fix: pass terminal_lifetime through to Modal sandbox timeout

The lifetime_seconds config was computed and stored as an env var but never reached Modal's Sandbox.create(timeout=...), causing sandboxes to always die after the hardcoded 1h default regardless of configuration.
fix: use hermes agent system prompt and nudges
2026-04-13 11:57:39 -07:00 · 2026-04-08 18:34:01 -07:00 · 2026-04-08 18:34:01 -07:00 · 2026-04-08 18:33:02 -07:00 · 2026-04-08 18:33:02 -07:00 · 2026-04-08 18:33:02 -07:00
259 changed files with 11758 additions and 2979 deletions
@@ -81,6 +81,14 @@
 # HF_TOKEN=
 # OPENCODE_GO_BASE_URL=https://opencode.ai/zen/go/v1  # Override default base URL

+# =============================================================================
+# LLM PROVIDER (Qwen OAuth)
+# =============================================================================
+# Qwen OAuth reuses your local Qwen CLI login (qwen auth qwen-oauth).
+# No API key needed — credentials come from ~/.qwen/oauth_creds.json.
+# Optional base URL override:
+# HERMES_QWEN_BASE_URL=https://portal.qwen.ai/v1
+
 # =============================================================================
 # TOOL API KEYS
 # =============================================================================
@@ -19,6 +19,9 @@ jobs:
      - name: Checkout code
        uses: actions/checkout@v4

+      - name: Install system dependencies
+        run: sudo apt-get update && sudo apt-get install -y ripgrep
+
      - name: Install uv
        uses: astral-sh/setup-uv@v5

@@ -0,0 +1,346 @@
+# Hermes Agent v0.8.0 (v2026.4.8)
+
+**Release Date:** April 8, 2026
+
+> The intelligence release — background task auto-notifications, free MiMo v2 Pro on Nous Portal, live model switching across all platforms, self-optimized GPT/Codex guidance, native Google AI Studio, smart inactivity timeouts, approval buttons, MCP OAuth 2.1, and 209 merged PRs with 82 resolved issues.
+
+---
+
+## ✨ Highlights
+
+- **Background Process Auto-Notifications (`notify_on_complete`)** — Background tasks can now automatically notify the agent when they finish. Start a long-running process (AI model training, test suites, deployments, builds) and the agent gets notified on completion — no polling needed. The agent can keep working on other things and pick up results when they land. ([#5779](https://github.com/NousResearch/hermes-agent/pull/5779))
+
+- **Free Xiaomi MiMo v2 Pro on Nous Portal** — Nous Portal now supports the free-tier Xiaomi MiMo v2 Pro model for auxiliary tasks (compression, vision, summarization), with free-tier model gating and pricing display in model selection. ([#6018](https://github.com/NousResearch/hermes-agent/pull/6018), [#5880](https://github.com/NousResearch/hermes-agent/pull/5880))
+
+- **Live Model Switching (`/model` Command)** — Switch models and providers mid-session from CLI, Telegram, Discord, Slack, or any gateway platform. Aggregator-aware resolution keeps you on OpenRouter/Nous when possible, with automatic cross-provider fallback when needed. Interactive model pickers on Telegram and Discord with inline buttons. ([#5181](https://github.com/NousResearch/hermes-agent/pull/5181), [#5742](https://github.com/NousResearch/hermes-agent/pull/5742))
+
+- **Self-Optimized GPT/Codex Tool-Use Guidance** — The agent diagnosed and patched 5 failure modes in GPT and Codex tool calling through automated behavioral benchmarking, dramatically improving reliability on OpenAI models. Includes execution discipline guidance and thinking-only prefill continuation for structured reasoning. ([#6120](https://github.com/NousResearch/hermes-agent/pull/6120), [#5414](https://github.com/NousResearch/hermes-agent/pull/5414), [#5931](https://github.com/NousResearch/hermes-agent/pull/5931))
+
+- **Google AI Studio (Gemini) Native Provider** — Direct access to Gemini models through Google's AI Studio API. Includes automatic models.dev registry integration for real-time context length detection across any provider. ([#5577](https://github.com/NousResearch/hermes-agent/pull/5577))
+
+- **Inactivity-Based Agent Timeouts** — Gateway and cron timeouts now track actual tool activity instead of wall-clock time. Long-running tasks that are actively working will never be killed — only truly idle agents time out. ([#5389](https://github.com/NousResearch/hermes-agent/pull/5389), [#5440](https://github.com/NousResearch/hermes-agent/pull/5440))
+
+- **Approval Buttons on Slack & Telegram** — Dangerous command approval via native platform buttons instead of typing `/approve`. Slack gets thread context preservation; Telegram gets emoji reactions for approval status. ([#5890](https://github.com/NousResearch/hermes-agent/pull/5890), [#5975](https://github.com/NousResearch/hermes-agent/pull/5975))
+
+- **MCP OAuth 2.1 PKCE + OSV Malware Scanning** — Full standards-compliant OAuth for MCP server authentication, plus automatic malware scanning of MCP extension packages via the OSV vulnerability database. ([#5420](https://github.com/NousResearch/hermes-agent/pull/5420), [#5305](https://github.com/NousResearch/hermes-agent/pull/5305))
+
+- **Centralized Logging & Config Validation** — Structured logging to `~/.hermes/logs/` (agent.log + errors.log) with the `hermes logs` command for tailing and filtering. Config structure validation catches malformed YAML at startup before it causes cryptic failures. ([#5430](https://github.com/NousResearch/hermes-agent/pull/5430), [#5426](https://github.com/NousResearch/hermes-agent/pull/5426))
+
+- **Plugin System Expansion** — Plugins can now register CLI subcommands, receive request-scoped API hooks with correlation IDs, prompt for required env vars during install, and hook into session lifecycle events (finalize/reset). ([#5295](https://github.com/NousResearch/hermes-agent/pull/5295), [#5427](https://github.com/NousResearch/hermes-agent/pull/5427), [#5470](https://github.com/NousResearch/hermes-agent/pull/5470), [#6129](https://github.com/NousResearch/hermes-agent/pull/6129))
+
+- **Matrix Tier 1 & Platform Hardening** — Matrix gets reactions, read receipts, rich formatting, and room management. Discord adds channel controls and ignored channels. Signal gets full MEDIA: tag delivery. Mattermost gets file attachments. Comprehensive reliability fixes across all platforms. ([#5275](https://github.com/NousResearch/hermes-agent/pull/5275), [#5975](https://github.com/NousResearch/hermes-agent/pull/5975), [#5602](https://github.com/NousResearch/hermes-agent/pull/5602))
+
+- **Security Hardening Pass** — Consolidated SSRF protections, timing attack mitigations, tar traversal prevention, credential leakage guards, cron path traversal hardening, and cross-session isolation. Terminal workdir sanitization across all backends. ([#5944](https://github.com/NousResearch/hermes-agent/pull/5944), [#5613](https://github.com/NousResearch/hermes-agent/pull/5613), [#5629](https://github.com/NousResearch/hermes-agent/pull/5629))
+
+---
+
+## 🏗️ Core Agent & Architecture
+
+### Provider & Model Support
+- **Native Google AI Studio (Gemini) provider** with models.dev integration for automatic context length detection ([#5577](https://github.com/NousResearch/hermes-agent/pull/5577))
+- **`/model` command — full provider+model system overhaul** — live switching across CLI and all gateway platforms with aggregator-aware resolution ([#5181](https://github.com/NousResearch/hermes-agent/pull/5181))
+- **Interactive model picker for Telegram and Discord** — inline button-based model selection ([#5742](https://github.com/NousResearch/hermes-agent/pull/5742))
+- **Nous Portal free-tier model gating** with pricing display in model selection ([#5880](https://github.com/NousResearch/hermes-agent/pull/5880))
+- **Model pricing display** for OpenRouter and Nous Portal providers ([#5416](https://github.com/NousResearch/hermes-agent/pull/5416))
+- **xAI (Grok) prompt caching** via `x-grok-conv-id` header ([#5604](https://github.com/NousResearch/hermes-agent/pull/5604))
+- **Grok added to tool-use enforcement models** for direct xAI usage ([#5595](https://github.com/NousResearch/hermes-agent/pull/5595))
+- **MiniMax TTS provider** (speech-2.8) ([#4963](https://github.com/NousResearch/hermes-agent/pull/4963))
+- **Non-agentic model warning** — warns users when loading Hermes LLM models not designed for tool use ([#5378](https://github.com/NousResearch/hermes-agent/pull/5378))
+- **Ollama Cloud auth, /model switch persistence**, and alias tab completion ([#5269](https://github.com/NousResearch/hermes-agent/pull/5269))
+- **Preserve dots in OpenCode Go model names** (minimax-m2.7, glm-4.5, kimi-k2.5) ([#5597](https://github.com/NousResearch/hermes-agent/pull/5597))
+- **MiniMax models 404 fix** — strip /v1 from Anthropic base URL for OpenCode Go ([#4918](https://github.com/NousResearch/hermes-agent/pull/4918))
+- **Provider credential reset windows** honored in pooled failover ([#5188](https://github.com/NousResearch/hermes-agent/pull/5188))
+- **OAuth token sync** between credential pool and credentials file ([#4981](https://github.com/NousResearch/hermes-agent/pull/4981))
+- **Stale OAuth credentials** no longer block OpenRouter users on auto-detect ([#5746](https://github.com/NousResearch/hermes-agent/pull/5746))
+- **Codex OAuth credential pool disconnect** + expired token import fix ([#5681](https://github.com/NousResearch/hermes-agent/pull/5681))
+- **Codex pool entry sync** from `~/.codex/auth.json` on exhaustion — @GratefulDave ([#5610](https://github.com/NousResearch/hermes-agent/pull/5610))
+- **Auxiliary client payment fallback** — retry with next provider on 402 ([#5599](https://github.com/NousResearch/hermes-agent/pull/5599))
+- **Auxiliary client resolves named custom providers** and 'main' alias ([#5978](https://github.com/NousResearch/hermes-agent/pull/5978))
+- **Use mimo-v2-pro** for non-vision auxiliary tasks on Nous free tier ([#6018](https://github.com/NousResearch/hermes-agent/pull/6018))
+- **Vision auto-detection** tries main provider first ([#6041](https://github.com/NousResearch/hermes-agent/pull/6041))
+- **Provider re-ordering and Quick Install** — @austinpickett ([#4664](https://github.com/NousResearch/hermes-agent/pull/4664))
+- **Nous OAuth access_token** no longer used as inference API key — @SHL0MS ([#5564](https://github.com/NousResearch/hermes-agent/pull/5564))
+- **HERMES_PORTAL_BASE_URL env var** respected during Nous login — @benbarclay ([#5745](https://github.com/NousResearch/hermes-agent/pull/5745))
+- **Env var overrides** for Nous portal/inference URLs ([#5419](https://github.com/NousResearch/hermes-agent/pull/5419))
+- **Z.AI endpoint auto-detect** via probe and cache ([#5763](https://github.com/NousResearch/hermes-agent/pull/5763))
+- **MiniMax context lengths, model catalog, thinking guard, aux model, and config base_url** corrections ([#6082](https://github.com/NousResearch/hermes-agent/pull/6082))
+- **Community provider/model resolution fixes** — salvaged 4 community PRs + MiniMax aux URL ([#5983](https://github.com/NousResearch/hermes-agent/pull/5983))
+
+### Agent Loop & Conversation
+- **Self-optimized GPT/Codex tool-use guidance** via automated behavioral benchmarking — agent self-diagnosed and patched 5 failure modes ([#6120](https://github.com/NousResearch/hermes-agent/pull/6120))
+- **GPT/Codex execution discipline guidance** in system prompts ([#5414](https://github.com/NousResearch/hermes-agent/pull/5414))
+- **Thinking-only prefill continuation** for structured reasoning responses ([#5931](https://github.com/NousResearch/hermes-agent/pull/5931))
+- **Accept reasoning-only responses** without retries — set content to "(empty)" instead of infinite retry ([#5278](https://github.com/NousResearch/hermes-agent/pull/5278))
+- **Jittered retry backoff** — exponential backoff with jitter for API retries ([#6048](https://github.com/NousResearch/hermes-agent/pull/6048))
+- **Smart thinking block signature management** — preserve and manage Anthropic thinking signatures across turns ([#6112](https://github.com/NousResearch/hermes-agent/pull/6112))
+- **Coerce tool call arguments** to match JSON Schema types — fixes models that send strings instead of numbers/booleans ([#5265](https://github.com/NousResearch/hermes-agent/pull/5265))
+- **Save oversized tool results to file** instead of destructive truncation ([#5210](https://github.com/NousResearch/hermes-agent/pull/5210))
+- **Sandbox-aware tool result persistence** ([#6085](https://github.com/NousResearch/hermes-agent/pull/6085))
+- **Streaming fallback** improved after edit failures ([#6110](https://github.com/NousResearch/hermes-agent/pull/6110))
+- **Codex empty-output gaps** covered in fallback + normalizer + auxiliary client ([#5724](https://github.com/NousResearch/hermes-agent/pull/5724), [#5730](https://github.com/NousResearch/hermes-agent/pull/5730), [#5734](https://github.com/NousResearch/hermes-agent/pull/5734))
+- **Codex stream output backfill** from output_item.done events ([#5689](https://github.com/NousResearch/hermes-agent/pull/5689))
+- **Stream consumer creates new message** after tool boundaries ([#5739](https://github.com/NousResearch/hermes-agent/pull/5739))
+- **Codex validation aligned** with normalization for empty stream output ([#5940](https://github.com/NousResearch/hermes-agent/pull/5940))
+- **Bridge tool-calls** in copilot-acp adapter ([#5460](https://github.com/NousResearch/hermes-agent/pull/5460))
+- **Filter transcript-only roles** from chat-completions payload ([#4880](https://github.com/NousResearch/hermes-agent/pull/4880))
+- **Context compaction failures fixed** on temperature-restricted models — @MadKangYu ([#5608](https://github.com/NousResearch/hermes-agent/pull/5608))
+- **Sanitize tool_calls for all strict APIs** (Fireworks, Mistral, etc.) — @lumethegreat ([#5183](https://github.com/NousResearch/hermes-agent/pull/5183))
+
+### Memory & Sessions
+- **Supermemory memory provider** — new memory plugin with multi-container, search_mode, identity template, and env var override ([#5737](https://github.com/NousResearch/hermes-agent/pull/5737), [#5933](https://github.com/NousResearch/hermes-agent/pull/5933))
+- **Shared thread sessions** by default — multi-user thread support across gateway platforms ([#5391](https://github.com/NousResearch/hermes-agent/pull/5391))
+- **Subagent sessions linked to parent** and hidden from session list ([#5309](https://github.com/NousResearch/hermes-agent/pull/5309))
+- **Profile-scoped memory isolation** and clone support ([#4845](https://github.com/NousResearch/hermes-agent/pull/4845))
+- **Thread gateway user_id to memory plugins** for per-user scoping ([#5895](https://github.com/NousResearch/hermes-agent/pull/5895))
+- **Honcho plugin drift overhaul** + plugin CLI registration system ([#5295](https://github.com/NousResearch/hermes-agent/pull/5295))
+- **Honcho holographic prompt and trust score** rendering preserved ([#4872](https://github.com/NousResearch/hermes-agent/pull/4872))
+- **Honcho doctor fix** — use recall_mode instead of memory_mode — @techguysimon ([#5645](https://github.com/NousResearch/hermes-agent/pull/5645))
+- **RetainDB** — API routes, write queue, dialectic, agent model, file tools fixes ([#5461](https://github.com/NousResearch/hermes-agent/pull/5461))
+- **Hindsight memory plugin overhaul** + memory setup wizard fixes ([#5094](https://github.com/NousResearch/hermes-agent/pull/5094))
+- **mem0 API v2 compat**, prefetch context fencing, secret redaction ([#5423](https://github.com/NousResearch/hermes-agent/pull/5423))
+- **mem0 env vars merged** with mem0.json instead of either/or ([#4939](https://github.com/NousResearch/hermes-agent/pull/4939))
+- **Clean user message** used for all memory provider operations ([#4940](https://github.com/NousResearch/hermes-agent/pull/4940))
+- **Silent memory flush failure** on /new and /resume fixed — @ryanautomated ([#5640](https://github.com/NousResearch/hermes-agent/pull/5640))
+- **OpenViking atexit safety net** for session commit ([#5664](https://github.com/NousResearch/hermes-agent/pull/5664))
+- **OpenViking tenant-scoping headers** for multi-tenant servers ([#4936](https://github.com/NousResearch/hermes-agent/pull/4936))
+- **ByteRover brv query** runs synchronously before LLM call ([#4831](https://github.com/NousResearch/hermes-agent/pull/4831))
+
+---
+
+## 📱 Messaging Platforms (Gateway)
+
+### Gateway Core
+- **Inactivity-based agent timeout** — replaces wall-clock timeout with smart activity tracking; long-running active tasks never killed ([#5389](https://github.com/NousResearch/hermes-agent/pull/5389))
+- **Approval buttons for Slack & Telegram** + Slack thread context preservation ([#5890](https://github.com/NousResearch/hermes-agent/pull/5890))
+- **Live-stream /update output** + forward interactive prompts to user ([#5180](https://github.com/NousResearch/hermes-agent/pull/5180))
+- **Infinite timeout support** + periodic notifications + actionable error messages ([#4959](https://github.com/NousResearch/hermes-agent/pull/4959))
+- **Duplicate message prevention** — gateway dedup + partial stream guard ([#4878](https://github.com/NousResearch/hermes-agent/pull/4878))
+- **Webhook delivery_info persistence** + full session id in /status ([#5942](https://github.com/NousResearch/hermes-agent/pull/5942))
+- **Tool preview truncation** respects tool_preview_length in all/new progress modes ([#5937](https://github.com/NousResearch/hermes-agent/pull/5937))
+- **Short preview truncation** restored for all/new tool progress modes ([#4935](https://github.com/NousResearch/hermes-agent/pull/4935))
+- **Update-pending state** written atomically to prevent corruption ([#4923](https://github.com/NousResearch/hermes-agent/pull/4923))
+- **Approval session key isolated** per turn ([#4884](https://github.com/NousResearch/hermes-agent/pull/4884))
+- **Active-session guard bypass** for /approve, /deny, /stop, /new ([#4926](https://github.com/NousResearch/hermes-agent/pull/4926), [#5765](https://github.com/NousResearch/hermes-agent/pull/5765))
+- **Typing indicator paused** during approval waits ([#5893](https://github.com/NousResearch/hermes-agent/pull/5893))
+- **Caption check** uses exact line-by-line match instead of substring (all platforms) ([#5939](https://github.com/NousResearch/hermes-agent/pull/5939))
+- **MEDIA: tags stripped** from streamed gateway messages ([#5152](https://github.com/NousResearch/hermes-agent/pull/5152))
+- **MEDIA: tags extracted** from cron delivery before sending ([#5598](https://github.com/NousResearch/hermes-agent/pull/5598))
+- **Profile-aware service units** + voice transcription cleanup ([#5972](https://github.com/NousResearch/hermes-agent/pull/5972))
+- **Thread-safe PairingStore** with atomic writes — @CharlieKerfoot ([#5656](https://github.com/NousResearch/hermes-agent/pull/5656))
+- **Sanitize media URLs** in base platform logs — @WAXLYY ([#5631](https://github.com/NousResearch/hermes-agent/pull/5631))
+- **Reduce Telegram fallback IP activation log noise** — @MadKangYu ([#5615](https://github.com/NousResearch/hermes-agent/pull/5615))
+- **Cron static method wrappers** to prevent self-binding ([#5299](https://github.com/NousResearch/hermes-agent/pull/5299))
+- **Stale 'hermes login' replaced** with 'hermes auth' + credential removal re-seeding fix ([#5670](https://github.com/NousResearch/hermes-agent/pull/5670))
+
+### Telegram
+- **Group topics skill binding** for supergroup forum topics ([#4886](https://github.com/NousResearch/hermes-agent/pull/4886))
+- **Emoji reactions** for approval status and notifications ([#5975](https://github.com/NousResearch/hermes-agent/pull/5975))
+- **Duplicate message delivery prevented** on send timeout ([#5153](https://github.com/NousResearch/hermes-agent/pull/5153))
+- **Command names sanitized** to strip invalid characters ([#5596](https://github.com/NousResearch/hermes-agent/pull/5596))
+- **Per-platform disabled skills** respected in Telegram menu and gateway dispatch ([#4799](https://github.com/NousResearch/hermes-agent/pull/4799))
+- **/approve and /deny** routed through running-agent guard ([#4798](https://github.com/NousResearch/hermes-agent/pull/4798))
+
+### Discord
+- **Channel controls** — ignored_channels and no_thread_channels config options ([#5975](https://github.com/NousResearch/hermes-agent/pull/5975))
+- **Skills registered as native slash commands** via shared gateway logic ([#5603](https://github.com/NousResearch/hermes-agent/pull/5603))
+- **/approve, /deny, /queue, /background, /btw** registered as native slash commands ([#4800](https://github.com/NousResearch/hermes-agent/pull/4800), [#5477](https://github.com/NousResearch/hermes-agent/pull/5477))
+- **Unnecessary members intent** removed on startup + token lock leak fix ([#5302](https://github.com/NousResearch/hermes-agent/pull/5302))
+
+### Slack
+- **Thread engagement** — auto-respond in bot-started and mentioned threads ([#5897](https://github.com/NousResearch/hermes-agent/pull/5897))
+- **mrkdwn in edit_message** + thread replies without @mentions ([#5733](https://github.com/NousResearch/hermes-agent/pull/5733))
+
+### Matrix
+- **Tier 1 feature parity** — reactions, read receipts, rich formatting, room management ([#5275](https://github.com/NousResearch/hermes-agent/pull/5275))
+- **MATRIX_REQUIRE_MENTION and MATRIX_AUTO_THREAD** support ([#5106](https://github.com/NousResearch/hermes-agent/pull/5106))
+- **Comprehensive reliability** — encrypted media, auth recovery, cron E2EE, Synapse compat ([#5271](https://github.com/NousResearch/hermes-agent/pull/5271))
+- **CJK input, E2EE, and reconnect** fixes ([#5665](https://github.com/NousResearch/hermes-agent/pull/5665))
+
+### Signal
+- **Full MEDIA: tag delivery** — send_image_file, send_voice, and send_video implemented ([#5602](https://github.com/NousResearch/hermes-agent/pull/5602))
+
+### Mattermost
+- **File attachments** — set message type to DOCUMENT when post has file attachments — @nericervin ([#5609](https://github.com/NousResearch/hermes-agent/pull/5609))
+
+### Feishu
+- **Interactive card approval buttons** ([#6043](https://github.com/NousResearch/hermes-agent/pull/6043))
+- **Reconnect and ACL** fixes ([#5665](https://github.com/NousResearch/hermes-agent/pull/5665))
+
+### Webhooks
+- **`{__raw__}` template token** and thread_id passthrough for forum topics ([#5662](https://github.com/NousResearch/hermes-agent/pull/5662))
+
+---
+
+## 🖥️ CLI & User Experience
+
+### Interactive CLI
+- **Defer response content** until reasoning block completes ([#5773](https://github.com/NousResearch/hermes-agent/pull/5773))
+- **Ghost status-bar lines cleared** on terminal resize ([#4960](https://github.com/NousResearch/hermes-agent/pull/4960))
+- **Normalise \r\n and \r line endings** in pasted text ([#4849](https://github.com/NousResearch/hermes-agent/pull/4849))
+- **ChatConsole errors, curses scroll, skin-aware banner, git state** banner fixes ([#5974](https://github.com/NousResearch/hermes-agent/pull/5974))
+- **Native Windows image paste** support ([#5917](https://github.com/NousResearch/hermes-agent/pull/5917))
+- **--yolo and other flags** no longer silently dropped when placed before 'chat' subcommand ([#5145](https://github.com/NousResearch/hermes-agent/pull/5145))
+
+### Setup & Configuration
+- **Config structure validation** — detect malformed YAML at startup with actionable error messages ([#5426](https://github.com/NousResearch/hermes-agent/pull/5426))
+- **Centralized logging** to `~/.hermes/logs/` — agent.log (INFO+), errors.log (WARNING+) with `hermes logs` command ([#5430](https://github.com/NousResearch/hermes-agent/pull/5430))
+- **Docs links added** to setup wizard sections ([#5283](https://github.com/NousResearch/hermes-agent/pull/5283))
+- **Doctor diagnostics** — sync provider checks, config migration, WAL and mem0 diagnostics ([#5077](https://github.com/NousResearch/hermes-agent/pull/5077))
+- **Timeout debug logging** and user-facing diagnostics improved ([#5370](https://github.com/NousResearch/hermes-agent/pull/5370))
+- **Reasoning effort unified** to config.yaml only ([#6118](https://github.com/NousResearch/hermes-agent/pull/6118))
+- **Permanent command allowlist** loaded on startup ([#5076](https://github.com/NousResearch/hermes-agent/pull/5076))
+- **`hermes auth remove`** now clears env-seeded credentials permanently ([#5285](https://github.com/NousResearch/hermes-agent/pull/5285))
+- **Bundled skills synced to all profiles** during update ([#5795](https://github.com/NousResearch/hermes-agent/pull/5795))
+- **`hermes update` no longer kills** freshly-restarted gateway service ([#5448](https://github.com/NousResearch/hermes-agent/pull/5448))
+- **Subprocess.run() timeouts** added to all gateway CLI commands ([#5424](https://github.com/NousResearch/hermes-agent/pull/5424))
+- **Actionable error message** when Codex refresh token is reused — @tymrtn ([#5612](https://github.com/NousResearch/hermes-agent/pull/5612))
+- **Google-workspace skill scripts** can now run directly — @xinbenlv ([#5624](https://github.com/NousResearch/hermes-agent/pull/5624))
+
+### Cron System
+- **Inactivity-based cron timeout** — replaces wall-clock; active tasks run indefinitely ([#5440](https://github.com/NousResearch/hermes-agent/pull/5440))
+- **Pre-run script injection** for data collection and change detection ([#5082](https://github.com/NousResearch/hermes-agent/pull/5082))
+- **Delivery failure tracking** in job status ([#6042](https://github.com/NousResearch/hermes-agent/pull/6042))
+- **Delivery guidance** in cron prompts — stops send_message thrashing ([#5444](https://github.com/NousResearch/hermes-agent/pull/5444))
+- **MEDIA files delivered** as native platform attachments ([#5921](https://github.com/NousResearch/hermes-agent/pull/5921))
+- **[SILENT] suppression** works anywhere in response — @auspic7 ([#5654](https://github.com/NousResearch/hermes-agent/pull/5654))
+- **Cron path traversal** hardening ([#5147](https://github.com/NousResearch/hermes-agent/pull/5147))
+
+---
+
+## 🔧 Tool System
+
+### Terminal & Execution
+- **Execute_code on remote backends** — code execution now works on Docker, SSH, Modal, and other remote terminal backends ([#5088](https://github.com/NousResearch/hermes-agent/pull/5088))
+- **Exit code context** for common CLI tools in terminal results — helps agent understand what went wrong ([#5144](https://github.com/NousResearch/hermes-agent/pull/5144))
+- **Progressive subdirectory hint discovery** — agent learns project structure as it navigates ([#5291](https://github.com/NousResearch/hermes-agent/pull/5291))
+- **notify_on_complete for background processes** — get notified when long-running tasks finish ([#5779](https://github.com/NousResearch/hermes-agent/pull/5779))
+- **Docker env config** — explicit container environment variables via docker_env config ([#4738](https://github.com/NousResearch/hermes-agent/pull/4738))
+- **Approval metadata included** in terminal tool results ([#5141](https://github.com/NousResearch/hermes-agent/pull/5141))
+- **Workdir parameter sanitized** in terminal tool across all backends ([#5629](https://github.com/NousResearch/hermes-agent/pull/5629))
+- **Detached process crash recovery** state corrected ([#6101](https://github.com/NousResearch/hermes-agent/pull/6101))
+- **Agent-browser paths with spaces** preserved — @Vasanthdev2004 ([#6077](https://github.com/NousResearch/hermes-agent/pull/6077))
+- **Portable base64 encoding** for image reading on macOS — @CharlieKerfoot ([#5657](https://github.com/NousResearch/hermes-agent/pull/5657))
+
+### Browser
+- **Switch managed browser provider** from Browserbase to Browser Use — @benbarclay ([#5750](https://github.com/NousResearch/hermes-agent/pull/5750))
+- **Firecrawl cloud browser** provider — @alt-glitch ([#5628](https://github.com/NousResearch/hermes-agent/pull/5628))
+- **JS evaluation** via browser_console expression parameter ([#5303](https://github.com/NousResearch/hermes-agent/pull/5303))
+- **Windows browser** fixes ([#5665](https://github.com/NousResearch/hermes-agent/pull/5665))
+
+### MCP
+- **MCP OAuth 2.1 PKCE** — full standards-compliant OAuth client support ([#5420](https://github.com/NousResearch/hermes-agent/pull/5420))
+- **OSV malware check** for MCP extension packages ([#5305](https://github.com/NousResearch/hermes-agent/pull/5305))
+- **Prefer structuredContent over text** + no_mcp sentinel ([#5979](https://github.com/NousResearch/hermes-agent/pull/5979))
+- **Unknown toolsets warning suppressed** for MCP server names ([#5279](https://github.com/NousResearch/hermes-agent/pull/5279))
+
+### Web & Files
+- **.zip document support** + auto-mount cache dirs into remote backends ([#4846](https://github.com/NousResearch/hermes-agent/pull/4846))
+- **Redact query secrets** in send_message errors — @WAXLYY ([#5650](https://github.com/NousResearch/hermes-agent/pull/5650))
+
+### Delegation
+- **Credential pool sharing** + workspace path hints for subagents ([#5748](https://github.com/NousResearch/hermes-agent/pull/5748))
+
+### ACP (VS Code / Zed / JetBrains)
+- **Aggregate ACP improvements** — auth compat, protocol fixes, command ads, delegation, SSE events ([#5292](https://github.com/NousResearch/hermes-agent/pull/5292))
+
+---
+
+## 🧩 Skills Ecosystem
+
+### Skills System
+- **Skill config interface** — skills can declare required config.yaml settings, prompted during setup, injected at load time ([#5635](https://github.com/NousResearch/hermes-agent/pull/5635))
+- **Plugin CLI registration system** — plugins register their own CLI subcommands without touching main.py ([#5295](https://github.com/NousResearch/hermes-agent/pull/5295))
+- **Request-scoped API hooks** with tool call correlation IDs for plugins ([#5427](https://github.com/NousResearch/hermes-agent/pull/5427))
+- **Session lifecycle hooks** — on_session_finalize and on_session_reset for CLI + gateway ([#6129](https://github.com/NousResearch/hermes-agent/pull/6129))
+- **Prompt for required env vars** during plugin install — @kshitijk4poor ([#5470](https://github.com/NousResearch/hermes-agent/pull/5470))
+- **Plugin name validation** — reject names that resolve to plugins root ([#5368](https://github.com/NousResearch/hermes-agent/pull/5368))
+- **pre_llm_call plugin context** moved to user message to preserve prompt cache ([#5146](https://github.com/NousResearch/hermes-agent/pull/5146))
+
+### New & Updated Skills
+- **popular-web-designs** — 54 production website design systems ([#5194](https://github.com/NousResearch/hermes-agent/pull/5194))
+- **p5js creative coding** — @SHL0MS ([#5600](https://github.com/NousResearch/hermes-agent/pull/5600))
+- **manim-video** — mathematical and technical animations — @SHL0MS ([#4930](https://github.com/NousResearch/hermes-agent/pull/4930))
+- **llm-wiki** — Karpathy's LLM Wiki skill ([#5635](https://github.com/NousResearch/hermes-agent/pull/5635))
+- **gitnexus-explorer** — codebase indexing and knowledge serving ([#5208](https://github.com/NousResearch/hermes-agent/pull/5208))
+- **research-paper-writing** — AI-Scientist & GPT-Researcher patterns — @SHL0MS ([#5421](https://github.com/NousResearch/hermes-agent/pull/5421))
+- **blogwatcher** updated to JulienTant's fork ([#5759](https://github.com/NousResearch/hermes-agent/pull/5759))
+- **claude-code skill** comprehensive rewrite v2.0 + v2.2 ([#5155](https://github.com/NousResearch/hermes-agent/pull/5155), [#5158](https://github.com/NousResearch/hermes-agent/pull/5158))
+- **Code verification skills** consolidated into one ([#4854](https://github.com/NousResearch/hermes-agent/pull/4854))
+- **Manim CE reference docs** expanded — geometry, animations, LaTeX — @leotrs ([#5791](https://github.com/NousResearch/hermes-agent/pull/5791))
+- **Manim-video references** — design thinking, updaters, paper explainer, decorations, production quality — @SHL0MS ([#5588](https://github.com/NousResearch/hermes-agent/pull/5588), [#5408](https://github.com/NousResearch/hermes-agent/pull/5408))
+
+---
+
+## 🔒 Security & Reliability
+
+### Security Hardening
+- **Consolidated security** — SSRF protections, timing attack mitigations, tar traversal prevention, credential leakage guards ([#5944](https://github.com/NousResearch/hermes-agent/pull/5944))
+- **Cross-session isolation** + cron path traversal hardening ([#5613](https://github.com/NousResearch/hermes-agent/pull/5613))
+- **Workdir parameter sanitized** in terminal tool across all backends ([#5629](https://github.com/NousResearch/hermes-agent/pull/5629))
+- **Approval 'once' session escalation** prevented + cron delivery platform validation ([#5280](https://github.com/NousResearch/hermes-agent/pull/5280))
+- **Profile-scoped Google Workspace OAuth tokens** protected ([#4910](https://github.com/NousResearch/hermes-agent/pull/4910))
+
+### Reliability
+- **Aggressive worktree and branch cleanup** to prevent accumulation ([#6134](https://github.com/NousResearch/hermes-agent/pull/6134))
+- **O(n²) catastrophic backtracking** in redact regex fixed — 100x improvement on large outputs ([#4962](https://github.com/NousResearch/hermes-agent/pull/4962))
+- **Runtime stability fixes** across core, web, delegate, and browser tools ([#4843](https://github.com/NousResearch/hermes-agent/pull/4843))
+- **API server streaming fix** + conversation history support ([#5977](https://github.com/NousResearch/hermes-agent/pull/5977))
+- **OpenViking API endpoint paths** and response parsing corrected ([#5078](https://github.com/NousResearch/hermes-agent/pull/5078))
+
+---
+
+## 🐛 Notable Bug Fixes
+
+- **9 community bugfixes salvaged** — gateway, cron, deps, macOS launchd in one batch ([#5288](https://github.com/NousResearch/hermes-agent/pull/5288))
+- **Batch core bug fixes** — model config, session reset, alias fallback, launchctl, delegation, atomic writes ([#5630](https://github.com/NousResearch/hermes-agent/pull/5630))
+- **Batch gateway/platform fixes** — matrix E2EE, CJK input, Windows browser, Feishu reconnect + ACL ([#5665](https://github.com/NousResearch/hermes-agent/pull/5665))
+- **Stale test skips removed**, regex backtracking, file search bug, and test flakiness ([#4969](https://github.com/NousResearch/hermes-agent/pull/4969))
+- **Nix flake** — read version, regen uv.lock, add hermes_logging — @alt-glitch ([#5651](https://github.com/NousResearch/hermes-agent/pull/5651))
+- **Lowercase variable redaction** regression tests ([#5185](https://github.com/NousResearch/hermes-agent/pull/5185))
+
+---
+
+## 🧪 Testing
+
+- **57 failing CI tests repaired** across 14 files ([#5823](https://github.com/NousResearch/hermes-agent/pull/5823))
+- **Test suite re-architecture** + CI failure fixes — @alt-glitch ([#5946](https://github.com/NousResearch/hermes-agent/pull/5946))
+- **Codebase-wide lint cleanup** — unused imports, dead code, and inefficient patterns ([#5821](https://github.com/NousResearch/hermes-agent/pull/5821))
+- **browser_close tool removed** — auto-cleanup handles it ([#5792](https://github.com/NousResearch/hermes-agent/pull/5792))
+
+---
+
+## 📚 Documentation
+
+- **Comprehensive documentation audit** — fix stale info, expand thin pages, add depth ([#5393](https://github.com/NousResearch/hermes-agent/pull/5393))
+- **40+ discrepancies fixed** between documentation and codebase ([#5818](https://github.com/NousResearch/hermes-agent/pull/5818))
+- **13 features documented** from last week's PRs ([#5815](https://github.com/NousResearch/hermes-agent/pull/5815))
+- **Guides section overhaul** — fix existing + add 3 new tutorials ([#5735](https://github.com/NousResearch/hermes-agent/pull/5735))
+- **Salvaged 4 docs PRs** — docker setup, post-update validation, local LLM guide, signal-cli install ([#5727](https://github.com/NousResearch/hermes-agent/pull/5727))
+- **Discord configuration reference** ([#5386](https://github.com/NousResearch/hermes-agent/pull/5386))
+- **Community FAQ entries** for common workflows and troubleshooting ([#4797](https://github.com/NousResearch/hermes-agent/pull/4797))
+- **WSL2 networking guide** for local model servers ([#5616](https://github.com/NousResearch/hermes-agent/pull/5616))
+- **Honcho CLI reference** + plugin CLI registration docs ([#5308](https://github.com/NousResearch/hermes-agent/pull/5308))
+- **Obsidian Headless setup** for servers in llm-wiki ([#5660](https://github.com/NousResearch/hermes-agent/pull/5660))
+- **Hermes Mod visual skin editor** added to skins page ([#6095](https://github.com/NousResearch/hermes-agent/pull/6095))
+
+---
+
+## 👥 Contributors
+
+### Core
+- **@teknium1** — 179 PRs
+
+### Top Community Contributors
+- **@SHL0MS** (7 PRs) — p5js creative coding skill, manim-video skill + 5 reference expansions, research-paper-writing, Nous OAuth fix, manim font fix
+- **@alt-glitch** (3 PRs) — Firecrawl cloud browser provider, test re-architecture + CI fixes, Nix flake fixes
+- **@benbarclay** (2 PRs) — Browser Use managed provider switch, Nous portal base URL fix
+- **@CharlieKerfoot** (2 PRs) — macOS portable base64 encoding, thread-safe PairingStore
+- **@WAXLYY** (2 PRs) — send_message secret redaction, gateway media URL sanitization
+- **@MadKangYu** (2 PRs) — Telegram log noise reduction, context compaction fix for temperature-restricted models
+
+### All Contributors
+@alt-glitch, @austinpickett, @auspic7, @benbarclay, @CharlieKerfoot, @GratefulDave, @kshitijk4poor, @leotrs, @lumethegreat, @MadKangYu, @nericervin, @ryanautomated, @SHL0MS, @techguysimon, @tymrtn, @Vasanthdev2004, @WAXLYY, @xinbenlv
+
+---
+
+**Full Changelog**: [v2026.4.3...v2026.4.8](https://github.com/NousResearch/hermes-agent/compare/v2026.4.3...v2026.4.8)
@@ -163,6 +163,17 @@ def _is_oauth_token(key: str) -> bool:
    return True


+def _normalize_base_url_text(base_url) -> str:
+    """Normalize SDK/base transport URL values to a plain string for inspection.
+
+    Some client objects expose ``base_url`` as an ``httpx.URL`` instead of a raw
+    string.  Provider/auth detection should accept either shape.
+    """
+    if not base_url:
+        return ""
+    return str(base_url).strip()
+
+
 def _is_third_party_anthropic_endpoint(base_url: str | None) -> bool:
    """Return True for non-Anthropic endpoints using the Anthropic Messages API.

@@ -170,9 +181,10 @@ def _is_third_party_anthropic_endpoint(base_url: str | None) -> bool:
    with their own API keys via x-api-key, not Anthropic OAuth tokens. OAuth
    detection should be skipped for these endpoints.
    """
-    if not base_url:
+    normalized = _normalize_base_url_text(base_url)
+    if not normalized:
        return False  # No base_url = direct Anthropic API
-    normalized = base_url.rstrip("/").lower()
+    normalized = normalized.rstrip("/").lower()
    if "anthropic.com" in normalized:
        return False  # Direct Anthropic API — OAuth applies
    return True  # Any other endpoint is a third-party proxy
@@ -182,12 +194,13 @@ def _requires_bearer_auth(base_url: str | None) -> bool:
    """Return True for Anthropic-compatible providers that require Bearer auth.

    Some third-party /anthropic endpoints implement Anthropic's Messages API but
-    require Authorization: Bearer instead of Anthropic's native x-api-key header.
+    require Authorization: Bearer *** of Anthropic's native x-api-key header.
    MiniMax's global and China Anthropic-compatible endpoints follow this pattern.
    """
-    if not base_url:
+    normalized = _normalize_base_url_text(base_url)
+    if not normalized:
        return False
-    normalized = base_url.rstrip("/").lower()
+    normalized = normalized.rstrip("/").lower()
    return normalized.startswith(("https://api.minimax.io/anthropic", "https://api.minimaxi.com/anthropic"))


@@ -203,13 +216,14 @@ def build_anthropic_client(api_key: str, base_url: str = None):
        )
    from httpx import Timeout

+    normalized_base_url = _normalize_base_url_text(base_url)
    kwargs = {
        "timeout": Timeout(timeout=900.0, connect=10.0),
    }
-    if base_url:
-        kwargs["base_url"] = base_url
+    if normalized_base_url:
+        kwargs["base_url"] = normalized_base_url

-    if _requires_bearer_auth(base_url):
+    if _requires_bearer_auth(normalized_base_url):
        # Some Anthropic-compatible providers (e.g. MiniMax) expect the API key in
        # Authorization: Bearer even for regular API keys. Route those endpoints
        # through auth_token so the SDK sends Bearer auth instead of x-api-key.
@@ -942,12 +956,18 @@ def _convert_content_to_anthropic(content: Any) -> Any:

 def convert_messages_to_anthropic(
    messages: List[Dict],
+    base_url: str | None = None,
 ) -> Tuple[Optional[Any], List[Dict]]:
    """Convert OpenAI-format messages to Anthropic format.

    Returns (system_prompt, anthropic_messages).
    System messages are extracted since Anthropic takes them as a separate param.
    system_prompt is a string or list of content blocks (when cache_control present).
+
+    When *base_url* is provided and points to a third-party Anthropic-compatible
+    endpoint, all thinking block signatures are stripped.  Signatures are
+    Anthropic-proprietary — third-party endpoints cannot validate them and will
+    reject them with HTTP 400 "Invalid signature in thinking block".
    """
    system = None
    result = []
@@ -1102,7 +1122,15 @@ def convert_messages_to_anthropic(
                        curr_content = [{"type": "text", "text": curr_content}]
                    fixed[-1]["content"] = prev_content + curr_content
            else:
-                # Consecutive assistant messages — merge text content
+                # Consecutive assistant messages — merge text content.
+                # Drop thinking blocks from the *second* message: their
+                # signature was computed against a different turn boundary
+                # and becomes invalid once merged.
+                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"))
+                    ]
                prev_blocks = fixed[-1]["content"]
                curr_blocks = m["content"]
                if isinstance(prev_blocks, list) and isinstance(curr_blocks, list):
@@ -1120,6 +1148,79 @@ def convert_messages_to_anthropic(
            fixed.append(m)
    result = fixed

+    # ── Thinking block signature management ──────────────────────────
+    # Anthropic signs thinking blocks against the full turn content.
+    # Any upstream mutation (context compression, session truncation,
+    # orphan stripping, message merging) invalidates the signature,
+    # causing HTTP 400 "Invalid signature in thinking block".
+    #
+    # Signatures are Anthropic-proprietary.  Third-party endpoints
+    # (MiniMax, Azure AI Foundry, self-hosted proxies) cannot validate
+    # them and will reject them outright.  When targeting a third-party
+    # endpoint, strip ALL thinking/redacted_thinking blocks from every
+    # assistant message — the third-party will generate its own
+    # thinking blocks if it supports extended thinking.
+    #
+    # For direct Anthropic (strategy following clawdbot/OpenClaw):
+    # 1. Strip thinking/redacted_thinking from all assistant messages
+    #    EXCEPT the last one — preserves reasoning continuity on the
+    #    current tool-use chain while avoiding stale signature errors.
+    # 2. Downgrade unsigned thinking blocks (no signature) to text —
+    #    Anthropic can't validate them and will reject them.
+    # 3. Strip cache_control from thinking/redacted_thinking blocks —
+    #    cache markers can interfere with signature validation.
+    _THINKING_TYPES = frozenset(("thinking", "redacted_thinking"))
+    _is_third_party = _is_third_party_anthropic_endpoint(base_url)
+
+    last_assistant_idx = None
+    for i in range(len(result) - 1, -1, -1):
+        if result[i].get("role") == "assistant":
+            last_assistant_idx = i
+            break
+
+    for idx, m in enumerate(result):
+        if m.get("role") != "assistant" or not isinstance(m.get("content"), list):
+            continue
+
+        if _is_third_party or idx != last_assistant_idx:
+            # Third-party endpoint: strip ALL thinking blocks from every
+            # assistant message — signatures are Anthropic-proprietary.
+            # Direct Anthropic: strip from non-latest assistant messages only.
+            stripped = [
+                b for b in m["content"]
+                if not (isinstance(b, dict) and b.get("type") in _THINKING_TYPES)
+            ]
+            m["content"] = stripped or [{"type": "text", "text": "(thinking elided)"}]
+        else:
+            # Latest assistant on direct Anthropic: keep signed thinking
+            # blocks for reasoning continuity; downgrade unsigned ones to
+            # plain text.
+            new_content = []
+            for b in m["content"]:
+                if not isinstance(b, dict) or b.get("type") not in _THINKING_TYPES:
+                    new_content.append(b)
+                    continue
+                if b.get("type") == "redacted_thinking":
+                    # Redacted blocks use 'data' for the signature payload
+                    if b.get("data"):
+                        new_content.append(b)
+                    # else: drop — no data means it can't be validated
+                elif b.get("signature"):
+                    # Signed thinking block — keep it
+                    new_content.append(b)
+                else:
+                    # Unsigned thinking — downgrade to text so it's not lost
+                    thinking_text = b.get("thinking", "")
+                    if thinking_text:
+                        new_content.append({"type": "text", "text": thinking_text})
+            m["content"] = new_content or [{"type": "text", "text": "(empty)"}]
+
+        # Strip cache_control from any remaining thinking/redacted_thinking
+        # blocks — cache markers interfere with signature validation.
+        for b in m["content"]:
+            if isinstance(b, dict) and b.get("type") in _THINKING_TYPES:
+                b.pop("cache_control", None)
+
    return system, result


@@ -1133,6 +1234,7 @@ def build_anthropic_kwargs(
    is_oauth: bool = False,
    preserve_dots: bool = False,
    context_length: Optional[int] = None,
+    base_url: str | None = None,
 ) -> Dict[str, Any]:
    """Build kwargs for anthropic.messages.create().

@@ -1146,8 +1248,11 @@ def build_anthropic_kwargs(

    When *preserve_dots* is True, model name dots are not converted to hyphens
    (for Alibaba/DashScope anthropic-compatible endpoints: qwen3.5-plus).
+
+    When *base_url* points to a third-party Anthropic-compatible endpoint,
+    thinking block signatures are stripped (they are Anthropic-proprietary).
    """
-    system, anthropic_messages = convert_messages_to_anthropic(messages)
+    system, anthropic_messages = convert_messages_to_anthropic(messages, base_url=base_url)
    anthropic_tools = convert_tools_to_anthropic(tools) if tools else []

    model = normalize_model_name(model, preserve_dots=preserve_dots)
@@ -1224,9 +1329,9 @@ def build_anthropic_kwargs(
    # Map reasoning_config to Anthropic's thinking parameter.
    # Claude 4.6 models use adaptive thinking + output_config.effort.
    # Older models use manual thinking with budget_tokens.
-    # Haiku models do NOT support extended thinking at all — skip entirely.
+    # Haiku and MiniMax models do NOT support extended thinking — skip entirely.
    if reasoning_config and isinstance(reasoning_config, dict):
-        if reasoning_config.get("enabled") is not False and "haiku" not in model.lower():
+        if reasoning_config.get("enabled") is not False and "haiku" not in model.lower() and "minimax" not in model.lower():
            effort = str(reasoning_config.get("effort", "medium")).lower()
            budget = THINKING_BUDGET.get(effort, 8000)
            if _supports_adaptive_thinking(model):
@@ -59,13 +59,48 @@ from hermes_constants import OPENROUTER_BASE_URL

 logger = logging.getLogger(__name__)

+_PROVIDER_ALIASES = {
+    "google": "gemini",
+    "google-gemini": "gemini",
+    "google-ai-studio": "gemini",
+    "glm": "zai",
+    "z-ai": "zai",
+    "z.ai": "zai",
+    "zhipu": "zai",
+    "kimi": "kimi-coding",
+    "moonshot": "kimi-coding",
+    "minimax-china": "minimax-cn",
+    "minimax_cn": "minimax-cn",
+    "claude": "anthropic",
+    "claude-code": "anthropic",
+}
+
+
+def _normalize_aux_provider(provider: Optional[str], *, for_vision: bool = False) -> str:
+    normalized = (provider or "auto").strip().lower()
+    if normalized.startswith("custom:"):
+        suffix = normalized.split(":", 1)[1].strip()
+        if not suffix:
+            return "custom"
+        normalized = suffix if not for_vision else "custom"
+    if normalized == "codex":
+        return "openai-codex"
+    if normalized == "main":
+        # Resolve to the user's actual main provider so named custom providers
+        # and non-aggregator providers (DeepSeek, Alibaba, etc.) work correctly.
+        main_prov = _read_main_provider()
+        if main_prov and main_prov not in ("auto", "main", ""):
+            return main_prov
+        return "custom"
+    return _PROVIDER_ALIASES.get(normalized, normalized)
+
 # Default auxiliary models for direct API-key providers (cheap/fast for side tasks)
 _API_KEY_PROVIDER_AUX_MODELS: Dict[str, str] = {
    "gemini": "gemini-3-flash-preview",
    "zai": "glm-4.5-flash",
    "kimi-coding": "kimi-k2-turbo-preview",
-    "minimax": "MiniMax-M2.7-highspeed",
-    "minimax-cn": "MiniMax-M2.7-highspeed",
+    "minimax": "MiniMax-M2.7",
+    "minimax-cn": "MiniMax-M2.7",
    "anthropic": "claude-haiku-4-5-20251001",
    "ai-gateway": "google/gemini-3-flash",
    "opencode-zen": "gemini-3-flash",
@@ -92,6 +127,7 @@ auxiliary_is_nous: bool = False
 _OPENROUTER_MODEL = "google/gemini-3-flash-preview"
 _NOUS_MODEL = "google/gemini-3-flash-preview"
 _NOUS_FREE_TIER_VISION_MODEL = "xiaomi/mimo-v2-omni"
+_NOUS_FREE_TIER_AUX_MODEL = "xiaomi/mimo-v2-pro"
 _NOUS_DEFAULT_BASE_URL = "https://inference-api.nousresearch.com/v1"
 _ANTHROPIC_DEFAULT_BASE_URL = "https://api.anthropic.com"
 _AUTH_JSON_PATH = get_hermes_home() / "auth.json"
@@ -105,6 +141,23 @@ _CODEX_AUX_MODEL = "gpt-5.2-codex"
 _CODEX_AUX_BASE_URL = "https://chatgpt.com/backend-api/codex"


+def _to_openai_base_url(base_url: str) -> str:
+    """Normalize an Anthropic-style base URL to OpenAI-compatible format.
+
+    Some providers (MiniMax, MiniMax-CN) expose an ``/anthropic`` endpoint for
+    the Anthropic Messages API and a separate ``/v1`` endpoint for OpenAI chat
+    completions.  The auxiliary client uses the OpenAI SDK, so it must hit the
+    ``/v1`` surface.  Passing the raw ``inference_base_url`` causes requests to
+    land on ``/anthropic/chat/completions`` — a 404.
+    """
+    url = str(base_url or "").strip().rstrip("/")
+    if url.endswith("/anthropic"):
+        rewritten = url[: -len("/anthropic")] + "/v1"
+        logger.debug("Auxiliary client: rewrote base URL %s → %s", url, rewritten)
+        return rewritten
+    return url
+
+
 def _select_pool_entry(provider: str) -> Tuple[bool, Optional[Any]]:
    """Return (pool_exists_for_provider, selected_entry)."""
    try:
@@ -634,7 +687,9 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
            if not api_key:
                continue

-            base_url = _pool_runtime_base_url(entry, pconfig.inference_base_url) or pconfig.inference_base_url
+            base_url = _to_openai_base_url(
+                _pool_runtime_base_url(entry, pconfig.inference_base_url) or pconfig.inference_base_url
+            )
            model = _API_KEY_PROVIDER_AUX_MODELS.get(provider_id, "default")
            logger.debug("Auxiliary text client: %s (%s) via pool", pconfig.name, model)
            extra = {}
@@ -651,7 +706,9 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
        if not api_key:
            continue

-        base_url = str(creds.get("base_url", "")).strip().rstrip("/") or pconfig.inference_base_url
+        base_url = _to_openai_base_url(
+            str(creds.get("base_url", "")).strip().rstrip("/") or pconfig.inference_base_url
+        )
        model = _API_KEY_PROVIDER_AUX_MODELS.get(provider_id, "default")
        logger.debug("Auxiliary text client: %s (%s)", pconfig.name, model)
        extra = {}
@@ -713,7 +770,7 @@ def _try_openrouter() -> Tuple[Optional[OpenAI], Optional[str]]:
                   default_headers=_OR_HEADERS), _OPENROUTER_MODEL


-def _try_nous() -> Tuple[Optional[OpenAI], Optional[str]]:
+def _try_nous(vision: bool = False) -> Tuple[Optional[OpenAI], Optional[str]]:
    nous = _read_nous_auth()
    if not nous:
        return None, None
@@ -725,12 +782,13 @@ def _try_nous() -> Tuple[Optional[OpenAI], Optional[str]]:
    else:
        model = _NOUS_MODEL
    # Free-tier users can't use paid auxiliary models — use the free
-    # multimodal model instead so vision/browser-vision still works.
+    # models instead: mimo-v2-omni for vision, mimo-v2-pro for text tasks.
    try:
        from hermes_cli.models import check_nous_free_tier
        if check_nous_free_tier():
-            model = _NOUS_FREE_TIER_VISION_MODEL
-            logger.debug("Free-tier Nous account — using %s for auxiliary/vision", model)
+            model = _NOUS_FREE_TIER_VISION_MODEL if vision else _NOUS_FREE_TIER_AUX_MODEL
+            logger.debug("Free-tier Nous account — using %s for auxiliary/%s",
+                         model, "vision" if vision else "text")
    except Exception:
        pass
    return (
@@ -1138,11 +1196,7 @@ def resolve_provider_client(
        (client, resolved_model) or (None, None) if auth is unavailable.
    """
    # Normalise aliases
-    provider = (provider or "auto").strip().lower()
-    if provider == "codex":
-        provider = "openai-codex"
-    if provider == "main":
-        provider = "custom"
+    provider = _normalize_aux_provider(provider)

    # ── Auto: try all providers in priority order ────────────────────
    if provider == "auto":
@@ -1238,6 +1292,28 @@ def resolve_provider_client(
                       "but no endpoint credentials found")
        return None, None

+    # ── Named custom providers (config.yaml custom_providers list) ───
+    try:
+        from hermes_cli.runtime_provider import _get_named_custom_provider
+        custom_entry = _get_named_custom_provider(provider)
+        if custom_entry:
+            custom_base = custom_entry.get("base_url", "").strip()
+            custom_key = custom_entry.get("api_key", "").strip() or "no-key-required"
+            if custom_base:
+                final_model = model or _read_main_model() or "gpt-4o-mini"
+                client = OpenAI(api_key=custom_key, base_url=custom_base)
+                logger.debug(
+                    "resolve_provider_client: named custom provider %r (%s)",
+                    provider, final_model)
+                return (_to_async_client(client, final_model) if async_mode
+                        else (client, final_model))
+            logger.warning(
+                "resolve_provider_client: named custom provider %r has no base_url",
+                provider)
+            return None, None
+    except ImportError:
+        pass
+
    # ── API-key providers from PROVIDER_REGISTRY ─────────────────────
    try:
        from hermes_cli.auth import PROVIDER_REGISTRY, resolve_api_key_provider_credentials
@@ -1270,7 +1346,9 @@ def resolve_provider_client(
                         provider, ", ".join(tried_sources))
            return None, None

-        base_url = str(creds.get("base_url", "")).strip().rstrip("/") or pconfig.inference_base_url
+        base_url = _to_openai_base_url(
+            str(creds.get("base_url", "")).strip().rstrip("/") or pconfig.inference_base_url
+        )

        default_model = _API_KEY_PROVIDER_AUX_MODELS.get(provider, "")
        final_model = model or default_model
@@ -1347,19 +1425,11 @@ def get_async_text_auxiliary_client(task: str = ""):
 _VISION_AUTO_PROVIDER_ORDER = (
    "openrouter",
    "nous",
-    "openai-codex",
-    "anthropic",
-    "custom",
 )


 def _normalize_vision_provider(provider: Optional[str]) -> str:
-    provider = (provider or "auto").strip().lower()
-    if provider == "codex":
-        return "openai-codex"
-    if provider == "main":
-        return "custom"
-    return provider
+    return _normalize_aux_provider(provider, for_vision=True)


 def _resolve_strict_vision_backend(provider: str) -> Tuple[Optional[Any], Optional[str]]:
@@ -1367,7 +1437,7 @@ def _resolve_strict_vision_backend(provider: str) -> Tuple[Optional[Any], Option
    if provider == "openrouter":
        return _try_openrouter()
    if provider == "nous":
-        return _try_nous()
+        return _try_nous(vision=True)
    if provider == "openai-codex":
        return _try_codex()
    if provider == "anthropic":
@@ -1400,17 +1470,26 @@ def _preferred_main_vision_provider() -> Optional[str]:
 def get_available_vision_backends() -> List[str]:
    """Return the currently available vision backends in auto-selection order.

-    This is the single source of truth for setup, tool gating, and runtime
-    auto-routing of vision tasks. The selected main provider is preferred when
-    it is also a known-good vision backend; otherwise Hermes falls back through
-    the standard conservative order.
+    Order: active provider → OpenRouter → Nous → stop.  This is the single
+    source of truth for setup, tool gating, and runtime auto-routing of
+    vision tasks.
    """
-    ordered = list(_VISION_AUTO_PROVIDER_ORDER)
-    preferred = _preferred_main_vision_provider()
-    if preferred in ordered:
-        ordered.remove(preferred)
-        ordered.insert(0, preferred)
-    return [provider for provider in ordered if _strict_vision_backend_available(provider)]
+    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 in _VISION_AUTO_PROVIDER_ORDER:
+            if _strict_vision_backend_available(main_provider):
+                available.append(main_provider)
+        else:
+            client, _ = resolve_provider_client(main_provider, _read_main_model())
+            if client is not None:
+                available.append(main_provider)
+    # 2. OpenRouter, 3. Nous — skip if already covered by main provider.
+    for p in _VISION_AUTO_PROVIDER_ORDER:
+        if p not in available and _strict_vision_backend_available(p):
+            available.append(p)
+    return available


 def resolve_vision_provider_client(
@@ -1455,16 +1534,39 @@ def resolve_vision_provider_client(
        return "custom", client, final_model

    if requested == "auto":
-        ordered = list(_VISION_AUTO_PROVIDER_ORDER)
-        preferred = _preferred_main_vision_provider()
-        if preferred in ordered:
-            ordered.remove(preferred)
-            ordered.insert(0, preferred)
+        # Vision auto-detection order:
+        #   1. Active provider + model (user's main chat config)
+        #   2. OpenRouter  (known vision-capable default model)
+        #   3. Nous Portal (known vision-capable default model)
+        #   4. Stop
+        main_provider = _read_main_provider()
+        main_model = _read_main_model()
+        if main_provider and main_provider not in ("auto", ""):
+            if main_provider in _VISION_AUTO_PROVIDER_ORDER:
+                # Known strict backend — use its defaults.
+                sync_client, default_model = _resolve_strict_vision_backend(main_provider)
+                if sync_client is not None:
+                    return _finalize(main_provider, sync_client, default_model)
+            else:
+                # Exotic provider (DeepSeek, Alibaba, named custom, etc.)
+                rpc_client, rpc_model = resolve_provider_client(
+                    main_provider, main_model)
+                if rpc_client is not None:
+                    logger.info(
+                        "Vision auto-detect: using active provider %s (%s)",
+                        main_provider, rpc_model or main_model,
+                    )
+                    return _finalize(
+                        main_provider, rpc_client, rpc_model or main_model)

-        for candidate in ordered:
+        # Fall back through aggregators.
+        for candidate in _VISION_AUTO_PROVIDER_ORDER:
+            if candidate == main_provider:
+                continue  # already tried above
            sync_client, default_model = _resolve_strict_vision_backend(candidate)
            if sync_client is not None:
                return _finalize(candidate, sync_client, default_model)
+
        logger.debug("Auxiliary vision client: none available")
        return None, None, None

@@ -26,12 +26,14 @@ _PROVIDER_PREFIXES: frozenset[str] = frozenset({
    "openrouter", "nous", "openai-codex", "copilot", "copilot-acp",
    "gemini", "zai", "kimi-coding", "minimax", "minimax-cn", "anthropic", "deepseek",
    "opencode-zen", "opencode-go", "ai-gateway", "kilocode", "alibaba",
+    "qwen-oauth",
    "custom", "local",
    # Common aliases
    "google", "google-gemini", "google-ai-studio",
    "glm", "z-ai", "z.ai", "zhipu", "github", "github-copilot",
    "github-models", "kimi", "moonshot", "claude", "deep-seek",
    "opencode", "zen", "go", "vercel", "kilo", "dashscope", "aliyun", "qwen",
+    "qwen-portal",
 })


@@ -113,8 +115,15 @@ DEFAULT_CONTEXT_LENGTHS = {
    "llama": 131072,
    # Qwen
    "qwen": 131072,
-    # MiniMax
-    "minimax": 204800,
+    # MiniMax (lowercase — lookup lowercases model names at line 973)
+    "minimax-m1-256k": 1000000,
+    "minimax-m1-128k": 1000000,
+    "minimax-m1-80k": 1000000,
+    "minimax-m1-40k": 1000000,
+    "minimax-m1": 1000000,
+    "minimax-m2.5": 1048576,
+    "minimax-m2.7": 1048576,
+    "minimax": 1048576,
    # GLM
    "glm": 202752,
    # Kimi
@@ -127,7 +136,7 @@ DEFAULT_CONTEXT_LENGTHS = {
    "deepseek-ai/DeepSeek-V3.2": 65536,
    "moonshotai/Kimi-K2.5": 262144,
    "moonshotai/Kimi-K2-Thinking": 262144,
-    "MiniMaxAI/MiniMax-M2.5": 204800,
+    "MiniMaxAI/MiniMax-M2.5": 1048576,
    "XiaomiMiMo/MiMo-V2-Flash": 32768,
    "mimo-v2-pro": 1048576,
    "mimo-v2-omni": 1048576,
@@ -180,6 +189,7 @@ _URL_TO_PROVIDER: Dict[str, str] = {
    "api.minimax": "minimax",
    "dashscope.aliyuncs.com": "alibaba",
    "dashscope-intl.aliyuncs.com": "alibaba",
+    "portal.qwen.ai": "qwen-oauth",
    "openrouter.ai": "openrouter",
    "generativelanguage.googleapis.com": "gemini",
    "inference-api.nousresearch.com": "nous",
@@ -611,6 +621,59 @@ def _model_id_matches(candidate_id: str, lookup_model: str) -> bool:
    return False


+def query_ollama_num_ctx(model: str, base_url: str) -> Optional[int]:
+    """Query an Ollama server for the model's context length.
+
+    Returns the model's maximum context from GGUF metadata via ``/api/show``,
+    or the explicit ``num_ctx`` from the Modelfile if set.  Returns None if
+    the server is unreachable or not Ollama.
+
+    This is the value that should be passed as ``num_ctx`` in Ollama chat
+    requests to override the default 2048.
+    """
+    import httpx
+
+    bare_model = _strip_provider_prefix(model)
+    server_url = base_url.rstrip("/")
+    if server_url.endswith("/v1"):
+        server_url = server_url[:-3]
+
+    try:
+        server_type = detect_local_server_type(base_url)
+    except Exception:
+        return None
+    if server_type != "ollama":
+        return None
+
+    try:
+        with httpx.Client(timeout=3.0) as client:
+            resp = client.post(f"{server_url}/api/show", json={"name": bare_model})
+            if resp.status_code != 200:
+                return None
+            data = resp.json()
+
+            # Prefer explicit num_ctx from Modelfile parameters (user override)
+            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:
+                                return int(parts[-1])
+                            except ValueError:
+                                pass
+
+            # Fall back to GGUF model_info context_length (training max)
+            model_info = data.get("model_info", {})
+            for key, value in model_info.items():
+                if "context_length" in key and isinstance(value, (int, float)):
+                    return int(value)
+    except Exception:
+        pass
+    return None
+
+
 def _query_local_context_length(model: str, base_url: str) -> Optional[int]:
    """Query a local server for the model's context length."""
    import httpx
@@ -153,6 +153,7 @@ PROVIDER_TO_MODELS_DEV: Dict[str, str] = {
    "minimax-cn": "minimax-cn",
    "deepseek": "deepseek",
    "alibaba": "alibaba",
+    "qwen-oauth": "alibaba",
    "copilot": "github-copilot",
    "ai-gateway": "vercel",
    "opencode-zen": "opencode",
@@ -204,6 +204,30 @@ OPENAI_MODEL_EXECUTION_GUIDANCE = (
    "the result.\n"
    "</tool_persistence>\n"
    "\n"
+    "<mandatory_tool_use>\n"
+    "NEVER answer these from memory or mental computation — ALWAYS use a tool:\n"
+    "- Arithmetic, math, calculations → use terminal or execute_code\n"
+    "- Hashes, encodings, checksums → use terminal (e.g. sha256sum, base64)\n"
+    "- Current time, date, timezone → use terminal (e.g. date)\n"
+    "- System state: OS, CPU, memory, disk, ports, processes → use terminal\n"
+    "- File contents, sizes, line counts → use read_file, search_files, or terminal\n"
+    "- Git history, branches, diffs → use terminal\n"
+    "- Current facts (weather, news, versions) → use web_search\n"
+    "Your memory and user profile describe the USER, not the system you are "
+    "running on. The execution environment may differ from what the user profile "
+    "says about their personal setup.\n"
+    "</mandatory_tool_use>\n"
+    "\n"
+    "<act_dont_ask>\n"
+    "When a question has an obvious default interpretation, act on it immediately "
+    "instead of asking for clarification. Examples:\n"
+    "- 'Is port 443 open?' → check THIS machine (don't ask 'open where?')\n"
+    "- 'What OS am I running?' → check the live system (don't use user profile)\n"
+    "- 'What time is it?' → run `date` (don't guess)\n"
+    "Only ask for clarification when the ambiguity genuinely changes what tool "
+    "you would call.\n"
+    "</act_dont_ask>\n"
+    "\n"
    "<prerequisite_checks>\n"
    "- Before taking an action, check whether prerequisite discovery, lookup, or "
    "context-gathering steps are needed.\n"
@@ -0,0 +1,57 @@
+"""Retry utilities — jittered backoff for decorrelated retries.
+
+Replaces fixed exponential backoff with jittered delays to prevent
+thundering-herd retry spikes when multiple sessions hit the same
+rate-limited provider concurrently.
+"""
+
+import random
+import threading
+import time
+
+# Monotonic counter for jitter seed uniqueness within the same process.
+# Protected by a lock to avoid race conditions in concurrent retry paths
+# (e.g. multiple gateway sessions retrying simultaneously).
+_jitter_counter = 0
+_jitter_lock = threading.Lock()
+
+
+def jittered_backoff(
+    attempt: int,
+    *,
+    base_delay: float = 5.0,
+    max_delay: float = 120.0,
+    jitter_ratio: float = 0.5,
+) -> float:
+    """Compute a jittered exponential backoff delay.
+
+    Args:
+        attempt: 1-based retry attempt number.
+        base_delay: Base delay in seconds for attempt 1.
+        max_delay: Maximum delay cap in seconds.
+        jitter_ratio: Fraction of computed delay to use as random jitter
+            range.  0.5 means jitter is uniform in [0, 0.5 * delay].
+
+    Returns:
+        Delay in seconds: min(base * 2^(attempt-1), max_delay) + jitter.
+
+    The jitter decorrelates concurrent retries so multiple sessions
+    hitting the same provider don't all retry at the same instant.
+    """
+    global _jitter_counter
+    with _jitter_lock:
+        _jitter_counter += 1
+        tick = _jitter_counter
+
+    exponent = max(0, attempt - 1)
+    if exponent >= 63 or base_delay <= 0:
+        delay = max_delay
+    else:
+        delay = min(base_delay * (2 ** exponent), max_delay)
+
+    # Seed from time + counter for decorrelation even with coarse clocks.
+    seed = (time.time_ns() ^ (tick * 0x9E3779B9)) & 0xFFFFFFFF
+    rng = random.Random(seed)
+    jitter = rng.uniform(0, jitter_ratio * delay)
+
+    return delay + jitter
@@ -644,10 +644,14 @@ platform_toolsets:
 # Voice Transcription (Speech-to-Text)
 # =============================================================================
 # Automatically transcribe voice messages on messaging platforms.
-# Requires OPENAI_API_KEY in .env (uses OpenAI Whisper API directly).
+# Providers: local (free, faster-whisper) | groq (free tier) | openai (Whisper API) | mistral (Voxtral Transcribe)
+# Set the corresponding API key in .env: GROQ_API_KEY, OPENAI_API_KEY, or MISTRAL_API_KEY.
 stt:
  enabled: true
+  # provider: "local"          # auto-detected if omitted
  model: "whisper-1"  # whisper-1 (cheapest) | gpt-4o-mini-transcribe | gpt-4o-transcribe
+  # mistral:
+  #   model: "voxtral-mini-latest"  # voxtral-mini-latest | voxtral-mini-2602

 # =============================================================================
 # Response Pacing (Messaging Platforms)
@@ -63,7 +63,7 @@ from agent.usage_pricing import (
    format_duration_compact,
    format_token_count_compact,
 )
-from hermes_cli.banner import _format_context_length
+from hermes_cli.banner import _format_context_length, format_banner_version_label

 _COMMAND_SPINNER_FRAMES = ("⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏")

@@ -612,6 +612,11 @@ def _run_cleanup():
        pass
    # Shut down memory provider (on_session_end + shutdown_all) at actual
    # session boundary — NOT per-turn inside run_conversation().
+    try:
+        from hermes_cli.plugins import invoke_hook as _invoke_hook
+        _invoke_hook("on_session_finalize", session_id=_active_agent_ref.session_id if _active_agent_ref else None, platform="cli")
+    except Exception:
+        pass
    try:
        if _active_agent_ref and hasattr(_active_agent_ref, 'shutdown_memory_provider'):
            _active_agent_ref.shutdown_memory_provider(
@@ -755,7 +760,10 @@ def _setup_worktree(repo_root: str = None) -> Optional[Dict[str, str]]:
 def _cleanup_worktree(info: Dict[str, str] = None) -> None:
    """Remove a worktree and its branch on exit.

-    If the worktree has uncommitted changes, warn and keep it.
+    Preserves the worktree only if it has unpushed commits (real work
+    that hasn't been pushed to any remote).  Uncommitted changes alone
+    (untracked files, test artifacts) are not enough to keep it — agent
+    work lives in commits/PRs, not the working tree.
    """
    global _active_worktree
    info = info or _active_worktree
@@ -771,23 +779,27 @@ def _cleanup_worktree(info: Dict[str, str] = None) -> None:
    if not Path(wt_path).exists():
        return

-    # Check for uncommitted changes
+    # Check for unpushed commits — commits reachable from HEAD but not
+    # from any remote branch.  These represent real work the agent did
+    # but didn't push.
+    has_unpushed = False
    try:
-        status = subprocess.run(
-            ["git", "status", "--porcelain"],
+        result = subprocess.run(
+            ["git", "log", "--oneline", "HEAD", "--not", "--remotes"],
            capture_output=True, text=True, timeout=10, cwd=wt_path,
        )
-        has_changes = bool(status.stdout.strip())
+        has_unpushed = bool(result.stdout.strip())
    except Exception:
-        has_changes = True  # Assume dirty on error — don't delete
+        has_unpushed = True  # Assume unpushed on error — don't delete

-    if has_changes:
-        print(f"\n\033[33m⚠ Worktree has uncommitted changes, keeping: {wt_path}\033[0m")
-        print(f"  To clean up manually: git worktree remove {wt_path}")
+    if has_unpushed:
+        print(f"\n\033[33m⚠ Worktree has unpushed commits, keeping: {wt_path}\033[0m")
+        print(f"  To clean up manually: git worktree remove --force {wt_path}")
        _active_worktree = None
        return

-    # Remove worktree
+    # Remove worktree (even if working tree is dirty — uncommitted
+    # changes without unpushed commits are just artifacts)
    try:
        subprocess.run(
            ["git", "worktree", "remove", wt_path, "--force"],
@@ -796,7 +808,7 @@ def _cleanup_worktree(info: Dict[str, str] = None) -> None:
    except Exception as e:
        logger.debug("Failed to remove worktree: %s", e)

-    # Delete the branch (only if it was never pushed / has no upstream)
+    # Delete the branch
    try:
        subprocess.run(
            ["git", "branch", "-D", branch],
@@ -810,19 +822,27 @@ def _cleanup_worktree(info: Dict[str, str] = None) -> None:


 def _prune_stale_worktrees(repo_root: str, max_age_hours: int = 24) -> None:
-    """Remove worktrees older than max_age_hours that have no uncommitted changes.
+    """Remove stale worktrees and orphaned branches on startup.

-    Runs silently on startup to clean up after crashed/killed sessions.
+    Age-based tiers:
+    - Under max_age_hours (24h): skip — session may still be active.
+    - 24h–72h: remove if no unpushed commits.
+    - Over 72h: force remove regardless (nothing should sit this long).
+
+    Also prunes orphaned ``hermes/*`` and ``pr-*`` local branches that
+    have no corresponding worktree.
    """
    import subprocess
    import time

    worktrees_dir = Path(repo_root) / ".worktrees"
    if not worktrees_dir.exists():
+        _prune_orphaned_branches(repo_root)
        return

    now = time.time()
-    cutoff = now - (max_age_hours * 3600)
+    soft_cutoff = now - (max_age_hours * 3600)       # 24h default
+    hard_cutoff = now - (max_age_hours * 3 * 3600)   # 72h default

    for entry in worktrees_dir.iterdir():
        if not entry.is_dir() or not entry.name.startswith("hermes-"):
@@ -831,21 +851,24 @@ def _prune_stale_worktrees(repo_root: str, max_age_hours: int = 24) -> None:
        # Check age
        try:
            mtime = entry.stat().st_mtime
-            if mtime > cutoff:
+            if mtime > soft_cutoff:
                continue  # Too recent — skip
        except Exception:
            continue

-        # Check for uncommitted changes
-        try:
-            status = subprocess.run(
-                ["git", "status", "--porcelain"],
-                capture_output=True, text=True, timeout=5, cwd=str(entry),
-            )
-            if status.stdout.strip():
-                continue  # Has changes — skip
-        except Exception:
-            continue  # Can't check — skip
+        force = mtime <= hard_cutoff  # Over 72h — force remove
+
+        if not force:
+            # 24h–72h tier: only remove if no unpushed commits
+            try:
+                result = subprocess.run(
+                    ["git", "log", "--oneline", "HEAD", "--not", "--remotes"],
+                    capture_output=True, text=True, timeout=5, cwd=str(entry),
+                )
+                if result.stdout.strip():
+                    continue  # Has unpushed commits — skip
+            except Exception:
+                continue  # Can't check — skip

        # Safe to remove
        try:
@@ -864,10 +887,81 @@ def _prune_stale_worktrees(repo_root: str, max_age_hours: int = 24) -> None:
                    ["git", "branch", "-D", branch],
                    capture_output=True, text=True, timeout=10, cwd=repo_root,
                )
-            logger.debug("Pruned stale worktree: %s", entry.name)
+            logger.debug("Pruned stale worktree: %s (force=%s)", entry.name, force)
        except Exception as e:
            logger.debug("Failed to prune worktree %s: %s", entry.name, e)

+    _prune_orphaned_branches(repo_root)
+
+
+def _prune_orphaned_branches(repo_root: str) -> None:
+    """Delete local ``hermes/hermes-*`` and ``pr-*`` branches with no worktree.
+
+    These are auto-generated by ``hermes -w`` sessions and PR review
+    workflows respectively.  Once their worktree is gone they serve no
+    purpose and just accumulate.
+    """
+    import subprocess
+
+    try:
+        result = subprocess.run(
+            ["git", "branch", "--format=%(refname:short)"],
+            capture_output=True, text=True, timeout=10, cwd=repo_root,
+        )
+        if result.returncode != 0:
+            return
+        all_branches = [b.strip() for b in result.stdout.strip().split("\n") if b.strip()]
+    except Exception:
+        return
+
+    # Collect branches that are actively checked out in a worktree
+    active_branches: set = set()
+    try:
+        wt_result = subprocess.run(
+            ["git", "worktree", "list", "--porcelain"],
+            capture_output=True, text=True, timeout=10, cwd=repo_root,
+        )
+        for line in wt_result.stdout.split("\n"):
+            if line.startswith("branch refs/heads/"):
+                active_branches.add(line.split("branch refs/heads/", 1)[-1].strip())
+    except Exception:
+        return  # Can't determine active branches — bail
+
+    # Also protect the currently checked-out branch and main
+    try:
+        head_result = subprocess.run(
+            ["git", "branch", "--show-current"],
+            capture_output=True, text=True, timeout=5, cwd=repo_root,
+        )
+        current = head_result.stdout.strip()
+        if current:
+            active_branches.add(current)
+    except Exception:
+        pass
+    active_branches.add("main")
+
+    orphaned = [
+        b for b in all_branches
+        if b not in active_branches
+        and (b.startswith("hermes/hermes-") or b.startswith("pr-"))
+    ]
+
+    if not orphaned:
+        return
+
+    # Delete in batches
+    for i in range(0, len(orphaned), 50):
+        batch = orphaned[i:i + 50]
+        try:
+            subprocess.run(
+                ["git", "branch", "-D"] + batch,
+                capture_output=True, text=True, timeout=30, cwd=repo_root,
+            )
+        except Exception as e:
+            logger.debug("Failed to prune orphaned branches: %s", e)
+
+    logger.debug("Pruned %d orphaned branches", len(orphaned))
+
 # ============================================================================
 # ASCII Art & Branding
 # ============================================================================
@@ -1036,21 +1130,44 @@ COMPACT_BANNER = """

 def _build_compact_banner() -> str:
    """Build a compact banner that fits the current terminal width."""
-    w = min(shutil.get_terminal_size().columns - 2, 64)
+    try:
+        from hermes_cli.skin_engine import get_active_skin
+        _skin = get_active_skin()
+    except Exception:
+        _skin = None
+
+    skin_name = getattr(_skin, "name", "default") if _skin else "default"
+    border_color = _skin.get_color("banner_border", "#FFD700") if _skin else "#FFD700"
+    title_color = _skin.get_color("banner_title", "#FFBF00") if _skin else "#FFBF00"
+    dim_color = _skin.get_color("banner_dim", "#B8860B") if _skin else "#B8860B"
+
+    if skin_name == "default":
+        line1 = "⚕ NOUS HERMES - AI Agent Framework"
+        tiny_line = "⚕ NOUS HERMES"
+    else:
+        agent_name = _skin.get_branding("agent_name", "Hermes Agent") if _skin else "Hermes Agent"
+        line1 = f"{agent_name} - AI Agent Framework"
+        tiny_line = agent_name
+
+    version_line = format_banner_version_label()
+
+    w = min(shutil.get_terminal_size().columns - 2, 88)
    if w < 30:
-        return "\n[#FFBF00]⚕ NOUS HERMES[/] [dim #B8860B]- Nous Research[/]\n"
+        return f"\n[{title_color}]{tiny_line}[/] [dim {dim_color}]- Nous Research[/]\n"
+
    inner = w - 2  # inside the box border
    bar = "═" * w
-    line1 = "⚕ NOUS HERMES - AI Agent Framework"
-    line2 = "Messenger of the Digital Gods  ·  Nous Research"
+    content_width = inner - 2
+
    # Truncate and pad to fit
-    line1 = line1[:inner - 2].ljust(inner - 2)
-    line2 = line2[:inner - 2].ljust(inner - 2)
+    line1 = line1[:content_width].ljust(content_width)
+    line2 = version_line[:content_width].ljust(content_width)
+
    return (
-        f"\n[bold #FFD700]╔{bar}╗[/]\n"
-        f"[bold #FFD700]║[/] [#FFBF00]{line1}[/] [bold #FFD700]║[/]\n"
-        f"[bold #FFD700]║[/] [dim #B8860B]{line2}[/] [bold #FFD700]║[/]\n"
-        f"[bold #FFD700]╚{bar}╝[/]\n"
+        f"\n[bold {border_color}]╔{bar}╗[/]\n"
+        f"[bold {border_color}]║[/] [{title_color}]{line1}[/] [bold {border_color}]║[/]\n"
+        f"[bold {border_color}]║[/] [dim {dim_color}]{line2}[/] [bold {border_color}]║[/]\n"
+        f"[bold {border_color}]╚{bar}╝[/]\n"
    )


@@ -2163,7 +2280,7 @@ class HermesCLI:
            )
        except Exception as exc:
            message = format_runtime_provider_error(exc)
-            self.console.print(f"[bold red]{message}[/]")
+            ChatConsole().print(f"[bold red]{message}[/]")
            return False

        api_key = runtime.get("api_key")
@@ -2378,7 +2495,7 @@ class HermesCLI:
                    self._pending_title = None
            return True
        except Exception as e:
-            self.console.print(f"[bold red]Failed to initialize agent: {e}[/]")
+            ChatConsole().print(f"[bold red]Failed to initialize agent: {e}[/]")
            return False
    
    def show_banner(self):
@@ -3291,6 +3408,22 @@ class HermesCLI:
        flush_tool_summary()
        print()
    
+    def _notify_session_boundary(self, event_type: str) -> None:
+        """Fire a session-boundary plugin hook (on_session_finalize or on_session_reset).
+
+        Non-blocking — errors are caught and logged.  Safe to call from any
+        lifecycle point (shutdown, /new, /reset).
+        """
+        try:
+            from hermes_cli.plugins import invoke_hook as _invoke_hook
+            _invoke_hook(
+                event_type,
+                session_id=self.agent.session_id if self.agent else None,
+                platform=getattr(self, "platform", None) or "cli",
+            )
+        except Exception:
+            pass
+
    def new_session(self, silent=False):
        """Start a fresh session with a new session ID and cleared agent state."""
        if self.agent and self.conversation_history:
@@ -3298,6 +3431,10 @@ class HermesCLI:
                self.agent.flush_memories(self.conversation_history)
            except (Exception, KeyboardInterrupt):
                pass
+            self._notify_session_boundary("on_session_finalize")
+        elif self.agent:
+            # First session or empty history — still finalize the old session
+            self._notify_session_boundary("on_session_finalize")

        old_session_id = self.session_id
        if self._session_db and old_session_id:
@@ -3342,6 +3479,7 @@ class HermesCLI:
                    )
                except Exception:
                    pass
+            self._notify_session_boundary("on_session_reset")

        if not silent:
            print("(^_^)v New session started!")
@@ -4572,7 +4710,7 @@ class HermesCLI:
                    if hasattr(self, '_pending_input'):
                        self._pending_input.put(msg)
                else:
-                    self.console.print(f"[bold red]Failed to load skill for {base_cmd}[/]")
+                    ChatConsole().print(f"[bold red]Failed to load skill for {base_cmd}[/]")
            else:
                # Prefix matching: if input uniquely identifies one command, execute it.
                # Matches against both built-in COMMANDS and installed skill commands so
@@ -4633,14 +4771,14 @@ class HermesCLI:
        )

        if not msg:
-            self.console.print("[bold red]Failed to load the bundled /plan skill[/]")
+            ChatConsole().print("[bold red]Failed to load the bundled /plan skill[/]")
            return

        _cprint(f"  📝 Plan mode queued via skill. Markdown plan target: {plan_path}")
        if hasattr(self, '_pending_input'):
            self._pending_input.put(msg)
        else:
-            self.console.print("[bold red]Plan mode unavailable: input queue not initialized[/]")
+            ChatConsole().print("[bold red]Plan mode unavailable: input queue not initialized[/]")
    
    def _handle_background_command(self, cmd: str):
        """Handle /background <prompt> — run a prompt in a separate background session.
@@ -574,12 +574,16 @@ def remove_job(job_id: str) -> bool:
    return False


-def mark_job_run(job_id: str, success: bool, error: Optional[str] = None):
+def mark_job_run(job_id: str, success: bool, error: Optional[str] = None,
+                 delivery_error: Optional[str] = None):
    """
    Mark a job as having been run.
    
    Updates last_run_at, last_status, increments completed count,
    computes next_run_at, and auto-deletes if repeat limit reached.
+
+    ``delivery_error`` is tracked separately from the agent error — a job
+    can succeed (agent produced output) but fail delivery (platform down).
    """
    jobs = load_jobs()
    for i, job in enumerate(jobs):
@@ -588,6 +592,8 @@ def mark_job_run(job_id: str, success: bool, error: Optional[str] = None):
            job["last_run_at"] = now
            job["last_status"] = "ok" if success else "error"
            job["last_error"] = error if not success else None
+            # Track delivery failures separately — cleared on successful delivery
+            job["last_delivery_error"] = delivery_error
            
            # Increment completed count
            if job.get("repeat"):
@@ -196,7 +196,7 @@ def _send_media_via_adapter(adapter, chat_id: str, media_files: list, metadata:
            logger.warning("Job '%s': failed to send media %s: %s", job.get("id", "?"), media_path, e)


-def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> None:
+def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> Optional[str]:
    """
    Deliver job output to the configured target (origin chat, specific platform, etc.).

@@ -204,16 +204,16 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> None:
    use the live adapter first — this supports E2EE rooms (e.g. Matrix) where
    the standalone HTTP path cannot encrypt.  Falls back to standalone send if
    the adapter path fails or is unavailable.
+
+    Returns None on success, or an error string on failure.
    """
    target = _resolve_delivery_target(job)
    if not target:
        if job.get("deliver", "local") != "local":
-            logger.warning(
-                "Job '%s' deliver=%s but no concrete delivery target could be resolved",
-                job["id"],
-                job.get("deliver", "local"),
-            )
-        return
+            msg = f"no delivery target resolved for deliver={job.get('deliver', 'local')}"
+            logger.warning("Job '%s': %s", job["id"], msg)
+            return msg
+        return None  # local-only jobs don't deliver — not a failure

    platform_name = target["platform"]
    chat_id = target["chat_id"]
@@ -239,19 +239,22 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> None:
    }
    platform = platform_map.get(platform_name.lower())
    if not platform:
-        logger.warning("Job '%s': unknown platform '%s' for delivery", job["id"], platform_name)
-        return
+        msg = f"unknown platform '{platform_name}'"
+        logger.warning("Job '%s': %s", job["id"], msg)
+        return msg

    try:
        config = load_gateway_config()
    except Exception as e:
-        logger.error("Job '%s': failed to load gateway config for delivery: %s", job["id"], e)
-        return
+        msg = f"failed to load gateway config: {e}"
+        logger.error("Job '%s': %s", job["id"], msg)
+        return msg

    pconfig = config.platforms.get(platform)
    if not pconfig or not pconfig.enabled:
-        logger.warning("Job '%s': platform '%s' not configured/enabled", job["id"], platform_name)
-        return
+        msg = f"platform '{platform_name}' not configured/enabled"
+        logger.warning("Job '%s': %s", job["id"], msg)
+        return msg

    # Optionally wrap the content with a header/footer so the user knows this
    # is a cron delivery.  Wrapping is on by default; set cron.wrap_response: false
@@ -307,7 +310,7 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> None:

            if adapter_ok:
                logger.info("Job '%s': delivered to %s:%s via live adapter", job["id"], platform_name, chat_id)
-                return
+                return None
        except Exception as e:
            logger.warning(
                "Job '%s': live adapter delivery to %s:%s failed (%s), falling back to standalone",
@@ -329,13 +332,17 @@ def _deliver_result(job: dict, content: str, adapters=None, loop=None) -> None:
            future = pool.submit(asyncio.run, _send_to_platform(platform, pconfig, chat_id, cleaned_delivery_content, thread_id=thread_id, media_files=media_files))
            result = future.result(timeout=30)
    except Exception as e:
-        logger.error("Job '%s': delivery to %s:%s failed: %s", job["id"], platform_name, chat_id, e)
-        return
+        msg = f"delivery to {platform_name}:{chat_id} failed: {e}"
+        logger.error("Job '%s': %s", job["id"], msg)
+        return msg

    if result and result.get("error"):
-        logger.error("Job '%s': delivery error: %s", job["id"], result["error"])
-    else:
-        logger.info("Job '%s': delivered to %s:%s", job["id"], platform_name, chat_id)
+        msg = f"delivery error: {result['error']}"
+        logger.error("Job '%s': %s", job["id"], msg)
+        return msg
+
+    logger.info("Job '%s': delivered to %s:%s", job["id"], platform_name, chat_id)
+    return None


 _SCRIPT_TIMEOUT = 120  # seconds
@@ -578,11 +585,9 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
        except Exception as e:
            logger.warning("Job '%s': failed to load config.yaml, using defaults: %s", job_id, e)

-        # Reasoning config from env or config.yaml
+        # Reasoning config from config.yaml
        from hermes_constants import parse_reasoning_effort
-        effort = os.getenv("HERMES_REASONING_EFFORT", "")
-        if not effort:
-            effort = str(_cfg.get("agent", {}).get("reasoning_effort", "")).strip()
+        effort = str(_cfg.get("agent", {}).get("reasoning_effort", "")).strip()
        reasoning_config = parse_reasoning_effort(effort)

        # Prefill messages from env or config.yaml
@@ -868,13 +873,15 @@ def tick(verbose: bool = True, adapters=None, loop=None) -> int:
                    logger.info("Job '%s': agent returned %s — skipping delivery", job["id"], SILENT_MARKER)
                    should_deliver = False

+                delivery_error = None
                if should_deliver:
                    try:
-                        _deliver_result(job, deliver_content, adapters=adapters, loop=loop)
+                        delivery_error = _deliver_result(job, deliver_content, adapters=adapters, loop=loop)
                    except Exception as de:
+                        delivery_error = str(de)
                        logger.error("Delivery failed for job %s: %s", job["id"], de)

-                mark_job_run(job["id"], success, error)
+                mark_job_run(job["id"], success, error, delivery_error=delivery_error)
                executed += 1

            except Exception as e:
@@ -21,6 +21,8 @@ from dataclasses import dataclass, field
 from typing import Any, Dict, List, Optional, Set

 from model_tools import handle_function_call
+from tools.terminal_tool import get_active_env
+from tools.tool_result_storage import maybe_persist_tool_result, enforce_turn_budget

 # Thread pool for running sync tool calls that internally use asyncio.run()
 # (e.g., the Modal/Docker/Daytona terminal backends). Running them in a separate
@@ -136,8 +138,10 @@ class HermesAgentLoop:
        max_turns: int = 30,
        task_id: Optional[str] = None,
        temperature: float = 1.0,
+        top_p: Optional[float] = None,
        max_tokens: Optional[int] = None,
        extra_body: Optional[Dict[str, Any]] = None,
+        budget_config: Optional["BudgetConfig"] = None,
    ):
        """
        Initialize the agent loop.
@@ -150,19 +154,26 @@ class HermesAgentLoop:
            max_turns: Maximum number of LLM calls before stopping
            task_id: Unique ID for terminal/browser session isolation
            temperature: Sampling temperature for generation
+            top_p: Nucleus sampling top_p (None = omit, use provider default)
            max_tokens: Max tokens per generation (None for server default)
            extra_body: Extra parameters passed to the OpenAI client's create() call.
                        Used for OpenRouter provider preferences, transforms, etc.
                        e.g. {"provider": {"ignore": ["DeepInfra"]}}
+            budget_config: Tool result persistence budget. Controls per-tool
+                        thresholds, per-turn aggregate budget, and preview size.
+                        If None, uses DEFAULT_BUDGET (current hardcoded values).
        """
+        from tools.budget_config import DEFAULT_BUDGET
        self.server = server
        self.tool_schemas = tool_schemas
        self.valid_tool_names = valid_tool_names
        self.max_turns = max_turns
        self.task_id = task_id or str(uuid.uuid4())
        self.temperature = temperature
+        self.top_p = top_p
        self.max_tokens = max_tokens
        self.extra_body = extra_body
+        self.budget_config = budget_config or DEFAULT_BUDGET

    async def run(self, messages: List[Dict[str, Any]]) -> AgentResult:
        """
@@ -203,6 +214,9 @@ class HermesAgentLoop:
                "temperature": self.temperature,
            }

+            if self.top_p is not None:
+                chat_kwargs["top_p"] = self.top_p
+
            # Only pass tools if we have them
            if self.tool_schemas:
                chat_kwargs["tools"] = self.tool_schemas
@@ -217,20 +231,35 @@ class HermesAgentLoop:
                chat_kwargs["extra_body"] = self.extra_body

            # Make the API call -- standard OpenAI spec
+            # Retry on timeout/connection errors (provider queuing, rate limits)
            api_start = _time.monotonic()
-            try:
-                response = await self.server.chat_completion(**chat_kwargs)
-            except Exception as e:
-                api_elapsed = _time.monotonic() - api_start
-                logger.error("API call failed on turn %d (%.1fs): %s", turn + 1, api_elapsed, e)
-                return AgentResult(
-                    messages=messages,
-                    managed_state=self._get_managed_state(),
-                    turns_used=turn + 1,
-                    finished_naturally=False,
-                    reasoning_per_turn=reasoning_per_turn,
-                    tool_errors=tool_errors,
-                )
+            response = None
+            max_retries = 3
+            for attempt in range(max_retries):
+                try:
+                    response = await self.server.chat_completion(**chat_kwargs)
+                    break
+                except Exception as e:
+                    api_elapsed = _time.monotonic() - api_start
+                    is_retryable = "timeout" in type(e).__name__.lower() or "connection" in type(e).__name__.lower()
+                    if is_retryable and attempt < max_retries - 1:
+                        wait = 2 ** attempt
+                        logger.warning(
+                            "[%s] API call timed out on turn %d attempt %d (%.1fs), retrying in %ds: %s",
+                            self.task_id[:8], turn + 1, attempt + 1, api_elapsed, wait, e,
+                        )
+                        await asyncio.sleep(wait)
+                        api_start = _time.monotonic()
+                        continue
+                    logger.error("API call failed on turn %d (%.1fs): %s", turn + 1, api_elapsed, e)
+                    return AgentResult(
+                        messages=messages,
+                        managed_state=self._get_managed_state(),
+                        turns_used=turn + 1,
+                        finished_naturally=False,
+                        reasoning_per_turn=reasoning_per_turn,
+                        tool_errors=tool_errors,
+                    )

            api_elapsed = _time.monotonic() - api_start

@@ -446,8 +475,15 @@ class HermesAgentLoop:
                        except (json.JSONDecodeError, TypeError):
                            pass

-                    # Add tool response to conversation
                    tc_id = tc.get("id", "") if isinstance(tc, dict) else tc.id
+                    tool_result = maybe_persist_tool_result(
+                        content=tool_result,
+                        tool_name=tool_name,
+                        tool_use_id=tc_id,
+                        env=get_active_env(self.task_id),
+                        config=self.budget_config,
+                    )
+
                    messages.append(
                        {
                            "role": "tool",
@@ -456,6 +492,14 @@ class HermesAgentLoop:
                        }
                    )

+                num_tcs = len(assistant_msg.tool_calls)
+                if num_tcs > 0:
+                    enforce_turn_budget(
+                        messages[-num_tcs:],
+                        env=get_active_env(self.task_id),
+                        config=self.budget_config,
+                    )
+
                turn_elapsed = _time.monotonic() - turn_start
                logger.info(
                    "[%s] turn %d: api=%.1fs, %d tools, turn_total=%.1fs",
@@ -1048,6 +1048,7 @@ class AgenticOPDEnv(HermesAgentBaseEnv):
                    temperature=0.0,
                    max_tokens=self.config.max_token_length,
                    extra_body=self.config.extra_body,
+                    budget_config=self.config.build_budget_config(),
                )
                result = await agent.run(messages)

@@ -15,15 +15,15 @@

 env:
  enabled_toolsets: ["terminal", "file"]
-  max_agent_turns: 60
+  max_agent_turns: 100
  max_token_length: 32000
-  agent_temperature: 0.8
+  agent_temperature: 1.0
  terminal_backend: "modal"
-  terminal_timeout: 300        # 5 min per command (builds, pip install)
-  tool_pool_size: 128          # thread pool for 89 parallel tasks
-  dataset_name: "NousResearch/terminal-bench-2"
+  terminal_timeout: 300 # 5 min per command (builds, pip install)
+  tool_pool_size: 128 # thread pool for 89 parallel tasks
+  dataset_name: "NousResearch/terminal-bench-2-verified-flattened"
  test_timeout: 600
-  task_timeout: 1800           # 30 min wall-clock per task, auto-FAIL if exceeded
+  task_timeout: 900 # 15 min wall-clock per task, auto-FAIL if exceeded
  tokenizer_name: "NousResearch/Hermes-3-Llama-3.1-8B"
  use_wandb: true
  wandb_name: "terminal-bench-2"
@@ -33,10 +33,15 @@ env:
  # Modal's blocking calls (App.lookup, etc.) deadlock when too many sandboxes
  # are created simultaneously inside thread pool workers via asyncio.run().
  max_concurrent_tasks: 8
+  extra_body:
+    provider:
+      order: ["DeepInfra"]
+      allow_fallbacks: false

 openai:
  base_url: "https://openrouter.ai/api/v1"
-  model_name: "anthropic/claude-opus-4.6"
+  model_name: "nvidia/nemotron-3-super-120b-a12b"
  server_type: "openai"
  health_check: false
+  timeout: 300 # 5 min per API call (default 1200s causes 20min stalls)
  # api_key loaded from OPENROUTER_API_KEY in .env
@@ -32,8 +32,8 @@ export PYTHONUNBUFFERED=1
 # These go to the log file; tqdm + [START]/[PASS]/[FAIL] go to terminal
 export LOGLEVEL=INFO

-python terminalbench2_env.py evaluate \
-  --config default.yaml \
+uv run python environments/benchmarks/terminalbench_2/terminalbench2_env.py evaluate \
+  --config environments/benchmarks/terminalbench_2/default.yaml \
  "$@" \
  2>&1 | tee "$LOG_FILE"

@@ -44,7 +44,7 @@ import tempfile
 import time
 import uuid
 from collections import defaultdict
-from pathlib import Path
+from pathlib import Path, PurePosixPath, PureWindowsPath
 from typing import Any, Dict, List, Optional, Tuple, Union

 # Ensure repo root is on sys.path for imports
@@ -52,18 +52,18 @@ _repo_root = Path(__file__).resolve().parent.parent.parent.parent
 if str(_repo_root) not in sys.path:
    sys.path.insert(0, str(_repo_root))

-from pydantic import Field
-
 from atroposlib.envs.base import EvalHandlingEnum
 from atroposlib.envs.server_handling.server_manager import APIServerConfig
+from pydantic import Field

+from agent.prompt_builder import DEFAULT_AGENT_IDENTITY
 from environments.agent_loop import AgentResult, HermesAgentLoop
 from environments.hermes_base_env import HermesAgentBaseEnv, HermesAgentEnvConfig
 from environments.tool_context import ToolContext
 from tools.terminal_tool import (
-    register_task_env_overrides,
-    clear_task_env_overrides,
    cleanup_vm,
+    clear_task_env_overrides,
+    register_task_env_overrides,
 )

 logger = logging.getLogger(__name__)
@@ -73,6 +73,7 @@ logger = logging.getLogger(__name__)
 # Configuration
 # =============================================================================

+
 class TerminalBench2EvalConfig(HermesAgentEnvConfig):
    """
    Configuration for the Terminal-Bench 2.0 evaluation environment.
@@ -138,16 +139,87 @@ class TerminalBench2EvalConfig(HermesAgentEnvConfig):

 # Tasks that cannot run properly on Modal and are excluded from scoring.
 MODAL_INCOMPATIBLE_TASKS = {
-    "qemu-startup",        # Needs KVM/hardware virtualization
-    "qemu-alpine-ssh",     # Needs KVM/hardware virtualization
-    "crack-7z-hash",       # Password brute-force -- too slow for cloud sandbox timeouts
+    "qemu-startup",  # Needs KVM/hardware virtualization
+    "qemu-alpine-ssh",  # Needs KVM/hardware virtualization
+    "crack-7z-hash",  # Password brute-force -- too slow for cloud sandbox timeouts
 }

+# Injected as a user message when the model responds with plain text instead of
+# calling a tool or including a <task_status> tag.
+_FORMAT_NUDGE_MESSAGE = (
+    "You wrote a plain text response instead of using your tools. "
+    "Plain text responses do not affect the environment — nothing was executed or saved.\n\n"
+    "You MUST use your tools (terminal, read_file, write_file) to actually complete the task. "
+    "Do not describe what you would do — execute it now by making tool calls.\n\n"
+    "If you have already completed all required work using tools in previous turns, "
+    "respond with exactly: <task_status>DONE</task_status>\n"
+    "If you have exhausted all approaches and cannot make further progress, "
+    "respond with exactly: <task_status>UNFINISHED</task_status>"
+)
+
+# Maximum number of format nudges before giving up and moving on to scoring.
+_MAX_FORMAT_NUDGES = 3
+

 # =============================================================================
 # Tar extraction helper
 # =============================================================================

+def _normalize_tar_member_parts(member_name: str) -> list:
+    """Return safe path components for a tar member or raise ValueError."""
+    normalized_name = member_name.replace("\\", "/")
+    posix_path = PurePosixPath(normalized_name)
+    windows_path = PureWindowsPath(member_name)
+
+    if (
+        not normalized_name
+        or posix_path.is_absolute()
+        or windows_path.is_absolute()
+        or windows_path.drive
+    ):
+        raise ValueError(f"Unsafe archive member path: {member_name}")
+
+    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
+
+
+def _safe_extract_tar(tar: tarfile.TarFile, target_dir: Path) -> None:
+    """Extract a tar archive without allowing traversal or link entries."""
+    target_dir.mkdir(parents=True, exist_ok=True)
+    target_root = target_dir.resolve()
+
+    for member in tar.getmembers():
+        parts = _normalize_tar_member_parts(member.name)
+        target = target_dir.joinpath(*parts)
+        target_real = target.resolve(strict=False)
+
+        try:
+            target_real.relative_to(target_root)
+        except ValueError as exc:
+            raise ValueError(f"Unsafe archive member path: {member.name}") from exc
+
+        if member.isdir():
+            target_real.mkdir(parents=True, exist_ok=True)
+            continue
+
+        if not member.isfile():
+            raise ValueError(f"Unsupported archive member type: {member.name}")
+
+        target_real.parent.mkdir(parents=True, exist_ok=True)
+        extracted = tar.extractfile(member)
+        if extracted is None:
+            raise ValueError(f"Cannot read archive member: {member.name}")
+
+        with extracted, open(target_real, "wb") as dst:
+            shutil.copyfileobj(extracted, dst)
+
+        try:
+            os.chmod(target_real, member.mode & 0o777)
+        except OSError:
+            pass
+
 def _extract_base64_tar(b64_data: str, target_dir: Path):
    """Extract a base64-encoded tar.gz archive into target_dir."""
    if not b64_data:
@@ -155,13 +227,14 @@ def _extract_base64_tar(b64_data: str, target_dir: Path):
    raw = base64.b64decode(b64_data)
    buf = io.BytesIO(raw)
    with tarfile.open(fileobj=buf, mode="r:gz") as tar:
-        tar.extractall(path=str(target_dir))
+        _safe_extract_tar(tar, target_dir)


 # =============================================================================
 # Main Environment
 # =============================================================================

+
 class TerminalBench2EvalEnv(HermesAgentBaseEnv):
    """
    Terminal-Bench 2.0 evaluation environment (eval-only, no training).
@@ -206,23 +279,18 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            enabled_toolsets=["terminal", "file"],
            disabled_toolsets=None,
            distribution=None,
-
            # Agent settings -- TB2 tasks are complex, need many turns
            max_agent_turns=60,
            max_token_length=16000,
            agent_temperature=0.6,
-            system_prompt=None,
-
+            system_prompt=DEFAULT_AGENT_IDENTITY,
            # Modal backend for per-task cloud-isolated sandboxes
            terminal_backend="modal",
-            terminal_timeout=300,   # 5 min per command (builds, pip install, etc.)
-
+            terminal_timeout=300,  # 5 min per command (builds, pip install, etc.)
            # Test execution timeout (TB2 test scripts can install deps like pytest)
            test_timeout=180,
-
            # 89 tasks run in parallel, each needs a thread for tool calls
            tool_pool_size=128,
-
            # --- Eval-only Atropos settings ---
            # These settings make the env work as an eval-only environment:
            #   - STOP_TRAIN: pauses training during eval (standard for eval envs)
@@ -232,7 +300,6 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            group_size=1,
            steps_per_eval=1,
            total_steps=1,
-
            tokenizer_name="NousResearch/Hermes-3-Llama-3.1-8B",
            use_wandb=True,
            wandb_name="terminal-bench-2",
@@ -280,7 +347,11 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):

        # Skip tasks incompatible with the current backend (e.g., QEMU on Modal)
        # plus any user-specified skip_tasks
-        skip = set(MODAL_INCOMPATIBLE_TASKS) if self.config.terminal_backend == "modal" else set()
+        skip = (
+            set(MODAL_INCOMPATIBLE_TASKS)
+            if self.config.terminal_backend == "modal"
+            else set()
+        )
        if self.config.skip_tasks:
            skip |= {name.strip() for name in self.config.skip_tasks.split(",")}
        if skip:
@@ -288,7 +359,9 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            tasks = [t for t in tasks if t["task_name"] not in skip]
            skipped = before - len(tasks)
            if skipped > 0:
-                print(f"  Skipped {skipped} incompatible tasks: {sorted(skip & {t['task_name'] for t in ds})}")
+                print(
+                    f"  Skipped {skipped} incompatible tasks: {sorted(skip & {t['task_name'] for t in ds})}"
+                )

        self.all_eval_items = tasks
        self.iter = 0
@@ -298,6 +371,16 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        for i, task in enumerate(self.all_eval_items):
            self.category_index[task.get("category", "unknown")].append(i)

+        # Pre-compute which tasks need Modal's add_python (avoids re-decoding
+        # multi-MB environment_tar blobs during per-task rollouts).
+        self._needs_add_python: Dict[str, bool] = {
+            task["task_name"]: self._image_needs_add_python(task)
+            for task in self.all_eval_items
+        }
+        add_py_count = sum(self._needs_add_python.values())
+        if add_py_count:
+            print(f"  {add_py_count} tasks need add_python (non-python base image)")
+
        # Reward tracking for wandb logging
        self.eval_metrics: List[Tuple[str, float]] = []

@@ -305,15 +388,30 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        # immediately on completion so data is preserved even on Ctrl+C.
        # Timestamped filename so each run produces a unique file.
        import datetime
+
        log_dir = os.path.join(os.path.dirname(__file__), "logs")
        os.makedirs(log_dir, exist_ok=True)
        run_ts = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
-        self._streaming_path = os.path.join(log_dir, f"samples_{run_ts}.jsonl")
+        model_name = self.server.servers[0].config.model_name
+        model_slug = model_name.replace("/", "_").replace(":", "_")
+        self._streaming_path = os.path.join(
+            log_dir, f"samples_{run_ts}_{model_slug}.jsonl"
+        )
        self._streaming_file = open(self._streaming_path, "w")
        self._streaming_lock = __import__("threading").Lock()
+        self._run_meta = {
+            "model_name": model_name,
+            "temperature": self.config.agent_temperature,
+            "top_p": self.config.agent_top_p,
+            "max_agent_turns": self.config.max_agent_turns,
+            "task_timeout": self.config.task_timeout,
+            "terminal_backend": self.config.terminal_backend,
+        }
        print(f"  Streaming results to: {self._streaming_path}")

-        print(f"TB2 ready: {len(self.all_eval_items)} tasks across {len(self.category_index)} categories")
+        print(
+            f"TB2 ready: {len(self.all_eval_items)} tasks across {len(self.category_index)} categories"
+        )
        for cat, indices in sorted(self.category_index.items()):
            print(f"  {cat}: {len(indices)} tasks")

@@ -322,7 +420,9 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        if not hasattr(self, "_streaming_file") or self._streaming_file.closed:
            return
        with self._streaming_lock:
-            self._streaming_file.write(json.dumps(result, ensure_ascii=False, default=str) + "\n")
+            self._streaming_file.write(
+                json.dumps(result, ensure_ascii=False, default=str) + "\n"
+            )
            self._streaming_file.flush()

    # =========================================================================
@@ -358,6 +458,36 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
    # Docker image resolution
    # =========================================================================

+    @staticmethod
+    def _image_needs_add_python(item: Dict[str, Any]) -> bool:
+        """Check if the task's base image lacks `python` on PATH.
+
+        Parses the Dockerfile FROM line in environment_tar. Returns True
+        for non-python base images (ubuntu, debian, etc.) that need
+        Modal's add_python parameter.
+        """
+        environment_tar = item.get("environment_tar", "")
+        if not environment_tar:
+            return False
+        try:
+            raw = base64.b64decode(environment_tar)
+            buf = io.BytesIO(raw)
+            with tarfile.open(fileobj=buf, mode="r:gz") as tar:
+                for member in tar:
+                    if not member.isfile() or "Dockerfile" not in member.name:
+                        continue
+                    f = tar.extractfile(member)
+                    if not f:
+                        continue
+                    for line in f.read().decode("utf-8", errors="ignore").splitlines():
+                        stripped = line.strip()
+                        if stripped.upper().startswith("FROM "):
+                            base = stripped.split()[1].lower()
+                            return not base.startswith("python:")
+        except Exception:
+            pass
+        return False
+
    def _resolve_task_image(
        self, item: Dict[str, Any], task_name: str
    ) -> Tuple[str, Optional[Path]]:
@@ -390,7 +520,9 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            if dockerfile_path.exists():
                logger.info(
                    "Task %s: building from Dockerfile (force_build=%s, docker_image=%s)",
-                    task_name, self.config.force_build, bool(docker_image),
+                    task_name,
+                    self.config.force_build,
+                    bool(docker_image),
                )
                return str(dockerfile_path), task_dir

@@ -398,12 +530,80 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        if docker_image:
            logger.warning(
                "Task %s: force_build=True but no environment_tar, "
-                "falling back to docker_image %s", task_name, docker_image,
+                "falling back to docker_image %s",
+                task_name,
+                docker_image,
            )
            return docker_image, None

        return "", None

+    # =========================================================================
+    # Agent loop with format nudging
+    # =========================================================================
+
+    async def _run_with_nudges(
+        self,
+        server,
+        tools: List[Dict[str, Any]],
+        valid_names: set,
+        messages: List[Dict[str, Any]],
+        task_id: str,
+        task_name: str,
+    ) -> Tuple["AgentResult", int]:
+        """Run the agent loop, nudging if the model returns plain text without task_status tag."""
+        total_turns_used = 0
+        nudge_count = 0
+        result = None
+
+        while total_turns_used < self.config.max_agent_turns:
+            remaining = self.config.max_agent_turns - total_turns_used
+            agent = HermesAgentLoop(
+                server=server,
+                tool_schemas=tools,
+                valid_tool_names=valid_names,
+                max_turns=remaining,
+                task_id=task_id,
+                temperature=self.config.agent_temperature,
+                top_p=self.config.agent_top_p,
+                max_tokens=self.config.max_token_length,
+                extra_body=self.config.extra_body,
+            )
+            result = await agent.run(messages)
+            total_turns_used += result.turns_used
+
+            if not result.finished_naturally:
+                break
+
+            last_content = next(
+                (
+                    m.get("content", "") or ""
+                    for m in reversed(messages)
+                    if m.get("role") == "assistant"
+                ),
+                "",
+            )
+            if "<task_status>" in last_content:
+                break
+
+            if nudge_count >= _MAX_FORMAT_NUDGES:
+                logger.warning(
+                    "Task %s: model ignored %d format nudges; stopping.",
+                    task_name,
+                    nudge_count,
+                )
+                break
+            nudge_count += 1
+            logger.info(
+                "Task %s: nudging model (nudge %d/%d) — no tool calls and no task_status",
+                task_name,
+                nudge_count,
+                _MAX_FORMAT_NUDGES,
+            )
+            messages.append({"role": "user", "content": _FORMAT_NUDGE_MESSAGE})
+
+        return result, total_turns_used
+
    # =========================================================================
    # Per-task evaluation -- agent loop + test verification
    # =========================================================================
@@ -432,6 +632,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        task_dir = None  # Set if we extract a Dockerfile (needs cleanup)

        from tqdm import tqdm
+
        tqdm.write(f"  [START] {task_name} (task_id={task_id[:8]})")
        task_start = time.time()

@@ -439,24 +640,32 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            # --- 1. Resolve Docker image ---
            modal_image, task_dir = self._resolve_task_image(eval_item, task_name)
            if not modal_image:
-                logger.error("Task %s: no docker_image or environment_tar, skipping", task_name)
+                logger.error(
+                    "Task %s: no docker_image or environment_tar, skipping", task_name
+                )
                return {
-                    "passed": False, "reward": 0.0,
-                    "task_name": task_name, "category": category,
+                    "passed": False,
+                    "reward": 0.0,
+                    "task_name": task_name,
+                    "category": category,
                    "error": "no_image",
                }

            # --- 2. Register per-task image override ---
            # Set both modal_image and docker_image so the task image is used
            # regardless of which backend is configured.
-            register_task_env_overrides(task_id, {
+            overrides = {
                "modal_image": modal_image,
                "docker_image": modal_image,
                "cwd": "/app",
-            })
+            }
+            if self._needs_add_python.get(task_name, False):
+                overrides["add_python"] = "3.12"
+            register_task_env_overrides(task_id, overrides)
            logger.info(
                "Task %s: registered image override for task_id %s",
-                task_name, task_id[:8],
+                task_name,
+                task_id[:8],
            )

            # --- 3. Resolve tools and build messages ---
@@ -464,51 +673,48 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):

            messages: List[Dict[str, Any]] = []
            if self.config.system_prompt:
-                messages.append({"role": "system", "content": self.config.system_prompt})
+                messages.append(
+                    {"role": "system", "content": self.config.system_prompt}
+                )
            messages.append({"role": "user", "content": self.format_prompt(eval_item)})

-            # --- 4. Run agent loop ---
-            # Use ManagedServer (Phase 2) for vLLM/SGLang backends to get
-            # token-level tracking via /generate. Falls back to direct
-            # ServerManager (Phase 1) for OpenAI endpoints.
+            # --- 4. Run agent loop with format enforcement ---
+            # The model must either call a tool or end with <task_status>DONE/UNFINISHED</task_status>.
+            # If it returns plain text without the tag, inject a nudge user message and
+            # continue with the remaining turn budget (up to _MAX_FORMAT_NUDGES times).
            if self._use_managed_server():
                async with self.server.managed_server(
                    tokenizer=self.tokenizer,
                    preserve_think_blocks=bool(self.config.thinking_mode),
                ) as managed:
-                    agent = HermesAgentLoop(
+                    result, total_turns_used = await self._run_with_nudges(
                        server=managed,
-                        tool_schemas=tools,
-                        valid_tool_names=valid_names,
-                        max_turns=self.config.max_agent_turns,
+                        tools=tools,
+                        valid_names=valid_names,
+                        messages=messages,
                        task_id=task_id,
-                        temperature=self.config.agent_temperature,
-                        max_tokens=self.config.max_token_length,
-                        extra_body=self.config.extra_body,
+                        task_name=task_name,
                    )
-                    result = await agent.run(messages)
            else:
-                agent = HermesAgentLoop(
+                result, total_turns_used = await self._run_with_nudges(
                    server=self.server,
-                    tool_schemas=tools,
-                    valid_tool_names=valid_names,
-                    max_turns=self.config.max_agent_turns,
+                    tools=tools,
+                    valid_names=valid_names,
+                    messages=messages,
                    task_id=task_id,
-                    temperature=self.config.agent_temperature,
-                    max_tokens=self.config.max_token_length,
-                    extra_body=self.config.extra_body,
+                    task_name=task_name,
                )
-                result = await agent.run(messages)

            # --- 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 messages
            )
-            if result.turns_used == 0 or only_system_and_user:
+            if total_turns_used == 0 or only_system_and_user:
                logger.warning(
                    "Task %s: agent produced no output (turns=%d). Reward=0.",
-                    task_name, result.turns_used,
+                    task_name,
+                    total_turns_used,
                )
                reward = 0.0
            else:
@@ -520,7 +726,10 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
                    loop = asyncio.get_event_loop()
                    reward = await loop.run_in_executor(
                        None,  # default thread pool
-                        self._run_tests, eval_item, ctx, task_name,
+                        self._run_tests,
+                        eval_item,
+                        ctx,
+                        task_name,
                    )
                except Exception as e:
                    logger.error("Task %s: test verification failed: %s", task_name, e)
@@ -531,20 +740,26 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            passed = reward == 1.0
            status = "PASS" if passed else "FAIL"
            elapsed = time.time() - task_start
-            tqdm.write(f"  [{status}] {task_name} (turns={result.turns_used}, {elapsed:.0f}s)")
+            tqdm.write(
+                f"  [{status}] {task_name} (turns={total_turns_used}, {elapsed:.0f}s)"
+            )
            logger.info(
                "Task %s: reward=%.1f, turns=%d, finished=%s",
-                task_name, reward, result.turns_used, result.finished_naturally,
+                task_name,
+                reward,
+                total_turns_used,
+                result.finished_naturally,
            )

            out = {
+                **self._run_meta,
                "passed": passed,
                "reward": reward,
                "task_name": task_name,
                "category": category,
-                "turns_used": result.turns_used,
+                "turns_used": total_turns_used,
                "finished_naturally": result.finished_naturally,
-                "messages": result.messages,
+                "messages": messages,
            }
            self._save_result(out)
            return out
@@ -554,8 +769,11 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            logger.error("Task %s: rollout failed: %s", task_name, e, exc_info=True)
            tqdm.write(f"  [ERROR] {task_name}: {e} ({elapsed:.0f}s)")
            out = {
-                "passed": False, "reward": 0.0,
-                "task_name": task_name, "category": category,
+                **self._run_meta,
+                "passed": False,
+                "reward": 0.0,
+                "task_name": task_name,
+                "category": category,
                "error": str(e),
            }
            self._save_result(out)
@@ -628,7 +846,8 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        # Execute the test suite
        logger.info(
            "Task %s: running test suite (timeout=%ds)",
-            task_name, self.config.test_timeout,
+            task_name,
+            self.config.test_timeout,
        )
        test_result = ctx.terminal(
            "bash /tests/test.sh",
@@ -661,7 +880,9 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
                        logger.warning(
                            "Task %s: reward.txt content unexpected (%r), "
                            "falling back to exit_code=%d",
-                            task_name, content, exit_code,
+                            task_name,
+                            content,
+                            exit_code,
                        )
                        reward = 1.0 if exit_code == 0 else 0.0
            else:
@@ -669,14 +890,17 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
                logger.warning(
                    "Task %s: reward.txt not found after download, "
                    "falling back to exit_code=%d",
-                    task_name, exit_code,
+                    task_name,
+                    exit_code,
                )
                reward = 1.0 if exit_code == 0 else 0.0
        except Exception as e:
            logger.warning(
                "Task %s: failed to download verifier dir: %s, "
                "falling back to exit_code=%d",
-                task_name, e, exit_code,
+                task_name,
+                e,
+                exit_code,
            )
            reward = 1.0 if exit_code == 0 else 0.0
        finally:
@@ -687,7 +911,9 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            output_preview = output[-500:] if output else "(no output)"
            logger.info(
                "Task %s: FAIL (exit_code=%d)\n%s",
-                task_name, exit_code, output_preview,
+                task_name,
+                exit_code,
+                output_preview,
            )

        return reward
@@ -712,12 +938,18 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            )
        except asyncio.TimeoutError:
            from tqdm import tqdm
+
            elapsed = self.config.task_timeout
-            tqdm.write(f"  [TIMEOUT] {task_name} (exceeded {elapsed}s wall-clock limit)")
+            tqdm.write(
+                f"  [TIMEOUT] {task_name} (exceeded {elapsed}s wall-clock limit)"
+            )
            logger.error("Task %s: wall-clock timeout after %ds", task_name, elapsed)
            out = {
-                "passed": False, "reward": 0.0,
-                "task_name": task_name, "category": category,
+                **self._run_meta,
+                "passed": False,
+                "reward": 0.0,
+                "task_name": task_name,
+                "category": category,
                "error": f"timeout ({elapsed}s)",
            }
            self._save_result(out)
@@ -751,23 +983,25 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
                    self.handleError(record)

        handler = _TqdmHandler()
-        handler.setFormatter(logging.Formatter(
-            "%(asctime)s [%(name)s] %(levelname)s: %(message)s",
-            datefmt="%H:%M:%S",
-        ))
+        handler.setFormatter(
+            logging.Formatter(
+                "%(asctime)s [%(name)s] %(levelname)s: %(message)s",
+                datefmt="%H:%M:%S",
+            )
+        )
        root = logging.getLogger()
        root.handlers = [handler]  # Replace any existing handlers
        root.setLevel(logging.INFO)

        # Silence noisy third-party loggers that flood the output
-        logging.getLogger("httpx").setLevel(logging.WARNING)      # Every HTTP request
-        logging.getLogger("openai").setLevel(logging.WARNING)     # OpenAI client retries
-        logging.getLogger("rex-deploy").setLevel(logging.WARNING) # Swerex deployment
+        logging.getLogger("httpx").setLevel(logging.WARNING)  # Every HTTP request
+        logging.getLogger("openai").setLevel(logging.WARNING)  # OpenAI client retries
+        logging.getLogger("rex-deploy").setLevel(logging.WARNING)  # Swerex deployment
        logging.getLogger("rex_image_builder").setLevel(logging.WARNING)  # Image builds

-        print(f"\n{'='*60}")
+        print(f"\n{'=' * 60}")
        print("Starting Terminal-Bench 2.0 Evaluation")
-        print(f"{'='*60}")
+        print(f"{'=' * 60}")
        print(f"  Dataset: {self.config.dataset_name}")
        print(f"  Total tasks: {len(self.all_eval_items)}")
        print(f"  Max agent turns: {self.config.max_agent_turns}")
@@ -775,9 +1009,11 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        print(f"  Terminal backend: {self.config.terminal_backend}")
        print(f"  Tool thread pool: {self.config.tool_pool_size}")
        print(f"  Terminal timeout: {self.config.terminal_timeout}s/cmd")
-        print(f"  Terminal lifetime: {self.config.terminal_lifetime}s (auto: task_timeout + 120)")
+        print(
+            f"  Terminal lifetime: {self.config.terminal_lifetime}s (auto: task_timeout + 120)"
+        )
        print(f"  Max concurrent tasks: {self.config.max_concurrent_tasks}")
-        print(f"{'='*60}\n")
+        print(f"{'=' * 60}\n")

        # Semaphore to limit concurrent Modal sandbox creations.
        # Without this, all 86 tasks fire simultaneously, each creating a Modal
@@ -819,6 +1055,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            await asyncio.gather(*eval_tasks, return_exceptions=True)
            # Belt-and-suspenders: clean up any remaining sandboxes
            from tools.terminal_tool import cleanup_all_environments
+
            cleanup_all_environments()
            print("All sandboxes cleaned up.")
            return
@@ -864,9 +1101,9 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        self.eval_metrics = [(k, v) for k, v in eval_metrics.items()]

        # ---- Print summary ----
-        print(f"\n{'='*60}")
+        print(f"\n{'=' * 60}")
        print("Terminal-Bench 2.0 Evaluation Results")
-        print(f"{'='*60}")
+        print(f"{'=' * 60}")
        print(f"Overall Pass Rate: {overall_pass_rate:.4f} ({passed}/{total})")
        print(f"Evaluation Time: {end_time - start_time:.1f} seconds")

@@ -886,7 +1123,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
            extra = f" (error: {error})" if error else ""
            print(f"  [{status}] {r['task_name']} (turns={turns}){extra}")

-        print(f"{'='*60}\n")
+        print(f"{'=' * 60}\n")

        # Build sample records for evaluate_log (includes full conversations)
        samples = [
@@ -911,6 +1148,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
                end_time=end_time,
                generation_parameters={
                    "temperature": self.config.agent_temperature,
+                    "top_p": self.config.agent_top_p,
                    "max_tokens": self.config.max_token_length,
                    "max_agent_turns": self.config.max_agent_turns,
                    "terminal_backend": self.config.terminal_backend,
@@ -927,6 +1165,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        # Kill all remaining sandboxes. Timed-out tasks leave orphaned thread
        # pool workers still executing commands -- cleanup_all stops them.
        from tools.terminal_tool import cleanup_all_environments
+
        print("\nCleaning up all sandboxes...")
        cleanup_all_environments()

@@ -934,6 +1173,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
        # tasks are killed immediately instead of retrying against dead
        # sandboxes and spamming the console with TimeoutError warnings.
        from environments.agent_loop import _tool_executor
+
        _tool_executor.shutdown(wait=False, cancel_futures=True)
        print("Done.")

@@ -549,6 +549,7 @@ class YCBenchEvalEnv(HermesAgentBaseEnv):
                temperature=self.config.agent_temperature,
                max_tokens=self.config.max_token_length,
                extra_body=self.config.extra_body,
+                budget_config=self.config.build_budget_config(),
            )
            result = await agent.run(messages)

@@ -62,6 +62,11 @@ from atroposlib.type_definitions import Item

 from environments.agent_loop import AgentResult, HermesAgentLoop
 from environments.tool_context import ToolContext
+from tools.budget_config import (
+    DEFAULT_RESULT_SIZE_CHARS,
+    DEFAULT_TURN_BUDGET_CHARS,
+    DEFAULT_PREVIEW_SIZE_CHARS,
+)

 # Import hermes-agent toolset infrastructure
 from model_tools import get_tool_definitions
@@ -110,6 +115,10 @@ class HermesAgentEnvConfig(BaseEnvConfig):
        default=1.0,
        description="Sampling temperature for agent generation during rollouts.",
    )
+    agent_top_p: Optional[float] = Field(
+        default=None,
+        description="Nucleus sampling top_p for agent generation. None = provider default.",
+    )

    # --- Terminal backend ---
    terminal_backend: str = Field(
@@ -160,6 +169,32 @@ class HermesAgentEnvConfig(BaseEnvConfig):
        "Options: hermes, mistral, llama3_json, qwen, deepseek_v3, etc.",
    )

+    # --- Tool result budget ---
+    # Defaults imported from tools.budget_config (single source of truth).
+    default_result_size_chars: int = Field(
+        default=DEFAULT_RESULT_SIZE_CHARS,
+        description="Default per-tool threshold (chars) for persisting large results "
+        "to sandbox. Results exceeding this are written to /tmp/hermes-results/ "
+        "and replaced with a preview. Per-tool registry values take precedence "
+        "unless overridden via tool_result_overrides.",
+    )
+    turn_budget_chars: int = Field(
+        default=DEFAULT_TURN_BUDGET_CHARS,
+        description="Aggregate char budget per assistant turn. If all tool results "
+        "in a single turn exceed this, the largest are persisted to disk first.",
+    )
+    preview_size_chars: int = Field(
+        default=DEFAULT_PREVIEW_SIZE_CHARS,
+        description="Size of the inline preview shown after a tool result is persisted.",
+    )
+    tool_result_overrides: Optional[Dict[str, int]] = Field(
+        default=None,
+        description="Per-tool threshold overrides (chars). Keys are tool names, "
+        "values are char thresholds. Overrides both the default and registry "
+        "per-tool values. Example: {'terminal': 10000, 'search_files': 5000}. "
+        "Note: read_file is pinned to infinity and cannot be overridden.",
+    )
+
    # --- Provider-specific parameters ---
    # Passed as extra_body to the OpenAI client's chat.completions.create() call.
    # Useful for OpenRouter provider preferences, transforms, route settings, etc.
@@ -176,6 +211,16 @@ class HermesAgentEnvConfig(BaseEnvConfig):
        "transforms, and other provider-specific settings.",
    )

+    def build_budget_config(self):
+        """Build a BudgetConfig from env config fields."""
+        from tools.budget_config import BudgetConfig
+        return BudgetConfig(
+            default_result_size=self.default_result_size_chars,
+            turn_budget=self.turn_budget_chars,
+            preview_size=self.preview_size_chars,
+            tool_overrides=dict(self.tool_result_overrides) if self.tool_result_overrides else {},
+        )
+

 class HermesAgentBaseEnv(BaseEnv):
    """
@@ -488,8 +533,10 @@ class HermesAgentBaseEnv(BaseEnv):
                        max_turns=self.config.max_agent_turns,
                        task_id=task_id,
                        temperature=self.config.agent_temperature,
+                        top_p=self.config.agent_top_p,
                        max_tokens=self.config.max_token_length,
                        extra_body=self.config.extra_body,
+                        budget_config=self.config.build_budget_config(),
                    )
                    result = await agent.run(messages)
            except NotImplementedError:
@@ -505,8 +552,10 @@ class HermesAgentBaseEnv(BaseEnv):
                    max_turns=self.config.max_agent_turns,
                    task_id=task_id,
                    temperature=self.config.agent_temperature,
+                    top_p=self.config.agent_top_p,
                    max_tokens=self.config.max_token_length,
                    extra_body=self.config.extra_body,
+                    budget_config=self.config.build_budget_config(),
                )
                result = await agent.run(messages)
        else:
@@ -518,8 +567,10 @@ class HermesAgentBaseEnv(BaseEnv):
                max_turns=self.config.max_agent_turns,
                task_id=task_id,
                temperature=self.config.agent_temperature,
+                top_p=self.config.agent_top_p,
                max_tokens=self.config.max_token_length,
                extra_body=self.config.extra_body,
+                budget_config=self.config.build_budget_config(),
            )
            result = await agent.run(messages)

@@ -472,6 +472,7 @@ class WebResearchEnv(HermesAgentBaseEnv):
                    temperature=0.0,  # Deterministic for eval
                    max_tokens=self.config.max_token_length,
                    extra_body=self.config.extra_body,
+                    budget_config=self.config.build_budget_config(),
                )
                result = await agent.run(messages)

@@ -556,6 +556,18 @@ def load_gateway_config() -> GatewayConfig:
                    os.environ["DISCORD_AUTO_THREAD"] = str(discord_cfg["auto_thread"]).lower()
                if "reactions" in discord_cfg and not os.getenv("DISCORD_REACTIONS"):
                    os.environ["DISCORD_REACTIONS"] = str(discord_cfg["reactions"]).lower()
+                # ignored_channels: channels where bot never responds (even when mentioned)
+                ic = discord_cfg.get("ignored_channels")
+                if ic is not None and not os.getenv("DISCORD_IGNORED_CHANNELS"):
+                    if isinstance(ic, list):
+                        ic = ",".join(str(v) for v in ic)
+                    os.environ["DISCORD_IGNORED_CHANNELS"] = str(ic)
+                # no_thread_channels: channels where bot responds directly without creating thread
+                ntc = discord_cfg.get("no_thread_channels")
+                if ntc is not None and not os.getenv("DISCORD_NO_THREAD_CHANNELS"):
+                    if isinstance(ntc, list):
+                        ntc = ",".join(str(v) for v in ntc)
+                    os.environ["DISCORD_NO_THREAD_CHANNELS"] = str(ntc)

            # Telegram settings → env vars (env vars take precedence)
            telegram_cfg = yaml_cfg.get("telegram", {})
@@ -570,6 +582,8 @@ def load_gateway_config() -> GatewayConfig:
                    if isinstance(frc, list):
                        frc = ",".join(str(v) for v in frc)
                    os.environ["TELEGRAM_FREE_RESPONSE_CHATS"] = str(frc)
+                if "reactions" in telegram_cfg and not os.getenv("TELEGRAM_REACTIONS"):
+                    os.environ["TELEGRAM_REACTIONS"] = str(telegram_cfg["reactions"]).lower()

            whatsapp_cfg = yaml_cfg.get("whatsapp", {})
            if isinstance(whatsapp_cfg, dict):
@@ -698,6 +712,13 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
            name=os.getenv("DISCORD_HOME_CHANNEL_NAME", "Home"),
        )
    
+    # 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 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")
    if whatsapp_enabled:
@@ -20,6 +20,7 @@ Requires:
 """

 import asyncio
+import hmac
 import json
 import logging
 import os
@@ -370,7 +371,7 @@ class APIServerAdapter(BasePlatformAdapter):
        auth_header = request.headers.get("Authorization", "")
        if auth_header.startswith("Bearer "):
            token = auth_header[7:].strip()
-            if token == self._api_key:
+            if hmac.compare_digest(token, self._api_key):
                return None  # Auth OK

        return web.json_response(
@@ -563,8 +564,10 @@ class APIServerAdapter(BasePlatformAdapter):
                if delta is not None:
                    _stream_q.put(delta)

-            def _on_tool_progress(name, preview, args):
+            def _on_tool_progress(event_type, name, preview, args, **kwargs):
                """Inject tool progress into the SSE stream for Open WebUI."""
+                if event_type != "tool.started":
+                    return  # Only show tool start events in chat stream
                if name.startswith("_"):
                    return  # Skip internal events (_thinking)
                from agent.display import get_tool_emoji
@@ -815,9 +818,29 @@ class APIServerAdapter(BasePlatformAdapter):
        else:
            return web.json_response(_openai_error("'input' must be a string or array"), status=400)

-        # Reconstruct conversation history from previous_response_id
+        # Accept explicit conversation_history from the request body.
+        # This lets stateless clients supply their own history instead of
+        # relying on server-side response chaining via previous_response_id.
+        # Precedence: explicit conversation_history > previous_response_id.
        conversation_history: List[Dict[str, str]] = []
-        if previous_response_id:
+        raw_history = body.get("conversation_history")
+        if raw_history:
+            if not isinstance(raw_history, list):
+                return web.json_response(
+                    _openai_error("'conversation_history' must be an array of message objects"),
+                    status=400,
+                )
+            for i, entry in enumerate(raw_history):
+                if not isinstance(entry, dict) or "role" not in entry or "content" not in entry:
+                    return web.json_response(
+                        _openai_error(f"conversation_history[{i}] must have 'role' and 'content' fields"),
+                        status=400,
+                    )
+                conversation_history.append({"role": str(entry["role"]), "content": str(entry["content"])})
+            if previous_response_id:
+                logger.debug("Both conversation_history and previous_response_id provided; using conversation_history")
+
+        if not conversation_history and previous_response_id:
            stored = self._response_store.get(previous_response_id)
            if stored is None:
                return web.json_response(_openai_error(f"Previous response not found: {previous_response_id}"), status=404)
@@ -1403,14 +1426,49 @@ class APIServerAdapter(BasePlatformAdapter):

        instructions = body.get("instructions")
        previous_response_id = body.get("previous_response_id")
+
+        # Accept explicit conversation_history from the request body.
+        # Precedence: explicit conversation_history > previous_response_id.
        conversation_history: List[Dict[str, str]] = []
-        if previous_response_id:
+        raw_history = body.get("conversation_history")
+        if raw_history:
+            if not isinstance(raw_history, list):
+                return web.json_response(
+                    _openai_error("'conversation_history' must be an array of message objects"),
+                    status=400,
+                )
+            for i, entry in enumerate(raw_history):
+                if not isinstance(entry, dict) or "role" not in entry or "content" not in entry:
+                    return web.json_response(
+                        _openai_error(f"conversation_history[{i}] must have 'role' and 'content' fields"),
+                        status=400,
+                    )
+                conversation_history.append({"role": str(entry["role"]), "content": str(entry["content"])})
+            if previous_response_id:
+                logger.debug("Both conversation_history and previous_response_id provided; using conversation_history")
+
+        if not conversation_history and previous_response_id:
            stored = self._response_store.get(previous_response_id)
            if stored:
                conversation_history = list(stored.get("conversation_history", []))
                if instructions is None:
                    instructions = stored.get("instructions")

+        # When input is a multi-message array, extract all but the last
+        # message as conversation history (the last becomes user_message).
+        # Only fires when no explicit history was provided.
+        if not conversation_history and isinstance(raw_input, list) and len(raw_input) > 1:
+            for msg in raw_input[:-1]:
+                if isinstance(msg, dict) and msg.get("role") and msg.get("content"):
+                    content = msg["content"]
+                    if isinstance(content, list):
+                        # Flatten multi-part content blocks to text
+                        content = " ".join(
+                            part.get("text", "") for part in content
+                            if isinstance(part, dict) and part.get("type") == "text"
+                        )
+                    conversation_history.append({"role": msg["role"], "content": str(content)})
+
        session_id = body.get("session_id") or run_id
        ephemeral_system_prompt = instructions

@@ -124,7 +124,14 @@ async def cache_image_from_url(url: str, ext: str = ".jpg", retries: int = 2) ->

    Returns:
        Absolute path to the cached image file as a string.
+
+    Raises:
+        ValueError: If the URL targets a private/internal network (SSRF protection).
    """
+    from tools.url_safety import is_safe_url
+    if not is_safe_url(url):
+        raise ValueError(f"Blocked unsafe URL (SSRF protection): {_safe_url_for_log(url)}")
+
    import asyncio
    import httpx
    import logging as _logging
@@ -232,7 +239,14 @@ async def cache_audio_from_url(url: str, ext: str = ".ogg", retries: int = 2) ->

    Returns:
        Absolute path to the cached audio file as a string.
+
+    Raises:
+        ValueError: If the URL targets a private/internal network (SSRF protection).
    """
+    from tools.url_safety import is_safe_url
+    if not is_safe_url(url):
+        raise ValueError(f"Blocked unsafe URL (SSRF protection): {_safe_url_for_log(url)}")
+
    import asyncio
    import httpx
    import logging as _logging
@@ -1105,6 +1119,22 @@ class BasePlatformAdapter(ABC):
            logger.error("[%s] Fallback send also failed: %s", self.name, fallback_result.error)
        return fallback_result

+    @staticmethod
+    def _merge_caption(existing_text: Optional[str], new_text: str) -> str:
+        """Merge a new caption into existing text, avoiding duplicates.
+
+        Uses line-by-line exact match (not substring) to prevent false positives
+        where a shorter caption is silently dropped because it appears as a
+        substring of a longer one (e.g. "Meeting" inside "Meeting agenda").
+        Whitespace is normalised for comparison.
+        """
+        if not existing_text:
+            return new_text
+        existing_captions = [c.strip() for c in existing_text.split("\n\n")]
+        if new_text.strip() not in existing_captions:
+            return f"{existing_text}\n\n{new_text}".strip()
+        return existing_text
+
    async def handle_message(self, event: MessageEvent) -> None:
        """
        Process an incoming message.
@@ -1164,10 +1194,7 @@ class BasePlatformAdapter(ABC):
                    existing.media_urls.extend(event.media_urls)
                    existing.media_types.extend(event.media_types)
                    if event.text:
-                        if not existing.text:
-                            existing.text = event.text
-                        elif event.text not in existing.text:
-                            existing.text = f"{existing.text}\n\n{event.text}".strip()
+                        existing.text = self._merge_caption(existing.text, event.text)
                else:
                    self._pending_messages[session_key] = event
                return  # Don't interrupt now - will run after current task completes
@@ -55,6 +55,7 @@ from gateway.platforms.base import (
    cache_document_from_bytes,
    SUPPORTED_DOCUMENT_TYPES,
 )
+from tools.url_safety import is_safe_url


 def _clean_discord_id(entry: str) -> str:
@@ -454,6 +455,9 @@ class DiscordAdapter(BasePlatformAdapter):
        self._seen_messages: Dict[str, float] = {}
        self._SEEN_TTL = 300   # 5 minutes
        self._SEEN_MAX = 2000  # prune threshold
+        # Reply threading mode: "off" (no replies), "first" (reply on first
+        # chunk only, default), "all" (reply-reference on every chunk).
+        self._reply_to_mode: str = getattr(config, 'reply_to_mode', 'first') or 'first'

    async def connect(self) -> bool:
        """Connect to Discord and start receiving events."""
@@ -773,7 +777,7 @@ class DiscordAdapter(BasePlatformAdapter):
            message_ids = []
            reference = None

-            if reply_to:
+            if reply_to and self._reply_to_mode != "off":
                try:
                    ref_msg = await channel.fetch_message(int(reply_to))
                    reference = ref_msg
@@ -781,7 +785,10 @@ class DiscordAdapter(BasePlatformAdapter):
                    logger.debug("Could not fetch reply-to message: %s", e)

            for i, chunk in enumerate(chunks):
-                chunk_reference = reference if i == 0 else None
+                if self._reply_to_mode == "all":
+                    chunk_reference = reference
+                else:  # "first" (default) or "off"
+                    chunk_reference = reference if i == 0 else None
                try:
                    msg = await channel.send(
                        content=chunk,
@@ -1285,6 +1292,10 @@ class DiscordAdapter(BasePlatformAdapter):
        if not self._client:
            return SendResult(success=False, error="Not connected")

+        if not is_safe_url(image_url):
+            logger.warning("[%s] Blocked unsafe image URL during Discord send_image", self.name)
+            return await super().send_image(chat_id, image_url, caption, reply_to, metadata=metadata)
+
        try:
            import aiohttp

@@ -2188,9 +2199,11 @@ class DiscordAdapter(BasePlatformAdapter):
        # UNLESS the channel is in the free-response list or the message is
        # in a thread where the bot has already participated.
        #
-        # Config (all settable via discord.* in config.yaml):
+        # Config (all settable via discord.* in config.yaml or DISCORD_* env vars):
        #   discord.require_mention: Require @mention in server channels (default: true)
        #   discord.free_response_channels: Channel IDs where bot responds without mention
+        #   discord.ignored_channels: Channel IDs where bot NEVER responds (even when mentioned)
+        #   discord.no_thread_channels: Channel IDs where bot responds directly without creating thread
        #   discord.auto_thread: Auto-create thread on @mention in channels (default: true)

        thread_id = None
@@ -2201,9 +2214,18 @@ class DiscordAdapter(BasePlatformAdapter):
            parent_channel_id = self._get_parent_channel_id(message.channel)

        if not isinstance(message.channel, discord.DMChannel):
+            # Check ignored channels first - never respond even when mentioned
+            ignored_channels_raw = os.getenv("DISCORD_IGNORED_CHANNELS", "")
+            ignored_channels = {ch.strip() for ch in ignored_channels_raw.split(",") if ch.strip()}
+            channel_ids = {str(message.channel.id)}
+            if parent_channel_id:
+                channel_ids.add(parent_channel_id)
+            if channel_ids & ignored_channels:
+                logger.debug("[%s] Ignoring message in ignored channel: %s", self.name, channel_ids)
+                return
+
            free_channels_raw = os.getenv("DISCORD_FREE_RESPONSE_CHANNELS", "")
            free_channels = {ch.strip() for ch in free_channels_raw.split(",") if ch.strip()}
-            channel_ids = {str(message.channel.id)}
            if parent_channel_id:
                channel_ids.add(parent_channel_id)

@@ -2225,10 +2247,14 @@ class DiscordAdapter(BasePlatformAdapter):
        # Auto-thread: when enabled, automatically create a thread for every
        # @mention in a text channel so each conversation is isolated (like Slack).
        # Messages already inside threads or DMs are unaffected.
+        # no_thread_channels: channels where bot responds directly without thread.
        auto_threaded_channel = None
        if not is_thread and not isinstance(message.channel, discord.DMChannel):
+            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")
-            if auto_thread:
+            if auto_thread and not skip_thread:
                thread = await self._auto_create_thread(message)
                if thread:
                    is_thread = True
@@ -20,6 +20,7 @@ from __future__ import annotations
 import asyncio
 import hashlib
 import hmac
+import itertools
 import json
 import logging
 import mimetypes
@@ -1052,6 +1053,9 @@ class FeishuAdapter(BasePlatformAdapter):
        self._media_batch_state = FeishuBatchState()
        self._pending_media_batches = self._media_batch_state.events
        self._pending_media_batch_tasks = self._media_batch_state.tasks
+        # Exec approval button state (approval_id → {session_key, message_id, chat_id})
+        self._approval_state: Dict[int, Dict[str, str]] = {}
+        self._approval_counter = itertools.count(1)
        self._load_seen_message_ids()

    @staticmethod
@@ -1394,6 +1398,104 @@ class FeishuAdapter(BasePlatformAdapter):
            logger.error("[Feishu] Failed to edit message %s: %s", message_id, exc, exc_info=True)
            return SendResult(success=False, error=str(exc))

+    async def send_exec_approval(
+        self, chat_id: str, command: str, session_key: str,
+        description: str = "dangerous command",
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> SendResult:
+        """Send an interactive card with approval buttons.
+
+        The buttons carry ``hermes_action`` in their value dict so that
+        ``_handle_card_action_event`` can intercept them and call
+        ``resolve_gateway_approval()`` to unblock the waiting agent thread.
+        """
+        if not self._client:
+            return SendResult(success=False, error="Not connected")
+
+        try:
+            approval_id = next(self._approval_counter)
+            cmd_preview = command[:3000] + "..." if len(command) > 3000 else command
+
+            def _btn(label: str, action_name: str, btn_type: str = "default") -> dict:
+                return {
+                    "tag": "button",
+                    "text": {"tag": "plain_text", "content": label},
+                    "type": btn_type,
+                    "value": {"hermes_action": action_name, "approval_id": approval_id},
+                }
+
+            card = {
+                "config": {"wide_screen_mode": True},
+                "header": {
+                    "title": {"content": "⚠️ Command Approval Required", "tag": "plain_text"},
+                    "template": "orange",
+                },
+                "elements": [
+                    {
+                        "tag": "markdown",
+                        "content": f"```\n{cmd_preview}\n```\n**Reason:** {description}",
+                    },
+                    {
+                        "tag": "action",
+                        "actions": [
+                            _btn("✅ Allow Once", "approve_once", "primary"),
+                            _btn("✅ Session", "approve_session"),
+                            _btn("✅ Always", "approve_always"),
+                            _btn("❌ Deny", "deny", "danger"),
+                        ],
+                    },
+                ],
+            }
+
+            payload = json.dumps(card, ensure_ascii=False)
+            response = await self._feishu_send_with_retry(
+                chat_id=chat_id,
+                msg_type="interactive",
+                payload=payload,
+                reply_to=None,
+                metadata=metadata,
+            )
+
+            result = self._finalize_send_result(response, "send_exec_approval failed")
+            if result.success:
+                self._approval_state[approval_id] = {
+                    "session_key": session_key,
+                    "message_id": result.message_id or "",
+                    "chat_id": chat_id,
+                }
+            return result
+        except Exception as exc:
+            logger.warning("[Feishu] send_exec_approval failed: %s", exc)
+            return SendResult(success=False, error=str(exc))
+
+    async def _update_approval_card(
+        self, message_id: str, label: str, user_name: str, choice: str,
+    ) -> None:
+        """Replace the approval card with a resolved status card."""
+        if not self._client or not message_id:
+            return
+        icon = "❌" if choice == "deny" else "✅"
+        card = {
+            "config": {"wide_screen_mode": True},
+            "header": {
+                "title": {"content": f"{icon} {label}", "tag": "plain_text"},
+                "template": "red" if choice == "deny" else "green",
+            },
+            "elements": [
+                {
+                    "tag": "markdown",
+                    "content": f"{icon} **{label}** by {user_name}",
+                },
+            ],
+        }
+        try:
+            payload = json.dumps(card, ensure_ascii=False)
+            body = self._build_update_message_body(msg_type="interactive", content=payload)
+            request = self._build_update_message_request(message_id=message_id, request_body=body)
+            await asyncio.to_thread(self._client.im.v1.message.update, request)
+        except Exception as exc:
+            logger.warning("[Feishu] Failed to update approval card %s: %s", message_id, exc)
+
    async def send_voice(
        self,
        chat_id: str,
@@ -1820,6 +1922,52 @@ class FeishuAdapter(BasePlatformAdapter):
        action = getattr(event, "action", None)
        action_tag = str(getattr(action, "tag", "") or "button")
        action_value = getattr(action, "value", {}) or {}
+
+        # --- Exec approval button intercept ---
+        hermes_action = action_value.get("hermes_action") if isinstance(action_value, dict) else None
+        if hermes_action:
+            approval_id = action_value.get("approval_id")
+            state = self._approval_state.pop(approval_id, None)
+            if not state:
+                logger.debug("[Feishu] Approval %s already resolved or unknown", approval_id)
+                return
+
+            choice_map = {
+                "approve_once": "once",
+                "approve_session": "session",
+                "approve_always": "always",
+                "deny": "deny",
+            }
+            choice = choice_map.get(hermes_action, "deny")
+
+            label_map = {
+                "once": "Approved once",
+                "session": "Approved for session",
+                "always": "Approved permanently",
+                "deny": "Denied",
+            }
+            label = label_map.get(choice, "Resolved")
+
+            # Resolve sender name for the status card
+            sender_id = SimpleNamespace(open_id=open_id, user_id=None, union_id=None)
+            sender_profile = await self._resolve_sender_profile(sender_id)
+            user_name = sender_profile.get("user_name") or open_id
+
+            # Resolve the approval — unblocks the agent thread
+            try:
+                from tools.approval import resolve_gateway_approval
+                count = resolve_gateway_approval(state["session_key"], choice)
+                logger.info(
+                    "Feishu button resolved %d approval(s) for session %s (choice=%s, user=%s)",
+                    count, state["session_key"], choice, user_name,
+                )
+            except Exception as exc:
+                logger.error("Failed to resolve gateway approval from Feishu button: %s", exc)
+
+            # Update the card to show the decision
+            await self._update_approval_card(state.get("message_id", ""), label, user_name, choice)
+            return
+
        synthetic_text = f"/card {action_tag}"
        if action_value:
            try:
@@ -2065,10 +2213,7 @@ class FeishuAdapter(BasePlatformAdapter):
        existing.media_urls.extend(event.media_urls)
        existing.media_types.extend(event.media_types)
        if event.text:
-            if not existing.text:
-                existing.text = event.text
-            elif event.text not in existing.text.split("\n\n"):
-                existing.text = f"{existing.text}\n\n{event.text}"
+            existing.text = self._merge_caption(existing.text, event.text)
        existing.timestamp = event.timestamp
        if event.message_id:
            existing.message_id = event.message_id
@@ -2112,6 +2257,10 @@ class FeishuAdapter(BasePlatformAdapter):
        default_ext: str,
        preferred_name: str,
    ) -> tuple[str, str]:
+        from tools.url_safety import is_safe_url
+        if not is_safe_url(file_url):
+            raise ValueError(f"Blocked unsafe URL (SSRF protection): {file_url[:80]}")
+
        import httpx

        async with httpx.AsyncClient(timeout=30.0, follow_redirects=True) as client:
@@ -586,6 +586,11 @@ class MatrixAdapter(BasePlatformAdapter):
        metadata: Optional[Dict[str, Any]] = None,
    ) -> SendResult:
        """Download an image URL and upload it to Matrix."""
+        from tools.url_safety import is_safe_url
+        if not is_safe_url(image_url):
+            logger.warning("Matrix: blocked unsafe image URL (SSRF protection)")
+            return await super().send_image(chat_id, image_url, caption, reply_to, metadata=metadata)
+
        try:
            # Try aiohttp first (always available), fall back to httpx
            try:
@@ -407,6 +407,11 @@ class MattermostAdapter(BasePlatformAdapter):
        kind: str = "file",
    ) -> SendResult:
        """Download a URL and upload it as a file attachment."""
+        from tools.url_safety import is_safe_url
+        if not is_safe_url(url):
+            logger.warning("Mattermost: blocked unsafe URL (SSRF protection)")
+            return await self.send(chat_id, f"{caption or ''}\n{url}".strip(), reply_to)
+
        import asyncio
        import aiohttp

@@ -647,7 +647,11 @@ class SignalAdapter(BasePlatformAdapter):

        if result is not None:
            self._track_sent_timestamp(result)
-            return SendResult(success=True)
+            # Use the timestamp from the RPC result as a pseudo message_id.
+            # Signal doesn't have real message IDs, but the stream consumer
+            # needs a truthy value to follow its edit→fallback path correctly.
+            _msg_id = str(result.get("timestamp", "")) if isinstance(result, dict) else None
+            return SendResult(success=True, message_id=_msg_id or None)
        return SendResult(success=False, error="RPC send failed")

    def _track_sent_timestamp(self, rpc_result) -> None:
@@ -837,6 +841,11 @@ class SignalAdapter(BasePlatformAdapter):
            except asyncio.CancelledError:
                pass

+    async def stop_typing(self, chat_id: str) -> None:
+        """Public interface for stopping typing — called by base adapter's
+        _keep_typing finally block to clean up platform-level typing tasks."""
+        await self._stop_typing_indicator(chat_id)
+
    # ------------------------------------------------------------------
    # Chat Info
    # ------------------------------------------------------------------
@@ -595,6 +595,11 @@ class SlackAdapter(BasePlatformAdapter):
        if not self._app:
            return SendResult(success=False, error="Not connected")

+        from tools.url_safety import is_safe_url
+        if not is_safe_url(image_url):
+            logger.warning("[Slack] Blocked unsafe image URL (SSRF protection)")
+            return await super().send_image(chat_id, image_url, caption, reply_to, metadata=metadata)
+
        try:
            import httpx

@@ -1632,7 +1632,12 @@ class TelegramAdapter(BasePlatformAdapter):
        """
        if not self._bot:
            return SendResult(success=False, error="Not connected")
-        
+
+        from tools.url_safety import is_safe_url
+        if not is_safe_url(image_url):
+            logger.warning("[%s] Blocked unsafe image URL (SSRF protection)", self.name)
+            return await super().send_image(chat_id, image_url, caption, reply_to, metadata=metadata)
+
        try:
            # Telegram can send photos directly from URLs (up to ~5MB)
            _photo_thread = metadata.get("thread_id") if metadata else None
@@ -2222,10 +2227,7 @@ class TelegramAdapter(BasePlatformAdapter):
            existing.media_urls.extend(event.media_urls)
            existing.media_types.extend(event.media_types)
            if event.text:
-                if not existing.text:
-                    existing.text = event.text
-                elif event.text not in existing.text:
-                    existing.text = f"{existing.text}\n\n{event.text}".strip()
+                existing.text = self._merge_caption(existing.text, event.text)

        prior_task = self._pending_photo_batch_tasks.get(batch_key)
        if prior_task and not prior_task.done():
@@ -2415,11 +2417,7 @@ class TelegramAdapter(BasePlatformAdapter):
            existing.media_urls.extend(event.media_urls)
            existing.media_types.extend(event.media_types)
            if event.text:
-                if existing.text:
-                    if event.text not in existing.text.split("\n\n"):
-                        existing.text = f"{existing.text}\n\n{event.text}"
-                else:
-                    existing.text = event.text
+                existing.text = self._merge_caption(existing.text, event.text)

        prior_task = self._media_group_tasks.get(media_group_id)
        if prior_task:
@@ -2675,3 +2673,46 @@ class TelegramAdapter(BasePlatformAdapter):
            auto_skill=topic_skill,
            timestamp=message.date,
        )
+
+    # ── Message reactions (processing lifecycle) ──────────────────────────
+
+    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")
+
+    async def _set_reaction(self, chat_id: str, message_id: str, emoji: str) -> bool:
+        """Set a single emoji reaction on a Telegram message."""
+        if not self._bot:
+            return False
+        try:
+            await self._bot.set_message_reaction(
+                chat_id=int(chat_id),
+                message_id=int(message_id),
+                reaction=emoji,
+            )
+            return True
+        except Exception as e:
+            logger.debug("[%s] set_message_reaction failed (%s): %s", self.name, emoji, e)
+            return False
+
+    async def on_processing_start(self, event: MessageEvent) -> None:
+        """Add an in-progress reaction when message processing begins."""
+        if not self._reactions_enabled():
+            return
+        chat_id = getattr(event.source, "chat_id", None)
+        message_id = getattr(event, "message_id", None)
+        if chat_id and message_id:
+            await self._set_reaction(chat_id, message_id, "\U0001f440")
+
+    async def on_processing_complete(self, event: MessageEvent, success: bool) -> None:
+        """Swap the in-progress reaction for a final success/failure reaction.
+
+        Unlike Discord (additive reactions), Telegram's set_message_reaction
+        replaces all existing reactions in one call — no remove step needed.
+        """
+        if not self._reactions_enabled():
+            return
+        chat_id = getattr(event.source, "chat_id", None)
+        message_id = getattr(event, "message_id", None)
+        if chat_id and message_id:
+            await self._set_reaction(chat_id, message_id, "\u2705" if success else "\u274c")
@@ -76,8 +76,17 @@ class WebhookAdapter(BasePlatformAdapter):
        self._routes: Dict[str, dict] = dict(self._static_routes)
        self._runner = None

-        # Delivery info keyed by session chat_id — consumed by send()
+        # Delivery info keyed by session chat_id.
+        #
+        # Read by every send() invocation for the chat_id (status messages
+        # AND the final response).  Cleaned up via TTL on each POST so the
+        # dict stays bounded — see _prune_delivery_info().  Do NOT pop on
+        # send(), or interim status messages (e.g. fallback notifications,
+        # context-pressure warnings) will consume the entry before the
+        # final response arrives, causing the response to silently fall
+        # back to the "log" deliver type.
        self._delivery_info: Dict[str, dict] = {}
+        self._delivery_info_created: Dict[str, float] = {}

        # Reference to gateway runner for cross-platform delivery (set externally)
        self.gateway_runner = None
@@ -160,10 +169,14 @@ class WebhookAdapter(BasePlatformAdapter):
    ) -> SendResult:
        """Deliver the agent's response to the configured destination.

-        chat_id is ``webhook:{route}:{delivery_id}`` — we pop the delivery
-        info stored during webhook receipt so it doesn't leak memory.
+        chat_id is ``webhook:{route}:{delivery_id}``.  The delivery info
+        stored during webhook receipt is read with ``.get()`` (not popped)
+        so that interim status messages emitted before the final response
+        — fallback-model notifications, context-pressure warnings, etc. —
+        do not consume the entry and silently downgrade the final response
+        to the ``log`` deliver type.  TTL cleanup happens on POST.
        """
-        delivery = self._delivery_info.pop(chat_id, {})
+        delivery = self._delivery_info.get(chat_id, {})
        deliver_type = delivery.get("deliver", "log")

        if deliver_type == "log":
@@ -190,6 +203,23 @@ class WebhookAdapter(BasePlatformAdapter):
            success=False, error=f"Unknown deliver type: {deliver_type}"
        )

+    def _prune_delivery_info(self, now: float) -> None:
+        """Drop delivery_info entries older than the idempotency TTL.
+
+        Mirrors the cleanup pattern used for ``_seen_deliveries``.  Called
+        on each POST so the dict size is bounded by ``rate_limit * TTL``
+        even if many webhooks fire and never receive a final response.
+        """
+        cutoff = now - self._idempotency_ttl
+        stale = [
+            k
+            for k, t in self._delivery_info_created.items()
+            if t < cutoff
+        ]
+        for k in stale:
+            self._delivery_info.pop(k, None)
+            self._delivery_info_created.pop(k, None)
+
    async def get_chat_info(self, chat_id: str) -> Dict[str, Any]:
        return {"name": chat_id, "type": "webhook"}

@@ -382,7 +412,9 @@ class WebhookAdapter(BasePlatformAdapter):
        # same route get independent agent runs (not queued/interrupted).
        session_chat_id = f"webhook:{route_name}:{delivery_id}"

-        # Store delivery info for send() — consumed (popped) on delivery
+        # Store delivery info for send().  Read by every send() invocation
+        # for this chat_id (interim status messages and the final response),
+        # so we do NOT pop on send.  TTL-based cleanup keeps the dict bounded.
        deliver_config = {
            "deliver": route_config.get("deliver", "log"),
            "deliver_extra": self._render_delivery_extra(
@@ -391,6 +423,8 @@ class WebhookAdapter(BasePlatformAdapter):
            "payload": payload,
        }
        self._delivery_info[session_chat_id] = deliver_config
+        self._delivery_info_created[session_chat_id] = now
+        self._prune_delivery_info(now)

        # Build source and event
        source = self.build_source(
@@ -910,6 +910,10 @@ class WeComAdapter(BasePlatformAdapter):
        url: str,
        max_bytes: int,
    ) -> Tuple[bytes, Dict[str, str]]:
+        from tools.url_safety import is_safe_url
+        if not is_safe_url(url):
+            raise ValueError(f"Blocked unsafe URL (SSRF protection): {url[:80]}")
+
        if not HTTPX_AVAILABLE:
            raise RuntimeError("httpx is required for WeCom media download")

@@ -921,12 +921,11 @@ class GatewayRunner:

    @staticmethod
    def _load_reasoning_config() -> dict | None:
-        """Load reasoning effort from config with env fallback.
+        """Load reasoning effort from config.yaml.

-        Checks agent.reasoning_effort in config.yaml first, then
-        HERMES_REASONING_EFFORT as a fallback. Valid: "xhigh", "high",
-        "medium", "low", "minimal", "none". Returns None to use default
-        (medium).
+        Reads agent.reasoning_effort from config.yaml. Valid: "xhigh",
+        "high", "medium", "low", "minimal", "none". Returns None to use
+        default (medium).
        """
        from hermes_constants import parse_reasoning_effort
        effort = ""
@@ -939,8 +938,6 @@ class GatewayRunner:
                effort = str(cfg.get("agent", {}).get("reasoning_effort", "") or "").strip()
        except Exception:
            pass
-        if not effort:
-            effort = os.getenv("HERMES_REASONING_EFFORT", "")
        result = parse_reasoning_effort(effort)
        if effort and effort.strip() and result is None:
            logger.warning("Unknown reasoning_effort '%s', using default (medium)", effort)
@@ -1484,6 +1481,14 @@ class GatewayRunner:
                logger.debug("Interrupted running agent for session %s during shutdown", session_key[:20])
            except Exception as e:
                logger.debug("Failed interrupting agent during shutdown: %s", e)
+            # Fire plugin on_session_finalize hook before memory shutdown
+            try:
+                from hermes_cli.plugins import invoke_hook as _invoke_hook
+                _invoke_hook("on_session_finalize",
+                             session_id=getattr(agent, 'session_id', None),
+                             platform="gateway")
+            except Exception:
+                pass
            # Shut down memory provider at actual session boundary
            try:
                if hasattr(agent, 'shutdown_memory_provider'):
@@ -1987,10 +1992,7 @@ class GatewayRunner:
                            existing.media_urls.extend(event.media_urls)
                            existing.media_types.extend(event.media_types)
                            if event.text:
-                                if not existing.text:
-                                    existing.text = event.text
-                                elif event.text not in existing.text:
-                                    existing.text = f"{existing.text}\n\n{event.text}".strip()
+                                existing.text = BasePlatformAdapter._merge_caption(existing.text, event.text)
                        else:
                            adapter._pending_messages[_quick_key] = event
                    else:
@@ -3280,6 +3282,15 @@ class GatewayRunner:
        # the configured default instead of the previously switched model.
        self._session_model_overrides.pop(session_key, None)

+        # Fire plugin on_session_finalize hook (session boundary)
+        try:
+            from hermes_cli.plugins import invoke_hook as _invoke_hook
+            _old_sid = old_entry.session_id if old_entry else None
+            _invoke_hook("on_session_finalize", session_id=_old_sid,
+                         platform=source.platform.value if source.platform else "")
+        except Exception:
+            pass
+
        # Emit session:end hook (session is ending)
        await self.hooks.emit("session:end", {
            "platform": source.platform.value if source.platform else "",
@@ -3293,7 +3304,7 @@ class GatewayRunner:
            "user_id": source.user_id,
            "session_key": session_key,
        })
-        
+
        # Resolve session config info to surface to the user
        try:
            session_info = self._format_session_info()
@@ -3304,9 +3315,18 @@ class GatewayRunner:
            header = "✨ Session reset! Starting fresh."
        else:
            # No existing session, just create one
-            self.session_store.get_or_create_session(source, force_new=True)
+            new_entry = self.session_store.get_or_create_session(source, force_new=True)
            header = "✨ New session started!"

+        # Fire plugin on_session_reset hook (new session guaranteed to exist)
+        try:
+            from hermes_cli.plugins import invoke_hook as _invoke_hook
+            _new_sid = new_entry.session_id if new_entry else None
+            _invoke_hook("on_session_reset", session_id=_new_sid,
+                         platform=source.platform.value if source.platform else "")
+        except Exception:
+            pass
+
        if session_info:
            return f"{header}\n\n{session_info}"
        return header
@@ -3345,25 +3365,36 @@ class GatewayRunner:
        """Handle /status command."""
        source = event.source
        session_entry = self.session_store.get_or_create_session(source)
-        
+
        connected_platforms = [p.value for p in self.adapters.keys()]
-        
+
        # Check if there's an active agent
        session_key = session_entry.session_key
        is_running = session_key in self._running_agents
-        
+
+        title = None
+        if self._session_db:
+            try:
+                title = self._session_db.get_session_title(session_entry.session_id)
+            except Exception:
+                title = None
+
        lines = [
            "📊 **Hermes Gateway Status**",
            "",
-            f"**Session ID:** `{session_entry.session_id[:12]}...`",
+            f"**Session ID:** `{session_entry.session_id}`",
+        ]
+        if title:
+            lines.append(f"**Title:** {title}")
+        lines.extend([
            f"**Created:** {session_entry.created_at.strftime('%Y-%m-%d %H:%M')}",
            f"**Last Activity:** {session_entry.updated_at.strftime('%Y-%m-%d %H:%M')}",
            f"**Tokens:** {session_entry.total_tokens:,}",
            f"**Agent Running:** {'Yes ⚡' if is_running else 'No'}",
            "",
            f"**Connected Platforms:** {', '.join(connected_platforms)}",
-        ]
-        
+        ])
+
        return "\n".join(lines)
    
    async def _handle_stop_command(self, event: MessageEvent) -> str:
@@ -4913,8 +4944,8 @@ class GatewayRunner:
        cycle = ["off", "new", "all", "verbose"]
        descriptions = {
            "off": "⚙️ Tool progress: **OFF** — no tool activity shown.",
-            "new": "⚙️ Tool progress: **NEW** — shown when tool changes (short previews).",
-            "all": "⚙️ Tool progress: **ALL** — every tool call shown (short previews).",
+            "new": "⚙️ Tool progress: **NEW** — shown when tool changes (preview length: `display.tool_preview_length`, default 40).",
+            "all": "⚙️ Tool progress: **ALL** — every tool call shown (preview length: `display.tool_preview_length`, default 40).",
            "verbose": "⚙️ Tool progress: **VERBOSE** — every tool call with full arguments.",
        }

@@ -6036,6 +6067,11 @@ class GatewayRunner:

        if enriched_parts:
            prefix = "\n\n".join(enriched_parts)
+            # Strip the empty-content placeholder from the Discord adapter
+            # when we successfully transcribed the audio — it's redundant.
+            _placeholder = "(The user sent a message with no text content)"
+            if user_text and user_text.strip() == _placeholder:
+                return prefix
            if user_text:
                return f"{prefix}\n\n{user_text}"
            return prefix
@@ -6272,7 +6308,15 @@ class GatewayRunner:
        # Falls back to env vars for backward compatibility.
        # YAML 1.1 parses bare `off` as boolean False — normalise before
        # the `or` chain so it doesn't silently fall through to "all".
-        _raw_tp = user_config.get("display", {}).get("tool_progress")
+        #
+        # Per-platform overrides (display.tool_progress_overrides) take
+        # priority over the global setting — e.g. Signal users can set
+        # tool_progress to "off" while keeping Telegram on "all".
+        _display_cfg = user_config.get("display", {})
+        _overrides = _display_cfg.get("tool_progress_overrides", {})
+        _raw_tp = _overrides.get(platform_key)
+        if _raw_tp is None:
+            _raw_tp = _display_cfg.get("tool_progress")
        if _raw_tp is False:
            _raw_tp = "off"
        progress_mode = (
@@ -6327,10 +6371,15 @@ class GatewayRunner:
                progress_queue.put(msg)
                return
            
-            # "all" / "new" modes: short preview, always truncated (40 chars)
+            # "all" / "new" modes: short preview, respects tool_preview_length
+            # config (defaults to 40 chars when unset to keep gateway messages
+            # compact — unlike CLI spinners, these persist as permanent messages).
            if preview:
-                if len(preview) > 40:
-                    preview = preview[:37] + "..."
+                from agent.display import get_tool_preview_max_len
+                _pl = get_tool_preview_max_len()
+                _cap = _pl if _pl > 0 else 40
+                if len(preview) > _cap:
+                    preview = preview[:_cap - 3] + "..."
                msg = f"{emoji} {tool_name}: \"{preview}\""
            else:
                msg = f"{emoji} {tool_name}..."
@@ -74,6 +74,8 @@ class GatewayStreamConsumer:
        self._edit_supported = True  # Disabled on first edit failure (Signal/Email/HA)
        self._last_edit_time = 0.0
        self._last_sent_text = ""   # Track last-sent text to skip redundant edits
+        self._fallback_final_send = False
+        self._fallback_prefix = ""

    @property
    def already_sent(self) -> bool:
@@ -138,12 +140,19 @@ class GatewayStreamConsumer:
                    while (
                        len(self._accumulated) > _safe_limit
                        and self._message_id is not None
+                        and self._edit_supported
                    ):
                        split_at = self._accumulated.rfind("\n", 0, _safe_limit)
                        if split_at < _safe_limit // 2:
                            split_at = _safe_limit
                        chunk = self._accumulated[:split_at]
                        await self._send_or_edit(chunk)
+                        if self._fallback_final_send:
+                            # Edit failed while attempting to split an oversized
+                            # message. Keep the full accumulated text intact so
+                            # the fallback final-send path can deliver the
+                            # remaining continuation without dropping content.
+                            break
                        self._accumulated = self._accumulated[split_at:].lstrip("\n")
                        self._message_id = None
                        self._last_sent_text = ""
@@ -156,9 +165,17 @@ class GatewayStreamConsumer:
                    self._last_edit_time = time.monotonic()

                if got_done:
-                    # Final edit without cursor
-                    if self._accumulated and self._message_id:
-                        await self._send_or_edit(self._accumulated)
+                    # Final edit without cursor. If progressive editing failed
+                    # mid-stream, send a single continuation/fallback message
+                    # here instead of letting the base gateway path send the
+                    # full response again.
+                    if self._accumulated:
+                        if self._fallback_final_send:
+                            await self._send_fallback_final(self._accumulated)
+                        elif self._message_id:
+                            await self._send_or_edit(self._accumulated)
+                        elif not self._already_sent:
+                            await self._send_or_edit(self._accumulated)
                    return

                # Tool boundary: the should_edit block above already flushed
@@ -169,6 +186,8 @@ class GatewayStreamConsumer:
                    self._message_id = None
                    self._accumulated = ""
                    self._last_sent_text = ""
+                    self._fallback_final_send = False
+                    self._fallback_prefix = ""

                await asyncio.sleep(0.05)  # Small yield to not busy-loop

@@ -207,6 +226,86 @@ class GatewayStreamConsumer:
        # Strip trailing whitespace/newlines but preserve leading content
        return cleaned.rstrip()

+    def _visible_prefix(self) -> str:
+        """Return the visible text already shown in the streamed message."""
+        prefix = self._last_sent_text or ""
+        if self.cfg.cursor and prefix.endswith(self.cfg.cursor):
+            prefix = prefix[:-len(self.cfg.cursor)]
+        return self._clean_for_display(prefix)
+
+    def _continuation_text(self, final_text: str) -> str:
+        """Return only the part of final_text the user has not already seen."""
+        prefix = self._fallback_prefix or self._visible_prefix()
+        if prefix and final_text.startswith(prefix):
+            return final_text[len(prefix):].lstrip()
+        return final_text
+
+    @staticmethod
+    def _split_text_chunks(text: str, limit: int) -> list[str]:
+        """Split text into reasonably sized chunks for fallback sends."""
+        if len(text) <= limit:
+            return [text]
+        chunks: list[str] = []
+        remaining = text
+        while len(remaining) > limit:
+            split_at = remaining.rfind("\n", 0, limit)
+            if split_at < limit // 2:
+                split_at = limit
+            chunks.append(remaining[:split_at])
+            remaining = remaining[split_at:].lstrip("\n")
+        if remaining:
+            chunks.append(remaining)
+        return chunks
+
+    async def _send_fallback_final(self, text: str) -> None:
+        """Send the final continuation after streaming edits stop working."""
+        final_text = self._clean_for_display(text)
+        continuation = self._continuation_text(final_text)
+        self._fallback_final_send = False
+        if not continuation.strip():
+            # Nothing new to send — the visible partial already matches final text.
+            self._already_sent = True
+            return
+
+        raw_limit = getattr(self.adapter, "MAX_MESSAGE_LENGTH", 4096)
+        safe_limit = max(500, raw_limit - 100)
+        chunks = self._split_text_chunks(continuation, safe_limit)
+
+        last_message_id: Optional[str] = None
+        last_successful_chunk = ""
+        sent_any_chunk = False
+        for chunk in chunks:
+            result = await self.adapter.send(
+                chat_id=self.chat_id,
+                content=chunk,
+                metadata=self.metadata,
+            )
+            if not result.success:
+                if sent_any_chunk:
+                    # Some continuation text already reached the user. Suppress
+                    # the base gateway final-send path so we don't resend the
+                    # full response and create another duplicate.
+                    self._already_sent = True
+                    self._message_id = last_message_id
+                    self._last_sent_text = last_successful_chunk
+                    self._fallback_prefix = ""
+                    return
+                # No fallback chunk reached the user — allow the normal gateway
+                # final-send path to try one more time.
+                self._already_sent = False
+                self._message_id = None
+                self._last_sent_text = ""
+                self._fallback_prefix = ""
+                return
+            sent_any_chunk = True
+            last_successful_chunk = chunk
+            last_message_id = result.message_id or last_message_id
+
+        self._message_id = last_message_id
+        self._already_sent = True
+        self._last_sent_text = chunks[-1]
+        self._fallback_prefix = ""
+
    async def _send_or_edit(self, text: str) -> None:
        """Send or edit the streaming message."""
        # Strip MEDIA: directives so they don't appear as visible text.
@@ -232,14 +331,16 @@ class GatewayStreamConsumer:
                        self._last_sent_text = text
                    else:
                        # If an edit fails mid-stream (especially Telegram flood control),
-                        # stop progressive edits and let the normal final send path deliver
-                        # the complete answer instead of leaving the user with a partial.
+                        # stop progressive edits and send only the missing tail once the
+                        # final response is available.
                        logger.debug("Edit failed, disabling streaming for this adapter")
+                        self._fallback_prefix = self._visible_prefix()
+                        self._fallback_final_send = True
                        self._edit_supported = False
-                        self._already_sent = False
+                        self._already_sent = True
                else:
                    # Editing not supported — skip intermediate updates.
-                    # The final response will be sent by the normal path.
+                    # The final response will be sent by the fallback path.
                    pass
            else:
                # First message — send new
@@ -252,6 +353,17 @@ class GatewayStreamConsumer:
                    self._message_id = result.message_id
                    self._already_sent = True
                    self._last_sent_text = text
+                elif result.success:
+                    # Platform accepted the message but returned no message_id
+                    # (e.g. Signal).  Can't edit without an ID — switch to
+                    # fallback mode: suppress intermediate deltas, send only
+                    # the missing tail once the final response is ready.
+                    self._already_sent = True
+                    self._edit_supported = False
+                    self._fallback_prefix = self._clean_for_display(text)
+                    self._fallback_final_send = True
+                    # Sentinel prevents re-entering this branch on every delta
+                    self._message_id = "__no_edit__"
                else:
                    # Initial send failed — disable streaming for this session
                    self._edit_supported = False
@@ -11,5 +11,5 @@ Provides subcommands for:
 - hermes cron          - Manage cron jobs
 """

-__version__ = "0.7.0"
-__release_date__ = "2026.4.3"
+__version__ = "0.8.0"
+__release_date__ = "2026.4.8"
@@ -37,7 +37,7 @@ from typing import Any, Dict, List, Optional
 import httpx
 import yaml

-from hermes_cli.config import get_hermes_home, get_config_path
+from hermes_cli.config import get_hermes_home, get_config_path, read_raw_config
 from hermes_constants import OPENROUTER_BASE_URL

 logger = logging.getLogger(__name__)
@@ -67,12 +67,16 @@ DEFAULT_AGENT_KEY_MIN_TTL_SECONDS = 30 * 60  # 30 minutes
 ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 120       # refresh 2 min before expiry
 DEVICE_AUTH_POLL_INTERVAL_CAP_SECONDS = 1     # poll at most every 1s
 DEFAULT_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
+DEFAULT_QWEN_BASE_URL = "https://portal.qwen.ai/v1"
 DEFAULT_GITHUB_MODELS_BASE_URL = "https://api.githubcopilot.com"
 DEFAULT_COPILOT_ACP_BASE_URL = "acp://copilot"
 DEFAULT_GEMINI_BASE_URL = "https://generativelanguage.googleapis.com/v1beta/openai"
 CODEX_OAUTH_CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann"
 CODEX_OAUTH_TOKEN_URL = "https://auth.openai.com/oauth/token"
 CODEX_ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 120
+QWEN_OAUTH_CLIENT_ID = "f0304373b74a44d2b584a3fb70ca9e56"
+QWEN_OAUTH_TOKEN_URL = "https://chat.qwen.ai/api/v1/oauth2/token"
+QWEN_ACCESS_TOKEN_REFRESH_SKEW_SECONDS = 120


 # =============================================================================
@@ -112,6 +116,12 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
        auth_type="oauth_external",
        inference_base_url=DEFAULT_CODEX_BASE_URL,
    ),
+    "qwen-oauth": ProviderConfig(
+        id="qwen-oauth",
+        name="Qwen OAuth",
+        auth_type="oauth_external",
+        inference_base_url=DEFAULT_QWEN_BASE_URL,
+    ),
    "copilot": ProviderConfig(
        id="copilot",
        name="GitHub Copilot",
@@ -817,6 +827,7 @@ def resolve_provider(
        "github-copilot-acp": "copilot-acp", "copilot-acp-agent": "copilot-acp",
        "aigateway": "ai-gateway", "vercel": "ai-gateway", "vercel-ai-gateway": "ai-gateway",
        "opencode": "opencode-zen", "zen": "opencode-zen",
+        "qwen-portal": "qwen-oauth", "qwen-cli": "qwen-oauth", "qwen-oauth": "qwen-oauth",
        "hf": "huggingface", "hugging-face": "huggingface", "huggingface-hub": "huggingface",
        "go": "opencode-go", "opencode-go-sub": "opencode-go",
        "kilo": "kilocode", "kilo-code": "kilocode", "kilo-gateway": "kilocode",
@@ -946,6 +957,176 @@ def _codex_access_token_is_expiring(access_token: Any, skew_seconds: int) -> boo
    return float(exp) <= (time.time() + max(0, int(skew_seconds)))


+def _qwen_cli_auth_path() -> Path:
+    return Path.home() / ".qwen" / "oauth_creds.json"
+
+
+def _read_qwen_cli_tokens() -> Dict[str, Any]:
+    auth_path = _qwen_cli_auth_path()
+    if not auth_path.exists():
+        raise AuthError(
+            "Qwen CLI credentials not found. Run 'qwen auth qwen-oauth' first.",
+            provider="qwen-oauth",
+            code="qwen_auth_missing",
+        )
+    try:
+        data = json.loads(auth_path.read_text(encoding="utf-8"))
+    except Exception as exc:
+        raise AuthError(
+            f"Failed to read Qwen CLI credentials from {auth_path}: {exc}",
+            provider="qwen-oauth",
+            code="qwen_auth_read_failed",
+        ) from exc
+    if not isinstance(data, dict):
+        raise AuthError(
+            f"Invalid Qwen CLI credentials in {auth_path}.",
+            provider="qwen-oauth",
+            code="qwen_auth_invalid",
+        )
+    return data
+
+
+def _save_qwen_cli_tokens(tokens: Dict[str, Any]) -> Path:
+    auth_path = _qwen_cli_auth_path()
+    auth_path.parent.mkdir(parents=True, exist_ok=True)
+    tmp_path = auth_path.with_suffix(".tmp")
+    tmp_path.write_text(json.dumps(tokens, indent=2, sort_keys=True) + "\n", encoding="utf-8")
+    os.chmod(tmp_path, stat.S_IRUSR | stat.S_IWUSR)
+    tmp_path.replace(auth_path)
+    return auth_path
+
+
+def _qwen_access_token_is_expiring(expiry_date_ms: Any, skew_seconds: int = QWEN_ACCESS_TOKEN_REFRESH_SKEW_SECONDS) -> bool:
+    try:
+        expiry_ms = int(expiry_date_ms)
+    except Exception:
+        return True
+    return (time.time() + max(0, int(skew_seconds))) * 1000 >= expiry_ms
+
+
+def _refresh_qwen_cli_tokens(tokens: Dict[str, Any], timeout_seconds: float = 20.0) -> Dict[str, Any]:
+    refresh_token = str(tokens.get("refresh_token", "") or "").strip()
+    if not refresh_token:
+        raise AuthError(
+            "Qwen OAuth refresh token missing. Re-run 'qwen auth qwen-oauth'.",
+            provider="qwen-oauth",
+            code="qwen_refresh_token_missing",
+        )
+
+    try:
+        response = httpx.post(
+            QWEN_OAUTH_TOKEN_URL,
+            headers={
+                "Content-Type": "application/x-www-form-urlencoded",
+                "Accept": "application/json",
+            },
+            data={
+                "grant_type": "refresh_token",
+                "refresh_token": refresh_token,
+                "client_id": QWEN_OAUTH_CLIENT_ID,
+            },
+            timeout=timeout_seconds,
+        )
+    except Exception as exc:
+        raise AuthError(
+            f"Qwen OAuth refresh failed: {exc}",
+            provider="qwen-oauth",
+            code="qwen_refresh_failed",
+        ) from exc
+
+    if response.status_code >= 400:
+        body = response.text.strip()
+        raise AuthError(
+            "Qwen OAuth refresh failed. Re-run 'qwen auth qwen-oauth'."
+            + (f" Response: {body}" if body else ""),
+            provider="qwen-oauth",
+            code="qwen_refresh_failed",
+        )
+
+    try:
+        payload = response.json()
+    except Exception as exc:
+        raise AuthError(
+            f"Qwen OAuth refresh returned invalid JSON: {exc}",
+            provider="qwen-oauth",
+            code="qwen_refresh_invalid_json",
+        ) from exc
+
+    if not isinstance(payload, dict) or not str(payload.get("access_token", "") or "").strip():
+        raise AuthError(
+            "Qwen OAuth refresh response missing access_token.",
+            provider="qwen-oauth",
+            code="qwen_refresh_invalid_response",
+        )
+
+    expires_in = payload.get("expires_in")
+    try:
+        expires_in_seconds = int(expires_in)
+    except Exception:
+        expires_in_seconds = 6 * 60 * 60
+
+    refreshed = {
+        "access_token": str(payload.get("access_token", "") or "").strip(),
+        "refresh_token": str(payload.get("refresh_token", refresh_token) or refresh_token).strip(),
+        "token_type": str(payload.get("token_type", tokens.get("token_type", "Bearer")) or "Bearer").strip() or "Bearer",
+        "resource_url": str(payload.get("resource_url", tokens.get("resource_url", "portal.qwen.ai")) or "portal.qwen.ai").strip(),
+        "expiry_date": int(time.time() * 1000) + max(1, expires_in_seconds) * 1000,
+    }
+    _save_qwen_cli_tokens(refreshed)
+    return refreshed
+
+
+def resolve_qwen_runtime_credentials(
+    *,
+    force_refresh: bool = False,
+    refresh_if_expiring: bool = True,
+    refresh_skew_seconds: int = QWEN_ACCESS_TOKEN_REFRESH_SKEW_SECONDS,
+) -> Dict[str, Any]:
+    tokens = _read_qwen_cli_tokens()
+    access_token = str(tokens.get("access_token", "") or "").strip()
+    should_refresh = bool(force_refresh)
+    if not should_refresh and refresh_if_expiring:
+        should_refresh = _qwen_access_token_is_expiring(tokens.get("expiry_date"), refresh_skew_seconds)
+    if should_refresh:
+        tokens = _refresh_qwen_cli_tokens(tokens)
+        access_token = str(tokens.get("access_token", "") or "").strip()
+    if not access_token:
+        raise AuthError(
+            "Qwen OAuth access token missing. Re-run 'qwen auth qwen-oauth'.",
+            provider="qwen-oauth",
+            code="qwen_access_token_missing",
+        )
+
+    base_url = os.getenv("HERMES_QWEN_BASE_URL", "").strip().rstrip("/") or DEFAULT_QWEN_BASE_URL
+    return {
+        "provider": "qwen-oauth",
+        "base_url": base_url,
+        "api_key": access_token,
+        "source": "qwen-cli",
+        "expires_at_ms": tokens.get("expiry_date"),
+        "auth_file": str(_qwen_cli_auth_path()),
+    }
+
+
+def get_qwen_auth_status() -> Dict[str, Any]:
+    auth_path = _qwen_cli_auth_path()
+    try:
+        creds = resolve_qwen_runtime_credentials(refresh_if_expiring=False)
+        return {
+            "logged_in": True,
+            "auth_file": str(auth_path),
+            "source": creds.get("source"),
+            "api_key": creds.get("api_key"),
+            "expires_at_ms": creds.get("expires_at_ms"),
+        }
+    except AuthError as exc:
+        return {
+            "logged_in": False,
+            "auth_file": str(auth_path),
+            "error": str(exc),
+        }
+
+
 # =============================================================================
 # SSH / remote session detection
 # =============================================================================
@@ -2072,6 +2253,8 @@ def get_auth_status(provider_id: Optional[str] = None) -> Dict[str, Any]:
        return get_nous_auth_status()
    if target == "openai-codex":
        return get_codex_auth_status()
+    if target == "qwen-oauth":
+        return get_qwen_auth_status()
    if target == "copilot-acp":
        return get_external_process_provider_status(target)
    # API-key providers
@@ -2214,14 +2397,7 @@ def _update_config_for_provider(
    config_path = get_config_path()
    config_path.parent.mkdir(parents=True, exist_ok=True)

-    config: Dict[str, Any] = {}
-    if config_path.exists():
-        try:
-            loaded = yaml.safe_load(config_path.read_text()) or {}
-            if isinstance(loaded, dict):
-                config = loaded
-        except Exception:
-            config = {}
+    config = read_raw_config()

    current_model = config.get("model")
    if isinstance(current_model, dict):
@@ -2258,12 +2434,8 @@ def _reset_config_provider() -> Path:
    if not config_path.exists():
        return config_path

-    try:
-        config = yaml.safe_load(config_path.read_text()) or {}
-    except Exception:
-        return config_path
-
-    if not isinstance(config, dict):
+    config = read_raw_config()
+    if not config:
        return config_path

    model = config.get("model")
@@ -32,7 +32,7 @@ from hermes_constants import OPENROUTER_BASE_URL


 # Providers that support OAuth login in addition to API keys.
-_OAUTH_CAPABLE_PROVIDERS = {"anthropic", "nous", "openai-codex"}
+_OAUTH_CAPABLE_PROVIDERS = {"anthropic", "nous", "openai-codex", "qwen-oauth"}


 def _get_custom_provider_names() -> list:
@@ -147,7 +147,7 @@ def auth_add_command(args) -> None:
        if provider.startswith(CUSTOM_POOL_PREFIX):
            requested_type = AUTH_TYPE_API_KEY
        else:
-            requested_type = AUTH_TYPE_OAUTH if provider in {"anthropic", "nous", "openai-codex"} else AUTH_TYPE_API_KEY
+            requested_type = AUTH_TYPE_OAUTH if provider in {"anthropic", "nous", "openai-codex", "qwen-oauth"} else AUTH_TYPE_API_KEY

    pool = load_pool(provider)

@@ -250,6 +250,26 @@ def auth_add_command(args) -> None:
        print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
        return

+    if provider == "qwen-oauth":
+        creds = auth_mod.resolve_qwen_runtime_credentials(refresh_if_expiring=False)
+        label = (getattr(args, "label", None) or "").strip() or label_from_token(
+            creds["api_key"],
+            _oauth_default_label(provider, len(pool.entries()) + 1),
+        )
+        entry = PooledCredential(
+            provider=provider,
+            id=uuid.uuid4().hex[:6],
+            label=label,
+            auth_type=AUTH_TYPE_OAUTH,
+            priority=0,
+            source=f"{SOURCE_MANUAL}:qwen_cli",
+            access_token=creds["api_key"],
+            base_url=creds.get("base_url"),
+        )
+        pool.add_entry(entry)
+        print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
+        return
+
    raise SystemExit(f"`hermes auth add {provider}` is not implemented for auth type {requested_type} yet.")


@@ -5,6 +5,7 @@ Pure display functions with no HermesCLI state dependency.

 import json
 import logging
+import os
 import shutil
 import subprocess
 import threading
@@ -189,6 +190,79 @@ def check_for_updates() -> Optional[int]:
    return behind


+def _resolve_repo_dir() -> Optional[Path]:
+    """Return the active Hermes git checkout, or None if this isn't a git install."""
+    hermes_home = get_hermes_home()
+    repo_dir = hermes_home / "hermes-agent"
+    if not (repo_dir / ".git").exists():
+        repo_dir = Path(__file__).parent.parent.resolve()
+    return repo_dir if (repo_dir / ".git").exists() else None
+
+
+def _git_short_hash(repo_dir: Path, rev: str) -> Optional[str]:
+    """Resolve a git revision to an 8-character short hash."""
+    try:
+        result = subprocess.run(
+            ["git", "rev-parse", "--short=8", rev],
+            capture_output=True,
+            text=True,
+            timeout=5,
+            cwd=str(repo_dir),
+        )
+    except Exception:
+        return None
+    if result.returncode != 0:
+        return None
+    value = (result.stdout or "").strip()
+    return value or None
+
+
+def get_git_banner_state(repo_dir: Optional[Path] = None) -> Optional[dict]:
+    """Return upstream/local git hashes for the startup banner."""
+    repo_dir = repo_dir or _resolve_repo_dir()
+    if repo_dir is None:
+        return None
+
+    upstream = _git_short_hash(repo_dir, "origin/main")
+    local = _git_short_hash(repo_dir, "HEAD")
+    if not upstream or not local:
+        return None
+
+    ahead = 0
+    try:
+        result = subprocess.run(
+            ["git", "rev-list", "--count", "origin/main..HEAD"],
+            capture_output=True,
+            text=True,
+            timeout=5,
+            cwd=str(repo_dir),
+        )
+        if result.returncode == 0:
+            ahead = int((result.stdout or "0").strip() or "0")
+    except Exception:
+        ahead = 0
+
+    return {"upstream": upstream, "local": local, "ahead": max(ahead, 0)}
+
+
+def format_banner_version_label() -> str:
+    """Return the version label shown in the startup banner title."""
+    base = f"Hermes Agent v{VERSION} ({RELEASE_DATE})"
+    state = get_git_banner_state()
+    if not state:
+        return base
+
+    upstream = state["upstream"]
+    local = state["local"]
+    ahead = int(state.get("ahead") or 0)
+
+    if ahead <= 0 or upstream == local:
+        return f"{base} · upstream {upstream}"
+
+    carried_word = "commit" if ahead == 1 else "commits"
+    return f"{base} · upstream {upstream} · local {local} (+{ahead} carried {carried_word})"
+
+
 # =========================================================================
 # Non-blocking update check
 # =========================================================================
@@ -448,7 +522,7 @@ def build_welcome_banner(console: Console, model: str, cwd: str,
    border_color = _skin_color("banner_border", "#CD7F32")
    outer_panel = Panel(
        layout_table,
-        title=f"[bold {title_color}]{agent_name} v{VERSION} ({RELEASE_DATE})[/]",
+        title=f"[bold {title_color}]{format_banner_version_label()}[/]",
        border_style=border_color,
        padding=(0, 2),
    )
@@ -293,14 +293,8 @@ def _resolve_config_gates() -> set[str]:
    if not gated:
        return set()
    try:
-        import yaml
-        from hermes_constants import get_hermes_home
-        config_path = str(get_hermes_home() / "config.yaml")
-        if os.path.exists(config_path):
-            with open(config_path, encoding="utf-8") as f:
-                cfg = yaml.safe_load(f) or {}
-        else:
-            cfg = {}
+        from hermes_cli.config import read_raw_config
+        cfg = read_raw_config()
    except Exception:
        return set()
    result: set[str] = set()
@@ -157,7 +157,14 @@ def get_project_root() -> Path:
    return Path(__file__).parent.parent.resolve()

 def _secure_dir(path):
-    """Set directory to owner-only access (0700). No-op on Windows."""
+    """Set directory to owner-only access (0700). No-op on Windows.
+
+    Skipped in managed mode — the NixOS module sets group-readable
+    permissions (0750) so interactive users in the hermes group can
+    share state with the gateway service.
+    """
+    if is_managed():
+        return
    try:
        os.chmod(path, 0o700)
    except (OSError, NotImplementedError):
@@ -165,7 +172,13 @@ def _secure_dir(path):


 def _secure_file(path):
-    """Set file to owner-only read/write (0600). No-op on Windows."""
+    """Set file to owner-only read/write (0600). No-op on Windows.
+
+    Skipped in managed mode — the NixOS activation script sets
+    group-readable permissions (0640) on config files.
+    """
+    if is_managed():
+        return
    try:
        if os.path.exists(str(path)):
            os.chmod(path, 0o600)
@@ -379,6 +392,7 @@ DEFAULT_CONFIG = {
        "show_cost": False,       # Show $ cost in the status bar (off by default)
        "skin": "default",
        "tool_progress_command": False,  # Enable /verbose command in messaging gateway
+        "tool_progress_overrides": {},  # Per-platform overrides: {"signal": "off", "telegram": "all"}
        "tool_preview_length": 0,  # Max chars for tool call previews (0 = no limit, show full paths/commands)
    },

@@ -413,13 +427,17 @@ DEFAULT_CONFIG = {
    
    "stt": {
        "enabled": True,
-        "provider": "local",  # "local" (free, faster-whisper) | "groq" | "openai" (Whisper API)
+        "provider": "local",  # "local" (free, faster-whisper) | "groq" | "openai" (Whisper API) | "mistral" (Voxtral Transcribe)
        "local": {
            "model": "base",  # tiny, base, small, medium, large-v3
+            "language": "",  # auto-detect by default; set to "en", "es", "fr", etc. to force
        },
        "openai": {
            "model": "whisper-1",  # whisper-1, gpt-4o-mini-transcribe, gpt-4o-transcribe
        },
+        "mistral": {
+            "model": "voxtral-mini-latest",  # voxtral-mini-latest, voxtral-mini-2602
+        },
    },

    "voice": {
@@ -723,6 +741,14 @@ OPTIONAL_ENV_VARS = {
        "category": "provider",
        "advanced": True,
    },
+    "HERMES_QWEN_BASE_URL": {
+        "description": "Qwen Portal base URL override (default: https://portal.qwen.ai/v1)",
+        "prompt": "Qwen Portal base URL (leave empty for default)",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
    "OPENCODE_ZEN_API_KEY": {
        "description": "OpenCode Zen API key (pay-as-you-go access to curated models)",
        "prompt": "OpenCode Zen API key",
@@ -974,6 +1000,13 @@ OPTIONAL_ENV_VARS = {
        "password": False,
        "category": "messaging",
    },
+    "DISCORD_REPLY_TO_MODE": {
+        "description": "Discord reply threading mode: 'off' (no reply references), 'first' (reply on first message only, default), 'all' (reply on every chunk)",
+        "prompt": "Discord reply mode (off/first/all)",
+        "url": None,
+        "password": False,
+        "category": "messaging",
+    },
    "SLACK_BOT_TOKEN": {
        "description": "Slack bot token (xoxb-). Get from OAuth & Permissions after installing your app. "
                       "Required scopes: chat:write, app_mentions:read, channels:history, groups:history, "
@@ -93,6 +93,21 @@ def cron_list(show_all: bool = False):
        script = job.get("script")
        if script:
            print(f"    Script:    {script}")
+
+        # Execution history
+        last_status = job.get("last_status")
+        if last_status:
+            last_run = job.get("last_run_at", "?")
+            if last_status == "ok":
+                status_display = color("ok", Colors.GREEN)
+            else:
+                status_display = color(f"{last_status}: {job.get('last_error', '?')}", Colors.RED)
+            print(f"    Last run:  {last_run}  {status_display}")
+
+        delivery_err = job.get("last_delivery_error")
+        if delivery_err:
+            print(f"    {color('⚠ Delivery failed:', Colors.YELLOW)} {delivery_err}")
+
        print()

    from hermes_cli.gateway import find_gateway_pids
@@ -812,69 +812,83 @@ def run_doctor(args):
        check_warn("No GITHUB_TOKEN", f"(60 req/hr rate limit — set in {_DHH}/.env for better rates)")

    # =========================================================================
-    # Honcho memory
+    # Memory Provider (only check the active provider, if any)
    # =========================================================================
    print()
-    print(color("◆ Honcho Memory", Colors.CYAN, Colors.BOLD))
+    print(color("◆ Memory Provider", Colors.CYAN, Colors.BOLD))

+    _active_memory_provider = ""
    try:
-        from plugins.memory.honcho.client import HonchoClientConfig, resolve_config_path
-        hcfg = HonchoClientConfig.from_global_config()
-        _honcho_cfg_path = resolve_config_path()
+        import yaml as _yaml
+        _mem_cfg_path = HERMES_HOME / "config.yaml"
+        if _mem_cfg_path.exists():
+            with open(_mem_cfg_path) as _f:
+                _raw_cfg = _yaml.safe_load(_f) or {}
+            _active_memory_provider = (_raw_cfg.get("memory") or {}).get("provider", "")
+    except Exception:
+        pass

-        if not _honcho_cfg_path.exists():
-            check_warn("Honcho config not found", "run: hermes memory setup")
-        elif not hcfg.enabled:
-            check_info(f"Honcho disabled (set enabled: true in {_honcho_cfg_path} to activate)")
-        elif not (hcfg.api_key or hcfg.base_url):
-            check_fail("Honcho API key or base URL not set", "run: hermes memory setup")
-            issues.append("No Honcho API key — run 'hermes memory setup'")
-        else:
-            from plugins.memory.honcho.client import get_honcho_client, reset_honcho_client
-            reset_honcho_client()
-            try:
-                get_honcho_client(hcfg)
-                check_ok(
-                    "Honcho connected",
-                    f"workspace={hcfg.workspace_id} mode={hcfg.recall_mode} freq={hcfg.write_frequency}",
-                )
-            except Exception as _e:
-                check_fail("Honcho connection failed", str(_e))
-                issues.append(f"Honcho unreachable: {_e}")
-    except ImportError:
-        check_warn("honcho-ai not installed", "pip install honcho-ai")
-    except Exception as _e:
-        check_warn("Honcho check failed", str(_e))
+    if not _active_memory_provider:
+        check_ok("Built-in memory active", "(no external provider configured — this is fine)")
+    elif _active_memory_provider == "honcho":
+        try:
+            from plugins.memory.honcho.client import HonchoClientConfig, resolve_config_path
+            hcfg = HonchoClientConfig.from_global_config()
+            _honcho_cfg_path = resolve_config_path()

-    # =========================================================================
-    # Mem0 memory
-    # =========================================================================
-    print()
-    print(color("◆ Mem0 Memory", Colors.CYAN, Colors.BOLD))
-
-    try:
-        from plugins.memory.mem0 import _load_config as _load_mem0_config
-        mem0_cfg = _load_mem0_config()
-        mem0_key = mem0_cfg.get("api_key", "")
-        if mem0_key:
-            check_ok("Mem0 API key configured")
-            check_info(f"user_id={mem0_cfg.get('user_id', '?')}  agent_id={mem0_cfg.get('agent_id', '?')}")
-            # Check if mem0.json exists but is missing api_key (the bug we fixed)
-            mem0_json = HERMES_HOME / "mem0.json"
-            if mem0_json.exists():
+            if not _honcho_cfg_path.exists():
+                check_warn("Honcho config not found", "run: hermes memory setup")
+            elif not hcfg.enabled:
+                check_info(f"Honcho disabled (set enabled: true in {_honcho_cfg_path} to activate)")
+            elif not (hcfg.api_key or hcfg.base_url):
+                check_fail("Honcho API key or base URL not set", "run: hermes memory setup")
+                issues.append("No Honcho API key — run 'hermes memory setup'")
+            else:
+                from plugins.memory.honcho.client import get_honcho_client, reset_honcho_client
+                reset_honcho_client()
                try:
-                    import json as _json
-                    file_cfg = _json.loads(mem0_json.read_text())
-                    if not file_cfg.get("api_key") and mem0_key:
-                        check_info("api_key from .env (not in mem0.json) — this is fine")
-                except Exception:
-                    pass
-        else:
-            check_warn("Mem0 not configured", "(set MEM0_API_KEY in .env or run hermes memory setup)")
-    except ImportError:
-        check_warn("Mem0 plugin not loadable", "(optional)")
-    except Exception as _e:
-        check_warn("Mem0 check failed", str(_e))
+                    get_honcho_client(hcfg)
+                    check_ok(
+                        "Honcho connected",
+                        f"workspace={hcfg.workspace_id} mode={hcfg.recall_mode} freq={hcfg.write_frequency}",
+                    )
+                except Exception as _e:
+                    check_fail("Honcho connection failed", str(_e))
+                    issues.append(f"Honcho unreachable: {_e}")
+        except ImportError:
+            check_fail("honcho-ai not installed", "pip install honcho-ai")
+            issues.append("Honcho is set as memory provider but honcho-ai is not installed")
+        except Exception as _e:
+            check_warn("Honcho check failed", str(_e))
+    elif _active_memory_provider == "mem0":
+        try:
+            from plugins.memory.mem0 import _load_config as _load_mem0_config
+            mem0_cfg = _load_mem0_config()
+            mem0_key = mem0_cfg.get("api_key", "")
+            if mem0_key:
+                check_ok("Mem0 API key configured")
+                check_info(f"user_id={mem0_cfg.get('user_id', '?')}  agent_id={mem0_cfg.get('agent_id', '?')}")
+            else:
+                check_fail("Mem0 API key not set", "(set MEM0_API_KEY in .env or run hermes memory setup)")
+                issues.append("Mem0 is set as memory provider but API key is missing")
+        except ImportError:
+            check_fail("Mem0 plugin not loadable", "pip install mem0ai")
+            issues.append("Mem0 is set as memory provider but mem0ai is not installed")
+        except Exception as _e:
+            check_warn("Mem0 check failed", str(_e))
+    else:
+        # Generic check for other memory providers (openviking, hindsight, etc.)
+        try:
+            from plugins.memory import load_memory_provider
+            _provider = load_memory_provider(_active_memory_provider)
+            if _provider and _provider.is_available():
+                check_ok(f"{_active_memory_provider} provider active")
+            elif _provider:
+                check_warn(f"{_active_memory_provider} configured but not available", "run: hermes memory status")
+            else:
+                check_warn(f"{_active_memory_provider} plugin not found", "run: hermes memory setup")
+        except Exception as _e:
+            check_warn(f"{_active_memory_provider} check failed", str(_e))

    # =========================================================================
    # Profiles
@@ -267,6 +267,34 @@ def _profile_suffix() -> str:
    return hashlib.sha256(str(home).encode()).hexdigest()[:8]


+def _profile_arg(hermes_home: str | None = None) -> str:
+    """Return ``--profile <name>`` only when HERMES_HOME is a named profile.
+
+    For ``~/.hermes/profiles/<name>``, returns ``"--profile <name>"``.
+    For the default profile or hash-based custom paths, returns the empty string.
+
+    Args:
+        hermes_home: Optional explicit HERMES_HOME path. Defaults to the current
+            ``get_hermes_home()`` value. Should be passed when generating a
+            service definition for a different user (e.g. system service).
+    """
+    import re
+    from pathlib import Path as _Path
+    home = Path(hermes_home or str(get_hermes_home())).resolve()
+    default = (_Path.home() / ".hermes").resolve()
+    if home == default:
+        return ""
+    profiles_root = (default / "profiles").resolve()
+    try:
+        rel = home.relative_to(profiles_root)
+        parts = rel.parts
+        if len(parts) == 1 and re.match(r"^[a-z0-9][a-z0-9_-]{0,63}$", parts[0]):
+            return f"--profile {parts[0]}"
+    except ValueError:
+        pass
+    return ""
+
+
 def get_service_name() -> str:
    """Derive a systemd service name scoped to this HERMES_HOME.

@@ -626,6 +654,7 @@ def generate_systemd_unit(system: bool = False, run_as_user: str | None = None)
    if system:
        username, group_name, home_dir = _system_service_identity(run_as_user)
        hermes_home = _hermes_home_for_target_user(home_dir)
+        profile_arg = _profile_arg(hermes_home)
        path_entries.extend(_build_user_local_paths(Path(home_dir), path_entries))
        path_entries.extend(common_bin_paths)
        sane_path = ":".join(path_entries)
@@ -640,7 +669,7 @@ StartLimitBurst=5
 Type=simple
 User={username}
 Group={group_name}
-ExecStart={python_path} -m hermes_cli.main gateway run --replace
+ExecStart={python_path} -m hermes_cli.main{f" {profile_arg}" if profile_arg else ""} gateway run --replace
 WorkingDirectory={working_dir}
 Environment="HOME={home_dir}"
 Environment="USER={username}"
@@ -661,6 +690,7 @@ WantedBy=multi-user.target
 """

    hermes_home = str(get_hermes_home().resolve())
+    profile_arg = _profile_arg(hermes_home)
    path_entries.extend(_build_user_local_paths(Path.home(), path_entries))
    path_entries.extend(common_bin_paths)
    sane_path = ":".join(path_entries)
@@ -672,7 +702,7 @@ StartLimitBurst=5

 [Service]
 Type=simple
-ExecStart={python_path} -m hermes_cli.main gateway run --replace
+ExecStart={python_path} -m hermes_cli.main{f" {profile_arg}" if profile_arg else ""} gateway run --replace
 WorkingDirectory={working_dir}
 Environment="PATH={sane_path}"
 Environment="VIRTUAL_ENV={venv_dir}"
@@ -965,6 +995,7 @@ def generate_launchd_plist() -> str:
    log_dir = get_hermes_home() / "logs"
    log_dir.mkdir(parents=True, exist_ok=True)
    label = get_launchd_label()
+    profile_arg = _profile_arg(hermes_home)
    # Build a sane PATH for the launchd plist.  launchd provides only a
    # minimal default (/usr/bin:/bin:/usr/sbin:/sbin) which misses Homebrew,
    # nvm, cargo, etc.  We prepend venv/bin and node_modules/.bin (matching
@@ -986,21 +1017,32 @@ def generate_launchd_plist() -> str:
        dict.fromkeys(priority_dirs + [p for p in os.environ.get("PATH", "").split(":") if p])
    )

+    # Build ProgramArguments array, including --profile when using a named profile
+    prog_args = [
+        f"<string>{python_path}</string>",
+        "<string>-m</string>",
+        "<string>hermes_cli.main</string>",
+    ]
+    if profile_arg:
+        for part in profile_arg.split():
+            prog_args.append(f"<string>{part}</string>")
+    prog_args.extend([
+        "<string>gateway</string>",
+        "<string>run</string>",
+        "<string>--replace</string>",
+    ])
+    prog_args_xml = "\n        ".join(prog_args)
+
    return f"""<?xml version="1.0" encoding="UTF-8"?>
 <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
 <plist version="1.0">
 <dict>
    <key>Label</key>
    <string>{label}</string>
-    
+
    <key>ProgramArguments</key>
    <array>
-        <string>{python_path}</string>
-        <string>-m</string>
-        <string>hermes_cli.main</string>
-        <string>gateway</string>
-        <string>run</string>
-        <string>--replace</string>
+        {prog_args_xml}
    </array>
    
    <key>WorkingDirectory</key>
@@ -918,6 +918,7 @@ def select_provider_and_model(args=None):
        "openrouter": "OpenRouter",
        "nous": "Nous Portal",
        "openai-codex": "OpenAI Codex",
+        "qwen-oauth": "Qwen OAuth",
        "copilot-acp": "GitHub Copilot ACP",
        "copilot": "GitHub Copilot",
        "anthropic": "Anthropic",
@@ -947,6 +948,7 @@ def select_provider_and_model(args=None):
        ("openrouter", "OpenRouter (100+ models, pay-per-use)"),
        ("anthropic", "Anthropic (Claude models — API key or Claude Code)"),
        ("openai-codex", "OpenAI Codex"),
+        ("qwen-oauth", "Qwen OAuth (reuses local Qwen CLI login)"),
        ("copilot", "GitHub Copilot (uses GITHUB_TOKEN or gh auth token)"),
        ("huggingface", "Hugging Face Inference Providers (20+ open models)"),
    ]
@@ -1043,6 +1045,8 @@ def select_provider_and_model(args=None):
        _model_flow_nous(config, current_model, args=args)
    elif selected_provider == "openai-codex":
        _model_flow_openai_codex(config, current_model)
+    elif selected_provider == "qwen-oauth":
+        _model_flow_qwen_oauth(config, current_model)
    elif selected_provider == "copilot-acp":
        _model_flow_copilot_acp(config, current_model)
    elif selected_provider == "copilot":
@@ -1359,6 +1363,56 @@ def _model_flow_openai_codex(config, current_model=""):



+_DEFAULT_QWEN_PORTAL_MODELS = [
+    "qwen3-coder-plus",
+    "qwen3-coder",
+]
+
+
+def _model_flow_qwen_oauth(_config, current_model=""):
+    """Qwen OAuth provider: reuse local Qwen CLI login, then pick model."""
+    from hermes_cli.auth import (
+        get_qwen_auth_status,
+        resolve_qwen_runtime_credentials,
+        _prompt_model_selection,
+        _save_model_choice,
+        _update_config_for_provider,
+        DEFAULT_QWEN_BASE_URL,
+    )
+    from hermes_cli.models import fetch_api_models
+
+    status = get_qwen_auth_status()
+    if not status.get("logged_in"):
+        print("Not logged into Qwen CLI OAuth.")
+        print("Run: qwen auth qwen-oauth")
+        auth_file = status.get("auth_file")
+        if auth_file:
+            print(f"Expected credentials file: {auth_file}")
+        if status.get("error"):
+            print(f"Error: {status.get('error')}")
+        return
+
+    # Try live model discovery, fall back to curated list.
+    models = None
+    try:
+        creds = resolve_qwen_runtime_credentials(refresh_if_expiring=True)
+        models = fetch_api_models(creds["api_key"], creds["base_url"])
+    except Exception:
+        pass
+    if not models:
+        models = list(_DEFAULT_QWEN_PORTAL_MODELS)
+
+    default = current_model or (models[0] if models else "qwen3-coder-plus")
+    selected = _prompt_model_selection(models, current_model=default)
+    if selected:
+        _save_model_choice(selected)
+        _update_config_for_provider("qwen-oauth", DEFAULT_QWEN_BASE_URL)
+        print(f"Default model set to: {selected} (via Qwen OAuth)")
+    else:
+        print("No change.")
+
+
+
 def _model_flow_custom(config):
    """Custom endpoint: collect URL, API key, and model name.

@@ -84,6 +84,7 @@ _PASSTHROUGH_PROVIDERS: frozenset[str] = frozenset({
    "minimax",
    "minimax-cn",
    "alibaba",
+    "qwen-oauth",
    "huggingface",
    "openai-codex",
    "custom",
@@ -791,12 +791,12 @@ def list_authenticated_providers(
        if overlay.auth_type in ("oauth_device_code", "oauth_external", "external_process"):
            # These use auth stores, not env vars — check for auth.json entries
            try:
-                from hermes_cli.auth import _read_auth_store
-                store = _read_auth_store()
-                if store and pid in store:
+                from hermes_cli.auth import _load_auth_store
+                store = _load_auth_store()
+                if store and (pid in store.get("providers", {}) or pid in store.get("credential_pool", {})):
                    has_creds = True
-            except Exception:
-                pass
+            except Exception as exc:
+                logger.debug("Auth store check failed for %s: %s", pid, exc)
        if not has_creds:
            continue

@@ -144,18 +144,22 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
        "kimi-k2-0905-preview",
    ],
    "minimax": [
-        "MiniMax-M2.7",
-        "MiniMax-M2.7-highspeed",
+        "MiniMax-M1",
+        "MiniMax-M1-40k",
+        "MiniMax-M1-80k",
+        "MiniMax-M1-128k",
+        "MiniMax-M1-256k",
        "MiniMax-M2.5",
-        "MiniMax-M2.5-highspeed",
-        "MiniMax-M2.1",
+        "MiniMax-M2.7",
    ],
    "minimax-cn": [
-        "MiniMax-M2.7",
-        "MiniMax-M2.7-highspeed",
+        "MiniMax-M1",
+        "MiniMax-M1-40k",
+        "MiniMax-M1-80k",
+        "MiniMax-M1-128k",
+        "MiniMax-M1-256k",
        "MiniMax-M2.5",
-        "MiniMax-M2.5-highspeed",
-        "MiniMax-M2.1",
+        "MiniMax-M2.7",
    ],
    "anthropic": [
        "claude-opus-4-6",
@@ -479,6 +483,7 @@ _PROVIDER_LABELS = {
    "ai-gateway": "AI Gateway",
    "kilocode": "Kilo Code",
    "alibaba": "Alibaba Cloud (DashScope)",
+    "qwen-oauth": "Qwen OAuth (Portal)",
    "huggingface": "Hugging Face",
    "custom": "Custom endpoint",
 }
@@ -518,6 +523,7 @@ _PROVIDER_ALIASES = {
    "aliyun": "alibaba",
    "qwen": "alibaba",
    "alibaba-cloud": "alibaba",
+    "qwen-portal": "qwen-oauth",
    "hf": "huggingface",
    "hugging-face": "huggingface",
    "huggingface-hub": "huggingface",
@@ -763,6 +769,7 @@ def list_available_providers() -> list[dict[str, str]]:
        "openrouter", "nous", "openai-codex", "copilot", "copilot-acp",
        "gemini", "huggingface",
        "zai", "kimi-coding", "minimax", "minimax-cn", "kilocode", "anthropic", "alibaba",
+        "qwen-oauth",
        "opencode-zen", "opencode-go",
        "ai-gateway", "deepseek", "custom",
    ]
@@ -61,6 +61,8 @@ VALID_HOOKS: Set[str] = {
    "post_api_request",
    "on_session_start",
    "on_session_end",
+    "on_session_finalize",
+    "on_session_reset",
 }

 ENTRY_POINTS_GROUP = "hermes_agent.plugins"
@@ -58,6 +58,12 @@ HERMES_OVERLAYS: Dict[str, HermesOverlay] = {
        auth_type="oauth_external",
        base_url_override="https://chatgpt.com/backend-api/codex",
    ),
+    "qwen-oauth": HermesOverlay(
+        transport="openai_chat",
+        auth_type="oauth_external",
+        base_url_override="https://portal.qwen.ai/v1",
+        base_url_env_var="HERMES_QWEN_BASE_URL",
+    ),
    "copilot-acp": HermesOverlay(
        transport="codex_responses",
        auth_type="external_process",
@@ -14,11 +14,13 @@ from agent.credential_pool import CredentialPool, PooledCredential, get_custom_p
 from hermes_cli.auth import (
    AuthError,
    DEFAULT_CODEX_BASE_URL,
+    DEFAULT_QWEN_BASE_URL,
    PROVIDER_REGISTRY,
    format_auth_error,
    resolve_provider,
    resolve_nous_runtime_credentials,
    resolve_codex_runtime_credentials,
+    resolve_qwen_runtime_credentials,
    resolve_api_key_provider_credentials,
    resolve_external_process_provider_credentials,
    has_usable_secret,
@@ -148,6 +150,9 @@ def _resolve_runtime_from_pool_entry(
    if provider == "openai-codex":
        api_mode = "codex_responses"
        base_url = base_url or DEFAULT_CODEX_BASE_URL
+    elif provider == "qwen-oauth":
+        api_mode = "chat_completions"
+        base_url = base_url or DEFAULT_QWEN_BASE_URL
    elif provider == "anthropic":
        api_mode = "anthropic_messages"
        cfg_provider = str(model_cfg.get("provider") or "").strip().lower()
@@ -163,6 +168,16 @@ def _resolve_runtime_from_pool_entry(
        api_mode = _copilot_runtime_api_mode(model_cfg, getattr(entry, "runtime_api_key", ""))
    else:
        configured_provider = str(model_cfg.get("provider") or "").strip().lower()
+        # Honour model.base_url from config.yaml when the configured provider
+        # matches this provider — same pattern as the Anthropic branch above.
+        # Only override when the pool entry has no explicit base_url (i.e. it
+        # fell back to the hardcoded default).  Env var overrides win (#6039).
+        pconfig = PROVIDER_REGISTRY.get(provider)
+        pool_url_is_default = pconfig and base_url.rstrip("/") == pconfig.inference_base_url.rstrip("/")
+        if configured_provider == provider and pool_url_is_default:
+            cfg_base_url = str(model_cfg.get("base_url") or "").strip().rstrip("/")
+            if cfg_base_url:
+                base_url = cfg_base_url
        configured_mode = _parse_api_mode(model_cfg.get("api_mode"))
        if configured_mode and _provider_supports_explicit_api_mode(provider, configured_provider):
            api_mode = configured_mode
@@ -681,6 +696,24 @@ def resolve_runtime_provider(
            logger.info("Auto-detected Codex provider but credentials failed; "
                        "falling through to next provider.")

+    if provider == "qwen-oauth":
+        try:
+            creds = resolve_qwen_runtime_credentials()
+            return {
+                "provider": "qwen-oauth",
+                "api_mode": "chat_completions",
+                "base_url": creds.get("base_url", "").rstrip("/"),
+                "api_key": creds.get("api_key", ""),
+                "source": creds.get("source", "qwen-cli"),
+                "expires_at_ms": creds.get("expires_at_ms"),
+                "requested_provider": requested_provider,
+            }
+        except AuthError:
+            if requested_provider != "auto":
+                raise
+            logger.info("Qwen OAuth credentials failed; "
+                        "falling through to next provider.")
+
    if provider == "copilot-acp":
        creds = resolve_external_process_provider_credentials(provider)
        return {
@@ -724,7 +757,15 @@ def resolve_runtime_provider(
    pconfig = PROVIDER_REGISTRY.get(provider)
    if pconfig and pconfig.auth_type == "api_key":
        creds = resolve_api_key_provider_credentials(provider)
-        base_url = creds.get("base_url", "").rstrip("/")
+        # Honour model.base_url from config.yaml when the configured provider
+        # matches this provider — mirrors the Anthropic path above.  Without
+        # this, users who set model.base_url to e.g. api.minimaxi.com/anthropic
+        # (China endpoint) still get the hardcoded api.minimax.io default (#6039).
+        cfg_provider = str(model_cfg.get("provider") or "").strip().lower()
+        cfg_base_url = ""
+        if cfg_provider == provider:
+            cfg_base_url = (model_cfg.get("base_url") or "").strip().rstrip("/")
+        base_url = cfg_base_url or creds.get("base_url", "").rstrip("/")
        api_mode = "chat_completions"
        if provider == "copilot":
            api_mode = _copilot_runtime_api_mode(model_cfg, creds.get("api_key", ""))
@@ -105,8 +105,8 @@ _DEFAULT_PROVIDER_MODELS = {
    ],
    "zai": ["glm-5", "glm-4.7", "glm-4.5", "glm-4.5-flash"],
    "kimi-coding": ["kimi-k2.5", "kimi-k2-thinking", "kimi-k2-turbo-preview"],
-    "minimax": ["MiniMax-M2.7", "MiniMax-M2.7-highspeed", "MiniMax-M2.5", "MiniMax-M2.5-highspeed", "MiniMax-M2.1"],
-    "minimax-cn": ["MiniMax-M2.7", "MiniMax-M2.7-highspeed", "MiniMax-M2.5", "MiniMax-M2.5-highspeed", "MiniMax-M2.1"],
+    "minimax": ["MiniMax-M1", "MiniMax-M1-40k", "MiniMax-M1-80k", "MiniMax-M1-128k", "MiniMax-M1-256k", "MiniMax-M2.5", "MiniMax-M2.7"],
+    "minimax-cn": ["MiniMax-M1", "MiniMax-M1-40k", "MiniMax-M1-80k", "MiniMax-M1-128k", "MiniMax-M1-256k", "MiniMax-M2.5", "MiniMax-M2.7"],
    "ai-gateway": ["anthropic/claude-opus-4.6", "anthropic/claude-sonnet-4.6", "openai/gpt-5", "google/gemini-3-flash"],
    "kilocode": ["anthropic/claude-opus-4.6", "anthropic/claude-sonnet-4.6", "openai/gpt-5.4", "google/gemini-3-pro-preview", "google/gemini-3-flash-preview"],
    "opencode-zen": ["gpt-5.4", "gpt-5.3-codex", "claude-sonnet-4-6", "gemini-3-flash", "glm-5", "kimi-k2.5", "minimax-m2.7"],
@@ -421,10 +421,22 @@ def _curses_prompt_choice(question: str, choices: list, default: int = 0) -> int
                curses.init_pair(1, curses.COLOR_GREEN, -1)
                curses.init_pair(2, curses.COLOR_YELLOW, -1)
            cursor = default
+            scroll_offset = 0

            while True:
                stdscr.clear()
                max_y, max_x = stdscr.getmaxyx()
+
+                # Rows available for list items: rows 2..(max_y-2) inclusive.
+                visible = max(1, max_y - 3)
+
+                # Scroll the viewport so the cursor is always visible.
+                if cursor < scroll_offset:
+                    scroll_offset = cursor
+                elif cursor >= scroll_offset + visible:
+                    scroll_offset = cursor - visible + 1
+                scroll_offset = max(0, min(scroll_offset, max(0, len(choices) - visible)))
+
                try:
                    stdscr.addnstr(
                        0,
@@ -436,12 +448,12 @@ def _curses_prompt_choice(question: str, choices: list, default: int = 0) -> int
                except curses.error:
                    pass

-                for i, choice in enumerate(choices):
-                    y = i + 2
+                for row, i in enumerate(range(scroll_offset, min(scroll_offset + visible, len(choices)))):
+                    y = row + 2
                    if y >= max_y - 1:
                        break
                    arrow = "→" if i == cursor else " "
-                    line = f" {arrow}  {choice}"
+                    line = f" {arrow}  {choices[i]}"
                    attr = curses.A_NORMAL
                    if i == cursor:
                        attr = curses.A_BOLD
@@ -153,12 +153,14 @@ def show_status(args):
    print(color("◆ Auth Providers", Colors.CYAN, Colors.BOLD))

    try:
-        from hermes_cli.auth import get_nous_auth_status, get_codex_auth_status
+        from hermes_cli.auth import get_nous_auth_status, get_codex_auth_status, get_qwen_auth_status
        nous_status = get_nous_auth_status()
        codex_status = get_codex_auth_status()
+        qwen_status = get_qwen_auth_status()
    except Exception:
        nous_status = {}
        codex_status = {}
+        qwen_status = {}

    nous_logged_in = bool(nous_status.get("logged_in"))
    print(
@@ -189,6 +191,21 @@ def show_status(args):
    if codex_status.get("error") and not codex_logged_in:
        print(f"    Error:      {codex_status.get('error')}")

+    qwen_logged_in = bool(qwen_status.get("logged_in"))
+    print(
+        f"  {'Qwen OAuth':<12}  {check_mark(qwen_logged_in)} "
+        f"{'logged in' if qwen_logged_in else 'not logged in (run: qwen auth qwen-oauth)'}"
+    )
+    qwen_auth_file = qwen_status.get("auth_file")
+    if qwen_auth_file:
+        print(f"    Auth file:  {qwen_auth_file}")
+    qwen_exp = qwen_status.get("expires_at_ms")
+    if qwen_exp:
+        from datetime import datetime, timezone
+        print(f"    Access exp: {datetime.fromtimestamp(int(qwen_exp) / 1000, tz=timezone.utc).isoformat()}")
+    if qwen_status.get("error") and not qwen_logged_in:
+        print(f"    Error:      {qwen_status.get('error')}")
+
    # =========================================================================
    # Nous Subscription Features
    # =========================================================================
@@ -554,6 +554,7 @@ def _get_platform_tools(
    # MCP servers are expected to be available on all platforms by default.
    # If the platform explicitly lists one or more MCP server names, treat that
    # as an allowlist. Otherwise include every globally enabled MCP server.
+    # Special sentinel: "no_mcp" in the toolset list disables all MCP servers.
    mcp_servers = config.get("mcp_servers") or {}
    enabled_mcp_servers = {
        name
@@ -561,10 +562,15 @@ def _get_platform_tools(
        if isinstance(server_cfg, dict)
        and _parse_enabled_flag(server_cfg.get("enabled", True), default=True)
    }
-    explicit_mcp_servers = explicit_passthrough & enabled_mcp_servers
-    enabled_toolsets.update(explicit_passthrough - enabled_mcp_servers)
+    # Allow "no_mcp" sentinel to opt out of all MCP servers for this platform
+    if "no_mcp" in toolset_names:
+        explicit_mcp_servers = set()
+        enabled_toolsets.update(explicit_passthrough - enabled_mcp_servers - {"no_mcp"})
+    else:
+        explicit_mcp_servers = explicit_passthrough & enabled_mcp_servers
+        enabled_toolsets.update(explicit_passthrough - enabled_mcp_servers)
    if include_default_mcp_servers:
-        if explicit_mcp_servers:
+        if explicit_mcp_servers or "no_mcp" in toolset_names:
            enabled_toolsets.update(explicit_mcp_servers)
        else:
            enabled_toolsets.update(enabled_mcp_servers)
@@ -464,7 +464,11 @@
      addToSystemPackages = mkOption {
        type = types.bool;
        default = false;
-        description = "Add hermes CLI to environment.systemPackages.";
+        description = ''
+          Add the hermes CLI to environment.systemPackages and export
+          HERMES_HOME system-wide (via environment.variables) so interactive
+          shells share state with the gateway service.
+        '';
      };

      # ── OCI Container (opt-in) ──────────────────────────────────────────
@@ -545,8 +549,12 @@
      })

      # ── Host CLI ──────────────────────────────────────────────────────
+      # Add the hermes CLI to system PATH and export HERMES_HOME system-wide
+      # so interactive shells share state (sessions, skills, cron) with the
+      # gateway service instead of creating a separate ~/.hermes/.
      (lib.mkIf cfg.addToSystemPackages {
        environment.systemPackages = [ cfg.package ];
+        environment.variables.HERMES_HOME = "${cfg.stateDir}/.hermes";
      })

      # ── Directories ───────────────────────────────────────────────────
@@ -601,7 +609,7 @@
          # so this is the single source of truth for both native and container mode.
          ${lib.optionalString (cfg.environment != {} || cfg.environmentFiles != []) ''
            ENV_FILE="${cfg.stateDir}/.hermes/.env"
-            install -o ${cfg.user} -g ${cfg.group} -m 0600 /dev/null "$ENV_FILE"
+            install -o ${cfg.user} -g ${cfg.group} -m 0640 /dev/null "$ENV_FILE"
            cat > "$ENV_FILE" <<'HERMES_NIX_ENV_EOF'
 ${envFileContent}
 HERMES_NIX_ENV_EOF
@@ -6,14 +6,68 @@
  uv2nix,
  pyproject-nix,
  pyproject-build-systems,
+  stdenv,
 }:
 let
  workspace = uv2nix.lib.workspace.loadWorkspace { workspaceRoot = ./..; };
+  hacks = callPackage pyproject-nix.build.hacks { };

  overlay = workspace.mkPyprojectOverlay {
    sourcePreference = "wheel";
  };

+  isAarch64Darwin = stdenv.hostPlatform.system == "aarch64-darwin";
+
+  # Keep the workspace locked through uv2nix, but supply the local voice stack
+  # from nixpkgs so wheel-only transitive artifacts do not break evaluation.
+  mkPrebuiltPassthru = dependencies: {
+    inherit dependencies;
+    optional-dependencies = { };
+    dependency-groups = { };
+  };
+
+  mkPrebuiltOverride = final: from: dependencies:
+    hacks.nixpkgsPrebuilt {
+      inherit from;
+      prev = {
+        nativeBuildInputs = [ final.pyprojectHook ];
+        passthru = mkPrebuiltPassthru dependencies;
+      };
+    };
+
+  pythonPackageOverrides = final: _prev:
+    if isAarch64Darwin then {
+      numpy = mkPrebuiltOverride final python311.pkgs.numpy { };
+
+      av = mkPrebuiltOverride final python311.pkgs.av { };
+
+      humanfriendly = mkPrebuiltOverride final python311.pkgs.humanfriendly { };
+
+      coloredlogs = mkPrebuiltOverride final python311.pkgs.coloredlogs {
+        humanfriendly = [ ];
+      };
+
+      onnxruntime = mkPrebuiltOverride final python311.pkgs.onnxruntime {
+        coloredlogs = [ ];
+        numpy = [ ];
+        packaging = [ ];
+      };
+
+      ctranslate2 = mkPrebuiltOverride final python311.pkgs.ctranslate2 {
+        numpy = [ ];
+        pyyaml = [ ];
+      };
+
+      faster-whisper = mkPrebuiltOverride final python311.pkgs.faster-whisper {
+        av = [ ];
+        ctranslate2 = [ ];
+        huggingface-hub = [ ];
+        onnxruntime = [ ];
+        tokenizers = [ ];
+        tqdm = [ ];
+      };
+    } else {};
+
  pythonSet =
    (callPackage pyproject-nix.build.packages {
      python = python311;
@@ -21,6 +75,7 @@ let
      (lib.composeManyExtensions [
        pyproject-build-systems.overlays.default
        overlay
+        pythonPackageOverrides
      ]);
 in
 pythonSet.mkVirtualEnv "hermes-agent-env" {
@@ -73,6 +73,7 @@ Config file: `~/.hermes/hindsight/config.json`
 |-----|---------|-------------|
 | `llm_provider` | `openai` | LLM provider: `openai`, `anthropic`, `gemini`, `groq`, `minimax`, `ollama` |
 | `llm_model` | per-provider | Model name (e.g. `gpt-4o-mini`, `openai/gpt-oss-120b`) |
+| `llm_base_url` | — | LLM Base URL override (e.g. `https://openrouter.ai/api/v1`) |

 The LLM API key is stored in `~/.hermes/.env` as `HINDSIGHT_LLM_API_KEY`.

@@ -92,6 +93,7 @@ Available in `hybrid` and `tools` memory modes:
 |----------|-------------|
 | `HINDSIGHT_API_KEY` | API key for Hindsight Cloud |
 | `HINDSIGHT_LLM_API_KEY` | LLM API key for local mode |
+| `HINDSIGHT_API_LLM_BASE_URL` | LLM Base URL for local mode (e.g. OpenRouter) |
 | `HINDSIGHT_API_URL` | Override API endpoint |
 | `HINDSIGHT_BANK_ID` | Override bank name |
 | `HINDSIGHT_BUDGET` | Override recall budget |
@@ -23,6 +23,8 @@ import json
 import logging
 import os
 import threading
+
+from hermes_constants import get_hermes_home
 from typing import Any, Dict, List

 from agent.memory_provider import MemoryProvider
@@ -142,7 +144,6 @@ def _load_config() -> dict:
      3. Environment variables
    """
    from pathlib import Path
-    from hermes_constants import get_hermes_home

    # Profile-scoped path (preferred)
    profile_path = get_hermes_home() / "hindsight" / "config.json"
@@ -234,6 +235,7 @@ class HindsightMemoryProvider(MemoryProvider):
            {"key": "api_key", "description": "Hindsight Cloud API key", "secret": True, "env_var": "HINDSIGHT_API_KEY", "url": "https://ui.hindsight.vectorize.io", "when": {"mode": "cloud"}},
            {"key": "llm_provider", "description": "LLM provider for local mode", "default": "openai", "choices": ["openai", "anthropic", "gemini", "groq", "minimax", "ollama"], "when": {"mode": "local"}},
            {"key": "llm_api_key", "description": "LLM API key for local Hindsight", "secret": True, "env_var": "HINDSIGHT_LLM_API_KEY", "when": {"mode": "local"}},
+            {"key": "llm_base_url", "description": "LLM Base URL (e.g. for OpenRouter)", "default": "", "env_var": "HINDSIGHT_API_LLM_BASE_URL", "when": {"mode": "local"}},
            {"key": "llm_model", "description": "LLM model for local mode", "default": "gpt-4o-mini", "default_from": {"field": "llm_provider", "map": _PROVIDER_DEFAULT_MODELS}, "when": {"mode": "local"}},
            {"key": "bank_id", "description": "Memory bank name", "default": "hermes"},
            {"key": "budget", "description": "Recall thoroughness", "default": "mid", "choices": ["low", "mid", "high"]},
@@ -250,12 +252,16 @@ class HindsightMemoryProvider(MemoryProvider):
                # different loop" errors during GC — we handle cleanup in
                # shutdown() instead.
                HindsightEmbedded.__del__ = lambda self: None
-                self._client = HindsightEmbedded(
+                kwargs = dict(
                    profile=self._config.get("profile", "hermes"),
                    llm_provider=self._config.get("llm_provider", ""),
-                    llm_api_key=self._config.get("llmApiKey") or os.environ.get("HINDSIGHT_LLM_API_KEY", ""),
+                    llm_api_key=self._config.get("llm_api_key") or os.environ.get("HINDSIGHT_LLM_API_KEY", ""),
                    llm_model=self._config.get("llm_model", ""),
                )
+                base_url = self._config.get("llm_base_url") or os.environ.get("HINDSIGHT_API_LLM_BASE_URL", "")
+                if base_url:
+                    kwargs["llm_base_url"] = base_url
+                self._client = HindsightEmbedded(**kwargs)
            else:
                from hindsight_client import Hindsight
                kwargs = {"base_url": self._api_url, "timeout": 30.0}
@@ -310,9 +316,10 @@ class HindsightMemoryProvider(MemoryProvider):
                    # If the config changed and the daemon is running, stop it.
                    from pathlib import Path as _Path
                    profile_env = _Path.home() / ".hindsight" / "profiles" / f"{profile}.env"
-                    current_key = self._config.get("llmApiKey") or os.environ.get("HINDSIGHT_LLM_API_KEY", "")
+                    current_key = self._config.get("llm_api_key") or os.environ.get("HINDSIGHT_LLM_API_KEY", "")
                    current_provider = self._config.get("llm_provider", "")
                    current_model = self._config.get("llm_model", "")
+                    current_base_url = self._config.get("llm_base_url") or os.environ.get("HINDSIGHT_API_LLM_BASE_URL", "")

                    # Read saved profile config
                    saved = {}
@@ -325,18 +332,22 @@ class HindsightMemoryProvider(MemoryProvider):
                    config_changed = (
                        saved.get("HINDSIGHT_API_LLM_PROVIDER") != current_provider or
                        saved.get("HINDSIGHT_API_LLM_MODEL") != current_model or
-                        saved.get("HINDSIGHT_API_LLM_API_KEY") != current_key
+                        saved.get("HINDSIGHT_API_LLM_API_KEY") != current_key or
+                        saved.get("HINDSIGHT_API_LLM_BASE_URL", "") != current_base_url
                    )

                    if config_changed:
                        # Write updated profile .env
                        profile_env.parent.mkdir(parents=True, exist_ok=True)
-                        profile_env.write_text(
+                        env_lines = (
                            f"HINDSIGHT_API_LLM_PROVIDER={current_provider}\n"
                            f"HINDSIGHT_API_LLM_API_KEY={current_key}\n"
                            f"HINDSIGHT_API_LLM_MODEL={current_model}\n"
                            f"HINDSIGHT_API_LOG_LEVEL=info\n"
                        )
+                        if current_base_url:
+                            env_lines += f"HINDSIGHT_API_LLM_BASE_URL={current_base_url}\n"
+                        profile_env.write_text(env_lines)
                        if client._manager.is_running(profile):
                            with open(log_path, "a") as f:
                                f.write("\n=== Config changed, restarting daemon ===\n")
@@ -17,7 +17,7 @@ Or manually:

 ```bash
 hermes config set memory.provider supermemory
-echo 'SUPERMEMORY_API_KEY=your-key-here' >> ~/.hermes/.env
+echo 'SUPERMEMORY_API_KEY=***' >> ~/.hermes/.env
 ```

 ## Config
@@ -26,15 +26,23 @@ Config file: `$HERMES_HOME/supermemory.json`

 | Key | Default | Description |
 |-----|---------|-------------|
-| `container_tag` | `hermes` | Container tag used for search and writes |
+| `container_tag` | `hermes` | Container tag used for search and writes. Supports `{identity}` template for profile-scoped tags (e.g. `hermes-{identity}` → `hermes-coder`). |
 | `auto_recall` | `true` | Inject relevant memory context before turns |
 | `auto_capture` | `true` | Store cleaned user-assistant turns after each response |
 | `max_recall_results` | `10` | Max recalled items to format into context |
 | `profile_frequency` | `50` | Include profile facts on first turn and every N turns |
 | `capture_mode` | `all` | Skip tiny or trivial turns by default |
+| `search_mode` | `hybrid` | Search mode: `hybrid` (profile + memories), `memories` (memories only), `documents` (documents only) |
 | `entity_context` | built-in default | Extraction guidance passed to Supermemory |
 | `api_timeout` | `5.0` | Timeout for SDK and ingest requests |

+### Environment Variables
+
+| Variable | Description |
+|----------|-------------|
+| `SUPERMEMORY_API_KEY` | API key (required) |
+| `SUPERMEMORY_CONTAINER_TAG` | Override container tag (takes priority over config file) |
+
 ## Tools

 | Tool | Description |
@@ -52,3 +60,40 @@ When enabled, Hermes can:
 - store cleaned conversation turns after each completed response
 - ingest the full session on session end for richer graph updates
 - expose explicit tools for search, store, forget, and profile access
+
+## Profile-Scoped Containers
+
+Use `{identity}` in the `container_tag` to scope memories per Hermes profile:
+
+```json
+{
+  "container_tag": "hermes-{identity}"
+}
+```
+
+For a profile named `coder`, this resolves to `hermes-coder`. The default profile resolves to `hermes-default`. Without `{identity}`, all profiles share the same container.
+
+## Multi-Container Mode
+
+For advanced setups (e.g. OpenClaw-style multi-workspace), you can enable custom container tags so the agent can read/write across multiple named containers:
+
+```json
+{
+  "container_tag": "hermes",
+  "enable_custom_container_tags": true,
+  "custom_containers": ["project-alpha", "project-beta", "shared-knowledge"],
+  "custom_container_instructions": "Use project-alpha for coding tasks, project-beta for research, and shared-knowledge for team-wide facts."
+}
+```
+
+When enabled:
+- `supermemory_search`, `supermemory_store`, `supermemory_forget`, and `supermemory_profile` accept an optional `container_tag` parameter
+- The tag must be in the whitelist: primary container + `custom_containers`
+- Automatic operations (turn sync, prefetch, memory write mirroring, session ingest) always use the **primary** container only
+- Custom container instructions are injected into the system prompt
+
+## Support
+
+- [Supermemory Discord](https://supermemory.link/discord)
+- [support@supermemory.com](mailto:support@supermemory.com)
+- [supermemory.ai](https://supermemory.ai)
@@ -26,6 +26,8 @@ _DEFAULT_CONTAINER_TAG = "hermes"
 _DEFAULT_MAX_RECALL_RESULTS = 10
 _DEFAULT_PROFILE_FREQUENCY = 50
 _DEFAULT_CAPTURE_MODE = "all"
+_DEFAULT_SEARCH_MODE = "hybrid"
+_VALID_SEARCH_MODES = ("hybrid", "memories", "documents")
 _DEFAULT_API_TIMEOUT = 5.0
 _MIN_CAPTURE_LENGTH = 10
 _MAX_ENTITY_CONTEXT_LENGTH = 1500
@@ -59,8 +61,12 @@ def _default_config() -> dict:
        "max_recall_results": _DEFAULT_MAX_RECALL_RESULTS,
        "profile_frequency": _DEFAULT_PROFILE_FREQUENCY,
        "capture_mode": _DEFAULT_CAPTURE_MODE,
+        "search_mode": _DEFAULT_SEARCH_MODE,
        "entity_context": _DEFAULT_ENTITY_CONTEXT,
        "api_timeout": _DEFAULT_API_TIMEOUT,
+        "enable_custom_container_tags": False,
+        "custom_containers": [],
+        "custom_container_instructions": "",
    }


@@ -100,7 +106,10 @@ def _load_supermemory_config(hermes_home: str) -> dict:
        except Exception:
            logger.debug("Failed to parse %s", config_path, exc_info=True)

-    config["container_tag"] = _sanitize_tag(str(config.get("container_tag", _DEFAULT_CONTAINER_TAG)))
+    # Keep raw container_tag — template variables like {identity} are resolved
+    # in initialize(), and _sanitize_tag runs AFTER resolution.
+    raw_tag = str(config.get("container_tag", _DEFAULT_CONTAINER_TAG)).strip()
+    config["container_tag"] = raw_tag if raw_tag else _DEFAULT_CONTAINER_TAG
    config["auto_recall"] = _as_bool(config.get("auto_recall"), True)
    config["auto_capture"] = _as_bool(config.get("auto_capture"), True)
    try:
@@ -112,11 +121,23 @@ def _load_supermemory_config(hermes_home: str) -> dict:
    except Exception:
        config["profile_frequency"] = _DEFAULT_PROFILE_FREQUENCY
    config["capture_mode"] = "everything" if config.get("capture_mode") == "everything" else "all"
+    raw_search_mode = str(config.get("search_mode", _DEFAULT_SEARCH_MODE)).strip().lower()
+    config["search_mode"] = raw_search_mode if raw_search_mode in _VALID_SEARCH_MODES else _DEFAULT_SEARCH_MODE
    config["entity_context"] = _clamp_entity_context(str(config.get("entity_context", _DEFAULT_ENTITY_CONTEXT)))
    try:
        config["api_timeout"] = max(0.5, min(15.0, float(config.get("api_timeout", _DEFAULT_API_TIMEOUT))))
    except Exception:
        config["api_timeout"] = _DEFAULT_API_TIMEOUT
+
+    # Multi-container support
+    config["enable_custom_container_tags"] = _as_bool(config.get("enable_custom_container_tags"), False)
+    raw_containers = config.get("custom_containers", [])
+    if isinstance(raw_containers, list):
+        config["custom_containers"] = [_sanitize_tag(str(t)) for t in raw_containers if t]
+    else:
+        config["custom_containers"] = []
+    config["custom_container_instructions"] = str(config.get("custom_container_instructions", "")).strip()
+
    return config


@@ -240,28 +261,41 @@ def _is_trivial_message(text: str) -> bool:


 class _SupermemoryClient:
-    def __init__(self, api_key: str, timeout: float, container_tag: str):
+    def __init__(self, api_key: str, timeout: float, container_tag: str, search_mode: str = "hybrid"):
        from supermemory import Supermemory

        self._api_key = api_key
        self._container_tag = container_tag
+        self._search_mode = search_mode if search_mode in _VALID_SEARCH_MODES else _DEFAULT_SEARCH_MODE
        self._timeout = timeout
        self._client = Supermemory(api_key=api_key, timeout=timeout, max_retries=0)

-    def add_memory(self, content: str, metadata: Optional[dict] = None, *, entity_context: str = "") -> dict:
-        kwargs = {
+    def add_memory(self, content: str, metadata: Optional[dict] = None, *,
+                   entity_context: str = "", container_tag: Optional[str] = None,
+                   custom_id: Optional[str] = None) -> dict:
+        tag = container_tag or self._container_tag
+        kwargs: dict[str, Any] = {
            "content": content.strip(),
-            "container_tags": [self._container_tag],
+            "container_tags": [tag],
        }
        if metadata:
            kwargs["metadata"] = metadata
        if entity_context:
            kwargs["entity_context"] = _clamp_entity_context(entity_context)
+        if custom_id:
+            kwargs["custom_id"] = custom_id
        result = self._client.documents.add(**kwargs)
        return {"id": getattr(result, "id", "")}

-    def search_memories(self, query: str, *, limit: int = 5) -> list[dict]:
-        response = self._client.search.memories(q=query, container_tag=self._container_tag, limit=limit)
+    def search_memories(self, query: str, *, limit: int = 5,
+                        container_tag: Optional[str] = None,
+                        search_mode: Optional[str] = None) -> list[dict]:
+        tag = container_tag or self._container_tag
+        mode = search_mode or self._search_mode
+        kwargs: dict[str, Any] = {"q": query, "container_tag": tag, "limit": limit}
+        if mode in _VALID_SEARCH_MODES:
+            kwargs["search_mode"] = mode
+        response = self._client.search.memories(**kwargs)
        results = []
        for item in (getattr(response, "results", None) or []):
            results.append({
@@ -273,8 +307,10 @@ class _SupermemoryClient:
            })
        return results

-    def get_profile(self, query: Optional[str] = None) -> dict:
-        kwargs = {"container_tag": self._container_tag}
+    def get_profile(self, query: Optional[str] = None, *,
+                    container_tag: Optional[str] = None) -> dict:
+        tag = container_tag or self._container_tag
+        kwargs: dict[str, Any] = {"container_tag": tag}
        if query:
            kwargs["q"] = query
        response = self._client.profile(**kwargs)
@@ -296,18 +332,19 @@ class _SupermemoryClient:
                    })
        return {"static": static, "dynamic": dynamic, "search_results": search_results}

-    def forget_memory(self, memory_id: str) -> None:
-        self._client.memories.forget(container_tag=self._container_tag, id=memory_id)
+    def forget_memory(self, memory_id: str, *, container_tag: Optional[str] = None) -> None:
+        tag = container_tag or self._container_tag
+        self._client.memories.forget(container_tag=tag, id=memory_id)

-    def forget_by_query(self, query: str) -> dict:
-        results = self.search_memories(query, limit=5)
+    def forget_by_query(self, query: str, *, container_tag: Optional[str] = None) -> dict:
+        results = self.search_memories(query, limit=5, container_tag=container_tag)
        if not results:
            return {"success": False, "message": "No matching memory found to forget."}
        target = results[0]
        memory_id = target.get("id", "")
        if not memory_id:
            return {"success": False, "message": "Best matching memory has no id."}
-        self.forget_memory(memory_id)
+        self.forget_memory(memory_id, container_tag=container_tag)
        preview = (target.get("memory") or "")[:100]
        return {"success": True, "message": f'Forgot: "{preview}"', "id": memory_id}

@@ -398,11 +435,17 @@ class SupermemoryMemoryProvider(MemoryProvider):
        self._max_recall_results = _DEFAULT_MAX_RECALL_RESULTS
        self._profile_frequency = _DEFAULT_PROFILE_FREQUENCY
        self._capture_mode = _DEFAULT_CAPTURE_MODE
+        self._search_mode = _DEFAULT_SEARCH_MODE
        self._entity_context = _DEFAULT_ENTITY_CONTEXT
        self._api_timeout = _DEFAULT_API_TIMEOUT
        self._hermes_home = ""
        self._write_enabled = True
        self._active = False
+        # Multi-container support
+        self._enable_custom_containers = False
+        self._custom_containers: List[str] = []
+        self._custom_container_instructions = ""
+        self._allowed_containers: List[str] = []

    @property
    def name(self) -> str:
@@ -419,16 +462,11 @@ class SupermemoryMemoryProvider(MemoryProvider):
            return False

    def get_config_schema(self):
+        # Only prompt for the API key during `hermes memory setup`.
+        # All other options are documented for $HERMES_HOME/supermemory.json
+        # or the SUPERMEMORY_CONTAINER_TAG env var.
        return [
            {"key": "api_key", "description": "Supermemory API key", "secret": True, "required": True, "env_var": "SUPERMEMORY_API_KEY", "url": "https://supermemory.ai"},
-            {"key": "container_tag", "description": "Container tag for reads and writes", "default": _DEFAULT_CONTAINER_TAG},
-            {"key": "auto_recall", "description": "Enable automatic recall before each turn", "default": "true", "choices": ["true", "false"]},
-            {"key": "auto_capture", "description": "Enable automatic capture after each completed turn", "default": "true", "choices": ["true", "false"]},
-            {"key": "max_recall_results", "description": "Maximum recalled items to inject", "default": str(_DEFAULT_MAX_RECALL_RESULTS)},
-            {"key": "profile_frequency", "description": "Include profile facts on first turn and every N turns", "default": str(_DEFAULT_PROFILE_FREQUENCY)},
-            {"key": "capture_mode", "description": "Capture mode", "default": _DEFAULT_CAPTURE_MODE, "choices": ["all", "everything"]},
-            {"key": "entity_context", "description": "Extraction guidance passed to Supermemory", "default": _DEFAULT_ENTITY_CONTEXT},
-            {"key": "api_timeout", "description": "Timeout in seconds for SDK and ingest calls", "default": str(_DEFAULT_API_TIMEOUT)},
        ]

    def save_config(self, values, hermes_home):
@@ -446,14 +484,29 @@ class SupermemoryMemoryProvider(MemoryProvider):
        self._turn_count = 0
        self._config = _load_supermemory_config(self._hermes_home)
        self._api_key = os.environ.get("SUPERMEMORY_API_KEY", "")
-        self._container_tag = self._config["container_tag"]
+
+        # Resolve container tag: env var > config > default.
+        # Supports {identity} template for profile-scoped containers.
+        env_tag = os.environ.get("SUPERMEMORY_CONTAINER_TAG", "").strip()
+        raw_tag = env_tag or self._config["container_tag"]
+        identity = kwargs.get("agent_identity", "default")
+        self._container_tag = _sanitize_tag(raw_tag.replace("{identity}", identity))
+
        self._auto_recall = self._config["auto_recall"]
        self._auto_capture = self._config["auto_capture"]
        self._max_recall_results = self._config["max_recall_results"]
        self._profile_frequency = self._config["profile_frequency"]
        self._capture_mode = self._config["capture_mode"]
+        self._search_mode = self._config["search_mode"]
        self._entity_context = self._config["entity_context"]
        self._api_timeout = self._config["api_timeout"]
+
+        # Multi-container setup
+        self._enable_custom_containers = self._config["enable_custom_container_tags"]
+        self._custom_containers = self._config["custom_containers"]
+        self._custom_container_instructions = self._config["custom_container_instructions"]
+        self._allowed_containers = [self._container_tag] + list(self._custom_containers)
+
        agent_context = kwargs.get("agent_context", "")
        self._write_enabled = agent_context not in ("cron", "flush", "subagent")
        self._active = bool(self._api_key)
@@ -464,6 +517,7 @@ class SupermemoryMemoryProvider(MemoryProvider):
                    api_key=self._api_key,
                    timeout=self._api_timeout,
                    container_tag=self._container_tag,
+                    search_mode=self._search_mode,
                )
            except Exception:
                logger.warning("Supermemory initialization failed", exc_info=True)
@@ -476,11 +530,18 @@ class SupermemoryMemoryProvider(MemoryProvider):
    def system_prompt_block(self) -> str:
        if not self._active:
            return ""
-        return (
-            "# Supermemory\n"
-            f"Active. Container: {self._container_tag}.\n"
-            "Use supermemory_search, supermemory_store, supermemory_forget, and supermemory_profile for explicit memory operations."
-        )
+        lines = [
+            "# Supermemory",
+            f"Active. Container: {self._container_tag}.",
+            "Use supermemory_search, supermemory_store, supermemory_forget, and supermemory_profile for explicit memory operations.",
+        ]
+        if self._enable_custom_containers and self._custom_containers:
+            tags_str = ", ".join(self._allowed_containers)
+            lines.append(f"\nMulti-container mode enabled. Available containers: {tags_str}.")
+            lines.append("Pass an optional container_tag to supermemory_search, supermemory_store, supermemory_forget, and supermemory_profile to target a specific container.")
+            if self._custom_container_instructions:
+                lines.append(f"\n{self._custom_container_instructions}")
+        return "\n".join(lines)

    def prefetch(self, query: str, *, session_id: str = "") -> str:
        if not self._active or not self._auto_recall or not self._client or not query.strip():
@@ -582,22 +643,62 @@ class SupermemoryMemoryProvider(MemoryProvider):
                thread.join(timeout=5.0)
            setattr(self, attr_name, None)

+    def _resolve_tool_container_tag(self, args: dict) -> Optional[str]:
+        """Validate and resolve container_tag from tool call args.
+
+        Returns None (use primary) if multi-container is disabled or no tag provided.
+        Returns the validated tag if it's in the allowed list.
+        Raises ValueError if the tag is not whitelisted.
+        """
+        if not self._enable_custom_containers:
+            return None
+        tag = str(args.get("container_tag") or "").strip()
+        if not tag:
+            return None
+        sanitized = _sanitize_tag(tag)
+        if sanitized not in self._allowed_containers:
+            raise ValueError(
+                f"Container tag '{sanitized}' is not allowed. "
+                f"Allowed: {', '.join(self._allowed_containers)}"
+            )
+        return sanitized
+
    def get_tool_schemas(self) -> List[Dict[str, Any]]:
-        return [STORE_SCHEMA, SEARCH_SCHEMA, FORGET_SCHEMA, PROFILE_SCHEMA]
+        if not self._enable_custom_containers:
+            return [STORE_SCHEMA, SEARCH_SCHEMA, FORGET_SCHEMA, PROFILE_SCHEMA]
+
+        # When multi-container is enabled, add optional container_tag to relevant tools
+        container_param = {
+            "type": "string",
+            "description": f"Optional container tag. Allowed: {', '.join(self._allowed_containers)}. Defaults to primary ({self._container_tag}).",
+        }
+        schemas = []
+        for base in [STORE_SCHEMA, SEARCH_SCHEMA, FORGET_SCHEMA, PROFILE_SCHEMA]:
+            schema = json.loads(json.dumps(base))  # deep copy
+            schema["parameters"]["properties"]["container_tag"] = container_param
+            schemas.append(schema)
+        return schemas

    def _tool_store(self, args: dict) -> str:
        content = str(args.get("content") or "").strip()
        if not content:
            return tool_error("content is required")
+        try:
+            tag = self._resolve_tool_container_tag(args)
+        except ValueError as exc:
+            return tool_error(str(exc))
        metadata = args.get("metadata") or {}
        if not isinstance(metadata, dict):
            metadata = {}
        metadata.setdefault("type", _detect_category(content))
        metadata["source"] = "hermes_tool"
        try:
-            result = self._client.add_memory(content, metadata=metadata, entity_context=self._entity_context)
+            result = self._client.add_memory(content, metadata=metadata, entity_context=self._entity_context, container_tag=tag)
            preview = content[:80] + ("..." if len(content) > 80 else "")
-            return json.dumps({"saved": True, "id": result.get("id", ""), "preview": preview})
+            resp: dict[str, Any] = {"saved": True, "id": result.get("id", ""), "preview": preview}
+            if tag:
+                resp["container_tag"] = tag
+            return json.dumps(resp)
        except Exception as exc:
            return tool_error(f"Failed to store memory: {exc}")

@@ -605,22 +706,29 @@ class SupermemoryMemoryProvider(MemoryProvider):
        query = str(args.get("query") or "").strip()
        if not query:
            return tool_error("query is required")
+        try:
+            tag = self._resolve_tool_container_tag(args)
+        except ValueError as exc:
+            return tool_error(str(exc))
        try:
            limit = max(1, min(20, int(args.get("limit", 5) or 5)))
        except Exception:
            limit = 5
        try:
-            results = self._client.search_memories(query, limit=limit)
+            results = self._client.search_memories(query, limit=limit, container_tag=tag)
            formatted = []
            for item in results:
-                entry = {"id": item.get("id", ""), "content": item.get("memory", "")}
+                entry: dict[str, Any] = {"id": item.get("id", ""), "content": item.get("memory", "")}
                if item.get("similarity") is not None:
                    try:
                        entry["similarity"] = round(float(item["similarity"]) * 100)
                    except Exception:
                        pass
                formatted.append(entry)
-            return json.dumps({"results": formatted, "count": len(formatted)})
+            resp: dict[str, Any] = {"results": formatted, "count": len(formatted)}
+            if tag:
+                resp["container_tag"] = tag
+            return json.dumps(resp)
        except Exception as exc:
            return tool_error(f"Search failed: {exc}")

@@ -629,28 +737,39 @@ class SupermemoryMemoryProvider(MemoryProvider):
        query = str(args.get("query") or "").strip()
        if not memory_id and not query:
            return tool_error("Provide either id or query")
+        try:
+            tag = self._resolve_tool_container_tag(args)
+        except ValueError as exc:
+            return tool_error(str(exc))
        try:
            if memory_id:
-                self._client.forget_memory(memory_id)
+                self._client.forget_memory(memory_id, container_tag=tag)
                return json.dumps({"forgotten": True, "id": memory_id})
-            return json.dumps(self._client.forget_by_query(query))
+            return json.dumps(self._client.forget_by_query(query, container_tag=tag))
        except Exception as exc:
            return tool_error(f"Forget failed: {exc}")

    def _tool_profile(self, args: dict) -> str:
        query = str(args.get("query") or "").strip() or None
        try:
-            profile = self._client.get_profile(query=query)
+            tag = self._resolve_tool_container_tag(args)
+        except ValueError as exc:
+            return tool_error(str(exc))
+        try:
+            profile = self._client.get_profile(query=query, container_tag=tag)
            sections = []
            if profile["static"]:
                sections.append("## User Profile (Persistent)\n" + "\n".join(f"- {item}" for item in profile["static"]))
            if profile["dynamic"]:
                sections.append("## Recent Context\n" + "\n".join(f"- {item}" for item in profile["dynamic"]))
-            return json.dumps({
+            resp: dict[str, Any] = {
                "profile": "\n\n".join(sections),
                "static_count": len(profile["static"]),
                "dynamic_count": len(profile["dynamic"]),
-            })
+            }
+            if tag:
+                resp["container_tag"] = tag
+            return json.dumps(resp)
        except Exception as exc:
            return tool_error(f"Profile failed: {exc}")

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"

 [project]
 name = "hermes-agent"
-version = "0.7.0"
+version = "0.8.0"
 description = "The self-improving AI agent — creates skills from experience, improves them during use, and runs anywhere"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -62,6 +62,7 @@ mcp = ["mcp>=1.2.0,<2"]
 homeassistant = ["aiohttp>=3.9.0,<4"]
 sms = ["aiohttp>=3.9.0,<4"]
 acp = ["agent-client-protocol>=0.9.0,<1.0"]
+mistral = ["mistralai>=2.3.0,<3"]
 dingtalk = ["dingtalk-stream>=0.1.0,<1"]
 feishu = ["lark-oapi>=1.5.3,<2"]
 rl = [
@@ -94,6 +95,7 @@ all = [
  "hermes-agent[voice]",
  "hermes-agent[dingtalk]",
  "hermes-agent[feishu]",
+  "hermes-agent[mistral]",
 ]

 [project.scripts]
@@ -66,7 +66,8 @@ from model_tools import (
    handle_function_call,
    check_toolset_requirements,
 )
-from tools.terminal_tool import cleanup_vm
+from tools.terminal_tool import cleanup_vm, get_active_env
+from tools.tool_result_storage import maybe_persist_tool_result, enforce_turn_budget
 from tools.interrupt import set_interrupt as _set_interrupt
 from tools.browser_tool import cleanup_browser

@@ -75,6 +76,7 @@ from hermes_constants import OPENROUTER_BASE_URL

 # Agent internals extracted to agent/ package for modularity
 from agent.memory_manager import build_memory_context_block
+from agent.retry_utils import jittered_backoff
 from agent.prompt_builder import (
    DEFAULT_AGENT_IDENTITY, PLATFORM_HINTS,
    MEMORY_GUIDANCE, SESSION_SEARCH_GUIDANCE, SKILLS_GUIDANCE,
@@ -85,6 +87,7 @@ from agent.model_metadata import (
    estimate_tokens_rough, estimate_messages_tokens_rough, estimate_request_tokens_rough,
    get_next_probe_tier, parse_context_limit_from_error,
    save_context_length, is_local_endpoint,
+    query_ollama_num_ctx,
 )
 from agent.context_compressor import ContextCompressor
 from agent.subdirectory_hints import SubdirectoryHintTracker
@@ -409,62 +412,26 @@ def _strip_budget_warnings_from_history(messages: list) -> None:
 # Large tool result handler — save oversized output to temp file
 # =========================================================================

-# Threshold at which tool results are saved to a file instead of kept inline.
-# 100K chars ≈ 25K tokens — generous for any reasonable output but prevents
-# catastrophic context explosions.
-_LARGE_RESULT_CHARS = 100_000

-# How many characters of the original result to include as an inline preview
-# so the model has immediate context about what the tool returned.
-_LARGE_RESULT_PREVIEW_CHARS = 1_500
+# =========================================================================
+# Qwen Portal headers — mimics QwenCode CLI for portal.qwen.ai compatibility.
+# Extracted as a module-level helper so both __init__ and
+# _apply_client_headers_for_base_url can share it.
+# =========================================================================
+_QWEN_CODE_VERSION = "0.14.1"


-def _save_oversized_tool_result(function_name: str, function_result: str) -> str:
-    """Replace oversized tool results with a file reference + preview.
+def _qwen_portal_headers() -> dict:
+    """Return default HTTP headers required by Qwen Portal API."""
+    import platform as _plat

-    When a tool returns more than ``_LARGE_RESULT_CHARS`` characters, the full
-    content is written to a temporary file under ``HERMES_HOME/cache/tool_responses/``
-    and the result sent to the model is replaced with:
-      • a brief head preview  (first ``_LARGE_RESULT_PREVIEW_CHARS`` chars)
-      • the file path so the model can use ``read_file`` / ``search_files``
-
-    Falls back to destructive truncation if the file write fails.
-    """
-    original_len = len(function_result)
-    if original_len <= _LARGE_RESULT_CHARS:
-        return function_result
-
-    # Build the target directory
-    try:
-        response_dir = os.path.join(get_hermes_home(), "cache", "tool_responses")
-        os.makedirs(response_dir, exist_ok=True)
-
-        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
-        # Sanitize tool name for use in filename
-        safe_name = re.sub(r"[^\w\-]", "_", function_name)[:40]
-        filename = f"{safe_name}_{timestamp}.txt"
-        filepath = os.path.join(response_dir, filename)
-
-        with open(filepath, "w", encoding="utf-8") as f:
-            f.write(function_result)
-
-        preview = function_result[:_LARGE_RESULT_PREVIEW_CHARS]
-        return (
-            f"{preview}\n\n"
-            f"[Large tool response: {original_len:,} characters total — "
-            f"only the first {_LARGE_RESULT_PREVIEW_CHARS:,} shown above. "
-            f"Full output saved to: {filepath}\n"
-            f"Use read_file or search_files on that path to access the rest.]"
-        )
-    except Exception as exc:
-        # Fall back to destructive truncation if file write fails
-        logger.warning("Failed to save large tool result to file: %s", exc)
-        return (
-            function_result[:_LARGE_RESULT_CHARS]
-            + f"\n\n[Truncated: tool response was {original_len:,} chars, "
-            f"exceeding the {_LARGE_RESULT_CHARS:,} char limit. "
-            f"File save failed: {exc}]"
-        )
+    _ua = f"QwenCode/{_QWEN_CODE_VERSION} ({_plat.system().lower()}; {_plat.machine()})"
+    return {
+        "User-Agent": _ua,
+        "X-DashScope-CacheControl": "enable",
+        "X-DashScope-UserAgent": _ua,
+        "X-DashScope-AuthType": "qwen-oauth",
+    }


 class AIAgent:
@@ -810,6 +777,8 @@ class AIAgent:
                    client_kwargs["default_headers"] = {
                        "User-Agent": "KimiCLI/1.3",
                    }
+                elif "portal.qwen.ai" in effective_base.lower():
+                    client_kwargs["default_headers"] = _qwen_portal_headers()
            else:
                # No explicit creds — use the centralized provider router
                from agent.auxiliary_client import resolve_provider_client
@@ -1216,6 +1185,33 @@ class AIAgent:
        self.session_cost_status = "unknown"
        self.session_cost_source = "none"
        
+        # ── Ollama num_ctx injection ──
+        # Ollama defaults to 2048 context regardless of the model's capabilities.
+        # When running against an Ollama server, detect the model's max context
+        # and pass num_ctx on every chat request so the full window is used.
+        # User override: set model.ollama_num_ctx in config.yaml to cap VRAM use.
+        self._ollama_num_ctx: int | None = None
+        _ollama_num_ctx_override = None
+        if isinstance(_model_cfg, dict):
+            _ollama_num_ctx_override = _model_cfg.get("ollama_num_ctx")
+        if _ollama_num_ctx_override is not None:
+            try:
+                self._ollama_num_ctx = int(_ollama_num_ctx_override)
+            except (TypeError, ValueError):
+                logger.debug("Invalid ollama_num_ctx config value: %r", _ollama_num_ctx_override)
+        if self._ollama_num_ctx is None and self.base_url and is_local_endpoint(self.base_url):
+            try:
+                _detected = query_ollama_num_ctx(self.model, self.base_url)
+                if _detected and _detected > 0:
+                    self._ollama_num_ctx = _detected
+            except Exception as exc:
+                logger.debug("Ollama num_ctx detection failed: %s", exc)
+        if self._ollama_num_ctx and not self.quiet_mode:
+            logger.info(
+                "Ollama num_ctx: will request %d tokens (model max from /api/show)",
+                self._ollama_num_ctx,
+            )
+
        if not self.quiet_mode:
            if compression_enabled:
                print(f"📊 Context limit: {self.context_compressor.context_length:,} tokens (compress at {int(compression_threshold*100)}% = {self.context_compressor.threshold_tokens:,})")
@@ -4107,6 +4103,8 @@ class AIAgent:
            self._client_kwargs["default_headers"] = copilot_default_headers()
        elif "api.kimi.com" in normalized:
            self._client_kwargs["default_headers"] = {"User-Agent": "KimiCLI/1.3"}
+        elif "portal.qwen.ai" in normalized:
+            self._client_kwargs["default_headers"] = _qwen_portal_headers()
        else:
            self._client_kwargs.pop("default_headers", None)

@@ -4897,7 +4895,7 @@ class AIAgent:
                effective_key = (fb_client.api_key or resolve_anthropic_token() or "") if fb_provider == "anthropic" else (fb_client.api_key or "")
                self.api_key = effective_key
                self._anthropic_api_key = effective_key
-                self._anthropic_base_url = getattr(fb_client, "base_url", None)
+                self._anthropic_base_url = fb_base_url
                self._anthropic_client = build_anthropic_client(effective_key, self._anthropic_base_url)
                self._is_anthropic_oauth = _is_oauth_token(effective_key)
                self.client = None
@@ -5253,6 +5251,71 @@ class AIAgent:
        base = (getattr(self, "base_url", "") or "").lower()
        return "dashscope" in base or "aliyuncs" in base or "opencode.ai/zen/go" in base

+    def _is_qwen_portal(self) -> bool:
+        """Return True when the base URL targets Qwen Portal."""
+        return "portal.qwen.ai" in self._base_url_lower
+
+    def _qwen_prepare_chat_messages(self, api_messages: list) -> list:
+        prepared = copy.deepcopy(api_messages)
+        if not prepared:
+            return prepared
+
+        for msg in prepared:
+            if not isinstance(msg, dict):
+                continue
+            content = msg.get("content")
+            if isinstance(content, str):
+                msg["content"] = [{"type": "text", "text": content}]
+            elif isinstance(content, list):
+                # Normalize: convert bare strings to text dicts, keep dicts as-is.
+                # deepcopy already created independent copies, no need for dict().
+                normalized_parts = []
+                for part in content:
+                    if isinstance(part, str):
+                        normalized_parts.append({"type": "text", "text": part})
+                    elif isinstance(part, dict):
+                        normalized_parts.append(part)
+                if normalized_parts:
+                    msg["content"] = normalized_parts
+
+        # Inject cache_control on the last part of the system message.
+        for msg in prepared:
+            if isinstance(msg, dict) and msg.get("role") == "system":
+                content = msg.get("content")
+                if isinstance(content, list) and content and isinstance(content[-1], dict):
+                    content[-1]["cache_control"] = {"type": "ephemeral"}
+                break
+
+        return prepared
+
+    def _qwen_prepare_chat_messages_inplace(self, messages: list) -> None:
+        """In-place variant — mutates an already-copied message list."""
+        if not messages:
+            return
+
+        for msg in messages:
+            if not isinstance(msg, dict):
+                continue
+            content = msg.get("content")
+            if isinstance(content, str):
+                msg["content"] = [{"type": "text", "text": content}]
+            elif isinstance(content, list):
+                normalized_parts = []
+                for part in content:
+                    if isinstance(part, str):
+                        normalized_parts.append({"type": "text", "text": part})
+                    elif isinstance(part, dict):
+                        normalized_parts.append(part)
+                if normalized_parts:
+                    msg["content"] = normalized_parts
+
+        for msg in messages:
+            if isinstance(msg, dict) and msg.get("role") == "system":
+                content = msg.get("content")
+                if isinstance(content, list) and content and isinstance(content[-1], dict):
+                    content[-1]["cache_control"] = {"type": "ephemeral"}
+                break
+
    def _build_api_kwargs(self, api_messages: list) -> dict:
        """Build the keyword arguments dict for the active API mode."""
        if self.api_mode == "anthropic_messages":
@@ -5271,6 +5334,7 @@ class AIAgent:
                is_oauth=self._is_anthropic_oauth,
                preserve_dots=self._anthropic_preserve_dots(),
                context_length=ctx_len,
+                base_url=getattr(self, "_anthropic_base_url", None),
            )

        if self.api_mode == "codex_responses":
@@ -5364,6 +5428,17 @@ class AIAgent:
                            tool_call.pop("call_id", None)
                            tool_call.pop("response_item_id", None)

+        # Qwen portal: normalize content to list-of-dicts, inject cache_control.
+        # Must run AFTER codex sanitization so we transform the final messages.
+        # If sanitization already deepcopied, reuse that copy (in-place).
+        if self._is_qwen_portal():
+            if sanitized_messages is api_messages:
+                # No sanitization was done — we need our own copy.
+                sanitized_messages = self._qwen_prepare_chat_messages(sanitized_messages)
+            else:
+                # Already a deepcopy — transform in place to avoid a second deepcopy.
+                self._qwen_prepare_chat_messages_inplace(sanitized_messages)
+
        # GPT-5 and Codex models respond better to 'developer' than 'system'
        # for instruction-following.  Swap the role at the API boundary so
        # internal message representation stays uniform ("system").
@@ -5396,11 +5471,17 @@ class AIAgent:
            "messages": sanitized_messages,
            "timeout": float(os.getenv("HERMES_API_TIMEOUT", 1800.0)),
        }
+        if self._is_qwen_portal():
+            api_kwargs["metadata"] = {
+                "sessionId": self.session_id or "hermes",
+                "promptId": str(uuid.uuid4()),
+            }
        if self.tools:
            api_kwargs["tools"] = self.tools

        if self.max_tokens is not None:
-            api_kwargs.update(self._max_tokens_param(self.max_tokens))
+            if not self._is_qwen_portal():
+                api_kwargs.update(self._max_tokens_param(self.max_tokens))
        elif self._is_openrouter_url() and "claude" in (self.model or "").lower():
            # OpenRouter translates requests to Anthropic's Messages API,
            # which requires max_tokens as a mandatory field.  When we omit
@@ -5456,6 +5537,18 @@ class AIAgent:
        if _is_nous:
            extra_body["tags"] = ["product=hermes-agent"]

+        # Ollama num_ctx: override the 2048 default so the model actually
+        # uses the context window it was trained for.  Passed via the OpenAI
+        # SDK's extra_body → options.num_ctx, which Ollama's OpenAI-compat
+        # endpoint forwards to the runner as --ctx-size.
+        if self._ollama_num_ctx:
+            options = extra_body.get("options", {})
+            options["num_ctx"] = self._ollama_num_ctx
+            extra_body["options"] = options
+
+        if self._is_qwen_portal():
+            extra_body["vl_high_resolution_images"] = True
+
        if extra_body:
            api_kwargs["extra_body"] = extra_body

@@ -6224,15 +6317,17 @@ class AIAgent:
                except Exception as cb_err:
                    logging.debug(f"Tool complete callback error: {cb_err}")

-            # Save oversized results to file instead of destructive truncation
-            function_result = _save_oversized_tool_result(name, function_result)
+            function_result = maybe_persist_tool_result(
+                content=function_result,
+                tool_name=name,
+                tool_use_id=tc.id,
+                env=get_active_env(effective_task_id),
+            )

-            # Discover subdirectory context files from tool arguments
            subdir_hints = self._subdirectory_hints.check_tool_call(name, args)
            if subdir_hints:
                function_result += subdir_hints

-            # Append tool result message in order
            tool_msg = {
                "role": "tool",
                "content": function_result,
@@ -6240,6 +6335,12 @@ class AIAgent:
            }
            messages.append(tool_msg)

+        # ── Per-turn aggregate budget enforcement ─────────────────────────
+        num_tools = len(parsed_calls)
+        if num_tools > 0:
+            turn_tool_msgs = messages[-num_tools:]
+            enforce_turn_budget(turn_tool_msgs, env=get_active_env(effective_task_id))
+
        # ── Budget pressure injection ────────────────────────────────────
        budget_warning = self._get_budget_warning(api_call_count)
        if budget_warning and messages and messages[-1].get("role") == "tool":
@@ -6524,8 +6625,12 @@ class AIAgent:
                except Exception as cb_err:
                    logging.debug(f"Tool complete callback error: {cb_err}")

-            # Save oversized results to file instead of destructive truncation
-            function_result = _save_oversized_tool_result(function_name, function_result)
+            function_result = maybe_persist_tool_result(
+                content=function_result,
+                tool_name=function_name,
+                tool_use_id=tool_call.id,
+                env=get_active_env(effective_task_id),
+            )

            # Discover subdirectory context files from tool arguments
            subdir_hints = self._subdirectory_hints.check_tool_call(function_name, function_args)
@@ -6563,6 +6668,11 @@ class AIAgent:
            if self.tool_delay > 0 and i < len(assistant_message.tool_calls):
                time.sleep(self.tool_delay)

+        # ── Per-turn aggregate budget enforcement ─────────────────────────
+        num_tools_seq = len(assistant_message.tool_calls)
+        if num_tools_seq > 0:
+            enforce_turn_budget(messages[-num_tools_seq:], env=get_active_env(effective_task_id))
+
        # ── Budget pressure injection ─────────────────────────────────
        # After all tool calls in this turn are processed, check if we're
        # approaching max_iterations. If so, inject a warning into the LAST
@@ -7289,6 +7399,7 @@ class AIAgent:
            codex_auth_retry_attempted=False
            anthropic_auth_retry_attempted=False
            nous_auth_retry_attempted=False
+            thinking_sig_retry_attempted = False
            has_retried_429 = False
            restart_with_compressed_messages = False
            restart_with_length_continuation = False
@@ -7391,20 +7502,30 @@ class AIAgent:
                            response_invalid = True
                            error_details.append("response.output is not a list")
                        elif not output_items:
-                            # If we reach here, _run_codex_stream's backfill
-                            # from output_item.done events and text-delta
-                            # synthesis both failed to populate output.
-                            _resp_status = getattr(response, "status", None)
-                            _resp_incomplete = getattr(response, "incomplete_details", None)
-                            logging.warning(
-                                "Codex response.output is empty after stream backfill "
-                                "(status=%s, incomplete_details=%s, model=%s). %s",
-                                _resp_status, _resp_incomplete,
-                                getattr(response, "model", None),
-                                f"api_mode={self.api_mode} provider={self.provider}",
-                            )
-                            response_invalid = True
-                            error_details.append("response.output is empty")
+                            # Stream backfill may have failed, but
+                            # _normalize_codex_response can still recover
+                            # from response.output_text. Only mark invalid
+                            # when that fallback is also absent.
+                            _out_text = getattr(response, "output_text", None)
+                            _out_text_stripped = _out_text.strip() if isinstance(_out_text, str) else ""
+                            if _out_text_stripped:
+                                logger.debug(
+                                    "Codex response.output is empty but output_text is present "
+                                    "(%d chars); deferring to normalization.",
+                                    len(_out_text_stripped),
+                                )
+                            else:
+                                _resp_status = getattr(response, "status", None)
+                                _resp_incomplete = getattr(response, "incomplete_details", None)
+                                logger.warning(
+                                    "Codex response.output is empty after stream backfill "
+                                    "(status=%s, incomplete_details=%s, model=%s). %s",
+                                    _resp_status, _resp_incomplete,
+                                    getattr(response, "model", None),
+                                    f"api_mode={self.api_mode} provider={self.provider}",
+                                )
+                                response_invalid = True
+                                error_details.append("response.output is empty")
                    elif self.api_mode == "anthropic_messages":
                        content_blocks = getattr(response, "content", None) if response is not None else None
                        if response is None:
@@ -7494,7 +7615,8 @@ class AIAgent:
                            }
                        
                        # Longer backoff for rate limiting (likely cause of None choices)
-                        wait_time = min(5 * (2 ** (retry_count - 1)), 120)  # 5s, 10s, 20s, 40s, 80s, 120s
+                        # Jittered exponential: 5s base, 120s cap + random jitter
+                        wait_time = jittered_backoff(retry_count, base_delay=5.0, max_delay=120.0)
                        self._vprint(f"{self.log_prefix}⏳ Retrying in {wait_time}s (extended backoff for possible rate limit)...", force=True)
                        logging.warning(f"Invalid API response (retry {retry_count}/{max_retries}): {', '.join(error_details)} | Provider: {provider_name}")
                        
@@ -7867,8 +7989,38 @@ class AIAgent:
                        print(f"{self.log_prefix}     • Check ANTHROPIC_API_KEY in {_dhh}/.env for API keys or legacy token values")
                        print(f"{self.log_prefix}     • For API keys: verify at https://console.anthropic.com/settings/keys")
                        print(f"{self.log_prefix}     • For Claude Code: run 'claude /login' to refresh, then retry")
-                        print(f"{self.log_prefix}     • Clear stale keys: hermes config set ANTHROPIC_TOKEN \"\"")
-                        print(f"{self.log_prefix}     • Legacy cleanup: hermes config set ANTHROPIC_API_KEY \"\"")
+                        print(f"{self.log_prefix}     • Legacy cleanup: hermes config set ANTHROPIC_TOKEN \"\"")
+                        print(f"{self.log_prefix}     • Clear stale keys: hermes config set ANTHROPIC_API_KEY \"\"")
+
+                    # ── Thinking block signature recovery ─────────────────
+                    # Anthropic signs thinking blocks against the full turn
+                    # content.  Any upstream mutation (context compression,
+                    # session truncation, message merging) invalidates the
+                    # signature → HTTP 400.  Recovery: strip reasoning_details
+                    # from all messages so the next retry sends no thinking
+                    # blocks at all.  One-shot — don't retry infinitely.
+                    if (
+                        self.api_mode == "anthropic_messages"
+                        and status_code == 400
+                        and not thinking_sig_retry_attempted
+                    ):
+                        _err_msg_lower = str(api_error).lower()
+                        if "signature" in _err_msg_lower and "thinking" in _err_msg_lower:
+                            thinking_sig_retry_attempted = True
+                            for _m in messages:
+                                if isinstance(_m, dict):
+                                    _m.pop("reasoning_details", None)
+                            self._vprint(
+                                f"{self.log_prefix}⚠️  Thinking block signature invalid — "
+                                f"stripped all thinking blocks, retrying...",
+                                force=True,
+                            )
+                            logging.warning(
+                                "%sThinking block signature recovery: stripped "
+                                "reasoning_details from %d messages",
+                                self.log_prefix, len(messages),
+                            )
+                            continue

                    retry_count += 1
                    elapsed_time = time.time() - api_start_time
@@ -8351,7 +8503,7 @@ class AIAgent:
                                    _retry_after = min(int(_ra_raw), 120)  # Cap at 2 minutes
                                except (TypeError, ValueError):
                                    pass
-                    wait_time = _retry_after if _retry_after else min(2 ** retry_count, 60)
+                    wait_time = _retry_after if _retry_after else jittered_backoff(retry_count, base_delay=2.0, max_delay=60.0)
                    if is_rate_limited:
                        self._emit_status(f"⏱️ Rate limit reached. Waiting {wait_time}s before retry (attempt {retry_count + 1}/{max_retries})...")
                    else:
@@ -1276,6 +1276,258 @@ class TestRoleAlternation:
        assert [m["role"] for m in result] == ["user", "assistant", "user"]


+# ---------------------------------------------------------------------------
+# Thinking block signature management
+# ---------------------------------------------------------------------------
+
+
+class TestThinkingBlockSignatureManagement:
+    """Tests for the thinking block handling strategy:
+    strip from old turns, preserve latest signed, downgrade unsigned."""
+
+    def test_thinking_stripped_from_non_last_assistant(self):
+        """Thinking blocks are removed from all assistant messages except the last."""
+        messages = [
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {"id": "tc_1", "function": {"name": "tool1", "arguments": "{}"}},
+                ],
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "Old reasoning.", "signature": "sig_old"},
+                ],
+            },
+            {"role": "tool", "tool_call_id": "tc_1", "content": "result 1"},
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {"id": "tc_2", "function": {"name": "tool2", "arguments": "{}"}},
+                ],
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "Latest reasoning.", "signature": "sig_new"},
+                ],
+            },
+            {"role": "tool", "tool_call_id": "tc_2", "content": "result 2"},
+        ]
+        _, result = convert_messages_to_anthropic(messages)
+
+        # Find both assistant messages
+        assistants = [m for m in result if m["role"] == "assistant"]
+        assert len(assistants) == 2
+
+        # First (non-last) assistant: no thinking blocks
+        first_types = [b.get("type") for b in assistants[0]["content"]]
+        assert "thinking" not in first_types
+        assert "redacted_thinking" not in first_types
+        assert "tool_use" in first_types  # tool_use should survive
+
+        # Last assistant: thinking block preserved with signature
+        last_blocks = assistants[1]["content"]
+        thinking_blocks = [b for b in last_blocks if b.get("type") == "thinking"]
+        assert len(thinking_blocks) == 1
+        assert thinking_blocks[0]["thinking"] == "Latest reasoning."
+        assert thinking_blocks[0]["signature"] == "sig_new"
+
+    def test_signed_thinking_preserved_on_last_turn(self):
+        """A signed thinking block on the last assistant message is kept."""
+        messages = [
+            {
+                "role": "assistant",
+                "content": "The answer is 42.",
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "Deep thought.", "signature": "sig_valid"},
+                ],
+            },
+        ]
+        _, result = convert_messages_to_anthropic(messages)
+        blocks = result[0]["content"]
+        thinking = [b for b in blocks if b.get("type") == "thinking"]
+        assert len(thinking) == 1
+        assert thinking[0]["signature"] == "sig_valid"
+
+    def test_unsigned_thinking_downgraded_to_text_on_last_turn(self):
+        """Unsigned thinking blocks on the last turn become text blocks."""
+        messages = [
+            {
+                "role": "assistant",
+                "content": "Response text.",
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "Unsigned reasoning."},
+                    # No 'signature' field
+                ],
+            },
+        ]
+        _, result = convert_messages_to_anthropic(messages)
+        blocks = result[0]["content"]
+
+        # No thinking blocks should remain
+        assert not any(b.get("type") == "thinking" for b in blocks)
+        # The reasoning text should be preserved as a text block
+        text_contents = [b.get("text", "") for b in blocks if b.get("type") == "text"]
+        assert "Unsigned reasoning." in text_contents
+
+    def test_redacted_thinking_with_data_preserved(self):
+        """Redacted thinking with 'data' field is kept on last turn."""
+        messages = [
+            {
+                "role": "assistant",
+                "content": "Response.",
+                "reasoning_details": [
+                    {"type": "redacted_thinking", "data": "opaque_signature_data"},
+                ],
+            },
+        ]
+        _, result = convert_messages_to_anthropic(messages)
+        blocks = result[0]["content"]
+        redacted = [b for b in blocks if b.get("type") == "redacted_thinking"]
+        assert len(redacted) == 1
+        assert redacted[0]["data"] == "opaque_signature_data"
+
+    def test_redacted_thinking_without_data_dropped(self):
+        """Redacted thinking without 'data' is dropped — can't be validated."""
+        messages = [
+            {
+                "role": "assistant",
+                "content": "Response.",
+                "reasoning_details": [
+                    {"type": "redacted_thinking"},
+                    # No 'data' field
+                ],
+            },
+        ]
+        _, result = convert_messages_to_anthropic(messages)
+        blocks = result[0]["content"]
+        assert not any(b.get("type") == "redacted_thinking" for b in blocks)
+
+    def test_cache_control_stripped_from_thinking_blocks(self):
+        """cache_control markers are removed from thinking/redacted_thinking blocks."""
+        messages = [
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {"id": "tc_1", "function": {"name": "t", "arguments": "{}"}},
+                ],
+                "reasoning_details": [
+                    {
+                        "type": "thinking",
+                        "thinking": "Reasoning.",
+                        "signature": "sig_1",
+                        "cache_control": {"type": "ephemeral"},
+                    },
+                ],
+            },
+            {"role": "tool", "tool_call_id": "tc_1", "content": "result"},
+        ]
+        _, result = convert_messages_to_anthropic(messages)
+        assistant = next(m for m in result if m["role"] == "assistant")
+        for block in assistant["content"]:
+            if block.get("type") in ("thinking", "redacted_thinking"):
+                assert "cache_control" not in block
+
+    def test_thinking_stripped_from_merged_consecutive_assistants(self):
+        """When consecutive assistants are merged, second one's thinking is dropped."""
+        messages = [
+            {
+                "role": "assistant",
+                "content": "First response.",
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "First thought.", "signature": "sig_1"},
+                ],
+            },
+            {
+                "role": "assistant",
+                "content": "Second response.",
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "Second thought.", "signature": "sig_2"},
+                ],
+            },
+        ]
+        _, result = convert_messages_to_anthropic(messages)
+
+        # Should be merged into one assistant message
+        assistants = [m for m in result if m["role"] == "assistant"]
+        assert len(assistants) == 1
+
+        # Only the first thinking block should remain (signed, on the last/only assistant)
+        blocks = assistants[0]["content"]
+        thinking = [b for b in blocks if b.get("type") == "thinking"]
+        assert len(thinking) == 1
+        assert thinking[0]["thinking"] == "First thought."
+
+    def test_empty_content_after_strip_gets_placeholder(self):
+        """If stripping thinking leaves an empty message, a placeholder is added."""
+        messages = [
+            {
+                "role": "assistant",
+                "content": "",
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "Only thinking, no text."},
+                    # Unsigned — will be downgraded, but content was empty string
+                ],
+            },
+            {"role": "user", "content": "Next message."},
+            {"role": "assistant", "content": "Final."},
+        ]
+        _, result = convert_messages_to_anthropic(messages)
+        # First assistant is non-last, so thinking is stripped completely.
+        # The original content was empty and thinking was unsigned → placeholder
+        first_assistant = result[0]
+        assert first_assistant["role"] == "assistant"
+        assert len(first_assistant["content"]) >= 1
+
+    def test_multi_turn_conversation_preserves_only_last(self):
+        """Full multi-turn conversation: only last assistant keeps thinking."""
+        messages = [
+            {"role": "user", "content": "Question 1"},
+            {
+                "role": "assistant",
+                "content": "Answer 1",
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "Thought 1", "signature": "sig_1"},
+                ],
+            },
+            {"role": "user", "content": "Question 2"},
+            {
+                "role": "assistant",
+                "content": "Answer 2",
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "Thought 2", "signature": "sig_2"},
+                ],
+            },
+            {"role": "user", "content": "Question 3"},
+            {
+                "role": "assistant",
+                "content": "Answer 3",
+                "reasoning_details": [
+                    {"type": "thinking", "thinking": "Thought 3", "signature": "sig_3"},
+                ],
+            },
+        ]
+        _, result = convert_messages_to_anthropic(messages)
+
+        assistants = [m for m in result if m["role"] == "assistant"]
+        assert len(assistants) == 3
+
+        # First two: no thinking blocks
+        for a in assistants[:2]:
+            assert not any(
+                b.get("type") in ("thinking", "redacted_thinking")
+                for b in a["content"]
+                if isinstance(b, dict)
+            )
+
+        # Last one: thinking preserved
+        last_thinking = [
+            b for b in assistants[2]["content"]
+            if isinstance(b, dict) and b.get("type") == "thinking"
+        ]
+        assert len(last_thinking) == 1
+        assert last_thinking[0]["signature"] == "sig_3"
+
+
 # ---------------------------------------------------------------------------
 # Tool choice
 # ---------------------------------------------------------------------------
@@ -471,6 +471,23 @@ class TestExplicitProviderRouting:
            client, model = resolve_provider_client("zai")
            assert client is not None

+    def test_explicit_google_alias_uses_gemini_credentials(self):
+        """provider='google' should route through the gemini API-key provider."""
+        with (
+            patch("hermes_cli.auth.resolve_api_key_provider_credentials", return_value={
+                "api_key": "gemini-key",
+                "base_url": "https://generativelanguage.googleapis.com/v1beta/openai",
+            }),
+            patch("agent.auxiliary_client.OpenAI") as mock_openai,
+        ):
+            mock_openai.return_value = MagicMock()
+            client, model = resolve_provider_client("google", model="gemini-3.1-pro-preview")
+
+        assert client is not None
+        assert model == "gemini-3.1-pro-preview"
+        assert mock_openai.call_args.kwargs["api_key"] == "gemini-key"
+        assert mock_openai.call_args.kwargs["base_url"] == "https://generativelanguage.googleapis.com/v1beta/openai"
+
    def test_explicit_unknown_returns_none(self, monkeypatch):
        """Unknown provider should return None."""
        client, model = resolve_provider_client("nonexistent-provider")
@@ -624,12 +641,15 @@ class TestVisionClientFallback:
        assert client is None
        assert model is None

-    def test_vision_auto_includes_anthropic_when_configured(self, monkeypatch):
-        monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-ant-api03-key")
+    def test_vision_auto_includes_active_provider_when_configured(self, monkeypatch):
+        """Active provider appears in available backends when credentials exist."""
+        monkeypatch.setenv("ANTHROPIC_API_KEY", "***")
        with (
            patch("agent.auxiliary_client._read_nous_auth", return_value=None),
+            patch("agent.auxiliary_client._read_main_provider", return_value="anthropic"),
+            patch("agent.auxiliary_client._read_main_model", return_value="claude-sonnet-4"),
            patch("agent.anthropic_adapter.build_anthropic_client", return_value=MagicMock()),
-            patch("agent.anthropic_adapter.resolve_anthropic_token", return_value="sk-ant-api03-key"),
+            patch("agent.anthropic_adapter.resolve_anthropic_token", return_value="***"),
        ):
            backends = get_available_vision_backends()

@@ -702,88 +722,51 @@ class TestAuxiliaryPoolAwareness:
        assert call_kwargs["base_url"] == "https://api.githubcopilot.com"
        assert call_kwargs["default_headers"]["Editor-Version"]

-    def test_vision_auto_uses_anthropic_when_no_higher_priority_backend(self, monkeypatch):
-        monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-ant-api03-key")
+    def test_vision_auto_uses_active_provider_as_fallback(self, monkeypatch):
+        """When no OpenRouter/Nous available, vision auto falls back to active provider."""
+        monkeypatch.setenv("ANTHROPIC_API_KEY", "***")
        with (
            patch("agent.auxiliary_client._read_nous_auth", return_value=None),
+            patch("agent.auxiliary_client._read_main_provider", return_value="anthropic"),
+            patch("agent.auxiliary_client._read_main_model", return_value="claude-sonnet-4"),
            patch("agent.anthropic_adapter.build_anthropic_client", return_value=MagicMock()),
-            patch("agent.anthropic_adapter.resolve_anthropic_token", return_value="sk-ant-api03-key"),
+            patch("agent.anthropic_adapter.resolve_anthropic_token", return_value="***"),
        ):
            client, model = get_vision_auxiliary_client()

        assert client is not None
        assert client.__class__.__name__ == "AnthropicAuxiliaryClient"
-        assert model == "claude-haiku-4-5-20251001"

-    def test_selected_anthropic_provider_is_preferred_for_vision_auto(self, monkeypatch):
+    def test_vision_auto_prefers_active_provider_over_openrouter(self, monkeypatch):
+        """Active provider is tried before OpenRouter in vision auto."""
        monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")
-        monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-ant-api03-key")
-
-        def fake_load_config():
-            return {"model": {"provider": "anthropic", "default": "claude-sonnet-4-6"}}
+        monkeypatch.setenv("ANTHROPIC_API_KEY", "***")

        with (
            patch("agent.auxiliary_client._read_nous_auth", return_value=None),
+            patch("agent.auxiliary_client._read_main_provider", return_value="anthropic"),
+            patch("agent.auxiliary_client._read_main_model", return_value="claude-sonnet-4"),
            patch("agent.anthropic_adapter.build_anthropic_client", return_value=MagicMock()),
-            patch("agent.anthropic_adapter.resolve_anthropic_token", return_value="sk-ant-api03-key"),
-            patch("agent.auxiliary_client.OpenAI") as mock_openai,
-            patch("hermes_cli.config.load_config", fake_load_config),
-        ):
-            client, model = get_vision_auxiliary_client()
-
-        assert client is not None
-        assert client.__class__.__name__ == "AnthropicAuxiliaryClient"
-        assert model == "claude-haiku-4-5-20251001"
-
-    def test_selected_codex_provider_short_circuits_vision_auto(self, monkeypatch):
-        def fake_load_config():
-            return {"model": {"provider": "openai-codex", "default": "gpt-5.2-codex"}}
-
-        codex_client = MagicMock()
-        with (
-            patch("hermes_cli.config.load_config", fake_load_config),
-            patch("agent.auxiliary_client._try_codex", return_value=(codex_client, "gpt-5.2-codex")) as mock_codex,
-            patch("agent.auxiliary_client._try_openrouter") as mock_openrouter,
-            patch("agent.auxiliary_client._try_nous") as mock_nous,
-            patch("agent.auxiliary_client._try_anthropic") as mock_anthropic,
-            patch("agent.auxiliary_client._try_custom_endpoint") as mock_custom,
+            patch("agent.anthropic_adapter.resolve_anthropic_token", return_value="***"),
        ):
            provider, client, model = resolve_vision_provider_client()

-        assert provider == "openai-codex"
-        assert client is codex_client
-        assert model == "gpt-5.2-codex"
-        mock_codex.assert_called_once()
-        mock_openrouter.assert_not_called()
-        mock_nous.assert_not_called()
-        mock_anthropic.assert_not_called()
-        mock_custom.assert_not_called()
+        # Active provider should win over OpenRouter
+        assert provider == "anthropic"

-    def test_vision_auto_includes_codex(self, codex_auth_dir):
-        """Codex supports vision (gpt-5.3-codex), so auto mode should use it."""
-        with patch("agent.auxiliary_client._read_nous_auth", return_value=None), \
-             patch("agent.auxiliary_client.OpenAI"):
-            client, model = get_vision_auxiliary_client()
-        from agent.auxiliary_client import CodexAuxiliaryClient
-        assert isinstance(client, CodexAuxiliaryClient)
-        assert model == "gpt-5.2-codex"
-
-    def test_vision_auto_falls_back_to_custom_endpoint(self, monkeypatch):
-        """Custom endpoint is used as fallback in vision auto mode.
-
-        Many local models (Qwen-VL, LLaVA, etc.) support vision.
-        When no OpenRouter/Nous/Codex is available, try the custom endpoint.
-        """
+    def test_vision_auto_uses_named_custom_as_active_provider(self, monkeypatch):
+        """Named custom provider works as active provider fallback in vision auto."""
        monkeypatch.delenv("OPENROUTER_API_KEY", raising=False)
        monkeypatch.delenv("ANTHROPIC_API_KEY", raising=False)
        with patch("agent.auxiliary_client._read_nous_auth", return_value=None), \
             patch("agent.auxiliary_client._select_pool_entry", return_value=(False, None)), \
-             patch("agent.auxiliary_client._read_codex_access_token", return_value=None), \
-             patch("agent.auxiliary_client._resolve_custom_runtime",
-                   return_value=("http://localhost:1234/v1", "local-key")), \
-             patch("agent.auxiliary_client.OpenAI") as mock_openai:
-            client, model = get_vision_auxiliary_client()
-        assert client is not None  # Custom endpoint picked up as fallback
+             patch("agent.auxiliary_client._read_main_provider", return_value="custom:local"), \
+             patch("agent.auxiliary_client._read_main_model", return_value="my-local-model"), \
+             patch("agent.auxiliary_client.resolve_provider_client",
+                   return_value=(MagicMock(), "my-local-model")) as mock_resolve:
+            provider, client, model = resolve_vision_provider_client()
+        assert client is not None
+        assert provider == "custom:local"

    def test_vision_direct_endpoint_override(self, monkeypatch):
        monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")
@@ -822,6 +805,31 @@ class TestAuxiliaryPoolAwareness:
        assert model == "google/gemini-3-flash-preview"
        assert client is not None

+    def test_vision_config_google_provider_uses_gemini_credentials(self, monkeypatch):
+        config = {
+            "auxiliary": {
+                "vision": {
+                    "provider": "google",
+                    "model": "gemini-3.1-pro-preview",
+                }
+            }
+        }
+        monkeypatch.setattr("hermes_cli.config.load_config", lambda: config)
+        with (
+            patch("hermes_cli.auth.resolve_api_key_provider_credentials", return_value={
+                "api_key": "gemini-key",
+                "base_url": "https://generativelanguage.googleapis.com/v1beta/openai",
+            }),
+            patch("agent.auxiliary_client.OpenAI") as mock_openai,
+        ):
+            resolved_provider, client, model = resolve_vision_provider_client()
+
+        assert resolved_provider == "gemini"
+        assert client is not None
+        assert model == "gemini-3.1-pro-preview"
+        assert mock_openai.call_args.kwargs["api_key"] == "gemini-key"
+        assert mock_openai.call_args.kwargs["base_url"] == "https://generativelanguage.googleapis.com/v1beta/openai"
+
    def test_vision_forced_main_uses_custom_endpoint(self, monkeypatch):
        """When explicitly forced to 'main', vision CAN use custom endpoint."""
        config = {
@@ -846,7 +854,14 @@ class TestAuxiliaryPoolAwareness:
        monkeypatch.setenv("AUXILIARY_VISION_PROVIDER", "main")
        monkeypatch.delenv("OPENAI_BASE_URL", raising=False)
        monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+        # Clear client cache to avoid stale entries from previous tests
+        from agent.auxiliary_client import _client_cache
+        _client_cache.clear()
        with patch("agent.auxiliary_client._read_nous_auth", return_value=None), \
+             patch("agent.auxiliary_client._read_main_provider", return_value=""), \
+             patch("agent.auxiliary_client._read_main_model", return_value=""), \
+             patch("agent.auxiliary_client._select_pool_entry", return_value=(False, None)), \
+             patch("agent.auxiliary_client._resolve_custom_runtime", return_value=(None, None)), \
             patch("agent.auxiliary_client._read_codex_access_token", return_value=None), \
             patch("agent.auxiliary_client._resolve_api_key_provider", return_value=(None, None)):
            client, model = get_vision_auxiliary_client()
@@ -13,7 +13,7 @@ from unittest.mock import patch, MagicMock
 import pytest
 import yaml

-sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))


 def _run_auxiliary_bridge(config_dict, monkeypatch):
@@ -199,7 +199,7 @@ class TestGatewayBridgeCodeParity:

    def test_gateway_has_auxiliary_bridge(self):
        """The gateway config bridge must include auxiliary.* bridging."""
-        gateway_path = Path(__file__).parent.parent / "gateway" / "run.py"
+        gateway_path = Path(__file__).parent.parent.parent / "gateway" / "run.py"
        content = gateway_path.read_text()
        # Check for key patterns that indicate the bridge is present
        assert "AUXILIARY_VISION_PROVIDER" in content
@@ -213,7 +213,7 @@ class TestGatewayBridgeCodeParity:

    def test_gateway_no_compression_env_bridge(self):
        """Gateway should NOT bridge compression config to env vars (config-only)."""
-        gateway_path = Path(__file__).parent.parent / "gateway" / "run.py"
+        gateway_path = Path(__file__).parent.parent.parent / "gateway" / "run.py"
        content = gateway_path.read_text()
        assert "CONTEXT_COMPRESSION_PROVIDER" not in content
        assert "CONTEXT_COMPRESSION_MODEL" not in content
@@ -0,0 +1,151 @@
+"""Tests for named custom provider and 'main' alias resolution in auxiliary_client."""
+
+import os
+from unittest.mock import patch, MagicMock
+
+import pytest
+
+
+@pytest.fixture(autouse=True)
+def _isolate(tmp_path, monkeypatch):
+    """Redirect HERMES_HOME and clear module caches."""
+    hermes_home = tmp_path / ".hermes"
+    hermes_home.mkdir()
+    monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+    # Write a minimal config so load_config doesn't fail
+    (hermes_home / "config.yaml").write_text("model:\n  default: test-model\n")
+
+
+def _write_config(tmp_path, config_dict):
+    """Write a config.yaml to the test HERMES_HOME."""
+    import yaml
+    config_path = tmp_path / ".hermes" / "config.yaml"
+    config_path.write_text(yaml.dump(config_dict))
+
+
+class TestNormalizeVisionProvider:
+    """_normalize_vision_provider should resolve 'main' to actual main provider."""
+
+    def test_main_resolves_to_named_custom(self, tmp_path):
+        _write_config(tmp_path, {
+            "model": {"default": "my-model", "provider": "custom:beans"},
+            "custom_providers": [{"name": "beans", "base_url": "http://localhost/v1"}],
+        })
+        from agent.auxiliary_client import _normalize_vision_provider
+        assert _normalize_vision_provider("main") == "custom:beans"
+
+    def test_main_resolves_to_openrouter(self, tmp_path):
+        _write_config(tmp_path, {
+            "model": {"default": "anthropic/claude-sonnet-4", "provider": "openrouter"},
+        })
+        from agent.auxiliary_client import _normalize_vision_provider
+        assert _normalize_vision_provider("main") == "openrouter"
+
+    def test_main_resolves_to_deepseek(self, tmp_path):
+        _write_config(tmp_path, {
+            "model": {"default": "deepseek-chat", "provider": "deepseek"},
+        })
+        from agent.auxiliary_client import _normalize_vision_provider
+        assert _normalize_vision_provider("main") == "deepseek"
+
+    def test_main_falls_back_to_custom_when_no_provider(self, tmp_path):
+        _write_config(tmp_path, {"model": {"default": "gpt-4o"}})
+        from agent.auxiliary_client import _normalize_vision_provider
+        assert _normalize_vision_provider("main") == "custom"
+
+    def test_bare_provider_name_unchanged(self):
+        from agent.auxiliary_client import _normalize_vision_provider
+        assert _normalize_vision_provider("beans") == "beans"
+        assert _normalize_vision_provider("deepseek") == "deepseek"
+
+    def test_codex_alias_still_works(self):
+        from agent.auxiliary_client import _normalize_vision_provider
+        assert _normalize_vision_provider("codex") == "openai-codex"
+
+    def test_auto_unchanged(self):
+        from agent.auxiliary_client import _normalize_vision_provider
+        assert _normalize_vision_provider("auto") == "auto"
+        assert _normalize_vision_provider(None) == "auto"
+
+
+class TestResolveProviderClientMainAlias:
+    """resolve_provider_client('main', ...) should resolve to actual main provider."""
+
+    def test_main_resolves_to_named_custom_provider(self, tmp_path):
+        _write_config(tmp_path, {
+            "model": {"default": "my-model", "provider": "beans"},
+            "custom_providers": [
+                {"name": "beans", "base_url": "http://beans.local/v1", "api_key": "k"},
+            ],
+        })
+        from agent.auxiliary_client import resolve_provider_client
+        client, model = resolve_provider_client("main", "override-model")
+        assert client is not None
+        assert model == "override-model"
+        assert "beans.local" in str(client.base_url)
+
+    def test_main_with_custom_colon_prefix(self, tmp_path):
+        _write_config(tmp_path, {
+            "model": {"default": "my-model", "provider": "custom:beans"},
+            "custom_providers": [
+                {"name": "beans", "base_url": "http://beans.local/v1", "api_key": "k"},
+            ],
+        })
+        from agent.auxiliary_client import resolve_provider_client
+        client, model = resolve_provider_client("main", "test")
+        assert client is not None
+        assert "beans.local" in str(client.base_url)
+
+
+class TestResolveProviderClientNamedCustom:
+    """resolve_provider_client should resolve named custom providers directly."""
+
+    def test_named_custom_provider(self, tmp_path):
+        _write_config(tmp_path, {
+            "model": {"default": "test-model"},
+            "custom_providers": [
+                {"name": "beans", "base_url": "http://beans.local/v1", "api_key": "k"},
+            ],
+        })
+        from agent.auxiliary_client import resolve_provider_client
+        client, model = resolve_provider_client("beans", "my-model")
+        assert client is not None
+        assert model == "my-model"
+        assert "beans.local" in str(client.base_url)
+
+    def test_named_custom_provider_default_model(self, tmp_path):
+        _write_config(tmp_path, {
+            "model": {"default": "main-model"},
+            "custom_providers": [
+                {"name": "beans", "base_url": "http://beans.local/v1", "api_key": "k"},
+            ],
+        })
+        from agent.auxiliary_client import resolve_provider_client
+        client, model = resolve_provider_client("beans")
+        assert client is not None
+        # Should use _read_main_model() fallback
+        assert model == "main-model"
+
+    def test_named_custom_no_api_key_uses_fallback(self, tmp_path):
+        _write_config(tmp_path, {
+            "model": {"default": "test"},
+            "custom_providers": [
+                {"name": "local", "base_url": "http://localhost:8080/v1"},
+            ],
+        })
+        from agent.auxiliary_client import resolve_provider_client
+        client, model = resolve_provider_client("local", "test")
+        assert client is not None
+        # no-key-required should be used
+
+    def test_nonexistent_named_custom_falls_through(self, tmp_path):
+        _write_config(tmp_path, {
+            "model": {"default": "test"},
+            "custom_providers": [
+                {"name": "beans", "base_url": "http://beans.local/v1"},
+            ],
+        })
+        from agent.auxiliary_client import resolve_provider_client
+        # "coffee" doesn't exist in custom_providers
+        client, model = resolve_provider_client("coffee", "test")
+        assert client is None
@@ -0,0 +1,42 @@
+"""Tests for MiniMax auxiliary client URL normalization.
+
+MiniMax and MiniMax-CN set inference_base_url to the /anthropic path.
+The auxiliary client uses the OpenAI SDK, which needs /v1 instead.
+"""
+
+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+from agent.auxiliary_client import _to_openai_base_url
+
+
+class TestToOpenaiBaseUrl:
+    def test_minimax_global_anthropic_suffix_replaced(self):
+        assert _to_openai_base_url("https://api.minimax.io/anthropic") == "https://api.minimax.io/v1"
+
+    def test_minimax_cn_anthropic_suffix_replaced(self):
+        assert _to_openai_base_url("https://api.minimaxi.com/anthropic") == "https://api.minimaxi.com/v1"
+
+    def test_trailing_slash_stripped_before_replace(self):
+        assert _to_openai_base_url("https://api.minimax.io/anthropic/") == "https://api.minimax.io/v1"
+
+    def test_v1_url_unchanged(self):
+        assert _to_openai_base_url("https://api.openai.com/v1") == "https://api.openai.com/v1"
+
+    def test_openrouter_url_unchanged(self):
+        assert _to_openai_base_url("https://openrouter.ai/api/v1") == "https://openrouter.ai/api/v1"
+
+    def test_anthropic_domain_unchanged(self):
+        """api.anthropic.com doesn't end with /anthropic — should be untouched."""
+        assert _to_openai_base_url("https://api.anthropic.com") == "https://api.anthropic.com"
+
+    def test_anthropic_in_subpath_unchanged(self):
+        assert _to_openai_base_url("https://example.com/anthropic/extra") == "https://example.com/anthropic/extra"
+
+    def test_empty_string(self):
+        assert _to_openai_base_url("") == ""
+
+    def test_none(self):
+        assert _to_openai_base_url(None) == ""
@@ -0,0 +1,105 @@
+"""Tests for MiniMax provider hardening — context lengths, thinking guard, catalog."""
+
+
+class TestMinimaxContextLengths:
+    """Verify per-model context length entries for MiniMax models."""
+
+    def test_m1_variants_have_1m_context(self):
+        from agent.model_metadata import DEFAULT_CONTEXT_LENGTHS
+        # Keys are lowercase because the lookup lowercases model names
+        for model in ("minimax-m1", "minimax-m1-40k", "minimax-m1-80k",
+                       "minimax-m1-128k", "minimax-m1-256k"):
+            assert model in DEFAULT_CONTEXT_LENGTHS, f"{model} missing from context lengths"
+            assert DEFAULT_CONTEXT_LENGTHS[model] == 1_000_000, f"{model} expected 1M"
+
+    def test_m2_variants_have_1m_context(self):
+        from agent.model_metadata import DEFAULT_CONTEXT_LENGTHS
+        # Keys are lowercase because the lookup lowercases model names
+        for model in ("minimax-m2.5", "minimax-m2.7"):
+            assert model in DEFAULT_CONTEXT_LENGTHS, f"{model} missing from context lengths"
+            assert DEFAULT_CONTEXT_LENGTHS[model] == 1_048_576, f"{model} expected 1048576"
+
+    def test_minimax_prefix_fallback(self):
+        from agent.model_metadata import DEFAULT_CONTEXT_LENGTHS
+        # The generic "minimax" prefix entry should be 1M for unknown models
+        assert DEFAULT_CONTEXT_LENGTHS["minimax"] == 1_048_576
+
+
+
+class TestMinimaxThinkingGuard:
+    """Verify that build_anthropic_kwargs does NOT add thinking params for MiniMax models."""
+
+    def test_no_thinking_for_minimax_m27(self):
+        from agent.anthropic_adapter import build_anthropic_kwargs
+        kwargs = build_anthropic_kwargs(
+            model="MiniMax-M2.7",
+            messages=[{"role": "user", "content": "hello"}],
+            tools=None,
+            max_tokens=4096,
+            reasoning_config={"enabled": True, "effort": "medium"},
+        )
+        assert "thinking" not in kwargs
+        assert "output_config" not in kwargs
+
+    def test_no_thinking_for_minimax_m1(self):
+        from agent.anthropic_adapter import build_anthropic_kwargs
+        kwargs = build_anthropic_kwargs(
+            model="MiniMax-M1-128k",
+            messages=[{"role": "user", "content": "hello"}],
+            tools=None,
+            max_tokens=4096,
+            reasoning_config={"enabled": True, "effort": "high"},
+        )
+        assert "thinking" not in kwargs
+
+    def test_thinking_still_works_for_claude(self):
+        from agent.anthropic_adapter import build_anthropic_kwargs
+        kwargs = build_anthropic_kwargs(
+            model="claude-sonnet-4-20250514",
+            messages=[{"role": "user", "content": "hello"}],
+            tools=None,
+            max_tokens=4096,
+            reasoning_config={"enabled": True, "effort": "medium"},
+        )
+        assert "thinking" in kwargs
+
+
+class TestMinimaxAuxModel:
+    """Verify auxiliary model is standard (not highspeed)."""
+
+    def test_minimax_aux_is_standard(self):
+        from agent.auxiliary_client import _API_KEY_PROVIDER_AUX_MODELS
+        assert _API_KEY_PROVIDER_AUX_MODELS["minimax"] == "MiniMax-M2.7"
+        assert _API_KEY_PROVIDER_AUX_MODELS["minimax-cn"] == "MiniMax-M2.7"
+
+    def test_minimax_aux_not_highspeed(self):
+        from agent.auxiliary_client import _API_KEY_PROVIDER_AUX_MODELS
+        assert "highspeed" not in _API_KEY_PROVIDER_AUX_MODELS["minimax"]
+        assert "highspeed" not in _API_KEY_PROVIDER_AUX_MODELS["minimax-cn"]
+
+
+class TestMinimaxModelCatalog:
+    """Verify the model catalog includes M1 family and excludes deprecated models."""
+
+    def test_catalog_includes_m1_family(self):
+        from hermes_cli.models import _PROVIDER_MODELS
+        for provider in ("minimax", "minimax-cn"):
+            models = _PROVIDER_MODELS[provider]
+            assert "MiniMax-M1" in models
+            assert "MiniMax-M1-40k" in models
+            assert "MiniMax-M1-80k" in models
+            assert "MiniMax-M1-128k" in models
+            assert "MiniMax-M1-256k" in models
+
+    def test_catalog_excludes_deprecated(self):
+        from hermes_cli.models import _PROVIDER_MODELS
+        for provider in ("minimax", "minimax-cn"):
+            models = _PROVIDER_MODELS[provider]
+            assert "MiniMax-M2.1" not in models
+
+    def test_catalog_excludes_highspeed(self):
+        from hermes_cli.models import _PROVIDER_MODELS
+        for provider in ("minimax", "minimax-cn"):
+            models = _PROVIDER_MODELS[provider]
+            assert "MiniMax-M2.7-highspeed" not in models
+            assert "MiniMax-M2.5-highspeed" not in models
@@ -330,7 +330,7 @@ def test_model_flow_nous_prints_subscription_guidance_without_mutating_explicit_
        "hermes_cli.auth.fetch_nous_models",
        lambda *args, **kwargs: ["claude-opus-4-6"],
    )
-    monkeypatch.setattr("hermes_cli.auth._prompt_model_selection", lambda model_ids, current_model="", pricing=None: "claude-opus-4-6")
+    monkeypatch.setattr("hermes_cli.auth._prompt_model_selection", lambda model_ids, current_model="", pricing=None, **kw: "claude-opus-4-6")
    monkeypatch.setattr("hermes_cli.auth._save_model_choice", lambda model: None)
    monkeypatch.setattr("hermes_cli.auth._update_config_for_provider", lambda provider, url: None)
    monkeypatch.setattr(
@@ -368,7 +368,7 @@ def test_model_flow_nous_applies_managed_tts_default_when_unconfigured(monkeypat
        "hermes_cli.auth.fetch_nous_models",
        lambda *args, **kwargs: ["claude-opus-4-6"],
    )
-    monkeypatch.setattr("hermes_cli.auth._prompt_model_selection", lambda model_ids, current_model="", pricing=None: "claude-opus-4-6")
+    monkeypatch.setattr("hermes_cli.auth._prompt_model_selection", lambda model_ids, current_model="", pricing=None, **kw: "claude-opus-4-6")
    monkeypatch.setattr("hermes_cli.auth._save_model_choice", lambda model: None)
    monkeypatch.setattr("hermes_cli.auth._update_config_for_provider", lambda provider, url: None)
    monkeypatch.setattr(
@@ -1,6 +1,6 @@
 """Regression tests for CLI /retry history replacement semantics."""

-from tests.test_cli_init import _make_cli
+from tests.cli.test_cli_init import _make_cli


 def test_retry_last_truncates_history_before_requeueing_message():
@@ -0,0 +1,98 @@
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+
+from cli import HermesCLI, _rich_text_from_ansi
+from hermes_cli.skin_engine import get_active_skin, set_active_skin
+
+
+def _make_cli_stub():
+    cli = HermesCLI.__new__(HermesCLI)
+    cli._sudo_state = None
+    cli._secret_state = None
+    cli._approval_state = None
+    cli._clarify_state = None
+    cli._clarify_freetext = False
+    cli._command_running = False
+    cli._agent_running = False
+    cli._voice_recording = False
+    cli._voice_processing = False
+    cli._voice_mode = False
+    cli._command_spinner_frame = lambda: "⟳"
+    cli._tui_style_base = {
+        "prompt": "#fff",
+        "input-area": "#fff",
+        "input-rule": "#aaa",
+        "prompt-working": "#888 italic",
+    }
+    cli._app = SimpleNamespace(style=None)
+    cli._invalidate = MagicMock()
+    return cli
+
+
+class TestCliSkinPromptIntegration:
+    def test_default_prompt_fragments_use_default_symbol(self):
+        cli = _make_cli_stub()
+
+        set_active_skin("default")
+        assert cli._get_tui_prompt_fragments() == [("class:prompt", "❯ ")]
+
+    def test_ares_prompt_fragments_use_skin_symbol(self):
+        cli = _make_cli_stub()
+
+        set_active_skin("ares")
+        assert cli._get_tui_prompt_fragments() == [("class:prompt", "⚔ ❯ ")]
+
+    def test_secret_prompt_fragments_preserve_secret_state(self):
+        cli = _make_cli_stub()
+        cli._secret_state = {"response_queue": object()}
+
+        set_active_skin("ares")
+        assert cli._get_tui_prompt_fragments() == [("class:sudo-prompt", "🔑 ❯ ")]
+
+    def test_icon_only_skin_symbol_still_visible_in_special_states(self):
+        cli = _make_cli_stub()
+        cli._secret_state = {"response_queue": object()}
+
+        with patch("hermes_cli.skin_engine.get_active_prompt_symbol", return_value="⚔ "):
+            assert cli._get_tui_prompt_fragments() == [("class:sudo-prompt", "🔑 ⚔ ")]
+
+    def test_build_tui_style_dict_uses_skin_overrides(self):
+        cli = _make_cli_stub()
+
+        set_active_skin("ares")
+        skin = get_active_skin()
+        style_dict = cli._build_tui_style_dict()
+
+        assert style_dict["prompt"] == skin.get_color("prompt")
+        assert style_dict["input-rule"] == skin.get_color("input_rule")
+        assert style_dict["prompt-working"] == f"{skin.get_color('banner_dim')} italic"
+        assert style_dict["approval-title"] == f"{skin.get_color('ui_warn')} bold"
+
+    def test_apply_tui_skin_style_updates_running_app(self):
+        cli = _make_cli_stub()
+
+        set_active_skin("ares")
+        assert cli._apply_tui_skin_style() is True
+        assert cli._app.style is not None
+        cli._invalidate.assert_called_once_with(min_interval=0.0)
+
+    def test_handle_skin_command_refreshes_live_tui(self, capsys):
+        cli = _make_cli_stub()
+
+        with patch("cli.save_config_value", return_value=True):
+            cli._handle_skin_command("/skin ares")
+
+        output = capsys.readouterr().out
+        assert "Skin set to: ares (saved)" in output
+        assert "Prompt + TUI colors updated." in output
+        assert cli._app.style is not None
+
+
+class TestAnsiRichTextHelper:
+    def test_preserves_literal_brackets(self):
+        text = _rich_text_from_ansi("[notatag] literal")
+        assert text.plain == "[notatag] literal"
+
+    def test_strips_ansi_but_keeps_plain_text(self):
+        text = _rich_text_from_ansi("\x1b[31mred\x1b[0m")
+        assert text.plain == "red"
--- a/Show More
+++ b/Show More