,6 +703,7 @@ def kanban_command(args: argparse.Namespace) -> int:
         kb.init_db()
     except Exception as exc:
         print(f"kanban: could not initialize database: {exc}", file=sys.stderr)
+        _restore_board_env()
         return 1
 
     handlers = {
@@ -726,12 +745,16 @@ def kanban_command(args: argparse.Namespace) -> int:
     handler = handlers.get(action)
     if not handler:
         print(f"kanban: unknown action {action!r}", file=sys.stderr)
+        _restore_board_env()
         return 2
     try:
         return int(handler(args) or 0)
     except (ValueError, RuntimeError) as exc:
         print(f"kanban: {exc}", file=sys.stderr)
+        _restore_board_env()
         return 1
+    finally:
+        _restore_board_env()
 
 
 # ---------------------------------------------------------------------------
@@ -765,15 +788,15 @@ def _dispatch_boards(args: argparse.Namespace) -> int:
     can still run ``boards create`` / ``boards list``.
     """
     sub = getattr(args, "boards_action", None) or "list"
-    if sub in ("list", "ls"):
+    if sub in {"list", "ls"}:
         return _cmd_boards_list(args)
-    if sub in ("create", "new"):
+    if sub in {"create", "new"}:
         return _cmd_boards_create(args)
-    if sub in ("rm", "remove", "delete"):
+    if sub in {"rm", "remove", "delete"}:
         return _cmd_boards_rm(args)
-    if sub in ("switch", "use"):
+    if sub in {"switch", "use"}:
         return _cmd_boards_switch(args)
-    if sub in ("show", "current"):
+    if sub in {"show", "current"}:
         return _cmd_boards_show(args)
     if sub == "rename":
         return _cmd_boards_rename(args)
@@ -1278,7 +1301,7 @@ def _cmd_show(args: argparse.Namespace) -> int:
 
 
 def _cmd_assign(args: argparse.Namespace) -> int:
-    profile = None if args.profile.lower() in ("none", "-", "null") else args.profile
+    profile = None if args.profile.lower() in {"none", "-", "null"} else args.profile
     with kb.connect() as conn:
         ok = kb.assign_task(conn, args.task_id, profile)
     if not ok:
@@ -1305,7 +1328,7 @@ def _cmd_reclaim(args: argparse.Namespace) -> int:
 
 
 def _cmd_reassign(args: argparse.Namespace) -> int:
-    profile = None if args.profile.lower() in ("none", "-", "null") else args.profile
+    profile = None if args.profile.lower() in {"none", "-", "null"} else args.profile
     with kb.connect() as conn:
         ok = kb.reassign_task(
             conn, args.task_id, profile,
@@ -1921,6 +1944,7 @@ def _cmd_notify_subscribe(args: argparse.Namespace) -> int:
             conn, task_id=args.task_id,
             platform=args.platform, chat_id=args.chat_id,
             thread_id=args.thread_id, user_id=args.user_id,
+            notifier_profile=args.notifier_profile or _profile_author(),
         )
     print(f"Subscribed {args.platform}:{args.chat_id}"
           + (f":{args.thread_id}" if args.thread_id else "")
@@ -1939,8 +1963,9 @@ def _cmd_notify_list(args: argparse.Namespace) -> int:
         return 0
     for s in subs:
         thr = f":{s['thread_id']}" if s.get("thread_id") else ""
+        owner = f"  owner={s['notifier_profile']}" if s.get("notifier_profile") else ""
         print(f"  {s['task_id']:10s}  {s['platform']}:{s['chat_id']}{thr}"
-              f"  (since event {s['last_event_id']})")
+              f"  (since event {s['last_event_id']}){owner}")
     return 0
 
 
@@ -2071,19 +2096,18 @@ def _cmd_specify(args: argparse.Namespace) -> int:
                 "reason": outcome.reason,
                 "new_title": outcome.new_title,
             }))
+        elif outcome.ok:
+            title_suffix = (
+                f" — retitled: {outcome.new_title!r}"
+                if outcome.new_title
+                else ""
+            )
+            print(f"Specified {outcome.task_id} → todo{title_suffix}")
         else:
-            if outcome.ok:
-                title_suffix = (
-                    f" — retitled: {outcome.new_title!r}"
-                    if outcome.new_title
-                    else ""
-                )
-                print(f"Specified {outcome.task_id} → todo{title_suffix}")
-            else:
-                print(
-                    f"kanban: specify {outcome.task_id}: {outcome.reason}",
-                    file=sys.stderr,
-                )
+            print(
+                f"kanban: specify {outcome.task_id}: {outcome.reason}",
+                file=sys.stderr,
+            )
     if not all_flag:
         return 0 if ok_count == 1 else 1
     # --all: succeed if at least one promotion landed; exit 1 only when
@@ -2136,6 +2160,29 @@ def _cmd_gc(args: argparse.Namespace) -> int:
 # Slash-command entry point (used by /kanban from CLI and gateway)
 # ---------------------------------------------------------------------------
 
+_SLASH_KANBAN_HELP = """\
+**/kanban** — manage the shared task board.
+
+Common subcommands:
+  `list` (alias `ls`)   List tasks on the current board
+  `show <id>`           Task details + comments + events
+  `stats`               Per-status / per-assignee counts
+  `create <title>…`     Create a task (auto-subscribes you to events)
+  `comment <id> <msg>`  Append a comment
+  `complete <id>…`      Mark task(s) done
+  `block <id> [reason]` Mark blocked; `unblock <id>` to revive
+  `assign <id> <profile>`  Reassign
+  `boards list`         Show all boards
+  `assignees`           Known profiles + counts
+  `context <id>`        Full worker-context dump
+  `runs <id>`           Attempt history
+  `log <id>`            Worker log
+
+Run `/kanban <subcommand> -h` for arguments. \
+Read-only commands are safe while an agent is running.\
+"""
+
+
 def run_slash(rest: str) -> str:
     """Execute a ``/kanban …`` string and return captured stdout/stderr.
 
@@ -2148,26 +2195,47 @@ def run_slash(rest: str) -> str:
 
     tokens = shlex.split(rest) if rest and rest.strip() else []
 
-    parser = argparse.ArgumentParser(prog="/kanban", add_help=False)
-    parser.exit_on_error = False  # type: ignore[attr-defined]
-    sub = parser.add_subparsers(dest="kanban_action")
-    # Reuse the argparse builder -- call it with a throwaway parent
-    # subparsers via a wrapping top-level parser.
-    wrap = argparse.ArgumentParser(prog="/", add_help=False)
-    wrap.exit_on_error = False  # type: ignore[attr-defined]
-    wrap_sub = wrap.add_subparsers(dest="_top")
-    build_parser(wrap_sub)
+    # Bare ``/kanban`` or ``/kanban help`` / ``--help`` / ``-h`` / ``?``:
+    # show the curated short-help block instead of dumping argparse's full
+    # usage tree (which is enormous and reads as garbage in a chat
+    # bubble).  Per-subcommand help still works via ``/kanban foo -h``.
+    if not tokens or tokens[0] in {"help", "--help", "-h", "?"}:
+        return _SLASH_KANBAN_HELP
+
+    # Single argparse tree rooted at "/kanban".  build_parser() expects a
+    # subparsers action to attach to, so build a throwaway one and pull
+    # the kanban_parser back out — then drive it directly so usage/error
+    # text reads as ``/kanban`` (not ``/kanban-wrap kanban``).
+    _wrap = argparse.ArgumentParser(prog="/kanban-wrap", add_help=False)
+    _wrap.exit_on_error = False  # type: ignore[attr-defined]
+    _top_sub = _wrap.add_subparsers(dest="_top")
+    kanban_parser = build_parser(_top_sub)
+    kanban_parser.prog = "/kanban"
+    kanban_parser.exit_on_error = False  # type: ignore[attr-defined]
+    for _action in kanban_parser._actions:
+        if isinstance(_action, argparse._SubParsersAction):
+            for _name, _choice in _action.choices.items():
+                _choice.prog = f"/kanban {_name}"
+                _choice.exit_on_error = False  # type: ignore[attr-defined]
 
     buf_out = io.StringIO()
     buf_err = io.StringIO()
+    # ``-h`` / ``--help`` makes argparse print to stdout and SystemExit(0).
+    # Capture both streams so neither the help text nor the error text
+    # bypasses our buffer.
     try:
-        # Prepend the "kanban" token so our top-level subparser routes here.
-        argv = ["kanban", *tokens] if tokens else ["kanban"]
-        args = wrap.parse_args(argv)
+        with contextlib.redirect_stdout(buf_out), contextlib.redirect_stderr(buf_err):
+            args = kanban_parser.parse_args(tokens)
     except SystemExit as exc:
-        return f"(usage error: {exc})"
+        out = buf_out.getvalue().rstrip()
+        err = buf_err.getvalue().rstrip()
+        # Help dump (exit 0) → return the captured help text directly.
+        if exc.code in {0, None} and out:
+            return out
+        body = err or out
+        return f"⚠ /kanban usage error\n{body}" if body else "⚠ /kanban usage error"
     except argparse.ArgumentError as exc:
-        return f"(usage error: {exc})"
+        return f"⚠ /kanban usage error: {exc}"
 
     with contextlib.redirect_stdout(buf_out), contextlib.redirect_stderr(buf_err):
         try:

hermes_cli/kanban_db.py

@@ -83,6 +83,8 @@ from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Iterable, Optional
 
+from toolsets import get_toolset_names
+
 
 # ---------------------------------------------------------------------------
 # Constants
@@ -90,6 +92,7 @@ from typing import Any, Iterable, Optional
 
 VALID_STATUSES = {"triage", "todo", "ready", "running", "blocked", "done", "archived"}
 VALID_WORKSPACE_KINDS = {"scratch", "worktree", "dir"}
+KNOWN_TOOLSET_NAMES = frozenset(name.casefold() for name in get_toolset_names())
 
 # A running task's claim is valid for 15 minutes; after that the next
 # dispatcher tick reclaims it.  Workers that outlive this window should call
@@ -858,6 +861,7 @@ CREATE TABLE IF NOT EXISTS kanban_notify_subs (
     chat_id       TEXT NOT NULL,
     thread_id     TEXT NOT NULL DEFAULT '',
     user_id       TEXT,
+    notifier_profile TEXT,
     created_at    INTEGER NOT NULL,
     last_event_id INTEGER NOT NULL DEFAULT 0,
     PRIMARY KEY (task_id, platform, chat_id, thread_id)
@@ -963,6 +967,25 @@ def init_db(
     return path
 
 
+def _add_column_if_missing(
+    conn: sqlite3.Connection, table: str, column: str, ddl: str
+) -> bool:
+    """Run ``ALTER TABLE <table> ADD COLUMN <ddl>``, idempotent across races.
+
+    Returns ``True`` when the column was actually added by this call.
+    Swallows ``duplicate column name`` errors so a concurrent connection
+    that ran the same migration first does not crash the dispatcher tick
+    (issue #21708).
+    """
+    try:
+        conn.execute(f"ALTER TABLE {table} ADD COLUMN {ddl}")
+        return True
+    except sqlite3.OperationalError as exc:
+        if "duplicate column name" in str(exc).lower():
+            return False
+        raise
+
+
 def _migrate_add_optional_columns(conn: sqlite3.Connection) -> None:
     """Add columns that were introduced after v1 release to legacy DBs.
 
@@ -970,11 +993,13 @@ def _migrate_add_optional_columns(conn: sqlite3.Connection) -> None:
     """
     cols = {row["name"] for row in conn.execute("PRAGMA table_info(tasks)")}
     if "tenant" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN tenant TEXT")
+        _add_column_if_missing(conn, "tasks", "tenant", "tenant TEXT")
     if "result" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN result TEXT")
+        _add_column_if_missing(conn, "tasks", "result", "result TEXT")
     if "idempotency_key" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN idempotency_key TEXT")
+        _add_column_if_missing(
+            conn, "tasks", "idempotency_key", "idempotency_key TEXT"
+        )
         conn.execute(
             "CREATE INDEX IF NOT EXISTS idx_tasks_idempotency "
             "ON tasks(idempotency_key)"
@@ -997,37 +1022,51 @@ def _migrate_add_optional_columns(conn: sqlite3.Connection) -> None:
     # the *original* snapshot; this is intentional and safe as long as
     # no step depends on a column added by a previous step in the same call.
     if "consecutive_failures" not in cols:
-        conn.execute(
-            "ALTER TABLE tasks ADD COLUMN consecutive_failures "
-            "INTEGER NOT NULL DEFAULT 0"
+        added = _add_column_if_missing(
+            conn,
+            "tasks",
+            "consecutive_failures",
+            "consecutive_failures INTEGER NOT NULL DEFAULT 0",
         )
-        if "spawn_failures" in cols:
+        if added and "spawn_failures" in cols:
             conn.execute(
                 "UPDATE tasks SET consecutive_failures = COALESCE(spawn_failures, 0)"
             )
     if "worker_pid" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN worker_pid INTEGER")
+        _add_column_if_missing(conn, "tasks", "worker_pid", "worker_pid INTEGER")
     if "last_failure_error" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN last_failure_error TEXT")
-        if "last_spawn_error" in cols:
+        added = _add_column_if_missing(
+            conn, "tasks", "last_failure_error", "last_failure_error TEXT"
+        )
+        if added and "last_spawn_error" in cols:
             conn.execute(
                 "UPDATE tasks SET last_failure_error = last_spawn_error"
             )
     if "max_runtime_seconds" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN max_runtime_seconds INTEGER")
+        _add_column_if_missing(
+            conn, "tasks", "max_runtime_seconds", "max_runtime_seconds INTEGER"
+        )
     if "last_heartbeat_at" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN last_heartbeat_at INTEGER")
+        _add_column_if_missing(
+            conn, "tasks", "last_heartbeat_at", "last_heartbeat_at INTEGER"
+        )
     if "current_run_id" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN current_run_id INTEGER")
+        _add_column_if_missing(
+            conn, "tasks", "current_run_id", "current_run_id INTEGER"
+        )
     if "workflow_template_id" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN workflow_template_id TEXT")
+        _add_column_if_missing(
+            conn, "tasks", "workflow_template_id", "workflow_template_id TEXT"
+        )
     if "current_step_key" not in cols:
-        conn.execute("ALTER TABLE tasks ADD COLUMN current_step_key TEXT")
+        _add_column_if_missing(
+            conn, "tasks", "current_step_key", "current_step_key TEXT"
+        )
     if "skills" not in cols:
         # JSON array of skill names the dispatcher force-loads into the
         # worker (additive to the built-in `kanban-worker`). NULL is fine
         # for existing rows.
-        conn.execute("ALTER TABLE tasks ADD COLUMN skills TEXT")
+        _add_column_if_missing(conn, "tasks", "skills", "skills TEXT")
 
     if "max_retries" not in cols:
         # Per-task override for the consecutive-failure circuit breaker.
@@ -1035,18 +1074,30 @@ def _migrate_add_optional_columns(conn: sqlite3.Connection) -> None:
         # config, then ``DEFAULT_FAILURE_LIMIT``. Existing rows get NULL,
         # which is the correct default (they keep the global behaviour
         # they were getting before the column existed).
-        conn.execute("ALTER TABLE tasks ADD COLUMN max_retries INTEGER")
+        _add_column_if_missing(conn, "tasks", "max_retries", "max_retries INTEGER")
 
     # task_events gained a run_id column; back-fill it as NULL for
     # historical events (they predate runs and can't be attributed).
     ev_cols = {row["name"] for row in conn.execute("PRAGMA table_info(task_events)")}
     if "run_id" not in ev_cols:
-        conn.execute("ALTER TABLE task_events ADD COLUMN run_id INTEGER")
+        _add_column_if_missing(conn, "task_events", "run_id", "run_id INTEGER")
         conn.execute(
             "CREATE INDEX IF NOT EXISTS idx_events_run "
             "ON task_events(run_id, id)"
         )
 
+    notify_table_exists = conn.execute(
+        "SELECT name FROM sqlite_master WHERE type='table' AND name='kanban_notify_subs'"
+    ).fetchone() is not None
+    if notify_table_exists:
+        notify_cols = {
+            row["name"] for row in conn.execute("PRAGMA table_info(kanban_notify_subs)")
+        }
+        if "notifier_profile" not in notify_cols:
+            _add_column_if_missing(
+                conn, "kanban_notify_subs", "notifier_profile", "notifier_profile TEXT"
+            )
+
     # One-shot backfill: any task that is 'running' before runs existed
     # had its claim_lock / claim_expires / worker_pid on the task row.
     # Synthesize a matching task_runs row so subsequent end-run / heartbeat
@@ -1237,6 +1288,12 @@ def create_task(
     if skills is not None:
         cleaned: list[str] = []
         seen: set[str] = set()
+        # Collect all toolset-name confusions up front so the user sees the
+        # whole list at once. Raising on the first hit is friendly when the
+        # input has one mistake, but agents that confuse skills with toolsets
+        # usually pass several at once (`skills=["web", "browser", "terminal"]`)
+        # and serial-correcting one per failure round-trips wastes tokens.
+        toolset_typos: list[str] = []
         for s in skills:
             if not s:
                 continue
@@ -1248,10 +1305,23 @@ def create_task(
                     f"skill name cannot contain comma: {name!r} "
                     f"(pass a list of separate names instead of a comma-joined string)"
                 )
+            if name.casefold() in KNOWN_TOOLSET_NAMES:
+                toolset_typos.append(name)
+                continue
             if name in seen:
                 continue
             seen.add(name)
             cleaned.append(name)
+        if toolset_typos:
+            quoted = ", ".join(repr(n) for n in toolset_typos)
+            noun = "is a toolset name" if len(toolset_typos) == 1 else "are toolset names"
+            raise ValueError(
+                f"{quoted} {noun}, not skill name(s). "
+                "Put toolsets in the assignee profile's `toolsets:` config "
+                "instead of per-task skills. Skills are named skill bundles "
+                "(e.g. `kanban-worker`, `blogwatcher`); toolsets are runtime "
+                "capabilities (e.g. `web`, `browser`, `terminal`)."
+            )
         skills_list = cleaned
 
     # Idempotency check — return the existing task instead of creating a
@@ -1504,7 +1574,14 @@ def unlink_tasks(conn: sqlite3.Connection, parent_id: str, child_id: str) -> boo
                 conn, child_id, "unlinked",
                 {"parent": parent_id, "child": child_id},
             )
-        return cur.rowcount > 0
+        removed = cur.rowcount > 0
+    if removed:
+        # Dependency edge removed — re-evaluate promotion eligibility for the
+        # child immediately.  Matches the contract of complete_task and
+        # unblock_task; without this the child stays stuck in todo until the
+        # next dispatcher tick or a manual `hermes kanban recompute` (issue #22459).
+        recompute_ready(conn)
+    return removed
 
 
 def parent_ids(conn: sqlite3.Connection, task_id: str) -> list[str]:
@@ -1749,7 +1826,7 @@ def _synthesize_ended_run(
 # ---------------------------------------------------------------------------
 
 def recompute_ready(conn: sqlite3.Connection) -> int:
-    """Promote ``todo`` tasks to ``ready`` when all parents are ``done``.
+    """Promote ``todo`` tasks to ``ready`` when all parents are ``done`` or ``archived``.
 
     Returns the number of tasks promoted.  Safe to call inside or outside
     an existing transaction; it opens its own IMMEDIATE txn.
@@ -1767,7 +1844,7 @@ def recompute_ready(conn: sqlite3.Connection) -> int:
                 "WHERE l.child_id = ?",
                 (task_id,),
             ).fetchall()
-            if all(p["status"] == "done" for p in parents):
+            if all(p["status"] in {"done", "archived"} for p in parents):
                 conn.execute(
                     "UPDATE tasks SET status = 'ready' WHERE id = ? AND status = 'todo'",
                     (task_id,),
@@ -1797,6 +1874,31 @@ def claim_task(
     lock = claimer or _claimer_id()
     expires = now + int(ttl_seconds)
     with write_txn(conn):
+        # Structural invariant: never transition ready -> running while any
+        # parent is not yet 'done'. This is the single enforcement point
+        # regardless of which writer (create_task, link_tasks, unblock_task,
+        # release_stale_claims, manual SQL) set status='ready'. If a racy
+        # writer promoted a task with undone parents, demote it back to
+        # 'todo' here — recompute_ready will re-promote when the parents
+        # actually finish. See RCA at
+        # kanban/boards/cookai/workspaces/t_a6acd07d/root-cause.md.
+        undone = conn.execute(
+            "SELECT 1 FROM task_links l "
+            "JOIN tasks p ON p.id = l.parent_id "
+            "WHERE l.child_id = ? AND p.status NOT IN ('done', 'archived') LIMIT 1",
+            (task_id,),
+        ).fetchone()
+        if undone:
+            conn.execute(
+                "UPDATE tasks SET status = 'todo' "
+                "WHERE id = ? AND status = 'ready'",
+                (task_id,),
+            )
+            _append_event(
+                conn, task_id, "claim_rejected",
+                {"reason": "parents_not_done"},
+            )
+            return None
         # Defensive: if a prior run somehow leaked (invariant violation from
         # an unknown code path), close it as 'reclaimed' so we don't strand
         # it when the CAS resets the pointer below. No-op when the invariant
@@ -1908,16 +2010,69 @@ def release_stale_claims(
 ) -> int:
     """Reset any ``running`` task whose claim has expired.
 
-    Returns the number of stale claims reclaimed.  Safe to call often.
+    A stale-by-TTL claim whose host-local worker PID is still alive is
+    *extended* (with a ``claim_extended`` event) instead of being
+    reclaimed. Reclaiming a live worker mid-flight produces the spawn-
+    then-immediately-reclaim loop seen on slow models that spend longer
+    than ``DEFAULT_CLAIM_TTL_SECONDS`` inside a single tool-free LLM
+    call (#23025): no tool calls means no ``kanban_heartbeat``, even
+    though the subprocess is healthy. ``enforce_max_runtime`` and
+    ``detect_crashed_workers`` remain the upper bounds for genuinely
+    wedged or dead workers.
+
+    Returns the number of stale claims actually reclaimed (live-pid
+    extensions don't count). Safe to call often.
     """
     now = int(time.time())
     reclaimed = 0
+    host_prefix = f"{_claimer_id().split(':', 1)[0]}:"
     stale = conn.execute(
-        "SELECT id, claim_lock, worker_pid FROM tasks "
-        "WHERE status = 'running' AND claim_expires IS NOT NULL AND claim_expires < ?",
+        "SELECT id, claim_lock, worker_pid, claim_expires, last_heartbeat_at "
+        "FROM tasks "
+        "WHERE status = 'running' AND claim_expires IS NOT NULL "
+        "  AND claim_expires < ?",
         (now,),
     ).fetchall()
     for row in stale:
+        lock = row["claim_lock"] or ""
+        host_local = lock.startswith(host_prefix)
+        if host_local and row["worker_pid"] and _pid_alive(row["worker_pid"]):
+            new_expires = now + int(DEFAULT_CLAIM_TTL_SECONDS)
+            with write_txn(conn):
+                cur = conn.execute(
+                    "UPDATE tasks SET claim_expires = ? "
+                    "WHERE id = ? AND status = 'running' "
+                    "  AND claim_lock IS ? "
+                    "  AND claim_expires IS NOT NULL "
+                    "  AND claim_expires < ?",
+                    (new_expires, row["id"], row["claim_lock"], now),
+                )
+                if cur.rowcount != 1:
+                    continue
+                run_id = _current_run_id(conn, row["id"])
+                if run_id is not None:
+                    conn.execute(
+                        "UPDATE task_runs SET claim_expires = ? WHERE id = ?",
+                        (new_expires, run_id),
+                    )
+                _append_event(
+                    conn, row["id"], "claim_extended",
+                    {
+                        "reason": "pid_alive",
+                        "worker_pid": int(row["worker_pid"]),
+                        "claim_lock": row["claim_lock"],
+                        "claim_expires_was": int(row["claim_expires"]),
+                        "claim_expires_now": new_expires,
+                        "last_heartbeat_at": (
+                            int(row["last_heartbeat_at"])
+                            if row["last_heartbeat_at"] is not None
+                            else None
+                        ),
+                    },
+                    run_id=run_id,
+                )
+            continue
+
         termination = _terminate_reclaimed_worker(
             row["worker_pid"], row["claim_lock"], signal_fn=signal_fn,
         )
@@ -1937,7 +2092,20 @@ def release_stale_claims(
                 error=f"stale_lock={row['claim_lock']}",
                 metadata=termination,
             )
-            payload = {"stale_lock": row["claim_lock"]}
+            payload = {
+                "stale_lock": row["claim_lock"],
+                "worker_pid": (
+                    int(row["worker_pid"])
+                    if row["worker_pid"] is not None else None
+                ),
+                "claim_expires": int(row["claim_expires"]),
+                "last_heartbeat_at": (
+                    int(row["last_heartbeat_at"])
+                    if row["last_heartbeat_at"] is not None else None
+                ),
+                "now": now,
+                "host_local": host_local,
+            }
             payload.update(termination)
             _append_event(
                 conn, row["id"], "reclaimed",
@@ -2496,14 +2664,30 @@ def unblock_task(conn: sqlite3.Connection, task_id: str) -> bool:
                 """,
                 (now, int(stale["current_run_id"])),
             )
-        cur = conn.execute(
-            "UPDATE tasks SET status = 'ready', current_run_id = NULL "
-            "WHERE id = ? AND status = 'blocked'",
+        # Re-gate on parent completion before flipping 'blocked' back to
+        # 'ready'. Unconditionally setting status='ready' here bypasses the
+        # parent-completion invariant (the dispatcher trusts that column);
+        # if parents are still in progress the task must wait in 'todo'
+        # until recompute_ready picks it up. RCA: Bug 2 at
+        # kanban/boards/cookai/workspaces/t_a6acd07d/root-cause.md.
+        undone_parents = conn.execute(
+            "SELECT 1 FROM task_links l "
+            "JOIN tasks p ON p.id = l.parent_id "
+            "WHERE l.child_id = ? AND p.status != 'done' LIMIT 1",
             (task_id,),
+        ).fetchone()
+        new_status = "todo" if undone_parents else "ready"
+        cur = conn.execute(
+            "UPDATE tasks SET status = ?, current_run_id = NULL "
+            "WHERE id = ? AND status = 'blocked'",
+            (new_status, task_id),
         )
         if cur.rowcount != 1:
             return False
-        _append_event(conn, task_id, "unblocked", None)
+        _append_event(
+            conn, task_id, "unblocked",
+            {"status": new_status} if new_status != "ready" else None,
+        )
         return True
 
 
@@ -3504,6 +3688,14 @@ def dispatch_once(
     failures the task is auto-blocked with the last error as its reason —
     prevents the dispatcher from thrashing forever on an unfixable task.
 
+    ``max_spawn`` is a **live concurrency cap**, not a per-tick spawn budget:
+    it counts tasks already in ``status='running'`` plus this tick's spawns
+    against the limit. So ``max_spawn=4`` means "at most 4 workers running
+    at any time across the whole board" — matching the gateway's stated
+    intent ("limit concurrent kanban tasks"). With a per-tick interpretation
+    a 60-second tick interval could grow concurrency by N every minute on a
+    busy board and accumulate without bound.
+
     ``spawn_fn`` defaults to ``_default_spawn``. Tests pass a stub.
     ``board`` pins workspace/log/db resolution for this tick to a specific
     board. When omitted, the current-board resolution chain is used.
@@ -3555,6 +3747,21 @@ def dispatch_once(
     result.timed_out = enforce_max_runtime(conn)
     result.promoted = recompute_ready(conn)
 
+    # Count tasks already running so max_spawn enforces concurrency rather
+    # than a per-tick spawn budget. See the docstring above for the full
+    # rationale; the short version is that a 60-second tick interval with a
+    # per-tick budget of N would grow concurrency by N every tick on a busy
+    # board, since "running" tasks aren't reclaimed by completion alone —
+    # they sit in status='running' until the worker calls
+    # kanban_complete/kanban_block (or the dispatcher TTL-reclaims them).
+    running_count = 0
+    if max_spawn is not None:
+        running_count = int(
+            conn.execute(
+                "SELECT COUNT(*) FROM tasks WHERE status = 'running'"
+            ).fetchone()[0]
+        )
+
     ready_rows = conn.execute(
         "SELECT id, assignee FROM tasks "
         "WHERE status = 'ready' AND claim_lock IS NULL "
@@ -3562,7 +3769,7 @@ def dispatch_once(
     ).fetchall()
     spawned = 0
     for row in ready_rows:
-        if max_spawn is not None and spawned >= max_spawn:
+        if max_spawn is not None and running_count + spawned >= max_spawn:
             break
         if not row["assignee"]:
             result.skipped_unassigned.append(row["id"])
@@ -3666,6 +3873,35 @@ def _rotate_worker_log(log_path: Path, max_bytes: int) -> None:
         pass
 
 
+def _resolve_hermes_argv() -> list[str]:
+    """Resolve the ``hermes`` invocation as argv parts for ``Popen``.
+
+    Tries in order:
+
+    1. ``shutil.which("hermes")`` — the console-script shim, the same form
+       that shows up in ``ps`` output and existing logs. Preferred so live
+       systems' diagnostics stay familiar.
+    2. ``sys.executable -m hermes_cli.main`` — fallback for setups where
+       Hermes is launched from a venv and the ``hermes`` shim is not on
+       the dispatcher's ``$PATH`` (cron, systemd ``User=`` services,
+       launchd jobs, detached processes, etc.). Goes through the running
+       interpreter so the result is independent of ``$PATH``.
+
+    Mirrors ``gateway.run._resolve_hermes_bin`` for the same reason. Kept
+    local (not imported from gateway) because ``hermes_cli`` sits below
+    ``gateway`` in the dependency order.
+    """
+    import shutil
+
+    hermes_bin = shutil.which("hermes")
+    if hermes_bin:
+        return [hermes_bin]
+    # Fallback to the module form. ``hermes_cli.main`` is the actual
+    # console-script target declared in pyproject.toml, NOT a top-level
+    # ``hermes`` package — there is no ``hermes`` package to import.
+    return [sys.executable, "-m", "hermes_cli.main"]
+
+
 def _default_spawn(
     task: Task,
     workspace: str,
@@ -3694,6 +3930,25 @@ def _default_spawn(
 
     prompt = f"work kanban task {task.id}"
     env = dict(os.environ)
+
+    # Inject HERMES_HOME so the worker reads the profile-scoped config.yaml
+    # (fallback_providers, toolsets, agent settings, etc.) instead of the root
+    # config.  Without this, `env = dict(os.environ)` copies only the parent's
+    # env, and when the child process starts `hermes -p <name>` the
+    # _apply_profile_override() runs *before* hermes_constants is imported.
+    # If HERMES_HOME is absent from the child's env, get_hermes_home() falls
+    # back to Path.home() / ".hermes" (the DEFAULT profile root), ignoring the
+    # profile-specific config entirely.  Fixes profile-scoped fallback_providers
+    # being invisible to kanban workers.
+    from hermes_cli.profiles import resolve_profile_env
+    try:
+        env["HERMES_HOME"] = resolve_profile_env(profile_arg)
+    except FileNotFoundError:
+        # Profile dir doesn't exist — defer resolution to the CLI's
+        # _apply_profile_override() via HERMES_PROFILE (set below).
+        # This only happens in test fixtures where the isolated
+        # HERMES_HOME never had profiles created.
+        pass
     if task.tenant:
         env["HERMES_TENANT"] = task.tenant
     env["HERMES_KANBAN_TASK"] = task.id
@@ -3722,7 +3977,7 @@ def _default_spawn(
     env["HERMES_PROFILE"] = profile_arg
 
     cmd = [
-        "hermes",
+        *_resolve_hermes_argv(),
         "-p", profile_arg,
         # Auto-load the kanban-worker skill so every dispatched worker
         # has the pattern library (good summary/metadata shapes, retry
@@ -4024,7 +4279,14 @@ def build_worker_context(conn: sqlite3.Connection, task_id: str) -> str:
             )
         for c in shown_c:
             ts = time.strftime("%Y-%m-%d %H:%M", time.localtime(c.created_at))
-            lines.append(f"**{c.author}** ({ts}):")
+            # Render author with explicit "comment from worker" framing so
+            # operator-controlled HERMES_PROFILE values like "hermes-system"
+            # or "operator" can't be misread by the next worker as a system
+            # directive above the (attacker-influenceable) comment body.
+            # Defense-in-depth — the LLM-controlled author-forgery surface
+            # was already closed in #22435. See #22452.
+            safe_author = (c.author or "").replace("`", "")
+            lines.append(f"comment from worker `{safe_author}` at {ts}:")
             lines.append(_cap(c.body, _CTX_MAX_COMMENT_BYTES))
             lines.append("")
 
@@ -4071,16 +4333,26 @@ def board_stats(conn: sqlite3.Connection) -> dict:
     }
 
 
+def _safe_int(val: Optional[str]) -> Optional[int]:
+    """Parse a timestamp field to int, returning None on garbage like '%s'."""
+    if val is None:
+        return None
+    try:
+        return int(val)
+    except (ValueError, TypeError):
+        return None
+
+
 def task_age(task: Task) -> dict:
     """Return age metrics for a single task. All values are seconds or None."""
     now = int(time.time())
-    age_since_created = now - int(task.created_at) if task.created_at else None
-    age_since_started = (
-        now - int(task.started_at) if task.started_at else None
-    )
+    created = _safe_int(task.created_at)
+    started = _safe_int(task.started_at)
+    completed = _safe_int(task.completed_at)
+    age_since_created = now - created if created else None
+    age_since_started = now - started if started else None
     time_to_complete = (
-        int(task.completed_at) - int(task.started_at or task.created_at)
-        if task.completed_at else None
+        completed - (started or created) if completed else None
     )
     return {
         "created_age_seconds": age_since_created,
@@ -4101,6 +4373,7 @@ def add_notify_sub(
     chat_id: str,
     thread_id: Optional[str] = None,
     user_id: Optional[str] = None,
+    notifier_profile: Optional[str] = None,
 ) -> None:
     """Register a gateway source that wants terminal-state notifications
     for ``task_id``. Idempotent on (task, platform, chat, thread)."""
@@ -4109,10 +4382,10 @@ def add_notify_sub(
         conn.execute(
             """
             INSERT OR IGNORE INTO kanban_notify_subs
-                (task_id, platform, chat_id, thread_id, user_id, created_at)
-            VALUES (?, ?, ?, ?, ?, ?)
+                (task_id, platform, chat_id, thread_id, user_id, notifier_profile, created_at)
+            VALUES (?, ?, ?, ?, ?, ?, ?)
             """,
-            (task_id, platform, chat_id, thread_id or "", user_id, now),
+            (task_id, platform, chat_id, thread_id or "", user_id, notifier_profile, now),
         )
 
 
@@ -4194,6 +4467,57 @@ def unseen_events_for_sub(
     return max_id, out
 
 
+def claim_unseen_events_for_sub(
+    conn: sqlite3.Connection,
+    *,
+    task_id: str,
+    platform: str,
+    chat_id: str,
+    thread_id: Optional[str] = None,
+    kinds: Optional[Iterable[str]] = None,
+) -> tuple[int, int, list[Event]]:
+    """Atomically claim unseen notification events for one subscription.
+
+    Returns ``(old_cursor, new_cursor, events)``. When events are returned,
+    ``kanban_notify_subs.last_event_id`` has already been advanced to
+    ``new_cursor`` inside a ``BEGIN IMMEDIATE`` transaction. That makes the
+    notifier's read/claim step single-owner across multiple gateway watcher
+    processes pointed at the same board DB: concurrent watchers serialize on
+    SQLite's writer lock, and only the first process sees and claims a given
+    event range.
+
+    Callers should send the claimed events, then either leave the cursor at
+    ``new_cursor`` on success or call :func:`rewind_notify_cursor` if delivery
+    failed before any terminal unsubscribe removed the row.
+    """
+    with write_txn(conn):
+        row = conn.execute(
+            "SELECT last_event_id FROM kanban_notify_subs "
+            "WHERE task_id = ? AND platform = ? AND chat_id = ? AND thread_id = ?",
+            (task_id, platform, chat_id, thread_id or ""),
+        ).fetchone()
+        if row is None:
+            return 0, 0, []
+        old_cursor = int(row["last_event_id"])
+        new_cursor, events = unseen_events_for_sub(
+            conn,
+            task_id=task_id,
+            platform=platform,
+            chat_id=chat_id,
+            thread_id=thread_id,
+            kinds=kinds,
+        )
+        if not events:
+            return old_cursor, old_cursor, []
+        conn.execute(
+            "UPDATE kanban_notify_subs SET last_event_id = ? "
+            "WHERE task_id = ? AND platform = ? AND chat_id = ? AND thread_id = ? "
+            "AND last_event_id = ?",
+            (int(new_cursor), task_id, platform, chat_id, thread_id or "", int(old_cursor)),
+        )
+        return old_cursor, new_cursor, events
+
+
 def advance_notify_cursor(
     conn: sqlite3.Connection,
     *,
@@ -4211,6 +4535,35 @@ def advance_notify_cursor(
         )
 
 
+def rewind_notify_cursor(
+    conn: sqlite3.Connection,
+    *,
+    task_id: str,
+    platform: str,
+    chat_id: str,
+    thread_id: Optional[str] = None,
+    claimed_cursor: int,
+    old_cursor: int,
+) -> bool:
+    """Undo a notification claim when delivery fails.
+
+    The CAS guard only rewinds if no later notifier advanced the row after our
+    claim. This keeps retry behavior for transient send failures without
+    clobbering newer progress.
+    """
+    with write_txn(conn):
+        cur = conn.execute(
+            "UPDATE kanban_notify_subs SET last_event_id = ? "
+            "WHERE task_id = ? AND platform = ? AND chat_id = ? AND thread_id = ? "
+            "AND last_event_id = ?",
+            (
+                int(old_cursor), task_id, platform, chat_id, thread_id or "",
+                int(claimed_cursor),
+            ),
+        )
+    return cur.rowcount > 0
+
+
 # ---------------------------------------------------------------------------
 # Retention + garbage collection
 # ---------------------------------------------------------------------------

hermes_cli/kanban_diagnostics.py

@@ -177,7 +177,7 @@ def _active_hallucination_events(
     active: list[Any] = []
     for ev in events:
         k = _event_kind(ev)
-        if k in ("completed", "edited"):
+        if k in {"completed", "edited"}:
             active.clear()
         elif k == kind:
             active.append(ev)
@@ -193,10 +193,9 @@ def _latest_clean_event_ts(events: Iterable[Any]) -> int:
     """
     latest = 0
     for ev in events:
-        if _event_kind(ev) in ("completed", "edited"):
+        if _event_kind(ev) in {"completed", "edited"}:
             t = _event_ts(ev)
-            if t > latest:
-                latest = t
+            latest = max(latest, t)
     return latest
 
 
@@ -356,7 +355,7 @@ def _rule_repeated_failures(task, events, runs, now, cfg) -> list[Diagnostic]:
     most_recent_outcome = None
     for r in reversed(ordered_runs):
         oc = _task_field(r, "outcome")
-        if oc in ("spawn_failed", "timed_out", "crashed"):
+        if oc in {"spawn_failed", "timed_out", "crashed"}:
             most_recent_outcome = oc
             break
 
@@ -374,7 +373,7 @@ def _rule_repeated_failures(task, events, runs, now, cfg) -> list[Diagnostic]:
             label=f"Fix profile auth: hermes -p {assignee} auth",
             payload={"command": f"hermes -p {assignee} auth"},
         ))
-    elif most_recent_outcome in ("timed_out", "crashed"):
+    elif most_recent_outcome in {"timed_out", "crashed"}:
         # Worker got off the ground but died. Logs are the right place
         # to diagnose; reclaim/reassign are the recovery levers.
         task_id = _task_field(task, "id")
@@ -467,7 +466,7 @@ def _rule_repeated_crashes(task, events, runs, now, cfg) -> list[Diagnostic]:
             consecutive += 1
             if last_err is None:
                 last_err = _task_field(r, "error")
-        elif outcome in ("completed", "reclaimed"):
+        elif outcome in {"completed", "reclaimed"}:
             # A success (or manual reclaim) breaks the streak.
             break
         else:
@@ -534,8 +533,7 @@ def _rule_stuck_in_blocked(task, events, runs, now, cfg) -> list[Diagnostic]:
     for ev in events:
         if _event_kind(ev) == "blocked":
             t = _event_ts(ev)
-            if t > last_blocked_ts:
-                last_blocked_ts = t
+            last_blocked_ts = max(last_blocked_ts, t)
     if last_blocked_ts == 0:
         return []
     age_hours = (now - last_blocked_ts) / 3600.0
@@ -543,7 +541,7 @@ def _rule_stuck_in_blocked(task, events, runs, now, cfg) -> list[Diagnostic]:
         return []
     # Any comment / unblock after the block breaks the "stale" signal.
     for ev in events:
-        if _event_kind(ev) in ("commented", "unblocked") and _event_ts(ev) > last_blocked_ts:
+        if _event_kind(ev) in {"commented", "unblocked"} and _event_ts(ev) > last_blocked_ts:
             return []
     actions: list[DiagnosticAction] = [
         DiagnosticAction(
@@ -570,6 +568,129 @@ def _rule_stuck_in_blocked(task, events, runs, now, cfg) -> list[Diagnostic]:
     )]
 
 
+def _rule_stranded_in_ready(task, events, runs, now, cfg) -> list[Diagnostic]:
+    """Task has been in ``ready`` status for too long without any worker
+    claiming it.
+
+    Threshold: cfg["stranded_threshold_seconds"] (default 1800 = 30 min).
+
+    Catches every "task waiting for a worker that never comes" case
+    without caring WHY:
+
+    * Operator typo'd the assignee — no profile or external worker matches.
+    * Profile was deleted, leaving its tasks stranded.
+    * External worker pool (Codex CLI, Claude Code lane, custom daemon)
+      is down, hung, or wasn't started.
+    * Dispatcher is misconfigured (wrong board, wrong HERMES_HOME).
+
+    Pre-rule, all of these silently rotted in ``skipped_nonspawnable`` —
+    the dispatcher correctly skipped them (good — no respawn loop) but
+    nobody surfaced the fact that operator-actionable work was
+    accumulating. The rule fires when a ready task's promoted-to-ready
+    timestamp is older than the threshold AND the assignee is non-empty
+    (truly unassigned tasks have their own ``skipped_unassigned`` signal
+    on the dispatcher and a different operator response).
+
+    The signal is age-based on purpose: it's identity-agnostic, so it
+    works for Hermes profiles, registered lanes, external workers, and
+    typos uniformly. No registry to curate, no per-board allowlist.
+    """
+    threshold_seconds = float(
+        cfg.get("stranded_threshold_seconds", 30 * 60)
+    )
+    status = _task_field(task, "status")
+    if status != "ready":
+        return []
+    # Skip tasks with a live claim — they're being worked on, even if
+    # the worker hasn't reported progress yet (run-level liveness
+    # extends the claim TTL; we don't want to second-guess that here).
+    if _task_field(task, "claim_lock"):
+        return []
+    assignee = _task_field(task, "assignee") or ""
+    if not assignee.strip():
+        # Unassigned tasks: the dispatcher's ``skipped_unassigned`` is
+        # already the right signal. A separate diagnostic here would
+        # double-flag the same condition.
+        return []
+
+    # Find the most recent event that put this task into ready.
+    # ``created`` covers tasks born ready; ``promoted`` covers parent-
+    # done auto-promotion; ``reclaimed`` covers TTL/crash recovery;
+    # ``unblocked`` covers human-driven resumes.
+    READY_TRANSITION_KINDS = {
+        "created", "promoted", "reclaimed", "unblocked",
+    }
+    last_ready_ts = 0
+    for ev in events:
+        if _event_kind(ev) in READY_TRANSITION_KINDS:
+            t = _event_ts(ev)
+            last_ready_ts = max(last_ready_ts, t)
+
+    # Fallback: if no qualifying event exists (very old task or events
+    # truncated), fall back to ``created_at`` on the task row. Better
+    # to occasionally over-flag an ancient task than miss a stranded one.
+    if last_ready_ts == 0:
+        last_ready_ts = int(_task_field(task, "created_at", default=0) or 0)
+    if last_ready_ts == 0:
+        return []
+
+    age_seconds = now - last_ready_ts
+    if age_seconds < threshold_seconds:
+        return []
+
+    # Format the age in the largest sensible unit.
+    if age_seconds >= 3600:
+        age_str = f"{age_seconds / 3600:.1f}h"
+    else:
+        age_str = f"{int(age_seconds / 60)}m"
+
+    # Severity escalates with age. Below 2x threshold = warning;
+    # 2x – 6x = error; beyond 6x = critical (something is clearly
+    # broken, not just slow).
+    if age_seconds >= threshold_seconds * 6:
+        severity = "critical"
+    elif age_seconds >= threshold_seconds * 2:
+        severity = "error"
+    else:
+        severity = "warning"
+
+    actions = [
+        DiagnosticAction(
+            kind="reassign",
+            label="Reassign to a different worker",
+            payload={"current_assignee": assignee},
+        ),
+        DiagnosticAction(
+            kind="cli_hint",
+            label="Check dispatcher status",
+            payload={"command": "hermes kanban diagnostics"},
+        ),
+    ]
+
+    return [Diagnostic(
+        kind="stranded_in_ready",
+        severity=severity,
+        title=f"Ready for {age_str} with no worker",
+        detail=(
+            f"This task has been ready for {age_str} but nothing has "
+            f"claimed it. Common causes: assignee {assignee!r} is "
+            f"misspelled, the profile was deleted, or the external "
+            f"worker pool for this lane is down. Confirm the assignee "
+            f"is correct and that a worker is actually polling for it."
+        ),
+        actions=actions,
+        first_seen_at=last_ready_ts,
+        last_seen_at=last_ready_ts,
+        count=1,
+        data={
+            "ready_since": last_ready_ts,
+            "age_seconds": int(age_seconds),
+            "assignee": assignee,
+            "threshold_seconds": int(threshold_seconds),
+        },
+    )]
+
+
 # Registry — order matters: rules higher on the list render first when
 # severity ties. Add new rules here.
 _RULES: list[RuleFn] = [
@@ -578,6 +699,7 @@ _RULES: list[RuleFn] = [
     _rule_repeated_failures,
     _rule_repeated_crashes,
     _rule_stuck_in_blocked,
+    _rule_stranded_in_ready,
 ]
 
 
@@ -589,6 +711,7 @@ DIAGNOSTIC_KINDS = (
     "repeated_failures",
     "repeated_crashes",
     "stuck_in_blocked",
+    "stranded_in_ready",
 )
 
 
@@ -598,6 +721,10 @@ DEFAULT_CONFIG = {
     "spawn_failure_threshold": 3,
     "crash_threshold": 2,
     "blocked_stale_hours": 24,
+    # Stranded-task threshold. 30 min by default — below that, the
+    # signal is dominated by tasks that are about to be claimed on the
+    # next dispatcher tick (default 60s) and would just be noise.
+    "stranded_threshold_seconds": 30 * 60,
 }