fix(nix): build web dashboard frontend in Nix package

* Add setuptools build dep for legacy alibabacloud packages and updated
stale npm-deps hash

* Add HERMES_NODE env var to pin Node.js version

The TUI requires Node.js 20+ for regex `/v` flag support (used by
string-width). Instead of relying on PATH lookup, explicitly set
HERMES_NODE to the bundled Node 22 in the Nix wrapper, and add a
fallback check in the Python code to use HERMES_NODE if available.

Also upgrade container provisioning to Node 22 via NodeSource (Ubuntu
24.04 ships Node 18 which is EOL) and add a Nix check to verify the
wrapper and Node version at build time.

.envrc

@@ -1 +1,5 @@
+watch_file pyproject.toml uv.lock
+watch_file ui-tui/package-lock.json ui-tui/package.json
+watch_file flake.nix flake.lock nix/devShell.nix nix/tui.nix nix/package.nix nix/python.nix
+
 use flake

.gitignore

@@ -60,5 +60,6 @@ mini-swe-agent/
 
 # Nix
 .direnv/
+.nix-stamps/
 result
 website/static/api/skills-index.json

AGENTS.md

@@ -56,6 +56,19 @@ hermes-agent/
 │   ├── run.py            # Main loop, slash commands, message dispatch
 │   ├── session.py        # SessionStore — conversation persistence
 │   └── platforms/        # Adapters: telegram, discord, slack, whatsapp, homeassistant, signal, qqbot
+├── ui-tui/               # Ink (React) terminal UI — `hermes --tui`
+│   ├── src/entry.tsx        # TTY gate + render()
+│   ├── src/app.tsx          # Main state machine and UI
+│   ├── src/gatewayClient.ts # Child process + JSON-RPC bridge
+│   ├── src/app/             # Decomposed app logic (event handler, slash handler, stores, hooks)
+│   ├── src/components/      # Ink components (branding, markdown, prompts, pickers, etc.)
+│   ├── src/hooks/           # useCompletion, useInputHistory, useQueue, useVirtualHistory
+│   └── src/lib/             # Pure helpers (history, osc52, text, rpc, messages)
+├── tui_gateway/          # Python JSON-RPC backend for the TUI
+│   ├── entry.py             # stdio entrypoint
+│   ├── server.py            # RPC handlers and session logic
+│   ├── render.py            # Optional rich/ANSI bridge
+│   └── slash_worker.py      # Persistent HermesCLI subprocess for slash commands
 ├── acp_adapter/          # ACP server (VS Code / Zed / JetBrains integration)
 ├── cron/                 # Scheduler (jobs.py, scheduler.py)
 ├── environments/         # RL training environments (Atropos)
@@ -179,6 +192,59 @@ if canonical == "mycommand":
 
 ---
 
+## TUI Architecture (ui-tui + tui_gateway)
+
+The TUI is a full replacement for the classic (prompt_toolkit) CLI, activated via `hermes --tui` or `HERMES_TUI=1`.
+
+### Process Model
+
+```
+hermes --tui
+  └─ Node (Ink)  ──stdio JSON-RPC──  Python (tui_gateway)
+       │                                  └─ AIAgent + tools + sessions
+       └─ renders transcript, composer, prompts, activity
+```
+
+TypeScript owns the screen. Python owns sessions, tools, model calls, and slash command logic.
+
+### Transport
+
+Newline-delimited JSON-RPC over stdio. Requests from Ink, events from Python. See `tui_gateway/server.py` for the full method/event catalog.
+
+### Key Surfaces
+
+| Surface | Ink component | Gateway method |
+|---------|---------------|----------------|
+| Chat streaming | `app.tsx` + `messageLine.tsx` | `prompt.submit` → `message.delta/complete` |
+| Tool activity | `thinking.tsx` | `tool.start/progress/complete` |
+| Approvals | `prompts.tsx` | `approval.respond` ← `approval.request` |
+| Clarify/sudo/secret | `prompts.tsx`, `maskedPrompt.tsx` | `clarify/sudo/secret.respond` |
+| Session picker | `sessionPicker.tsx` | `session.list/resume` |
+| Slash commands | Local handler + fallthrough | `slash.exec` → `_SlashWorker`, `command.dispatch` |
+| Completions | `useCompletion` hook | `complete.slash`, `complete.path` |
+| Theming | `theme.ts` + `branding.tsx` | `gateway.ready` with skin data |
+
+### Slash Command Flow
+
+1. Built-in client commands (`/help`, `/quit`, `/clear`, `/resume`, `/copy`, `/paste`, etc.) handled locally in `app.tsx`
+2. Everything else → `slash.exec` (runs in persistent `_SlashWorker` subprocess) → `command.dispatch` fallback
+
+### Dev Commands
+
+```bash
+cd ui-tui
+npm install       # first time
+npm run dev       # watch mode (rebuilds hermes-ink + tsx --watch)
+npm start         # production
+npm run build     # full build (hermes-ink + tsc)
+npm run type-check # typecheck only (tsc --noEmit)
+npm run lint      # eslint
+npm run fmt       # prettier
+npm test          # vitest
+```
+
+---
+
 ## Adding New Tools
 
 Requires changes in **2 files**:

README.md

@@ -13,7 +13,7 @@
 
 **The self-improving AI agent built by [Nous Research](https://nousresearch.com).** It's the only agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions. Run it on a $5 VPS, a GPU cluster, or serverless infrastructure that costs nearly nothing when idle. It's not tied to your laptop — talk to it from Telegram while it works on a cloud VM.
 
-Use any model you want — [Nous Portal](https://portal.nousresearch.com), [OpenRouter](https://openrouter.ai) (200+ models), [Xiaomi MiMo](https://platform.xiaomimimo.com), [z.ai/GLM](https://z.ai), [Kimi/Moonshot](https://platform.moonshot.ai), [MiniMax](https://www.minimax.io), [Hugging Face](https://huggingface.co), OpenAI, or your own endpoint. Switch with `hermes model` — no code changes, no lock-in.
+Use any model you want — [Nous Portal](https://portal.nousresearch.com), [OpenRouter](https://openrouter.ai) (200+ models), [NVIDIA NIM](https://build.nvidia.com) (Nemotron), [Xiaomi MiMo](https://platform.xiaomimimo.com), [z.ai/GLM](https://z.ai), [Kimi/Moonshot](https://platform.moonshot.ai), [MiniMax](https://www.minimax.io), [Hugging Face](https://huggingface.co), OpenAI, or your own endpoint. Switch with `hermes model` — no code changes, no lock-in.
 
 <table>
 <tr><td><b>A real terminal interface</b></td><td>Full TUI with multiline editing, slash-command autocomplete, conversation history, interrupt-and-redirect, and streaming tool output.</td></tr>
@@ -141,11 +141,18 @@ See `hermes claw migrate --help` for all options, or use the `openclaw-migration
 
 We welcome contributions! See the [Contributing Guide](https://hermes-agent.nousresearch.com/docs/developer-guide/contributing) for development setup, code style, and PR process.
 
-Quick start for contributors:
+Quick start for contributors — clone and go with `setup-hermes.sh`:
 
 ```bash
 git clone https://github.com/NousResearch/hermes-agent.git
 cd hermes-agent
+./setup-hermes.sh     # installs uv, creates venv, installs .[all], symlinks ~/.local/bin/hermes
+./hermes              # auto-detects the venv, no need to `source` first
+```
+
+Manual path (equivalent to the above):
+
+```bash
 curl -LsSf https://astral.sh/uv/install.sh | sh
 uv venv venv --python 3.11
 source venv/bin/activate

acp_adapter/events.py

@@ -49,6 +49,7 @@ def make_tool_progress_cb(
     session_id: str,
     loop: asyncio.AbstractEventLoop,
     tool_call_ids: Dict[str, Deque[str]],
+    tool_call_meta: Dict[str, Dict[str, Any]],
 ) -> Callable:
     """Create a ``tool_progress_callback`` for AIAgent.
 
@@ -84,6 +85,16 @@ def make_tool_progress_cb(
             tool_call_ids[name] = queue
         queue.append(tc_id)
 
+        snapshot = None
+        if name in {"write_file", "patch", "skill_manage"}:
+            try:
+                from agent.display import capture_local_edit_snapshot
+
+                snapshot = capture_local_edit_snapshot(name, args)
+            except Exception:
+                logger.debug("Failed to capture ACP edit snapshot for %s", name, exc_info=True)
+        tool_call_meta[tc_id] = {"args": args, "snapshot": snapshot}
+
         update = build_tool_start(tc_id, name, args)
         _send_update(conn, session_id, loop, update)
 
@@ -119,6 +130,7 @@ def make_step_cb(
     session_id: str,
     loop: asyncio.AbstractEventLoop,
     tool_call_ids: Dict[str, Deque[str]],
+    tool_call_meta: Dict[str, Dict[str, Any]],
 ) -> Callable:
     """Create a ``step_callback`` for AIAgent.
 
@@ -132,10 +144,12 @@ def make_step_cb(
             for tool_info in prev_tools:
                 tool_name = None
                 result = None
+                function_args = None
 
                 if isinstance(tool_info, dict):
                     tool_name = tool_info.get("name") or tool_info.get("function_name")
                     result = tool_info.get("result") or tool_info.get("output")
+                    function_args = tool_info.get("arguments") or tool_info.get("args")
                 elif isinstance(tool_info, str):
                     tool_name = tool_info
 
@@ -145,8 +159,13 @@ def make_step_cb(
                     tool_call_ids[tool_name] = queue
                 if tool_name and queue:
                     tc_id = queue.popleft()
+                    meta = tool_call_meta.pop(tc_id, {})
                     update = build_tool_complete(
-                        tc_id, tool_name, result=str(result) if result is not None else None
+                        tc_id,
+                        tool_name,
+                        result=str(result) if result is not None else None,
+                        function_args=function_args or meta.get("args"),
+                        snapshot=meta.get("snapshot"),
                     )
                     _send_update(conn, session_id, loop, update)
                     if not queue:

acp_adapter/server.py

@@ -26,6 +26,7 @@ from acp.schema import (
     McpServerHttp,
     McpServerSse,
     McpServerStdio,
+    ModelInfo,
     NewSessionResponse,
     PromptResponse,
     ResumeSessionResponse,
@@ -36,6 +37,7 @@ from acp.schema import (
     SessionCapabilities,
     SessionForkCapabilities,
     SessionListCapabilities,
+    SessionModelState,
     SessionResumeCapabilities,
     SessionInfo,
     TextContentBlock,
@@ -147,6 +149,98 @@ class HermesACPAgent(acp.Agent):
         self._conn = conn
         logger.info("ACP client connected")
 
+    @staticmethod
+    def _encode_model_choice(provider: str | None, model: str | None) -> str:
+        """Encode a model selection so ACP clients can keep provider context."""
+        raw_model = str(model or "").strip()
+        if not raw_model:
+            return ""
+        raw_provider = str(provider or "").strip().lower()
+        if not raw_provider:
+            return raw_model
+        return f"{raw_provider}:{raw_model}"
+
+    def _build_model_state(self, state: SessionState) -> SessionModelState | None:
+        """Return the ACP model selector payload for editors like Zed."""
+        model = str(state.model or getattr(state.agent, "model", "") or "").strip()
+        provider = getattr(state.agent, "provider", None) or detect_provider() or "openrouter"
+
+        try:
+            from hermes_cli.models import curated_models_for_provider, normalize_provider, provider_label
+
+            normalized_provider = normalize_provider(provider)
+            provider_name = provider_label(normalized_provider)
+            available_models: list[ModelInfo] = []
+            seen_ids: set[str] = set()
+
+            for model_id, description in curated_models_for_provider(normalized_provider):
+                rendered_model = str(model_id or "").strip()
+                if not rendered_model:
+                    continue
+                choice_id = self._encode_model_choice(normalized_provider, rendered_model)
+                if choice_id in seen_ids:
+                    continue
+                desc_parts = [f"Provider: {provider_name}"]
+                if description:
+                    desc_parts.append(str(description).strip())
+                if rendered_model == model:
+                    desc_parts.append("current")
+                available_models.append(
+                    ModelInfo(
+                        model_id=choice_id,
+                        name=rendered_model,
+                        description=" • ".join(part for part in desc_parts if part),
+                    )
+                )
+                seen_ids.add(choice_id)
+
+            current_model_id = self._encode_model_choice(normalized_provider, model)
+            if current_model_id and current_model_id not in seen_ids:
+                available_models.insert(
+                    0,
+                    ModelInfo(
+                        model_id=current_model_id,
+                        name=model,
+                        description=f"Provider: {provider_name} • current",
+                    ),
+                )
+
+            if available_models:
+                return SessionModelState(
+                    available_models=available_models,
+                    current_model_id=current_model_id or available_models[0].model_id,
+                )
+        except Exception:
+            logger.debug("Could not build ACP model state", exc_info=True)
+
+        if not model:
+            return None
+
+        fallback_choice = self._encode_model_choice(provider, model)
+        return SessionModelState(
+            available_models=[ModelInfo(model_id=fallback_choice, name=model)],
+            current_model_id=fallback_choice,
+        )
+
+    @staticmethod
+    def _resolve_model_selection(raw_model: str, current_provider: str) -> tuple[str, str]:
+        """Resolve ``provider:model`` input into the provider and normalized model id."""
+        target_provider = current_provider
+        new_model = raw_model.strip()
+
+        try:
+            from hermes_cli.models import detect_provider_for_model, parse_model_input
+
+            target_provider, new_model = parse_model_input(new_model, current_provider)
+            if target_provider == current_provider:
+                detected = detect_provider_for_model(new_model, current_provider)
+                if detected:
+                    target_provider, new_model = detected
+        except Exception:
+            logger.debug("Provider detection failed, using model as-is", exc_info=True)
+
+        return target_provider, new_model
+
     async def _register_session_mcp_servers(
         self,
         state: SessionState,
@@ -273,7 +367,10 @@ class HermesACPAgent(acp.Agent):
         await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("New session %s (cwd=%s)", state.session_id, cwd)
         self._schedule_available_commands_update(state.session_id)
-        return NewSessionResponse(session_id=state.session_id)
+        return NewSessionResponse(
+            session_id=state.session_id,
+            models=self._build_model_state(state),
+        )
 
     async def load_session(
         self,
@@ -289,7 +386,7 @@ class HermesACPAgent(acp.Agent):
         await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("Loaded session %s", session_id)
         self._schedule_available_commands_update(session_id)
-        return LoadSessionResponse()
+        return LoadSessionResponse(models=self._build_model_state(state))
 
     async def resume_session(
         self,
@@ -305,7 +402,7 @@ class HermesACPAgent(acp.Agent):
         await self._register_session_mcp_servers(state, mcp_servers)
         logger.info("Resumed session %s", state.session_id)
         self._schedule_available_commands_update(state.session_id)
-        return ResumeSessionResponse()
+        return ResumeSessionResponse(models=self._build_model_state(state))
 
     async def cancel(self, session_id: str, **kwargs: Any) -> None:
         state = self.session_manager.get_session(session_id)
@@ -340,11 +437,20 @@ class HermesACPAgent(acp.Agent):
         cwd: str | None = None,
         **kwargs: Any,
     ) -> ListSessionsResponse:
-        infos = self.session_manager.list_sessions()
-        sessions = [
-            SessionInfo(session_id=s["session_id"], cwd=s["cwd"])
-            for s in infos
-        ]
+        infos = self.session_manager.list_sessions(cwd=cwd)
+        sessions = []
+        for s in infos:
+            updated_at = s.get("updated_at")
+            if updated_at is not None and not isinstance(updated_at, str):
+                updated_at = str(updated_at)
+            sessions.append(
+                SessionInfo(
+                    session_id=s["session_id"],
+                    cwd=s["cwd"],
+                    title=s.get("title"),
+                    updated_at=updated_at,
+                )
+            )
         return ListSessionsResponse(sessions=sessions)
 
     # ---- Prompt (core) ------------------------------------------------------
@@ -389,12 +495,13 @@ class HermesACPAgent(acp.Agent):
             state.cancel_event.clear()
 
         tool_call_ids: dict[str, Deque[str]] = defaultdict(deque)
+        tool_call_meta: dict[str, dict[str, Any]] = {}
         previous_approval_cb = None
 
         if conn:
-            tool_progress_cb = make_tool_progress_cb(conn, session_id, loop, tool_call_ids)
+            tool_progress_cb = make_tool_progress_cb(conn, session_id, loop, tool_call_ids, tool_call_meta)
             thinking_cb = make_thinking_cb(conn, session_id, loop)
-            step_cb = make_step_cb(conn, session_id, loop, tool_call_ids)
+            step_cb = make_step_cb(conn, session_id, loop, tool_call_ids, tool_call_meta)
             message_cb = make_message_cb(conn, session_id, loop)
             approval_cb = make_approval_callback(conn.request_permission, loop, session_id)
         else:
@@ -449,6 +556,19 @@ class HermesACPAgent(acp.Agent):
             self.session_manager.save_session(session_id)
 
         final_response = result.get("final_response", "")
+        if final_response:
+            try:
+                from agent.title_generator import maybe_auto_title
+
+                maybe_auto_title(
+                    self.session_manager._get_db(),
+                    session_id,
+                    user_text,
+                    final_response,
+                    state.history,
+                )
+            except Exception:
+                logger.debug("Failed to auto-title ACP session %s", session_id, exc_info=True)
         if final_response and conn:
             update = acp.update_agent_message_text(final_response)
             await conn.session_update(session_id, update)
@@ -556,27 +676,15 @@ class HermesACPAgent(acp.Agent):
             provider = getattr(state.agent, "provider", None) or "auto"
             return f"Current model: {model}\nProvider: {provider}"
 
-        new_model = args.strip()
-        target_provider = None
         current_provider = getattr(state.agent, "provider", None) or "openrouter"
-
-        # Auto-detect provider for the requested model
-        try:
-            from hermes_cli.models import parse_model_input, detect_provider_for_model
-            target_provider, new_model = parse_model_input(new_model, current_provider)
-            if target_provider == current_provider:
-                detected = detect_provider_for_model(new_model, current_provider)
-                if detected:
-                    target_provider, new_model = detected
-        except Exception:
-            logger.debug("Provider detection failed, using model as-is", exc_info=True)
+        target_provider, new_model = self._resolve_model_selection(args, current_provider)
 
         state.model = new_model
         state.agent = self.session_manager._make_agent(
             session_id=state.session_id,
             cwd=state.cwd,
             model=new_model,
-            requested_provider=target_provider or current_provider,
+            requested_provider=target_provider,
         )
         self.session_manager.save_session(state.session_id)
         provider_label = getattr(state.agent, "provider", None) or target_provider or current_provider
@@ -678,20 +786,30 @@ class HermesACPAgent(acp.Agent):
         """Switch the model for a session (called by ACP protocol)."""
         state = self.session_manager.get_session(session_id)
         if state:
-            state.model = model_id
             current_provider = getattr(state.agent, "provider", None)
-            current_base_url = getattr(state.agent, "base_url", None)
-            current_api_mode = getattr(state.agent, "api_mode", None)
+            requested_provider, resolved_model = self._resolve_model_selection(
+                model_id,
+                current_provider or "openrouter",
+            )
+            state.model = resolved_model
+            provider_changed = bool(current_provider and requested_provider != current_provider)
+            current_base_url = None if provider_changed else getattr(state.agent, "base_url", None)
+            current_api_mode = None if provider_changed else getattr(state.agent, "api_mode", None)
             state.agent = self.session_manager._make_agent(
                 session_id=session_id,
                 cwd=state.cwd,
-                model=model_id,
-                requested_provider=current_provider,
+                model=resolved_model,
+                requested_provider=requested_provider,
                 base_url=current_base_url,
                 api_mode=current_api_mode,
             )
             self.session_manager.save_session(session_id)
-            logger.info("Session %s: model switched to %s", session_id, model_id)
+            logger.info(
+                "Session %s: model switched to %s via provider %s",
+                session_id,
+                resolved_model,
+                requested_provider,
+            )
             return SetSessionModelResponse()
         logger.warning("Session %s: model switch requested for missing session", session_id)
         return None

acp_adapter/session.py

@@ -13,8 +13,12 @@ from hermes_constants import get_hermes_home
 import copy
 import json
 import logging
+import os
+import re
 import sys
+import time
 import uuid
+from datetime import datetime, timezone
 from dataclasses import dataclass, field
 from threading import Lock
 from typing import Any, Dict, List, Optional
@@ -22,6 +26,64 @@ from typing import Any, Dict, List, Optional
 logger = logging.getLogger(__name__)
 
 
+def _normalize_cwd_for_compare(cwd: str | None) -> str:
+    raw = str(cwd or ".").strip()
+    if not raw:
+        raw = "."
+    expanded = os.path.expanduser(raw)
+
+    # Normalize Windows drive paths into the equivalent WSL mount form so
+    # ACP history filters match the same workspace across Windows and WSL.
+    match = re.match(r"^([A-Za-z]):[\\/](.*)$", expanded)
+    if match:
+        drive = match.group(1).lower()
+        tail = match.group(2).replace("\\", "/")
+        expanded = f"/mnt/{drive}/{tail}"
+    elif re.match(r"^/mnt/[A-Za-z]/", expanded):
+        expanded = f"/mnt/{expanded[5].lower()}/{expanded[7:]}"
+
+    return os.path.normpath(expanded)
+
+
+def _build_session_title(title: Any, preview: Any, cwd: str | None) -> str:
+    explicit = str(title or "").strip()
+    if explicit:
+        return explicit
+    preview_text = str(preview or "").strip()
+    if preview_text:
+        return preview_text
+    leaf = os.path.basename(str(cwd or "").rstrip("/\\"))
+    return leaf or "New thread"
+
+
+def _format_updated_at(value: Any) -> str | None:
+    if value is None:
+        return None
+    if isinstance(value, str) and value.strip():
+        return value
+    try:
+        return datetime.fromtimestamp(float(value), tz=timezone.utc).isoformat()
+    except Exception:
+        return None
+
+
+def _updated_at_sort_key(value: Any) -> float:
+    if value is None:
+        return float("-inf")
+    if isinstance(value, (int, float)):
+        return float(value)
+    raw = str(value).strip()
+    if not raw:
+        return float("-inf")
+    try:
+        return datetime.fromisoformat(raw.replace("Z", "+00:00")).timestamp()
+    except Exception:
+        try:
+            return float(raw)
+        except Exception:
+            return float("-inf")
+
+
 def _acp_stderr_print(*args, **kwargs) -> None:
     """Best-effort human-readable output sink for ACP stdio sessions.
 
@@ -162,47 +224,78 @@ class SessionManager:
         logger.info("Forked ACP session %s -> %s", session_id, new_id)
         return state
 
-    def list_sessions(self) -> List[Dict[str, Any]]:
+    def list_sessions(self, cwd: str | None = None) -> List[Dict[str, Any]]:
         """Return lightweight info dicts for all sessions (memory + database)."""
+        normalized_cwd = _normalize_cwd_for_compare(cwd) if cwd else None
+        db = self._get_db()
+        persisted_rows: dict[str, dict[str, Any]] = {}
+
+        if db is not None:
+            try:
+                for row in db.list_sessions_rich(source="acp", limit=1000):
+                    persisted_rows[str(row["id"])] = dict(row)
+            except Exception:
+                logger.debug("Failed to load ACP sessions from DB", exc_info=True)
+
         # Collect in-memory sessions first.
         with self._lock:
             seen_ids = set(self._sessions.keys())
-            results = [
-                {
-                    "session_id": s.session_id,
-                    "cwd": s.cwd,
-                    "model": s.model,
-                    "history_len": len(s.history),
-                }
-                for s in self._sessions.values()
-            ]
+            results = []
+            for s in self._sessions.values():
+                history_len = len(s.history)
+                if history_len <= 0:
+                    continue
+                if normalized_cwd and _normalize_cwd_for_compare(s.cwd) != normalized_cwd:
+                    continue
+                persisted = persisted_rows.get(s.session_id, {})
+                preview = next(
+                    (
+                        str(msg.get("content") or "").strip()
+                        for msg in s.history
+                        if msg.get("role") == "user" and str(msg.get("content") or "").strip()
+                    ),
+                    persisted.get("preview") or "",
+                )
+                results.append(
+                    {
+                        "session_id": s.session_id,
+                        "cwd": s.cwd,
+                        "model": s.model,
+                        "history_len": history_len,
+                        "title": _build_session_title(persisted.get("title"), preview, s.cwd),
+                        "updated_at": _format_updated_at(
+                            persisted.get("last_active") or persisted.get("started_at") or time.time()
+                        ),
+                    }
+                )
 
         # Merge any persisted sessions not currently in memory.
-        db = self._get_db()
-        if db is not None:
-            try:
-                rows = db.search_sessions(source="acp", limit=1000)
-                for row in rows:
-                    sid = row["id"]
-                    if sid in seen_ids:
-                        continue
-                    # Extract cwd from model_config JSON.
-                    cwd = "."
-                    mc = row.get("model_config")
-                    if mc:
-                        try:
-                            cwd = json.loads(mc).get("cwd", ".")
-                        except (json.JSONDecodeError, TypeError):
-                            pass
-                    results.append({
-                        "session_id": sid,
-                        "cwd": cwd,
-                        "model": row.get("model") or "",
-                        "history_len": row.get("message_count") or 0,
-                    })
-            except Exception:
-                logger.debug("Failed to list ACP sessions from DB", exc_info=True)
+        for sid, row in persisted_rows.items():
+            if sid in seen_ids:
+                continue
+            message_count = int(row.get("message_count") or 0)
+            if message_count <= 0:
+                continue
+            # Extract cwd from model_config JSON.
+            session_cwd = "."
+            mc = row.get("model_config")
+            if mc:
+                try:
+                    session_cwd = json.loads(mc).get("cwd", ".")
+                except (json.JSONDecodeError, TypeError):
+                    pass
+            if normalized_cwd and _normalize_cwd_for_compare(session_cwd) != normalized_cwd:
+                continue
+            results.append({
+                "session_id": sid,
+                "cwd": session_cwd,
+                "model": row.get("model") or "",
+                "history_len": message_count,
+                "title": _build_session_title(row.get("title"), row.get("preview"), session_cwd),
+                "updated_at": _format_updated_at(row.get("last_active") or row.get("started_at")),
+            })
 
+        results.sort(key=lambda item: _updated_at_sort_key(item.get("updated_at")), reverse=True)
         return results
 
     def update_cwd(self, session_id: str, cwd: str) -> Optional[SessionState]:

acp_adapter/tools.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import json
 import uuid
 from typing import Any, Dict, List, Optional
 
@@ -96,6 +97,170 @@ def build_tool_title(tool_name: str, args: Dict[str, Any]) -> str:
     return tool_name
 
 
+def _build_patch_mode_content(patch_text: str) -> List[Any]:
+    """Parse V4A patch mode input into ACP diff blocks when possible."""
+    if not patch_text:
+        return [acp.tool_content(acp.text_block(""))]
+
+    try:
+        from tools.patch_parser import OperationType, parse_v4a_patch
+
+        operations, error = parse_v4a_patch(patch_text)
+        if error or not operations:
+            return [acp.tool_content(acp.text_block(patch_text))]
+
+        content: List[Any] = []
+        for op in operations:
+            if op.operation == OperationType.UPDATE:
+                old_chunks: list[str] = []
+                new_chunks: list[str] = []
+                for hunk in op.hunks:
+                    old_lines = [line.content for line in hunk.lines if line.prefix in (" ", "-")]
+                    new_lines = [line.content for line in hunk.lines if line.prefix in (" ", "+")]
+                    if old_lines or new_lines:
+                        old_chunks.append("\n".join(old_lines))
+                        new_chunks.append("\n".join(new_lines))
+
+                old_text = "\n...\n".join(chunk for chunk in old_chunks if chunk)
+                new_text = "\n...\n".join(chunk for chunk in new_chunks if chunk)
+                if old_text or new_text:
+                    content.append(
+                        acp.tool_diff_content(
+                            path=op.file_path,
+                            old_text=old_text or None,
+                            new_text=new_text or "",
+                        )
+                    )
+                continue
+
+            if op.operation == OperationType.ADD:
+                added_lines = [line.content for hunk in op.hunks for line in hunk.lines if line.prefix == "+"]
+                content.append(
+                    acp.tool_diff_content(
+                        path=op.file_path,
+                        new_text="\n".join(added_lines),
+                    )
+                )
+                continue
+
+            if op.operation == OperationType.DELETE:
+                content.append(
+                    acp.tool_diff_content(
+                        path=op.file_path,
+                        old_text=f"Delete file: {op.file_path}",
+                        new_text="",
+                    )
+                )
+                continue
+
+            if op.operation == OperationType.MOVE:
+                content.append(
+                    acp.tool_content(acp.text_block(f"Move file: {op.file_path} -> {op.new_path}"))
+                )
+
+        return content or [acp.tool_content(acp.text_block(patch_text))]
+    except Exception:
+        return [acp.tool_content(acp.text_block(patch_text))]
+
+
+def _strip_diff_prefix(path: str) -> str:
+    raw = str(path or "").strip()
+    if raw.startswith(("a/", "b/")):
+        return raw[2:]
+    return raw
+
+
+def _parse_unified_diff_content(diff_text: str) -> List[Any]:
+    """Convert unified diff text into ACP diff content blocks."""
+    if not diff_text:
+        return []
+
+    content: List[Any] = []
+    current_old_path: Optional[str] = None
+    current_new_path: Optional[str] = None
+    old_lines: list[str] = []
+    new_lines: list[str] = []
+
+    def _flush() -> None:
+        nonlocal current_old_path, current_new_path, old_lines, new_lines
+        if current_old_path is None and current_new_path is None:
+            return
+        path = current_new_path if current_new_path and current_new_path != "/dev/null" else current_old_path
+        if not path or path == "/dev/null":
+            current_old_path = None
+            current_new_path = None
+            old_lines = []
+            new_lines = []
+            return
+        content.append(
+            acp.tool_diff_content(
+                path=_strip_diff_prefix(path),
+                old_text="\n".join(old_lines) if old_lines else None,
+                new_text="\n".join(new_lines),
+            )
+        )
+        current_old_path = None
+        current_new_path = None
+        old_lines = []
+        new_lines = []
+
+    for line in diff_text.splitlines():
+        if line.startswith("--- "):
+            _flush()
+            current_old_path = line[4:].strip()
+            continue
+        if line.startswith("+++ "):
+            current_new_path = line[4:].strip()
+            continue
+        if line.startswith("@@"):
+            continue
+        if current_old_path is None and current_new_path is None:
+            continue
+        if line.startswith("+"):
+            new_lines.append(line[1:])
+        elif line.startswith("-"):
+            old_lines.append(line[1:])
+        elif line.startswith(" "):
+            shared = line[1:]
+            old_lines.append(shared)
+            new_lines.append(shared)
+
+    _flush()
+    return content
+
+
+def _build_tool_complete_content(
+    tool_name: str,
+    result: Optional[str],
+    *,
+    function_args: Optional[Dict[str, Any]] = None,
+    snapshot: Any = None,
+) -> List[Any]:
+    """Build structured ACP completion content, falling back to plain text."""
+    display_result = result or ""
+    if len(display_result) > 5000:
+        display_result = display_result[:4900] + f"\n... ({len(result)} chars total, truncated)"
+
+    if tool_name in {"write_file", "patch", "skill_manage"}:
+        try:
+            from agent.display import extract_edit_diff
+
+            diff_text = extract_edit_diff(
+                tool_name,
+                result,
+                function_args=function_args,
+                snapshot=snapshot,
+            )
+            if isinstance(diff_text, str) and diff_text.strip():
+                diff_content = _parse_unified_diff_content(diff_text)
+                if diff_content:
+                    return diff_content
+        except Exception:
+            pass
+
+    return [acp.tool_content(acp.text_block(display_result))]
+
+
 # ---------------------------------------------------------------------------
 # Build ACP content objects for tool-call events
 # ---------------------------------------------------------------------------
@@ -119,9 +284,8 @@ def build_tool_start(
             new = arguments.get("new_string", "")
             content = [acp.tool_diff_content(path=path, new_text=new, old_text=old)]
         else:
-            # Patch mode — show the patch content as text
             patch_text = arguments.get("patch", "")
-            content = [acp.tool_content(acp.text_block(patch_text))]
+            content = _build_patch_mode_content(patch_text)
         return acp.start_tool_call(
             tool_call_id, title, kind=kind, content=content, locations=locations,
             raw_input=arguments,
@@ -178,16 +342,17 @@ def build_tool_complete(
     tool_call_id: str,
     tool_name: str,
     result: Optional[str] = None,
+    function_args: Optional[Dict[str, Any]] = None,
+    snapshot: Any = None,
 ) -> ToolCallProgress:
     """Create a ToolCallUpdate (progress) event for a completed tool call."""
     kind = get_tool_kind(tool_name)
-
-    # Truncate very large results for the UI
-    display_result = result or ""
-    if len(display_result) > 5000:
-        display_result = display_result[:4900] + f"\n... ({len(result)} chars total, truncated)"
-
-    content = [acp.tool_content(acp.text_block(display_result))]
+    content = _build_tool_complete_content(
+        tool_name,
+        result,
+        function_args=function_args,
+        snapshot=snapshot,
+    )
     return acp.update_tool_call(
         tool_call_id,
         kind=kind,

agent/auxiliary_client.py

@@ -94,6 +94,17 @@ def _normalize_aux_provider(provider: Optional[str]) -> str:
         return "custom"
     return _PROVIDER_ALIASES.get(normalized, normalized)
 
+
+_FIXED_TEMPERATURE_MODELS: Dict[str, float] = {
+    "kimi-for-coding": 0.6,
+}
+
+
+def _fixed_temperature_for_model(model: Optional[str]) -> Optional[float]:
+    """Return a required temperature override for models with strict contracts."""
+    normalized = (model or "").strip().lower()
+    return _FIXED_TEMPERATURE_MODELS.get(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",
@@ -734,6 +745,15 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
                 from hermes_cli.models import copilot_default_headers
 
                 extra["default_headers"] = copilot_default_headers()
+            elif "generativelanguage.googleapis.com" in base_url.lower():
+                # Google's OpenAI-compatible endpoint only accepts x-goog-api-key.
+                # Passing api_key= causes the SDK to inject Authorization: Bearer,
+                # which Google rejects with HTTP 400 "Multiple authentication
+                # credentials received". Use a placeholder for api_key and pass
+                # the real key via x-goog-api-key header instead.
+                # Fixes: https://github.com/NousResearch/hermes-agent/issues/7893
+                extra["default_headers"] = {"x-goog-api-key": api_key}
+                api_key = "not-used"
             return OpenAI(api_key=api_key, base_url=base_url, **extra), model
 
         creds = resolve_api_key_provider_credentials(provider_id)
@@ -755,6 +775,15 @@ def _resolve_api_key_provider() -> Tuple[Optional[OpenAI], Optional[str]]:
             from hermes_cli.models import copilot_default_headers
 
             extra["default_headers"] = copilot_default_headers()
+        elif "generativelanguage.googleapis.com" in base_url.lower():
+            # Google's OpenAI-compatible endpoint only accepts x-goog-api-key.
+            # Passing api_key= causes the SDK to inject Authorization: Bearer,
+            # which Google rejects with HTTP 400 "Multiple authentication
+            # credentials received". Use a placeholder for api_key and pass
+            # the real key via x-goog-api-key header instead.
+            # Fixes: https://github.com/NousResearch/hermes-agent/issues/7893
+            extra["default_headers"] = {"x-goog-api-key": api_key}
+            api_key = "not-used"
         return OpenAI(api_key=api_key, base_url=base_url, **extra), model
 
     return None, None
@@ -1064,8 +1093,6 @@ _AUTO_PROVIDER_LABELS = {
     "_resolve_api_key_provider": "api-key",
 }
 
-_AGGREGATOR_PROVIDERS = frozenset({"openrouter", "nous"})
-
 _MAIN_RUNTIME_FIELDS = ("provider", "model", "base_url", "api_key", "api_mode")
 
 
@@ -1196,11 +1223,15 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
     """Full auto-detection chain.
 
     Priority:
-      1. If the user's main provider is NOT an aggregator (OpenRouter / Nous),
-         use their main provider + main model directly.  This ensures users on
-         Alibaba, DeepSeek, ZAI, etc. get auxiliary tasks handled by the same
-         provider they already have credentials for — no OpenRouter key needed.
-      2. OpenRouter → Nous → custom → Codex → API-key providers (original chain).
+      1. User's main provider + main model, regardless of provider type.
+         This means auxiliary tasks (compression, vision, web extraction,
+         session search, etc.) use the same model the user configured for
+         chat.  Users on OpenRouter/Nous get their chosen chat model; users
+         on DeepSeek/ZAI/Alibaba get theirs; etc.  Running aux tasks on the
+         user's picked model keeps behavior predictable — no surprise
+         switches to a cheap fallback model for side tasks.
+      2. OpenRouter → Nous → custom → Codex → API-key providers (fallback
+         chain, only used when the main provider has no working client).
     """
     global auxiliary_is_nous, _stale_base_url_warned
     auxiliary_is_nous = False  # Reset — _try_nous() will set True if it wins
@@ -1230,11 +1261,16 @@ def _resolve_auto(main_runtime: Optional[Dict[str, Any]] = None) -> Tuple[Option
             )
             _stale_base_url_warned = True
 
-    # ── Step 1: non-aggregator main provider → use main model directly ──
+    # ── Step 1: main provider + main model → use them directly ──
+    #
+    # This is the primary aux backend for every user.  "auto" means
+    # "use my main chat model for side tasks as well" — including users
+    # on aggregators (OpenRouter, Nous) who previously got routed to a
+    # cheap provider-side default.  Explicit per-task overrides set via
+    # config.yaml (auxiliary.<task>.provider) still win over this.
     main_provider = runtime_provider or _read_main_provider()
     main_model = runtime_model or _read_main_model()
     if (main_provider and main_model
-            and main_provider not in _AGGREGATOR_PROVIDERS
             and main_provider not in ("auto", "")):
         resolved_provider = main_provider
         explicit_base_url = None
@@ -1593,6 +1629,15 @@ def resolve_provider_client(
             from hermes_cli.models import copilot_default_headers
 
             headers.update(copilot_default_headers())
+        elif "generativelanguage.googleapis.com" in base_url.lower():
+            # Google's OpenAI-compatible endpoint only accepts x-goog-api-key.
+            # Passing api_key= causes the OpenAI SDK to inject Authorization: Bearer,
+            # which Google rejects with HTTP 400 "Multiple authentication credentials
+            # received". Use a placeholder for api_key and pass the real key via
+            # x-goog-api-key header instead.
+            # Fixes: https://github.com/NousResearch/hermes-agent/issues/7893
+            headers["x-goog-api-key"] = api_key
+            api_key = "not-used"
 
         client = OpenAI(api_key=api_key, base_url=base_url,
                         **({"default_headers": headers} if headers else {}))
@@ -1817,34 +1862,31 @@ def resolve_vision_provider_client(
 
     if requested == "auto":
         # 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)
+        #   1. User's main provider + main model (including aggregators).
+        #      _PROVIDER_VISION_MODELS provides per-provider vision model
+        #      overrides when the provider has a dedicated multimodal model
+        #      that differs from the chat model (e.g. xiaomi → mimo-v2-omni,
+        #      zai → glm-5v-turbo).
+        #   2. OpenRouter  (vision-capable aggregator fallback)
+        #   3. Nous Portal (vision-capable aggregator fallback)
         #   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, Xiaomi, named custom, etc.)
-                # Use provider-specific vision model if available, otherwise main model.
-                vision_model = _PROVIDER_VISION_MODELS.get(main_provider, main_model)
-                rpc_client, rpc_model = resolve_provider_client(
-                    main_provider, vision_model,
-                    api_mode=resolved_api_mode)
-                if rpc_client is not None:
-                    logger.info(
-                        "Vision auto-detect: using active provider %s (%s)",
-                        main_provider, rpc_model or vision_model,
-                    )
-                    return _finalize(
-                        main_provider, rpc_client, rpc_model or vision_model)
+            vision_model = _PROVIDER_VISION_MODELS.get(main_provider, main_model)
+            rpc_client, rpc_model = resolve_provider_client(
+                main_provider, vision_model,
+                api_mode=resolved_api_mode)
+            if rpc_client is not None:
+                logger.info(
+                    "Vision auto-detect: using main provider %s (%s)",
+                    main_provider, rpc_model or vision_model,
+                )
+                return _finalize(
+                    main_provider, rpc_client, rpc_model or vision_model)
 
-        # Fall back through aggregators.
+        # Fall back through aggregators (uses their dedicated vision model,
+        # not the user's main model) when main provider has no client.
         for candidate in _VISION_AUTO_PROVIDER_ORDER:
             if candidate == main_provider:
                 continue  # already tried above
@@ -2293,6 +2335,10 @@ def _build_call_kwargs(
         "timeout": timeout,
     }
 
+    fixed_temperature = _fixed_temperature_for_model(model)
+    if fixed_temperature is not None:
+        temperature = fixed_temperature
+
     # Opus 4.7+ rejects any non-default temperature/top_p/top_k — silently
     # drop here so auxiliary callers that hardcode temperature (e.g. 0.3 on
     # flush_memories, 0 on structured-JSON extraction) don't 400 the moment

agent/credential_pool.py

@@ -1130,6 +1130,14 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
         state = _load_provider_state(auth_store, "nous")
         if state:
             active_sources.add("device_code")
+            # Prefer a user-supplied label embedded in the singleton state
+            # (set by persist_nous_credentials(label=...) when the user ran
+            # `hermes auth add nous --label <name>`).  Fall back to the
+            # auto-derived token fingerprint for logins that didn't supply one.
+            custom_label = str(state.get("label") or "").strip()
+            seeded_label = custom_label or label_from_token(
+                state.get("access_token", ""), "device_code"
+            )
             changed |= _upsert_entry(
                 entries,
                 provider,
@@ -1148,7 +1156,7 @@ def _seed_from_singletons(provider: str, entries: List[PooledCredential]) -> Tup
                     "agent_key": state.get("agent_key"),
                     "agent_key_expires_at": state.get("agent_key_expires_at"),
                     "tls": state.get("tls") if isinstance(state.get("tls"), dict) else None,
-                    "label": label_from_token(state.get("access_token", ""), "device_code"),
+                    "label": seeded_label,
                 },
             )
 

agent/gemini_cloudcode_adapter.py

@@ -747,18 +747,149 @@ class GeminiCloudCodeClient:
 
 
 def _gemini_http_error(response: httpx.Response) -> CodeAssistError:
+    """Translate an httpx response into a CodeAssistError with rich metadata.
+
+    Parses Google's error envelope (``{"error": {"code", "message", "status",
+    "details": [...]}}``) so the agent's error classifier can reason about
+    the failure — ``status_code`` enables the rate_limit / auth classification
+    paths, and ``response`` lets the main loop honor ``Retry-After`` just
+    like it does for OpenAI SDK exceptions.
+
+    Also lifts a few recognizable Google conditions into human-readable
+    messages so the user sees something better than a 500-char JSON dump:
+
+        MODEL_CAPACITY_EXHAUSTED → "Gemini model capacity exhausted for
+            <model>. This is a Google-side throttle..."
+        RESOURCE_EXHAUSTED w/o reason → quota-style message
+        404 → "Model <name> not found at cloudcode-pa..."
+    """
     status = response.status_code
+
+    # Parse the body once, surviving any weird encodings.
+    body_text = ""
+    body_json: Dict[str, Any] = {}
     try:
-        body = response.text[:500]
+        body_text = response.text
     except Exception:
-        body = ""
-    # Let run_agent's retry logic see auth errors as rotatable via `api_key`
+        body_text = ""
+    if body_text:
+        try:
+            parsed = json.loads(body_text)
+            if isinstance(parsed, dict):
+                body_json = parsed
+        except (ValueError, TypeError):
+            body_json = {}
+
+    # Dig into Google's error envelope.  Shape is:
+    #   {"error": {"code": 429, "message": "...", "status": "RESOURCE_EXHAUSTED",
+    #              "details": [{"@type": ".../ErrorInfo", "reason": "MODEL_CAPACITY_EXHAUSTED",
+    #                           "metadata": {...}},
+    #                          {"@type": ".../RetryInfo", "retryDelay": "30s"}]}}
+    err_obj = body_json.get("error") if isinstance(body_json, dict) else None
+    if not isinstance(err_obj, dict):
+        err_obj = {}
+    err_status = str(err_obj.get("status") or "").strip()
+    err_message = str(err_obj.get("message") or "").strip()
+    err_details_list = err_obj.get("details") if isinstance(err_obj.get("details"), list) else []
+
+    # Extract google.rpc.ErrorInfo reason + metadata.  There may be more
+    # than one ErrorInfo (rare), so we pick the first one with a reason.
+    error_reason = ""
+    error_metadata: Dict[str, Any] = {}
+    retry_delay_seconds: Optional[float] = None
+    for detail in err_details_list:
+        if not isinstance(detail, dict):
+            continue
+        type_url = str(detail.get("@type") or "")
+        if not error_reason and type_url.endswith("/google.rpc.ErrorInfo"):
+            reason = detail.get("reason")
+            if isinstance(reason, str) and reason:
+                error_reason = reason
+            md = detail.get("metadata")
+            if isinstance(md, dict):
+                error_metadata = md
+        elif retry_delay_seconds is None and type_url.endswith("/google.rpc.RetryInfo"):
+            # retryDelay is a google.protobuf.Duration string like "30s" or "1.5s".
+            delay_raw = detail.get("retryDelay")
+            if isinstance(delay_raw, str) and delay_raw.endswith("s"):
+                try:
+                    retry_delay_seconds = float(delay_raw[:-1])
+                except ValueError:
+                    pass
+            elif isinstance(delay_raw, (int, float)):
+                retry_delay_seconds = float(delay_raw)
+
+    # Fall back to the Retry-After header if the body didn't include RetryInfo.
+    if retry_delay_seconds is None:
+        try:
+            header_val = response.headers.get("Retry-After") or response.headers.get("retry-after")
+        except Exception:
+            header_val = None
+        if header_val:
+            try:
+                retry_delay_seconds = float(header_val)
+            except (TypeError, ValueError):
+                retry_delay_seconds = None
+
+    # Classify the error code.  ``code_assist_rate_limited`` stays the default
+    # for 429s; a more specific reason tag helps downstream callers (e.g. tests,
+    # logs) without changing the rate_limit classification path.
     code = f"code_assist_http_{status}"
     if status == 401:
         code = "code_assist_unauthorized"
     elif status == 429:
         code = "code_assist_rate_limited"
+        if error_reason == "MODEL_CAPACITY_EXHAUSTED":
+            code = "code_assist_capacity_exhausted"
+
+    # Build a human-readable message.  Keep the status + a raw-body tail for
+    # debugging, but lead with a friendlier summary when we recognize the
+    # Google signal.
+    model_hint = ""
+    if isinstance(error_metadata, dict):
+        model_hint = str(error_metadata.get("model") or error_metadata.get("modelId") or "").strip()
+
+    if status == 429 and error_reason == "MODEL_CAPACITY_EXHAUSTED":
+        target = model_hint or "this Gemini model"
+        message = (
+            f"Gemini capacity exhausted for {target} (Google-side throttle, "
+            f"not a Hermes issue). Try a different Gemini model or set a "
+            f"fallback_providers entry to a non-Gemini provider."
+        )
+        if retry_delay_seconds is not None:
+            message += f" Google suggests retrying in {retry_delay_seconds:g}s."
+    elif status == 429 and err_status == "RESOURCE_EXHAUSTED":
+        message = (
+            f"Gemini quota exhausted ({err_message or 'RESOURCE_EXHAUSTED'}). "
+            f"Check /gquota for remaining daily requests."
+        )
+        if retry_delay_seconds is not None:
+            message += f" Retry suggested in {retry_delay_seconds:g}s."
+    elif status == 404:
+        # Google returns 404 when a model has been retired or renamed.
+        target = model_hint or (err_message or "model")
+        message = (
+            f"Code Assist 404: {target} is not available at "
+            f"cloudcode-pa.googleapis.com. It may have been renamed or "
+            f"retired. Check hermes_cli/models.py for the current list."
+        )
+    elif err_message:
+        # Generic fallback with the parsed message.
+        message = f"Code Assist HTTP {status} ({err_status or 'error'}): {err_message}"
+    else:
+        # Last-ditch fallback — raw body snippet.
+        message = f"Code Assist returned HTTP {status}: {body_text[:500]}"
+
     return CodeAssistError(
-        f"Code Assist returned HTTP {status}: {body}",
+        message,
         code=code,
+        status_code=status,
+        response=response,
+        retry_after=retry_delay_seconds,
+        details={
+            "status": err_status,
+            "reason": error_reason,
+            "metadata": error_metadata,
+            "message": err_message,
+        },
     )

agent/google_code_assist.py

@@ -68,9 +68,45 @@ _ONBOARDING_POLL_INTERVAL_SECONDS = 5.0
 
 
 class CodeAssistError(RuntimeError):
-    def __init__(self, message: str, *, code: str = "code_assist_error") -> None:
+    """Exception raised by the Code Assist (``cloudcode-pa``) integration.
+
+    Carries HTTP status / response / retry-after metadata so the agent's
+    ``error_classifier._extract_status_code`` and the main loop's Retry-After
+    handling (which walks ``error.response.headers``) pick up the right
+    signals.  Without these, 429s from the OAuth path look like opaque
+    ``RuntimeError`` and skip the rate-limit path.
+    """
+
+    def __init__(
+        self,
+        message: str,
+        *,
+        code: str = "code_assist_error",
+        status_code: Optional[int] = None,
+        response: Any = None,
+        retry_after: Optional[float] = None,
+        details: Optional[Dict[str, Any]] = None,
+    ) -> None:
         super().__init__(message)
         self.code = code
+        # ``status_code`` is picked up by ``agent.error_classifier._extract_status_code``
+        # so a 429 from Code Assist classifies as FailoverReason.rate_limit and
+        # triggers the main loop's fallback_providers chain the same way SDK
+        # errors do.
+        self.status_code = status_code
+        # ``response`` is the underlying ``httpx.Response`` (or a shim with a
+        # ``.headers`` mapping and ``.json()`` method).  The main loop reads
+        # ``error.response.headers["Retry-After"]`` to honor Google's retry
+        # hints when the backend throttles us.
+        self.response = response
+        # Parsed ``Retry-After`` seconds (kept separately for convenience —
+        # Google returns retry hints in both the header and the error body's
+        # ``google.rpc.RetryInfo`` details, and we pick whichever we found).
+        self.retry_after = retry_after
+        # Parsed structured error details from the Google error envelope
+        # (e.g. ``{"reason": "MODEL_CAPACITY_EXHAUSTED", "status": "RESOURCE_EXHAUSTED"}``).
+        # Useful for logging and for tests that want to assert on specifics.
+        self.details = details or {}
 
 
 class ProjectIdRequiredError(CodeAssistError):

agent/model_metadata.py

@@ -38,6 +38,7 @@ _PROVIDER_PREFIXES: frozenset[str] = frozenset({
     "mimo", "xiaomi-mimo",
     "arcee-ai", "arceeai",
     "xai", "x-ai", "x.ai", "grok",
+    "nvidia", "nim", "nvidia-nim", "nemotron",
     "qwen-portal",
 })
 
@@ -124,7 +125,6 @@ DEFAULT_CONTEXT_LENGTHS = {
     "gemini": 1048576,
     # Gemma (open models served via AI Studio)
     "gemma-4-31b": 256000,
-    "gemma-4-26b": 256000,
     "gemma-3": 131072,
     "gemma": 8192,  # fallback for older gemma models
     # DeepSeek
@@ -158,6 +158,8 @@ DEFAULT_CONTEXT_LENGTHS = {
     "grok": 131072,             # catch-all (grok-beta, unknown grok-*)
     # Kimi
     "kimi": 262144,
+    # Nemotron — NVIDIA's open-weights series (128K context across all sizes)
+    "nemotron": 131072,
     # Arcee
     "trinity": 262144,
     # OpenRouter
@@ -240,6 +242,7 @@ _URL_TO_PROVIDER: Dict[str, str] = {
     "api.fireworks.ai": "fireworks",
     "opencode.ai": "opencode-go",
     "api.x.ai": "xai",
+    "integrate.api.nvidia.com": "nvidia",
     "api.xiaomimimo.com": "xiaomi",
     "xiaomimimo.com": "xiaomi",
     "ollama.com": "ollama-cloud",

agent/prompt_builder.py

@@ -654,7 +654,7 @@ def build_skills_system_prompt(
             ):
                 continue
             skills_by_category.setdefault(category, []).append(
-                (skill_name, entry.get("description", ""))
+                (frontmatter_name, entry.get("description", ""))
             )
         category_descriptions = {
             str(k): str(v)
@@ -679,7 +679,7 @@ def build_skills_system_prompt(
             ):
                 continue
             skills_by_category.setdefault(entry["category"], []).append(
-                (skill_name, entry["description"])
+                (entry["frontmatter_name"], entry["description"])
             )
 
         # Read category-level DESCRIPTION.md files
@@ -722,9 +722,10 @@ def build_skills_system_prompt(
                     continue
                 entry = _build_snapshot_entry(skill_file, ext_dir, frontmatter, desc)
                 skill_name = entry["skill_name"]
-                if skill_name in seen_skill_names:
+                frontmatter_name = entry["frontmatter_name"]
+                if frontmatter_name in seen_skill_names:
                     continue
-                if entry["frontmatter_name"] in disabled or skill_name in disabled:
+                if frontmatter_name in disabled or skill_name in disabled:
                     continue
                 if not _skill_should_show(
                     extract_skill_conditions(frontmatter),
@@ -732,9 +733,9 @@ def build_skills_system_prompt(
                     available_toolsets,
                 ):
                     continue
-                seen_skill_names.add(skill_name)
+                seen_skill_names.add(frontmatter_name)
                 skills_by_category.setdefault(entry["category"], []).append(
-                    (skill_name, entry["description"])
+                    (frontmatter_name, entry["description"])
                 )
             except Exception as e:
                 logger.debug("Error reading external skill %s: %s", skill_file, e)

cli-config.yaml.example

@@ -24,6 +24,7 @@ model:
   #   "minimax"      - MiniMax global (requires: MINIMAX_API_KEY)
   #   "minimax-cn"   - MiniMax China (requires: MINIMAX_CN_API_KEY)
   #   "huggingface"  - Hugging Face Inference (requires: HF_TOKEN)
+  #   "nvidia"       - NVIDIA NIM / build.nvidia.com (requires: NVIDIA_API_KEY)
   #   "xiaomi"       - Xiaomi MiMo (requires: XIAOMI_API_KEY)
   #   "arcee"        - Arcee AI Trinity models (requires: ARCEEAI_API_KEY)
   #   "ollama-cloud" - Ollama Cloud (requires: OLLAMA_API_KEY — https://ollama.com/settings)

cli.py

@@ -18,6 +18,8 @@ import os
 import shutil
 import sys
 import json
+import re
+import base64
 import atexit
 import tempfile
 import time
@@ -78,6 +80,42 @@ _project_env = Path(__file__).parent / '.env'
 load_hermes_dotenv(hermes_home=_hermes_home, project_env=_project_env)
 
 
+_REASONING_TAGS = (
+    "REASONING_SCRATCHPAD",
+    "think",
+    "reasoning",
+    "THINKING",
+    "thinking",
+)
+
+
+def _strip_reasoning_tags(text: str) -> str:
+    cleaned = text
+    for tag in _REASONING_TAGS:
+        cleaned = re.sub(rf"<{tag}>.*?</{tag}>\s*", "", cleaned, flags=re.DOTALL)
+        cleaned = re.sub(rf"<{tag}>.*$", "", cleaned, flags=re.DOTALL)
+    return cleaned.strip()
+
+
+def _assistant_content_as_text(content: Any) -> str:
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = [
+            str(part.get("text", ""))
+            for part in content
+            if isinstance(part, dict) and part.get("type") == "text"
+        ]
+        return "\n".join(p for p in parts if p)
+    return str(content)
+
+
+def _assistant_copy_text(content: Any) -> str:
+    return _strip_reasoning_tags(_assistant_content_as_text(content))
+
+
 # =============================================================================
 # Configuration Loading
 # =============================================================================
@@ -1172,6 +1210,10 @@ def _resolve_attachment_path(raw_path: str) -> Path | None:
         return None
 
     expanded = os.path.expandvars(os.path.expanduser(token))
+    if os.name != "nt":
+        normalized = expanded.replace("\\", "/")
+        if len(normalized) >= 3 and normalized[1] == ":" and normalized[2] == "/" and normalized[0].isalpha():
+            expanded = f"/mnt/{normalized[0].lower()}/{normalized[3:]}"
     path = Path(expanded)
     if not path.is_absolute():
         base_dir = Path(os.getenv("TERMINAL_CWD", os.getcwd()))
@@ -1254,10 +1296,12 @@ def _detect_file_drop(user_input: str) -> "dict | None":
         or stripped.startswith("~")
         or stripped.startswith("./")
         or stripped.startswith("../")
+        or (len(stripped) >= 3 and stripped[1] == ":" and stripped[2] in ("\\", "/") and stripped[0].isalpha())
         or stripped.startswith('"/')
         or stripped.startswith('"~')
         or stripped.startswith("'/")
         or stripped.startswith("'~")
+        or (len(stripped) >= 4 and stripped[0] in ("'", '"') and stripped[2] == ":" and stripped[3] in ("\\", "/") and stripped[1].isalpha())
     )
     if not starts_like_path:
         return None
@@ -3125,21 +3169,6 @@ class HermesCLI:
         MAX_ASST_LEN = 200           # truncate assistant text
         MAX_ASST_LINES = 3           # max lines of assistant text
 
-        def _strip_reasoning(text: str) -> str:
-            """Remove <REASONING_SCRATCHPAD>...</REASONING_SCRATCHPAD> blocks
-            from displayed text (reasoning model internal thoughts)."""
-            import re
-            cleaned = re.sub(
-                r"<REASONING_SCRATCHPAD>.*?</REASONING_SCRATCHPAD>\s*",
-                "", text, flags=re.DOTALL,
-            )
-            # Also strip unclosed reasoning tags at the end
-            cleaned = re.sub(
-                r"<REASONING_SCRATCHPAD>.*$",
-                "", cleaned, flags=re.DOTALL,
-            )
-            return cleaned.strip()
-
         # Collect displayable entries (skip system, tool-result messages)
         entries = []  # list of (role, display_text)
         _last_asst_idx = None       # index of last assistant entry
@@ -3171,7 +3200,7 @@ class HermesCLI:
 
             elif role == "assistant":
                 text = "" if content is None else str(content)
-                text = _strip_reasoning(text)
+                text = _strip_reasoning_tags(text)
                 parts = []
                 full_parts = []  # un-truncated version
                 if text:
@@ -3510,6 +3539,26 @@ class HermesCLI:
         killed = process_registry.kill_all()
         print(f"  ✅ Stopped {killed} process(es).")
 
+    def _handle_agents_command(self):
+        """Handle /agents — show background processes and agent status."""
+        from tools.process_registry import format_uptime_short, process_registry
+
+        processes = process_registry.list_sessions()
+        running = [p for p in processes if p.get("status") == "running"]
+        finished = [p for p in processes if p.get("status") != "running"]
+
+        _cprint(f"  Running processes: {len(running)}")
+        for p in running:
+            cmd = p.get("command", "")[:80]
+            up = format_uptime_short(p.get("uptime_seconds", 0))
+            _cprint(f"    {p.get('session_id', '?')} · {up} · {cmd}")
+
+        if finished:
+            _cprint(f"  Recently finished: {len(finished)}")
+
+        agent_running = getattr(self, "_agent_running", False)
+        _cprint(f"  Agent: {'running' if agent_running else 'idle'}")
+
     def _handle_paste_command(self):
         """Handle /paste — explicitly check clipboard for an image.
 
@@ -3535,6 +3584,61 @@ class HermesCLI:
         else:
             _cprint(f"  {_DIM}(._.) No image found in clipboard{_RST}")
 
+    def _write_osc52_clipboard(self, text: str) -> None:
+        """Copy *text* to terminal clipboard via OSC 52."""
+        payload = base64.b64encode(text.encode("utf-8")).decode("ascii")
+        seq = f"\x1b]52;c;{payload}\x07"
+        out = getattr(self, "_app", None)
+        output = getattr(out, "output", None) if out else None
+        if output and hasattr(output, "write_raw"):
+            output.write_raw(seq)
+            output.flush()
+            return
+        if output and hasattr(output, "write"):
+            output.write(seq)
+            output.flush()
+            return
+        sys.stdout.write(seq)
+        sys.stdout.flush()
+
+    def _handle_copy_command(self, cmd_original: str) -> None:
+        """Handle /copy [number] — copy assistant output to clipboard."""
+        parts = cmd_original.split(maxsplit=1)
+        arg = parts[1].strip() if len(parts) > 1 else ""
+
+        assistant = [m for m in self.conversation_history if m.get("role") == "assistant"]
+        if not assistant:
+            _cprint("  Nothing to copy yet.")
+            return
+
+        if arg:
+            try:
+                idx = int(arg) - 1
+            except ValueError:
+                _cprint("  Usage: /copy [number]")
+                return
+            if idx < 0 or idx >= len(assistant):
+                _cprint(f"  Invalid response number. Use 1-{len(assistant)}.")
+                return
+        else:
+            idx = len(assistant) - 1
+            while idx >= 0 and not _assistant_copy_text(assistant[idx].get("content")):
+                idx -= 1
+            if idx < 0:
+                _cprint("  Nothing to copy in assistant responses yet.")
+                return
+
+        text = _assistant_copy_text(assistant[idx].get("content"))
+        if not text:
+            _cprint("  Nothing to copy in that assistant response.")
+            return
+
+        try:
+            self._write_osc52_clipboard(text)
+            _cprint(f"  Copied assistant response #{idx + 1} to clipboard")
+        except Exception as e:
+            _cprint(f"  Clipboard copy failed: {e}")
+
     def _handle_image_command(self, cmd_original: str):
         """Handle /image <path> — attach a local image file for the next prompt."""
         raw_args = (cmd_original.split(None, 1)[1].strip() if " " in cmd_original else "")
@@ -3671,7 +3775,7 @@ class HermesCLI:
             skin = get_active_skin()
             separator_color = skin.get_color("banner_dim", "#B8860B")
             accent_color = skin.get_color("ui_accent", "#FFBF00")
-            label_color = skin.get_color("ui_label", "#4dd0e1")
+            label_color = skin.get_color("ui_label", "#DAA520")
         except Exception:
             separator_color, accent_color, label_color = "#B8860B", "#FFBF00", "cyan"
         toolsets_info = ""
@@ -5553,6 +5657,8 @@ class HermesCLI:
             self._show_usage()
         elif canonical == "insights":
             self._show_insights(cmd_original)
+        elif canonical == "copy":
+            self._handle_copy_command(cmd_original)
         elif canonical == "debug":
             self._handle_debug_command()
         elif canonical == "paste":
@@ -5596,6 +5702,8 @@ class HermesCLI:
             self._handle_snapshot_command(cmd_original)
         elif canonical == "stop":
             self._handle_stop_command()
+        elif canonical == "agents":
+            self._handle_agents_command()
         elif canonical == "background":
             self._handle_background_command(cmd_original)
         elif canonical == "btw":
@@ -5612,6 +5720,30 @@ class HermesCLI:
                     _cprint(f"  Queued for the next turn: {payload[:80]}{'...' if len(payload) > 80 else ''}")
                 else:
                     _cprint(f"  Queued: {payload[:80]}{'...' if len(payload) > 80 else ''}")
+        elif canonical == "steer":
+            # Inject a message after the next tool call without interrupting.
+            # If the agent is actively running, push the text into the agent's
+            # pending_steer slot — the drain hook in _execute_tool_calls_*
+            # will append it to the next tool result's content. If no agent
+            # is running, fall back to queue semantics (same as /queue).
+            parts = cmd_original.split(None, 1)
+            payload = parts[1].strip() if len(parts) > 1 else ""
+            if not payload:
+                _cprint("  Usage: /steer <prompt>")
+            elif self._agent_running and self.agent is not None and hasattr(self.agent, "steer"):
+                try:
+                    accepted = self.agent.steer(payload)
+                except Exception as exc:
+                    _cprint(f"  Steer failed: {exc}")
+                else:
+                    if accepted:
+                        _cprint(f"  ⏩ Steer queued — arrives after the next tool call: {payload[:80]}{'...' if len(payload) > 80 else ''}")
+                    else:
+                        _cprint("  Steer rejected (empty payload).")
+            else:
+                # No active run — treat as a normal next-turn message.
+                self._pending_input.put(payload)
+                _cprint(f"  No agent running; queued as next turn: {payload[:80]}{'...' if len(payload) > 80 else ''}")
         elif canonical == "skin":
             self._handle_skin_command(cmd_original)
         elif canonical == "voice":
@@ -6909,8 +7041,7 @@ class HermesCLI:
                 )
             raise RuntimeError(
                 "Voice mode requires sounddevice and numpy.\n"
-                "Install with: pip install sounddevice numpy\n"
-                "Or: pip install hermes-agent[voice]"
+                f"Install with: {sys.executable} -m pip install sounddevice numpy"
             )
         if not reqs.get("stt_available", reqs.get("stt_key_set")):
             raise RuntimeError(
@@ -7186,8 +7317,7 @@ class HermesCLI:
                     _cprint(f"  {_DIM}Then install/update the Termux:API Android app for microphone capture{_RST}")
                     _cprint(f"  {_BOLD}Option 2: pkg install python-numpy portaudio && python -m pip install sounddevice{_RST}")
                 else:
-                    _cprint(f"\n  {_BOLD}Install: pip install {' '.join(reqs['missing_packages'])}{_RST}")
-                    _cprint(f"  {_DIM}Or: pip install hermes-agent[voice]{_RST}")
+                    _cprint(f"\n  {_BOLD}Install: {sys.executable} -m pip install {' '.join(reqs['missing_packages'])}{_RST}")
             return
 
         with self._voice_lock:
@@ -8138,7 +8268,15 @@ class HermesCLI:
                 else:
                     print(f"\n⚡ Sending after interrupt: '{preview}'")
                 self._pending_input.put(combined)
-            
+
+            # If a /steer was left over (agent finished before another tool
+            # batch could absorb it), deliver it as the next user turn.
+            _leftover_steer = result.get("pending_steer") if result else None
+            if _leftover_steer and hasattr(self, '_pending_input'):
+                preview = _leftover_steer[:60] + ("..." if len(_leftover_steer) > 60 else "")
+                print(f"\n⏩ Delivering leftover /steer as next turn: '{preview}'")
+                self._pending_input.put(_leftover_steer)
+
             return response
             
         except Exception as e:
@@ -9959,8 +10097,36 @@ class HermesCLI:
         
         # Register signal handlers for graceful shutdown on SSH disconnect / SIGTERM
         def _signal_handler(signum, frame):
-            """Handle SIGHUP/SIGTERM by triggering graceful cleanup."""
+            """Handle SIGHUP/SIGTERM by triggering graceful cleanup.
+
+            Calls ``self.agent.interrupt()`` first so the agent daemon
+            thread's poll loop sees the per-thread interrupt and kills the
+            tool's subprocess group via ``_kill_process`` (os.killpg).
+            Without this, the main thread dies from KeyboardInterrupt and
+            the daemon thread is killed with it — before it can run one
+            more poll iteration to clean up the subprocess, which was
+            spawned with ``os.setsid`` and therefore survives as an orphan
+            with PPID=1.
+
+            Grace window (``HERMES_SIGTERM_GRACE``, default 1.5 s) gives
+            the daemon time to: detect the interrupt (next 200 ms poll) →
+            call _kill_process (SIGTERM + 1 s wait + SIGKILL if needed) →
+            return from _wait_for_process.  ``time.sleep`` releases the
+            GIL so the daemon actually runs during the window.
+            """
             logger.debug("Received signal %s, triggering graceful shutdown", signum)
+            try:
+                if getattr(self, "agent", None) and getattr(self, "_agent_running", False):
+                    self.agent.interrupt(f"received signal {signum}")
+                    import time as _t
+                    try:
+                        _grace = float(os.getenv("HERMES_SIGTERM_GRACE", "1.5"))
+                    except (TypeError, ValueError):
+                        _grace = 1.5
+                    if _grace > 0:
+                        _t.sleep(_grace)
+            except Exception:
+                pass  # never block signal handling
             raise KeyboardInterrupt()
         
         try:
@@ -10263,6 +10429,45 @@ def main(
     
     # Register cleanup for single-query mode (interactive mode registers in run())
     atexit.register(_run_cleanup)
+
+    # Also install signal handlers in single-query / `-q` mode.  Interactive
+    # mode registers its own inside HermesCLI.run(), but `-q` runs
+    # cli.agent.run_conversation() below and AIAgent spawns worker threads
+    # for tools — so when SIGTERM arrives on the main thread, raising
+    # KeyboardInterrupt only unwinds the main thread, not the worker
+    # running _wait_for_process.  Python then exits, the child subprocess
+    # (spawned with os.setsid, its own process group) is reparented to
+    # init and keeps running as an orphan.
+    #
+    # Fix: route SIGTERM/SIGHUP through agent.interrupt() which sets the
+    # per-thread interrupt flag the worker's poll loop checks every 200 ms.
+    # Give the worker a grace window to call _kill_process (SIGTERM to the
+    # process group, then SIGKILL after 1 s), then raise KeyboardInterrupt
+    # so main unwinds normally.  HERMES_SIGTERM_GRACE overrides the 1.5 s
+    # default for debugging.
+    def _signal_handler_q(signum, frame):
+        logger.debug("Received signal %s in single-query mode", signum)
+        try:
+            _agent = getattr(cli, "agent", None)
+            if _agent is not None:
+                _agent.interrupt(f"received signal {signum}")
+                import time as _t
+                try:
+                    _grace = float(os.getenv("HERMES_SIGTERM_GRACE", "1.5"))
+                except (TypeError, ValueError):
+                    _grace = 1.5
+                if _grace > 0:
+                    _t.sleep(_grace)
+        except Exception:
+            pass  # never block signal handling
+        raise KeyboardInterrupt()
+    try:
+        import signal as _signal
+        _signal.signal(_signal.SIGTERM, _signal_handler_q)
+        if hasattr(_signal, "SIGHUP"):
+            _signal.signal(_signal.SIGHUP, _signal_handler_q)
+    except Exception:
+        pass  # signal handler may fail in restricted environments
     
     # Handle single query mode
     if query or image:

cron/scheduler.py

@@ -65,7 +65,15 @@ _HOME_TARGET_ENV_VARS = {
     "wecom": "WECOM_HOME_CHANNEL",
     "weixin": "WEIXIN_HOME_CHANNEL",
     "bluebubbles": "BLUEBUBBLES_HOME_CHANNEL",
-    "qqbot": "QQ_HOME_CHANNEL",
+    "qqbot": "QQBOT_HOME_CHANNEL",
+}
+
+# Legacy env var names kept for back-compat.  Each entry is the current
+# primary env var → the previous name.  _get_home_target_chat_id falls
+# back to the legacy name if the primary is unset, so users who set the
+# old name before the rename keep working until they migrate.
+_LEGACY_HOME_TARGET_ENV_VARS = {
+    "QQBOT_HOME_CHANNEL": "QQ_HOME_CHANNEL",
 }
 
 from cron.jobs import get_due_jobs, mark_job_run, save_job_output, advance_next_run
@@ -100,7 +108,12 @@ def _get_home_target_chat_id(platform_name: str) -> str:
     env_var = _HOME_TARGET_ENV_VARS.get(platform_name.lower())
     if not env_var:
         return ""
-    return os.getenv(env_var, "")
+    value = os.getenv(env_var, "")
+    if not value:
+        legacy = _LEGACY_HOME_TARGET_ENV_VARS.get(env_var)
+        if legacy:
+            value = os.getenv(legacy, "")
+    return value
 
 
 def _resolve_single_delivery_target(job: dict, deliver_value: str) -> Optional[dict]:

docs/plans/2026-04-01-ink-gateway-tui-migration-plan.md

@@ -0,0 +1,108 @@
+# Ink Gateway TUI Migration — Post-mortem
+
+Planned: 2026-04-01 · Delivered: 2026-04 · Status: shipped, classic (prompt_toolkit) CLI still present
+
+## What Shipped
+
+Three layers, same repo, Python runtime unchanged.
+
+```
+ui-tui (Node/TS)  ──stdio JSON-RPC──▶  tui_gateway (Py)  ──▶  AIAgent (run_agent.py)
+```
+
+### Backend — `tui_gateway/`
+
+```
+tui_gateway/
+├── entry.py          # subprocess entrypoint, stdio read/write loop
+├── server.py         # everything: sessions dict, @method handlers, _emit
+├── render.py         # stream renderer, diff rendering, message rendering
+├── slash_worker.py   # subprocess that runs hermes_cli slash commands
+└── __init__.py
+```
+
+`server.py` owns the full runtime-control surface: session store (`_sessions: dict[str, dict]`), method registry (`@method("…")` decorator), event emitter (`_emit`), agent lifecycle (`_make_agent`, `_init_session`, `_wire_callbacks`), approval/sudo/clarify round-trips, and JSON-RPC dispatch.
+
+Protocol methods (`@method(...)` in `server.py`):
+
+- session: `session.{create, resume, list, close, interrupt, usage, history, compress, branch, title, save, undo}`
+- prompt: `prompt.{submit, background, btw}`
+- tools: `tools.{list, show, configure}`
+- slash: `slash.exec`, `command.{dispatch, resolve}`, `commands.catalog`, `complete.{path, slash}`
+- approvals: `approval.respond`, `sudo.respond`, `clarify.respond`, `secret.respond`
+- config/state: `config.{get, set, show}`, `model.options`, `reload.mcp`
+- ops: `shell.exec`, `cli.exec`, `terminal.resize`, `input.detect_drop`, `clipboard.paste`, `paste.collapse`, `image.attach`, `process.stop`
+- misc: `agents.list`, `skills.manage`, `plugins.list`, `cron.manage`, `insights.get`, `rollback.{list, diff, restore}`, `browser.manage`
+
+Protocol events (`_emit(…)` → handled in `ui-tui/src/app/createGatewayEventHandler.ts`):
+
+- lifecycle: `gateway.{ready, stderr}`, `session.info`, `skin.changed`
+- stream: `message.{start, delta, complete}`, `thinking.delta`, `reasoning.{delta, available}`, `status.update`
+- tools: `tool.{start, progress, complete, generating}`, `subagent.{start, thinking, tool, progress, complete}`
+- interactive: `approval.request`, `sudo.request`, `clarify.request`, `secret.request`
+- async: `background.complete`, `btw.complete`, `error`
+
+### Frontend — `ui-tui/src/`
+
+```
+src/
+├── entry.tsx            # node bootstrap: bootBanner → spawn python → dynamic-import Ink → render(<App/>)
+├── app.tsx              # <GatewayProvider> wraps <AppLayout>
+├── bootBanner.ts        # raw-ANSI banner to stdout in ~2ms, pre-React
+├── gatewayClient.ts     # JSON-RPC client over child_process stdio
+├── gatewayTypes.ts      # typed RPC responses + GatewayEvent union
+├── theme.ts             # DEFAULT_THEME + fromSkin
+│
+├── app/                 # hooks + stores — the orchestration layer
+│   ├── uiStore.ts             # nanostore: sid, info, busy, usage, theme, status…
+│   ├── turnStore.ts           # nanostore: per-turn activity / reasoning / tools
+│   ├── turnController.ts      # imperative singleton for stream-time operations
+│   ├── overlayStore.ts        # nanostore: modal/overlay state
+│   ├── useMainApp.ts          # top-level composition hook
+│   ├── useSessionLifecycle.ts # session.create/resume/close/reset
+│   ├── useSubmission.ts       # shell/slash/prompt dispatch + interpolation
+│   ├── useConfigSync.ts       # config.get + mtime poll
+│   ├── useComposerState.ts    # input buffer, paste snippets, editor mode
+│   ├── useInputHandlers.ts    # key bindings
+│   ├── createGatewayEventHandler.ts  # event-stream dispatcher
+│   ├── createSlashHandler.ts         # slash command router (registry + python fallback)
+│   └── slash/commands/        # core.ts, ops.ts, session.ts — TS-owned slash commands
+│
+├── components/          # AppLayout, AppChrome, AppOverlays, MessageLine, Thinking, Markdown, pickers, prompts, Banner, SessionPanel
+├── config/              # env, limits, timing constants
+├── content/             # charms, faces, fortunes, hotkeys, placeholders, verbs
+├── domain/              # details, messages, paths, roles, slash, usage, viewport
+├── protocol/            # interpolation, paste regex
+├── hooks/               # useCompletion, useInputHistory, useQueue, useVirtualHistory
+└── lib/                 # history, messages, osc52, rpc, text
+```
+
+### CLI entry points — `hermes_cli/main.py`
+
+- `hermes --tui`      → `node dist/entry.js` (auto-builds when `.ts`/`.tsx` newer than `dist/entry.js`)
+- `hermes --tui --dev` → `tsx src/entry.tsx` (skip build)
+- `HERMES_TUI_DIR=…`  → external prebuilt dist (nix, distro packaging)
+
+## Diverged From Original Plan
+
+| Plan | Reality | Why |
+|---|---|---|
+| `tui_gateway/{controller,session_state,events,protocol}.py` | all collapsed into `server.py` | no second consumer ever emerged, keeping one file cheaper than four |
+| `ui-tui/src/main.tsx` | split into `entry.tsx` (bootstrap) + `app.tsx` (shell) | boot banner + early python spawn wanted a pre-React moment |
+| `ui-tui/src/state/store.ts` | three nanostores (`uiStore`, `turnStore`, `overlayStore`) | separate lifetimes: ui persists, turn resets per reply, overlay is modal |
+| `approval.requested` / `sudo.requested` / `clarify.requested` | `*.request` (no `-ed`) | cosmetic |
+| `session.cancel` | dropped | `session.interrupt` covers it |
+| `HERMES_EXPERIMENTAL_TUI=1`, `display.experimental_tui: true`, `/tui on/off/status` | none shipped | `--tui` went from opt-in to first-class without an experimental phase |
+
+## Post-migration Additions (not in original plan)
+
+- **Async `session.create`** — returns sid in ~1ms, agent builds on a background thread, `session.info` broadcasts when ready; `_wait_agent()` gates every agent-touching handler via `_sess`
+- **`bootBanner`** — raw-ANSI logo painted to stdout at T≈2ms, before Ink loads; `<AlternateScreen>` wipes it seamlessly when React mounts
+- **Selection uniform bg** — `theme.color.selectionBg` wired via `useSelection().setSelectionBgColor`; replaces SGR-inverse per-cell swap that fragmented over amber/gold fg
+- **Slash command registry** — TS-owned commands in `app/slash/commands/{core,ops,session}.ts`, everything else falls through to `slash.exec` (python worker)
+- **Turn store + controller split** — imperative singleton (`turnController`) holds refs/timers, nanostore (`turnStore`) holds render-visible state
+
+## What's Still Open
+
+- **Classic CLI not deleted.** `cli.py` still has ~80 `prompt_toolkit` references; classic REPL is still the default when `--tui` is absent. The original plan's "Cut 4 · prompt_toolkit removal later" hasn't happened.
+- **No config-file opt-in.** `HERMES_EXPERIMENTAL_TUI` and `display.experimental_tui` were never built; only the CLI flag exists. Fine for now — if we want "default to TUI", a single line in `main.py` flips it.

docs/skins/example-skin.yaml

@@ -6,6 +6,11 @@
 # All fields are optional — missing values inherit from the default skin.
 # Activate with: /skin <name>  or  display.skin: <name> in config.yaml
 #
+# Keys are marked:
+#   (both)    — applies to both the classic CLI and the TUI
+#   (classic) — classic CLI only (see hermes --tui in user-guide/tui.md)
+#   (tui)     — TUI only
+#
 # See hermes_cli/skin_engine.py for the full schema reference.
 # ============================================================================
 
@@ -14,43 +19,47 @@ name: example
 description: An example custom skin — copy and modify this template
 
 # ── Colors ──────────────────────────────────────────────────────────────────
-# Hex color values for Rich markup. These control the CLI's visual palette.
+# Hex color values. These control the visual palette.
 colors:
-  # Banner panel (the startup welcome box)
+  # Banner panel (the startup welcome box) — (both)
   banner_border: "#CD7F32"        # Panel border
   banner_title: "#FFD700"         # Panel title text
   banner_accent: "#FFBF00"        # Section headers (Available Tools, Skills, etc.)
   banner_dim: "#B8860B"           # Dim/muted text (separators, model info)
   banner_text: "#FFF8DC"          # Body text (tool names, skill names)
 
-  # UI elements
-  ui_accent: "#FFBF00"            # General accent color
+  # UI elements — (both)
+  ui_accent: "#FFBF00"            # General accent (falls back to banner_accent)
   ui_label: "#4dd0e1"             # Labels
   ui_ok: "#4caf50"                # Success indicators
   ui_error: "#ef5350"             # Error indicators
   ui_warn: "#ffa726"              # Warning indicators
 
   # Input area
-  prompt: "#FFF8DC"               # Prompt text color
-  input_rule: "#CD7F32"           # Horizontal rule around input
+  prompt: "#FFF8DC"               # Prompt text / `❯` glyph color (both)
+  input_rule: "#CD7F32"           # Horizontal rule above input (classic)
 
-  # Response box
-  response_border: "#FFD700"      # Response box border (ANSI color)
+  # Response box — (classic)
+  response_border: "#FFD700"      # Response box border
 
-  # Session display
-  session_label: "#DAA520"        # Session label
-  session_border: "#8B8682"       # Session ID dim color
+  # Session display — (both)
+  session_label: "#DAA520"        # "Session: " label
+  session_border: "#8B8682"       # Session ID text
 
-  # TUI surfaces
-  status_bar_bg: "#1a1a2e"              # Status / usage bar background
-  voice_status_bg: "#1a1a2e"            # Voice-mode badge background
-  completion_menu_bg: "#1a1a2e"         # Completion list background
-  completion_menu_current_bg: "#333355" # Active completion row background
-  completion_menu_meta_bg: "#1a1a2e"    # Completion meta column background
-  completion_menu_meta_current_bg: "#333355"  # Active completion meta background
+  # TUI / CLI surfaces — (classic: status bar, voice badge, completion meta)
+  status_bar_bg: "#1a1a2e"              # Status / usage bar background (classic)
+  voice_status_bg: "#1a1a2e"            # Voice-mode badge background (classic)
+  completion_menu_bg: "#1a1a2e"         # Completion list background (both)
+  completion_menu_current_bg: "#333355" # Active completion row background (both)
+  completion_menu_meta_bg: "#1a1a2e"    # Completion meta column bg (classic)
+  completion_menu_meta_current_bg: "#333355"  # Active meta bg (classic)
+
+  # Drag-to-select background — (tui)
+  selection_bg: "#3a3a55"               # Uniform selection highlight in the TUI
 
 # ── Spinner ─────────────────────────────────────────────────────────────────
-# Customize the animated spinner shown during API calls and tool execution.
+# (classic) — the TUI uses its own animated indicators; spinner config here
+# is only read by the classic prompt_toolkit CLI.
 spinner:
   # Faces shown while waiting for the API response
   waiting_faces:
@@ -78,17 +87,17 @@ spinner:
   #   - ["⟪▲", "▲⟫"]
 
 # ── Branding ────────────────────────────────────────────────────────────────
-# Text strings used throughout the CLI interface.
+# Text strings used throughout the interface.
 branding:
-  agent_name: "Hermes Agent"          # Banner title, about display
-  welcome: "Welcome! Type your message or /help for commands."
-  goodbye: "Goodbye! ⚕"              # Exit message
-  response_label: " ⚕ Hermes "       # Response box header label
-  prompt_symbol: "❯ "                 # Input prompt symbol
-  help_header: "(^_^)? Available Commands"  # /help header text
+  agent_name: "Hermes Agent"                  # (both) Banner title, about display
+  welcome: "Welcome! Type your message or /help for commands."  # (both)
+  goodbye: "Goodbye! ⚕"                       # (both) Exit message
+  response_label: " ⚕ Hermes "                # (classic) Response box header label
+  prompt_symbol: "❯ "                          # (both) Input prompt glyph
+  help_header: "(^_^)? Available Commands"     # (both) /help overlay title
 
 # ── Tool Output ─────────────────────────────────────────────────────────────
-# Character used as the prefix for tool output lines.
+# Character used as the prefix for tool output lines. (both)
 # Default is "┊" (thin dotted vertical line). Some alternatives:
 #   "╎" (light triple dash vertical)
 #   "▏" (left one-eighth block)

flake.lock

@@ -36,6 +36,26 @@
         "type": "github"
       }
     },
+    "npm-lockfile-fix": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1775903712,
+        "narHash": "sha256-2GV79U6iVH4gKAPWYrxUReB0S41ty/Y3dBLquU8AlaA=",
+        "owner": "jeslie0",
+        "repo": "npm-lockfile-fix",
+        "rev": "c6093acb0c0548e0f9b8b3d82918823721930fe8",
+        "type": "github"
+      },
+      "original": {
+        "owner": "jeslie0",
+        "repo": "npm-lockfile-fix",
+        "type": "github"
+      }
+    },
     "pyproject-build-systems": {
       "inputs": {
         "nixpkgs": [
@@ -124,6 +144,7 @@
       "inputs": {
         "flake-parts": "flake-parts",
         "nixpkgs": "nixpkgs",
+        "npm-lockfile-fix": "npm-lockfile-fix",
         "pyproject-build-systems": "pyproject-build-systems",
         "pyproject-nix": "pyproject-nix_2",
         "uv2nix": "uv2nix_2"

flake.nix

@@ -19,11 +19,20 @@
       url = "github:pyproject-nix/build-system-pkgs";
       inputs.nixpkgs.follows = "nixpkgs";
     };
+    npm-lockfile-fix = {
+      url = "github:jeslie0/npm-lockfile-fix";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
   };
 
-  outputs = inputs:
+  outputs =
+    inputs:
     inputs.flake-parts.lib.mkFlake { inherit inputs; } {
-      systems = [ "x86_64-linux" "aarch64-linux" "aarch64-darwin" ];
+      systems = [
+        "x86_64-linux"
+        "aarch64-linux"
+        "aarch64-darwin"
+      ];
 
       imports = [
         ./nix/packages.nix

gateway/channel_directory.py

@@ -100,7 +100,7 @@ def build_channel_directory(adapters: Dict[Any, Any]) -> Dict[str, Any]:
 
 
 def _build_discord(adapter) -> List[Dict[str, str]]:
-    """Enumerate all text channels the Discord bot can see."""
+    """Enumerate all text channels and forum channels the Discord bot can see."""
     channels = []
     client = getattr(adapter, "_client", None)
     if not client:
@@ -119,6 +119,15 @@ def _build_discord(adapter) -> List[Dict[str, str]]:
                 "guild": guild.name,
                 "type": "channel",
             })
+        # Forum channels (type 15) — creating a message auto-spawns a thread post.
+        forums = getattr(guild, "forum_channels", None) or []
+        for ch in forums:
+            channels.append({
+                "id": str(ch.id),
+                "name": ch.name,
+                "guild": guild.name,
+                "type": "forum",
+            })
         # Also include DM-capable users we've interacted with is not
         # feasible via guild enumeration; those come from sessions.
 
@@ -191,6 +200,15 @@ def load_directory() -> Dict[str, Any]:
         return {"updated_at": None, "platforms": {}}
 
 
+def lookup_channel_type(platform_name: str, chat_id: str) -> Optional[str]:
+    """Return the channel ``type`` string (e.g. ``"channel"``, ``"forum"``) for *chat_id*, or *None* if unknown."""
+    directory = load_directory()
+    for ch in directory.get("platforms", {}).get(platform_name, []):
+        if ch.get("id") == chat_id:
+            return ch.get("type")
+    return None
+
+
 def resolve_channel_name(platform_name: str, name: str) -> Optional[str]:
     """
     Resolve a human-friendly channel name to a numeric ID.

gateway/config.py

@@ -258,6 +258,13 @@ class GatewayConfig:
     # Streaming configuration
     streaming: StreamingConfig = field(default_factory=StreamingConfig)
 
+    # Session store pruning: drop SessionEntry records older than this many
+    # days from the in-memory dict and sessions.json.  Keeps the store from
+    # growing unbounded in gateways serving many chats/threads/users over
+    # months.  Pruning is invisible to users — if they resume, they get a
+    # fresh session exactly as if the reset policy had fired.  0 = disabled.
+    session_store_max_age_days: int = 90
+
     def get_connected_platforms(self) -> List[Platform]:
         """Return list of platforms that are enabled and configured."""
         connected = []
@@ -365,6 +372,7 @@ class GatewayConfig:
             "thread_sessions_per_user": self.thread_sessions_per_user,
             "unauthorized_dm_behavior": self.unauthorized_dm_behavior,
             "streaming": self.streaming.to_dict(),
+            "session_store_max_age_days": self.session_store_max_age_days,
         }
     
     @classmethod
@@ -412,6 +420,13 @@ class GatewayConfig:
             "pair",
         )
 
+        try:
+            session_store_max_age_days = int(data.get("session_store_max_age_days", 90))
+            if session_store_max_age_days < 0:
+                session_store_max_age_days = 0
+        except (TypeError, ValueError):
+            session_store_max_age_days = 90
+
         return cls(
             platforms=platforms,
             default_reset_policy=default_policy,
@@ -426,6 +441,7 @@ class GatewayConfig:
             thread_sessions_per_user=_coerce_bool(thread_sessions_per_user, False),
             unauthorized_dm_behavior=unauthorized_dm_behavior,
             streaming=StreamingConfig.from_dict(data.get("streaming", {})),
+            session_store_max_age_days=session_store_max_age_days,
         )
 
     def get_unauthorized_dm_behavior(self, platform: Optional[Platform] = None) -> str:
@@ -1213,12 +1229,24 @@ def _apply_env_overrides(config: GatewayConfig) -> None:
         qq_group_allowed = os.getenv("QQ_GROUP_ALLOWED_USERS", "").strip()
         if qq_group_allowed:
             extra["group_allow_from"] = qq_group_allowed
-        qq_home = os.getenv("QQ_HOME_CHANNEL", "").strip()
+        qq_home = os.getenv("QQBOT_HOME_CHANNEL", "").strip()
+        qq_home_name_env = "QQBOT_HOME_CHANNEL_NAME"
+        if not qq_home:
+            # Back-compat: accept the pre-rename name and log a one-time warning.
+            legacy_home = os.getenv("QQ_HOME_CHANNEL", "").strip()
+            if legacy_home:
+                qq_home = legacy_home
+                qq_home_name_env = "QQ_HOME_CHANNEL_NAME"
+                import logging
+                logging.getLogger(__name__).warning(
+                    "QQ_HOME_CHANNEL is deprecated; rename to QQBOT_HOME_CHANNEL "
+                    "in your .env for consistency with the platform key."
+                )
         if qq_home:
             config.platforms[Platform.QQBOT].home_channel = HomeChannel(
                 platform=Platform.QQBOT,
                 chat_id=qq_home,
-                name=os.getenv("QQ_HOME_CHANNEL_NAME", "Home"),
+                name=os.getenv("QQBOT_HOME_CHANNEL_NAME") or os.getenv(qq_home_name_env, "Home"),
             )
 
     # Session settings

gateway/platforms/base.py

@@ -669,6 +669,15 @@ class MessageEvent:
     # Original platform data
     raw_message: Any = None
     message_id: Optional[str] = None
+
+    # Platform-specific update identifier.  For Telegram this is the
+    # ``update_id`` from the PTB Update wrapper; other platforms currently
+    # ignore it.  Used by ``/restart`` to record the triggering update so the
+    # new gateway can advance the Telegram offset past it and avoid processing
+    # the same ``/restart`` twice if PTB's graceful-shutdown ACK times out
+    # ("Error while calling `get_updates` one more time to mark all fetched
+    # updates" in gateway.log).
+    platform_update_id: Optional[int] = None
     
     # Media attachments
     # media_urls: local file paths (for vision tool access)
@@ -1045,16 +1054,40 @@ class BasePlatformAdapter(ABC):
         """
         pass
 
+    # Default: the adapter treats ``finalize=True`` on edit_message as a
+    # no-op and is happy to have the stream consumer skip redundant final
+    # edits.  Subclasses that *require* an explicit finalize call to close
+    # out the message lifecycle (e.g. rich card / AI assistant surfaces
+    # such as DingTalk AI Cards) override this to True (class attribute or
+    # property) so the stream consumer knows not to short-circuit.
+    REQUIRES_EDIT_FINALIZE: bool = False
+
     async def edit_message(
         self,
         chat_id: str,
         message_id: str,
         content: str,
+        *,
+        finalize: bool = False,
     ) -> SendResult:
         """
         Edit a previously sent message. Optional — platforms that don't
         support editing return success=False and callers fall back to
         sending a new message.
+
+        ``finalize`` signals that this is the last edit in a streaming
+        sequence.  Most platforms (Telegram, Slack, Discord, Matrix,
+        etc.) treat it as a no-op because their edit APIs have no notion
+        of message lifecycle state — an edit is an edit.  Platforms that
+        render streaming updates with a distinct "in progress" state and
+        require explicit closure (e.g. rich card / AI assistant surfaces
+        such as DingTalk AI Cards) use it to finalize the message and
+        transition the UI out of the streaming indicator — those should
+        also set ``REQUIRES_EDIT_FINALIZE = True`` so callers route a
+        final edit through even when content is unchanged.  Callers
+        should set ``finalize=True`` on the final edit of a streamed
+        response (typically when ``got_done`` fires in the stream
+        consumer) and leave it ``False`` on intermediate edits.
         """
         return SendResult(success=False, error="Not supported")
 
@@ -1579,7 +1612,9 @@ class BasePlatformAdapter(ABC):
             # session lifecycle and its cleanup races with the running task
             # (see PR #4926).
             cmd = event.get_command()
-            if cmd in ("approve", "deny", "status", "stop", "new", "reset", "background", "restart", "queue", "q"):
+            from hermes_cli.commands import should_bypass_active_session
+
+            if should_bypass_active_session(cmd):
                 logger.debug(
                     "[%s] Command '/%s' bypassing active-session guard for %s",
                     self.name, cmd, session_key,

gateway/platforms/discord.py

@@ -857,6 +857,9 @@ class DiscordAdapter(BasePlatformAdapter):
 
         When metadata contains a thread_id, the message is sent to that
         thread instead of the parent channel identified by chat_id.
+
+        Forum channels (type 15) reject direct messages — a thread post is
+        created automatically.
         """
         if not self._client:
             return SendResult(success=False, error="Not connected")
@@ -882,6 +885,10 @@ class DiscordAdapter(BasePlatformAdapter):
                 if not channel:
                     return SendResult(success=False, error=f"Channel {chat_id} not found")
 
+            # Forum channels reject channel.send() — create a thread post instead.
+            if self._is_forum_parent(channel):
+                return await self._send_to_forum(channel, content)
+
             # Format and split message if needed
             formatted = self.format_message(content)
             chunks = self.truncate_message(formatted, self.MAX_MESSAGE_LENGTH)
@@ -945,6 +952,120 @@ class DiscordAdapter(BasePlatformAdapter):
             logger.error("[%s] Failed to send Discord message: %s", self.name, e, exc_info=True)
             return SendResult(success=False, error=str(e))
 
+    async def _send_to_forum(self, forum_channel: Any, content: str) -> SendResult:
+        """Create a thread post in a forum channel with the message as starter content.
+
+        Forum channels (type 15) don't support direct messages.  Instead we
+        POST to /channels/{forum_id}/threads with a thread name derived from
+        the first line of the message.  Any follow-up chunk failures are
+        reported in ``raw_response['warnings']`` so the caller can surface
+        partial-send issues.
+        """
+        from tools.send_message_tool import _derive_forum_thread_name
+
+        formatted = self.format_message(content)
+        chunks = self.truncate_message(formatted, self.MAX_MESSAGE_LENGTH)
+
+        thread_name = _derive_forum_thread_name(content)
+
+        starter_content = chunks[0] if chunks else thread_name
+
+        try:
+            thread = await forum_channel.create_thread(
+                name=thread_name,
+                content=starter_content,
+            )
+        except Exception as e:
+            logger.error("[%s] Failed to create forum thread in %s: %s", self.name, forum_channel.id, e)
+            return SendResult(success=False, error=f"Forum thread creation failed: {e}")
+
+        thread_channel = thread if hasattr(thread, "send") else getattr(thread, "thread", None)
+        thread_id = str(getattr(thread_channel, "id", getattr(thread, "id", "")))
+        starter_msg = getattr(thread, "message", None)
+        message_id = str(getattr(starter_msg, "id", thread_id)) if starter_msg else thread_id
+
+        # Send remaining chunks into the newly created thread.  Track any
+        # per-chunk failures so the caller sees partial-send outcomes.
+        message_ids = [message_id]
+        warnings: list[str] = []
+        for chunk in chunks[1:]:
+            try:
+                msg = await thread_channel.send(content=chunk)
+                message_ids.append(str(msg.id))
+            except Exception as e:
+                warning = f"Failed to send follow-up chunk to forum thread {thread_id}: {e}"
+                logger.warning("[%s] %s", self.name, warning)
+                warnings.append(warning)
+
+        raw_response: Dict[str, Any] = {"message_ids": message_ids, "thread_id": thread_id}
+        if warnings:
+            raw_response["warnings"] = warnings
+
+        return SendResult(
+            success=True,
+            message_id=message_ids[0],
+            raw_response=raw_response,
+        )
+
+    async def _forum_post_file(
+        self,
+        forum_channel: Any,
+        *,
+        thread_name: Optional[str] = None,
+        content: str = "",
+        file: Any = None,
+        files: Optional[list] = None,
+    ) -> SendResult:
+        """Create a forum thread whose starter message carries file attachments.
+
+        Used by the send_voice / send_image_file / send_document paths when
+        the target channel is a forum (type 15).  ``create_thread`` on a
+        ForumChannel accepts the same file/files/content kwargs as
+        ``channel.send``, creating the thread and starter message atomically.
+        """
+        from tools.send_message_tool import _derive_forum_thread_name
+
+        if not thread_name:
+            # Prefer the text content, fall back to the first attached
+            # filename, fall back to the generic default.
+            hint = content or ""
+            if not hint.strip():
+                if file is not None:
+                    hint = getattr(file, "filename", "") or ""
+                elif files:
+                    hint = getattr(files[0], "filename", "") or ""
+            thread_name = _derive_forum_thread_name(hint) if hint.strip() else "New Post"
+
+        kwargs: Dict[str, Any] = {"name": thread_name}
+        if content:
+            kwargs["content"] = content
+        if file is not None:
+            kwargs["file"] = file
+        if files:
+            kwargs["files"] = files
+
+        try:
+            thread = await forum_channel.create_thread(**kwargs)
+        except Exception as e:
+            logger.error(
+                "[%s] Failed to create forum thread with file in %s: %s",
+                self.name,
+                getattr(forum_channel, "id", "?"),
+                e,
+            )
+            return SendResult(success=False, error=f"Forum thread creation failed: {e}")
+
+        thread_channel = thread if hasattr(thread, "send") else getattr(thread, "thread", None)
+        thread_id = str(getattr(thread_channel, "id", getattr(thread, "id", "")))
+        starter_msg = getattr(thread, "message", None)
+        message_id = str(getattr(starter_msg, "id", thread_id)) if starter_msg else thread_id
+
+        return SendResult(
+            success=True,
+            message_id=message_id,
+            raw_response={"thread_id": thread_id},
+        )
+
     async def edit_message(
         self,
         chat_id: str,
@@ -975,7 +1096,11 @@ class DiscordAdapter(BasePlatformAdapter):
         caption: Optional[str] = None,
         file_name: Optional[str] = None,
     ) -> SendResult:
-        """Send a local file as a Discord attachment."""
+        """Send a local file as a Discord attachment.
+
+        Forum channels (type 15) get a new thread whose starter message
+        carries the file — they reject direct POST /messages.
+        """
         if not self._client:
             return SendResult(success=False, error="Not connected")
 
@@ -988,6 +1113,12 @@ class DiscordAdapter(BasePlatformAdapter):
         filename = file_name or os.path.basename(file_path)
         with open(file_path, "rb") as fh:
             file = discord.File(fh, filename=filename)
+            if self._is_forum_parent(channel):
+                return await self._forum_post_file(
+                    channel,
+                    content=(caption or "").strip(),
+                    file=file,
+                )
             msg = await channel.send(content=caption if caption else None, file=file)
         return SendResult(success=True, message_id=str(msg.id))
 
@@ -1036,6 +1167,18 @@ class DiscordAdapter(BasePlatformAdapter):
             with open(audio_path, "rb") as f:
                 file_data = f.read()
 
+            # Forum channels (type 15) reject direct POST /messages — the
+            # native voice flag path also targets /messages so it would fail
+            # too.  Create a thread post with the audio as the starter
+            # attachment instead.
+            if self._is_forum_parent(channel):
+                forum_file = discord.File(io.BytesIO(file_data), filename=filename)
+                return await self._forum_post_file(
+                    channel,
+                    content=(caption or "").strip(),
+                    file=forum_file,
+                )
+
             # Try sending as a native voice message via raw API (flags=8192).
             try:
                 import base64
@@ -1488,6 +1631,13 @@ class DiscordAdapter(BasePlatformAdapter):
                     import io
                     file = discord.File(io.BytesIO(image_data), filename=f"image.{ext}")
 
+                    if self._is_forum_parent(channel):
+                        return await self._forum_post_file(
+                            channel,
+                            content=(caption or "").strip(),
+                            file=file,
+                        )
+
                     msg = await channel.send(
                         content=caption if caption else None,
                         file=file,
@@ -1550,6 +1700,13 @@ class DiscordAdapter(BasePlatformAdapter):
                     import io
                     file = discord.File(io.BytesIO(animation_data), filename="animation.gif")
 
+                    if self._is_forum_parent(channel):
+                        return await self._forum_post_file(
+                            channel,
+                            content=(caption or "").strip(),
+                            file=file,
+                        )
+
                     msg = await channel.send(
                         content=caption if caption else None,
                         file=file,
@@ -1837,6 +1994,11 @@ class DiscordAdapter(BasePlatformAdapter):
         async def slash_stop(interaction: discord.Interaction):
             await self._run_simple_slash(interaction, "/stop", "Stop requested~")
 
+        @tree.command(name="steer", description="Inject a message after the next tool call (no interrupt)")
+        @discord.app_commands.describe(prompt="Text to inject into the agent's next tool result")
+        async def slash_steer(interaction: discord.Interaction, prompt: str):
+            await self._run_simple_slash(interaction, f"/steer {prompt}".strip())
+
         @tree.command(name="compress", description="Compress conversation context")
         async def slash_compress(interaction: discord.Interaction):
             await self._run_simple_slash(interaction, "/compress")

gateway/platforms/feishu.py

@@ -1228,6 +1228,10 @@ class FeishuAdapter(BasePlatformAdapter):
             .register_p2_im_chat_member_bot_deleted_v1(self._on_bot_removed_from_chat)
             .register_p2_im_chat_access_event_bot_p2p_chat_entered_v1(self._on_p2p_chat_entered)
             .register_p2_im_message_recalled_v1(self._on_message_recalled)
+            .register_p2_customized_event(
+                "drive.notice.comment_add_v1",
+                self._on_drive_comment_event,
+            )
             .build()
         )
 
@@ -1965,6 +1969,25 @@ class FeishuAdapter(BasePlatformAdapter):
     def _on_message_recalled(self, data: Any) -> None:
         logger.debug("[Feishu] Message recalled by user")
 
+    def _on_drive_comment_event(self, data: Any) -> None:
+        """Handle drive document comment notification (drive.notice.comment_add_v1).
+
+        Delegates to :mod:`gateway.platforms.feishu_comment` for parsing,
+        logging, and reaction.  Scheduling follows the same
+        ``run_coroutine_threadsafe`` pattern used by ``_on_message_event``.
+        """
+        from gateway.platforms.feishu_comment import handle_drive_comment_event
+
+        loop = self._loop
+        if not self._loop_accepts_callbacks(loop):
+            logger.warning("[Feishu] Dropping drive comment event before adapter loop is ready")
+            return
+        future = asyncio.run_coroutine_threadsafe(
+            handle_drive_comment_event(self._client, data, self_open_id=self._bot_open_id),
+            loop,
+        )
+        future.add_done_callback(self._log_background_failure)
+
     def _on_reaction_event(self, event_type: str, data: Any) -> None:
         """Route user reactions on bot messages as synthetic text events."""
         event = getattr(data, "event", None)
@@ -2590,6 +2613,8 @@ class FeishuAdapter(BasePlatformAdapter):
             self._on_reaction_event(event_type, data)
         elif event_type == "card.action.trigger":
             self._on_card_action_trigger(data)
+        elif event_type == "drive.notice.comment_add_v1":
+            self._on_drive_comment_event(data)
         else:
             logger.debug("[Feishu] Ignoring webhook event type: %s", event_type or "unknown")
         return web.json_response({"code": 0, "msg": "ok"})

gateway/platforms/feishu_comment_rules.py

@@ -0,0 +1,429 @@
+"""
+Feishu document comment access-control rules.
+
+3-tier rule resolution: exact doc > wildcard "*" > top-level > code defaults.
+Each field (enabled/policy/allow_from) falls back independently.
+Config: ~/.hermes/feishu_comment_rules.json (mtime-cached, hot-reload).
+Pairing store: ~/.hermes/feishu_comment_pairing.json.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from hermes_constants import get_hermes_home
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Paths
+# ---------------------------------------------------------------------------
+#
+# Uses the canonical ``get_hermes_home()`` helper (HERMES_HOME-aware and
+# profile-safe). Resolved at import time; this module is lazy-imported by
+# the Feishu comment event handler, which runs long after profile overrides
+# have been applied, so freezing paths here is safe.
+
+RULES_FILE = get_hermes_home() / "feishu_comment_rules.json"
+PAIRING_FILE = get_hermes_home() / "feishu_comment_pairing.json"
+
+# ---------------------------------------------------------------------------
+# Data models
+# ---------------------------------------------------------------------------
+
+_VALID_POLICIES = ("allowlist", "pairing")
+
+
+@dataclass(frozen=True)
+class CommentDocumentRule:
+    """Per-document rule.  ``None`` means 'inherit from lower tier'."""
+    enabled: Optional[bool] = None
+    policy: Optional[str] = None
+    allow_from: Optional[frozenset] = None
+
+
+@dataclass(frozen=True)
+class CommentsConfig:
+    """Top-level comment access config."""
+    enabled: bool = True
+    policy: str = "pairing"
+    allow_from: frozenset = field(default_factory=frozenset)
+    documents: Dict[str, CommentDocumentRule] = field(default_factory=dict)
+
+
+@dataclass(frozen=True)
+class ResolvedCommentRule:
+    """Fully resolved rule after field-by-field fallback."""
+    enabled: bool
+    policy: str
+    allow_from: frozenset
+    match_source: str  # e.g. "exact:docx:xxx" | "wildcard" | "top" | "default"
+
+
+# ---------------------------------------------------------------------------
+# Mtime-cached file loading
+# ---------------------------------------------------------------------------
+
+class _MtimeCache:
+    """Generic mtime-based file cache.  ``stat()`` per access, re-read only on change."""
+
+    def __init__(self, path: Path):
+        self._path = path
+        self._mtime: float = 0.0
+        self._data: Optional[dict] = None
+
+    def load(self) -> dict:
+        try:
+            st = self._path.stat()
+            mtime = st.st_mtime
+        except FileNotFoundError:
+            self._mtime = 0.0
+            self._data = {}
+            return {}
+
+        if mtime == self._mtime and self._data is not None:
+            return self._data
+
+        try:
+            with open(self._path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+            if not isinstance(data, dict):
+                data = {}
+        except (json.JSONDecodeError, OSError):
+            logger.warning("[Feishu-Rules] Failed to read %s, using empty config", self._path)
+            data = {}
+
+        self._mtime = mtime
+        self._data = data
+        return data
+
+
+_rules_cache = _MtimeCache(RULES_FILE)
+_pairing_cache = _MtimeCache(PAIRING_FILE)
+
+
+# ---------------------------------------------------------------------------
+# Config parsing
+# ---------------------------------------------------------------------------
+
+def _parse_frozenset(raw: Any) -> Optional[frozenset]:
+    """Parse a list of strings into a frozenset; return None if key absent."""
+    if raw is None:
+        return None
+    if isinstance(raw, (list, tuple)):
+        return frozenset(str(u).strip() for u in raw if str(u).strip())
+    return None
+
+
+def _parse_document_rule(raw: dict) -> CommentDocumentRule:
+    enabled = raw.get("enabled")
+    if enabled is not None:
+        enabled = bool(enabled)
+    policy = raw.get("policy")
+    if policy is not None:
+        policy = str(policy).strip().lower()
+        if policy not in _VALID_POLICIES:
+            policy = None
+    allow_from = _parse_frozenset(raw.get("allow_from"))
+    return CommentDocumentRule(enabled=enabled, policy=policy, allow_from=allow_from)
+
+
+def load_config() -> CommentsConfig:
+    """Load comment rules from disk (mtime-cached)."""
+    raw = _rules_cache.load()
+    if not raw:
+        return CommentsConfig()
+
+    documents: Dict[str, CommentDocumentRule] = {}
+    raw_docs = raw.get("documents", {})
+    if isinstance(raw_docs, dict):
+        for key, rule_raw in raw_docs.items():
+            if isinstance(rule_raw, dict):
+                documents[str(key)] = _parse_document_rule(rule_raw)
+
+    policy = str(raw.get("policy", "pairing")).strip().lower()
+    if policy not in _VALID_POLICIES:
+        policy = "pairing"
+
+    return CommentsConfig(
+        enabled=raw.get("enabled", True),
+        policy=policy,
+        allow_from=_parse_frozenset(raw.get("allow_from")) or frozenset(),
+        documents=documents,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Rule resolution  (§8.4 field-by-field fallback)
+# ---------------------------------------------------------------------------
+
+def has_wiki_keys(cfg: CommentsConfig) -> bool:
+    """Check if any document rule key starts with 'wiki:'."""
+    return any(k.startswith("wiki:") for k in cfg.documents)
+
+
+def resolve_rule(
+    cfg: CommentsConfig,
+    file_type: str,
+    file_token: str,
+    wiki_token: str = "",
+) -> ResolvedCommentRule:
+    """Resolve effective rule: exact doc → wiki key → wildcard → top-level → defaults."""
+    exact_key = f"{file_type}:{file_token}"
+
+    exact = cfg.documents.get(exact_key)
+    exact_src = f"exact:{exact_key}"
+    if exact is None and wiki_token:
+        wiki_key = f"wiki:{wiki_token}"
+        exact = cfg.documents.get(wiki_key)
+        exact_src = f"exact:{wiki_key}"
+
+    wildcard = cfg.documents.get("*")
+
+    layers = []
+    if exact is not None:
+        layers.append((exact, exact_src))
+    if wildcard is not None:
+        layers.append((wildcard, "wildcard"))
+
+    def _pick(field_name: str):
+        for layer, source in layers:
+            val = getattr(layer, field_name)
+            if val is not None:
+                return val, source
+        return getattr(cfg, field_name), "top"
+
+    enabled, en_src = _pick("enabled")
+    policy, pol_src = _pick("policy")
+    allow_from, _ = _pick("allow_from")
+
+    # match_source = highest-priority tier that contributed any field
+    priority_order = {"exact": 0, "wildcard": 1, "top": 2}
+    best_src = min(
+        [en_src, pol_src],
+        key=lambda s: priority_order.get(s.split(":")[0], 3),
+    )
+
+    return ResolvedCommentRule(
+        enabled=enabled,
+        policy=policy,
+        allow_from=allow_from,
+        match_source=best_src,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Pairing store
+# ---------------------------------------------------------------------------
+
+def _load_pairing_approved() -> set:
+    """Return set of approved user open_ids (mtime-cached)."""
+    data = _pairing_cache.load()
+    approved = data.get("approved", {})
+    if isinstance(approved, dict):
+        return set(approved.keys())
+    if isinstance(approved, list):
+        return set(str(u) for u in approved if u)
+    return set()
+
+
+def _save_pairing(data: dict) -> None:
+    PAIRING_FILE.parent.mkdir(parents=True, exist_ok=True)
+    tmp = PAIRING_FILE.with_suffix(".tmp")
+    with open(tmp, "w", encoding="utf-8") as f:
+        json.dump(data, f, indent=2, ensure_ascii=False)
+    tmp.replace(PAIRING_FILE)
+    # Invalidate cache so next load picks up change
+    _pairing_cache._mtime = 0.0
+    _pairing_cache._data = None
+
+
+def pairing_add(user_open_id: str) -> bool:
+    """Add a user to the pairing-approved list. Returns True if newly added."""
+    data = _pairing_cache.load()
+    approved = data.get("approved", {})
+    if not isinstance(approved, dict):
+        approved = {}
+    if user_open_id in approved:
+        return False
+    approved[user_open_id] = {"approved_at": time.time()}
+    data["approved"] = approved
+    _save_pairing(data)
+    return True
+
+
+def pairing_remove(user_open_id: str) -> bool:
+    """Remove a user from the pairing-approved list. Returns True if removed."""
+    data = _pairing_cache.load()
+    approved = data.get("approved", {})
+    if not isinstance(approved, dict):
+        return False
+    if user_open_id not in approved:
+        return False
+    del approved[user_open_id]
+    data["approved"] = approved
+    _save_pairing(data)
+    return True
+
+
+def pairing_list() -> Dict[str, Any]:
+    """Return the approved dict  {user_open_id: {approved_at: ...}}."""
+    data = _pairing_cache.load()
+    approved = data.get("approved", {})
+    return dict(approved) if isinstance(approved, dict) else {}
+
+
+# ---------------------------------------------------------------------------
+# Access check  (public API for feishu_comment.py)
+# ---------------------------------------------------------------------------
+
+def is_user_allowed(rule: ResolvedCommentRule, user_open_id: str) -> bool:
+    """Check if user passes the resolved rule's policy gate."""
+    if user_open_id in rule.allow_from:
+        return True
+    if rule.policy == "pairing":
+        return user_open_id in _load_pairing_approved()
+    return False
+
+
+# ---------------------------------------------------------------------------
+# CLI
+# ---------------------------------------------------------------------------
+
+def _print_status() -> None:
+    cfg = load_config()
+    print(f"Rules file: {RULES_FILE}")
+    print(f"  exists: {RULES_FILE.exists()}")
+    print(f"Pairing file: {PAIRING_FILE}")
+    print(f"  exists: {PAIRING_FILE.exists()}")
+    print()
+    print(f"Top-level:")
+    print(f"  enabled:    {cfg.enabled}")
+    print(f"  policy:     {cfg.policy}")
+    print(f"  allow_from: {sorted(cfg.allow_from) if cfg.allow_from else '[]'}")
+    print()
+    if cfg.documents:
+        print(f"Document rules ({len(cfg.documents)}):")
+        for key, rule in sorted(cfg.documents.items()):
+            parts = []
+            if rule.enabled is not None:
+                parts.append(f"enabled={rule.enabled}")
+            if rule.policy is not None:
+                parts.append(f"policy={rule.policy}")
+            if rule.allow_from is not None:
+                parts.append(f"allow_from={sorted(rule.allow_from)}")
+            print(f"  [{key}] {', '.join(parts) if parts else '(empty — inherits all)'}")
+    else:
+        print("Document rules: (none)")
+    print()
+    approved = pairing_list()
+    print(f"Pairing approved ({len(approved)}):")
+    for uid, meta in sorted(approved.items()):
+        ts = meta.get("approved_at", 0)
+        print(f"  {uid}  (approved_at={ts})")
+
+
+def _do_check(doc_key: str, user_open_id: str) -> None:
+    cfg = load_config()
+    parts = doc_key.split(":", 1)
+    if len(parts) != 2:
+        print(f"Error: doc_key must be 'fileType:fileToken', got '{doc_key}'")
+        return
+    file_type, file_token = parts
+    rule = resolve_rule(cfg, file_type, file_token)
+    allowed = is_user_allowed(rule, user_open_id)
+    print(f"Document:     {doc_key}")
+    print(f"User:         {user_open_id}")
+    print(f"Resolved rule:")
+    print(f"  enabled:      {rule.enabled}")
+    print(f"  policy:       {rule.policy}")
+    print(f"  allow_from:   {sorted(rule.allow_from) if rule.allow_from else '[]'}")
+    print(f"  match_source: {rule.match_source}")
+    print(f"Result:       {'ALLOWED' if allowed else 'DENIED'}")
+
+
+def _main() -> int:
+    import sys
+
+    try:
+        from hermes_cli.env_loader import load_hermes_dotenv
+        load_hermes_dotenv()
+    except Exception:
+        pass
+
+    usage = (
+        "Usage: python -m gateway.platforms.feishu_comment_rules <command> [args]\n"
+        "\n"
+        "Commands:\n"
+        "  status                              Show rules config and pairing state\n"
+        "  check <fileType:token> <user>        Simulate access check\n"
+        "  pairing add <user_open_id>           Add user to pairing-approved list\n"
+        "  pairing remove <user_open_id>        Remove user from pairing-approved list\n"
+        "  pairing list                         List pairing-approved users\n"
+        "\n"
+        f"Rules config file: {RULES_FILE}\n"
+        "  Edit this JSON file directly to configure policies and document rules.\n"
+        "  Changes take effect on the next comment event (no restart needed).\n"
+    )
+
+    args = sys.argv[1:]
+    if not args:
+        print(usage)
+        return 1
+
+    cmd = args[0]
+
+    if cmd == "status":
+        _print_status()
+
+    elif cmd == "check":
+        if len(args) < 3:
+            print("Usage: check <fileType:fileToken> <user_open_id>")
+            return 1
+        _do_check(args[1], args[2])
+
+    elif cmd == "pairing":
+        if len(args) < 2:
+            print("Usage: pairing <add|remove|list> [args]")
+            return 1
+        sub = args[1]
+        if sub == "add":
+            if len(args) < 3:
+                print("Usage: pairing add <user_open_id>")
+                return 1
+            if pairing_add(args[2]):
+                print(f"Added: {args[2]}")
+            else:
+                print(f"Already approved: {args[2]}")
+        elif sub == "remove":
+            if len(args) < 3:
+                print("Usage: pairing remove <user_open_id>")
+                return 1
+            if pairing_remove(args[2]):
+                print(f"Removed: {args[2]}")
+            else:
+                print(f"Not in approved list: {args[2]}")
+        elif sub == "list":
+            approved = pairing_list()
+            if not approved:
+                print("(no approved users)")
+            for uid, meta in sorted(approved.items()):
+                print(f"  {uid}  approved_at={meta.get('approved_at', '?')}")
+        else:
+            print(f"Unknown pairing subcommand: {sub}")
+            return 1
+    else:
+        print(f"Unknown command: {cmd}\n")
+        print(usage)
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(_main())

gateway/platforms/qqbot/__init__.py

@@ -0,0 +1,57 @@
+"""
+QQBot platform package.
+
+Re-exports the main adapter symbols from ``adapter.py`` (the original
+``qqbot.py``) so that **all existing import paths remain unchanged**::
+
+    from gateway.platforms.qqbot import QQAdapter          # works
+    from gateway.platforms.qqbot import check_qq_requirements  # works
+
+New modules:
+    - ``constants`` — shared constants (API URLs, timeouts, message types)
+    - ``utils`` — User-Agent builder, config helpers
+    - ``crypto`` — AES-256-GCM key generation and decryption
+    - ``onboard`` — QR-code scan-to-configure flow
+"""
+
+# -- Adapter (original qqbot.py) ------------------------------------------
+from .adapter import (  # noqa: F401
+    QQAdapter,
+    QQCloseError,
+    check_qq_requirements,
+    _coerce_list,
+    _ssrf_redirect_guard,
+)
+
+# -- Onboard (QR-code scan-to-configure) -----------------------------------
+from .onboard import (  # noqa: F401
+    BindStatus,
+    create_bind_task,
+    poll_bind_result,
+    build_connect_url,
+)
+from .crypto import decrypt_secret, generate_bind_key  # noqa: F401
+
+# -- Utils -----------------------------------------------------------------
+from .utils import build_user_agent, get_api_headers, coerce_list  # noqa: F401
+
+__all__ = [
+    # adapter
+    "QQAdapter",
+    "QQCloseError",
+    "check_qq_requirements",
+    "_coerce_list",
+    "_ssrf_redirect_guard",
+    # onboard
+    "BindStatus",
+    "create_bind_task",
+    "poll_bind_result",
+    "build_connect_url",
+    # crypto
+    "decrypt_secret",
+    "generate_bind_key",
+    # utils
+    "build_user_agent",
+    "get_api_headers",
+    "coerce_list",
+]

gateway/platforms/qqbot/constants.py

@@ -0,0 +1,74 @@
+"""QQBot package-level constants shared across adapter, onboard, and other modules."""
+
+from __future__ import annotations
+
+import os
+
+# ---------------------------------------------------------------------------
+# QQBot adapter version — bump on functional changes to the adapter package.
+# ---------------------------------------------------------------------------
+
+QQBOT_VERSION = "1.1.0"
+
+# ---------------------------------------------------------------------------
+# API endpoints
+# ---------------------------------------------------------------------------
+
+# The portal domain is configurable via QQ_API_HOST for corporate proxies
+# or test environments.  Default: q.qq.com (production).
+PORTAL_HOST = os.getenv("QQ_PORTAL_HOST", "q.qq.com")
+
+API_BASE = "https://api.sgroup.qq.com"
+TOKEN_URL = "https://bots.qq.com/app/getAppAccessToken"
+GATEWAY_URL_PATH = "/gateway"
+
+# QR-code onboard endpoints (on the portal host)
+ONBOARD_CREATE_PATH = "/lite/create_bind_task"
+ONBOARD_POLL_PATH = "/lite/poll_bind_result"
+QR_URL_TEMPLATE = (
+    "https://q.qq.com/qqbot/openclaw/connect.html"
+    "?task_id={task_id}&_wv=2&source=hermes"
+)
+
+# ---------------------------------------------------------------------------
+# Timeouts & retry
+# ---------------------------------------------------------------------------
+
+DEFAULT_API_TIMEOUT = 30.0
+FILE_UPLOAD_TIMEOUT = 120.0
+CONNECT_TIMEOUT_SECONDS = 20.0
+
+RECONNECT_BACKOFF = [2, 5, 10, 30, 60]
+MAX_RECONNECT_ATTEMPTS = 100
+RATE_LIMIT_DELAY = 60  # seconds
+QUICK_DISCONNECT_THRESHOLD = 5.0  # seconds
+MAX_QUICK_DISCONNECT_COUNT = 3
+
+ONBOARD_POLL_INTERVAL = 2.0  # seconds between poll_bind_result calls
+ONBOARD_API_TIMEOUT = 10.0
+
+# ---------------------------------------------------------------------------
+# Message limits
+# ---------------------------------------------------------------------------
+
+MAX_MESSAGE_LENGTH = 4000
+DEDUP_WINDOW_SECONDS = 300
+DEDUP_MAX_SIZE = 1000
+
+# ---------------------------------------------------------------------------
+# QQ Bot message types
+# ---------------------------------------------------------------------------
+
+MSG_TYPE_TEXT = 0
+MSG_TYPE_MARKDOWN = 2
+MSG_TYPE_MEDIA = 7
+MSG_TYPE_INPUT_NOTIFY = 6
+
+# ---------------------------------------------------------------------------
+# QQ Bot file media types
+# ---------------------------------------------------------------------------
+
+MEDIA_TYPE_IMAGE = 1
+MEDIA_TYPE_VIDEO = 2
+MEDIA_TYPE_VOICE = 3
+MEDIA_TYPE_FILE = 4

gateway/platforms/qqbot/crypto.py

@@ -0,0 +1,45 @@
+"""AES-256-GCM utilities for QQBot scan-to-configure credential decryption."""
+
+from __future__ import annotations
+
+import base64
+import os
+
+
+def generate_bind_key() -> str:
+    """Generate a 256-bit random AES key and return it as base64.
+
+    The key is passed to ``create_bind_task`` so the server can encrypt
+    the bot's *client_secret* before returning it.  Only this CLI holds
+    the key, ensuring the secret never travels in plaintext.
+    """
+    return base64.b64encode(os.urandom(32)).decode()
+
+
+def decrypt_secret(encrypted_base64: str, key_base64: str) -> str:
+    """Decrypt a base64-encoded AES-256-GCM ciphertext.
+
+    Ciphertext layout (after base64-decoding)::
+
+        IV (12 bytes) ‖ ciphertext (N bytes) ‖ AuthTag (16 bytes)
+
+    Args:
+        encrypted_base64: The ``bot_encrypt_secret`` value from
+            ``poll_bind_result``.
+        key_base64: The base64 AES key generated by
+            :func:`generate_bind_key`.
+
+    Returns:
+        The decrypted *client_secret* as a UTF-8 string.
+    """
+    from cryptography.hazmat.primitives.ciphers.aead import AESGCM
+
+    key = base64.b64decode(key_base64)
+    raw = base64.b64decode(encrypted_base64)
+
+    iv = raw[:12]
+    ciphertext_with_tag = raw[12:]  # AESGCM expects ciphertext + tag concatenated
+
+    aesgcm = AESGCM(key)
+    plaintext = aesgcm.decrypt(iv, ciphertext_with_tag, None)
+    return plaintext.decode("utf-8")

gateway/platforms/qqbot/onboard.py

@@ -0,0 +1,124 @@
+"""
+QQBot scan-to-configure (QR code onboard) module.
+
+Calls the ``q.qq.com`` ``create_bind_task`` / ``poll_bind_result`` APIs to
+generate a QR-code URL and poll for scan completion.  On success the caller
+receives the bot's *app_id*, *client_secret* (decrypted locally), and the
+scanner's *user_openid* — enough to fully configure the QQBot gateway.
+
+Reference: https://bot.q.qq.com/wiki/develop/api-v2/
+"""
+
+from __future__ import annotations
+
+import logging
+from enum import IntEnum
+from typing import Tuple
+from urllib.parse import quote
+
+from .constants import (
+    ONBOARD_API_TIMEOUT,
+    ONBOARD_CREATE_PATH,
+    ONBOARD_POLL_PATH,
+    PORTAL_HOST,
+    QR_URL_TEMPLATE,
+)
+from .crypto import generate_bind_key
+from .utils import get_api_headers
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Bind status
+# ---------------------------------------------------------------------------
+
+
+class BindStatus(IntEnum):
+    """Status codes returned by ``poll_bind_result``."""
+
+    NONE = 0
+    PENDING = 1
+    COMPLETED = 2
+    EXPIRED = 3
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+
+async def create_bind_task(
+    timeout: float = ONBOARD_API_TIMEOUT,
+) -> Tuple[str, str]:
+    """Create a bind task and return *(task_id, aes_key_base64)*.
+
+    The AES key is generated locally and sent to the server so it can
+    encrypt the bot credentials before returning them.
+
+    Raises:
+        RuntimeError: If the API returns a non-zero ``retcode``.
+    """
+    import httpx
+
+    url = f"https://{PORTAL_HOST}{ONBOARD_CREATE_PATH}"
+    key = generate_bind_key()
+
+    async with httpx.AsyncClient(timeout=timeout, follow_redirects=True) as client:
+        resp = await client.post(url, json={"key": key}, headers=get_api_headers())
+        resp.raise_for_status()
+        data = resp.json()
+
+    if data.get("retcode") != 0:
+        raise RuntimeError(data.get("msg", "create_bind_task failed"))
+
+    task_id = data.get("data", {}).get("task_id")
+    if not task_id:
+        raise RuntimeError("create_bind_task: missing task_id in response")
+
+    logger.debug("create_bind_task ok: task_id=%s", task_id)
+    return task_id, key
+
+
+async def poll_bind_result(
+    task_id: str,
+    timeout: float = ONBOARD_API_TIMEOUT,
+) -> Tuple[BindStatus, str, str, str]:
+    """Poll the bind result for *task_id*.
+
+    Returns:
+        A 4-tuple of ``(status, bot_appid, bot_encrypt_secret, user_openid)``.
+
+        * ``bot_encrypt_secret`` is AES-256-GCM encrypted — decrypt it with
+          :func:`~gateway.platforms.qqbot.crypto.decrypt_secret` using the
+          key from :func:`create_bind_task`.
+        * ``user_openid`` is the OpenID of the person who scanned the code
+          (available when ``status == COMPLETED``).
+
+    Raises:
+        RuntimeError: If the API returns a non-zero ``retcode``.
+    """
+    import httpx
+
+    url = f"https://{PORTAL_HOST}{ONBOARD_POLL_PATH}"
+
+    async with httpx.AsyncClient(timeout=timeout, follow_redirects=True) as client:
+        resp = await client.post(url, json={"task_id": task_id}, headers=get_api_headers())
+        resp.raise_for_status()
+        data = resp.json()
+
+    if data.get("retcode") != 0:
+        raise RuntimeError(data.get("msg", "poll_bind_result failed"))
+
+    d = data.get("data", {})
+    return (
+        BindStatus(d.get("status", 0)),
+        str(d.get("bot_appid", "")),
+        d.get("bot_encrypt_secret", ""),
+        d.get("user_openid", ""),
+    )
+
+
+def build_connect_url(task_id: str) -> str:
+    """Build the QR-code target URL for a given *task_id*."""
+    return QR_URL_TEMPLATE.format(task_id=quote(task_id))

gateway/platforms/qqbot/utils.py

@@ -0,0 +1,71 @@
+"""QQBot shared utilities — User-Agent, HTTP helpers, config coercion."""
+
+from __future__ import annotations
+
+import platform
+import sys
+from typing import Any, Dict, List
+
+from .constants import QQBOT_VERSION
+
+
+# ---------------------------------------------------------------------------
+# User-Agent
+# ---------------------------------------------------------------------------
+
+def _get_hermes_version() -> str:
+    """Return the hermes-agent package version, or 'dev' if unavailable."""
+    try:
+        from importlib.metadata import version
+        return version("hermes-agent")
+    except Exception:
+        return "dev"
+
+
+def build_user_agent() -> str:
+    """Build a descriptive User-Agent string.
+
+    Format::
+
+        QQBotAdapter/<qqbot_version> (Python/<py_version>; <os>; Hermes/<hermes_version>)
+
+    Example::
+
+        QQBotAdapter/1.0.0 (Python/3.11.15; darwin; Hermes/0.9.0)
+    """
+    py_version = f"{sys.version_info.major}.{sys.version_info.minor}.{sys.version_info.micro}"
+    os_name = platform.system().lower()
+    hermes_version = _get_hermes_version()
+    return f"QQBotAdapter/{QQBOT_VERSION} (Python/{py_version}; {os_name}; Hermes/{hermes_version})"
+
+
+def get_api_headers() -> Dict[str, str]:
+    """Return standard HTTP headers for QQBot API requests.
+
+    Includes ``Content-Type``, ``Accept``, and a dynamic ``User-Agent``.
+    ``q.qq.com`` requires ``Accept: application/json`` — without it,
+    the server returns a JavaScript anti-bot challenge page.
+    """
+    return {
+        "Content-Type": "application/json",
+        "Accept": "application/json",
+        "User-Agent": build_user_agent(),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Config helpers
+# ---------------------------------------------------------------------------
+
+def coerce_list(value: Any) -> List[str]:
+    """Coerce config values into a trimmed string list.
+
+    Accepts comma-separated strings, lists, tuples, sets, or single values.
+    """
+    if value is None:
+        return []
+    if isinstance(value, str):
+        return [item.strip() for item in value.split(",") if item.strip()]
+    if isinstance(value, (list, tuple, set)):
+        return [str(item).strip() for item in value if str(item).strip()]
+    return [str(value).strip()] if str(value).strip() else []

gateway/platforms/signal.py

@@ -160,6 +160,14 @@ class SignalAdapter(BasePlatformAdapter):
         self._sse_task: Optional[asyncio.Task] = None
         self._health_monitor_task: Optional[asyncio.Task] = None
         self._typing_tasks: Dict[str, asyncio.Task] = {}
+        # Per-chat typing-indicator backoff. When signal-cli reports
+        # NETWORK_FAILURE (recipient offline / unroutable), base.py's
+        # _keep_typing refresh loop would otherwise hammer sendTyping every
+        # ~2s indefinitely, producing WARNING-level log spam and pointless
+        # RPC traffic. We track consecutive failures per chat and skip the
+        # RPC during a cooldown window instead.
+        self._typing_failures: Dict[str, int] = {}
+        self._typing_skip_until: Dict[str, float] = {}
         self._running = False
         self._last_sse_activity = 0.0
         self._sse_response: Optional[httpx.Response] = None
@@ -548,8 +556,22 @@ class SignalAdapter(BasePlatformAdapter):
     # JSON-RPC Communication
     # ------------------------------------------------------------------
 
-    async def _rpc(self, method: str, params: dict, rpc_id: str = None) -> Any:
-        """Send a JSON-RPC 2.0 request to signal-cli daemon."""
+    async def _rpc(
+        self,
+        method: str,
+        params: dict,
+        rpc_id: str = None,
+        *,
+        log_failures: bool = True,
+    ) -> Any:
+        """Send a JSON-RPC 2.0 request to signal-cli daemon.
+
+        When ``log_failures=False``, error and exception paths log at DEBUG
+        instead of WARNING — used by the typing-indicator path to silence
+        repeated NETWORK_FAILURE spam for unreachable recipients while
+        still preserving visibility for the first occurrence and for
+        unrelated RPCs.
+        """
         if not self.client:
             logger.warning("Signal: RPC called but client not connected")
             return None
@@ -574,13 +596,19 @@ class SignalAdapter(BasePlatformAdapter):
             data = resp.json()
 
             if "error" in data:
-                logger.warning("Signal RPC error (%s): %s", method, data["error"])
+                if log_failures:
+                    logger.warning("Signal RPC error (%s): %s", method, data["error"])
+                else:
+                    logger.debug("Signal RPC error (%s): %s", method, data["error"])
                 return None
 
             return data.get("result")
 
         except Exception as e:
-            logger.warning("Signal RPC %s failed: %s", method, e)
+            if log_failures:
+                logger.warning("Signal RPC %s failed: %s", method, e)
+            else:
+                logger.debug("Signal RPC %s failed: %s", method, e)
             return None
 
     # ------------------------------------------------------------------
@@ -627,7 +655,28 @@ class SignalAdapter(BasePlatformAdapter):
                 self._recent_sent_timestamps.pop()
 
     async def send_typing(self, chat_id: str, metadata=None) -> None:
-        """Send a typing indicator."""
+        """Send a typing indicator.
+
+        base.py's ``_keep_typing`` refresh loop calls this every ~2s while
+        the agent is processing. If signal-cli returns NETWORK_FAILURE for
+        this recipient (offline, unroutable, group membership lost, etc.)
+        the unmitigated behaviour is: a WARNING log every 2 seconds for as
+        long as the agent keeps running. Instead we:
+
+        - silence the WARNING after the first consecutive failure (subsequent
+          attempts log at DEBUG) so transport issues are still visible once
+          but don't flood the log,
+        - skip the RPC entirely during an exponential cooldown window once
+          three consecutive failures have happened, so we stop hammering
+          signal-cli with requests it can't deliver.
+
+        A successful sendTyping clears the counters.
+        """
+        now = time.monotonic()
+        skip_until = self._typing_skip_until.get(chat_id, 0.0)
+        if now < skip_until:
+            return
+
         params: Dict[str, Any] = {
             "account": self.account,
         }
@@ -637,7 +686,26 @@ class SignalAdapter(BasePlatformAdapter):
         else:
             params["recipient"] = [chat_id]
 
-        await self._rpc("sendTyping", params, rpc_id="typing")
+        fails = self._typing_failures.get(chat_id, 0)
+        result = await self._rpc(
+            "sendTyping",
+            params,
+            rpc_id="typing",
+            log_failures=(fails == 0),
+        )
+
+        if result is None:
+            fails += 1
+            self._typing_failures[chat_id] = fails
+            # After 3 consecutive failures, back off exponentially (16s,
+            # 32s, 60s cap) to stop spamming signal-cli for a recipient
+            # that clearly isn't reachable right now.
+            if fails >= 3:
+                backoff = min(60.0, 16.0 * (2 ** (fails - 3)))
+                self._typing_skip_until[chat_id] = now + backoff
+        else:
+            self._typing_failures.pop(chat_id, None)
+            self._typing_skip_until.pop(chat_id, None)
 
     async def send_image(
         self,
@@ -789,6 +857,10 @@ class SignalAdapter(BasePlatformAdapter):
                 await task
             except asyncio.CancelledError:
                 pass
+        # Reset per-chat typing backoff state so the next agent turn starts
+        # fresh rather than inheriting a cooldown from a prior conversation.
+        self._typing_failures.pop(chat_id, None)
+        self._typing_skip_until.pop(chat_id, None)
 
     async def stop_typing(self, chat_id: str) -> None:
         """Public interface for stopping typing — called by base adapter's

gateway/platforms/telegram.py

@@ -118,6 +118,84 @@ def _strip_mdv2(text: str) -> str:
     return cleaned
 
 
+# ---------------------------------------------------------------------------
+# Markdown table → code block conversion
+# ---------------------------------------------------------------------------
+# Telegram's MarkdownV2 has no table syntax — '|' is just an escaped literal,
+# so pipe tables render as noisy backslash-pipe text with no alignment.
+# Wrapping the table in a fenced code block makes Telegram render it as
+# monospace preformatted text with columns intact.
+
+# Matches a GFM table delimiter row: optional outer pipes, cells containing
+# only dashes (with optional leading/trailing colons for alignment) separated
+# by '|'.  Requires at least one internal '|' so lone '---' horizontal rules
+# are NOT matched.
+_TABLE_SEPARATOR_RE = re.compile(
+    r'^\s*\|?\s*:?-+:?\s*(?:\|\s*:?-+:?\s*){1,}\|?\s*$'
+)
+
+
+def _is_table_row(line: str) -> bool:
+    """Return True if *line* could plausibly be a table data row."""
+    stripped = line.strip()
+    return bool(stripped) and '|' in stripped
+
+
+def _wrap_markdown_tables(text: str) -> str:
+    """Wrap GFM-style pipe tables in ``` fences so Telegram renders them.
+
+    Detected by a row containing '|' immediately followed by a delimiter
+    row matching :data:`_TABLE_SEPARATOR_RE`.  Subsequent pipe-containing
+    non-blank lines are consumed as the table body and included in the
+    wrapped block.  Tables inside existing fenced code blocks are left
+    alone.
+    """
+    if '|' not in text or '-' not in text:
+        return text
+
+    lines = text.split('\n')
+    out: list[str] = []
+    in_fence = False
+    i = 0
+    while i < len(lines):
+        line = lines[i]
+        stripped = line.lstrip()
+
+        # Track existing fenced code blocks — never touch content inside.
+        if stripped.startswith('```'):
+            in_fence = not in_fence
+            out.append(line)
+            i += 1
+            continue
+        if in_fence:
+            out.append(line)
+            i += 1
+            continue
+
+        # Look for a header row (contains '|') immediately followed by a
+        # delimiter row.
+        if (
+            '|' in line
+            and i + 1 < len(lines)
+            and _TABLE_SEPARATOR_RE.match(lines[i + 1])
+        ):
+            table_block = [line, lines[i + 1]]
+            j = i + 2
+            while j < len(lines) and _is_table_row(lines[j]):
+                table_block.append(lines[j])
+                j += 1
+            out.append('```')
+            out.extend(table_block)
+            out.append('```')
+            i = j
+            continue
+
+        out.append(line)
+        i += 1
+
+    return '\n'.join(out)
+
+
 class TelegramAdapter(BasePlatformAdapter):
     """
     Telegram bot adapter.
@@ -1916,6 +1994,12 @@ class TelegramAdapter(BasePlatformAdapter):
 
         text = content
 
+        # 0) Pre-wrap GFM-style pipe tables in ``` fences.  Telegram can't
+        #    render tables natively, but fenced code blocks render as
+        #    monospace preformatted text with columns intact.  The wrapped
+        #    tables then flow through step (1) below as protected regions.
+        text = _wrap_markdown_tables(text)
+
         # 1) Protect fenced code blocks (``` ... ```)
         #    Per MarkdownV2 spec, \ and ` inside pre/code must be escaped.
         def _protect_fenced(m):
@@ -2242,7 +2326,7 @@ class TelegramAdapter(BasePlatformAdapter):
         if not self._should_process_message(update.message):
             return
 
-        event = self._build_message_event(update.message, MessageType.TEXT)
+        event = self._build_message_event(update.message, MessageType.TEXT, update_id=update.update_id)
         event.text = self._clean_bot_trigger_text(event.text)
         self._enqueue_text_event(event)
     
@@ -2253,7 +2337,7 @@ class TelegramAdapter(BasePlatformAdapter):
         if not self._should_process_message(update.message, is_command=True):
             return
         
-        event = self._build_message_event(update.message, MessageType.COMMAND)
+        event = self._build_message_event(update.message, MessageType.COMMAND, update_id=update.update_id)
         await self.handle_message(event)
     
     async def _handle_location_message(self, update: Update, context: ContextTypes.DEFAULT_TYPE) -> None:
@@ -2289,7 +2373,7 @@ class TelegramAdapter(BasePlatformAdapter):
         parts.append(f"Map: https://www.google.com/maps/search/?api=1&query={lat},{lon}")
         parts.append("Ask what they'd like to find nearby (restaurants, cafes, etc.) and any preferences.")
 
-        event = self._build_message_event(msg, MessageType.LOCATION)
+        event = self._build_message_event(msg, MessageType.LOCATION, update_id=update.update_id)
         event.text = "\n".join(parts)
         await self.handle_message(event)
 
@@ -2440,7 +2524,7 @@ class TelegramAdapter(BasePlatformAdapter):
         else:
             msg_type = MessageType.DOCUMENT
         
-        event = self._build_message_event(msg, msg_type)
+        event = self._build_message_event(msg, msg_type, update_id=update.update_id)
         
         # Add caption as text
         if msg.caption:
@@ -2779,8 +2863,19 @@ class TelegramAdapter(BasePlatformAdapter):
                 self.name, cache_key, thread_id,
             )
 
-    def _build_message_event(self, message: Message, msg_type: MessageType) -> MessageEvent:
-        """Build a MessageEvent from a Telegram message."""
+    def _build_message_event(
+        self,
+        message: Message,
+        msg_type: MessageType,
+        update_id: Optional[int] = None,
+    ) -> MessageEvent:
+        """Build a MessageEvent from a Telegram message.
+
+        ``update_id`` is the ``Update.update_id`` from PTB; passing it through
+        lets ``/restart`` record the triggering offset so the new gateway
+        process can advance past it (prevents ``/restart`` being re-delivered
+        when PTB's graceful-shutdown ACK fails).
+        """
         chat = message.chat
         user = message.from_user
         
@@ -2859,6 +2954,7 @@ class TelegramAdapter(BasePlatformAdapter):
             source=source,
             raw_message=message,
             message_id=str(message.message_id),
+            platform_update_id=update_id,
             reply_to_message_id=reply_to_id,
             reply_to_text=reply_to_text,
             auto_skill=topic_skill,

gateway/platforms/wecom.py

@@ -180,6 +180,8 @@ class WeComAdapter(BasePlatformAdapter):
         self._text_batch_split_delay_seconds = float(os.getenv("HERMES_WECOM_TEXT_BATCH_SPLIT_DELAY_SECONDS", "2.0"))
         self._pending_text_batches: Dict[str, MessageEvent] = {}
         self._pending_text_batch_tasks: Dict[str, asyncio.Task] = {}
+        self._device_id = uuid.uuid4().hex
+        self._last_chat_req_ids: Dict[str, str] = {}
 
     # ------------------------------------------------------------------
     # Connection lifecycle
@@ -277,7 +279,11 @@ class WeComAdapter(BasePlatformAdapter):
             {
                 "cmd": APP_CMD_SUBSCRIBE,
                 "headers": {"req_id": req_id},
-                "body": {"bot_id": self._bot_id, "secret": self._secret},
+                "body": {
+                    "bot_id": self._bot_id,
+                    "secret": self._secret,
+                    "device_id": self._device_id,
+                },
             }
         )
 
@@ -496,6 +502,11 @@ class WeComAdapter(BasePlatformAdapter):
             logger.debug("[%s] DM sender %s blocked by policy", self.name, sender_id)
             return
 
+        # Cache the inbound req_id after policy checks so proactive sends to
+        # this chat can fall back to APP_CMD_RESPONSE (required for groups —
+        # WeCom AI Bots cannot initiate APP_CMD_SEND in group chats).
+        self._remember_chat_req_id(chat_id, self._payload_req_id(payload))
+
         text, reply_text = self._extract_text(body)
         media_urls, media_types = await self._extract_media(body)
         message_type = self._derive_message_type(body, text, media_types)
@@ -847,6 +858,23 @@ class WeComAdapter(BasePlatformAdapter):
         while len(self._reply_req_ids) > DEDUP_MAX_SIZE:
             self._reply_req_ids.pop(next(iter(self._reply_req_ids)))
 
+    def _remember_chat_req_id(self, chat_id: str, req_id: str) -> None:
+        """Cache the most recent inbound req_id per chat.
+
+        Used as a fallback reply target when we need to send into a group
+        without an explicit ``reply_to`` — WeCom AI Bots are blocked from
+        APP_CMD_SEND in groups and must use APP_CMD_RESPONSE bound to some
+        prior req_id. Bounded like _reply_req_ids so long-running gateways
+        don't leak memory across many chats.
+        """
+        normalized_chat_id = str(chat_id or "").strip()
+        normalized_req_id = str(req_id or "").strip()
+        if not normalized_chat_id or not normalized_req_id:
+            return
+        self._last_chat_req_ids[normalized_chat_id] = normalized_req_id
+        while len(self._last_chat_req_ids) > DEDUP_MAX_SIZE:
+            self._last_chat_req_ids.pop(next(iter(self._last_chat_req_ids)))
+
     def _reply_req_id_for_message(self, reply_to: Optional[str]) -> Optional[str]:
         normalized = str(reply_to or "").strip()
         if not normalized or normalized.startswith("quote:"):
@@ -1163,19 +1191,15 @@ class WeComAdapter(BasePlatformAdapter):
         self._raise_for_wecom_error(response, "send media message")
         return response
 
-    async def _send_reply_stream(self, reply_req_id: str, content: str) -> Dict[str, Any]:
+    async def _send_reply_markdown(self, reply_req_id: str, content: str) -> Dict[str, Any]:
         response = await self._send_reply_request(
             reply_req_id,
             {
-                "msgtype": "stream",
-                "stream": {
-                    "id": self._new_req_id("stream"),
-                    "finish": True,
-                    "content": content[:self.MAX_MESSAGE_LENGTH],
-                },
+                "msgtype": "markdown",
+                "markdown": {"content": content[:self.MAX_MESSAGE_LENGTH]},
             },
         )
-        self._raise_for_wecom_error(response, "send reply stream")
+        self._raise_for_wecom_error(response, "send reply markdown")
         return response
 
     async def _send_reply_media_message(
@@ -1235,6 +1259,9 @@ class WeComAdapter(BasePlatformAdapter):
             return SendResult(success=False, error=prepared["reject_reason"])
 
         reply_req_id = self._reply_req_id_for_message(reply_to)
+        if not reply_req_id and chat_id in self._last_chat_req_ids:
+            reply_req_id = self._last_chat_req_ids[chat_id]
+
         try:
             upload_result = await self._upload_media_bytes(
                 prepared["data"],
@@ -1302,8 +1329,12 @@ class WeComAdapter(BasePlatformAdapter):
 
         try:
             reply_req_id = self._reply_req_id_for_message(reply_to)
+
+            if not reply_req_id and chat_id in self._last_chat_req_ids:
+                reply_req_id = self._last_chat_req_ids[chat_id]
+
             if reply_req_id:
-                response = await self._send_reply_stream(reply_req_id, content)
+                response = await self._send_reply_markdown(reply_req_id, content)
             else:
                 response = await self._send_request(
                     APP_CMD_SEND,

gateway/run.py

@@ -2178,6 +2178,30 @@ class GatewayRunner:
                         )
                 except Exception as _e:
                     logger.debug("Idle agent sweep failed: %s", _e)
+
+                # Periodically prune stale SessionStore entries.  The
+                # in-memory dict (and sessions.json) would otherwise grow
+                # unbounded in gateways serving many rotating chats /
+                # threads / users over long time windows.  Pruning is
+                # invisible to users — a resumed session just gets a
+                # fresh session_id, exactly as if the reset policy fired.
+                _last_prune_ts = getattr(self, "_last_session_store_prune_ts", 0.0)
+                _prune_interval = 3600.0  # once per hour
+                if time.time() - _last_prune_ts > _prune_interval:
+                    try:
+                        _max_age = int(
+                            getattr(self.config, "session_store_max_age_days", 0) or 0
+                        )
+                        if _max_age > 0:
+                            _pruned = self.session_store.prune_old_entries(_max_age)
+                            if _pruned:
+                                logger.info(
+                                    "SessionStore prune: dropped %d stale entries",
+                                    _pruned,
+                                )
+                    except Exception as _e:
+                        logger.debug("SessionStore prune failed: %s", _e)
+                    self._last_session_store_prune_ts = time.time()
             except Exception as e:
                 logger.debug("Session expiry watcher error: %s", e)
             # Sleep in small increments so we can stop quickly
@@ -2384,6 +2408,7 @@ class GatewayRunner:
 
             self.adapters.clear()
             self._running_agents.clear()
+            self._running_agents_ts.clear()
             self._pending_messages.clear()
             self._pending_approvals.clear()
             if hasattr(self, '_busy_ack_ts'):
@@ -2408,6 +2433,20 @@ class GatewayRunner:
             except Exception:
                 pass
 
+            # Close SQLite session DBs so the WAL write lock is released.
+            # Without this, --replace and similar restart flows leave the
+            # old gateway's connection holding the WAL lock until Python
+            # actually exits — causing 'database is locked' errors when
+            # the new gateway tries to open the same file.
+            for _db_holder in (self, getattr(self, "session_store", None)):
+                _db = getattr(_db_holder, "_db", None) if _db_holder else None
+                if _db is None or not hasattr(_db, "close"):
+                    continue
+                try:
+                    _db.close()
+                except Exception as _e:
+                    logger.debug("SessionDB close error: %s", _e)
+
             from gateway.status import remove_pid_file
             remove_pid_file()
 
@@ -2906,16 +2945,17 @@ class GatewayRunner:
                     _quick_key[:30], _stale_age, _stale_idle,
                     _raw_stale_timeout, _stale_detail,
                 )
-                del self._running_agents[_quick_key]
-                self._running_agents_ts.pop(_quick_key, None)
-                self._busy_ack_ts.pop(_quick_key, None)
+                self._release_running_agent_state(_quick_key)
 
         if _quick_key in self._running_agents:
             if event.get_command() == "status":
                 return await self._handle_status_command(event)
 
             # Resolve the command once for all early-intercept checks below.
-            from hermes_cli.commands import resolve_command as _resolve_cmd_inner
+            from hermes_cli.commands import (
+                resolve_command as _resolve_cmd_inner,
+                should_bypass_active_session as _should_bypass_active_inner,
+            )
             _evt_cmd = event.get_command()
             _cmd_def_inner = _resolve_cmd_inner(_evt_cmd) if _evt_cmd else None
 
@@ -2936,8 +2976,7 @@ class GatewayRunner:
                 if adapter and hasattr(adapter, 'get_pending_message'):
                     adapter.get_pending_message(_quick_key)  # consume and discard
                 self._pending_messages.pop(_quick_key, None)
-                if _quick_key in self._running_agents:
-                    del self._running_agents[_quick_key]
+                self._release_running_agent_state(_quick_key)
                 logger.info("STOP for session %s — agent interrupted, session lock released", _quick_key[:20])
                 return "⚡ Stopped. You can continue this session."
 
@@ -2959,8 +2998,7 @@ class GatewayRunner:
                 self._pending_messages.pop(_quick_key, None)
                 # Clean up the running agent entry so the reset handler
                 # doesn't think an agent is still active.
-                if _quick_key in self._running_agents:
-                    del self._running_agents[_quick_key]
+                self._release_running_agent_state(_quick_key)
                 return await self._handle_reset_command(event)
 
             # /queue <prompt> — queue without interrupting
@@ -2981,6 +3019,54 @@ class GatewayRunner:
                     adapter._pending_messages[_quick_key] = queued_event
                 return "Queued for the next turn."
 
+            # /steer <prompt> — inject mid-run after the next tool call.
+            # Unlike /queue (turn boundary), /steer lands BETWEEN tool-call
+            # iterations inside the same agent run, by appending to the
+            # last tool result's content. No interrupt, no new user turn,
+            # no role-alternation violation.
+            if _cmd_def_inner and _cmd_def_inner.name == "steer":
+                steer_text = event.get_command_args().strip()
+                if not steer_text:
+                    return "Usage: /steer <prompt>"
+                running_agent = self._running_agents.get(_quick_key)
+                if running_agent is _AGENT_PENDING_SENTINEL:
+                    # Agent hasn't started yet — queue as turn-boundary fallback.
+                    adapter = self.adapters.get(source.platform)
+                    if adapter:
+                        from gateway.platforms.base import MessageEvent as _ME, MessageType as _MT
+                        queued_event = _ME(
+                            text=steer_text,
+                            message_type=_MT.TEXT,
+                            source=event.source,
+                            message_id=event.message_id,
+                            channel_prompt=event.channel_prompt,
+                        )
+                        adapter._pending_messages[_quick_key] = queued_event
+                    return "Agent still starting — /steer queued for the next turn."
+                if running_agent and hasattr(running_agent, "steer"):
+                    try:
+                        accepted = running_agent.steer(steer_text)
+                    except Exception as exc:
+                        logger.warning("Steer failed for session %s: %s", _quick_key[:20], exc)
+                        return f"⚠️ Steer failed: {exc}"
+                    if accepted:
+                        preview = steer_text[:60] + ("..." if len(steer_text) > 60 else "")
+                        return f"⏩ Steer queued — arrives after the next tool call: '{preview}'"
+                    return "Steer rejected (empty payload)."
+                # Running agent is missing or lacks steer() — fall back to queue.
+                adapter = self.adapters.get(source.platform)
+                if adapter:
+                    from gateway.platforms.base import MessageEvent as _ME, MessageType as _MT
+                    queued_event = _ME(
+                        text=steer_text,
+                        message_type=_MT.TEXT,
+                        source=event.source,
+                        message_id=event.message_id,
+                        channel_prompt=event.channel_prompt,
+                    )
+                    adapter._pending_messages[_quick_key] = queued_event
+                return "No active agent — /steer queued for the next turn."
+
             # /model must not be used while the agent is running.
             if _cmd_def_inner and _cmd_def_inner.name == "model":
                 return "Agent is running — wait or /stop first, then switch models."
@@ -2994,11 +3080,29 @@ class GatewayRunner:
                     return await self._handle_approve_command(event)
                 return await self._handle_deny_command(event)
 
+            # /agents (/tasks alias) should be query-only and never interrupt.
+            if _cmd_def_inner and _cmd_def_inner.name == "agents":
+                return await self._handle_agents_command(event)
+
             # /background must bypass the running-agent guard — it starts a
             # parallel task and must never interrupt the active conversation.
             if _cmd_def_inner and _cmd_def_inner.name == "background":
                 return await self._handle_background_command(event)
 
+            # Gateway-handled info/control commands must never fall through to
+            # the interrupt path. If they are queued as pending text, the
+            # slash-command safety net discards them before the user sees any
+            # response.
+            if _cmd_def_inner and _should_bypass_active_inner(_cmd_def_inner.name):
+                if _cmd_def_inner.name == "help":
+                    return await self._handle_help_command(event)
+                if _cmd_def_inner.name == "commands":
+                    return await self._handle_commands_command(event)
+                if _cmd_def_inner.name == "profile":
+                    return await self._handle_profile_command(event)
+                if _cmd_def_inner.name == "update":
+                    return await self._handle_update_command(event)
+
             if event.message_type == MessageType.PHOTO:
                 logger.debug("PRIORITY photo follow-up for session %s — queueing without interrupt", _quick_key[:20])
                 adapter = self.adapters.get(source.platform)
@@ -3037,8 +3141,7 @@ class GatewayRunner:
                 # Agent is being set up but not ready yet.
                 if event.get_command() == "stop":
                     # Force-clean the sentinel so the session is unlocked.
-                    if _quick_key in self._running_agents:
-                        del self._running_agents[_quick_key]
+                    self._release_running_agent_state(_quick_key)
                     logger.info("HARD STOP (pending) for session %s — sentinel cleared", _quick_key[:20])
                     return "⚡ Force-stopped. The agent was still starting — session unlocked."
                 # Queue the message so it will be picked up after the
@@ -3102,6 +3205,9 @@ class GatewayRunner:
         if canonical == "status":
             return await self._handle_status_command(event)
 
+        if canonical == "agents":
+            return await self._handle_agents_command(event)
+
         if canonical == "restart":
             return await self._handle_restart_command(event)
         
@@ -3202,6 +3308,21 @@ class GatewayRunner:
         if canonical == "btw":
             return await self._handle_btw_command(event)
 
+        if canonical == "steer":
+            # No active agent — /steer has no tool call to inject into.
+            # Strip the prefix so downstream treats it as a normal user
+            # message. If the payload is empty, surface the usage hint.
+            steer_payload = event.get_command_args().strip()
+            if not steer_payload:
+                return "Usage: /steer <prompt>  (no agent is running; sending as a normal message)"
+            try:
+                event.text = steer_payload
+            except Exception:
+                pass
+            # Do NOT return — fall through to _handle_message_with_agent
+            # at the end of this function so the rewritten text is sent
+            # to the agent as a regular user turn.
+
         if canonical == "voice":
             return await self._handle_voice_command(event)
 
@@ -3354,8 +3475,13 @@ class GatewayRunner:
             # (exception, command fallthrough, etc.) the sentinel must
             # not linger or the session would be permanently locked out.
             if self._running_agents.get(_quick_key) is _AGENT_PENDING_SENTINEL:
-                del self._running_agents[_quick_key]
-            self._running_agents_ts.pop(_quick_key, None)
+                self._release_running_agent_state(_quick_key)
+            else:
+                # Agent path already cleaned _running_agents; make sure
+                # the paired metadata dicts are gone too.
+                self._running_agents_ts.pop(_quick_key, None)
+                if hasattr(self, "_busy_ack_ts"):
+                    self._busy_ack_ts.pop(_quick_key, None)
 
     async def _prepare_inbound_message_text(
         self,
@@ -4552,6 +4678,96 @@ class GatewayRunner:
         ])
 
         return "\n".join(lines)
+
+    async def _handle_agents_command(self, event: MessageEvent) -> str:
+        """Handle /agents command - list active agents and running tasks."""
+        from tools.process_registry import format_uptime_short, process_registry
+
+        now = time.time()
+        current_session_key = self._session_key_for_source(event.source)
+
+        running_agents: dict = getattr(self, "_running_agents", {}) or {}
+        running_started: dict = getattr(self, "_running_agents_ts", {}) or {}
+
+        agent_rows: list[dict] = []
+        for session_key, agent in running_agents.items():
+            started = float(running_started.get(session_key, now))
+            elapsed = max(0, int(now - started))
+            is_pending = agent is _AGENT_PENDING_SENTINEL
+            agent_rows.append(
+                {
+                    "session_key": session_key,
+                    "elapsed": elapsed,
+                    "state": "starting" if is_pending else "running",
+                    "session_id": "" if is_pending else str(getattr(agent, "session_id", "") or ""),
+                    "model": "" if is_pending else str(getattr(agent, "model", "") or ""),
+                }
+            )
+
+        agent_rows.sort(key=lambda row: row["elapsed"], reverse=True)
+
+        running_processes: list[dict] = []
+        try:
+            running_processes = [
+                p for p in process_registry.list_sessions()
+                if p.get("status") == "running"
+            ]
+        except Exception:
+            running_processes = []
+
+        background_tasks = [
+            t for t in (getattr(self, "_background_tasks", set()) or set())
+            if hasattr(t, "done") and not t.done()
+        ]
+
+        lines = [
+            "🤖 **Active Agents & Tasks**",
+            "",
+            f"**Active agents:** {len(agent_rows)}",
+        ]
+
+        if agent_rows:
+            for idx, row in enumerate(agent_rows[:12], 1):
+                current = " · this chat" if row["session_key"] == current_session_key else ""
+                sid = f" · `{row['session_id']}`" if row["session_id"] else ""
+                model = f" · `{row['model']}`" if row["model"] else ""
+                lines.append(
+                    f"{idx}. `{row['session_key']}` · {row['state']} · "
+                    f"{format_uptime_short(row['elapsed'])}{sid}{model}{current}"
+                )
+            if len(agent_rows) > 12:
+                lines.append(f"... and {len(agent_rows) - 12} more")
+
+        lines.extend(
+            [
+                "",
+                f"**Running background processes:** {len(running_processes)}",
+            ]
+        )
+        if running_processes:
+            for proc in running_processes[:12]:
+                cmd = " ".join(str(proc.get("command", "")).split())
+                if len(cmd) > 90:
+                    cmd = cmd[:87] + "..."
+                lines.append(
+                    f"- `{proc.get('session_id', '?')}` · "
+                    f"{format_uptime_short(int(proc.get('uptime_seconds', 0)))} · `{cmd}`"
+                )
+            if len(running_processes) > 12:
+                lines.append(f"... and {len(running_processes) - 12} more")
+
+        lines.extend(
+            [
+                "",
+                f"**Gateway async jobs:** {len(background_tasks)}",
+            ]
+        )
+
+        if not agent_rows and not running_processes and not background_tasks:
+            lines.append("")
+            lines.append("No active agents or running tasks.")
+
+        return "\n".join(lines)
     
     async def _handle_stop_command(self, event: MessageEvent) -> str:
         """Handle /stop command - interrupt a running agent.
@@ -4571,22 +4787,40 @@ class GatewayRunner:
         agent = self._running_agents.get(session_key)
         if agent is _AGENT_PENDING_SENTINEL:
             # Force-clean the sentinel so the session is unlocked.
-            if session_key in self._running_agents:
-                del self._running_agents[session_key]
+            self._release_running_agent_state(session_key)
             logger.info("STOP (pending) for session %s — sentinel cleared", session_key[:20])
             return "⚡ Stopped. The agent hadn't started yet — you can continue this session."
         if agent:
             agent.interrupt("Stop requested")
             # Force-clean the session lock so a truly hung agent doesn't
             # keep it locked forever.
-            if session_key in self._running_agents:
-                del self._running_agents[session_key]
+            self._release_running_agent_state(session_key)
             return "⚡ Stopped. You can continue this session."
         else:
             return "No active task to stop."
 
     async def _handle_restart_command(self, event: MessageEvent) -> str:
         """Handle /restart command - drain active work, then restart the gateway."""
+        # Defensive idempotency check: if the previous gateway process
+        # recorded this same /restart (same platform + update_id) and the new
+        # process is seeing it *again*, this is a re-delivery caused by PTB's
+        # graceful-shutdown `get_updates` ACK failing on the way out ("Error
+        # while calling `get_updates` one more time to mark all fetched
+        # updates. Suppressing error to ensure graceful shutdown. When
+        # polling for updates is restarted, updates may be received twice."
+        # in gateway.log).  Ignoring the stale redelivery prevents a
+        # self-perpetuating restart loop where every fresh gateway
+        # re-processes the same /restart command and immediately restarts
+        # again.
+        if self._is_stale_restart_redelivery(event):
+            logger.info(
+                "Ignoring redelivered /restart (platform=%s, update_id=%s) — "
+                "already processed by a previous gateway instance.",
+                event.source.platform.value if event.source and event.source.platform else "?",
+                event.platform_update_id,
+            )
+            return ""
+
         if self._restart_requested or self._draining:
             count = self._running_agent_count()
             if count:
@@ -4609,6 +4843,26 @@ class GatewayRunner:
         except Exception as e:
             logger.debug("Failed to write restart notify file: %s", e)
 
+        # Record the triggering platform + update_id in a dedicated dedup
+        # marker.  Unlike .restart_notify.json (which gets unlinked once the
+        # new gateway sends the "gateway restarted" notification), this
+        # marker persists so the new gateway can still detect a delayed
+        # /restart redelivery from Telegram.  Overwritten on every /restart.
+        try:
+            import json as _json
+            import time as _time
+            dedup_data = {
+                "platform": event.source.platform.value if event.source.platform else None,
+                "requested_at": _time.time(),
+            }
+            if event.platform_update_id is not None:
+                dedup_data["update_id"] = event.platform_update_id
+            (_hermes_home / ".restart_last_processed.json").write_text(
+                _json.dumps(dedup_data)
+            )
+        except Exception as e:
+            logger.debug("Failed to write restart dedup marker: %s", e)
+
         active_agents = self._running_agent_count()
         # When running under a service manager (systemd/launchd), use the
         # service restart path: exit with code 75 so the service manager
@@ -4624,6 +4878,58 @@ class GatewayRunner:
             return f"⏳ Draining {active_agents} active agent(s) before restart..."
         return "♻ Restarting gateway. If you aren't notified within 60 seconds, restart from the console with `hermes gateway restart`."
 
+    def _is_stale_restart_redelivery(self, event: MessageEvent) -> bool:
+        """Return True if this /restart is a Telegram re-delivery we already handled.
+
+        The previous gateway wrote ``.restart_last_processed.json`` with the
+        triggering platform + update_id when it processed the /restart.  If
+        we now see a /restart on the same platform with an update_id <= that
+        recorded value AND the marker is recent (< 5 minutes), it's a
+        redelivery and should be ignored.
+
+        Only applies to Telegram today (the only platform that exposes a
+        numeric cross-session update ordering); other platforms return False.
+        """
+        if event is None or event.source is None:
+            return False
+        if event.platform_update_id is None:
+            return False
+        if event.source.platform is None:
+            return False
+        # Only Telegram populates platform_update_id currently; be explicit
+        # so future platforms aren't accidentally gated by this check.
+        try:
+            platform_value = event.source.platform.value
+        except Exception:
+            return False
+        if platform_value != "telegram":
+            return False
+
+        try:
+            import json as _json
+            import time as _time
+            marker_path = _hermes_home / ".restart_last_processed.json"
+            if not marker_path.exists():
+                return False
+            data = _json.loads(marker_path.read_text())
+        except Exception:
+            return False
+
+        if data.get("platform") != platform_value:
+            return False
+        recorded_uid = data.get("update_id")
+        if not isinstance(recorded_uid, int):
+            return False
+        # Staleness guard: ignore markers older than 5 minutes.  A legitimately
+        # old marker (e.g. crash recovery where notify never fired) should not
+        # swallow a fresh /restart from the user.
+        requested_at = data.get("requested_at")
+        if isinstance(requested_at, (int, float)):
+            if _time.time() - requested_at > 300:
+                return False
+        return event.platform_update_id <= recorded_uid
+
+
     async def _handle_help_command(self, event: MessageEvent) -> str:
         """Handle /help command - list available commands."""
         from hermes_cli.commands import gateway_help_lines
@@ -5369,8 +5675,7 @@ class GatewayRunner:
             if "pynacl" in err_lower or "nacl" in err_lower or "davey" in err_lower:
                 return (
                     "Voice dependencies are missing (PyNaCl / davey). "
-                    "Install or reinstall Hermes with the messaging extra, e.g. "
-                    "`pip install hermes-agent[messaging]`."
+                    f"Install with: `{sys.executable} -m pip install PyNaCl`"
                 )
             return f"Failed to join voice channel: {e}"
 
@@ -6496,8 +6801,7 @@ class GatewayRunner:
             logger.debug("Memory flush on resume failed: %s", e)
 
         # Clear any running agent for this session key
-        if session_key in self._running_agents:
-            del self._running_agents[session_key]
+        self._release_running_agent_state(session_key)
 
         # Switch the session entry to point at the old session
         new_entry = self.session_store.switch_session(session_key, target_id)
@@ -7913,6 +8217,30 @@ class GatewayRunner:
         override = self._session_model_overrides.get(session_key)
         return override is not None and override.get("model") == agent_model
 
+    def _release_running_agent_state(self, session_key: str) -> None:
+        """Pop ALL per-running-agent state entries for ``session_key``.
+
+        Replaces ad-hoc ``del self._running_agents[key]`` calls scattered
+        across the gateway.  Those sites had drifted: some popped only
+        ``_running_agents``; some also ``_running_agents_ts``; only one
+        path also cleared ``_busy_ack_ts``.  Each missed entry was a
+        small, persistent leak — a (str_key → float) tuple per session
+        per gateway lifetime.
+
+        Use this at every site that ends a running turn, regardless of
+        cause (normal completion, /stop, /reset, /resume, sentinel
+        cleanup, stale-eviction).  Per-session state that PERSISTS
+        across turns (``_session_model_overrides``, ``_voice_mode``,
+        ``_pending_approvals``, ``_update_prompt_pending``) is NOT
+        touched here — those have their own lifecycles.
+        """
+        if not session_key:
+            return
+        self._running_agents.pop(session_key, None)
+        self._running_agents_ts.pop(session_key, None)
+        if hasattr(self, "_busy_ack_ts"):
+            self._busy_ack_ts.pop(session_key, None)
+
     def _evict_cached_agent(self, session_key: str) -> None:
         """Remove a cached agent for a session (called on /new, /model, etc)."""
         _lock = getattr(self, "_agent_cache_lock", None)
@@ -9748,10 +10076,8 @@ class GatewayRunner:
             
             # Clean up tracking
             tracking_task.cancel()
-            if session_key and session_key in self._running_agents:
-                del self._running_agents[session_key]
             if session_key:
-                self._running_agents_ts.pop(session_key, None)
+                self._release_running_agent_state(session_key)
             if self._draining:
                 self._update_runtime_status("draining")
             
@@ -9880,6 +10206,16 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
                 "Replacing existing gateway instance (PID %d) with --replace.",
                 existing_pid,
             )
+            # Record a takeover marker so the target's shutdown handler
+            # recognises its SIGTERM as a planned takeover and exits 0
+            # (rather than exit 1, which would trigger systemd's
+            # Restart=on-failure and start a flap loop against us).
+            # Best-effort — proceed even if the write fails.
+            try:
+                from gateway.status import write_takeover_marker
+                write_takeover_marker(existing_pid)
+            except Exception as e:
+                logger.debug("Could not write takeover marker: %s", e)
             try:
                 terminate_pid(existing_pid, force=False)
             except ProcessLookupError:
@@ -9889,6 +10225,13 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
                     "Permission denied killing PID %d. Cannot replace.",
                     existing_pid,
                 )
+                # Marker is scoped to a specific target; clean it up on
+                # give-up so it doesn't grief an unrelated future shutdown.
+                try:
+                    from gateway.status import clear_takeover_marker
+                    clear_takeover_marker()
+                except Exception:
+                    pass
                 return False
             # Wait up to 10 seconds for the old process to exit
             for _ in range(20):
@@ -9909,6 +10252,13 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
                 except (ProcessLookupError, PermissionError, OSError):
                     pass
             remove_pid_file()
+            # Clean up any takeover marker the old process didn't consume
+            # (e.g. SIGKILL'd before its shutdown handler could read it).
+            try:
+                from gateway.status import clear_takeover_marker
+                clear_takeover_marker()
+            except Exception:
+                pass
             # Also release all scoped locks left by the old process.
             # Stopped (Ctrl+Z) processes don't release locks on exit,
             # leaving stale lock files that block the new gateway from starting.
@@ -9976,8 +10326,27 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
     # Set up signal handlers
     def shutdown_signal_handler():
         nonlocal _signal_initiated_shutdown
-        _signal_initiated_shutdown = True
-        logger.info("Received SIGTERM/SIGINT — initiating shutdown")
+        # Planned --replace takeover check: when a sibling gateway is
+        # taking over via --replace, it wrote a marker naming this PID
+        # before sending SIGTERM. If present, treat the signal as a
+        # planned shutdown and exit 0 so systemd's Restart=on-failure
+        # doesn't revive us (which would flap-fight the replacer when
+        # both services are enabled, e.g. hermes.service + hermes-
+        # gateway.service from pre-rename installs).
+        planned_takeover = False
+        try:
+            from gateway.status import consume_takeover_marker_for_self
+            planned_takeover = consume_takeover_marker_for_self()
+        except Exception as e:
+            logger.debug("Takeover marker check failed: %s", e)
+
+        if planned_takeover:
+            logger.info(
+                "Received SIGTERM as a planned --replace takeover — exiting cleanly"
+            )
+        else:
+            _signal_initiated_shutdown = True
+            logger.info("Received SIGTERM/SIGINT — initiating shutdown")
         # Diagnostic: log all hermes-related processes so we can identify
         # what triggered the signal (hermes update, hermes gateway restart,
         # a stale detached subprocess, etc.).

gateway/session.py

@@ -802,6 +802,57 @@ class SessionStore:
                 return True
         return False
 
+    def prune_old_entries(self, max_age_days: int) -> int:
+        """Drop SessionEntry records older than max_age_days.
+
+        Pruning is based on ``updated_at`` (last activity), not ``created_at``.
+        A session that's been active within the window is kept regardless of
+        how old it is.  Entries marked ``suspended`` are kept — the user
+        explicitly paused them for later resume.  Entries held by an active
+        process (via has_active_processes_fn) are also kept so long-running
+        background work isn't orphaned.
+
+        Pruning is functionally identical to a natural reset-policy expiry:
+        the transcript in SQLite stays, but the session_key → session_id
+        mapping is dropped and the user starts a fresh session on return.
+
+        ``max_age_days <= 0`` disables pruning; returns 0 immediately.
+        Returns the number of entries removed.
+        """
+        if max_age_days is None or max_age_days <= 0:
+            return 0
+        from datetime import timedelta
+
+        cutoff = _now() - timedelta(days=max_age_days)
+        removed_keys: list[str] = []
+
+        with self._lock:
+            self._ensure_loaded_locked()
+            for key, entry in list(self._entries.items()):
+                if entry.suspended:
+                    continue
+                # Never prune sessions with an active background process
+                # attached — the user may still be waiting on output.
+                if self._has_active_processes_fn is not None:
+                    try:
+                        if self._has_active_processes_fn(entry.session_id):
+                            continue
+                    except Exception:
+                        pass
+                if entry.updated_at < cutoff:
+                    removed_keys.append(key)
+            for key in removed_keys:
+                self._entries.pop(key, None)
+            if removed_keys:
+                self._save()
+
+        if removed_keys:
+            logger.info(
+                "SessionStore pruned %d entries older than %d days",
+                len(removed_keys), max_age_days,
+            )
+        return len(removed_keys)
+
     def suspend_recently_active(self, max_age_seconds: int = 120) -> int:
         """Mark recently-active sessions as suspended.
 

gateway/status.py

@@ -188,8 +188,8 @@ def _write_json_file(path: Path, payload: dict[str, Any]) -> None:
     path.write_text(json.dumps(payload))
 
 
-def _read_pid_record() -> Optional[dict]:
-    pid_path = _get_pid_path()
+def _read_pid_record(pid_path: Optional[Path] = None) -> Optional[dict]:
+    pid_path = pid_path or _get_pid_path()
     if not pid_path.exists():
         return None
 
@@ -212,6 +212,18 @@ def _read_pid_record() -> Optional[dict]:
     return None
 
 
+def _cleanup_invalid_pid_path(pid_path: Path, *, cleanup_stale: bool) -> None:
+    if not cleanup_stale:
+        return
+    try:
+        if pid_path == _get_pid_path():
+            remove_pid_file()
+        else:
+            pid_path.unlink(missing_ok=True)
+    except Exception:
+        pass
+
+
 def write_pid_file() -> None:
     """Write the current process PID and metadata to the gateway PID file."""
     _write_json_file(_get_pid_path(), _build_pid_record())
@@ -413,43 +425,179 @@ def release_all_scoped_locks() -> int:
     return removed
 
 
-def get_running_pid() -> Optional[int]:
+# ── --replace takeover marker ─────────────────────────────────────────
+#
+# When a new gateway starts with ``--replace``, it SIGTERMs the existing
+# gateway so it can take over the bot token. PR #5646 made SIGTERM exit
+# the gateway with code 1 so ``Restart=on-failure`` can revive it after
+# unexpected kills — but that also means a --replace takeover target
+# exits 1, which tricks systemd into reviving it 30 seconds later,
+# starting a flap loop against the replacer when both services are
+# enabled in the user's systemd (e.g. ``hermes.service`` + ``hermes-
+# gateway.service``).
+#
+# The takeover marker breaks the loop: the replacer writes a short-lived
+# file naming the target PID + start_time BEFORE sending SIGTERM.
+# The target's shutdown handler reads the marker and, if it names
+# this process, treats the SIGTERM as a planned takeover and exits 0.
+# The marker is unlinked after the target has consumed it, so a stale
+# marker left by a crashed replacer can grief at most one future
+# shutdown on the same PID — and only within _TAKEOVER_MARKER_TTL_S.
+
+_TAKEOVER_MARKER_FILENAME = ".gateway-takeover.json"
+_TAKEOVER_MARKER_TTL_S = 60  # Marker older than this is treated as stale
+
+
+def _get_takeover_marker_path() -> Path:
+    """Return the path to the --replace takeover marker file."""
+    home = get_hermes_home()
+    return home / _TAKEOVER_MARKER_FILENAME
+
+
+def write_takeover_marker(target_pid: int) -> bool:
+    """Record that ``target_pid`` is being replaced by the current process.
+
+    Captures the target's ``start_time`` so that PID reuse after the
+    target exits cannot later match the marker. Also records the
+    replacer's PID and a UTC timestamp for TTL-based staleness checks.
+
+    Returns True on successful write, False on any failure. The caller
+    should proceed with the SIGTERM even if the write fails (the marker
+    is a best-effort signal, not a correctness requirement).
+    """
+    try:
+        target_start_time = _get_process_start_time(target_pid)
+        record = {
+            "target_pid": target_pid,
+            "target_start_time": target_start_time,
+            "replacer_pid": os.getpid(),
+            "written_at": _utc_now_iso(),
+        }
+        _write_json_file(_get_takeover_marker_path(), record)
+        return True
+    except (OSError, PermissionError):
+        return False
+
+
+def consume_takeover_marker_for_self() -> bool:
+    """Check & unlink the takeover marker if it names the current process.
+
+    Returns True only when a valid (non-stale) marker names this PID +
+    start_time. A returning True indicates the current SIGTERM is a
+    planned --replace takeover; the caller should exit 0 instead of
+    signalling ``_signal_initiated_shutdown``.
+
+    Always unlinks the marker on match (and on detected staleness) so
+    subsequent unrelated signals don't re-trigger.
+    """
+    path = _get_takeover_marker_path()
+    record = _read_json_file(path)
+    if not record:
+        return False
+
+    # Any malformed or stale marker → drop it and return False
+    try:
+        target_pid = int(record["target_pid"])
+        target_start_time = record.get("target_start_time")
+        written_at = record.get("written_at") or ""
+    except (KeyError, TypeError, ValueError):
+        try:
+            path.unlink(missing_ok=True)
+        except OSError:
+            pass
+        return False
+
+    # TTL guard: a stale marker older than _TAKEOVER_MARKER_TTL_S is ignored.
+    stale = False
+    try:
+        written_dt = datetime.fromisoformat(written_at)
+        age = (datetime.now(timezone.utc) - written_dt).total_seconds()
+        if age > _TAKEOVER_MARKER_TTL_S:
+            stale = True
+    except (TypeError, ValueError):
+        stale = True  # Unparseable timestamp — treat as stale
+
+    if stale:
+        try:
+            path.unlink(missing_ok=True)
+        except OSError:
+            pass
+        return False
+
+    # Does the marker name THIS process?
+    our_pid = os.getpid()
+    our_start_time = _get_process_start_time(our_pid)
+    matches = (
+        target_pid == our_pid
+        and target_start_time is not None
+        and our_start_time is not None
+        and target_start_time == our_start_time
+    )
+
+    # Consume the marker whether it matched or not — a marker that doesn't
+    # match our identity is stale-for-us anyway.
+    try:
+        path.unlink(missing_ok=True)
+    except OSError:
+        pass
+
+    return matches
+
+
+def clear_takeover_marker() -> None:
+    """Remove the takeover marker unconditionally. Safe to call repeatedly."""
+    try:
+        _get_takeover_marker_path().unlink(missing_ok=True)
+    except OSError:
+        pass
+
+
+def get_running_pid(
+    pid_path: Optional[Path] = None,
+    *,
+    cleanup_stale: bool = True,
+) -> Optional[int]:
     """Return the PID of a running gateway instance, or ``None``.
 
     Checks the PID file and verifies the process is actually alive.
     Cleans up stale PID files automatically.
     """
-    record = _read_pid_record()
+    resolved_pid_path = pid_path or _get_pid_path()
+    record = _read_pid_record(resolved_pid_path)
     if not record:
-        remove_pid_file()
+        _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
         return None
 
     try:
         pid = int(record["pid"])
     except (KeyError, TypeError, ValueError):
-        remove_pid_file()
+        _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
         return None
 
     try:
         os.kill(pid, 0)  # signal 0 = existence check, no actual signal sent
     except (ProcessLookupError, PermissionError):
-        remove_pid_file()
+        _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
         return None
 
     recorded_start = record.get("start_time")
     current_start = _get_process_start_time(pid)
     if recorded_start is not None and current_start is not None and current_start != recorded_start:
-        remove_pid_file()
+        _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
         return None
 
     if not _looks_like_gateway_process(pid):
         if not _record_looks_like_gateway(record):
-            remove_pid_file()
+            _cleanup_invalid_pid_path(resolved_pid_path, cleanup_stale=cleanup_stale)
             return None
 
     return pid
 
 
-def is_gateway_running() -> bool:
+def is_gateway_running(
+    pid_path: Optional[Path] = None,
+    *,
+    cleanup_stale: bool = True,
+) -> bool:
     """Check if the gateway daemon is currently running."""
-    return get_running_pid() is not None
+    return get_running_pid(pid_path, cleanup_stale=cleanup_stale) is not None

gateway/stream_consumer.py

@@ -100,6 +100,14 @@ class GatewayStreamConsumer:
         self._flood_strikes = 0         # Consecutive flood-control edit failures
         self._current_edit_interval = self.cfg.edit_interval  # Adaptive backoff
         self._final_response_sent = False
+        # Cache adapter lifecycle capability: only platforms that need an
+        # explicit finalize call (e.g. DingTalk AI Cards) force us to make
+        # a redundant final edit.  Everyone else keeps the fast path.
+        # Use ``is True`` (not ``bool(...)``) so MagicMock attribute access
+        # in tests doesn't incorrectly enable this path.
+        self._adapter_requires_finalize: bool = (
+            getattr(adapter, "REQUIRES_EDIT_FINALIZE", False) is True
+        )
 
         # Think-block filter state (mirrors CLI's _stream_delta tag suppression)
         self._in_think_block = False
@@ -361,7 +369,16 @@ class GatewayStreamConsumer:
                     if not got_done and not got_segment_break and commentary_text is None:
                         display_text += self.cfg.cursor
 
-                    current_update_visible = await self._send_or_edit(display_text)
+                    # Segment break: finalize the current message so platforms
+                    # that need explicit closure (e.g. DingTalk AI Cards) don't
+                    # leave the previous segment stuck in a loading state when
+                    # the next segment (tool progress, next chunk) creates a
+                    # new message below it.  got_done has its own finalize
+                    # path below so we don't finalize here for it.
+                    current_update_visible = await self._send_or_edit(
+                        display_text,
+                        finalize=got_segment_break,
+                    )
                     self._last_edit_time = time.monotonic()
 
                 if got_done:
@@ -372,10 +389,22 @@ class GatewayStreamConsumer:
                     if self._accumulated:
                         if self._fallback_final_send:
                             await self._send_fallback_final(self._accumulated)
-                        elif current_update_visible:
+                        elif (
+                            current_update_visible
+                            and not self._adapter_requires_finalize
+                        ):
+                            # Mid-stream edit above already delivered the
+                            # final accumulated content.  Skip the redundant
+                            # final edit — but only for adapters that don't
+                            # need an explicit finalize signal.
                             self._final_response_sent = True
                         elif self._message_id:
-                            self._final_response_sent = await self._send_or_edit(self._accumulated)
+                            # Either the mid-stream edit didn't run (no
+                            # visible update this tick) OR the adapter needs
+                            # explicit finalize=True to close the stream.
+                            self._final_response_sent = await self._send_or_edit(
+                                self._accumulated, finalize=True,
+                            )
                         elif not self._already_sent:
                             self._final_response_sent = await self._send_or_edit(self._accumulated)
                     return
@@ -633,12 +662,15 @@ class GatewayStreamConsumer:
             logger.error("Commentary send error: %s", e)
             return False
 
-    async def _send_or_edit(self, text: str) -> bool:
+    async def _send_or_edit(self, text: str, *, finalize: bool = False) -> bool:
         """Send or edit the streaming message.
 
         Returns True if the text was successfully delivered (sent or edited),
         False otherwise.  Callers like the overflow split loop use this to
         decide whether to advance past the delivered chunk.
+
+        ``finalize`` is True when this is the last edit in a streaming
+        sequence.
         """
         # Strip MEDIA: directives so they don't appear as visible text.
         # Media files are delivered as native attachments after the stream
@@ -672,14 +704,22 @@ class GatewayStreamConsumer:
         try:
             if self._message_id is not None:
                 if self._edit_supported:
-                    # Skip if text is identical to what we last sent
-                    if text == self._last_sent_text:
+                    # Skip if text is identical to what we last sent.
+                    # Exception: adapters that require an explicit finalize
+                    # call (REQUIRES_EDIT_FINALIZE) must still receive the
+                    # finalize=True edit even when content is unchanged, so
+                    # their streaming UI can transition out of the in-
+                    # progress state.  Everyone else short-circuits.
+                    if text == self._last_sent_text and not (
+                        finalize and self._adapter_requires_finalize
+                    ):
                         return True
                     # Edit existing message
                     result = await self.adapter.edit_message(
                         chat_id=self.chat_id,
                         message_id=self._message_id,
                         content=text,
+                        finalize=finalize,
                     )
                     if result.success:
                         self._already_sent = True

hermes_cli/auth.py

@@ -233,6 +233,14 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         api_key_env_vars=("XAI_API_KEY",),
         base_url_env_var="XAI_BASE_URL",
     ),
+    "nvidia": ProviderConfig(
+        id="nvidia",
+        name="NVIDIA NIM",
+        auth_type="api_key",
+        inference_base_url="https://integrate.api.nvidia.com/v1",
+        api_key_env_vars=("NVIDIA_API_KEY",),
+        base_url_env_var="NVIDIA_BASE_URL",
+    ),
     "ai-gateway": ProviderConfig(
         id="ai-gateway",
         name="Vercel AI Gateway",
@@ -2151,6 +2159,62 @@ def refresh_nous_oauth_from_state(
     )
 
 
+NOUS_DEVICE_CODE_SOURCE = "device_code"
+
+
+def persist_nous_credentials(
+    creds: Dict[str, Any],
+    *,
+    label: Optional[str] = None,
+):
+    """Persist minted Nous OAuth credentials as the singleton provider state
+    and ensure the credential pool is in sync.
+
+    Nous credentials are read at runtime from two independent locations:
+
+    - ``providers.nous``: singleton state read by
+      ``resolve_nous_runtime_credentials()`` during 401 recovery and by
+      ``_seed_from_singletons()`` during pool load.
+    - ``credential_pool.nous``: used by the runtime ``pool.select()`` path.
+
+    Historically ``hermes auth add nous`` wrote a ``manual:device_code`` pool
+    entry only, skipping ``providers.nous``.  When the 24h agent_key TTL
+    expired, the recovery path read the empty singleton state and raised
+    ``AuthError`` silently (``logger.debug`` at INFO level).
+
+    This helper writes ``providers.nous`` then calls ``load_pool("nous")`` so
+    ``_seed_from_singletons`` materialises the canonical ``device_code`` pool
+    entry from the singleton.  Re-running login upserts the same entry in
+    place; the pool never accumulates duplicate device_code rows.
+
+    ``label`` is an optional user-chosen display name (from
+    ``hermes auth add nous --label <name>``).  It gets embedded in the
+    singleton state so that ``_seed_from_singletons`` uses it as the pool
+    entry's label on every subsequent ``load_pool("nous")`` instead of the
+    auto-derived token fingerprint.  When ``None``, the auto-derived label
+    via ``label_from_token`` is used (unchanged default behaviour).
+
+    Returns the upserted :class:`PooledCredential` entry (or ``None`` if
+    seeding somehow produced no match — shouldn't happen).
+    """
+    from agent.credential_pool import load_pool
+
+    state = dict(creds)
+    if label and str(label).strip():
+        state["label"] = str(label).strip()
+
+    with _auth_store_lock():
+        auth_store = _load_auth_store()
+        _save_provider_state(auth_store, "nous", state)
+        _save_auth_store(auth_store)
+
+    pool = load_pool("nous")
+    return next(
+        (e for e in pool.entries() if e.source == NOUS_DEVICE_CODE_SOURCE),
+        None,
+    )
+
+
 def resolve_nous_runtime_credentials(
     *,
     min_key_ttl_seconds: int = DEFAULT_AGENT_KEY_MIN_TTL_SECONDS,

hermes_cli/auth_commands.py

@@ -217,19 +217,15 @@ def auth_add_command(args) -> None:
             ca_bundle=getattr(args, "ca_bundle", None),
             min_key_ttl_seconds=max(60, int(getattr(args, "min_key_ttl_seconds", 5 * 60))),
         )
-        label = (getattr(args, "label", None) or "").strip() or label_from_token(
-            creds.get("access_token", ""),
-            _oauth_default_label(provider, len(pool.entries()) + 1),
+        # Honor `--label <name>` so nous matches other providers' UX.  The
+        # helper embeds this into providers.nous so that label_from_token
+        # doesn't overwrite it on every subsequent load_pool("nous").
+        custom_label = (getattr(args, "label", None) or "").strip() or None
+        entry = auth_mod.persist_nous_credentials(creds, label=custom_label)
+        shown_label = entry.label if entry is not None else label_from_token(
+            creds.get("access_token", ""), _oauth_default_label(provider, 1),
         )
-        entry = PooledCredential.from_dict(provider, {
-            **creds,
-            "label": label,
-            "auth_type": AUTH_TYPE_OAUTH,
-            "source": f"{SOURCE_MANUAL}:device_code",
-            "base_url": creds.get("inference_base_url"),
-        })
-        pool.add_entry(entry)
-        print(f'Added {provider} OAuth credential #{len(pool.entries())}: "{entry.label}"')
+        print(f'Saved {provider} OAuth device-code credentials: "{shown_label}"')
         return
 
     if provider == "openai-codex":

hermes_cli/clipboard.py

@@ -7,8 +7,8 @@ CLI tools that ship with the platform (or are commonly installed).
 
 Platform support:
   macOS   — osascript (always available), pngpaste (if installed)
-  Windows — PowerShell via .NET System.Windows.Forms.Clipboard
-  WSL2    — powershell.exe via .NET System.Windows.Forms.Clipboard
+  Windows — PowerShell via WinForms, Get-Clipboard, file-drop fallback
+  WSL2    — powershell.exe via WinForms, Get-Clipboard, file-drop fallback
   Linux   — wl-paste (Wayland), xclip (X11)
 """
 
@@ -46,10 +46,11 @@ def has_clipboard_image() -> bool:
         return _macos_has_image()
     if sys.platform == "win32":
         return _windows_has_image()
-    if _is_wsl():
-        return _wsl_has_image()
-    if os.environ.get("WAYLAND_DISPLAY"):
-        return _wayland_has_image()
+    # Match _linux_save fallthrough order: WSL → Wayland → X11
+    if _is_wsl() and _wsl_has_image():
+        return True
+    if os.environ.get("WAYLAND_DISPLAY") and _wayland_has_image():
+        return True
     return _xclip_has_image()
 
 
@@ -135,6 +136,114 @@ _PS_EXTRACT_IMAGE = (
     "[System.Convert]::ToBase64String($ms.ToArray())"
 )
 
+_PS_CHECK_IMAGE_GET_CLIPBOARD = (
+    "try { "
+    "$img = Get-Clipboard -Format Image -ErrorAction Stop;"
+    "if ($null -ne $img) { 'True' } else { 'False' }"
+    "} catch { 'False' }"
+)
+
+_PS_EXTRACT_IMAGE_GET_CLIPBOARD = (
+    "try { "
+    "Add-Type -AssemblyName System.Drawing;"
+    "Add-Type -AssemblyName PresentationCore;"
+    "Add-Type -AssemblyName WindowsBase;"
+    "$img = Get-Clipboard -Format Image -ErrorAction Stop;"
+    "if ($null -eq $img) { exit 1 }"
+    "$ms = New-Object System.IO.MemoryStream;"
+    "if ($img -is [System.Drawing.Image]) {"
+    "$img.Save($ms, [System.Drawing.Imaging.ImageFormat]::Png)"
+    "} elseif ($img -is [System.Windows.Media.Imaging.BitmapSource]) {"
+    "$enc = New-Object System.Windows.Media.Imaging.PngBitmapEncoder;"
+    "$enc.Frames.Add([System.Windows.Media.Imaging.BitmapFrame]::Create($img));"
+    "$enc.Save($ms)"
+    "} else { exit 2 }"
+    "[System.Convert]::ToBase64String($ms.ToArray())"
+    "} catch { exit 1 }"
+)
+
+_FILEDROP_IMAGE_EXTS = "'.png','.jpg','.jpeg','.gif','.webp','.bmp','.tiff','.tif'"
+
+_PS_CHECK_FILEDROP_IMAGE = (
+    "try { "
+    "$files = Get-Clipboard -Format FileDropList -ErrorAction Stop;"
+    f"$exts = @({_FILEDROP_IMAGE_EXTS});"
+    "$hit = $files | Where-Object { $exts -contains ([System.IO.Path]::GetExtension($_).ToLowerInvariant()) } | Select-Object -First 1;"
+    "if ($null -ne $hit) { 'True' } else { 'False' }"
+    "} catch { 'False' }"
+)
+
+_PS_EXTRACT_FILEDROP_IMAGE = (
+    "try { "
+    "$files = Get-Clipboard -Format FileDropList -ErrorAction Stop;"
+    f"$exts = @({_FILEDROP_IMAGE_EXTS});"
+    "$hit = $files | Where-Object { $exts -contains ([System.IO.Path]::GetExtension($_).ToLowerInvariant()) } | Select-Object -First 1;"
+    "if ($null -eq $hit) { exit 1 }"
+    "[System.Convert]::ToBase64String([System.IO.File]::ReadAllBytes($hit))"
+    "} catch { exit 1 }"
+)
+
+_POWERSHELL_HAS_IMAGE_SCRIPTS = (
+    _PS_CHECK_IMAGE,
+    _PS_CHECK_IMAGE_GET_CLIPBOARD,
+    _PS_CHECK_FILEDROP_IMAGE,
+)
+
+_POWERSHELL_EXTRACT_IMAGE_SCRIPTS = (
+    _PS_EXTRACT_IMAGE,
+    _PS_EXTRACT_IMAGE_GET_CLIPBOARD,
+    _PS_EXTRACT_FILEDROP_IMAGE,
+)
+
+
+def _run_powershell(exe: str, script: str, timeout: int) -> subprocess.CompletedProcess:
+    return subprocess.run(
+        [exe, "-NoProfile", "-NonInteractive", "-Command", script],
+        capture_output=True, text=True, timeout=timeout,
+    )
+
+
+def _write_base64_image(dest: Path, b64_data: str) -> bool:
+    image_bytes = base64.b64decode(b64_data, validate=True)
+    dest.write_bytes(image_bytes)
+    return dest.exists() and dest.stat().st_size > 0
+
+
+def _powershell_has_image(exe: str, *, timeout: int, label: str) -> bool:
+    for script in _POWERSHELL_HAS_IMAGE_SCRIPTS:
+        try:
+            r = _run_powershell(exe, script, timeout=timeout)
+            if r.returncode == 0 and "True" in r.stdout:
+                return True
+        except FileNotFoundError:
+            logger.debug("%s not found — clipboard unavailable", exe)
+            return False
+        except Exception as e:
+            logger.debug("%s clipboard image check failed: %s", label, e)
+    return False
+
+
+def _powershell_save_image(exe: str, dest: Path, *, timeout: int, label: str) -> bool:
+    for script in _POWERSHELL_EXTRACT_IMAGE_SCRIPTS:
+        try:
+            r = _run_powershell(exe, script, timeout=timeout)
+            if r.returncode != 0:
+                continue
+
+            b64_data = r.stdout.strip()
+            if not b64_data:
+                continue
+
+            if _write_base64_image(dest, b64_data):
+                return True
+        except FileNotFoundError:
+            logger.debug("%s not found — clipboard unavailable", exe)
+            return False
+        except Exception as e:
+            logger.debug("%s clipboard image extraction failed: %s", label, e)
+            dest.unlink(missing_ok=True)
+    return False
+
 
 # ── Native Windows ────────────────────────────────────────────────────────
 
@@ -175,15 +284,7 @@ def _windows_has_image() -> bool:
     ps = _get_ps_exe()
     if ps is None:
         return False
-    try:
-        r = subprocess.run(
-            [ps, "-NoProfile", "-NonInteractive", "-Command", _PS_CHECK_IMAGE],
-            capture_output=True, text=True, timeout=5,
-        )
-        return r.returncode == 0 and "True" in r.stdout
-    except Exception as e:
-        logger.debug("Windows clipboard image check failed: %s", e)
-    return False
+    return _powershell_has_image(ps, timeout=5, label="Windows")
 
 
 def _windows_save(dest: Path) -> bool:
@@ -192,26 +293,7 @@ def _windows_save(dest: Path) -> bool:
     if ps is None:
         logger.debug("No PowerShell found — Windows clipboard image paste unavailable")
         return False
-    try:
-        r = subprocess.run(
-            [ps, "-NoProfile", "-NonInteractive", "-Command", _PS_EXTRACT_IMAGE],
-            capture_output=True, text=True, timeout=15,
-        )
-        if r.returncode != 0:
-            return False
-
-        b64_data = r.stdout.strip()
-        if not b64_data:
-            return False
-
-        png_bytes = base64.b64decode(b64_data)
-        dest.write_bytes(png_bytes)
-        return dest.exists() and dest.stat().st_size > 0
-
-    except Exception as e:
-        logger.debug("Windows clipboard image extraction failed: %s", e)
-        dest.unlink(missing_ok=True)
-    return False
+    return _powershell_save_image(ps, dest, timeout=15, label="Windows")
 
 
 # ── Linux ────────────────────────────────────────────────────────────────
@@ -235,45 +317,12 @@ def _linux_save(dest: Path) -> bool:
 
 def _wsl_has_image() -> bool:
     """Check if Windows clipboard has an image (via powershell.exe)."""
-    try:
-        r = subprocess.run(
-            ["powershell.exe", "-NoProfile", "-NonInteractive", "-Command",
-             _PS_CHECK_IMAGE],
-            capture_output=True, text=True, timeout=8,
-        )
-        return r.returncode == 0 and "True" in r.stdout
-    except FileNotFoundError:
-        logger.debug("powershell.exe not found — WSL clipboard unavailable")
-    except Exception as e:
-        logger.debug("WSL clipboard check failed: %s", e)
-    return False
+    return _powershell_has_image("powershell.exe", timeout=8, label="WSL")
 
 
 def _wsl_save(dest: Path) -> bool:
     """Extract clipboard image via powershell.exe → base64 → decode to PNG."""
-    try:
-        r = subprocess.run(
-            ["powershell.exe", "-NoProfile", "-NonInteractive", "-Command",
-             _PS_EXTRACT_IMAGE],
-            capture_output=True, text=True, timeout=15,
-        )
-        if r.returncode != 0:
-            return False
-
-        b64_data = r.stdout.strip()
-        if not b64_data:
-            return False
-
-        png_bytes = base64.b64decode(b64_data)
-        dest.write_bytes(png_bytes)
-        return dest.exists() and dest.stat().st_size > 0
-
-    except FileNotFoundError:
-        logger.debug("powershell.exe not found — WSL clipboard unavailable")
-    except Exception as e:
-        logger.debug("WSL clipboard extraction failed: %s", e)
-        dest.unlink(missing_ok=True)
-    return False
+    return _powershell_save_image("powershell.exe", dest, timeout=15, label="WSL")
 
 
 # ── Wayland (wl-paste) ──────────────────────────────────────────────────

hermes_cli/commands.py

@@ -87,8 +87,12 @@ COMMAND_REGISTRY: list[CommandDef] = [
                aliases=("bg",), args_hint="<prompt>"),
     CommandDef("btw", "Ephemeral side question using session context (no tools, not persisted)", "Session",
                args_hint="<question>"),
+    CommandDef("agents", "Show active agents and running tasks", "Session",
+               aliases=("tasks",)),
     CommandDef("queue", "Queue a prompt for the next turn (doesn't interrupt)", "Session",
                aliases=("q",), args_hint="<prompt>"),
+    CommandDef("steer", "Inject a message after the next tool call without interrupting", "Session",
+               args_hint="<prompt>"),
     CommandDef("status", "Show session info", "Session"),
     CommandDef("profile", "Show active profile name and home directory", "Info"),
     CommandDef("sethome", "Set this chat as the home channel", "Session",
@@ -99,7 +103,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
     # Configuration
     CommandDef("config", "Show current configuration", "Configuration",
                cli_only=True),
-    CommandDef("model", "Switch model for this session", "Configuration", args_hint="[model] [--global]"),
+    CommandDef("model", "Switch model for this session", "Configuration", args_hint="[model] [--provider name] [--global]"),
     CommandDef("provider", "Show available providers and current provider",
                "Configuration"),
     CommandDef("gquota", "Show Google Gemini Code Assist quota usage", "Info"),
@@ -120,7 +124,7 @@ COMMAND_REGISTRY: list[CommandDef] = [
                args_hint="[normal|fast|status]",
                subcommands=("normal", "fast", "status", "on", "off")),
     CommandDef("skin", "Show or change the display skin/theme", "Configuration",
-               cli_only=True, args_hint="[name]"),
+               args_hint="[name]"),
     CommandDef("voice", "Toggle voice mode", "Configuration",
                args_hint="[on|off|tts|status]", subcommands=("on", "off", "tts", "status")),
 
@@ -155,7 +159,9 @@ COMMAND_REGISTRY: list[CommandDef] = [
                args_hint="[days]"),
     CommandDef("platforms", "Show gateway/messaging platform status", "Info",
                cli_only=True, aliases=("gateway",)),
-    CommandDef("paste", "Check clipboard for an image and attach it", "Info",
+    CommandDef("copy", "Copy the last assistant response to clipboard", "Info",
+               cli_only=True, args_hint="[number]"),
+    CommandDef("paste", "Attach clipboard image from your clipboard", "Info",
                cli_only=True),
     CommandDef("image", "Attach a local image file for your next prompt", "Info",
                cli_only=True, args_hint="<path>"),
@@ -254,6 +260,36 @@ GATEWAY_KNOWN_COMMANDS: frozenset[str] = frozenset(
 )
 
 
+# Commands that must never be queued behind an active gateway session.
+# These are explicit control/info commands handled by the gateway itself;
+# if they get queued as pending text, the safety net in gateway.run will
+# discard them before they ever reach the user.
+ACTIVE_SESSION_BYPASS_COMMANDS: frozenset[str] = frozenset(
+    {
+        "agents",
+        "approve",
+        "background",
+        "commands",
+        "deny",
+        "help",
+        "new",
+        "profile",
+        "queue",
+        "restart",
+        "status",
+        "steer",
+        "stop",
+        "update",
+    }
+)
+
+
+def should_bypass_active_session(command_name: str | None) -> bool:
+    """Return True when a slash command must bypass active-session queuing."""
+    cmd = resolve_command(command_name) if command_name else None
+    return bool(cmd and cmd.name in ACTIVE_SESSION_BYPASS_COMMANDS)
+
+
 def _resolve_config_gates() -> set[str]:
     """Return canonical names of commands whose ``gateway_config_gate`` is truthy.
 
@@ -1044,6 +1080,51 @@ class SlashCommandCompleter(Completer):
                 display_meta=f"{fp}  {meta}" if meta else fp,
             )
 
+    @staticmethod
+    def _skin_completions(sub_text: str, sub_lower: str):
+        """Yield completions for /skin from available skins."""
+        try:
+            from hermes_cli.skin_engine import list_skins
+            for s in list_skins():
+                name = s["name"]
+                if name.startswith(sub_lower) and name != sub_lower:
+                    yield Completion(
+                        name,
+                        start_position=-len(sub_text),
+                        display=name,
+                        display_meta=s.get("description", "") or s.get("source", ""),
+                    )
+        except Exception:
+            pass
+
+    @staticmethod
+    def _personality_completions(sub_text: str, sub_lower: str):
+        """Yield completions for /personality from configured personalities."""
+        try:
+            from hermes_cli.config import load_config
+            personalities = load_config().get("agent", {}).get("personalities", {})
+            if "none".startswith(sub_lower) and "none" != sub_lower:
+                yield Completion(
+                    "none",
+                    start_position=-len(sub_text),
+                    display="none",
+                    display_meta="clear personality overlay",
+                )
+            for name, prompt in personalities.items():
+                if name.startswith(sub_lower) and name != sub_lower:
+                    if isinstance(prompt, dict):
+                        meta = prompt.get("description") or prompt.get("system_prompt", "")[:50]
+                    else:
+                        meta = str(prompt)[:50]
+                    yield Completion(
+                        name,
+                        start_position=-len(sub_text),
+                        display=name,
+                        display_meta=meta,
+                    )
+        except Exception:
+            pass
+
     def _model_completions(self, sub_text: str, sub_lower: str):
         """Yield completions for /model from config aliases + built-in aliases."""
         seen = set()
@@ -1098,10 +1179,17 @@ class SlashCommandCompleter(Completer):
             sub_text = parts[1] if len(parts) > 1 else ""
             sub_lower = sub_text.lower()
 
-            # Dynamic model alias completions for /model
-            if " " not in sub_text and base_cmd == "/model":
-                yield from self._model_completions(sub_text, sub_lower)
-                return
+            # Dynamic completions for commands with runtime lists
+            if " " not in sub_text:
+                if base_cmd == "/model":
+                    yield from self._model_completions(sub_text, sub_lower)
+                    return
+                if base_cmd == "/skin":
+                    yield from self._skin_completions(sub_text, sub_lower)
+                    return
+                if base_cmd == "/personality":
+                    yield from self._personality_completions(sub_text, sub_lower)
+                    return
 
             # Static subcommand completions
             if " " not in sub_text and base_cmd in SUBCOMMANDS and self._command_allowed(base_cmd):

hermes_cli/config.py

@@ -12,6 +12,7 @@ This module provides:
 - hermes config wizard   - Re-run setup wizard
 """
 
+import copy
 import os
 import platform
 import re
@@ -26,6 +27,7 @@ from typing import Dict, Any, Optional, List, Tuple
 
 _IS_WINDOWS = platform.system() == "Windows"
 _ENV_VAR_NAME_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*$")
+_LAST_EXPANDED_CONFIG_BY_PATH: Dict[str, Any] = {}
 # Env var names written to .env that aren't in OPTIONAL_ENV_VARS
 # (managed by setup/provider flows directly).
 _EXTRA_ENV_KEYS = frozenset({
@@ -44,7 +46,8 @@ _EXTRA_ENV_KEYS = frozenset({
     "WEIXIN_HOME_CHANNEL", "WEIXIN_HOME_CHANNEL_NAME", "WEIXIN_DM_POLICY", "WEIXIN_GROUP_POLICY",
     "WEIXIN_ALLOWED_USERS", "WEIXIN_GROUP_ALLOWED_USERS", "WEIXIN_ALLOW_ALL_USERS",
     "BLUEBUBBLES_SERVER_URL", "BLUEBUBBLES_PASSWORD",
-    "QQ_APP_ID", "QQ_CLIENT_SECRET", "QQ_HOME_CHANNEL", "QQ_HOME_CHANNEL_NAME",
+    "QQ_APP_ID", "QQ_CLIENT_SECRET", "QQBOT_HOME_CHANNEL", "QQBOT_HOME_CHANNEL_NAME",
+    "QQ_HOME_CHANNEL", "QQ_HOME_CHANNEL_NAME",  # legacy aliases (pre-rename, still read for back-compat)
     "QQ_ALLOWED_USERS", "QQ_GROUP_ALLOWED_USERS", "QQ_ALLOW_ALL_USERS", "QQ_MARKDOWN_SUPPORT",
     "QQ_STT_API_KEY", "QQ_STT_BASE_URL", "QQ_STT_MODEL",
     "TERMINAL_ENV", "TERMINAL_SSH_KEY", "TERMINAL_SSH_PORT",
@@ -417,6 +420,7 @@ DEFAULT_CONFIG = {
         "command_timeout": 30,  # Timeout for browser commands in seconds (screenshot, navigate, etc.)
         "record_sessions": False,  # Auto-record browser sessions as WebM videos
         "allow_private_urls": False,  # Allow navigating to private/internal IPs (localhost, 192.168.x.x, etc.)
+        "cdp_url": "",  # Optional persistent CDP endpoint for attaching to an existing Chromium/Chrome
         "camofox": {
             # When true, Hermes sends a stable profile-scoped userId to Camofox
             # so the server maps it to a persistent Firefox profile automatically.
@@ -537,6 +541,13 @@ DEFAULT_CONFIG = {
             "api_key": "",
             "timeout": 30,
         },
+        "title_generation": {
+            "provider": "auto",
+            "model": "",
+            "base_url": "",
+            "api_key": "",
+            "timeout": 30,
+        },
     },
     
     "display": {
@@ -760,6 +771,20 @@ DEFAULT_CONFIG = {
         "wrap_response": True,
     },
 
+    # execute_code settings — controls the tool used for programmatic tool calls.
+    "code_execution": {
+        # Execution mode:
+        #   project (default) — scripts run in the session's working directory
+        #     with the active virtualenv/conda env's python, so project deps
+        #     (pandas, torch, project packages) and relative paths resolve.
+        #   strict            — scripts run in an isolated temp directory with
+        #     hermes-agent's own python (sys.executable). Maximum isolation
+        #     and reproducibility; project deps and relative paths won't work.
+        # Env scrubbing (strips *_API_KEY, *_TOKEN, *_SECRET, ...) and the
+        # tool whitelist apply identically in both modes.
+        "mode": "project",
+    },
+
     # Logging — controls file logging to ~/.hermes/logs/.
     # agent.log captures INFO+ (all agent activity); errors.log captures WARNING+.
     "logging": {
@@ -777,7 +802,7 @@ DEFAULT_CONFIG = {
     },
 
     # Config schema version - bump this when adding new required fields
-    "_config_version": 18,
+    "_config_version": 19,
 }
 
 # =============================================================================
@@ -861,6 +886,22 @@ OPTIONAL_ENV_VARS = {
         "category": "provider",
         "advanced": True,
     },
+    "NVIDIA_API_KEY": {
+        "description": "NVIDIA NIM API key (build.nvidia.com or local NIM endpoint)",
+        "prompt": "NVIDIA NIM API key",
+        "url": "https://build.nvidia.com/",
+        "password": True,
+        "category": "provider",
+        "advanced": True,
+    },
+    "NVIDIA_BASE_URL": {
+        "description": "NVIDIA NIM base URL override (e.g. http://localhost:8000/v1 for local NIM)",
+        "prompt": "NVIDIA NIM base URL (leave empty for default)",
+        "url": None,
+        "password": False,
+        "category": "provider",
+        "advanced": True,
+    },
     "GLM_API_KEY": {
         "description": "Z.AI / GLM API key (also recognized as ZAI_API_KEY / Z_AI_API_KEY)",
         "prompt": "Z.AI / GLM API key",
@@ -1518,12 +1559,12 @@ OPTIONAL_ENV_VARS = {
         "prompt": "Allow All QQ Users",
         "category": "messaging",
     },
-    "QQ_HOME_CHANNEL": {
+    "QQBOT_HOME_CHANNEL": {
         "description": "Default QQ channel/group for cron delivery and notifications",
         "prompt": "QQ Home Channel",
         "category": "messaging",
     },
-    "QQ_HOME_CHANNEL_NAME": {
+    "QQBOT_HOME_CHANNEL_NAME": {
         "description": "Display name for the QQ home channel",
         "prompt": "QQ Home Channel Name",
         "category": "messaging",
@@ -2610,6 +2651,85 @@ def _expand_env_vars(obj):
     return obj
 
 
+def _items_by_unique_name(items):
+    """Return a name-indexed dict only when all items have unique string names."""
+    if not isinstance(items, list):
+        return None
+    indexed = {}
+    for item in items:
+        if not isinstance(item, dict) or not isinstance(item.get("name"), str):
+            return None
+        name = item["name"]
+        if name in indexed:
+            return None
+        indexed[name] = item
+    return indexed
+
+
+def _preserve_env_ref_templates(current, raw, loaded_expanded=None):
+    """Restore raw ``${VAR}`` templates when a value is otherwise unchanged.
+
+    ``load_config()`` expands env refs for runtime use. When a caller later
+    persists that config after modifying some unrelated setting, keep the
+    original on-disk template instead of writing the expanded plaintext
+    secret back to ``config.yaml``.
+
+    Prefer preserving the raw template when ``current`` still matches either
+    the value previously returned by ``load_config()`` for this config path or
+    the current environment expansion of ``raw``. This handles env-var
+    rotation between load and save while still treating mixed literal/template
+    string edits as caller-owned once their rendered value diverges.
+    """
+    if isinstance(current, str) and isinstance(raw, str) and re.search(r"\${[^}]+}", raw):
+        if current == raw:
+            return raw
+        if isinstance(loaded_expanded, str) and current == loaded_expanded:
+            return raw
+        if _expand_env_vars(raw) == current:
+            return raw
+        return current
+
+    if isinstance(current, dict) and isinstance(raw, dict):
+        return {
+            key: _preserve_env_ref_templates(
+                value,
+                raw.get(key),
+                loaded_expanded.get(key) if isinstance(loaded_expanded, dict) else None,
+            )
+            for key, value in current.items()
+        }
+
+    if isinstance(current, list) and isinstance(raw, list):
+        # Prefer matching named config objects (e.g. custom_providers) by name
+        # so harmless reordering doesn't drop the original template. If names
+        # are duplicated, fall back to positional matching instead of silently
+        # shadowing one entry.
+        current_by_name = _items_by_unique_name(current)
+        raw_by_name = _items_by_unique_name(raw)
+        loaded_by_name = _items_by_unique_name(loaded_expanded)
+        if current_by_name is not None and raw_by_name is not None:
+            return [
+                _preserve_env_ref_templates(
+                    item,
+                    raw_by_name.get(item.get("name")),
+                    loaded_by_name.get(item.get("name")) if loaded_by_name is not None else None,
+                )
+                for item in current
+            ]
+        return [
+            _preserve_env_ref_templates(
+                item,
+                raw[index] if index < len(raw) else None,
+                loaded_expanded[index]
+                if isinstance(loaded_expanded, list) and index < len(loaded_expanded)
+                else None,
+            )
+            for index, item in enumerate(current)
+        ]
+
+    return current
+
+
 def _normalize_root_model_keys(config: Dict[str, Any]) -> Dict[str, Any]:
     """Move stale root-level provider/base_url into model section.
 
@@ -2677,7 +2797,6 @@ def read_raw_config() -> Dict[str, Any]:
 
 def load_config() -> Dict[str, Any]:
     """Load configuration from ~/.hermes/config.yaml."""
-    import copy
     ensure_hermes_home()
     config_path = get_config_path()
     
@@ -2698,8 +2817,11 @@ def load_config() -> Dict[str, Any]:
             config = _deep_merge(config, user_config)
         except Exception as e:
             print(f"Warning: Failed to load config: {e}")
-    
-    return _expand_env_vars(_normalize_root_model_keys(_normalize_max_turns_config(config)))
+
+    normalized = _normalize_root_model_keys(_normalize_max_turns_config(config))
+    expanded = _expand_env_vars(normalized)
+    _LAST_EXPANDED_CONFIG_BY_PATH[str(config_path)] = copy.deepcopy(expanded)
+    return expanded
 
 
 _SECURITY_COMMENT = """
@@ -2808,7 +2930,15 @@ def save_config(config: Dict[str, Any]):
 
     ensure_hermes_home()
     config_path = get_config_path()
-    normalized = _normalize_root_model_keys(_normalize_max_turns_config(config))
+    current_normalized = _normalize_root_model_keys(_normalize_max_turns_config(config))
+    normalized = current_normalized
+    raw_existing = _normalize_root_model_keys(_normalize_max_turns_config(read_raw_config()))
+    if raw_existing:
+        normalized = _preserve_env_ref_templates(
+            normalized,
+            raw_existing,
+            _LAST_EXPANDED_CONFIG_BY_PATH.get(str(config_path)),
+        )
 
     # Build optional commented-out sections for features that are off by
     # default or only relevant when explicitly configured.
@@ -2826,6 +2956,7 @@ def save_config(config: Dict[str, Any]):
         extra_content="".join(parts) if parts else None,
     )
     _secure_file(config_path)
+    _LAST_EXPANDED_CONFIG_BY_PATH[str(config_path)] = copy.deepcopy(current_normalized)
 
 
 def load_env() -> Dict[str, str]:

hermes_cli/debug.py

@@ -6,7 +6,10 @@ Currently supports:
 """
 
 import io
+import json
+import os
 import sys
+import time
 import urllib.error
 import urllib.parse
 import urllib.request
@@ -31,6 +34,119 @@ _MAX_LOG_BYTES = 512_000
 _AUTO_DELETE_SECONDS = 21600
 
 
+# ---------------------------------------------------------------------------
+# Pending-deletion tracking (replaces the old fork-and-sleep subprocess).
+# ---------------------------------------------------------------------------
+
+def _pending_file() -> Path:
+    """Path to ``~/.hermes/pastes/pending.json``.
+
+    Each entry: ``{"url": "...", "expire_at": <unix_ts>}``.  Scheduled
+    DELETEs used to be handled by spawning a detached Python process per
+    paste that slept for 6 hours; those accumulated forever if the user
+    ran ``hermes debug share`` repeatedly.  We now persist the schedule
+    to disk and sweep expired entries on the next debug invocation.
+    """
+    return get_hermes_home() / "pastes" / "pending.json"
+
+
+def _load_pending() -> list[dict]:
+    path = _pending_file()
+    if not path.exists():
+        return []
+    try:
+        data = json.loads(path.read_text(encoding="utf-8"))
+        if isinstance(data, list):
+            # Filter to well-formed entries only
+            return [
+                e for e in data
+                if isinstance(e, dict) and "url" in e and "expire_at" in e
+            ]
+    except (OSError, ValueError, json.JSONDecodeError):
+        pass
+    return []
+
+
+def _save_pending(entries: list[dict]) -> None:
+    path = _pending_file()
+    try:
+        path.parent.mkdir(parents=True, exist_ok=True)
+        tmp = path.with_suffix(".json.tmp")
+        tmp.write_text(json.dumps(entries, indent=2), encoding="utf-8")
+        os.replace(tmp, path)
+    except OSError:
+        # Non-fatal — worst case the user has to run ``hermes debug delete``
+        # manually.
+        pass
+
+
+def _record_pending(urls: list[str], delay_seconds: int = _AUTO_DELETE_SECONDS) -> None:
+    """Record *urls* for deletion at ``now + delay_seconds``.
+
+    Only paste.rs URLs are recorded (dpaste.com auto-expires).  Entries
+    are merged into any existing pending.json.
+    """
+    paste_rs_urls = [u for u in urls if _extract_paste_id(u)]
+    if not paste_rs_urls:
+        return
+
+    entries = _load_pending()
+    # Dedupe by URL: keep the later expire_at if same URL appears twice
+    by_url: dict[str, float] = {e["url"]: float(e["expire_at"]) for e in entries}
+    expire_at = time.time() + delay_seconds
+    for u in paste_rs_urls:
+        by_url[u] = max(expire_at, by_url.get(u, 0.0))
+    merged = [{"url": u, "expire_at": ts} for u, ts in by_url.items()]
+    _save_pending(merged)
+
+
+def _sweep_expired_pastes(now: Optional[float] = None) -> tuple[int, int]:
+    """Synchronously DELETE any pending pastes whose ``expire_at`` has passed.
+
+    Returns ``(deleted, remaining)``.  Best-effort: failed deletes stay in
+    the pending file and will be retried on the next sweep.  Silent —
+    intended to be called from every ``hermes debug`` invocation with
+    minimal noise.
+    """
+    entries = _load_pending()
+    if not entries:
+        return (0, 0)
+
+    current = time.time() if now is None else now
+    deleted = 0
+    remaining: list[dict] = []
+
+    for entry in entries:
+        try:
+            expire_at = float(entry.get("expire_at", 0))
+        except (TypeError, ValueError):
+            continue  # drop malformed entries
+        if expire_at > current:
+            remaining.append(entry)
+            continue
+
+        url = entry.get("url", "")
+        try:
+            if delete_paste(url):
+                deleted += 1
+                continue
+        except Exception:
+            # Network hiccup, 404 (already gone), etc. — drop the entry
+            # after a grace period; don't retry forever.
+            pass
+
+        # Retain failed deletes for up to 24h past expiration, then give up.
+        if expire_at + 86400 > current:
+            remaining.append(entry)
+        else:
+            deleted += 1  # count as reaped (paste.rs will GC eventually)
+
+    if deleted:
+        _save_pending(remaining)
+
+    return (deleted, len(remaining))
+
+
 # ---------------------------------------------------------------------------
 # Privacy / delete helpers
 # ---------------------------------------------------------------------------
@@ -90,37 +206,19 @@ def delete_paste(url: str) -> bool:
 
 
 def _schedule_auto_delete(urls: list[str], delay_seconds: int = _AUTO_DELETE_SECONDS):
-    """Spawn a detached process to delete paste.rs pastes after *delay_seconds*.
+    """Record *urls* for deletion ``delay_seconds`` from now.
 
-    The child process is fully detached (``start_new_session=True``) so it
-    survives the parent exiting (important for CLI mode).  Only paste.rs
-    URLs are attempted — dpaste.com pastes auto-expire on their own.
+    Previously this spawned a detached Python subprocess per call that slept
+    for 6 hours and then issued DELETE requests.  Those subprocesses leaked —
+    every ``hermes debug share`` invocation added ~20 MB of resident Python
+    interpreters that never exited until the sleep completed.
+
+    The replacement is stateless: we append to ``~/.hermes/pastes/pending.json``
+    and rely on opportunistic sweeps (``_sweep_expired_pastes``) called from
+    every ``hermes debug`` invocation.  If the user never runs ``hermes debug``
+    again, paste.rs's own retention policy handles cleanup.
     """
-    import subprocess
-
-    paste_rs_urls = [u for u in urls if _extract_paste_id(u)]
-    if not paste_rs_urls:
-        return
-
-    # Build a tiny inline Python script.  No imports beyond stdlib.
-    url_list = ", ".join(f'"{u}"' for u in paste_rs_urls)
-    script = (
-        "import time, urllib.request; "
-        f"time.sleep({delay_seconds}); "
-        f"[urllib.request.urlopen(urllib.request.Request(u, method='DELETE', "
-        f"headers={{'User-Agent': 'hermes-agent/auto-delete'}}), timeout=15) "
-        f"for u in [{url_list}]]"
-    )
-
-    try:
-        subprocess.Popen(
-            [sys.executable, "-c", script],
-            start_new_session=True,
-            stdout=subprocess.DEVNULL,
-            stderr=subprocess.DEVNULL,
-        )
-    except Exception:
-        pass  # Best-effort; manual delete still available.
+    _record_pending(urls, delay_seconds=delay_seconds)
 
 
 def _delete_hint(url: str) -> str:
@@ -455,6 +553,16 @@ def run_debug_delete(args):
 
 def run_debug(args):
     """Route debug subcommands."""
+    # Opportunistic sweep of expired pastes on every ``hermes debug`` call.
+    # Replaces the old per-paste sleeping subprocess that used to leak as
+    # one orphaned Python interpreter per scheduled deletion.  Silent and
+    # best-effort — any failure is swallowed so ``hermes debug`` stays
+    # reliable even when offline.
+    try:
+        _sweep_expired_pastes()
+    except Exception:
+        pass
+
     subcmd = getattr(args, "debug_command", None)
     if subcmd == "share":
         run_debug_share(args)

hermes_cli/doctor.py

@@ -825,6 +825,7 @@ def run_doctor(args):
         ("Arcee AI",         ("ARCEEAI_API_KEY",),                            "https://api.arcee.ai/api/v1/models",  "ARCEE_BASE_URL", True),
         ("DeepSeek",         ("DEEPSEEK_API_KEY",),                           "https://api.deepseek.com/v1/models",  "DEEPSEEK_BASE_URL", True),
         ("Hugging Face",     ("HF_TOKEN",),                                   "https://router.huggingface.co/v1/models", "HF_BASE_URL", True),
+        ("NVIDIA NIM",       ("NVIDIA_API_KEY",),                             "https://integrate.api.nvidia.com/v1/models", "NVIDIA_BASE_URL", True),
         ("Alibaba/DashScope", ("DASHSCOPE_API_KEY",),                         "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/models", "DASHSCOPE_BASE_URL", True),
         # MiniMax: the /anthropic endpoint doesn't support /models, but the /v1 endpoint does.
         ("MiniMax",          ("MINIMAX_API_KEY",),                            "https://api.minimax.io/v1/models",    "MINIMAX_BASE_URL", True),
@@ -894,8 +895,8 @@ def run_doctor(args):
                 _model_count = len(_br_resp.get("modelSummaries", []))
                 print(f"\r  {color('✓', Colors.GREEN)} {_label} {color(f'({_auth_var}, {_region}, {_model_count} models)', Colors.DIM)}           ")
             except ImportError:
-                print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color('(boto3 not installed — pip install hermes-agent[bedrock])', Colors.DIM)}           ")
-                issues.append("Install boto3 for Bedrock: pip install hermes-agent[bedrock]")
+                print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color(f'(boto3 not installed — {sys.executable} -m pip install boto3)', Colors.DIM)}           ")
+                issues.append(f"Install boto3 for Bedrock: {sys.executable} -m pip install boto3")
             except Exception as _e:
                 _err_name = type(_e).__name__
                 print(f"\r  {color('⚠', Colors.YELLOW)} {_label} {color(f'({_err_name}: {_e})', Colors.DIM)}           ")

hermes_cli/dump.py

@@ -43,41 +43,20 @@ def _redact(value: str) -> str:
 
 def _gateway_status() -> str:
     """Return a short gateway status string."""
-    if sys.platform.startswith("linux"):
-        from hermes_constants import is_container
-        if is_container():
-            try:
-                from hermes_cli.gateway import find_gateway_pids
-                pids = find_gateway_pids()
-                if pids:
-                    return f"running (docker, pid {pids[0]})"
-                return "stopped (docker)"
-            except Exception:
-                return "stopped (docker)"
-        try:
-            from hermes_cli.gateway import get_service_name
-            svc = get_service_name()
-        except Exception:
-            svc = "hermes-gateway"
-        try:
-            r = subprocess.run(
-                ["systemctl", "--user", "is-active", svc],
-                capture_output=True, text=True, timeout=5,
-            )
-            return "running (systemd)" if r.stdout.strip() == "active" else "stopped"
-        except Exception:
-            return "unknown"
-    elif sys.platform == "darwin":
-        try:
-            from hermes_cli.gateway import get_launchd_label
-            r = subprocess.run(
-                ["launchctl", "list", get_launchd_label()],
-                capture_output=True, text=True, timeout=5,
-            )
-            return "loaded (launchd)" if r.returncode == 0 else "not loaded"
-        except Exception:
-            return "unknown"
-    return "N/A"
+    try:
+        from hermes_cli.gateway import get_gateway_runtime_snapshot
+
+        snapshot = get_gateway_runtime_snapshot()
+        if snapshot.running:
+            mode = snapshot.manager
+            if snapshot.has_process_service_mismatch:
+                mode = "manual"
+            return f"running ({mode}, pid {snapshot.gateway_pids[0]})"
+        if snapshot.service_installed and not snapshot.service_running:
+            return f"stopped ({snapshot.manager})"
+        return f"stopped ({snapshot.manager})"
+    except Exception:
+        return "unknown" if sys.platform.startswith(("linux", "darwin")) else "N/A"
 
 
 def _count_skills(hermes_home: Path) -> int:
@@ -296,6 +275,7 @@ def run_dump(args):
         ("DEEPSEEK_API_KEY", "deepseek"),
         ("DASHSCOPE_API_KEY", "dashscope"),
         ("HF_TOKEN", "huggingface"),
+        ("NVIDIA_API_KEY", "nvidia"),
         ("AI_GATEWAY_API_KEY", "ai_gateway"),
         ("OPENCODE_ZEN_API_KEY", "opencode_zen"),
         ("OPENCODE_GO_API_KEY", "opencode_go"),

hermes_cli/gateway.py

@@ -10,6 +10,7 @@ import shutil
 import signal
 import subprocess
 import sys
+from dataclasses import dataclass
 from pathlib import Path
 
 PROJECT_ROOT = Path(__file__).parent.parent.resolve()
@@ -41,6 +42,23 @@ from hermes_cli.colors import Colors, color
 # Process Management (for manual gateway runs)
 # =============================================================================
 
+
+@dataclass(frozen=True)
+class GatewayRuntimeSnapshot:
+    manager: str
+    service_installed: bool = False
+    service_running: bool = False
+    gateway_pids: tuple[int, ...] = ()
+    service_scope: str | None = None
+
+    @property
+    def running(self) -> bool:
+        return self.service_running or bool(self.gateway_pids)
+
+    @property
+    def has_process_service_mismatch(self) -> bool:
+        return self.service_installed and self.running and not self.service_running
+
 def _get_service_pids() -> set:
     """Return PIDs currently managed by systemd or launchd gateway services.
 
@@ -157,20 +175,22 @@ def _request_gateway_self_restart(pid: int) -> bool:
     return True
 
 
-def find_gateway_pids(exclude_pids: set | None = None, all_profiles: bool = False) -> list:
-    """Find PIDs of running gateway processes.
+def _append_unique_pid(pids: list[int], pid: int | None, exclude_pids: set[int]) -> None:
+    if pid is None or pid <= 0:
+        return
+    if pid == os.getpid() or pid in exclude_pids or pid in pids:
+        return
+    pids.append(pid)
 
-    Args:
-        exclude_pids: PIDs to exclude from the result (e.g. service-managed
-            PIDs that should not be killed during a stale-process sweep).
-        all_profiles: When ``True``, return gateway PIDs across **all**
-            profiles (the pre-7923 global behaviour).  ``hermes update``
-            needs this because a code update affects every profile.
-            When ``False`` (default), only PIDs belonging to the current
-            Hermes profile are returned.
+
+def _scan_gateway_pids(exclude_pids: set[int], all_profiles: bool = False) -> list[int]:
+    """Best-effort process-table scan for gateway PIDs.
+
+    This supplements the profile-scoped PID file so status views can still spot
+    a live gateway when the PID file is stale/missing, and ``--all`` sweeps can
+    discover gateways outside the current profile.
     """
-    _exclude = exclude_pids or set()
-    pids = [pid for pid in _get_service_pids() if pid not in _exclude]
+    pids: list[int] = []
     patterns = [
         "hermes_cli.main gateway",
         "hermes_cli.main --profile",
@@ -203,20 +223,24 @@ def find_gateway_pids(exclude_pids: set | None = None, all_profiles: bool = Fals
         if is_windows():
             result = subprocess.run(
                 ["wmic", "process", "get", "ProcessId,CommandLine", "/FORMAT:LIST"],
-                capture_output=True, text=True, timeout=10
+                capture_output=True,
+                text=True,
+                timeout=10,
             )
+            if result.returncode != 0:
+                return []
             current_cmd = ""
-            for line in result.stdout.split('\n'):
+            for line in result.stdout.split("\n"):
                 line = line.strip()
                 if line.startswith("CommandLine="):
                     current_cmd = line[len("CommandLine="):]
                 elif line.startswith("ProcessId="):
                     pid_str = line[len("ProcessId="):]
-                    if any(p in current_cmd for p in patterns) and (all_profiles or _matches_current_profile(current_cmd)):
+                    if any(p in current_cmd for p in patterns) and (
+                        all_profiles or _matches_current_profile(current_cmd)
+                    ):
                         try:
-                            pid = int(pid_str)
-                            if pid != os.getpid() and pid not in pids and pid not in _exclude:
-                                pids.append(pid)
+                            _append_unique_pid(pids, int(pid_str), exclude_pids)
                         except ValueError:
                             pass
                     current_cmd = ""
@@ -227,9 +251,11 @@ def find_gateway_pids(exclude_pids: set | None = None, all_profiles: bool = Fals
                 text=True,
                 timeout=10,
             )
-            for line in result.stdout.split('\n'):
+            if result.returncode != 0:
+                return []
+            for line in result.stdout.split("\n"):
                 stripped = line.strip()
-                if not stripped or 'grep' in stripped:
+                if not stripped or "grep" in stripped:
                     continue
 
                 pid = None
@@ -251,16 +277,137 @@ def find_gateway_pids(exclude_pids: set | None = None, all_profiles: bool = Fals
 
                 if pid is None:
                     continue
-                if pid == os.getpid() or pid in pids or pid in _exclude:
-                    continue
-                if any(pattern in command for pattern in patterns) and (all_profiles or _matches_current_profile(command)):
-                    pids.append(pid)
+                if any(pattern in command for pattern in patterns) and (
+                    all_profiles or _matches_current_profile(command)
+                ):
+                    _append_unique_pid(pids, pid, exclude_pids)
     except (OSError, subprocess.TimeoutExpired):
-        pass
+        return []
 
     return pids
 
 
+def find_gateway_pids(exclude_pids: set | None = None, all_profiles: bool = False) -> list:
+    """Find PIDs of running gateway processes.
+
+    Args:
+        exclude_pids: PIDs to exclude from the result (e.g. service-managed
+            PIDs that should not be killed during a stale-process sweep).
+        all_profiles: When ``True``, return gateway PIDs across **all**
+            profiles (the pre-7923 global behaviour).  ``hermes update``
+            needs this because a code update affects every profile.
+            When ``False`` (default), only PIDs belonging to the current
+            Hermes profile are returned.
+    """
+    _exclude = set(exclude_pids or set())
+    pids: list[int] = []
+    if not all_profiles:
+        try:
+            from gateway.status import get_running_pid
+
+            _append_unique_pid(pids, get_running_pid(), _exclude)
+        except Exception:
+            pass
+    for pid in _get_service_pids():
+        _append_unique_pid(pids, pid, _exclude)
+    for pid in _scan_gateway_pids(_exclude, all_profiles=all_profiles):
+        _append_unique_pid(pids, pid, _exclude)
+    return pids
+
+
+def _probe_systemd_service_running(system: bool = False) -> tuple[bool, bool]:
+    selected_system = _select_systemd_scope(system)
+    unit_exists = get_systemd_unit_path(system=selected_system).exists()
+    if not unit_exists:
+        return selected_system, False
+    try:
+        result = _run_systemctl(
+            ["is-active", get_service_name()],
+            system=selected_system,
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+    except (RuntimeError, subprocess.TimeoutExpired):
+        return selected_system, False
+    return selected_system, result.stdout.strip() == "active"
+
+
+def _probe_launchd_service_running() -> bool:
+    if not get_launchd_plist_path().exists():
+        return False
+    try:
+        result = subprocess.run(
+            ["launchctl", "list", get_launchd_label()],
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+    except subprocess.TimeoutExpired:
+        return False
+    return result.returncode == 0
+
+
+def get_gateway_runtime_snapshot(system: bool = False) -> GatewayRuntimeSnapshot:
+    """Return a unified view of gateway liveness for the current profile."""
+    gateway_pids = tuple(find_gateway_pids())
+    if is_termux():
+        return GatewayRuntimeSnapshot(
+            manager="Termux / manual process",
+            gateway_pids=gateway_pids,
+        )
+
+    from hermes_constants import is_container
+
+    if is_linux() and is_container():
+        return GatewayRuntimeSnapshot(
+            manager="docker (foreground)",
+            gateway_pids=gateway_pids,
+        )
+
+    if supports_systemd_services():
+        selected_system, service_running = _probe_systemd_service_running(system=system)
+        scope_label = _service_scope_label(selected_system)
+        return GatewayRuntimeSnapshot(
+            manager=f"systemd ({scope_label})",
+            service_installed=get_systemd_unit_path(system=selected_system).exists(),
+            service_running=service_running,
+            gateway_pids=gateway_pids,
+            service_scope=scope_label,
+        )
+
+    if is_macos():
+        return GatewayRuntimeSnapshot(
+            manager="launchd",
+            service_installed=get_launchd_plist_path().exists(),
+            service_running=_probe_launchd_service_running(),
+            gateway_pids=gateway_pids,
+            service_scope="launchd",
+        )
+
+    return GatewayRuntimeSnapshot(
+        manager="manual process",
+        gateway_pids=gateway_pids,
+    )
+
+
+def _format_gateway_pids(pids: tuple[int, ...] | list[int], *, limit: int | None = 3) -> str:
+    rendered = [str(pid) for pid in pids[:limit] if pid > 0] if limit is not None else [str(pid) for pid in pids if pid > 0]
+    if limit is not None and len(pids) > limit:
+        rendered.append("...")
+    return ", ".join(rendered)
+
+
+def _print_gateway_process_mismatch(snapshot: GatewayRuntimeSnapshot) -> None:
+    if not snapshot.has_process_service_mismatch:
+        return
+    print()
+    print("⚠ Gateway process is running for this profile, but the service is not active")
+    print(f"  PID(s): {_format_gateway_pids(snapshot.gateway_pids, limit=None)}")
+    print("  This is usually a manual foreground/tmux/nohup run, so `hermes gateway`")
+    print("  can refuse to start another copy until this process stops.")
+
+
 def kill_gateway_processes(force: bool = False, exclude_pids: set | None = None,
                            all_profiles: bool = False) -> int:
     """Kill any running gateway processes. Returns count killed.
@@ -340,25 +487,44 @@ def _wsl_systemd_operational() -> bool:
     WSL2 with ``systemd=true`` in wsl.conf has working systemd.
     WSL2 without it (or WSL1) does not — systemctl commands fail.
     """
+    return _systemd_operational(system=True)
+
+
+def _systemd_operational(system: bool = False) -> bool:
+    """Return True when the requested systemd scope is usable."""
     try:
-        result = subprocess.run(
-            ["systemctl", "is-system-running"],
-            capture_output=True, text=True, timeout=5,
+        result = _run_systemctl(
+            ["is-system-running"],
+            system=system,
+            capture_output=True,
+            text=True,
+            timeout=5,
         )
         # "running", "degraded", "starting" all mean systemd is PID 1
         status = result.stdout.strip().lower()
         return status in ("running", "degraded", "starting", "initializing")
-    except (FileNotFoundError, subprocess.TimeoutExpired, OSError):
+    except (RuntimeError, subprocess.TimeoutExpired, OSError):
         return False
 
 
+def _container_systemd_operational() -> bool:
+    """Return True when a container exposes working user or system systemd."""
+    if _systemd_operational(system=False):
+        return True
+    if _systemd_operational(system=True):
+        return True
+    return False
+
+
 def supports_systemd_services() -> bool:
-    if not is_linux() or is_termux() or is_container():
+    if not is_linux() or is_termux():
         return False
     if shutil.which("systemctl") is None:
         return False
     if is_wsl():
         return _wsl_systemd_operational()
+    if is_container():
+        return _container_systemd_operational()
     return True
 
 
@@ -521,6 +687,195 @@ def has_conflicting_systemd_units() -> bool:
     return len(get_installed_systemd_scopes()) > 1
 
 
+# Legacy service names from older Hermes installs that predate the
+# hermes-gateway rename. Kept as an explicit allowlist (NOT a glob) so
+# profile units (hermes-gateway-*.service) and unrelated third-party
+# "hermes" units are never matched.
+_LEGACY_SERVICE_NAMES: tuple[str, ...] = ("hermes.service",)
+
+# ExecStart content markers that identify a unit as running our gateway.
+# A legacy unit is only flagged when its file contains one of these.
+_LEGACY_UNIT_EXECSTART_MARKERS: tuple[str, ...] = (
+    "hermes_cli.main gateway",
+    "hermes_cli/main.py gateway",
+    "gateway/run.py",
+    " hermes gateway ",
+    "/hermes gateway ",
+)
+
+
+def _legacy_unit_search_paths() -> list[tuple[bool, Path]]:
+    """Return ``[(is_system, base_dir), ...]`` — directories to scan for legacy units.
+
+    Factored out so tests can monkeypatch the search roots without touching
+    real filesystem paths.
+    """
+    return [
+        (False, Path.home() / ".config" / "systemd" / "user"),
+        (True, Path("/etc/systemd/system")),
+    ]
+
+
+def _find_legacy_hermes_units() -> list[tuple[str, Path, bool]]:
+    """Return ``[(unit_name, unit_path, is_system)]`` for legacy Hermes gateway units.
+
+    Detects unit files installed by older Hermes versions that used a
+    different service name (e.g. ``hermes.service`` before the rename to
+    ``hermes-gateway.service``). When both a legacy unit and the current
+    ``hermes-gateway.service`` are active, they fight over the same bot
+    token — the PR #5646 signal-recovery change turns this into a 30-second
+    SIGTERM flap loop.
+
+    Safety guards:
+
+    * Explicit allowlist of legacy names (no globbing). Profile units such
+      as ``hermes-gateway-coder.service`` and unrelated third-party
+      ``hermes-*`` services are never matched.
+    * ExecStart content check — only flag units that invoke our gateway
+      entrypoint. A user-created ``hermes.service`` running an unrelated
+      binary is left untouched.
+    * Results are returned purely for caller inspection; this function
+      never mutates or removes anything.
+    """
+    results: list[tuple[str, Path, bool]] = []
+    for is_system, base in _legacy_unit_search_paths():
+        for name in _LEGACY_SERVICE_NAMES:
+            unit_path = base / name
+            try:
+                if not unit_path.exists():
+                    continue
+                text = unit_path.read_text(encoding="utf-8", errors="ignore")
+            except (OSError, PermissionError):
+                continue
+            if not any(marker in text for marker in _LEGACY_UNIT_EXECSTART_MARKERS):
+                # Not our gateway — leave alone
+                continue
+            results.append((name, unit_path, is_system))
+    return results
+
+
+def has_legacy_hermes_units() -> bool:
+    """Return True when any legacy Hermes gateway unit files exist."""
+    return bool(_find_legacy_hermes_units())
+
+
+def print_legacy_unit_warning() -> None:
+    """Warn about legacy Hermes gateway unit files if any are installed.
+
+    Idempotent: prints nothing when no legacy units are detected. Safe to
+    call from any status/install/setup path.
+    """
+    legacy = _find_legacy_hermes_units()
+    if not legacy:
+        return
+    print_warning("Legacy Hermes gateway unit(s) detected from an older install:")
+    for name, path, is_system in legacy:
+        scope = "system" if is_system else "user"
+        print_info(f"    {path}  ({scope} scope)")
+    print_info("  These run alongside the current hermes-gateway service and")
+    print_info("  cause SIGTERM flap loops — both try to use the same bot token.")
+    print_info("  Remove them with:")
+    print_info("    hermes gateway migrate-legacy")
+
+
+def remove_legacy_hermes_units(
+    interactive: bool = True,
+    dry_run: bool = False,
+) -> tuple[int, list[Path]]:
+    """Stop, disable, and remove legacy Hermes gateway unit files.
+
+    Iterates over whatever ``_find_legacy_hermes_units()`` returns — which is
+    an explicit allowlist of legacy names (not a glob). Profile units and
+    unrelated third-party services are never touched.
+
+    Args:
+        interactive: When True, prompt before removing. When False, remove
+            without asking (used when another prompt has already confirmed,
+            e.g. from the install flow).
+        dry_run: When True, list what would be removed and return.
+
+    Returns:
+        ``(removed_count, remaining_paths)`` — remaining includes units we
+        couldn't remove (typically system-scope when not running as root).
+    """
+    legacy = _find_legacy_hermes_units()
+    if not legacy:
+        print("No legacy Hermes gateway units found.")
+        return 0, []
+
+    user_units = [(n, p) for n, p, is_sys in legacy if not is_sys]
+    system_units = [(n, p) for n, p, is_sys in legacy if is_sys]
+
+    print()
+    print("Legacy Hermes gateway unit(s) found:")
+    for name, path, is_system in legacy:
+        scope = "system" if is_system else "user"
+        print(f"  {path}  ({scope} scope)")
+    print()
+
+    if dry_run:
+        print("(dry-run — nothing removed)")
+        return 0, [p for _, p, _ in legacy]
+
+    if interactive and not prompt_yes_no("Remove these legacy units?", True):
+        print("Skipped. Run again with: hermes gateway migrate-legacy")
+        return 0, [p for _, p, _ in legacy]
+
+    removed = 0
+    remaining: list[Path] = []
+
+    # User-scope removal
+    for name, path in user_units:
+        try:
+            _run_systemctl(["stop", name], system=False, check=False, timeout=90)
+            _run_systemctl(["disable", name], system=False, check=False, timeout=30)
+            path.unlink(missing_ok=True)
+            print(f"  ✓ Removed {path}")
+            removed += 1
+        except (OSError, RuntimeError) as e:
+            print(f"  ⚠ Could not remove {path}: {e}")
+            remaining.append(path)
+
+    if user_units:
+        try:
+            _run_systemctl(["daemon-reload"], system=False, check=False, timeout=30)
+        except RuntimeError:
+            pass
+
+    # System-scope removal (needs root)
+    if system_units:
+        if os.geteuid() != 0:
+            print()
+            print_warning("System-scope legacy units require root to remove.")
+            print_info("  Re-run with: sudo hermes gateway migrate-legacy")
+            for _, path in system_units:
+                remaining.append(path)
+        else:
+            for name, path in system_units:
+                try:
+                    _run_systemctl(["stop", name], system=True, check=False, timeout=90)
+                    _run_systemctl(["disable", name], system=True, check=False, timeout=30)
+                    path.unlink(missing_ok=True)
+                    print(f"  ✓ Removed {path}")
+                    removed += 1
+                except (OSError, RuntimeError) as e:
+                    print(f"  ⚠ Could not remove {path}: {e}")
+                    remaining.append(path)
+
+            try:
+                _run_systemctl(["daemon-reload"], system=True, check=False, timeout=30)
+            except RuntimeError:
+                pass
+
+    print()
+    if remaining:
+        print_warning(f"{len(remaining)} legacy unit(s) still present — see messages above.")
+    else:
+        print_success(f"Removed {removed} legacy unit(s).")
+
+    return removed, remaining
+
+
 def print_systemd_scope_conflict_warning() -> None:
     scopes = get_installed_systemd_scopes()
     if len(scopes) < 2:
@@ -1054,6 +1409,19 @@ def systemd_install(force: bool = False, system: bool = False, run_as_user: str
     if system:
         _require_root_for_system_service("install")
 
+    # Offer to remove legacy units (hermes.service from pre-rename installs)
+    # before installing the new hermes-gateway.service. If both remain, they
+    # flap-fight for the Telegram bot token on every gateway startup.
+    # Only removes units matching _LEGACY_SERVICE_NAMES + our ExecStart
+    # signature — profile units are never touched.
+    if has_legacy_hermes_units():
+        print()
+        print_legacy_unit_warning()
+        print()
+        if prompt_yes_no("Remove the legacy unit(s) before installing?", True):
+            remove_legacy_hermes_units(interactive=False)
+            print()
+
     unit_path = get_systemd_unit_path(system=system)
     scope_flag = " --system" if system else ""
 
@@ -1092,6 +1460,7 @@ def systemd_install(force: bool = False, system: bool = False, run_as_user: str
         _ensure_linger_enabled()
 
     print_systemd_scope_conflict_warning()
+    print_legacy_unit_warning()
 
 
 def systemd_uninstall(system: bool = False):
@@ -1215,6 +1584,10 @@ def systemd_status(deep: bool = False, system: bool = False):
         print_systemd_scope_conflict_warning()
         print()
 
+    if has_legacy_hermes_units():
+        print_legacy_unit_warning()
+        print()
+
     if not systemd_unit_is_current(system=system):
         print("⚠ Installed gateway service definition is outdated")
         print(f"  Run: {'sudo ' if system else ''}hermes gateway restart{scope_flag}  # auto-refreshes the unit")
@@ -1998,7 +2371,7 @@ _PLATFORMS = [
             {"name": "QQ_ALLOWED_USERS", "prompt": "Allowed user OpenIDs (comma-separated, leave empty for open access)", "password": False,
              "is_allowlist": True,
              "help": "Optional — restrict DM access to specific user OpenIDs."},
-            {"name": "QQ_HOME_CHANNEL", "prompt": "Home channel (user/group OpenID for cron delivery, or empty)", "password": False,
+            {"name": "QQBOT_HOME_CHANNEL", "prompt": "Home channel (user/group OpenID for cron delivery, or empty)", "password": False,
              "help": "OpenID to deliver cron results and notifications to."},
         ],
     },
@@ -2625,6 +2998,215 @@ def _setup_feishu():
         print_info(f"  Bot: {bot_name}")
 
 
+def _setup_qqbot():
+    """Interactive setup for QQ Bot — scan-to-configure or manual credentials."""
+    print()
+    print(color("  ─── 🐧 QQ Bot Setup ───", Colors.CYAN))
+
+    existing_app_id = get_env_value("QQ_APP_ID")
+    existing_secret = get_env_value("QQ_CLIENT_SECRET")
+    if existing_app_id and existing_secret:
+        print()
+        print_success("QQ Bot is already configured.")
+        if not prompt_yes_no("  Reconfigure QQ Bot?", False):
+            return
+
+    # ── Choose setup method ──
+    print()
+    method_choices = [
+        "Scan QR code to add bot automatically (recommended)",
+        "Enter existing App ID and App Secret manually",
+    ]
+    method_idx = prompt_choice("  How would you like to set up QQ Bot?", method_choices, 0)
+
+    credentials = None
+    used_qr = False
+
+    if method_idx == 0:
+        # ── QR scan-to-configure ──
+        try:
+            credentials = _qqbot_qr_flow()
+        except KeyboardInterrupt:
+            print()
+            print_warning("  QQ Bot setup cancelled.")
+            return
+        if credentials:
+            used_qr = True
+        if not credentials:
+            print_info("  QR setup did not complete. Continuing with manual input.")
+
+    # ── Manual credential input ──
+    if not credentials:
+        print()
+        print_info("  Go to https://q.qq.com to register a QQ Bot application.")
+        print_info("  Note your App ID and App Secret from the application page.")
+        print()
+        app_id = prompt("  App ID", password=False)
+        if not app_id:
+            print_warning("  Skipped — QQ Bot won't work without an App ID.")
+            return
+        app_secret = prompt("  App Secret", password=True)
+        if not app_secret:
+            print_warning("  Skipped — QQ Bot won't work without an App Secret.")
+            return
+        credentials = {"app_id": app_id.strip(), "client_secret": app_secret.strip(), "user_openid": ""}
+
+    # ── Save core credentials ──
+    save_env_value("QQ_APP_ID", credentials["app_id"])
+    save_env_value("QQ_CLIENT_SECRET", credentials["client_secret"])
+
+    user_openid = credentials.get("user_openid", "")
+
+    # ── DM security policy ──
+    print()
+    access_choices = [
+        "Use DM pairing approval (recommended)",
+        "Allow all direct messages",
+        "Only allow listed user OpenIDs",
+    ]
+    access_idx = prompt_choice("  How should direct messages be authorized?", access_choices, 0)
+    if access_idx == 0:
+        save_env_value("QQ_ALLOW_ALL_USERS", "false")
+        if user_openid:
+            print()
+            if prompt_yes_no(f"  Add yourself ({user_openid}) to the allow list?", True):
+                save_env_value("QQ_ALLOWED_USERS", user_openid)
+                print_success(f"  Allow list set to {user_openid}")
+            else:
+                save_env_value("QQ_ALLOWED_USERS", "")
+        else:
+            save_env_value("QQ_ALLOWED_USERS", "")
+        print_success("  DM pairing enabled.")
+        print_info("  Unknown users can request access; approve with `hermes pairing approve`.")
+    elif access_idx == 1:
+        save_env_value("QQ_ALLOW_ALL_USERS", "true")
+        save_env_value("QQ_ALLOWED_USERS", "")
+        print_warning("  Open DM access enabled for QQ Bot.")
+    else:
+        default_allow = user_openid or ""
+        allowlist = prompt("  Allowed user OpenIDs (comma-separated)", default_allow, password=False).replace(" ", "")
+        save_env_value("QQ_ALLOW_ALL_USERS", "false")
+        save_env_value("QQ_ALLOWED_USERS", allowlist)
+        print_success("  Allowlist saved.")
+
+    # ── Home channel ──
+    if user_openid:
+        print()
+        if prompt_yes_no(f"  Use your QQ user ID ({user_openid}) as the home channel?", True):
+            save_env_value("QQBOT_HOME_CHANNEL", user_openid)
+            print_success(f"  Home channel set to {user_openid}")
+    else:
+        print()
+        home_channel = prompt("  Home channel OpenID (for cron/notifications, or empty)", password=False)
+        if home_channel:
+            save_env_value("QQBOT_HOME_CHANNEL", home_channel.strip())
+            print_success(f"  Home channel set to {home_channel.strip()}")
+
+    print()
+    print_success("🐧 QQ Bot configured!")
+    print_info(f"  App ID: {credentials['app_id']}")
+
+
+def _qqbot_render_qr(url: str) -> bool:
+    """Try to render a QR code in the terminal. Returns True if successful."""
+    try:
+        import qrcode as _qr
+        qr = _qr.QRCode(border=1,error_correction=_qr.constants.ERROR_CORRECT_L)
+        qr.add_data(url)
+        qr.make(fit=True)
+        qr.print_ascii(invert=True)
+        return True
+    except Exception:
+        return False
+
+
+def _qqbot_qr_flow():
+    """Run the QR-code scan-to-configure flow.
+
+    Returns a dict with app_id, client_secret, user_openid on success,
+    or None on failure/cancel.
+    """
+    try:
+        from gateway.platforms.qqbot import (
+            create_bind_task, poll_bind_result, build_connect_url,
+            decrypt_secret, BindStatus,
+        )
+        from gateway.platforms.qqbot.constants import ONBOARD_POLL_INTERVAL
+    except Exception as exc:
+        print_error(f"  QQBot onboard import failed: {exc}")
+        return None
+
+    import asyncio
+    import time
+
+    MAX_REFRESHES = 3
+    refresh_count = 0
+
+    while refresh_count <= MAX_REFRESHES:
+        loop = asyncio.new_event_loop()
+
+        # ── Create bind task ──
+        try:
+            task_id, aes_key = loop.run_until_complete(create_bind_task())
+        except Exception as e:
+            print_warning(f"  Failed to create bind task: {e}")
+            loop.close()
+            return None
+
+        url = build_connect_url(task_id)
+
+        # ── Display QR code + URL ──
+        print()
+        if _qqbot_render_qr(url):
+            print(f"  Scan the QR code above, or open this URL directly:\n  {url}")
+        else:
+            print(f"  Open this URL in QQ on your phone:\n  {url}")
+            print_info("  Tip: pip install qrcode  to show a scannable QR code here")
+
+        # ── Poll loop (silent — keep QR visible at bottom) ──
+        try:
+            while True:
+                try:
+                    status, app_id, encrypted_secret, user_openid = loop.run_until_complete(
+                        poll_bind_result(task_id)
+                    )
+                except Exception:
+                    time.sleep(ONBOARD_POLL_INTERVAL)
+                    continue
+
+                if status == BindStatus.COMPLETED:
+                    client_secret = decrypt_secret(encrypted_secret, aes_key)
+                    print()
+                    print_success(f"  QR scan complete! (App ID: {app_id})")
+                    if user_openid:
+                        print_info(f"  Scanner's OpenID: {user_openid}")
+                    return {
+                        "app_id": app_id,
+                        "client_secret": client_secret,
+                        "user_openid": user_openid,
+                    }
+
+                if status == BindStatus.EXPIRED:
+                    refresh_count += 1
+                    if refresh_count > MAX_REFRESHES:
+                        print()
+                        print_warning(f"  QR code expired {MAX_REFRESHES} times — giving up.")
+                        return None
+                    print()
+                    print_warning(f"  QR code expired, refreshing... ({refresh_count}/{MAX_REFRESHES})")
+                    loop.close()
+                    break  # outer while creates a new task
+
+                time.sleep(ONBOARD_POLL_INTERVAL)
+        except KeyboardInterrupt:
+            loop.close()
+            raise
+        finally:
+            loop.close()
+
+    return None
+
+
 def _setup_signal():
     """Interactive setup for Signal messenger."""
     import shutil
@@ -2762,6 +3344,10 @@ def gateway_setup():
         print_systemd_scope_conflict_warning()
         print()
 
+    if supports_systemd_services() and has_legacy_hermes_units():
+        print_legacy_unit_warning()
+        print()
+
     if service_installed and service_running:
         print_success("Gateway service is installed and running.")
     elif service_installed:
@@ -2806,6 +3392,8 @@ def gateway_setup():
             _setup_dingtalk()
         elif platform["key"] == "feishu":
             _setup_feishu()
+        elif platform["key"] == "qqbot":
+            _setup_qqbot()
         else:
             _setup_standard_platform(platform)
 
@@ -3165,15 +3753,18 @@ def gateway_command(args):
     elif subcmd == "status":
         deep = getattr(args, 'deep', False)
         system = getattr(args, 'system', False)
+        snapshot = get_gateway_runtime_snapshot(system=system)
         
         # Check for service first
         if supports_systemd_services() and (get_systemd_unit_path(system=False).exists() or get_systemd_unit_path(system=True).exists()):
             systemd_status(deep, system=system)
+            _print_gateway_process_mismatch(snapshot)
         elif is_macos() and get_launchd_plist_path().exists():
             launchd_status(deep)
+            _print_gateway_process_mismatch(snapshot)
         else:
             # Check for manually running processes
-            pids = find_gateway_pids()
+            pids = list(snapshot.gateway_pids)
             if pids:
                 print(f"✓ Gateway is running (PID: {', '.join(map(str, pids))})")
                 print("  (Running manually, not as a system service)")
@@ -3214,3 +3805,14 @@ def gateway_command(args):
                 else:
                     print("  hermes gateway install  # Install as user service")
                     print("  sudo hermes gateway install --system  # Install as boot-time system service")
+
+    elif subcmd == "migrate-legacy":
+        # Stop, disable, and remove legacy Hermes gateway unit files from
+        # pre-rename installs (e.g. hermes.service). Profile units and
+        # unrelated third-party services are never touched.
+        dry_run = getattr(args, 'dry_run', False)
+        yes = getattr(args, 'yes', False)
+        if not supports_systemd_services() and not is_macos():
+            print("Legacy unit migration only applies to systemd-based Linux hosts.")
+            return
+        remove_legacy_hermes_units(interactive=not yes, dry_run=dry_run)

hermes_cli/model_switch.py

@@ -692,12 +692,12 @@ def switch_model(
             api_key=api_key,
             base_url=base_url,
         )
-    except Exception:
+    except Exception as e:
         validation = {
-            "accepted": True,
-            "persist": True,
+            "accepted": False,
+            "persist": False,
             "recognized": False,
-            "message": None,
+            "message": f"Could not validate `{new_model}`: {e}",
         }
 
     if not validation.get("accepted"):

hermes_cli/models.py

@@ -26,7 +26,8 @@ COPILOT_REASONING_EFFORTS_O_SERIES = ["low", "medium", "high"]
 # Fallback OpenRouter snapshot used when the live catalog is unavailable.
 # (model_id, display description shown in menus)
 OPENROUTER_MODELS: list[tuple[str, str]] = [
-    ("anthropic/claude-opus-4.7",       "recommended"),
+    ("moonshotai/kimi-k2.5",            "recommended"),
+    ("anthropic/claude-opus-4.7",       ""),
     ("anthropic/claude-opus-4.6",       ""),
     ("anthropic/claude-sonnet-4.6",     ""),
     ("qwen/qwen3.6-plus",               ""),
@@ -49,7 +50,6 @@ OPENROUTER_MODELS: list[tuple[str, str]] = [
     ("z-ai/glm-5.1",                    ""),
     ("z-ai/glm-5v-turbo",               ""),
     ("z-ai/glm-5-turbo",                ""),
-    ("moonshotai/kimi-k2.5",            ""),
     ("x-ai/grok-4.20",                  ""),
     ("nvidia/nemotron-3-super-120b-a12b",      ""),
     ("nvidia/nemotron-3-super-120b-a12b:free", "free"),
@@ -75,6 +75,7 @@ def _codex_curated_models() -> list[str]:
 
 _PROVIDER_MODELS: dict[str, list[str]] = {
     "nous": [
+        "moonshotai/kimi-k2.5",
         "xiaomi/mimo-v2-pro",
         "anthropic/claude-opus-4.7",
         "anthropic/claude-opus-4.6",
@@ -96,7 +97,6 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "z-ai/glm-5.1",
         "z-ai/glm-5v-turbo",
         "z-ai/glm-5-turbo",
-        "moonshotai/kimi-k2.5",
         "x-ai/grok-4.20-beta",
         "nvidia/nemotron-3-super-120b-a12b",
         "nvidia/nemotron-3-super-120b-a12b:free",
@@ -135,7 +135,6 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "gemini-2.5-flash-lite",
         # Gemma open models (also served via AI Studio)
         "gemma-4-31b-it",
-        "gemma-4-26b-it",
     ],
     "google-gemini-cli": [
         "gemini-2.5-pro",
@@ -155,9 +154,23 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "grok-4.20-reasoning",
         "grok-4-1-fast-reasoning",
     ],
+    "nvidia": [
+        # NVIDIA flagship reasoning models
+        "nvidia/nemotron-3-super-120b-a12b",
+        "nvidia/nemotron-3-nano-30b-a3b",
+        "nvidia/llama-3.3-nemotron-super-49b-v1.5",
+        # Third-party agentic models hosted on build.nvidia.com
+        # (map to OpenRouter defaults — users get familiar picks on NIM)
+        "qwen/qwen3.5-397b-a17b",
+        "deepseek-ai/deepseek-v3.2",
+        "moonshotai/kimi-k2.5",
+        "minimaxai/minimax-m2.5",
+        "z-ai/glm5",
+        "openai/gpt-oss-120b",
+    ],
     "kimi-coding": [
-        "kimi-for-coding",
         "kimi-k2.5",
+        "kimi-for-coding",
         "kimi-k2-thinking",
         "kimi-k2-thinking-turbo",
         "kimi-k2-turbo-preview",
@@ -212,6 +225,7 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "trinity-mini",
     ],
     "opencode-zen": [
+        "kimi-k2.5",
         "gpt-5.4-pro",
         "gpt-5.4",
         "gpt-5.3-codex",
@@ -243,16 +257,15 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
         "glm-5",
         "glm-4.7",
         "glm-4.6",
-        "kimi-k2.5",
         "kimi-k2-thinking",
         "kimi-k2",
         "qwen3-coder",
         "big-pickle",
     ],
     "opencode-go": [
+        "kimi-k2.5",
         "glm-5.1",
         "glm-5",
-        "kimi-k2.5",
         "mimo-v2-pro",
         "mimo-v2-omni",
         "minimax-m2.7",
@@ -285,21 +298,21 @@ _PROVIDER_MODELS: dict[str, list[str]] = {
     # to https://dashscope-intl.aliyuncs.com/compatible-mode/v1 (OpenAI-compat)
     # or https://dashscope-intl.aliyuncs.com/apps/anthropic (Anthropic-compat).
     "alibaba": [
+        "kimi-k2.5",
         "qwen3.5-plus",
         "qwen3-coder-plus",
         "qwen3-coder-next",
         # Third-party models available on coding-intl
         "glm-5",
         "glm-4.7",
-        "kimi-k2.5",
         "MiniMax-M2.5",
     ],
     # Curated HF model list — only agentic models that map to OpenRouter defaults.
     "huggingface": [
+        "moonshotai/Kimi-K2.5",
         "Qwen/Qwen3.5-397B-A17B",
         "Qwen/Qwen3.5-35B-A3B",
         "deepseek-ai/DeepSeek-V3.2",
-        "moonshotai/Kimi-K2.5",
         "MiniMaxAI/MiniMax-M2.5",
         "zai-org/GLM-5",
         "XiaomiMiMo/MiMo-V2-Flash",
@@ -536,6 +549,7 @@ CANONICAL_PROVIDERS: list[ProviderEntry] = [
     ProviderEntry("anthropic",      "Anthropic",                "Anthropic (Claude models — API key or Claude Code)"),
     ProviderEntry("openai-codex",   "OpenAI Codex",             "OpenAI Codex"),
     ProviderEntry("xiaomi",         "Xiaomi MiMo",              "Xiaomi MiMo (MiMo-V2 models — pro, omni, flash)"),
+    ProviderEntry("nvidia",         "NVIDIA NIM",               "NVIDIA NIM (Nemotron models — build.nvidia.com or local NIM)"),
     ProviderEntry("qwen-oauth",     "Qwen OAuth (Portal)",      "Qwen OAuth (reuses local Qwen CLI login)"),
     ProviderEntry("copilot",        "GitHub Copilot",           "GitHub Copilot (uses GITHUB_TOKEN or gh auth token)"),
     ProviderEntry("copilot-acp",    "GitHub Copilot ACP",       "GitHub Copilot ACP (spawns `copilot --acp --stdio`)"),
@@ -618,6 +632,10 @@ _PROVIDER_ALIASES = {
     "grok": "xai",
     "x-ai": "xai",
     "x.ai": "xai",
+    "nim": "nvidia",
+    "nvidia-nim": "nvidia",
+    "build-nvidia": "nvidia",
+    "nemotron": "nvidia",
     "ollama": "custom",  # bare "ollama" = local; use "ollama-cloud" for cloud
     "ollama_cloud": "ollama-cloud",
 }
@@ -2032,8 +2050,8 @@ def validate_requested_model(
                 )
 
             return {
-                "accepted": True,
-                "persist": True,
+                "accepted": False,
+                "persist": False,
                 "recognized": False,
                 "message": message,
             }
@@ -2046,8 +2064,8 @@ def validate_requested_model(
             message += f"\n  If this server expects `/v1`, try base URL: `{probe.get('suggested_base_url')}`"
 
         return {
-            "accepted": True,
-            "persist": True,
+            "accepted": False,
+            "persist": False,
             "recognized": False,
             "message": message,
         }
@@ -2081,12 +2099,11 @@ def validate_requested_model(
             if suggestions:
                 suggestion_text = "\n  Similar models: " + ", ".join(f"`{s}`" for s in suggestions)
             return {
-                "accepted": True,
-                "persist": True,
+                "accepted": False,
+                "persist": False,
                 "recognized": False,
                 "message": (
-                    f"Note: `{requested}` was not found in the OpenAI Codex model listing. "
-                    f"It may still work if your account has access to it."
+                    f"Model `{requested}` was not found in the OpenAI Codex model listing."
                     f"{suggestion_text}"
                 ),
             }
@@ -2125,16 +2142,15 @@ def validate_requested_model(
             if suggestions:
                 suggestion_text = "\n  Similar models: " + ", ".join(f"`{s}`" for s in suggestions)
 
-            return {
-                "accepted": True,
-                "persist": True,
-                "recognized": False,
-                "message": (
-                    f"Note: `{requested}` was not found in this provider's model listing. "
-                    f"It may still work if your plan supports it."
-                    f"{suggestion_text}"
-                ),
-            }
+        return {
+            "accepted": False,
+            "persist": False,
+            "recognized": False,
+            "message": (
+                f"Model `{requested}` was not found in this provider's model listing."
+                f"{suggestion_text}"
+            ),
+        }
 
     # api_models is None — couldn't reach API.  Accept and persist,
     # but warn so typos don't silently break things.
@@ -2176,8 +2192,8 @@ def validate_requested_model(
 
     provider_label = _PROVIDER_LABELS.get(normalized, normalized)
     return {
-        "accepted": True,
-        "persist": True,
+        "accepted": False,
+        "persist": False,
         "recognized": False,
         "message": (
             f"Could not reach the {provider_label} API to validate `{requested}`. "

hermes_cli/profiles.py

@@ -300,19 +300,10 @@ def _read_config_model(profile_dir: Path) -> tuple:
 
 def _check_gateway_running(profile_dir: Path) -> bool:
     """Check if a gateway is running for a given profile directory."""
-    pid_file = profile_dir / "gateway.pid"
-    if not pid_file.exists():
-        return False
     try:
-        raw = pid_file.read_text().strip()
-        if not raw:
-            return False
-        data = json.loads(raw) if raw.startswith("{") else {"pid": int(raw)}
-        pid = int(data["pid"])
-        os.kill(pid, 0)  # existence check
-        return True
-    except (json.JSONDecodeError, KeyError, ValueError, TypeError,
-            ProcessLookupError, PermissionError, OSError):
+        from gateway.status import get_running_pid
+        return get_running_pid(profile_dir / "gateway.pid", cleanup_stale=False) is not None
+    except Exception:
         return False
 
 

hermes_cli/providers.py

@@ -137,6 +137,11 @@ HERMES_OVERLAYS: Dict[str, HermesOverlay] = {
         base_url_override="https://api.x.ai/v1",
         base_url_env_var="XAI_BASE_URL",
     ),
+    "nvidia": HermesOverlay(
+        transport="openai_chat",
+        base_url_override="https://integrate.api.nvidia.com/v1",
+        base_url_env_var="NVIDIA_BASE_URL",
+    ),
     "xiaomi": HermesOverlay(
         transport="openai_chat",
         base_url_env_var="XIAOMI_BASE_URL",
@@ -191,6 +196,12 @@ ALIASES: Dict[str, str] = {
     "x.ai": "xai",
     "grok": "xai",
 
+    # nvidia
+    "nim": "nvidia",
+    "nvidia-nim": "nvidia",
+    "build-nvidia": "nvidia",
+    "nemotron": "nvidia",
+
     # kimi-for-coding (models.dev ID)
     "kimi": "kimi-for-coding",
     "kimi-coding": "kimi-for-coding",

hermes_cli/setup.py

@@ -91,7 +91,7 @@ _DEFAULT_PROVIDER_MODELS = {
     "gemini": [
         "gemini-3.1-pro-preview", "gemini-3-flash-preview", "gemini-3.1-flash-lite-preview",
         "gemini-2.5-pro", "gemini-2.5-flash", "gemini-2.5-flash-lite",
-        "gemma-4-31b-it", "gemma-4-26b-it",
+        "gemma-4-31b-it",
     ],
     "zai": ["glm-5.1", "glm-5", "glm-4.7", "glm-4.5", "glm-4.5-flash"],
     "kimi-coding": ["kimi-k2.5", "kimi-k2-thinking", "kimi-k2-turbo-preview"],
@@ -2005,52 +2005,6 @@ def _setup_wecom_callback():
     _gw_setup()
 
 
-def _setup_qqbot():
-    """Configure QQ Bot gateway."""
-    print_header("QQ Bot")
-    existing = get_env_value("QQ_APP_ID")
-    if existing:
-        print_info("QQ Bot: already configured")
-        if not prompt_yes_no("Reconfigure QQ Bot?", False):
-            return
-
-    print_info("Connects Hermes to QQ via the Official QQ Bot API (v2).")
-    print_info("   Requires a QQ Bot application at q.qq.com")
-    print_info("   Reference: https://bot.q.qq.com/wiki/develop/api-v2/")
-    print()
-
-    app_id = prompt("QQ Bot App ID")
-    if not app_id:
-        print_warning("App ID is required — skipping QQ Bot setup")
-        return
-    save_env_value("QQ_APP_ID", app_id.strip())
-
-    client_secret = prompt("QQ Bot App Secret", password=True)
-    if not client_secret:
-        print_warning("App Secret is required — skipping QQ Bot setup")
-        return
-    save_env_value("QQ_CLIENT_SECRET", client_secret)
-    print_success("QQ Bot credentials saved")
-
-    print()
-    print_info("🔒 Security: Restrict who can DM your bot")
-    print_info("   Use QQ user OpenIDs (found in event payloads)")
-    print()
-    allowed_users = prompt("Allowed user OpenIDs (comma-separated, leave empty for open access)")
-    if allowed_users:
-        save_env_value("QQ_ALLOWED_USERS", allowed_users.replace(" ", ""))
-        print_success("QQ Bot allowlist configured")
-    else:
-        print_info("⚠️  No allowlist set — anyone can DM the bot!")
-
-    print()
-    print_info("📬 Home Channel: OpenID for cron job delivery and notifications.")
-    home_channel = prompt("Home channel OpenID (leave empty to set later)")
-    if home_channel:
-        save_env_value("QQ_HOME_CHANNEL", home_channel)
-
-    print()
-    print_success("QQ Bot configured!")
 
 
 def _setup_bluebubbles():
@@ -2119,12 +2073,9 @@ def _setup_bluebubbles():
 
 
 def _setup_qqbot():
-    """Configure QQ Bot (Official API v2) via standard platform setup."""
-    from hermes_cli.gateway import _PLATFORMS
-    qq_platform = next((p for p in _PLATFORMS if p["key"] == "qqbot"), None)
-    if qq_platform:
-        from hermes_cli.gateway import _setup_standard_platform
-        _setup_standard_platform(qq_platform)
+    """Configure QQ Bot (Official API v2) via gateway setup."""
+    from hermes_cli.gateway import _setup_qqbot as _gateway_setup_qqbot
+    _gateway_setup_qqbot()
 
 
 def _setup_webhooks():
@@ -2264,7 +2215,9 @@ def setup_gateway(config: dict):
             missing_home.append("Slack")
         if get_env_value("BLUEBUBBLES_SERVER_URL") and not get_env_value("BLUEBUBBLES_HOME_CHANNEL"):
             missing_home.append("BlueBubbles")
-        if get_env_value("QQ_APP_ID") and not get_env_value("QQ_HOME_CHANNEL"):
+        if get_env_value("QQ_APP_ID") and not (
+            get_env_value("QQBOT_HOME_CHANNEL") or get_env_value("QQ_HOME_CHANNEL")
+        ):
             missing_home.append("QQBot")
 
         if missing_home:
@@ -2289,8 +2242,10 @@ def setup_gateway(config: dict):
             _is_service_running,
             supports_systemd_services,
             has_conflicting_systemd_units,
+            has_legacy_hermes_units,
             install_linux_gateway_from_setup,
             print_systemd_scope_conflict_warning,
+            print_legacy_unit_warning,
             systemd_start,
             systemd_restart,
             launchd_install,
@@ -2308,6 +2263,10 @@ def setup_gateway(config: dict):
             print_systemd_scope_conflict_warning()
             print()
 
+        if supports_systemd and has_legacy_hermes_units():
+            print_legacy_unit_warning()
+            print()
+
         if service_running:
             if prompt_yes_no("  Restart the gateway to pick up changes?", True):
                 try:

hermes_cli/skills_hub.py

@@ -515,6 +515,90 @@ def do_inspect(identifier: str, console: Optional[Console] = None) -> None:
     c.print()
 
 
+def browse_skills(page: int = 1, page_size: int = 20, source: str = "all") -> dict:
+    """Paginated hub browse for programmatic callers (e.g. TUI gateway).
+
+    Returns ``{"items": [...], "page": int, "total_pages": int, "total": int}``.
+    """
+    from tools.skills_hub import GitHubAuth, create_source_router
+
+    page_size = max(1, min(page_size, 100))
+    _TRUST_RANK = {"builtin": 3, "trusted": 2, "community": 1}
+    _PER_SOURCE_LIMIT = {"official": 100, "skills-sh": 100, "well-known": 25, "github": 100, "clawhub": 50,
+                         "claude-marketplace": 50, "lobehub": 50}
+    auth = GitHubAuth()
+    sources = create_source_router(auth)
+    all_results: list = []
+    for src in sources:
+        sid = src.source_id()
+        if source != "all" and sid != source and sid != "official":
+            continue
+        try:
+            limit = _PER_SOURCE_LIMIT.get(sid, 50)
+            all_results.extend(src.search("", limit=limit))
+        except Exception:
+            continue
+    if not all_results:
+        return {"items": [], "page": 1, "total_pages": 1, "total": 0}
+    seen: dict = {}
+    for r in all_results:
+        rank = _TRUST_RANK.get(r.trust_level, 0)
+        if r.name not in seen or rank > _TRUST_RANK.get(seen[r.name].trust_level, 0):
+            seen[r.name] = r
+    deduped = list(seen.values())
+    deduped.sort(key=lambda r: (-_TRUST_RANK.get(r.trust_level, 0), r.source != "official", r.name.lower()))
+    total = len(deduped)
+    total_pages = max(1, (total + page_size - 1) // page_size)
+    page = max(1, min(page, total_pages))
+    start = (page - 1) * page_size
+    page_items = deduped[start : min(start + page_size, total)]
+    return {
+        "items": [{"name": r.name, "description": r.description, "source": r.source,
+                    "trust": r.trust_level} for r in page_items],
+        "page": page,
+        "total_pages": total_pages,
+        "total": total,
+    }
+
+
+def inspect_skill(identifier: str) -> Optional[dict]:
+    """Skill metadata (+ SKILL.md preview) for programmatic callers."""
+    from tools.skills_hub import GitHubAuth, create_source_router
+
+    class _Q:
+        def print(self, *a, **k):
+            pass
+
+    c = _Q()
+    auth = GitHubAuth()
+    sources = create_source_router(auth)
+    ident = identifier
+    if "/" not in ident:
+        ident = _resolve_short_name(ident, sources, c)
+        if not ident:
+            return None
+    meta, bundle, _ = _resolve_source_meta_and_bundle(ident, sources)
+    if not meta:
+        return None
+    out: dict = {
+        "name": meta.name,
+        "description": meta.description,
+        "source": meta.source,
+        "identifier": meta.identifier,
+        "tags": list(meta.tags) if meta.tags else [],
+    }
+    if bundle and "SKILL.md" in bundle.files:
+        content = bundle.files["SKILL.md"]
+        if isinstance(content, bytes):
+            content = content.decode("utf-8", errors="replace")
+        lines = content.split("\n")
+        preview = "\n".join(lines[:50])
+        if len(lines) > 50:
+            preview += f"\n\n... ({len(lines) - 50} more lines)"
+        out["skill_md_preview"] = preview
+    return out
+
+
 def do_list(source_filter: str = "all", console: Optional[Console] = None) -> None:
     """List installed skills, distinguishing hub, builtin, and local skills."""
     from tools.skills_hub import HubLockFile, ensure_hub_dirs

hermes_cli/skin_engine.py

@@ -23,7 +23,7 @@ All fields are optional. Missing values inherit from the ``default`` skin.
       banner_dim: "#B8860B"               # Dim/muted text (separators, labels)
       banner_text: "#FFF8DC"              # Body text (tool names, skill names)
       ui_accent: "#FFBF00"               # General UI accent
-      ui_label: "#4dd0e1"                # UI labels
+      ui_label: "#DAA520"                # UI labels (warm gold; teal clashed w/ default banner gold)
       ui_ok: "#4caf50"                   # Success indicators
       ui_error: "#ef5350"                # Error indicators
       ui_warn: "#ffa726"                 # Warning indicators
@@ -163,7 +163,7 @@ _BUILTIN_SKINS: Dict[str, Dict[str, Any]] = {
             "banner_dim": "#B8860B",
             "banner_text": "#FFF8DC",
             "ui_accent": "#FFBF00",
-            "ui_label": "#4dd0e1",
+            "ui_label": "#DAA520",
             "ui_ok": "#4caf50",
             "ui_error": "#ef5350",
             "ui_warn": "#ffa726",

hermes_cli/status.py

@@ -317,7 +317,7 @@ def show_status(args):
         "WeCom Callback": ("WECOM_CALLBACK_CORP_ID", None),
         "Weixin": ("WEIXIN_ACCOUNT_ID", "WEIXIN_HOME_CHANNEL"),
         "BlueBubbles": ("BLUEBUBBLES_SERVER_URL", "BLUEBUBBLES_HOME_CHANNEL"),
-        "QQBot": ("QQ_APP_ID", "QQ_HOME_CHANNEL"),
+        "QQBot": ("QQ_APP_ID", "QQBOT_HOME_CHANNEL"),
     }
     
     for name, (token_var, home_var) in platforms.items():
@@ -327,6 +327,9 @@ def show_status(args):
         home_channel = ""
         if home_var:
             home_channel = os.getenv(home_var, "")
+        # Back-compat: QQBot home channel was renamed from QQ_HOME_CHANNEL to QQBOT_HOME_CHANNEL
+        if not home_channel and home_var == "QQBOT_HOME_CHANNEL":
+            home_channel = os.getenv("QQ_HOME_CHANNEL", "")
         
         status = "configured" if has_token else "not configured"
         if home_channel:
@@ -339,73 +342,36 @@ def show_status(args):
     # =========================================================================
     print()
     print(color("◆ Gateway Service", Colors.CYAN, Colors.BOLD))
-    
-    if _is_termux():
-        try:
-            from hermes_cli.gateway import find_gateway_pids
-            gateway_pids = find_gateway_pids()
-        except Exception:
-            gateway_pids = []
-        is_running = bool(gateway_pids)
+
+    try:
+        from hermes_cli.gateway import get_gateway_runtime_snapshot, _format_gateway_pids
+
+        snapshot = get_gateway_runtime_snapshot()
+        is_running = snapshot.running
         print(f"  Status:       {check_mark(is_running)} {'running' if is_running else 'stopped'}")
-        print("  Manager:      Termux / manual process")
-        if gateway_pids:
-            rendered = ", ".join(str(pid) for pid in gateway_pids[:3])
-            if len(gateway_pids) > 3:
-                rendered += ", ..."
-            print(f"  PID(s):       {rendered}")
-        else:
+        print(f"  Manager:      {snapshot.manager}")
+        if snapshot.gateway_pids:
+            print(f"  PID(s):       {_format_gateway_pids(snapshot.gateway_pids)}")
+        if snapshot.has_process_service_mismatch:
+            print("  Service:      installed but not managing the current running gateway")
+        elif _is_termux() and not snapshot.gateway_pids:
             print("  Start with:   hermes gateway")
             print("  Note:         Android may stop background jobs when Termux is suspended")
-
-    elif sys.platform.startswith('linux'):
-        from hermes_constants import is_container
-        if is_container():
-            # Docker/Podman: no systemd — check for running gateway processes
-            try:
-                from hermes_cli.gateway import find_gateway_pids
-                gateway_pids = find_gateway_pids()
-                is_active = len(gateway_pids) > 0
-            except Exception:
-                is_active = False
-            print(f"  Status:       {check_mark(is_active)} {'running' if is_active else 'stopped'}")
-            print("  Manager:      docker (foreground)")
+        elif snapshot.service_installed and not snapshot.service_running:
+            print("  Service:      installed but stopped")
+    except Exception:
+        if _is_termux():
+            print(f"  Status:       {color('unknown', Colors.DIM)}")
+            print("  Manager:      Termux / manual process")
+        elif sys.platform.startswith('linux'):
+            print(f"  Status:       {color('unknown', Colors.DIM)}")
+            print("  Manager:      systemd/manual")
+        elif sys.platform == 'darwin':
+            print(f"  Status:       {color('unknown', Colors.DIM)}")
+            print("  Manager:      launchd")
         else:
-            try:
-                from hermes_cli.gateway import get_service_name
-                _gw_svc = get_service_name()
-            except Exception:
-                _gw_svc = "hermes-gateway"
-            try:
-                result = subprocess.run(
-                    ["systemctl", "--user", "is-active", _gw_svc],
-                    capture_output=True,
-                    text=True,
-                    timeout=5
-                )
-                is_active = result.stdout.strip() == "active"
-            except (FileNotFoundError, subprocess.TimeoutExpired):
-                is_active = False
-            print(f"  Status:       {check_mark(is_active)} {'running' if is_active else 'stopped'}")
-            print("  Manager:      systemd (user)")
-        
-    elif sys.platform == 'darwin':
-        from hermes_cli.gateway import get_launchd_label
-        try:
-            result = subprocess.run(
-                ["launchctl", "list", get_launchd_label()],
-                capture_output=True,
-                text=True,
-                timeout=5
-            )
-            is_loaded = result.returncode == 0
-        except subprocess.TimeoutExpired:
-            is_loaded = False
-        print(f"  Status:       {check_mark(is_loaded)} {'loaded' if is_loaded else 'not loaded'}")
-        print("  Manager:      launchd")
-    else:
-        print(f"  Status:       {color('N/A', Colors.DIM)}")
-        print("  Manager:      (not supported on this platform)")
+            print(f"  Status:       {color('N/A', Colors.DIM)}")
+            print("  Manager:      (not supported on this platform)")
     
     # =========================================================================
     # Cron Jobs

hermes_cli/web_server.py

@@ -56,10 +56,10 @@ try:
 except ImportError:
     raise SystemExit(
         "Web UI requires fastapi and uvicorn.\n"
-        "Run 'hermes web' to auto-install, or: pip install hermes-agent[web]"
+        f"Install with: {sys.executable} -m pip install 'fastapi' 'uvicorn[standard]'"
     )
 
-WEB_DIST = Path(__file__).parent / "web_dist"
+WEB_DIST = Path(os.environ["HERMES_WEB_DIST"]) if "HERMES_WEB_DIST" in os.environ else Path(__file__).parent / "web_dist"
 _log = logging.getLogger(__name__)
 
 app = FastAPI(title="Hermes Agent", version=__version__)
@@ -1444,38 +1444,8 @@ def _nous_poller(session_id: str) -> None:
             auth_state, min_key_ttl_seconds=300, timeout_seconds=15.0,
             force_refresh=False, force_mint=True,
         )
-        # Save into credential pool same as auth_commands.py does
-        from agent.credential_pool import (
-            PooledCredential,
-            load_pool,
-            AUTH_TYPE_OAUTH,
-            SOURCE_MANUAL,
-        )
-        pool = load_pool("nous")
-        entry = PooledCredential.from_dict("nous", {
-            **full_state,
-            "label": "dashboard device_code",
-            "auth_type": AUTH_TYPE_OAUTH,
-            "source": f"{SOURCE_MANUAL}:dashboard_device_code",
-            "base_url": full_state.get("inference_base_url"),
-        })
-        pool.add_entry(entry)
-        # Also persist to auth store so get_nous_auth_status() sees it
-        # (matches what _login_nous in auth.py does for the CLI flow).
-        try:
-            from hermes_cli.auth import (
-                _load_auth_store, _save_provider_state, _save_auth_store,
-                _auth_store_lock,
-            )
-            with _auth_store_lock():
-                auth_store = _load_auth_store()
-                _save_provider_state(auth_store, "nous", full_state)
-                _save_auth_store(auth_store)
-        except Exception as store_exc:
-            _log.warning(
-                "oauth/device: credential pool saved but auth store write failed "
-                "(session=%s): %s", session_id, store_exc,
-            )
+        from hermes_cli.auth import persist_nous_credentials
+        persist_nous_credentials(full_state)
         with _oauth_sessions_lock:
             sess["status"] = "approved"
         _log.info("oauth/device: nous login completed (session=%s)", session_id)

hermes_constants.py

@@ -14,7 +14,8 @@ def get_hermes_home() -> Path:
     Reads HERMES_HOME env var, falls back to ~/.hermes.
     This is the single source of truth — all other copies should import this.
     """
-    return Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+    val = os.environ.get("HERMES_HOME", "").strip()
+    return Path(val) if val else Path.home() / ".hermes"
 
 
 def get_default_hermes_root() -> Path:

hermes_state.py

@@ -987,6 +987,22 @@ class SessionDB:
 
         return sanitized.strip()
 
+
+    @staticmethod
+    def _contains_cjk(text: str) -> bool:
+        """Check if text contains CJK (Chinese, Japanese, Korean) characters."""
+        for ch in text:
+            cp = ord(ch)
+            if (0x4E00 <= cp <= 0x9FFF or    # CJK Unified Ideographs
+                0x3400 <= cp <= 0x4DBF or    # CJK Extension A
+                0x20000 <= cp <= 0x2A6DF or  # CJK Extension B
+                0x3000 <= cp <= 0x303F or    # CJK Symbols
+                0x3040 <= cp <= 0x309F or    # Hiragana
+                0x30A0 <= cp <= 0x30FF or    # Katakana
+                0xAC00 <= cp <= 0xD7AF):     # Hangul Syllables
+                return True
+        return False
+
     def search_messages(
         self,
         query: str,
@@ -1062,8 +1078,47 @@ class SessionDB:
                 cursor = self._conn.execute(sql, params)
             except sqlite3.OperationalError:
                 # FTS5 query syntax error despite sanitization — return empty
-                return []
-            matches = [dict(row) for row in cursor.fetchall()]
+                # unless query contains CJK (fall back to LIKE below)
+                if not self._contains_cjk(query):
+                    return []
+                matches = []
+            else:
+                matches = [dict(row) for row in cursor.fetchall()]
+
+        # LIKE fallback for CJK queries: FTS5 default tokenizer splits CJK
+        # characters individually, causing multi-character queries to fail.
+        if not matches and self._contains_cjk(query):
+            raw_query = query.strip('"').strip()
+            like_where = ["m.content LIKE ?"]
+            like_params: list = [f"%{raw_query}%"]
+            if source_filter is not None:
+                like_where.append(f"s.source IN ({','.join('?' for _ in source_filter)})")
+                like_params.extend(source_filter)
+            if exclude_sources is not None:
+                like_where.append(f"s.source NOT IN ({','.join('?' for _ in exclude_sources)})")
+                like_params.extend(exclude_sources)
+            if role_filter:
+                like_where.append(f"m.role IN ({','.join('?' for _ in role_filter)})")
+                like_params.extend(role_filter)
+            like_sql = f"""
+                SELECT m.id, m.session_id, m.role,
+                       substr(m.content,
+                              max(1, instr(m.content, ?) - 40),
+                              120) AS snippet,
+                       m.content, m.timestamp, m.tool_name,
+                       s.source, s.model, s.started_at AS session_started
+                FROM messages m
+                JOIN sessions s ON s.id = m.session_id
+                WHERE {' AND '.join(like_where)}
+                ORDER BY m.timestamp DESC
+                LIMIT ? OFFSET ?
+            """
+            like_params.extend([limit, offset])
+            # instr() parameter goes first in the bound list
+            like_params = [raw_query] + like_params
+            with self._lock:
+                like_cursor = self._conn.execute(like_sql, like_params)
+                matches = [dict(row) for row in like_cursor.fetchall()]
 
         # Add surrounding context (1 message before + after each match).
         # Done outside the lock so we don't hold it across N sequential queries.

mcp_serve.py

@@ -433,7 +433,7 @@ def create_mcp_server(event_bridge: Optional[EventBridge] = None) -> "FastMCP":
     if not _MCP_SERVER_AVAILABLE:
         raise ImportError(
             "MCP server requires the 'mcp' package. "
-            "Install with: pip install 'hermes-agent[mcp]'"
+            f"Install with: {sys.executable} -m pip install 'mcp'"
         )
 
     mcp = FastMCP(
@@ -838,7 +838,7 @@ def run_mcp_server(verbose: bool = False) -> None:
     if not _MCP_SERVER_AVAILABLE:
         print(
             "Error: MCP server requires the 'mcp' package.\n"
-            "Install with: pip install 'hermes-agent[mcp]'",
+            f"Install with: {sys.executable} -m pip install 'mcp'",
             file=sys.stderr,
         )
         sys.exit(1)

mini_swe_runner.py

@@ -43,6 +43,15 @@ from dotenv import load_dotenv
 load_dotenv()
 
 
+def _effective_temperature_for_model(model: str) -> Optional[float]:
+    """Return a fixed temperature for models with strict sampling contracts."""
+    try:
+        from agent.auxiliary_client import _fixed_temperature_for_model
+    except Exception:
+        return None
+    return _fixed_temperature_for_model(model)
+
+
 
 
 # ============================================================================
@@ -442,12 +451,17 @@ Complete the user's task step by step."""
                 
                 # Make API call
                 try:
-                    response = self.client.chat.completions.create(
-                        model=self.model,
-                        messages=api_messages,
-                        tools=self.tools,
-                        timeout=300.0
-                    )
+                    api_kwargs = {
+                        "model": self.model,
+                        "messages": api_messages,
+                        "tools": self.tools,
+                        "timeout": 300.0,
+                    }
+                    fixed_temperature = _effective_temperature_for_model(self.model)
+                    if fixed_temperature is not None:
+                        api_kwargs["temperature"] = fixed_temperature
+
+                    response = self.client.chat.completions.create(**api_kwargs)
                 except Exception as e:
                     self.logger.error(f"API call failed: {e}")
                     break

model_tools.py

@@ -274,9 +274,9 @@ def get_tool_definitions(
     # execute_code" even when the API key isn't configured or the toolset is
     # disabled (#560-discord).
     if "execute_code" in available_tool_names:
-        from tools.code_execution_tool import SANDBOX_ALLOWED_TOOLS, build_execute_code_schema
+        from tools.code_execution_tool import SANDBOX_ALLOWED_TOOLS, build_execute_code_schema, _get_execution_mode
         sandbox_enabled = SANDBOX_ALLOWED_TOOLS & available_tool_names
-        dynamic_schema = build_execute_code_schema(sandbox_enabled)
+        dynamic_schema = build_execute_code_schema(sandbox_enabled, mode=_get_execution_mode())
         for i, td in enumerate(filtered_tools):
             if td.get("function", {}).get("name") == "execute_code":
                 filtered_tools[i] = {"type": "function", "function": dynamic_schema}

nix/checks.nix

@@ -103,6 +103,51 @@ json.dump(sorted(leaf_paths(DEFAULT_CONFIG)), sys.stdout, indent=2)
           echo "ok" > $out/result
         '';
 
+        # Verify bundled TUI is present and compiled
+        bundled-tui = pkgs.runCommand "hermes-bundled-tui" { } ''
+          set -e
+          echo "=== Checking bundled TUI ==="
+          test -d ${hermes-agent}/ui-tui || (echo "FAIL: ui-tui directory missing"; exit 1)
+          echo "PASS: ui-tui directory exists"
+
+          test -f ${hermes-agent}/ui-tui/dist/entry.js || (echo "FAIL: compiled entry.js missing"; exit 1)
+          echo "PASS: compiled entry.js present"
+
+          test -d ${hermes-agent}/ui-tui/node_modules || (echo "FAIL: node_modules missing"; exit 1)
+          echo "PASS: node_modules present"
+
+          grep -q "HERMES_TUI_DIR" ${hermes-agent}/bin/hermes || \
+            (echo "FAIL: HERMES_TUI_DIR not in wrapper"; exit 1)
+          echo "PASS: HERMES_TUI_DIR set in wrapper"
+
+          echo "=== All bundled TUI checks passed ==="
+          mkdir -p $out
+          echo "ok" > $out/result
+        '';
+
+        # Verify HERMES_NODE is set in wrapper and points to Node 20+
+        # (string-width uses the /v regex flag which requires Node 20+)
+        hermes-node = pkgs.runCommand "hermes-node-version" { } ''
+          set -e
+          echo "=== Checking HERMES_NODE in wrapper ==="
+          grep -q "HERMES_NODE" ${hermes-agent}/bin/hermes || \
+            (echo "FAIL: HERMES_NODE not set in wrapper"; exit 1)
+          echo "PASS: HERMES_NODE present in wrapper"
+
+          HERMES_NODE=$(sed -n "s/^export HERMES_NODE='\(.*\)'/\1/p" ${hermes-agent}/bin/hermes)
+          test -x "$HERMES_NODE" || (echo "FAIL: HERMES_NODE=$HERMES_NODE not executable"; exit 1)
+          echo "PASS: HERMES_NODE executable at $HERMES_NODE"
+
+          NODE_MAJOR=$("$HERMES_NODE" --version | sed 's/^v//' | cut -d. -f1)
+          test "$NODE_MAJOR" -ge 20 || \
+            (echo "FAIL: Node v$NODE_MAJOR < 20, TUI needs /v regex flag support"; exit 1)
+          echo "PASS: Node v$NODE_MAJOR >= 20"
+
+          echo "=== All HERMES_NODE checks passed ==="
+          mkdir -p $out
+          echo "ok" > $out/result
+        '';
+
         # Verify HERMES_MANAGED guard works on all mutation commands
         managed-guard = pkgs.runCommand "hermes-managed-guard" { } ''
           set -e

nix/devShell.nix

@@ -1,49 +1,26 @@
-# nix/devShell.nix — Fast dev shell with stamp-file optimization
+# nix/devShell.nix — Dev shell that delegates setup to each package
+#
+# Each package in inputsFrom exposes passthru.devShellHook — a bash snippet
+# with stamp-checked setup logic. This file collects and runs them all.
 { inputs, ... }: {
-  perSystem = { pkgs, ... }:
+  perSystem = { pkgs, system, ... }:
     let
-      python = pkgs.python311;
+      hermes-agent = inputs.self.packages.${system}.default;
+      hermes-tui = inputs.self.packages.${system}.tui;
+      packages = [ hermes-agent hermes-tui ];
     in {
       devShells.default = pkgs.mkShell {
+        inputsFrom = packages;
         packages = with pkgs; [
-          python uv nodejs_20 ripgrep git openssh ffmpeg
+          python311 uv nodejs_22 ripgrep git openssh ffmpeg
         ];
 
-        shellHook = ''
+        shellHook = let
+          hooks = map (p: p.passthru.devShellHook or "") packages;
+          combined = pkgs.lib.concatStringsSep "\n" (builtins.filter (h: h != "") hooks);
+        in ''
           echo "Hermes Agent dev shell"
-
-          # Composite stamp: changes when nix python or uv change
-          STAMP_VALUE="${python}:${pkgs.uv}"
-          STAMP_FILE=".venv/.nix-stamp"
-
-          # Create venv if missing
-          if [ ! -d .venv ]; then
-            echo "Creating Python 3.11 venv..."
-            uv venv .venv --python ${python}/bin/python3
-          fi
-
-          source .venv/bin/activate
-
-          # Only install if stamp is stale or missing
-          if [ ! -f "$STAMP_FILE" ] || [ "$(cat "$STAMP_FILE")" != "$STAMP_VALUE" ]; then
-            echo "Installing Python dependencies..."
-            uv pip install -e ".[all]"
-            if [ -d mini-swe-agent ]; then
-              uv pip install -e ./mini-swe-agent 2>/dev/null || true
-            fi
-            if [ -d tinker-atropos ]; then
-              uv pip install -e ./tinker-atropos 2>/dev/null || true
-            fi
-
-            # Install npm deps
-            if [ -f package.json ] && [ ! -d node_modules ]; then
-              echo "Installing npm dependencies..."
-              npm install
-            fi
-
-            echo "$STAMP_VALUE" > "$STAMP_FILE"
-          fi
-
+          ${combined}
           echo "Ready. Run 'hermes' to start."
         '';
       };

nix/nixosModules.nix

@@ -121,11 +121,19 @@
       # ── Provision apt packages (first boot only, cached in writable layer) ──
       # sudo: agent self-modification
       # nodejs/npm: writable node so npm i -g works (nix store copies are read-only)
-      # curl: needed for uv installer
+      #   Node 22 via NodeSource — Ubuntu 24.04 ships Node 18 which is EOL.
+      # curl: needed for uv installer + NodeSource setup
       if [ ! -f /var/lib/hermes-tools-provisioned ] && command -v apt-get >/dev/null 2>&1; then
         echo "First boot: provisioning agent tools..."
         apt-get update -qq
-        apt-get install -y -qq sudo nodejs npm curl
+        apt-get install -y -qq sudo curl ca-certificates gnupg
+        mkdir -p /etc/apt/keyrings
+        curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key \
+          | gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg
+        echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_22.x nodistro main" \
+          > /etc/apt/sources.list.d/nodesource.list
+        apt-get update -qq
+        apt-get install -y -qq nodejs
         touch /var/lib/hermes-tools-provisioned
       fi
 
@@ -171,7 +179,7 @@
     # Package and entrypoint use stable symlinks (current-package, current-entrypoint)
     # so they can update without recreation. Env vars go through $HERMES_HOME/.env.
     containerIdentity = builtins.hashString "sha256" (builtins.toJSON {
-      schema = 3; # bump when identity inputs change
+      schema = 4; # bump when identity inputs change (4: Node 18→22 via NodeSource)
       image = cfg.container.image;
       extraVolumes = cfg.container.extraVolumes;
       extraOptions = cfg.container.extraOptions;

nix/packages.nix

@@ -1,54 +1,116 @@
 # nix/packages.nix — Hermes Agent package built with uv2nix
-{ inputs, ... }: {
-  perSystem = { pkgs, system, ... }:
+{ inputs, ... }:
+{
+  perSystem =
+    { pkgs, inputs', ... }:
     let
       hermesVenv = pkgs.callPackage ./python.nix {
         inherit (inputs) uv2nix pyproject-nix pyproject-build-systems;
       };
 
+      hermesTui = pkgs.callPackage ./tui.nix {
+        npm-lockfile-fix = inputs'.npm-lockfile-fix.packages.default;
+      };
+
       # Import bundled skills, excluding runtime caches
       bundledSkills = pkgs.lib.cleanSourceWith {
         src = ../skills;
-        filter = path: _type:
-          !(pkgs.lib.hasInfix "/index-cache/" path);
+        filter = path: _type: !(pkgs.lib.hasInfix "/index-cache/" path);
+      };
+
+      hermesWeb = pkgs.callPackage ./web.nix {
+        npm-lockfile-fix = inputs'.npm-lockfile-fix.packages.default;
       };
 
       runtimeDeps = with pkgs; [
-        nodejs_20 ripgrep git openssh ffmpeg tirith
+        nodejs_22
+        ripgrep
+        git
+        openssh
+        ffmpeg
+        tirith
       ];
 
       runtimePath = pkgs.lib.makeBinPath runtimeDeps;
-    in {
-      packages.default = pkgs.stdenv.mkDerivation {
-        pname = "hermes-agent";
-        version = (builtins.fromTOML (builtins.readFile ../pyproject.toml)).project.version;
 
-        dontUnpack = true;
-        dontBuild = true;
-        nativeBuildInputs = [ pkgs.makeWrapper ];
+      # Lockfile hashes for dev shell stamps
+      pyprojectHash = builtins.hashString "sha256" (builtins.readFile ../pyproject.toml);
+      uvLockHash =
+        if builtins.pathExists ../uv.lock then
+          builtins.hashString "sha256" (builtins.readFile ../uv.lock)
+        else
+          "none";
+    in
+    {
+      packages = {
+        default = pkgs.stdenv.mkDerivation {
+          pname = "hermes-agent";
+          version = (fromTOML (builtins.readFile ../pyproject.toml)).project.version;
 
-        installPhase = ''
-          runHook preInstall
+          dontUnpack = true;
+          dontBuild = true;
+          nativeBuildInputs = [ pkgs.makeWrapper ];
 
-          mkdir -p $out/share/hermes-agent $out/bin
-          cp -r ${bundledSkills} $out/share/hermes-agent/skills
+          installPhase = ''
+            runHook preInstall
 
-          ${pkgs.lib.concatMapStringsSep "\n" (name: ''
-            makeWrapper ${hermesVenv}/bin/${name} $out/bin/${name} \
-              --suffix PATH : "${runtimePath}" \
-              --set HERMES_BUNDLED_SKILLS $out/share/hermes-agent/skills
-          '') [ "hermes" "hermes-agent" "hermes-acp" ]}
+            mkdir -p $out/share/hermes-agent $out/bin
+            cp -r ${bundledSkills} $out/share/hermes-agent/skills
+            cp -r ${hermesWeb} $out/share/hermes-agent/web_dist
 
-          runHook postInstall
-        '';
+            # copy pre-built TUI (same layout as dev: ui-tui/dist/ + node_modules/)
+            mkdir -p $out/ui-tui
+            cp -r ${hermesTui}/lib/hermes-tui/* $out/ui-tui/
 
-        meta = with pkgs.lib; {
-          description = "AI agent with advanced tool-calling capabilities";
-          homepage = "https://github.com/NousResearch/hermes-agent";
-          mainProgram = "hermes";
-          license = licenses.mit;
-          platforms = platforms.unix;
+            ${pkgs.lib.concatMapStringsSep "\n"
+              (name: ''
+                makeWrapper ${hermesVenv}/bin/${name} $out/bin/${name} \
+                  --suffix PATH : "${runtimePath}" \
+                  --set HERMES_BUNDLED_SKILLS $out/share/hermes-agent/skills \
+                  --set HERMES_WEB_DIST $out/share/hermes-agent/web_dist \
+                  --set HERMES_TUI_DIR $out/ui-tui \
+                  --set HERMES_PYTHON ${hermesVenv}/bin/python3 \
+                  --set HERMES_NODE ${pkgs.nodejs_22}/bin/node
+              '')
+              [
+                "hermes"
+                "hermes-agent"
+                "hermes-acp"
+              ]
+            }
+
+            runHook postInstall
+          '';
+
+          passthru.devShellHook = ''
+            STAMP=".nix-stamps/hermes-agent"
+            STAMP_VALUE="${pyprojectHash}:${uvLockHash}"
+            if [ ! -f "$STAMP" ] || [ "$(cat "$STAMP")" != "$STAMP_VALUE" ]; then
+              echo "hermes-agent: installing Python dependencies..."
+              uv venv .venv --python ${pkgs.python311}/bin/python3 2>/dev/null || true
+              source .venv/bin/activate
+              uv pip install -e ".[all]"
+              [ -d mini-swe-agent ] && uv pip install -e ./mini-swe-agent 2>/dev/null || true
+              [ -d tinker-atropos ] && uv pip install -e ./tinker-atropos 2>/dev/null || true
+              mkdir -p .nix-stamps
+              echo "$STAMP_VALUE" > "$STAMP"
+            else
+              source .venv/bin/activate
+              export HERMES_PYTHON=${hermesVenv}/bin/python3
+            fi
+          '';
+
+          meta = with pkgs.lib; {
+            description = "AI agent with advanced tool-calling capabilities";
+            homepage = "https://github.com/NousResearch/hermes-agent";
+            mainProgram = "hermes";
+            license = licenses.mit;
+            platforms = platforms.unix;
+          };
         };
+
+        tui = hermesTui;
+        web = hermesWeb;
       };
     };
 }

nix/python.nix

@@ -35,6 +35,20 @@ let
       };
     };
 
+  # Legacy alibabacloud packages ship only sdists with setup.py/setup.cfg
+  # and no pyproject.toml, so setuptools isn't declared as a build dep.
+  buildSystemOverrides = final: prev: builtins.mapAttrs
+    (name: _: prev.${name}.overrideAttrs (old: {
+      nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [ final.setuptools ];
+    }))
+    (lib.genAttrs [
+      "alibabacloud-credentials-api"
+      "alibabacloud-endpoint-util"
+      "alibabacloud-gateway-dingtalk"
+      "alibabacloud-gateway-spi"
+      "alibabacloud-tea"
+    ] (_: null));
+
   pythonPackageOverrides = final: _prev:
     if isAarch64Darwin then {
       numpy = mkPrebuiltOverride final python311.pkgs.numpy { };
@@ -75,6 +89,7 @@ let
       (lib.composeManyExtensions [
         pyproject-build-systems.overlays.default
         overlay
+        buildSystemOverrides
         pythonPackageOverrides
       ]);
 in

nix/tui.nix

@@ -0,0 +1,77 @@
+# nix/tui.nix — Hermes TUI (Ink/React) compiled with tsc and bundled
+{ pkgs, npm-lockfile-fix, ... }:
+let
+  src = ../ui-tui;
+  npmDeps = pkgs.fetchNpmDeps {
+    inherit src;
+    hash = "sha256-mG3vpgGi4ljt4X3XIf3I/5mIcm+rVTUAmx2DQ6YVA90=";
+  };
+
+  packageJson = builtins.fromJSON (builtins.readFile (src + "/package.json"));
+  version = packageJson.version;
+
+  npmLockHash = builtins.hashString "sha256" (builtins.readFile ../ui-tui/package-lock.json);
+in
+pkgs.buildNpmPackage {
+  pname = "hermes-tui";
+  inherit src npmDeps version;
+
+  doCheck = false;
+
+  installPhase = ''
+    runHook preInstall
+
+    mkdir -p $out/lib/hermes-tui
+
+    cp -r dist $out/lib/hermes-tui/dist
+
+    # runtime node_modules
+    cp -r node_modules $out/lib/hermes-tui/node_modules
+
+    # @hermes/ink is a file: dependency, we need to copy it in fr
+    rm -f $out/lib/hermes-tui/node_modules/@hermes/ink
+    cp -r packages/hermes-ink $out/lib/hermes-tui/node_modules/@hermes/ink
+
+    # package.json needed for "type": "module" resolution
+    cp package.json $out/lib/hermes-tui/
+
+    runHook postInstall
+  '';
+
+  nativeBuildInputs = [
+    (pkgs.writeShellScriptBin "update_tui_lockfile" ''
+      set -euox pipefail
+
+      # get root of repo
+      REPO_ROOT=$(git rev-parse --show-toplevel)
+
+      # cd into ui-tui and reinstall
+      cd "$REPO_ROOT/ui-tui"
+      rm -rf node_modules/
+      npm cache clean --force
+      CI=true npm install # ci env var to suppress annoying unicode install banner lag
+      ${pkgs.lib.getExe npm-lockfile-fix} ./package-lock.json
+
+      NIX_FILE="$REPO_ROOT/nix/tui.nix"
+      # compute the new hash
+      sed -i "s/hash = \"[^\"]*\";/hash = \"\";/" $NIX_FILE
+      NIX_OUTPUT=$(nix build .#tui 2>&1 || true)
+      NEW_HASH=$(echo "$NIX_OUTPUT" | grep 'got:' | awk '{print $2}') 
+      echo got new hash $NEW_HASH
+      sed -i "s|hash = \"[^\"]*\";|hash = \"$NEW_HASH\";|" $NIX_FILE
+      nix build .#tui
+      echo "Updated npm hash in $NIX_FILE to $NEW_HASH"
+    '')
+  ];
+
+  passthru.devShellHook = ''
+    STAMP=".nix-stamps/hermes-tui"
+    STAMP_VALUE="${npmLockHash}"
+    if [ ! -f "$STAMP" ] || [ "$(cat "$STAMP")" != "$STAMP_VALUE" ]; then
+      echo "hermes-tui: installing npm dependencies..."
+      cd ui-tui && CI=true npm install --silent --no-fund --no-audit 2>/dev/null && cd ..
+      mkdir -p .nix-stamps
+      echo "$STAMP_VALUE" > "$STAMP"
+    fi
+  '';
+}

nix/web.nix

@@ -0,0 +1,63 @@
+# nix/web.nix — Hermes Web Dashboard (Vite/React) frontend build
+{ pkgs, npm-lockfile-fix, ... }:
+let
+  src = ../web;
+  npmDeps = pkgs.fetchNpmDeps {
+    inherit src;
+    hash = "sha256-Y0pOzdFG8BLjfvCLmsvqYpjxFjAQabXp1i7X9W/cCU4=";
+  };
+
+  npmLockHash = builtins.hashString "sha256" (builtins.readFile ../web/package-lock.json);
+in
+pkgs.buildNpmPackage {
+  pname = "hermes-web";
+  version = "0.0.0";
+  inherit src npmDeps;
+
+  doCheck = false;
+
+  buildPhase = ''
+    npx tsc -b
+    npx vite build --outDir dist
+  '';
+
+  installPhase = ''
+    runHook preInstall
+    cp -r dist $out
+    runHook postInstall
+  '';
+
+  nativeBuildInputs = [
+    (pkgs.writeShellScriptBin "update_web_lockfile" ''
+      set -euox pipefail
+
+      REPO_ROOT=$(git rev-parse --show-toplevel)
+
+      cd "$REPO_ROOT/web"
+      rm -rf node_modules/
+      npm cache clean --force
+      CI=true npm install
+      ${pkgs.lib.getExe npm-lockfile-fix} ./package-lock.json
+
+      NIX_FILE="$REPO_ROOT/nix/web.nix"
+      sed -i "s/hash = \"[^\"]*\";/hash = \"\";/" $NIX_FILE
+      NIX_OUTPUT=$(nix build .#web 2>&1 || true)
+      NEW_HASH=$(echo "$NIX_OUTPUT" | grep 'got:' | awk '{print $2}')
+      echo got new hash $NEW_HASH
+      sed -i "s|hash = \"[^\"]*\";|hash = \"$NEW_HASH\";|" $NIX_FILE
+      nix build .#web
+      echo "Updated npm hash in $NIX_FILE to $NEW_HASH"
+    '')
+  ];
+
+  passthru.devShellHook = ''
+    STAMP=".nix-stamps/hermes-web"
+    STAMP_VALUE="${npmLockHash}"
+    if [ ! -f "$STAMP" ] || [ "$(cat "$STAMP")" != "$STAMP_VALUE" ]; then
+      echo "hermes-web: installing npm dependencies..."
+      cd web && CI=true npm install --silent --no-fund --no-audit 2>/dev/null && cd ..
+      mkdir -p .nix-stamps
+      echo "$STAMP_VALUE" > "$STAMP"
+    fi
+  '';
+}

optional-skills/mlops/hermes-atropos-environments/SKILL.md

@@ -7,7 +7,7 @@ license: MIT
 metadata:
   hermes:
     tags: [atropos, rl, environments, training, reinforcement-learning, reward-functions]
-    related_skills: [axolotl, grpo-rl-training, trl-fine-tuning, lm-evaluation-harness]
+    related_skills: [axolotl, fine-tuning-with-trl, lm-evaluation-harness]
 ---
 
 # Hermes Agent Atropos Environments

pyproject.toml

@@ -76,8 +76,8 @@ termux = [
   "hermes-agent[honcho]",
   "hermes-agent[acp]",
 ]
-dingtalk = ["dingtalk-stream>=0.1.0,<1"]
-feishu = ["lark-oapi>=1.5.3,<2"]
+dingtalk = ["dingtalk-stream>=0.20,<1", "alibabacloud-dingtalk>=2.0.0", "qrcode>=7.0,<8"]
+feishu = ["lark-oapi>=1.5.3,<2", "qrcode>=7.0,<8"]
 web = ["fastapi>=0.104.0,<1", "uvicorn[standard]>=0.24.0,<1"]
 rl = [
   "atroposlib @ git+https://github.com/NousResearch/atropos.git@c20c85256e5a45ad31edf8b7276e9c5ee1995a30",
@@ -126,7 +126,7 @@ py-modules = ["run_agent", "model_tools", "toolsets", "batch_runner", "trajector
 hermes_cli = ["web_dist/**/*"]
 
 [tool.setuptools.packages.find]
-include = ["agent", "tools", "tools.*", "hermes_cli", "gateway", "gateway.*", "cron", "acp_adapter", "plugins", "plugins.*"]
+include = ["agent", "tools", "tools.*", "hermes_cli", "gateway", "gateway.*", "tui_gateway", "tui_gateway.*", "cron", "acp_adapter", "plugins", "plugins.*"]
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]

run_agent.py

@@ -353,12 +353,50 @@ def _sanitize_surrogates(text: str) -> str:
     return text
 
 
+def _sanitize_structure_surrogates(payload: Any) -> bool:
+    """Replace surrogate code points in nested dict/list payloads in-place.
+
+    Mirror of ``_sanitize_structure_non_ascii`` but for surrogate recovery.
+    Used to scrub nested structured fields (e.g. ``reasoning_details`` — an
+    array of dicts with ``summary``/``text`` strings) that flat per-field
+    checks don't reach.  Returns True if any surrogates were replaced.
+    """
+    found = False
+
+    def _walk(node):
+        nonlocal found
+        if isinstance(node, dict):
+            for key, value in node.items():
+                if isinstance(value, str):
+                    if _SURROGATE_RE.search(value):
+                        node[key] = _SURROGATE_RE.sub('\ufffd', value)
+                        found = True
+                elif isinstance(value, (dict, list)):
+                    _walk(value)
+        elif isinstance(node, list):
+            for idx, value in enumerate(node):
+                if isinstance(value, str):
+                    if _SURROGATE_RE.search(value):
+                        node[idx] = _SURROGATE_RE.sub('\ufffd', value)
+                        found = True
+                elif isinstance(value, (dict, list)):
+                    _walk(value)
+
+    _walk(payload)
+    return found
+
+
 def _sanitize_messages_surrogates(messages: list) -> bool:
     """Sanitize surrogate characters from all string content in a messages list.
 
     Walks message dicts in-place. Returns True if any surrogates were found
-    and replaced, False otherwise. Covers content/text, name, and tool call
-    metadata/arguments so retries don't fail on a non-content field.
+    and replaced, False otherwise. Covers content/text, name, tool call
+    metadata/arguments, AND any additional string or nested structured fields
+    (``reasoning``, ``reasoning_content``, ``reasoning_details``, etc.) so
+    retries don't fail on a non-content field.  Byte-level reasoning models
+    (xiaomi/mimo, kimi, glm) can emit lone surrogates in reasoning output
+    that flow through to ``api_messages["reasoning_content"]`` on the next
+    turn and crash json.dumps inside the OpenAI SDK.
     """
     found = False
     for msg in messages:
@@ -398,6 +436,21 @@ def _sanitize_messages_surrogates(messages: list) -> bool:
                     if isinstance(fn_args, str) and _SURROGATE_RE.search(fn_args):
                         fn["arguments"] = _SURROGATE_RE.sub('\ufffd', fn_args)
                         found = True
+        # Walk any additional string / nested fields (reasoning,
+        # reasoning_content, reasoning_details, etc.) — surrogates from
+        # byte-level reasoning models (xiaomi/mimo, kimi, glm) can lurk
+        # in these fields and aren't covered by the per-field checks above.
+        # Matches _sanitize_messages_non_ascii's coverage (PR #10537).
+        for key, value in msg.items():
+            if key in {"content", "name", "tool_calls", "role"}:
+                continue
+            if isinstance(value, str):
+                if _SURROGATE_RE.search(value):
+                    msg[key] = _SURROGATE_RE.sub('\ufffd', value)
+                    found = True
+            elif isinstance(value, (dict, list)):
+                if _sanitize_structure_surrogates(value):
+                    found = True
     return found
 
 
@@ -778,6 +831,26 @@ class AIAgent:
         self._execution_thread_id: int | None = None  # Set at run_conversation() start
         self._interrupt_thread_signal_pending = False
         self._client_lock = threading.RLock()
+
+        # /steer mechanism — inject a user note into the next tool result
+        # without interrupting the agent. Unlike interrupt(), steer() does
+        # NOT set _interrupt_requested; it waits for the current tool batch
+        # to finish naturally, then the drain hook appends the text to the
+        # last tool result's content so the model sees it on its next
+        # iteration. Message-role alternation is preserved (we modify an
+        # existing tool message rather than inserting a new user turn).
+        self._pending_steer: Optional[str] = None
+        self._pending_steer_lock = threading.Lock()
+
+        # Concurrent-tool worker thread tracking.  `_execute_tool_calls_concurrent`
+        # runs each tool on its own ThreadPoolExecutor worker — those worker
+        # threads have tids distinct from `_execution_thread_id`, so
+        # `_set_interrupt(True, _execution_thread_id)` alone does NOT cause
+        # `is_interrupted()` inside the worker to return True.  Track the
+        # workers here so `interrupt()` / `clear_interrupt()` can fan out to
+        # their tids explicitly.
+        self._tool_worker_threads: set[int] = set()
+        self._tool_worker_threads_lock = threading.Lock()
         
         # Subagent delegation state
         self._delegate_depth = 0        # 0 = top-level agent, incremented for children
@@ -981,6 +1054,16 @@ class AIAgent:
                     }
                 elif "portal.qwen.ai" in effective_base.lower():
                     client_kwargs["default_headers"] = _qwen_portal_headers()
+                elif "generativelanguage.googleapis.com" in effective_base.lower():
+                    # Google's OpenAI-compatible endpoint only accepts x-goog-api-key.
+                    # The OpenAI SDK auto-injects Authorization: Bearer when api_key= is
+                    # set to a real value, causing HTTP 400 "Multiple authentication
+                    # credentials received".  Pass a placeholder so the SDK does not
+                    # emit Bearer, and carry the real key via x-goog-api-key instead.
+                    # Fixes: https://github.com/NousResearch/hermes-agent/issues/7893
+                    real_key = client_kwargs["api_key"]
+                    client_kwargs["api_key"] = "not-used"
+                    client_kwargs["default_headers"] = {"x-goog-api-key": real_key}
             else:
                 # No explicit creds — use the centralized provider router
                 from agent.auxiliary_client import resolve_provider_client
@@ -3138,6 +3221,25 @@ class AIAgent:
             # interrupt signal until startup completes instead of targeting
             # the caller thread by mistake.
             self._interrupt_thread_signal_pending = True
+        # Fan out to concurrent-tool worker threads.  Those workers run tools
+        # on their own tids (ThreadPoolExecutor workers), so `is_interrupted()`
+        # inside a tool only sees an interrupt when their specific tid is in
+        # the `_interrupted_threads` set.  Without this propagation, an
+        # already-running concurrent tool (e.g. a terminal command hung on
+        # network I/O) never notices the interrupt and has to run to its own
+        # timeout.  See `_run_tool` for the matching entry/exit bookkeeping.
+        # `getattr` fallback covers test stubs that build AIAgent via
+        # object.__new__ and skip __init__.
+        _tracker = getattr(self, "_tool_worker_threads", None)
+        _tracker_lock = getattr(self, "_tool_worker_threads_lock", None)
+        if _tracker is not None and _tracker_lock is not None:
+            with _tracker_lock:
+                _worker_tids = list(_tracker)
+            for _wtid in _worker_tids:
+                try:
+                    _set_interrupt(True, _wtid)
+                except Exception:
+                    pass
         # Propagate interrupt to any running child agents (subagent delegation)
         with self._active_children_lock:
             children_copy = list(self._active_children)
@@ -3156,6 +3258,146 @@ class AIAgent:
         self._interrupt_thread_signal_pending = False
         if self._execution_thread_id is not None:
             _set_interrupt(False, self._execution_thread_id)
+        # Also clear any concurrent-tool worker thread bits.  Tracked
+        # workers normally clear their own bit on exit, but an explicit
+        # clear here guarantees no stale interrupt can survive a turn
+        # boundary and fire on a subsequent, unrelated tool call that
+        # happens to get scheduled onto the same recycled worker tid.
+        # `getattr` fallback covers test stubs that build AIAgent via
+        # object.__new__ and skip __init__.
+        _tracker = getattr(self, "_tool_worker_threads", None)
+        _tracker_lock = getattr(self, "_tool_worker_threads_lock", None)
+        if _tracker is not None and _tracker_lock is not None:
+            with _tracker_lock:
+                _worker_tids = list(_tracker)
+            for _wtid in _worker_tids:
+                try:
+                    _set_interrupt(False, _wtid)
+                except Exception:
+                    pass
+        # A hard interrupt supersedes any pending /steer — the steer was
+        # meant for the agent's next tool-call iteration, which will no
+        # longer happen. Drop it instead of surprising the user with a
+        # late injection on the post-interrupt turn.
+        _steer_lock = getattr(self, "_pending_steer_lock", None)
+        if _steer_lock is not None:
+            with _steer_lock:
+                self._pending_steer = None
+
+    def steer(self, text: str) -> bool:
+        """
+        Inject a user message into the next tool result without interrupting.
+
+        Unlike interrupt(), this does NOT stop the current tool call. The
+        text is stashed and the agent loop appends it to the LAST tool
+        result's content once the current tool batch finishes. The model
+        sees the steer as part of the tool output on its next iteration.
+
+        Thread-safe: callable from gateway/CLI/TUI threads. Multiple calls
+        before the drain point concatenate with newlines.
+
+        Args:
+            text: The user text to inject. Empty strings are ignored.
+
+        Returns:
+            True if the steer was accepted, False if the text was empty.
+        """
+        if not text or not text.strip():
+            return False
+        cleaned = text.strip()
+        _lock = getattr(self, "_pending_steer_lock", None)
+        if _lock is None:
+            # Test stubs that built AIAgent via object.__new__ skip __init__.
+            # Fall back to direct attribute set; no concurrent callers expected
+            # in those stubs.
+            existing = getattr(self, "_pending_steer", None)
+            self._pending_steer = (existing + "\n" + cleaned) if existing else cleaned
+            return True
+        with _lock:
+            if self._pending_steer:
+                self._pending_steer = self._pending_steer + "\n" + cleaned
+            else:
+                self._pending_steer = cleaned
+        return True
+
+    def _drain_pending_steer(self) -> Optional[str]:
+        """Return the pending steer text (if any) and clear the slot.
+
+        Safe to call from the agent execution thread after appending tool
+        results. Returns None when no steer is pending.
+        """
+        _lock = getattr(self, "_pending_steer_lock", None)
+        if _lock is None:
+            text = getattr(self, "_pending_steer", None)
+            self._pending_steer = None
+            return text
+        with _lock:
+            text = self._pending_steer
+            self._pending_steer = None
+        return text
+
+    def _apply_pending_steer_to_tool_results(self, messages: list, num_tool_msgs: int) -> None:
+        """Append any pending /steer text to the last tool result in this turn.
+
+        Called at the end of a tool-call batch, before the next API call.
+        The steer is appended to the last ``role:"tool"`` message's content
+        with a clear marker so the model understands it came from the user
+        and NOT from the tool itself. Role alternation is preserved —
+        nothing new is inserted, we only modify existing content.
+
+        Args:
+            messages: The running messages list.
+            num_tool_msgs: Number of tool results appended in this batch;
+                used to locate the tail slice safely.
+        """
+        if num_tool_msgs <= 0 or not messages:
+            return
+        steer_text = self._drain_pending_steer()
+        if not steer_text:
+            return
+        # Find the last tool-role message in the recent tail. Skipping
+        # non-tool messages defends against future code appending
+        # something else at the boundary.
+        target_idx = None
+        for j in range(len(messages) - 1, max(len(messages) - num_tool_msgs - 1, -1), -1):
+            msg = messages[j]
+            if isinstance(msg, dict) and msg.get("role") == "tool":
+                target_idx = j
+                break
+        if target_idx is None:
+            # No tool result in this batch (e.g. all skipped by interrupt);
+            # put the steer back so the caller's fallback path can deliver
+            # it as a normal next-turn user message.
+            _lock = getattr(self, "_pending_steer_lock", None)
+            if _lock is not None:
+                with _lock:
+                    if self._pending_steer:
+                        self._pending_steer = self._pending_steer + "\n" + steer_text
+                    else:
+                        self._pending_steer = steer_text
+            else:
+                existing = getattr(self, "_pending_steer", None)
+                self._pending_steer = (existing + "\n" + steer_text) if existing else steer_text
+            return
+        marker = f"\n\n[USER STEER (injected mid-run, not tool output): {steer_text}]"
+        existing_content = messages[target_idx].get("content", "")
+        if not isinstance(existing_content, str):
+            # Anthropic multimodal content blocks — preserve them and append
+            # a text block at the end.
+            try:
+                blocks = list(existing_content) if existing_content else []
+                blocks.append({"type": "text", "text": marker.lstrip()})
+                messages[target_idx]["content"] = blocks
+            except Exception:
+                # Fall back to string replacement if content shape is unexpected.
+                messages[target_idx]["content"] = f"{existing_content}{marker}"
+        else:
+            messages[target_idx]["content"] = existing_content + marker
+        logger.info(
+            "Delivered /steer to agent after tool batch (%d chars): %s",
+            len(steer_text),
+            steer_text[:120] + ("..." if len(steer_text) > 120 else ""),
+        )
 
     def _touch_activity(self, desc: str) -> None:
         """Update the last-activity timestamp and description (thread-safe)."""
@@ -5003,6 +5245,17 @@ class AIAgent:
             self._client_kwargs["default_headers"] = {"User-Agent": "KimiCLI/1.30.0"}
         elif "portal.qwen.ai" in normalized:
             self._client_kwargs["default_headers"] = _qwen_portal_headers()
+        elif "generativelanguage.googleapis.com" in normalized:
+            # Google's endpoint rejects Bearer tokens; use x-goog-api-key instead.
+            # Swap the real key out of api_key and into the header so the OpenAI
+            # SDK does not emit Authorization: Bearer.
+            # Fixes: https://github.com/NousResearch/hermes-agent/issues/7893
+            real_key = self._client_kwargs.get("api_key", "")
+            if real_key and real_key != "not-used":
+                self._client_kwargs["api_key"] = "not-used"
+            self._client_kwargs["default_headers"] = {
+                "x-goog-api-key": real_key or self._client_kwargs.get("api_key", ""),
+            }
         else:
             self._client_kwargs.pop("default_headers", None)
 
@@ -5459,7 +5712,7 @@ class AIAgent:
                 raise result["error"]
             return result["response"]
 
-        result = {"response": None, "error": None}
+        result = {"response": None, "error": None, "partial_tool_names": []}
         request_client_holder = {"client": None}
         first_delta_fired = {"done": False}
         deltas_were_sent = {"yes": False}  # Track if any deltas were fired (for fallback)
@@ -5631,6 +5884,14 @@ class AIAgent:
                             tool_gen_notified.add(idx)
                             _fire_first_delta()
                             self._fire_tool_gen_started(name)
+                            # Record the partial tool-call name so the outer
+                            # stub-builder can surface a user-visible warning
+                            # if streaming dies before this tool's arguments
+                            # are fully delivered.  Without this, a stall
+                            # during tool-call JSON generation lets the stub
+                            # at line ~6107 return `tool_calls=None`, silently
+                            # discarding the attempted action.
+                            result["partial_tool_names"].append(name)
 
                 if chunk.choices[0].finish_reason:
                     finish_reason = chunk.choices[0].finish_reason
@@ -5841,6 +6102,7 @@ class AIAgent:
                                     )
                                 except Exception:
                                     pass
+                                self._emit_status("🔄 Reconnected — resuming…")
                                 continue
                             self._emit_status(
                                 "❌ Connection to provider failed after "
@@ -5996,13 +6258,44 @@ class AIAgent:
                 _partial_text = (
                     getattr(self, "_current_streamed_assistant_text", "") or ""
                 ).strip() or None
-                logger.warning(
-                    "Partial stream delivered before error; returning stub "
-                    "response with %s chars of recovered content to prevent "
-                    "duplicate messages: %s",
-                    len(_partial_text or ""),
-                    result["error"],
-                )
+
+                # If the stream died while the model was emitting a tool call,
+                # the stub below will silently set `tool_calls=None` and the
+                # agent loop will treat the turn as complete — the attempted
+                # action is lost with no user-facing signal.  Append a
+                # human-visible warning to the stub content so (a) the user
+                # knows something failed, and (b) the next turn's model sees
+                # in conversation history what was attempted and can retry.
+                _partial_names = list(result.get("partial_tool_names") or [])
+                if _partial_names:
+                    _name_str = ", ".join(_partial_names[:3])
+                    if len(_partial_names) > 3:
+                        _name_str += f", +{len(_partial_names) - 3} more"
+                    _warn = (
+                        f"\n\n⚠ Stream stalled mid tool-call "
+                        f"({_name_str}); the action was not executed. "
+                        f"Ask me to retry if you want to continue."
+                    )
+                    _partial_text = (_partial_text or "") + _warn
+                    # Also fire as a streaming delta so the user sees it now
+                    # instead of only in the persisted transcript.
+                    try:
+                        self._fire_stream_delta(_warn)
+                    except Exception:
+                        pass
+                    logger.warning(
+                        "Partial stream dropped tool call(s) %s after %s chars "
+                        "of text; surfaced warning to user: %s",
+                        _partial_names, len(_partial_text or ""), result["error"],
+                    )
+                else:
+                    logger.warning(
+                        "Partial stream delivered before error; returning stub "
+                        "response with %s chars of recovered content to prevent "
+                        "duplicate messages: %s",
+                        len(_partial_text or ""),
+                        result["error"],
+                    )
                 _stub_msg = SimpleNamespace(
                     role="assistant", content=_partial_text, tool_calls=None,
                     reasoning_content=None,
@@ -6744,6 +7037,14 @@ class AIAgent:
             "messages": sanitized_messages,
             "timeout": float(os.getenv("HERMES_API_TIMEOUT", 1800.0)),
         }
+        try:
+            from agent.auxiliary_client import _fixed_temperature_for_model
+        except Exception:
+            _fixed_temperature_for_model = None
+        if _fixed_temperature_for_model is not None:
+            fixed_temperature = _fixed_temperature_for_model(self.model)
+            if fixed_temperature is not None:
+                api_kwargs["temperature"] = fixed_temperature
         if self._is_qwen_portal():
             api_kwargs["metadata"] = {
                 "sessionId": self.session_id or "hermes",
@@ -6949,7 +7250,7 @@ class AIAgent:
             # (gateway, batch, quiet) still get reasoning.
             # Any reasoning that wasn't shown during streaming is caught by the
             # CLI post-response display fallback (cli.py _reasoning_shown_this_turn).
-            if not self.stream_delta_callback:
+            if not self.stream_delta_callback and not self._stream_callback:
                 try:
                     self.reasoning_callback(reasoning_text)
                 except Exception:
@@ -7154,14 +7455,22 @@ class AIAgent:
 
             # Use auxiliary client for the flush call when available --
             # it's cheaper and avoids Codex Responses API incompatibility.
-            from agent.auxiliary_client import call_llm as _call_llm
+            from agent.auxiliary_client import (
+                call_llm as _call_llm,
+                _fixed_temperature_for_model,
+            )
             _aux_available = True
+            # Use the fixed-temperature override (e.g. kimi-for-coding → 0.6) if
+            # the model has a strict contract; otherwise the historical 0.3 default.
+            _flush_temperature = _fixed_temperature_for_model(self.model)
+            if _flush_temperature is None:
+                _flush_temperature = 0.3
             try:
                 response = _call_llm(
                     task="flush_memories",
                     messages=api_messages,
                     tools=[memory_tool_def],
-                    temperature=0.3,
+                    temperature=_flush_temperature,
                     max_tokens=5120,
                     # timeout resolved from auxiliary.flush_memories.timeout config
                 )
@@ -7173,7 +7482,7 @@ class AIAgent:
                 # No auxiliary client -- use the Codex Responses path directly
                 codex_kwargs = self._build_api_kwargs(api_messages)
                 codex_kwargs["tools"] = self._responses_tools([memory_tool_def])
-                codex_kwargs["temperature"] = 0.3
+                codex_kwargs["temperature"] = _flush_temperature
                 if "max_output_tokens" in codex_kwargs:
                     codex_kwargs["max_output_tokens"] = 5120
                 response = self._run_codex_stream(codex_kwargs)
@@ -7192,7 +7501,7 @@ class AIAgent:
                     "model": self.model,
                     "messages": api_messages,
                     "tools": [memory_tool_def],
-                    "temperature": 0.3,
+                    "temperature": _flush_temperature,
                     **self._max_tokens_param(5120),
                 }
                 from agent.auxiliary_client import _get_task_timeout
@@ -7583,6 +7892,22 @@ class AIAgent:
 
         def _run_tool(index, tool_call, function_name, function_args):
             """Worker function executed in a thread."""
+            # Register this worker tid so the agent can fan out an interrupt
+            # to it — see AIAgent.interrupt().  Must happen first thing, and
+            # must be paired with discard + clear in the finally block.
+            _worker_tid = threading.current_thread().ident
+            with self._tool_worker_threads_lock:
+                self._tool_worker_threads.add(_worker_tid)
+            # Race: if the agent was interrupted between fan-out (which
+            # snapshotted an empty/earlier set) and our registration, apply
+            # the interrupt to our own tid now so is_interrupted() inside
+            # the tool returns True on the next poll.
+            if self._interrupt_requested:
+                try:
+                    from tools.interrupt import set_interrupt as _sif
+                    _sif(True, _worker_tid)
+                except Exception:
+                    pass
             # Set the activity callback on THIS worker thread so
             # _wait_for_process (terminal commands) can fire heartbeats.
             # The callback is thread-local; the main thread's callback
@@ -7605,6 +7930,16 @@ class AIAgent:
             else:
                 logger.info("tool %s completed (%.2fs, %d chars)", function_name, duration, len(result))
             results[index] = (function_name, function_args, result, duration, is_error)
+            # Tear down worker-tid tracking.  Clear any interrupt bit we may
+            # have set so the next task scheduled onto this recycled tid
+            # starts with a clean slate.
+            with self._tool_worker_threads_lock:
+                self._tool_worker_threads.discard(_worker_tid)
+            try:
+                from tools.interrupt import set_interrupt as _sif
+                _sif(False, _worker_tid)
+            except Exception:
+                pass
 
         # Start spinner for CLI mode (skip when TUI handles tool progress)
         spinner = None
@@ -7749,6 +8084,13 @@ class AIAgent:
             turn_tool_msgs = messages[-num_tools:]
             enforce_turn_budget(turn_tool_msgs, env=get_active_env(effective_task_id))
 
+        # ── /steer injection ──────────────────────────────────────────────
+        # Append any pending user steer text to the last tool result so the
+        # agent sees it on its next iteration. Runs AFTER budget enforcement
+        # so the steer marker is never truncated. See steer() for details.
+        if num_tools > 0:
+            self._apply_pending_steer_to_tool_results(messages, num_tools)
+
     def _execute_tool_calls_sequential(self, assistant_message, messages: list, effective_task_id: str, api_call_count: int = 0) -> None:
         """Execute tool calls sequentially (original behavior). Used for single calls or interactive tools."""
         for i, tool_call in enumerate(assistant_message.tool_calls, 1):
@@ -8128,6 +8470,12 @@ class AIAgent:
         if num_tools_seq > 0:
             enforce_turn_budget(messages[-num_tools_seq:], env=get_active_env(effective_task_id))
 
+        # ── /steer injection ──────────────────────────────────────────────
+        # See _execute_tool_calls_parallel for the rationale. Same hook,
+        # applied to sequential execution as well.
+        if num_tools_seq > 0:
+            self._apply_pending_steer_to_tool_results(messages, num_tools_seq)
+
 
 
     def _handle_max_iterations(self, messages: list, api_call_count: int) -> str:
@@ -8165,6 +8513,15 @@ class AIAgent:
                     api_messages.insert(sys_offset + idx, pfm.copy())
 
             summary_extra_body = {}
+            try:
+                from agent.auxiliary_client import _fixed_temperature_for_model
+            except Exception:
+                _fixed_temperature_for_model = None
+            _summary_temperature = (
+                _fixed_temperature_for_model(self.model)
+                if _fixed_temperature_for_model is not None
+                else None
+            )
             _is_nous = "nousresearch" in self._base_url_lower
             if self._supports_reasoning_extra_body():
                 if self.reasoning_config is not None:
@@ -8188,6 +8545,8 @@ class AIAgent:
                     "model": self.model,
                     "messages": api_messages,
                 }
+                if _summary_temperature is not None:
+                    summary_kwargs["temperature"] = _summary_temperature
                 if self.max_tokens is not None:
                     summary_kwargs.update(self._max_tokens_param(self.max_tokens))
 
@@ -8253,6 +8612,8 @@ class AIAgent:
                         "model": self.model,
                         "messages": api_messages,
                     }
+                    if _summary_temperature is not None:
+                        summary_kwargs["temperature"] = _summary_temperature
                     if self.max_tokens is not None:
                         summary_kwargs.update(self._max_tokens_param(self.max_tokens))
                     if summary_extra_body:
@@ -8688,6 +9049,7 @@ class AIAgent:
                                 {
                                     "name": tc["function"]["name"],
                                     "result": _results_by_id.get(tc.get("id")),
+                                    "arguments": tc["function"].get("arguments"),
                                 }
                                 for tc in _m["tool_calls"]
                                 if isinstance(tc, dict)
@@ -9302,8 +9664,7 @@ class AIAgent:
                                 "and had none left for the actual response.\n\n"
                                 "To fix this:\n"
                                 "→ Lower reasoning effort: `/thinkon low` or `/thinkon minimal`\n"
-                                "→ Increase the output token limit: "
-                                "set `model.max_tokens` in config.yaml"
+                                "→ Or switch to a larger/non-reasoning model with `/model`"
                             )
                             self._cleanup_task_resources(effective_task_id)
                             self._persist_session(messages, conversation_history)
@@ -9570,13 +9931,51 @@ class AIAgent:
                     if isinstance(api_error, UnicodeEncodeError) and getattr(self, '_unicode_sanitization_passes', 0) < 2:
                         _err_str = str(api_error).lower()
                         _is_ascii_codec = "'ascii'" in _err_str or "ascii" in _err_str
+                        # Detect surrogate errors — utf-8 codec refusing to
+                        # encode U+D800..U+DFFF.  The error text is:
+                        #   "'utf-8' codec can't encode characters in position
+                        #    N-M: surrogates not allowed"
+                        _is_surrogate_error = (
+                            "surrogate" in _err_str
+                            or ("'utf-8'" in _err_str and not _is_ascii_codec)
+                        )
+                        # Sanitize surrogates from both the canonical `messages`
+                        # list AND `api_messages` (the API-copy, which may carry
+                        # `reasoning_content`/`reasoning_details` transformed
+                        # from `reasoning` — fields the canonical list doesn't
+                        # have directly).  Also clean `api_kwargs` if built and
+                        # `prefill_messages` if present.  Mirrors the ASCII
+                        # codec recovery below.
                         _surrogates_found = _sanitize_messages_surrogates(messages)
-                        if _surrogates_found:
+                        if isinstance(api_messages, list):
+                            if _sanitize_messages_surrogates(api_messages):
+                                _surrogates_found = True
+                        if isinstance(api_kwargs, dict):
+                            if _sanitize_structure_surrogates(api_kwargs):
+                                _surrogates_found = True
+                        if isinstance(getattr(self, "prefill_messages", None), list):
+                            if _sanitize_messages_surrogates(self.prefill_messages):
+                                _surrogates_found = True
+                        # Gate the retry on the error type, not on whether we
+                        # found anything — _force_ascii_payload / the extended
+                        # surrogate walker above cover all known paths, but a
+                        # new transformed field could still slip through.  If
+                        # the error was a surrogate encode failure, always let
+                        # the retry run; the proactive sanitizer at line ~8781
+                        # runs again on the next iteration.  Bounded by
+                        # _unicode_sanitization_passes < 2 (outer guard).
+                        if _surrogates_found or _is_surrogate_error:
                             self._unicode_sanitization_passes += 1
-                            self._vprint(
-                                f"{self.log_prefix}⚠️  Stripped invalid surrogate characters from messages. Retrying...",
-                                force=True,
-                            )
+                            if _surrogates_found:
+                                self._vprint(
+                                    f"{self.log_prefix}⚠️  Stripped invalid surrogate characters from messages. Retrying...",
+                                    force=True,
+                                )
+                            else:
+                                self._vprint(
+                                    f"{self.log_prefix}⚠️  Surrogate encoding error — retrying after full-payload sanitization...",
+                                    force=True,
+                                )
                             continue
                         if _is_ascii_codec:
                             self._force_ascii_payload = True
@@ -10344,9 +10743,9 @@ class AIAgent:
                                     pass
                     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})...")
+                        self._emit_status(f"⏱️ Rate limited. Waiting {wait_time:.1f}s (attempt {retry_count + 1}/{max_retries})...")
                     else:
-                        self._emit_status(f"⏳ Retrying in {wait_time}s (attempt {retry_count}/{max_retries})...")
+                        self._emit_status(f"⏳ Retrying in {wait_time:.1f}s (attempt {retry_count}/{max_retries})...")
                     logger.warning(
                         "Retrying API call in %ss (attempt %s/%s) %s error=%s",
                         wait_time,
@@ -10762,7 +11161,14 @@ class AIAgent:
                         elif self.quiet_mode:
                             clean = self._strip_think_blocks(turn_content).strip()
                             if clean:
-                                self._vprint(f"  ┊ 💬 {clean}")
+                                relayed = False
+                                if (
+                                    self.tool_progress_callback
+                                    and getattr(self, "platform", "") == "tui"
+                                ):
+                                    relayed = True
+                                if not relayed:
+                                    self._vprint(f"  ┊ 💬 {clean}")
                     
                     # Pop thinking-only prefill message(s) before appending
                     # (tool-call path — same rationale as the final-response path).
@@ -11350,6 +11756,12 @@ class AIAgent:
             "cost_status": self.session_cost_status,
             "cost_source": self.session_cost_source,
         }
+        # If a /steer landed after the final assistant turn (no more tool
+        # batches to drain into), hand it back to the caller so it can be
+        # delivered as the next user turn instead of being silently lost.
+        _leftover_steer = self._drain_pending_steer()
+        if _leftover_steer:
+            result["pending_steer"] = _leftover_steer
         self._response_was_previewed = False
         
         # Include interrupt message if one triggered the interrupt

scripts/install.ps1

@@ -721,6 +721,20 @@ function Install-NodeDeps {
         }
     }
     
+    # Install TUI dependencies
+    $tuiDir = "$InstallDir\ui-tui"
+    if (Test-Path "$tuiDir\package.json") {
+        Write-Info "Installing TUI dependencies..."
+        Push-Location $tuiDir
+        try {
+            npm install --silent 2>&1 | Out-Null
+            Write-Success "TUI dependencies installed"
+        } catch {
+            Write-Warn "TUI npm install failed (hermes --tui may not work)"
+        }
+        Pop-Location
+    }
+
     # Install WhatsApp bridge dependencies
     $bridgeDir = "$InstallDir\scripts\whatsapp-bridge"
     if (Test-Path "$bridgeDir\package.json") {

scripts/install.sh

@@ -1194,6 +1194,16 @@ install_node_deps() {
         log_success "Browser engine setup complete"
     fi
 
+    # Install TUI dependencies
+    if [ -f "$INSTALL_DIR/ui-tui/package.json" ]; then
+        log_info "Installing TUI dependencies..."
+        cd "$INSTALL_DIR/ui-tui"
+        npm install --silent 2>/dev/null || {
+            log_warn "TUI npm install failed (hermes --tui may not work)"
+        }
+        log_success "TUI dependencies installed"
+    fi
+
     # Install WhatsApp bridge dependencies
     if [ -f "$INSTALL_DIR/scripts/whatsapp-bridge/package.json" ]; then
         log_info "Installing WhatsApp bridge dependencies..."

scripts/lib/node-bootstrap.sh

@@ -0,0 +1,238 @@
+#!/usr/bin/env bash
+# ============================================================================
+# scripts/lib/node-bootstrap.sh
+# ----------------------------------------------------------------------------
+# Sourceable helper: ensure Node.js >= MIN_VERSION is available for the TUI
+# (React + Ink), browser tools, and the WhatsApp bridge.
+#
+# Strategy (first hit wins — respects the user's existing tooling):
+#   1. modern `node` already on PATH
+#   2. ~/.hermes/node/ from a prior Hermes-managed install
+#   3. fnm, proto, nvm (in that order) if the user already uses a version manager
+#   4. Termux `pkg`, macOS Homebrew
+#   5. pinned nodejs.org tarball into ~/.hermes/node/ (always works, zero shell rc edits)
+#
+# Usage:
+#   source scripts/lib/node-bootstrap.sh
+#   ensure_node   # returns 0 on success, non-zero on failure
+#   if [ "$HERMES_NODE_AVAILABLE" = true ]; then ...; fi
+#
+# Env inputs (set before sourcing to override defaults):
+#   HERMES_NODE_MIN_VERSION   (default: 20)   — accepted on PATH
+#   HERMES_NODE_TARGET_MAJOR  (default: 22)   — installed when we install
+#   HERMES_HOME               (default: $HOME/.hermes)
+# ============================================================================
+
+HERMES_NODE_MIN_VERSION="${HERMES_NODE_MIN_VERSION:-20}"
+HERMES_NODE_TARGET_MAJOR="${HERMES_NODE_TARGET_MAJOR:-22}"
+HERMES_HOME="${HERMES_HOME:-$HOME/.hermes}"
+HERMES_NODE_AVAILABLE=false
+
+# ---------------------------------------------------------------------------
+# Logging — prefer the host script's log_* helpers when present
+# ---------------------------------------------------------------------------
+
+_nb_log()  { declare -F log_info    >/dev/null 2>&1 && log_info    "$*" || printf '→ %s\n' "$*" >&2; }
+_nb_ok()   { declare -F log_success >/dev/null 2>&1 && log_success "$*" || printf '✓ %s\n' "$*" >&2; }
+_nb_warn() { declare -F log_warn    >/dev/null 2>&1 && log_warn    "$*" || printf '⚠ %s\n' "$*" >&2; }
+
+# ---------------------------------------------------------------------------
+# Platform + version helpers
+# ---------------------------------------------------------------------------
+
+_nb_is_termux() {
+    [ -n "${TERMUX_VERSION:-}" ] || [[ "${PREFIX:-}" == *"com.termux/files/usr"* ]]
+}
+
+_nb_node_major() {
+    local v
+    v=$(node --version 2>/dev/null | sed 's/^v//' | cut -d. -f1)
+    [[ "$v" =~ ^[0-9]+$ ]] && echo "$v" || echo 0
+}
+
+_nb_have_modern_node() {
+    command -v node >/dev/null 2>&1 || return 1
+    [ "$(_nb_node_major)" -ge "$HERMES_NODE_MIN_VERSION" ]
+}
+
+# ---------------------------------------------------------------------------
+# Version-manager paths — respect what the user already uses
+# ---------------------------------------------------------------------------
+
+_nb_try_fnm() {
+    command -v fnm >/dev/null 2>&1 || return 1
+    _nb_log "fnm detected — installing Node $HERMES_NODE_TARGET_MAJOR..."
+    eval "$(fnm env 2>/dev/null)" || true
+    fnm install "$HERMES_NODE_TARGET_MAJOR" >/dev/null 2>&1 || return 1
+    fnm use     "$HERMES_NODE_TARGET_MAJOR" >/dev/null 2>&1 || return 1
+    _nb_have_modern_node || return 1
+    _nb_ok "Node $(node --version) activated via fnm"
+    return 0
+}
+
+_nb_try_proto() {
+    command -v proto >/dev/null 2>&1 || return 1
+    _nb_log "proto detected — installing Node $HERMES_NODE_TARGET_MAJOR..."
+    proto install node "$HERMES_NODE_TARGET_MAJOR" >/dev/null 2>&1 || return 1
+    _nb_have_modern_node || return 1
+    _nb_ok "Node $(node --version) activated via proto"
+    return 0
+}
+
+_nb_try_nvm() {
+    local nvm_sh="${NVM_DIR:-$HOME/.nvm}/nvm.sh"
+    [ -s "$nvm_sh" ] || return 1
+    # shellcheck source=/dev/null
+    \. "$nvm_sh" >/dev/null 2>&1 || return 1
+    _nb_log "nvm detected — installing Node $HERMES_NODE_TARGET_MAJOR..."
+    nvm install "$HERMES_NODE_TARGET_MAJOR" >/dev/null 2>&1 || return 1
+    nvm use     "$HERMES_NODE_TARGET_MAJOR" >/dev/null 2>&1 || return 1
+    _nb_have_modern_node || return 1
+    _nb_ok "Node $(node --version) activated via nvm"
+    return 0
+}
+
+# ---------------------------------------------------------------------------
+# Platform package managers
+# ---------------------------------------------------------------------------
+
+_nb_try_termux_pkg() {
+    _nb_is_termux || return 1
+    _nb_log "Installing Node.js via pkg..."
+    pkg install -y nodejs >/dev/null 2>&1 || return 1
+    _nb_have_modern_node || return 1
+    _nb_ok "Node $(node --version) installed via pkg"
+    return 0
+}
+
+_nb_try_brew() {
+    [ "$(uname -s)" = "Darwin" ] || return 1
+    command -v brew >/dev/null 2>&1 || return 1
+    _nb_log "Installing Node via Homebrew..."
+    brew install "node@${HERMES_NODE_TARGET_MAJOR}" >/dev/null 2>&1 \
+        || brew install node >/dev/null 2>&1 \
+        || return 1
+    brew link --overwrite --force "node@${HERMES_NODE_TARGET_MAJOR}" >/dev/null 2>&1 || true
+    _nb_have_modern_node || return 1
+    _nb_ok "Node $(node --version) installed via Homebrew"
+    return 0
+}
+
+# ---------------------------------------------------------------------------
+# Bundled binary fallback — always works, no shell rc edits
+# ---------------------------------------------------------------------------
+
+_nb_install_bundled_node() {
+    local arch node_arch os_name node_os
+    arch=$(uname -m)
+    case "$arch" in
+        x86_64)        node_arch="x64"    ;;
+        aarch64|arm64) node_arch="arm64"  ;;
+        armv7l)        node_arch="armv7l" ;;
+        *)
+            _nb_warn "Unsupported arch ($arch) — install Node.js manually: https://nodejs.org/"
+            return 1
+            ;;
+    esac
+
+    os_name=$(uname -s)
+    case "$os_name" in
+        Linux*)  node_os="linux"  ;;
+        Darwin*) node_os="darwin" ;;
+        *)
+            _nb_warn "Unsupported OS ($os_name) — install Node.js manually: https://nodejs.org/"
+            return 1
+            ;;
+    esac
+
+    local index_url="https://nodejs.org/dist/latest-v${HERMES_NODE_TARGET_MAJOR}.x/"
+    local tarball
+    tarball=$(curl -fsSL "$index_url" \
+        | grep -oE "node-v${HERMES_NODE_TARGET_MAJOR}\.[0-9]+\.[0-9]+-${node_os}-${node_arch}\.tar\.xz" \
+        | head -1)
+    if [ -z "$tarball" ]; then
+        tarball=$(curl -fsSL "$index_url" \
+            | grep -oE "node-v${HERMES_NODE_TARGET_MAJOR}\.[0-9]+\.[0-9]+-${node_os}-${node_arch}\.tar\.gz" \
+            | head -1)
+    fi
+    if [ -z "$tarball" ]; then
+        _nb_warn "Could not resolve Node $HERMES_NODE_TARGET_MAJOR binary for $node_os-$node_arch"
+        return 1
+    fi
+
+    local tmp
+    tmp=$(mktemp -d)
+    _nb_log "Downloading $tarball..."
+    curl -fsSL "${index_url}${tarball}" -o "$tmp/$tarball" || {
+        _nb_warn "Download failed"; rm -rf "$tmp"; return 1
+    }
+
+    _nb_log "Extracting to $HERMES_HOME/node/..."
+    if [[ "$tarball" == *.tar.xz ]]; then
+        tar xf  "$tmp/$tarball" -C "$tmp" || { rm -rf "$tmp"; return 1; }
+    else
+        tar xzf "$tmp/$tarball" -C "$tmp" || { rm -rf "$tmp"; return 1; }
+    fi
+
+    local extracted
+    extracted=$(find "$tmp" -maxdepth 1 -type d -name 'node-v*' 2>/dev/null | head -1)
+    if [ ! -d "$extracted" ]; then
+        _nb_warn "Extraction produced no node-v* directory"
+        rm -rf "$tmp"
+        return 1
+    fi
+
+    mkdir -p "$HERMES_HOME"
+    rm -rf "$HERMES_HOME/node"
+    mv "$extracted" "$HERMES_HOME/node"
+    rm -rf "$tmp"
+
+    mkdir -p "$HOME/.local/bin"
+    ln -sf "$HERMES_HOME/node/bin/node" "$HOME/.local/bin/node"
+    ln -sf "$HERMES_HOME/node/bin/npm"  "$HOME/.local/bin/npm"
+    ln -sf "$HERMES_HOME/node/bin/npx"  "$HOME/.local/bin/npx"
+    export PATH="$HERMES_HOME/node/bin:$PATH"
+
+    _nb_have_modern_node || return 1
+    _nb_ok "Node $(node --version) installed to $HERMES_HOME/node/"
+    return 0
+}
+
+# ---------------------------------------------------------------------------
+# Public entry point
+# ---------------------------------------------------------------------------
+
+ensure_node() {
+    HERMES_NODE_AVAILABLE=false
+
+    if _nb_have_modern_node; then
+        _nb_ok "Node $(node --version) found"
+        HERMES_NODE_AVAILABLE=true
+        return 0
+    fi
+
+    if [ -x "$HERMES_HOME/node/bin/node" ]; then
+        export PATH="$HERMES_HOME/node/bin:$PATH"
+        if _nb_have_modern_node; then
+            _nb_ok "Node $(node --version) found (Hermes-managed)"
+            HERMES_NODE_AVAILABLE=true
+            return 0
+        fi
+    fi
+
+    # Version managers first — respect the user's existing setup.
+    _nb_try_fnm   && { HERMES_NODE_AVAILABLE=true; return 0; }
+    _nb_try_proto && { HERMES_NODE_AVAILABLE=true; return 0; }
+    _nb_try_nvm   && { HERMES_NODE_AVAILABLE=true; return 0; }
+
+    # Platform package managers.
+    _nb_try_termux_pkg && { HERMES_NODE_AVAILABLE=true; return 0; }
+    _nb_try_brew       && { HERMES_NODE_AVAILABLE=true; return 0; }
+
+    # Last resort: pinned nodejs.org tarball.
+    _nb_install_bundled_node && { HERMES_NODE_AVAILABLE=true; return 0; }
+
+    _nb_warn "Node.js install failed — TUI and browser tools will be unavailable."
+    _nb_warn "Install manually: https://nodejs.org/en/download/  (or: \`brew install node\`, \`fnm install $HERMES_NODE_TARGET_MAJOR\`, etc.)"
+    return 1
+}

scripts/release.py

@@ -44,6 +44,7 @@ AUTHOR_MAP = {
     "teknium@nousresearch.com": "teknium1",
     "127238744+teknium1@users.noreply.github.com": "teknium1",
     # contributors (from noreply pattern)
+    "snreynolds2506@gmail.com": "snreynolds",
     "35742124+0xbyt4@users.noreply.github.com": "0xbyt4",
     "82637225+kshitijk4poor@users.noreply.github.com": "kshitijk4poor",
     "kshitijk4poor@users.noreply.github.com": "kshitijk4poor",
@@ -75,6 +76,7 @@ AUTHOR_MAP = {
     "Asunfly@users.noreply.github.com": "Asunfly",
     # contributors (manual mapping from git names)
     "ahmedsherif95@gmail.com": "asheriif",
+    "liujinkun@bytedance.com": "liujinkun2025",
     "dmayhem93@gmail.com": "dmahan93",
     "samherring99@gmail.com": "samherring99",
     "desaiaum08@gmail.com": "Aum08Desai",
@@ -95,6 +97,7 @@ AUTHOR_MAP = {
     "4317663+helix4u@users.noreply.github.com": "helix4u",
     "331214+counterposition@users.noreply.github.com": "counterposition",
     "blspear@gmail.com": "BrennerSpear",
+    "akhater@gmail.com": "akhater",
     "239876380+handsdiff@users.noreply.github.com": "handsdiff",
     "gpickett00@gmail.com": "gpickett00",
     "mcosma@gmail.com": "wakamex",
@@ -103,6 +106,7 @@ AUTHOR_MAP = {
     "dangtc94@gmail.com": "dieutx",
     "jaisehgal11299@gmail.com": "jaisup",
     "percydikec@gmail.com": "PercyDikec",
+    "noonou7@gmail.com": "HenkDz",
     "dean.kerr@gmail.com": "deankerr",
     "socrates1024@gmail.com": "socrates1024",
     "satelerd@gmail.com": "satelerd",
@@ -115,6 +119,7 @@ AUTHOR_MAP = {
     "vincentcharlebois@gmail.com": "vincentcharlebois",
     "aryan@synvoid.com": "aryansingh",
     "johnsonblake1@gmail.com": "blakejohnson",
+    "hcn518@gmail.com": "pedh",
     "greer.guthrie@gmail.com": "g-guthrie",
     "kennyx102@gmail.com": "bobashopcashier",
     "shokatalishaikh95@gmail.com": "areu01or00",
@@ -202,6 +207,7 @@ AUTHOR_MAP = {
     "cola-runner@users.noreply.github.com": "cola-runner",
     "ygd58@users.noreply.github.com": "ygd58",
     "vominh1919@users.noreply.github.com": "vominh1919",
+    "iamagenius00@users.noreply.github.com": "iamagenius00",
     "trevmanthony@gmail.com": "trevthefoolish",
     "ziliangpeng@users.noreply.github.com": "ziliangpeng",
     "centripetal-star@users.noreply.github.com": "centripetal-star",
@@ -255,6 +261,9 @@ AUTHOR_MAP = {
     "anthhub@163.com": "anthhub",
     "shenuu@gmail.com": "shenuu",
     "xiayh17@gmail.com": "xiayh0107",
+    "asurla@nvidia.com": "anniesurla",
+    "limkuan24@gmail.com": "WideLee",
+    "aviralarora002@gmail.com": "AviArora02-commits",
 }
 
 

skills/mlops/inference/gguf/SKILL.md

@@ -1,430 +0,0 @@
----
-name: gguf-quantization
-description: GGUF format and llama.cpp quantization for efficient CPU/GPU inference. Use when deploying models on consumer hardware, Apple Silicon, or when needing flexible quantization from 2-8 bit without GPU requirements.
-version: 1.0.0
-author: Orchestra Research
-license: MIT
-dependencies: [llama-cpp-python>=0.2.0]
-metadata:
-  hermes:
-    tags: [GGUF, Quantization, llama.cpp, CPU Inference, Apple Silicon, Model Compression, Optimization]
-
----
-
-# GGUF - Quantization Format for llama.cpp
-
-The GGUF (GPT-Generated Unified Format) is the standard file format for llama.cpp, enabling efficient inference on CPUs, Apple Silicon, and GPUs with flexible quantization options.
-
-## When to use GGUF
-
-**Use GGUF when:**
-- Deploying on consumer hardware (laptops, desktops)
-- Running on Apple Silicon (M1/M2/M3) with Metal acceleration
-- Need CPU inference without GPU requirements
-- Want flexible quantization (Q2_K to Q8_0)
-- Using local AI tools (LM Studio, Ollama, text-generation-webui)
-
-**Key advantages:**
-- **Universal hardware**: CPU, Apple Silicon, NVIDIA, AMD support
-- **No Python runtime**: Pure C/C++ inference
-- **Flexible quantization**: 2-8 bit with various methods (K-quants)
-- **Ecosystem support**: LM Studio, Ollama, koboldcpp, and more
-- **imatrix**: Importance matrix for better low-bit quality
-
-**Use alternatives instead:**
-- **AWQ/GPTQ**: Maximum accuracy with calibration on NVIDIA GPUs
-- **HQQ**: Fast calibration-free quantization for HuggingFace
-- **bitsandbytes**: Simple integration with transformers library
-- **TensorRT-LLM**: Production NVIDIA deployment with maximum speed
-
-## Quick start
-
-### Installation
-
-```bash
-# Clone llama.cpp
-git clone https://github.com/ggml-org/llama.cpp
-cd llama.cpp
-
-# Build (CPU)
-make
-
-# Build with CUDA (NVIDIA)
-make GGML_CUDA=1
-
-# Build with Metal (Apple Silicon)
-make GGML_METAL=1
-
-# Install Python bindings (optional)
-pip install llama-cpp-python
-```
-
-### Convert model to GGUF
-
-```bash
-# Install requirements
-pip install -r requirements.txt
-
-# Convert HuggingFace model to GGUF (FP16)
-python convert_hf_to_gguf.py ./path/to/model --outfile model-f16.gguf
-
-# Or specify output type
-python convert_hf_to_gguf.py ./path/to/model \
-    --outfile model-f16.gguf \
-    --outtype f16
-```
-
-### Quantize model
-
-```bash
-# Basic quantization to Q4_K_M
-./llama-quantize model-f16.gguf model-q4_k_m.gguf Q4_K_M
-
-# Quantize with importance matrix (better quality)
-./llama-imatrix -m model-f16.gguf -f calibration.txt -o model.imatrix
-./llama-quantize --imatrix model.imatrix model-f16.gguf model-q4_k_m.gguf Q4_K_M
-```
-
-### Run inference
-
-```bash
-# CLI inference
-./llama-cli -m model-q4_k_m.gguf -p "Hello, how are you?"
-
-# Interactive mode
-./llama-cli -m model-q4_k_m.gguf --interactive
-
-# With GPU offload
-./llama-cli -m model-q4_k_m.gguf -ngl 35 -p "Hello!"
-```
-
-## Quantization types
-
-### K-quant methods (recommended)
-
-| Type | Bits | Size (7B) | Quality | Use Case |
-|------|------|-----------|---------|----------|
-| Q2_K | 2.5 | ~2.8 GB | Low | Extreme compression |
-| Q3_K_S | 3.0 | ~3.0 GB | Low-Med | Memory constrained |
-| Q3_K_M | 3.3 | ~3.3 GB | Medium | Balance |
-| Q4_K_S | 4.0 | ~3.8 GB | Med-High | Good balance |
-| Q4_K_M | 4.5 | ~4.1 GB | High | **Recommended default** |
-| Q5_K_S | 5.0 | ~4.6 GB | High | Quality focused |
-| Q5_K_M | 5.5 | ~4.8 GB | Very High | High quality |
-| Q6_K | 6.0 | ~5.5 GB | Excellent | Near-original |
-| Q8_0 | 8.0 | ~7.2 GB | Best | Maximum quality |
-
-### Legacy methods
-
-| Type | Description |
-|------|-------------|
-| Q4_0 | 4-bit, basic |
-| Q4_1 | 4-bit with delta |
-| Q5_0 | 5-bit, basic |
-| Q5_1 | 5-bit with delta |
-
-**Recommendation**: Use K-quant methods (Q4_K_M, Q5_K_M) for best quality/size ratio.
-
-## Conversion workflows
-
-### Workflow 1: HuggingFace to GGUF
-
-```bash
-# 1. Download model
-huggingface-cli download meta-llama/Llama-3.1-8B --local-dir ./llama-3.1-8b
-
-# 2. Convert to GGUF (FP16)
-python convert_hf_to_gguf.py ./llama-3.1-8b \
-    --outfile llama-3.1-8b-f16.gguf \
-    --outtype f16
-
-# 3. Quantize
-./llama-quantize llama-3.1-8b-f16.gguf llama-3.1-8b-q4_k_m.gguf Q4_K_M
-
-# 4. Test
-./llama-cli -m llama-3.1-8b-q4_k_m.gguf -p "Hello!" -n 50
-```
-
-### Workflow 2: With importance matrix (better quality)
-
-```bash
-# 1. Convert to GGUF
-python convert_hf_to_gguf.py ./model --outfile model-f16.gguf
-
-# 2. Create calibration text (diverse samples)
-cat > calibration.txt << 'EOF'
-The quick brown fox jumps over the lazy dog.
-Machine learning is a subset of artificial intelligence.
-Python is a popular programming language.
-# Add more diverse text samples...
-EOF
-
-# 3. Generate importance matrix
-./llama-imatrix -m model-f16.gguf \
-    -f calibration.txt \
-    --chunk 512 \
-    -o model.imatrix \
-    -ngl 35  # GPU layers if available
-
-# 4. Quantize with imatrix
-./llama-quantize --imatrix model.imatrix \
-    model-f16.gguf \
-    model-q4_k_m.gguf \
-    Q4_K_M
-```
-
-### Workflow 3: Multiple quantizations
-
-```bash
-#!/bin/bash
-MODEL="llama-3.1-8b-f16.gguf"
-IMATRIX="llama-3.1-8b.imatrix"
-
-# Generate imatrix once
-./llama-imatrix -m $MODEL -f wiki.txt -o $IMATRIX -ngl 35
-
-# Create multiple quantizations
-for QUANT in Q4_K_M Q5_K_M Q6_K Q8_0; do
-    OUTPUT="llama-3.1-8b-${QUANT,,}.gguf"
-    ./llama-quantize --imatrix $IMATRIX $MODEL $OUTPUT $QUANT
-    echo "Created: $OUTPUT ($(du -h $OUTPUT | cut -f1))"
-done
-```
-
-## Python usage
-
-### llama-cpp-python
-
-```python
-from llama_cpp import Llama
-
-# Load model
-llm = Llama(
-    model_path="./model-q4_k_m.gguf",
-    n_ctx=4096,          # Context window
-    n_gpu_layers=35,     # GPU offload (0 for CPU only)
-    n_threads=8          # CPU threads
-)
-
-# Generate
-output = llm(
-    "What is machine learning?",
-    max_tokens=256,
-    temperature=0.7,
-    stop=["</s>", "\n\n"]
-)
-print(output["choices"][0]["text"])
-```
-
-### Chat completion
-
-```python
-from llama_cpp import Llama
-
-llm = Llama(
-    model_path="./model-q4_k_m.gguf",
-    n_ctx=4096,
-    n_gpu_layers=35,
-    chat_format="llama-3"  # Or "chatml", "mistral", etc.
-)
-
-messages = [
-    {"role": "system", "content": "You are a helpful assistant."},
-    {"role": "user", "content": "What is Python?"}
-]
-
-response = llm.create_chat_completion(
-    messages=messages,
-    max_tokens=256,
-    temperature=0.7
-)
-print(response["choices"][0]["message"]["content"])
-```
-
-### Streaming
-
-```python
-from llama_cpp import Llama
-
-llm = Llama(model_path="./model-q4_k_m.gguf", n_gpu_layers=35)
-
-# Stream tokens
-for chunk in llm(
-    "Explain quantum computing:",
-    max_tokens=256,
-    stream=True
-):
-    print(chunk["choices"][0]["text"], end="", flush=True)
-```
-
-## Server mode
-
-### Start OpenAI-compatible server
-
-```bash
-# Start server
-./llama-server -m model-q4_k_m.gguf \
-    --host 0.0.0.0 \
-    --port 8080 \
-    -ngl 35 \
-    -c 4096
-
-# Or with Python bindings
-python -m llama_cpp.server \
-    --model model-q4_k_m.gguf \
-    --n_gpu_layers 35 \
-    --host 0.0.0.0 \
-    --port 8080
-```
-
-### Use with OpenAI client
-
-```python
-from openai import OpenAI
-
-client = OpenAI(
-    base_url="http://localhost:8080/v1",
-    api_key="not-needed"
-)
-
-response = client.chat.completions.create(
-    model="local-model",
-    messages=[{"role": "user", "content": "Hello!"}],
-    max_tokens=256
-)
-print(response.choices[0].message.content)
-```
-
-## Hardware optimization
-
-### Apple Silicon (Metal)
-
-```bash
-# Build with Metal
-make clean && make GGML_METAL=1
-
-# Run with Metal acceleration
-./llama-cli -m model.gguf -ngl 99 -p "Hello"
-
-# Python with Metal
-llm = Llama(
-    model_path="model.gguf",
-    n_gpu_layers=99,     # Offload all layers
-    n_threads=1          # Metal handles parallelism
-)
-```
-
-### NVIDIA CUDA
-
-```bash
-# Build with CUDA
-make clean && make GGML_CUDA=1
-
-# Run with CUDA
-./llama-cli -m model.gguf -ngl 35 -p "Hello"
-
-# Specify GPU
-CUDA_VISIBLE_DEVICES=0 ./llama-cli -m model.gguf -ngl 35
-```
-
-### CPU optimization
-
-```bash
-# Build with AVX2/AVX512
-make clean && make
-
-# Run with optimal threads
-./llama-cli -m model.gguf -t 8 -p "Hello"
-
-# Python CPU config
-llm = Llama(
-    model_path="model.gguf",
-    n_gpu_layers=0,      # CPU only
-    n_threads=8,         # Match physical cores
-    n_batch=512          # Batch size for prompt processing
-)
-```
-
-## Integration with tools
-
-### Ollama
-
-```bash
-# Create Modelfile
-cat > Modelfile << 'EOF'
-FROM ./model-q4_k_m.gguf
-TEMPLATE """{{ .System }}
-{{ .Prompt }}"""
-PARAMETER temperature 0.7
-PARAMETER num_ctx 4096
-EOF
-
-# Create Ollama model
-ollama create mymodel -f Modelfile
-
-# Run
-ollama run mymodel "Hello!"
-```
-
-### LM Studio
-
-1. Place GGUF file in `~/.cache/lm-studio/models/`
-2. Open LM Studio and select the model
-3. Configure context length and GPU offload
-4. Start inference
-
-### text-generation-webui
-
-```bash
-# Place in models folder
-cp model-q4_k_m.gguf text-generation-webui/models/
-
-# Start with llama.cpp loader
-python server.py --model model-q4_k_m.gguf --loader llama.cpp --n-gpu-layers 35
-```
-
-## Best practices
-
-1. **Use K-quants**: Q4_K_M offers best quality/size balance
-2. **Use imatrix**: Always use importance matrix for Q4 and below
-3. **GPU offload**: Offload as many layers as VRAM allows
-4. **Context length**: Start with 4096, increase if needed
-5. **Thread count**: Match physical CPU cores, not logical
-6. **Batch size**: Increase n_batch for faster prompt processing
-
-## Common issues
-
-**Model loads slowly:**
-```bash
-# Use mmap for faster loading
-./llama-cli -m model.gguf --mmap
-```
-
-**Out of memory:**
-```bash
-# Reduce GPU layers
-./llama-cli -m model.gguf -ngl 20  # Reduce from 35
-
-# Or use smaller quantization
-./llama-quantize model-f16.gguf model-q3_k_m.gguf Q3_K_M
-```
-
-**Poor quality at low bits:**
-```bash
-# Always use imatrix for Q4 and below
-./llama-imatrix -m model-f16.gguf -f calibration.txt -o model.imatrix
-./llama-quantize --imatrix model.imatrix model-f16.gguf model-q4_k_m.gguf Q4_K_M
-```
-
-## References
-
-- **[Advanced Usage](references/advanced-usage.md)** - Batching, speculative decoding, custom builds
-- **[Troubleshooting](references/troubleshooting.md)** - Common issues, debugging, benchmarks
-
-## Resources
-
-- **Repository**: https://github.com/ggml-org/llama.cpp
-- **Python Bindings**: https://github.com/abetlen/llama-cpp-python
-- **Pre-quantized Models**: https://huggingface.co/TheBloke
-- **GGUF Converter**: https://huggingface.co/spaces/ggml-org/gguf-my-repo
-- **License**: MIT

skills/mlops/inference/llama-cpp/SKILL.md

@@ -1,138 +1,271 @@
 ---
 name: llama-cpp
-description: Runs LLM inference on CPU, Apple Silicon, and consumer GPUs without NVIDIA hardware. Use for edge deployment, M1/M2/M3 Macs, AMD/Intel GPUs, or when CUDA is unavailable. Supports GGUF quantization (1.5-8 bit) for reduced memory and 4-10× speedup vs PyTorch on CPU.
-version: 1.0.0
+description: Run LLM inference with llama.cpp on CPU, Apple Silicon, AMD/Intel GPUs, or NVIDIA — plus GGUF model conversion and quantization (2–8 bit with K-quants and imatrix). Covers CLI, Python bindings, OpenAI-compatible server, and Ollama/LM Studio integration. Use for edge deployment, M1/M2/M3/M4 Macs, CUDA-less environments, or flexible local quantization.
+version: 2.0.0
 author: Orchestra Research
 license: MIT
-dependencies: [llama-cpp-python]
+dependencies: [llama-cpp-python>=0.2.0]
 metadata:
   hermes:
-    tags: [Inference Serving, Llama.cpp, CPU Inference, Apple Silicon, Edge Deployment, GGUF, Quantization, Non-NVIDIA, AMD GPUs, Intel GPUs, Embedded]
-
+    tags: [llama.cpp, GGUF, Quantization, CPU Inference, Apple Silicon, Edge Deployment, Non-NVIDIA, AMD GPUs, Intel GPUs, Embedded, Model Compression]
 ---
 
-# llama.cpp
+# llama.cpp + GGUF
 
-Pure C/C++ LLM inference with minimal dependencies, optimized for CPUs and non-NVIDIA hardware.
+Pure C/C++ LLM inference with minimal dependencies, plus the GGUF (GPT-Generated Unified Format) standard used for quantized weights. One toolchain covers conversion, quantization, and serving.
 
-## When to use llama.cpp
+## When to use
 
-**Use llama.cpp when:**
-- Running on CPU-only machines
-- Deploying on Apple Silicon (M1/M2/M3/M4)
-- Using AMD or Intel GPUs (no CUDA)
-- Edge deployment (Raspberry Pi, embedded systems)
-- Need simple deployment without Docker/Python
+**Use llama.cpp + GGUF when:**
+- Running on CPU-only machines or Apple Silicon (M1/M2/M3/M4) with Metal acceleration
+- Using AMD (ROCm) or Intel GPUs where CUDA isn't available
+- Edge deployment (Raspberry Pi, embedded systems, consumer laptops)
+- Need flexible quantization (2–8 bit with K-quants)
+- Want local AI tools (LM Studio, Ollama, text-generation-webui, koboldcpp)
+- Want a single binary deploy without Docker/Python
 
-**Use TensorRT-LLM instead when:**
-- Have NVIDIA GPUs (A100/H100)
-- Need maximum throughput (100K+ tok/s)
-- Running in datacenter with CUDA
+**Key advantages:**
+- Universal hardware: CPU, Apple Silicon, NVIDIA, AMD, Intel
+- No Python runtime required (pure C/C++)
+- K-quants + imatrix for better low-bit quality
+- OpenAI-compatible server built in
+- Rich ecosystem (Ollama, LM Studio, llama-cpp-python)
 
-**Use vLLM instead when:**
-- Have NVIDIA GPUs
-- Need Python-first API
-- Want PagedAttention
+**Use alternatives instead:**
+- **vLLM** — NVIDIA GPUs, PagedAttention, Python-first, max throughput
+- **TensorRT-LLM** — Production NVIDIA (A100/H100), maximum speed
+- **AWQ/GPTQ** — Calibrated quantization for NVIDIA-only deployments
+- **bitsandbytes** — Simple HuggingFace transformers integration
+- **HQQ** — Fast calibration-free quantization
 
 ## Quick start
 
-### Installation
+### Install
 
 ```bash
-# macOS/Linux
+# macOS / Linux (simplest)
 brew install llama.cpp
 
 # Or build from source
-git clone https://github.com/ggerganov/llama.cpp
+git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
-make
+make                        # CPU
+make GGML_METAL=1           # Apple Silicon
+make GGML_CUDA=1            # NVIDIA CUDA
+make LLAMA_HIP=1            # AMD ROCm
 
-# With Metal (Apple Silicon)
-make LLAMA_METAL=1
-
-# With CUDA (NVIDIA)
-make LLAMA_CUDA=1
-
-# With ROCm (AMD)
-make LLAMA_HIP=1
+# Python bindings (optional)
+pip install llama-cpp-python
+# With CUDA:   CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --force-reinstall --no-cache-dir
+# With Metal:  CMAKE_ARGS="-DGGML_METAL=on" pip install llama-cpp-python --force-reinstall --no-cache-dir
 ```
 
-### Download model
+### Download a pre-quantized GGUF
 
 ```bash
-# Download from HuggingFace (GGUF format)
+# TheBloke hosts most popular models pre-quantized
 huggingface-cli download \
     TheBloke/Llama-2-7B-Chat-GGUF \
     llama-2-7b-chat.Q4_K_M.gguf \
     --local-dir models/
+```
 
-# Or convert from HuggingFace
-python convert_hf_to_gguf.py models/llama-2-7b-chat/
+### Or convert a HuggingFace model to GGUF
+
+```bash
+# 1. Download HF model
+huggingface-cli download meta-llama/Llama-3.1-8B --local-dir ./llama-3.1-8b
+
+# 2. Convert to FP16 GGUF
+python convert_hf_to_gguf.py ./llama-3.1-8b \
+    --outfile llama-3.1-8b-f16.gguf \
+    --outtype f16
+
+# 3. Quantize to Q4_K_M
+./llama-quantize llama-3.1-8b-f16.gguf llama-3.1-8b-q4_k_m.gguf Q4_K_M
 ```
 
 ### Run inference
 
 ```bash
-# Simple chat
-./llama-cli \
-    -m models/llama-2-7b-chat.Q4_K_M.gguf \
-    -p "Explain quantum computing" \
-    -n 256  # Max tokens
+# One-shot prompt
+./llama-cli -m model.Q4_K_M.gguf -p "Explain quantum computing" -n 256
 
 # Interactive chat
-./llama-cli \
-    -m models/llama-2-7b-chat.Q4_K_M.gguf \
-    --interactive
+./llama-cli -m model.Q4_K_M.gguf --interactive
+
+# With GPU offload
+./llama-cli -m model.Q4_K_M.gguf -ngl 35 -p "Hello!"
 ```
 
-### Server mode
+### Serve an OpenAI-compatible API
 
 ```bash
-# Start OpenAI-compatible server
 ./llama-server \
-    -m models/llama-2-7b-chat.Q4_K_M.gguf \
+    -m model.Q4_K_M.gguf \
     --host 0.0.0.0 \
     --port 8080 \
-    -ngl 32  # Offload 32 layers to GPU
+    -ngl 35 \
+    -c 4096 \
+    --parallel 4 \
+    --cont-batching
+```
 
-# Client request
+```bash
 curl http://localhost:8080/v1/chat/completions \
   -H "Content-Type: application/json" \
   -d '{
-    "model": "llama-2-7b-chat",
+    "model": "local",
     "messages": [{"role": "user", "content": "Hello!"}],
     "temperature": 0.7,
     "max_tokens": 100
   }'
 ```
 
-## Quantization formats
+## Quantization formats (GGUF)
 
-### GGUF format overview
+### K-quant methods (recommended)
 
-| Format | Bits | Size (7B) | Speed | Quality | Use Case |
-|--------|------|-----------|-------|---------|----------|
-| **Q4_K_M** | 4.5 | 4.1 GB | Fast | Good | **Recommended default** |
-| Q4_K_S | 4.3 | 3.9 GB | Faster | Lower | Speed critical |
-| Q5_K_M | 5.5 | 4.8 GB | Medium | Better | Quality critical |
-| Q6_K | 6.5 | 5.5 GB | Slower | Best | Maximum quality |
-| Q8_0 | 8.0 | 7.0 GB | Slow | Excellent | Minimal degradation |
-| Q2_K | 2.5 | 2.7 GB | Fastest | Poor | Testing only |
+| Type | Bits | Size (7B) | Quality | Use Case |
+|------|------|-----------|---------|----------|
+| Q2_K | 2.5 | ~2.8 GB | Low | Extreme compression (testing only) |
+| Q3_K_S | 3.0 | ~3.0 GB | Low-Med | Memory constrained |
+| Q3_K_M | 3.3 | ~3.3 GB | Medium | Fits small devices |
+| Q4_K_S | 4.0 | ~3.8 GB | Med-High | Speed critical |
+| **Q4_K_M** | 4.5 | ~4.1 GB | High | **Recommended default** |
+| Q5_K_S | 5.0 | ~4.6 GB | High | Quality focused |
+| Q5_K_M | 5.5 | ~4.8 GB | Very High | High quality |
+| Q6_K | 6.0 | ~5.5 GB | Excellent | Near-original |
+| Q8_0 | 8.0 | ~7.2 GB | Best | Maximum quality, minimal degradation |
 
-### Choosing quantization
+**Variant suffixes** — `_S` (Small, faster, lower quality), `_M` (Medium, balanced), `_L` (Large, better quality).
+
+**Legacy (Q4_0/Q4_1/Q5_0/Q5_1) exist** but always prefer K-quants for better quality/size ratio.
+
+**IQ quantization** — ultra-low-bit with importance-aware methods: IQ2_XXS, IQ2_XS, IQ2_S, IQ3_XXS, IQ3_XS, IQ3_S, IQ4_XS. Require `--imatrix`.
+
+**Task-specific defaults:**
+- General chat / assistants: Q4_K_M, or Q5_K_M if RAM allows
+- Code generation: Q5_K_M or Q6_K (higher precision helps)
+- Technical / medical: Q6_K or Q8_0
+- Very large (70B, 405B) on consumer hardware: Q3_K_M or Q4_K_S
+- Raspberry Pi / edge: Q2_K or Q3_K_S
+
+## Conversion workflows
+
+### Basic: HF → GGUF → quantized
 
 ```bash
-# General use (balanced)
-Q4_K_M  # 4-bit, medium quality
+python convert_hf_to_gguf.py ./model --outfile model-f16.gguf --outtype f16
+./llama-quantize model-f16.gguf model-q4_k_m.gguf Q4_K_M
+./llama-cli -m model-q4_k_m.gguf -p "Hello!" -n 50
+```
 
-# Maximum speed (more degradation)
-Q2_K or Q3_K_M
+### With importance matrix (imatrix) — better low-bit quality
 
-# Maximum quality (slower)
-Q6_K or Q8_0
+`imatrix` gives 10–20% perplexity improvement at Q4, essential at Q3 and below.
 
-# Very large models (70B, 405B)
-Q3_K_M or Q4_K_S  # Lower bits to fit in memory
+```bash
+# 1. Convert to FP16 GGUF
+python convert_hf_to_gguf.py ./model --outfile model-f16.gguf
+
+# 2. Prepare calibration data (diverse text, ~100MB is ideal)
+cat > calibration.txt << 'EOF'
+The quick brown fox jumps over the lazy dog.
+Machine learning is a subset of artificial intelligence.
+# Add more diverse text samples...
+EOF
+
+# 3. Generate importance matrix
+./llama-imatrix -m model-f16.gguf \
+    -f calibration.txt \
+    --chunk 512 \
+    -o model.imatrix \
+    -ngl 35
+
+# 4. Quantize with imatrix
+./llama-quantize --imatrix model.imatrix \
+    model-f16.gguf model-q4_k_m.gguf Q4_K_M
+```
+
+### Multi-quant batch
+
+```bash
+#!/bin/bash
+MODEL="llama-3.1-8b-f16.gguf"
+IMATRIX="llama-3.1-8b.imatrix"
+
+./llama-imatrix -m $MODEL -f wiki.txt -o $IMATRIX -ngl 35
+
+for QUANT in Q4_K_M Q5_K_M Q6_K Q8_0; do
+    OUTPUT="llama-3.1-8b-${QUANT,,}.gguf"
+    ./llama-quantize --imatrix $IMATRIX $MODEL $OUTPUT $QUANT
+    echo "Created: $OUTPUT ($(du -h $OUTPUT | cut -f1))"
+done
+```
+
+### Quality testing (perplexity)
+
+```bash
+./llama-perplexity -m model.gguf -f wikitext-2-raw/wiki.test.raw -c 512
+# Baseline FP16: ~5.96  |  Q4_K_M: ~6.06 (+1.7%)  |  Q2_K: ~6.87 (+15.3%)
+```
+
+## Python bindings (llama-cpp-python)
+
+### Basic generation
+
+```python
+from llama_cpp import Llama
+
+llm = Llama(
+    model_path="./model-q4_k_m.gguf",
+    n_ctx=4096,
+    n_gpu_layers=35,     # 0 for CPU only, 99 to offload everything
+    n_threads=8,
+)
+
+output = llm(
+    "What is machine learning?",
+    max_tokens=256,
+    temperature=0.7,
+    stop=["</s>", "\n\n"],
+)
+print(output["choices"][0]["text"])
+```
+
+### Chat completion + streaming
+
+```python
+llm = Llama(
+    model_path="./model-q4_k_m.gguf",
+    n_ctx=4096,
+    n_gpu_layers=35,
+    chat_format="llama-3",    # Or "chatml", "mistral", etc.
+)
+
+# Non-streaming
+response = llm.create_chat_completion(
+    messages=[
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": "What is Python?"},
+    ],
+    max_tokens=256,
+    temperature=0.7,
+)
+print(response["choices"][0]["message"]["content"])
+
+# Streaming
+for chunk in llm("Explain quantum computing:", max_tokens=256, stream=True):
+    print(chunk["choices"][0]["text"], end="", flush=True)
+```
+
+### Embeddings
+
+```python
+llm = Llama(model_path="./model-q4_k_m.gguf", embedding=True, n_gpu_layers=35)
+vec = llm.embed("This is a test sentence.")
+print(f"Embedding dimension: {len(vec)}")
 ```
 
 ## Hardware acceleration
@@ -140,122 +273,166 @@ Q3_K_M or Q4_K_S  # Lower bits to fit in memory
 ### Apple Silicon (Metal)
 
 ```bash
-# Build with Metal
-make LLAMA_METAL=1
-
-# Run with GPU acceleration (automatic)
-./llama-cli -m model.gguf -ngl 999  # Offload all layers
-
-# Performance: M3 Max 40-60 tokens/sec (Llama 2-7B Q4_K_M)
+make clean && make GGML_METAL=1
+./llama-cli -m model.gguf -ngl 99 -p "Hello"   # offload all layers
 ```
 
-### NVIDIA GPUs (CUDA)
-
-```bash
-# Build with CUDA
-make LLAMA_CUDA=1
-
-# Offload layers to GPU
-./llama-cli -m model.gguf -ngl 35  # Offload 35/40 layers
-
-# Hybrid CPU+GPU for large models
-./llama-cli -m llama-70b.Q4_K_M.gguf -ngl 20  # GPU: 20 layers, CPU: rest
+```python
+llm = Llama(
+    model_path="model.gguf",
+    n_gpu_layers=99,     # Offload everything
+    n_threads=1,         # Metal handles parallelism
+)
 ```
 
-### AMD GPUs (ROCm)
+Performance: M3 Max ~40–60 tok/s on Llama 2-7B Q4_K_M.
+
+### NVIDIA (CUDA)
+
+```bash
+make clean && make GGML_CUDA=1
+./llama-cli -m model.gguf -ngl 35 -p "Hello"
+
+# Hybrid for large models
+./llama-cli -m llama-70b.Q4_K_M.gguf -ngl 20   # GPU: 20 layers, CPU: rest
+
+# Multi-GPU split
+./llama-cli -m large-model.gguf --tensor-split 0.5,0.5 -ngl 60
+```
+
+### AMD (ROCm)
 
 ```bash
-# Build with ROCm
 make LLAMA_HIP=1
-
-# Run with AMD GPU
 ./llama-cli -m model.gguf -ngl 999
 ```
 
-## Common patterns
-
-### Batch processing
+### CPU
 
 ```bash
-# Process multiple prompts from file
-cat prompts.txt | ./llama-cli \
-    -m model.gguf \
-    --batch-size 512 \
-    -n 100
+# Match PHYSICAL cores, not logical
+./llama-cli -m model.gguf -t 8 -p "Hello"
+
+# BLAS acceleration (2–3× speedup)
+make LLAMA_OPENBLAS=1
 ```
 
-### Constrained generation
-
-```bash
-# JSON output with grammar
-./llama-cli \
-    -m model.gguf \
-    -p "Generate a person: " \
-    --grammar-file grammars/json.gbnf
-
-# Outputs valid JSON only
-```
-
-### Context size
-
-```bash
-# Increase context (default 512)
-./llama-cli \
-    -m model.gguf \
-    -c 4096  # 4K context window
-
-# Very long context (if model supports)
-./llama-cli -m model.gguf -c 32768  # 32K context
+```python
+llm = Llama(
+    model_path="model.gguf",
+    n_gpu_layers=0,
+    n_threads=8,
+    n_batch=512,         # Larger batch = faster prompt processing
+)
 ```
 
 ## Performance benchmarks
 
-### CPU performance (Llama 2-7B Q4_K_M)
+### CPU (Llama 2-7B Q4_K_M)
 
-| CPU | Threads | Speed | Cost |
-|-----|---------|-------|------|
-| Apple M3 Max | 16 | 50 tok/s | $0 (local) |
-| AMD Ryzen 9 7950X | 32 | 35 tok/s | $0.50/hour |
-| Intel i9-13900K | 32 | 30 tok/s | $0.40/hour |
-| AWS c7i.16xlarge | 64 | 40 tok/s | $2.88/hour |
+| CPU | Threads | Speed |
+|-----|---------|-------|
+| Apple M3 Max (Metal) | 16 | 50 tok/s |
+| AMD Ryzen 9 7950X | 32 | 35 tok/s |
+| Intel i9-13900K | 32 | 30 tok/s |
 
-### GPU acceleration (Llama 2-7B Q4_K_M)
+### GPU offloading on RTX 4090
 
-| GPU | Speed | vs CPU | Cost |
-|-----|-------|--------|------|
-| NVIDIA RTX 4090 | 120 tok/s | 3-4× | $0 (local) |
-| NVIDIA A10 | 80 tok/s | 2-3× | $1.00/hour |
-| AMD MI250 | 70 tok/s | 2× | $2.00/hour |
-| Apple M3 Max (Metal) | 50 tok/s | ~Same | $0 (local) |
+| Layers GPU | Speed | VRAM |
+|------------|-------|------|
+| 0 (CPU only) | 30 tok/s | 0 GB |
+| 20 (hybrid) | 80 tok/s | 8 GB |
+| 35 (all) | 120 tok/s | 12 GB |
 
 ## Supported models
 
-**LLaMA family**:
-- Llama 2 (7B, 13B, 70B)
-- Llama 3 (8B, 70B, 405B)
-- Code Llama
+- **LLaMA family**: Llama 2 (7B/13B/70B), Llama 3 (8B/70B/405B), Code Llama
+- **Mistral family**: Mistral 7B, Mixtral 8x7B/8x22B
+- **Other**: Falcon, BLOOM, GPT-J, Phi-3, Gemma, Qwen, LLaVA (vision), Whisper (audio)
 
-**Mistral family**:
-- Mistral 7B
-- Mixtral 8x7B, 8x22B
+Find GGUF models: https://huggingface.co/models?library=gguf
 
-**Other**:
-- Falcon, BLOOM, GPT-J
-- Phi-3, Gemma, Qwen
-- LLaVA (vision), Whisper (audio)
+## Ecosystem integrations
 
-**Find models**: https://huggingface.co/models?library=gguf
+### Ollama
+
+```bash
+cat > Modelfile << 'EOF'
+FROM ./model-q4_k_m.gguf
+TEMPLATE """{{ .System }}
+{{ .Prompt }}"""
+PARAMETER temperature 0.7
+PARAMETER num_ctx 4096
+EOF
+
+ollama create mymodel -f Modelfile
+ollama run mymodel "Hello!"
+```
+
+### LM Studio
+
+1. Place GGUF file in `~/.cache/lm-studio/models/`
+2. Open LM Studio and select the model
+3. Configure context length and GPU offload, start inference
+
+### text-generation-webui
+
+```bash
+cp model-q4_k_m.gguf text-generation-webui/models/
+python server.py --model model-q4_k_m.gguf --loader llama.cpp --n-gpu-layers 35
+```
+
+### OpenAI client → llama-server
+
+```python
+from openai import OpenAI
+
+client = OpenAI(base_url="http://localhost:8080/v1", api_key="not-needed")
+response = client.chat.completions.create(
+    model="local-model",
+    messages=[{"role": "user", "content": "Hello!"}],
+    max_tokens=256,
+)
+print(response.choices[0].message.content)
+```
+
+## Best practices
+
+1. **Use K-quants** — Q4_K_M is the recommended default
+2. **Use imatrix** for Q4 and below (calibration improves quality substantially)
+3. **Offload as many layers as VRAM allows** — start high, reduce by 5 on OOM
+4. **Thread count** — match physical cores, not logical
+5. **Batch size** — increase `n_batch` (e.g. 512) for faster prompt processing
+6. **Context** — start at 4096, grow only as needed (memory scales with ctx)
+7. **Flash Attention** — add `--flash-attn` if your build supports it
+
+## Common issues (quick fixes)
+
+**Model loads slowly** — use `--mmap` for memory-mapped loading.
+
+**Out of memory (GPU)** — reduce `-ngl`, use a smaller quant (Q4_K_S / Q3_K_M), or quantize the KV cache:
+```python
+Llama(model_path="...", type_k=2, type_v=2, n_gpu_layers=35)  # Q4_0 KV cache
+```
+
+**Garbage output** — wrong `chat_format`, temperature too high, or model file corrupted. Test with `temperature=0.1` and verify FP16 baseline works.
+
+**Connection refused (server)** — bind to `--host 0.0.0.0`, check `lsof -i :8080`.
+
+See `references/troubleshooting.md` for the full playbook.
 
 ## References
 
-- **[Quantization Guide](references/quantization.md)** - GGUF formats, conversion, quality comparison
-- **[Server Deployment](references/server.md)** - API endpoints, Docker, monitoring
-- **[Optimization](references/optimization.md)** - Performance tuning, hybrid CPU+GPU
+- **[advanced-usage.md](references/advanced-usage.md)** — speculative decoding, batched inference, grammar-constrained generation, LoRA, multi-GPU, custom builds, benchmark scripts
+- **[quantization.md](references/quantization.md)** — perplexity tables, use-case guide, model size scaling (7B/13B/70B RAM needs), imatrix deep dive
+- **[server.md](references/server.md)** — OpenAI API endpoints, Docker deployment, NGINX load balancing, monitoring
+- **[optimization.md](references/optimization.md)** — CPU threading, BLAS, GPU offload heuristics, batch tuning, benchmarks
+- **[troubleshooting.md](references/troubleshooting.md)** — install/convert/quantize/inference/server issues, Apple Silicon, debugging
 
 ## Resources
 
-- **GitHub**: https://github.com/ggerganov/llama.cpp
-- **Models**: https://huggingface.co/models?library=gguf
-- **Discord**: https://discord.gg/llama-cpp
-
-
+- **GitHub**: https://github.com/ggml-org/llama.cpp
+- **Python bindings**: https://github.com/abetlen/llama-cpp-python
+- **Pre-quantized models**: https://huggingface.co/TheBloke
+- **GGUF converter Space**: https://huggingface.co/spaces/ggml-org/gguf-my-repo
+- **License**: MIT

skills/mlops/training/grpo-rl-training/README.md

@@ -1,97 +0,0 @@
-# GRPO/RL Training Skill
-
-**Expert-level guidance for Group Relative Policy Optimization with TRL**
-
-## 📁 Skill Structure
-
-```
-grpo-rl-training/
-├── SKILL.md                              # Main skill documentation (READ THIS FIRST)
-├── README.md                             # This file
-├── templates/
-│   └── basic_grpo_training.py            # Production-ready training template
-└── examples/
-    └── reward_functions_library.py       # 20+ reward function examples
-```
-
-## 🚀 Quick Start
-
-1. **Read SKILL.md** - Comprehensive guide with all concepts and patterns
-2. **Copy `templates/basic_grpo_training.py`** - Start with working code
-3. **Browse `examples/reward_functions_library.py`** - Pick reward functions for your task
-4. **Modify for your use case** - Adapt dataset, rewards, and config
-
-## 💡 What's Inside
-
-### SKILL.md (Main Documentation)
-- Core GRPO concepts and algorithm fundamentals
-- Complete implementation workflow (dataset → rewards → training → deployment)
-- 10+ reward function examples with code
-- Hyperparameter tuning guide
-- Training insights (loss behavior, metrics, debugging)
-- Troubleshooting guide
-- Production best practices
-
-### Templates
-- **basic_grpo_training.py**: Minimal, production-ready training script
-  - Uses Qwen 2.5 1.5B Instruct
-  - 3 reward functions (format + correctness)
-  - LoRA for efficient training
-  - Fully documented and ready to run
-
-### Examples
-- **reward_functions_library.py**: 20+ battle-tested reward functions
-  - Correctness rewards (exact match, fuzzy match, numeric, code execution)
-  - Format rewards (XML, JSON, strict/soft)
-  - Length rewards (ideal length, min/max)
-  - Style rewards (reasoning quality, citations, repetition penalty)
-  - Combined rewards (multi-objective optimization)
-  - Preset collections for common tasks
-
-## 📖 Usage for Agents
-
-When this skill is loaded in your agent's context:
-
-1. **Always read SKILL.md first** before implementing
-2. **Start simple** - Use length-based reward to validate setup
-3. **Build incrementally** - Add one reward function at a time
-4. **Reference examples** - Copy patterns from reward_functions_library.py
-5. **Monitor training** - Watch reward metrics (not loss!)
-
-## 🎯 Common Use Cases
-
-| Task Type | Recommended Rewards | Template |
-|-----------|---------------------|----------|
-| Math reasoning | `MATH_REASONING_REWARDS` preset | basic_grpo_training.py |
-| Code generation | `CODE_GENERATION_REWARDS` preset | Modify dataset in template |
-| Summarization | `SUMMARIZATION_REWARDS` preset | Adjust prompts + rewards |
-| Q&A | `QA_REWARDS` preset | Use fuzzy match + citations |
-
-## ⚠️ Critical Reminders
-
-- **Loss goes UP during training** - This is normal (it's KL divergence)
-- **Use 3-5 reward functions** - Single rewards often fail
-- **Test rewards before training** - Debug each function independently
-- **Monitor reward_std** - Should stay > 0.1 (avoid mode collapse)
-- **Start with num_generations=4-8** - Scale up if GPU allows
-
-## 🔗 External Resources
-
-- [TRL Documentation](https://huggingface.co/docs/trl)
-- [DeepSeek R1 Paper](https://arxiv.org/abs/2501.12948)
-- [Open R1 Implementation](https://github.com/huggingface/open-r1)
-- [Unsloth (2-3x faster)](https://docs.unsloth.ai/)
-
-## 📝 Version
-
-**v1.0.0** - Initial release (January 2025)
-
-## 👨‍💻 Maintained By
-
-Orchestra Research
-For questions or improvements, see https://orchestra.com
-
----
-
-**License:** MIT
-**Last Updated:** January 2025

skills/mlops/training/trl-fine-tuning/SKILL.md

@@ -252,6 +252,8 @@ trl dpo \
 
 Train with reinforcement learning using minimal memory.
 
+For in-depth GRPO guidance — reward function design, critical training insights (loss behavior, mode collapse, tuning), and advanced multi-stage patterns — see **[references/grpo-training.md](references/grpo-training.md)**. A production-ready training script is in **[templates/basic_grpo_training.py](templates/basic_grpo_training.py)**.
+
 Copy this checklist:
 
 ```
@@ -428,6 +430,8 @@ config = PPOConfig(
 
 **Online RL methods**: See [references/online-rl.md](references/online-rl.md) for PPO, GRPO, RLOO, and OnlineDPO with detailed configurations.
 
+**GRPO deep dive**: See [references/grpo-training.md](references/grpo-training.md) for expert-level GRPO patterns — reward function design philosophy, training insights (why loss increases, mode collapse detection), hyperparameter tuning, multi-stage training, and troubleshooting. Production-ready template in [templates/basic_grpo_training.py](templates/basic_grpo_training.py).
+
 ## Hardware requirements
 
 - **GPU**: NVIDIA (CUDA required)

skills/mlops/training/trl-fine-tuning/references/grpo-training.md

@@ -1,51 +1,36 @@
----
-name: grpo-rl-training
-description: Expert guidance for GRPO/RL fine-tuning with TRL for reasoning and task-specific model training
-version: 1.0.0
-author: Orchestra Research
-license: MIT
-dependencies: [transformers>=4.47.0, trl>=0.14.0, datasets>=3.2.0, peft>=0.14.0, torch]
-metadata:
-  hermes:
-    tags: [Post-Training, Reinforcement Learning, GRPO, TRL, RLHF, Reward Modeling, Reasoning, DPO, PPO, Structured Output]
+# GRPO (Group Relative Policy Optimization) — Deep Guide
 
----
+Expert-level patterns, critical insights, and production-ready workflows for fine-tuning language models with custom reward functions using TRL's `GRPOTrainer`. This is the deep reference for the GRPO workflow summarized in the main skill.
 
-# GRPO/RL Training with TRL
+## When to use GRPO
 
-Expert-level guidance for implementing Group Relative Policy Optimization (GRPO) using the Transformer Reinforcement Learning (TRL) library. This skill provides battle-tested patterns, critical insights, and production-ready workflows for fine-tuning language models with custom reward functions.
-
-## When to Use This Skill
-
-Use GRPO training when you need to:
-- **Enforce specific output formats** (e.g., XML tags, JSON, structured reasoning)
+Use GRPO when you need to:
+- **Enforce specific output formats** (XML tags, JSON, structured reasoning)
 - **Teach verifiable tasks** with objective correctness metrics (math, coding, fact-checking)
 - **Improve reasoning capabilities** by rewarding chain-of-thought patterns
 - **Align models to domain-specific behaviors** without labeled preference data
 - **Optimize for multiple objectives** simultaneously (format + correctness + style)
 
 **Do NOT use GRPO for:**
-- Simple supervised fine-tuning tasks (use SFT instead)
+- Simple supervised fine-tuning tasks → use SFT
 - Tasks without clear reward signals
-- When you already have high-quality preference pairs (use DPO/PPO instead)
+- When you already have high-quality preference pairs → use DPO/PPO
 
----
+## Core concepts
 
-## Core Concepts
+### 1. GRPO algorithm fundamentals
 
-### 1. GRPO Algorithm Fundamentals
-
-**Key Mechanism:**
-- Generates **multiple completions** for each prompt (group size: 4-16)
+**Key mechanism:**
+- Generates **multiple completions** per prompt (group size: 4–16)
 - Compares completions within each group using reward functions
 - Updates policy to favor higher-rewarded responses relative to the group
 
-**Critical Difference from PPO:**
+**Critical differences from PPO:**
 - No separate reward model needed
 - More sample-efficient (learns from within-group comparisons)
 - Simpler to implement and debug
 
-**Mathematical Intuition:**
+**Mathematical intuition:**
 ```
 For each prompt p:
   1. Generate N completions: {c₁, c₂, ..., cₙ}
@@ -54,35 +39,32 @@ For each prompt p:
      relative to low-reward ones in the same group
 ```
 
-### 2. Reward Function Design Philosophy
+### 2. Reward function design philosophy
 
-**Golden Rules:**
-1. **Compose multiple reward functions** - Each handles one aspect (format, correctness, style)
-2. **Scale rewards appropriately** - Higher weight = stronger signal
-3. **Use incremental rewards** - Partial credit for partial compliance
-4. **Test rewards independently** - Debug each reward function in isolation
+**Golden rules:**
+1. **Compose multiple reward functions** — each handles one aspect (format, correctness, style)
+2. **Scale rewards appropriately** — higher weight = stronger signal
+3. **Use incremental rewards** — partial credit for partial compliance
+4. **Test rewards independently** — debug each reward function in isolation
 
-**Reward Function Types:**
+**Reward function types:**
 
 | Type | Use Case | Example Weight |
 |------|----------|----------------|
 | **Correctness** | Verifiable tasks (math, code) | 2.0 (highest) |
-| **Format** | Strict structure enforcement | 0.5-1.0 |
-| **Length** | Encourage verbosity/conciseness | 0.1-0.5 |
-| **Style** | Penalize unwanted patterns | -0.5 to 0.5 |
+| **Format** | Strict structure enforcement | 0.5–1.0 |
+| **Length** | Encourage verbosity/conciseness | 0.1–0.5 |
+| **Style** | Penalize unwanted patterns | −0.5 to 0.5 |
 
----
+## Implementation workflow
 
-## Implementation Workflow
+### Step 1: Dataset preparation
 
-### Step 1: Dataset Preparation
-
-**Critical Requirements:**
-- Prompts in chat format (list of dicts with 'role' and 'content')
+**Critical requirements:**
+- Prompts in chat format (list of dicts with `role` and `content`)
 - Include system prompts to set expectations
 - For verifiable tasks, include ground truth answers as additional columns
 
-**Example Structure:**
 ```python
 from datasets import load_dataset, Dataset
 
@@ -97,8 +79,7 @@ Respond in the following format:
 """
 
 def prepare_dataset(raw_data):
-    """
-    Transform raw data into GRPO-compatible format.
+    """Transform raw data into GRPO-compatible format.
 
     Returns: Dataset with columns:
     - 'prompt': List[Dict] with role/content (system + user messages)
@@ -113,14 +94,14 @@ def prepare_dataset(raw_data):
     })
 ```
 
-**Pro Tips:**
-- Use one-shot or few-shot examples in system prompt for complex formats
-- Keep prompts concise (max_prompt_length: 256-512 tokens)
+**Pro tips:**
+- Use one-shot or few-shot examples in the system prompt for complex formats
+- Keep prompts concise (max_prompt_length: 256–512 tokens)
 - Validate data quality before training (garbage in = garbage out)
 
-### Step 2: Reward Function Implementation
+### Step 2: Reward function implementation
 
-**Template Structure:**
+**Template structure:**
 ```python
 def reward_function_name(
     prompts,        # List[List[Dict]]: Original prompts
@@ -128,24 +109,16 @@ def reward_function_name(
     answer=None,    # Optional: Ground truth from dataset
     **kwargs        # Additional dataset columns
 ) -> list[float]:
-    """
-    Evaluate completions and return rewards.
-
-    Returns: List of floats (one per completion)
-    """
-    # Extract completion text
+    """Evaluate completions and return rewards (one per completion)."""
     responses = [comp[0]['content'] for comp in completions]
-
-    # Compute rewards
     rewards = []
     for response in responses:
         score = compute_score(response)
         rewards.append(score)
-
     return rewards
 ```
 
-**Example 1: Correctness Reward (Math/Coding)**
+**Example 1: correctness reward (math/coding)**
 ```python
 def correctness_reward(prompts, completions, answer, **kwargs):
     """Reward correct answers with high score."""
@@ -155,7 +128,7 @@ def correctness_reward(prompts, completions, answer, **kwargs):
             for ans, gt in zip(extracted, answer)]
 ```
 
-**Example 2: Format Reward (Structured Output)**
+**Example 2: format reward (structured output)**
 ```python
 import re
 
@@ -167,7 +140,7 @@ def format_reward(completions, **kwargs):
             for r in responses]
 ```
 
-**Example 3: Incremental Format Reward (Partial Credit)**
+**Example 3: incremental format reward (partial credit)**
 ```python
 def incremental_format_reward(completions, **kwargs):
     """Award partial credit for format compliance."""
@@ -176,14 +149,10 @@ def incremental_format_reward(completions, **kwargs):
 
     for r in responses:
         score = 0.0
-        if '<reasoning>' in r:
-            score += 0.25
-        if '</reasoning>' in r:
-            score += 0.25
-        if '<answer>' in r:
-            score += 0.25
-        if '</answer>' in r:
-            score += 0.25
+        if '<reasoning>' in r:  score += 0.25
+        if '</reasoning>' in r: score += 0.25
+        if '<answer>' in r:     score += 0.25
+        if '</answer>' in r:    score += 0.25
         # Penalize extra text after closing tag
         if r.count('</answer>') == 1:
             extra_text = r.split('</answer>')[-1].strip()
@@ -193,12 +162,11 @@ def incremental_format_reward(completions, **kwargs):
     return rewards
 ```
 
-**Critical Insight:**
-Combine 3-5 reward functions for robust training. Order matters less than diversity of signals.
+**Critical insight:** Combine 3–5 reward functions for robust training. Order matters less than diversity of signals.
 
-### Step 3: Training Configuration
+### Step 3: Training configuration
 
-**Memory-Optimized Config (Small GPU)**
+**Memory-optimized config (small GPU)**
 ```python
 from trl import GRPOConfig
 
@@ -218,13 +186,13 @@ training_args = GRPOConfig(
     gradient_accumulation_steps=4,  # Effective batch = 4
 
     # GRPO-specific
-    num_generations=8,            # Group size: 8-16 recommended
+    num_generations=8,            # Group size: 8–16 recommended
     max_prompt_length=256,
     max_completion_length=512,
 
     # Training duration
     num_train_epochs=1,
-    max_steps=None,               # Or set fixed steps (e.g., 500)
+    max_steps=None,
 
     # Optimization
     bf16=True,                    # Faster on A100/H100
@@ -234,11 +202,11 @@ training_args = GRPOConfig(
     # Logging
     logging_steps=1,
     save_steps=100,
-    report_to="wandb",            # Or "none" for no logging
+    report_to="wandb",
 )
 ```
 
-**High-Performance Config (Large GPU)**
+**High-performance config (large GPU)**
 ```python
 training_args = GRPOConfig(
     output_dir="outputs/grpo-model",
@@ -255,31 +223,30 @@ training_args = GRPOConfig(
 )
 ```
 
-**Critical Hyperparameters:**
+**Critical hyperparameters:**
 
 | Parameter | Impact | Tuning Advice |
 |-----------|--------|---------------|
-| `num_generations` | Group size for comparison | Start with 8, increase to 16 if GPU allows |
+| `num_generations` | Group size for comparison | Start 8, increase to 16 if GPU allows |
 | `learning_rate` | Convergence speed/stability | 5e-6 (safe), 1e-5 (faster, riskier) |
-| `max_completion_length` | Output verbosity | Match your task (512 for reasoning, 256 for short answers) |
+| `max_completion_length` | Output verbosity | Match your task (512 reasoning, 256 short answers) |
 | `gradient_accumulation_steps` | Effective batch size | Increase if GPU memory limited |
 
-### Step 4: Model Setup and Training
+### Step 4: Model setup and training
 
-**Standard Setup (Transformers)**
+**Standard setup (Transformers + TRL)**
 ```python
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from peft import LoraConfig
 from trl import GRPOTrainer
 
-# Load model
 model_name = "Qwen/Qwen2.5-1.5B-Instruct"
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
     torch_dtype=torch.bfloat16,
-    attn_implementation="flash_attention_2",  # 2-3x faster
-    device_map="auto"
+    attn_implementation="flash_attention_2",  # 2–3× faster
+    device_map="auto",
 )
 
 tokenizer = AutoTokenizer.from_pretrained(model_name)
@@ -287,17 +254,16 @@ tokenizer.pad_token = tokenizer.eos_token
 
 # Optional: LoRA for parameter-efficient training
 peft_config = LoraConfig(
-    r=16,                         # Rank (higher = more capacity)
-    lora_alpha=32,               # Scaling factor (typically 2*r)
+    r=16,
+    lora_alpha=32,
     target_modules=[
         "q_proj", "k_proj", "v_proj", "o_proj",
-        "gate_proj", "up_proj", "down_proj"
+        "gate_proj", "up_proj", "down_proj",
     ],
     task_type="CAUSAL_LM",
     lora_dropout=0.05,
 )
 
-# Initialize trainer
 trainer = GRPOTrainer(
     model=model,
     processing_class=tokenizer,
@@ -308,17 +274,14 @@ trainer = GRPOTrainer(
     ],
     args=training_args,
     train_dataset=dataset,
-    peft_config=peft_config,      # Remove for full fine-tuning
+    peft_config=peft_config,   # Remove for full fine-tuning
 )
 
-# Train
 trainer.train()
-
-# Save
 trainer.save_model("final_model")
 ```
 
-**Unsloth Setup (2-3x Faster)**
+**Unsloth setup (2–3× faster)**
 ```python
 from unsloth import FastLanguageModel
 
@@ -339,28 +302,26 @@ model = FastLanguageModel.get_peft_model(
     use_gradient_checkpointing="unsloth",
 )
 
-# Rest is identical to standard setup
+# Rest is identical to the standard setup
 trainer = GRPOTrainer(model=model, ...)
 trainer.train()
 ```
 
----
+## Critical training insights
 
-## Critical Training Insights
+### 1. Loss behavior (EXPECTED pattern)
+- **Loss starts near 0 and INCREASES during training** — this is CORRECT
+- Loss measures KL divergence from initial policy; the model is learning (diverging from original behavior to optimize rewards)
+- **Monitor reward metrics, not loss, for progress**
 
-### 1. Loss Behavior (EXPECTED PATTERN)
-- **Loss starts near 0 and INCREASES during training**
-- This is CORRECT - loss measures KL divergence from initial policy
-- Model is learning (diverging from original behavior to optimize rewards)
-- Monitor reward metrics instead of loss for progress
+### 2. Reward tracking
 
-### 2. Reward Tracking
 Key metrics to watch:
-- `reward`: Average across all completions
-- `reward_std`: Diversity within groups (should remain > 0)
-- `kl`: KL divergence from reference (should grow moderately)
+- `reward` — average across all completions
+- `reward_std` — diversity within groups (should remain > 0)
+- `kl` — KL divergence from reference (should grow moderately)
 
-**Healthy Training Pattern:**
+**Healthy pattern:**
 ```
 Step   Reward    Reward_Std   KL
 100    0.5       0.3          0.02
@@ -369,12 +330,12 @@ Step   Reward    Reward_Std   KL
 400    1.5       0.15         0.12
 ```
 
-**Warning Signs:**
-- Reward std → 0 (model collapsing to single response)
-- KL exploding (> 0.5) (diverging too much, reduce LR)
-- Reward stuck (reward functions too harsh or model capacity issue)
+**Warning signs:**
+- `reward_std` → 0 (model collapsing to a single response)
+- `kl` exploding (> 0.5) — diverging too much, reduce LR
+- Reward stuck — reward functions too harsh or model capacity issue
 
-### 3. Common Pitfalls and Solutions
+### 3. Common pitfalls and solutions
 
 | Problem | Symptom | Solution |
 |---------|---------|----------|
@@ -384,15 +345,14 @@ Step   Reward    Reward_Std   KL
 | **Slow training** | < 1 it/s | Enable `use_vllm=True`, use Unsloth, reduce seq length |
 | **Format ignored** | Model doesn't follow structure | Increase format reward weight, add incremental rewards |
 
----
+## Advanced patterns
 
-## Advanced Patterns
+### 1. Multi-stage training
 
-### 1. Multi-Stage Training
 For complex tasks, train in stages:
 
 ```python
-# Stage 1: Format compliance (epochs=1)
+# Stage 1: Format compliance
 trainer_stage1 = GRPOTrainer(
     model=model,
     reward_funcs=[incremental_format_reward, format_reward],
@@ -400,7 +360,7 @@ trainer_stage1 = GRPOTrainer(
 )
 trainer_stage1.train()
 
-# Stage 2: Correctness (epochs=1)
+# Stage 2: Correctness
 trainer_stage2 = GRPOTrainer(
     model=model,
     reward_funcs=[format_reward, correctness_reward],
@@ -409,7 +369,8 @@ trainer_stage2 = GRPOTrainer(
 trainer_stage2.train()
 ```
 
-### 2. Adaptive Reward Scaling
+### 2. Adaptive reward scaling
+
 ```python
 class AdaptiveReward:
     def __init__(self, base_reward_func, initial_weight=1.0):
@@ -428,148 +389,116 @@ class AdaptiveReward:
             self.weight *= 0.9
 ```
 
-### 3. Custom Dataset Integration
+### 3. Custom dataset integration
+
 ```python
 def load_custom_knowledge_base(csv_path):
-    """Example: School communication platform docs."""
     import pandas as pd
     df = pd.read_csv(csv_path)
-
-    dataset = Dataset.from_pandas(df).map(lambda x: {
+    return Dataset.from_pandas(df).map(lambda x: {
         'prompt': [
             {'role': 'system', 'content': CUSTOM_SYSTEM_PROMPT},
             {'role': 'user', 'content': x['question']}
         ],
         'answer': x['expert_answer']
     })
-    return dataset
 ```
 
----
+## Deployment and inference
 
-## Deployment and Inference
-
-### Save and Merge LoRA
+### Save and merge LoRA
 ```python
-# Merge LoRA adapters into base model
 if hasattr(trainer.model, 'merge_and_unload'):
     merged_model = trainer.model.merge_and_unload()
     merged_model.save_pretrained("production_model")
     tokenizer.save_pretrained("production_model")
 ```
 
-### Inference Example
+### Inference
 ```python
 from transformers import pipeline
 
-generator = pipeline(
-    "text-generation",
-    model="production_model",
-    tokenizer=tokenizer
-)
+generator = pipeline("text-generation", model="production_model", tokenizer=tokenizer)
 
 result = generator(
     [
         {'role': 'system', 'content': SYSTEM_PROMPT},
-        {'role': 'user', 'content': "What is 15 + 27?"}
+        {'role': 'user', 'content': "What is 15 + 27?"},
     ],
     max_new_tokens=256,
     do_sample=True,
     temperature=0.7,
-    top_p=0.9
+    top_p=0.9,
 )
 print(result[0]['generated_text'])
 ```
 
----
+## Best practices checklist
 
-## Best Practices Checklist
-
-**Before Training:**
+**Before training:**
 - [ ] Validate dataset format (prompts as List[Dict])
 - [ ] Test reward functions on sample data
-- [ ] Calculate expected max_prompt_length from data
-- [ ] Choose appropriate num_generations based on GPU memory
+- [ ] Calculate expected `max_prompt_length` from data
+- [ ] Choose `num_generations` based on GPU memory
 - [ ] Set up logging (wandb recommended)
 
-**During Training:**
+**During training:**
 - [ ] Monitor reward progression (should increase)
-- [ ] Check reward_std (should stay > 0.1)
+- [ ] Check `reward_std` (should stay > 0.1)
 - [ ] Watch for OOM errors (reduce batch size if needed)
-- [ ] Sample generations every 50-100 steps
+- [ ] Sample generations every 50–100 steps
 - [ ] Validate format compliance on holdout set
 
-**After Training:**
+**After training:**
 - [ ] Merge LoRA weights if using PEFT
 - [ ] Test on diverse prompts
 - [ ] Compare to baseline model
 - [ ] Document reward weights and hyperparameters
 - [ ] Save reproducibility config
 
----
+## Troubleshooting
 
-## Troubleshooting Guide
+### Debugging workflow
+1. **Isolate reward functions** — test each independently
+2. **Check data distribution** — ensure diversity in prompts
+3. **Reduce complexity** — start with single reward, add gradually
+4. **Monitor generations** — print samples every N steps
+5. **Validate extraction logic** — ensure answer parsing works
 
-### Debugging Workflow
-1. **Isolate reward functions** - Test each independently
-2. **Check data distribution** - Ensure diversity in prompts
-3. **Reduce complexity** - Start with single reward, add gradually
-4. **Monitor generations** - Print samples every N steps
-5. **Validate extraction logic** - Ensure answer parsing works
-
-### Quick Fixes
+### Quick debug reward
 ```python
-# Debug reward function
 def debug_reward(completions, **kwargs):
     responses = [comp[0]['content'] for comp in completions]
-    for i, r in enumerate(responses[:2]):  # Print first 2
+    for i, r in enumerate(responses[:2]):
         print(f"Response {i}: {r[:200]}...")
-    return [1.0] * len(responses)  # Dummy rewards
+    return [1.0] * len(responses)
 
 # Test without training
 trainer = GRPOTrainer(..., reward_funcs=[debug_reward])
-trainer.generate_completions(dataset[:1])  # Generate without updating
+trainer.generate_completions(dataset[:1])
 ```
 
----
+## Template
 
-## References and Resources
+A production-ready training script lives at **`../templates/basic_grpo_training.py`**. It uses Qwen 2.5-1.5B-Instruct with LoRA and three reward functions (incremental format, strict format, correctness) on GSM8K. Copy and adapt:
+1. `get_dataset()` — swap in your data loader
+2. Reward functions — tune to your task
+3. `SYSTEM_PROMPT` — match your output format
+4. `GRPOConfig` — adjust hyperparameters for your GPU
+
+## References and resources
 
-**Official Documentation:**
 - TRL GRPO Trainer: https://huggingface.co/docs/trl/grpo_trainer
-- DeepSeek R1 Paper: https://arxiv.org/abs/2501.12948
-- Unsloth Docs: https://docs.unsloth.ai/
-
-**Example Repositories:**
-- Open R1 Implementation: https://github.com/huggingface/open-r1
-- TRL Examples: https://github.com/huggingface/trl/tree/main/examples
-
-**Recommended Reading:**
-- Progressive Disclosure Pattern for agent instructions
-- Reward shaping in RL (Ng et al.)
-- LoRA paper (Hu et al., 2021)
-
----
-
-## Usage Instructions for Agents
-
-When this skill is loaded:
-
-1. **Read this entire file** before implementing GRPO training
-2. **Start with the simplest reward function** (e.g., length-based) to validate setup
-3. **Use the templates** in `templates/` directory as starting points
-4. **Reference examples** in `examples/` for task-specific implementations
-5. **Follow the workflow** sequentially (don't skip steps)
-6. **Debug incrementally** - add one reward function at a time
-
-**Critical Reminders:**
-- Always use multiple reward functions (3-5 is optimal)
-- Monitor reward metrics, not loss
-- Test reward functions before training
-- Start small (num_generations=4), scale up gradually
-- Save checkpoints frequently (every 100 steps)
-
-This skill is designed for **expert-level implementation**. Beginners should start with supervised fine-tuning before attempting GRPO.
-
+- GRPO paper (DeepSeek): https://arxiv.org/abs/2402.03300
+- DeepSeek R1 paper: https://arxiv.org/abs/2501.12948
+- Open R1 implementation: https://github.com/huggingface/open-r1
+- TRL examples: https://github.com/huggingface/trl/tree/main/examples
+- Unsloth (faster training): https://docs.unsloth.ai/
 
+## Critical reminders
 
+- **Loss goes UP during training** — this is normal (it's KL divergence)
+- **Use 3–5 reward functions** — single rewards often fail
+- **Test rewards before training** — debug each function independently
+- **Monitor `reward_std`** — should stay > 0.1 (avoid mode collapse)
+- **Start with `num_generations=4–8`** — scale up if GPU allows

tests/acp/test_events.py

@@ -42,9 +42,10 @@ class TestToolProgressCallback:
     def test_emits_tool_call_start(self, mock_conn, event_loop_fixture):
         """Tool progress should emit a ToolCallStart update."""
         tool_call_ids = {}
+        tool_call_meta = {}
         loop = event_loop_fixture
 
-        cb = make_tool_progress_cb(mock_conn, "session-1", loop, tool_call_ids)
+        cb = make_tool_progress_cb(mock_conn, "session-1", loop, tool_call_ids, tool_call_meta)
 
         # Run callback in the event loop context
         with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts:
@@ -66,9 +67,10 @@ class TestToolProgressCallback:
     def test_handles_string_args(self, mock_conn, event_loop_fixture):
         """If args is a JSON string, it should be parsed."""
         tool_call_ids = {}
+        tool_call_meta = {}
         loop = event_loop_fixture
 
-        cb = make_tool_progress_cb(mock_conn, "session-1", loop, tool_call_ids)
+        cb = make_tool_progress_cb(mock_conn, "session-1", loop, tool_call_ids, tool_call_meta)
 
         with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts:
             future = MagicMock(spec=Future)
@@ -82,9 +84,10 @@ class TestToolProgressCallback:
     def test_handles_non_dict_args(self, mock_conn, event_loop_fixture):
         """If args is not a dict, it should be wrapped."""
         tool_call_ids = {}
+        tool_call_meta = {}
         loop = event_loop_fixture
 
-        cb = make_tool_progress_cb(mock_conn, "session-1", loop, tool_call_ids)
+        cb = make_tool_progress_cb(mock_conn, "session-1", loop, tool_call_ids, tool_call_meta)
 
         with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts:
             future = MagicMock(spec=Future)
@@ -98,10 +101,11 @@ class TestToolProgressCallback:
     def test_duplicate_same_name_tool_calls_use_fifo_ids(self, mock_conn, event_loop_fixture):
         """Multiple same-name tool calls should be tracked independently in order."""
         tool_call_ids = {}
+        tool_call_meta = {}
         loop = event_loop_fixture
 
-        progress_cb = make_tool_progress_cb(mock_conn, "session-1", loop, tool_call_ids)
-        step_cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids)
+        progress_cb = make_tool_progress_cb(mock_conn, "session-1", loop, tool_call_ids, tool_call_meta)
+        step_cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids, tool_call_meta)
 
         with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts:
             future = MagicMock(spec=Future)
@@ -163,7 +167,7 @@ class TestStepCallback:
         tool_call_ids = {"terminal": "tc-abc123"}
         loop = event_loop_fixture
 
-        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids)
+        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids, {})
 
         with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts:
             future = MagicMock(spec=Future)
@@ -181,7 +185,7 @@ class TestStepCallback:
         tool_call_ids = {}
         loop = event_loop_fixture
 
-        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids)
+        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids, {})
 
         with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts:
             cb(1, [{"name": "unknown_tool", "result": "ok"}])
@@ -193,7 +197,7 @@ class TestStepCallback:
         tool_call_ids = {"read_file": "tc-def456"}
         loop = event_loop_fixture
 
-        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids)
+        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids, {})
 
         with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts:
             future = MagicMock(spec=Future)
@@ -212,7 +216,7 @@ class TestStepCallback:
         tool_call_ids = {"terminal": deque(["tc-xyz789"])}
         loop = event_loop_fixture
 
-        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids)
+        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids, {})
 
         with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts, \
              patch("acp_adapter.events.build_tool_complete") as mock_btc:
@@ -224,7 +228,7 @@ class TestStepCallback:
             cb(1, [{"name": "terminal", "result": '{"output": "hello"}'}])
 
         mock_btc.assert_called_once_with(
-            "tc-xyz789", "terminal", result='{"output": "hello"}'
+            "tc-xyz789", "terminal", result='{"output": "hello"}', function_args=None, snapshot=None
         )
 
     def test_none_result_passed_through(self, mock_conn, event_loop_fixture):
@@ -234,7 +238,7 @@ class TestStepCallback:
         tool_call_ids = {"web_search": deque(["tc-aaa"])}
         loop = event_loop_fixture
 
-        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids)
+        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids, {})
 
         with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts, \
              patch("acp_adapter.events.build_tool_complete") as mock_btc:
@@ -244,7 +248,50 @@ class TestStepCallback:
 
             cb(1, [{"name": "web_search", "result": None}])
 
-        mock_btc.assert_called_once_with("tc-aaa", "web_search", result=None)
+        mock_btc.assert_called_once_with("tc-aaa", "web_search", result=None, function_args=None, snapshot=None)
+
+    def test_step_callback_passes_arguments_and_snapshot(self, mock_conn, event_loop_fixture):
+        from collections import deque
+
+        tool_call_ids = {"write_file": deque(["tc-write"])}
+        tool_call_meta = {"tc-write": {"args": {"path": "fallback.txt"}, "snapshot": "snap"}}
+        loop = event_loop_fixture
+
+        cb = make_step_cb(mock_conn, "session-1", loop, tool_call_ids, tool_call_meta)
+
+        with patch("acp_adapter.events.asyncio.run_coroutine_threadsafe") as mock_rcts, \
+             patch("acp_adapter.events.build_tool_complete") as mock_btc:
+            future = MagicMock(spec=Future)
+            future.result.return_value = None
+            mock_rcts.return_value = future
+
+            cb(1, [{"name": "write_file", "result": '{"bytes_written": 23}', "arguments": {"path": "diff-test.txt"}}])
+
+        mock_btc.assert_called_once_with(
+            "tc-write",
+            "write_file",
+            result='{"bytes_written": 23}',
+            function_args={"path": "diff-test.txt"},
+            snapshot="snap",
+        )
+
+    def test_tool_progress_captures_snapshot_metadata(self, mock_conn, event_loop_fixture):
+        tool_call_ids = {}
+        tool_call_meta = {}
+        loop = event_loop_fixture
+
+        with patch("acp_adapter.events.make_tool_call_id", return_value="tc-meta"), \
+             patch("acp_adapter.events._send_update") as mock_send, \
+             patch("agent.display.capture_local_edit_snapshot", return_value="snapshot"):
+            cb = make_tool_progress_cb(mock_conn, "session-1", loop, tool_call_ids, tool_call_meta)
+            cb("tool.started", "write_file", None, {"path": "diff-test.txt", "content": "hello"})
+
+        assert list(tool_call_ids["write_file"]) == ["tc-meta"]
+        assert tool_call_meta["tc-meta"] == {
+            "args": {"path": "diff-test.txt", "content": "hello"},
+            "snapshot": "snapshot",
+        }
+        mock_send.assert_called_once()
 
 
 # ---------------------------------------------------------------------------

tests/acp/test_mcp_e2e.py

@@ -29,6 +29,7 @@ from acp.schema import (
 
 from acp_adapter.server import HermesACPAgent
 from acp_adapter.session import SessionManager
+from acp_adapter.tools import build_tool_start
 
 
 # ---------------------------------------------------------------------------
@@ -181,6 +182,25 @@ class TestMcpRegistrationE2E:
         assert complete_event.raw_output is not None
         assert "hello" in str(complete_event.raw_output)
 
+    def test_patch_mode_tool_start_emits_diff_blocks_for_v4a_patch(self):
+        update = build_tool_start(
+            "tc-1",
+            "patch",
+            {
+                "mode": "patch",
+                "patch": "*** Begin Patch\n*** Update File: src/app.py\n@@\n-old line\n+new line\n*** Add File: src/new.py\n+hello\n*** End Patch",
+            },
+        )
+
+        assert len(update.content) == 2
+        assert update.content[0].type == "diff"
+        assert update.content[0].path == "src/app.py"
+        assert update.content[0].old_text == "old line"
+        assert update.content[0].new_text == "new line"
+        assert update.content[1].type == "diff"
+        assert update.content[1].path == "src/new.py"
+        assert update.content[1].new_text == "hello"
+
     @pytest.mark.asyncio
     async def test_prompt_tool_results_paired_by_call_id(self, acp_agent, mock_manager):
         """The ToolCallUpdate's toolCallId must match the ToolCallStart's."""

tests/acp/test_server.py

@@ -20,7 +20,9 @@ from acp.schema import (
     NewSessionResponse,
     PromptResponse,
     ResumeSessionResponse,
+    SessionModelState,
     SetSessionConfigOptionResponse,
+    SetSessionModelResponse,
     SetSessionModeResponse,
     SessionInfo,
     TextContentBlock,
@@ -127,6 +129,25 @@ class TestSessionOps:
         assert state is not None
         assert state.cwd == "/home/user/project"
 
+    @pytest.mark.asyncio
+    async def test_new_session_returns_model_state(self):
+        manager = SessionManager(
+            agent_factory=lambda: SimpleNamespace(model="gpt-5.4", provider="openai-codex")
+        )
+        acp_agent = HermesACPAgent(session_manager=manager)
+
+        with patch(
+            "hermes_cli.models.curated_models_for_provider",
+            return_value=[("gpt-5.4", "recommended"), ("gpt-5.4-mini", "")],
+        ):
+            resp = await acp_agent.new_session(cwd="/tmp")
+
+        assert isinstance(resp.models, SessionModelState)
+        assert resp.models.current_model_id == "openai-codex:gpt-5.4"
+        assert resp.models.available_models[0].model_id == "openai-codex:gpt-5.4"
+        assert resp.models.available_models[0].description is not None
+        assert "Provider:" in resp.models.available_models[0].description
+
     @pytest.mark.asyncio
     async def test_available_commands_include_help(self, agent):
         help_cmd = next(
@@ -204,6 +225,33 @@ class TestListAndFork:
         assert fork_resp.session_id
         assert fork_resp.session_id != new_resp.session_id
 
+    @pytest.mark.asyncio
+    async def test_list_sessions_includes_title_and_updated_at(self, agent):
+        with patch.object(
+            agent.session_manager,
+            "list_sessions",
+            return_value=[
+                {
+                    "session_id": "session-1",
+                    "cwd": "/tmp/project",
+                    "title": "Fix Zed session history",
+                    "updated_at": 123.0,
+                }
+            ],
+        ):
+            resp = await agent.list_sessions(cwd="/tmp/project")
+
+        assert isinstance(resp.sessions[0], SessionInfo)
+        assert resp.sessions[0].title == "Fix Zed session history"
+        assert resp.sessions[0].updated_at == "123.0"
+
+    @pytest.mark.asyncio
+    async def test_list_sessions_passes_cwd_filter(self, agent):
+        with patch.object(agent.session_manager, "list_sessions", return_value=[]) as mock_list:
+            await agent.list_sessions(cwd="/mnt/e/Projects/AI/browser-link-3")
+
+        mock_list.assert_called_once_with(cwd="/mnt/e/Projects/AI/browser-link-3")
+
 # ---------------------------------------------------------------------------
 # session configuration / model routing
 # ---------------------------------------------------------------------------
@@ -257,6 +305,53 @@ class TestSessionConfiguration:
         assert result == {}
         assert state.model == "gpt-5.4"
 
+    @pytest.mark.asyncio
+    async def test_set_session_model_accepts_provider_prefixed_choice(self, tmp_path, monkeypatch):
+        runtime_calls = []
+
+        def fake_resolve_runtime_provider(requested=None, **kwargs):
+            runtime_calls.append(requested)
+            provider = requested or "openrouter"
+            return {
+                "provider": provider,
+                "api_mode": "anthropic_messages" if provider == "anthropic" else "chat_completions",
+                "base_url": f"https://{provider}.example/v1",
+                "api_key": f"{provider}-key",
+                "command": None,
+                "args": [],
+            }
+
+        def fake_agent(**kwargs):
+            return SimpleNamespace(
+                model=kwargs.get("model"),
+                provider=kwargs.get("provider"),
+                base_url=kwargs.get("base_url"),
+                api_mode=kwargs.get("api_mode"),
+            )
+
+        monkeypatch.setattr("hermes_cli.config.load_config", lambda: {
+            "model": {"provider": "openrouter", "default": "openrouter/gpt-5"}
+        })
+        monkeypatch.setattr(
+            "hermes_cli.runtime_provider.resolve_runtime_provider",
+            fake_resolve_runtime_provider,
+        )
+        manager = SessionManager(db=SessionDB(tmp_path / "state.db"))
+
+        with patch("run_agent.AIAgent", side_effect=fake_agent):
+            acp_agent = HermesACPAgent(session_manager=manager)
+            state = manager.create_session(cwd="/tmp")
+            result = await acp_agent.set_session_model(
+                model_id="anthropic:claude-sonnet-4-6",
+                session_id=state.session_id,
+            )
+
+        assert isinstance(result, SetSessionModelResponse)
+        assert state.model == "claude-sonnet-4-6"
+        assert state.agent.provider == "anthropic"
+        assert state.agent.base_url == "https://anthropic.example/v1"
+        assert runtime_calls[-1] == "anthropic"
+
 
 # ---------------------------------------------------------------------------
 # prompt
@@ -354,6 +449,31 @@ class TestPrompt:
         update = last_call[1].get("update") or last_call[0][1]
         assert update.session_update == "agent_message_chunk"
 
+    @pytest.mark.asyncio
+    async def test_prompt_auto_titles_session(self, agent):
+        new_resp = await agent.new_session(cwd=".")
+        state = agent.session_manager.get_session(new_resp.session_id)
+        state.agent.run_conversation = MagicMock(return_value={
+            "final_response": "Here is the fix.",
+            "messages": [
+                {"role": "user", "content": "fix the broken ACP history"},
+                {"role": "assistant", "content": "Here is the fix."},
+            ],
+        })
+
+        mock_conn = MagicMock(spec=acp.Client)
+        mock_conn.session_update = AsyncMock()
+        agent._conn = mock_conn
+
+        with patch("agent.title_generator.maybe_auto_title") as mock_title:
+            prompt = [TextContentBlock(type="text", text="fix the broken ACP history")]
+            await agent.prompt(prompt=prompt, session_id=new_resp.session_id)
+
+        mock_title.assert_called_once()
+        assert mock_title.call_args.args[1] == new_resp.session_id
+        assert mock_title.call_args.args[2] == "fix the broken ACP history"
+        assert mock_title.call_args.args[3] == "Here is the fix."
+
     @pytest.mark.asyncio
     async def test_prompt_populates_usage_from_top_level_run_conversation_fields(self, agent):
         """ACP should map top-level token fields into PromptResponse.usage."""

tests/acp/test_session.py

@@ -3,6 +3,7 @@
 import contextlib
 import io
 import json
+import time
 from types import SimpleNamespace
 import pytest
 from unittest.mock import MagicMock, patch
@@ -100,15 +101,23 @@ class TestListAndCleanup:
     def test_list_sessions_returns_created(self, manager):
         s1 = manager.create_session(cwd="/a")
         s2 = manager.create_session(cwd="/b")
+        s1.history.append({"role": "user", "content": "hello from a"})
+        s2.history.append({"role": "user", "content": "hello from b"})
         listing = manager.list_sessions()
         ids = {s["session_id"] for s in listing}
         assert s1.session_id in ids
         assert s2.session_id in ids
         assert len(listing) == 2
 
+    def test_list_sessions_hides_empty_threads(self, manager):
+        manager.create_session(cwd="/empty")
+        assert manager.list_sessions() == []
+
     def test_cleanup_clears_all(self, manager):
-        manager.create_session()
-        manager.create_session()
+        s1 = manager.create_session()
+        s2 = manager.create_session()
+        s1.history.append({"role": "user", "content": "one"})
+        s2.history.append({"role": "user", "content": "two"})
         assert len(manager.list_sessions()) == 2
         manager.cleanup()
         assert manager.list_sessions() == []
@@ -194,6 +203,8 @@ class TestPersistence:
     def test_list_sessions_includes_db_only(self, manager):
         """Sessions only in DB (not in memory) appear in list_sessions."""
         state = manager.create_session(cwd="/db-only")
+        state.history.append({"role": "user", "content": "database only thread"})
+        manager.save_session(state.session_id)
         sid = state.session_id
 
         # Drop from memory.
@@ -204,6 +215,53 @@ class TestPersistence:
         ids = {s["session_id"] for s in listing}
         assert sid in ids
 
+    def test_list_sessions_filters_by_cwd(self, manager):
+        keep = manager.create_session(cwd="/keep")
+        drop = manager.create_session(cwd="/drop")
+        keep.history.append({"role": "user", "content": "keep me"})
+        drop.history.append({"role": "user", "content": "drop me"})
+
+        listing = manager.list_sessions(cwd="/keep")
+        ids = {s["session_id"] for s in listing}
+        assert keep.session_id in ids
+        assert drop.session_id not in ids
+
+    def test_list_sessions_matches_windows_and_wsl_paths(self, manager):
+        state = manager.create_session(cwd="/mnt/e/Projects/AI/browser-link-3")
+        state.history.append({"role": "user", "content": "same project from WSL"})
+
+        listing = manager.list_sessions(cwd=r"E:\Projects\AI\browser-link-3")
+        ids = {s["session_id"] for s in listing}
+        assert state.session_id in ids
+
+    def test_list_sessions_prefers_title_then_preview(self, manager):
+        state = manager.create_session(cwd="/named")
+        state.history.append({"role": "user", "content": "Investigate broken ACP history in Zed"})
+        manager.save_session(state.session_id)
+        db = manager._get_db()
+        db.set_session_title(state.session_id, "Fix Zed ACP history")
+
+        listing = manager.list_sessions(cwd="/named")
+        assert listing[0]["title"] == "Fix Zed ACP history"
+
+        db.set_session_title(state.session_id, "")
+        listing = manager.list_sessions(cwd="/named")
+        assert listing[0]["title"].startswith("Investigate broken ACP history")
+
+    def test_list_sessions_sorted_by_most_recent_activity(self, manager):
+        older = manager.create_session(cwd="/ordered")
+        older.history.append({"role": "user", "content": "older"})
+        manager.save_session(older.session_id)
+        time.sleep(0.02)
+        newer = manager.create_session(cwd="/ordered")
+        newer.history.append({"role": "user", "content": "newer"})
+        manager.save_session(newer.session_id)
+
+        listing = manager.list_sessions(cwd="/ordered")
+        assert [item["session_id"] for item in listing[:2]] == [newer.session_id, older.session_id]
+        assert listing[0]["updated_at"]
+        assert listing[1]["updated_at"]
+
     def test_fork_restores_source_from_db(self, manager):
         """Forking a session that is only in DB should work."""
         original = manager.create_session()

tests/acp/test_tools.py

@@ -215,6 +215,46 @@ class TestBuildToolComplete:
         assert len(display_text) < 6000
         assert "truncated" in display_text
 
+    def test_build_tool_complete_for_patch_uses_diff_blocks(self):
+        """Completed patch calls should keep structured diff content for Zed."""
+        patch_result = (
+            '{"success": true, "diff": "--- a/README.md\\n+++ b/README.md\\n@@ -1 +1,2 @@\\n old line\\n+new line\\n", '
+            '"files_modified": ["README.md"]}'
+        )
+        result = build_tool_complete("tc-p1", "patch", patch_result)
+        assert isinstance(result, ToolCallProgress)
+        assert len(result.content) == 1
+        diff_item = result.content[0]
+        assert isinstance(diff_item, FileEditToolCallContent)
+        assert diff_item.path == "README.md"
+        assert diff_item.old_text == "old line"
+        assert diff_item.new_text == "old line\nnew line"
+
+    def test_build_tool_complete_for_patch_falls_back_to_text_when_no_diff(self):
+        result = build_tool_complete("tc-p2", "patch", '{"success": true}')
+        assert isinstance(result, ToolCallProgress)
+        assert isinstance(result.content[0], ContentToolCallContent)
+
+    def test_build_tool_complete_for_write_file_uses_snapshot_diff(self, tmp_path):
+        target = tmp_path / "diff-test.txt"
+        snapshot = type("Snapshot", (), {"paths": [target], "before": {str(target): None}})()
+        target.write_text("hello from hermes\n", encoding="utf-8")
+
+        result = build_tool_complete(
+            "tc-wf1",
+            "write_file",
+            '{"bytes_written": 18, "dirs_created": false}',
+            function_args={"path": str(target), "content": "hello from hermes\n"},
+            snapshot=snapshot,
+        )
+        assert isinstance(result, ToolCallProgress)
+        assert len(result.content) == 1
+        diff_item = result.content[0]
+        assert isinstance(diff_item, FileEditToolCallContent)
+        assert diff_item.path.endswith("diff-test.txt")
+        assert diff_item.old_text is None
+        assert diff_item.new_text == "hello from hermes"
+
 
 # ---------------------------------------------------------------------------
 # extract_locations

tests/agent/test_auxiliary_client.py

@@ -696,6 +696,95 @@ class TestIsConnectionError:
         assert _is_connection_error(err) is False
 
 
+class TestKimiForCodingTemperature:
+    """kimi-for-coding now requires temperature=0.6 exactly."""
+
+    def test_build_call_kwargs_forces_fixed_temperature(self):
+        from agent.auxiliary_client import _build_call_kwargs
+
+        kwargs = _build_call_kwargs(
+            provider="kimi-coding",
+            model="kimi-for-coding",
+            messages=[{"role": "user", "content": "hello"}],
+            temperature=0.3,
+        )
+
+        assert kwargs["temperature"] == 0.6
+
+    def test_build_call_kwargs_injects_temperature_when_missing(self):
+        from agent.auxiliary_client import _build_call_kwargs
+
+        kwargs = _build_call_kwargs(
+            provider="kimi-coding",
+            model="kimi-for-coding",
+            messages=[{"role": "user", "content": "hello"}],
+            temperature=None,
+        )
+
+        assert kwargs["temperature"] == 0.6
+
+    def test_auto_routed_kimi_for_coding_sync_call_uses_fixed_temperature(self):
+        client = MagicMock()
+        client.base_url = "https://api.kimi.com/coding/v1"
+        response = MagicMock()
+        client.chat.completions.create.return_value = response
+
+        with patch(
+            "agent.auxiliary_client._get_cached_client",
+            return_value=(client, "kimi-for-coding"),
+        ), patch(
+            "agent.auxiliary_client._resolve_task_provider_model",
+            return_value=("auto", "kimi-for-coding", None, None, None),
+        ):
+            result = call_llm(
+                task="session_search",
+                messages=[{"role": "user", "content": "hello"}],
+                temperature=0.1,
+            )
+
+        assert result is response
+        kwargs = client.chat.completions.create.call_args.kwargs
+        assert kwargs["model"] == "kimi-for-coding"
+        assert kwargs["temperature"] == 0.6
+
+    @pytest.mark.asyncio
+    async def test_auto_routed_kimi_for_coding_async_call_uses_fixed_temperature(self):
+        client = MagicMock()
+        client.base_url = "https://api.kimi.com/coding/v1"
+        response = MagicMock()
+        client.chat.completions.create = AsyncMock(return_value=response)
+
+        with patch(
+            "agent.auxiliary_client._get_cached_client",
+            return_value=(client, "kimi-for-coding"),
+        ), patch(
+            "agent.auxiliary_client._resolve_task_provider_model",
+            return_value=("auto", "kimi-for-coding", None, None, None),
+        ):
+            result = await async_call_llm(
+                task="session_search",
+                messages=[{"role": "user", "content": "hello"}],
+                temperature=0.1,
+            )
+
+        assert result is response
+        kwargs = client.chat.completions.create.call_args.kwargs
+        assert kwargs["model"] == "kimi-for-coding"
+        assert kwargs["temperature"] == 0.6
+
+    def test_non_kimi_model_still_preserves_temperature(self):
+        from agent.auxiliary_client import _build_call_kwargs
+
+        kwargs = _build_call_kwargs(
+            provider="kimi-coding",
+            model="kimi-k2.5",
+            messages=[{"role": "user", "content": "hello"}],
+            temperature=0.3,
+        )
+
+        assert kwargs["temperature"] == 0.3
+
+
 # ---------------------------------------------------------------------------
 # async_call_llm payment / connection fallback (#7512 bug 2)
 # ---------------------------------------------------------------------------

tests/agent/test_auxiliary_main_first.py

@@ -0,0 +1,311 @@
+"""Regression tests for the ``auto`` → main-model-first policy.
+
+Prior to this change, aggregator users (OpenRouter / Nous Portal) had aux
+tasks routed through a cheap provider-side default (Gemini Flash) while
+non-aggregator users got their main model.  This made behavior inconsistent
+and surprising — users picked Claude but got Gemini Flash summaries.
+
+The current policy: ``auto`` means "use my main chat model" for every user,
+regardless of provider type.  Explicit per-task overrides in ``config.yaml``
+(``auxiliary.<task>.provider``) still win.  The cheap fallback chain only
+runs when the main provider has no working client.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+# ── Text aux tasks — _resolve_auto ──────────────────────────────────────────
+
+
+class TestResolveAutoMainFirst:
+    """_resolve_auto() must prefer main provider + main model for every user."""
+
+    def test_openrouter_main_uses_main_model_for_aux(self, monkeypatch):
+        """OpenRouter main user → aux uses their picked OR model, not Gemini Flash."""
+        monkeypatch.setenv("OPENROUTER_API_KEY", "or-test-key")
+
+        with patch(
+            "agent.auxiliary_client._read_main_provider",
+            return_value="openrouter",
+        ), patch(
+            "agent.auxiliary_client._read_main_model",
+            return_value="anthropic/claude-sonnet-4.6",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client"
+        ) as mock_resolve:
+            mock_client = MagicMock()
+            mock_resolve.return_value = (mock_client, "anthropic/claude-sonnet-4.6")
+
+            from agent.auxiliary_client import _resolve_auto
+
+            client, model = _resolve_auto()
+
+        assert client is mock_client
+        assert model == "anthropic/claude-sonnet-4.6"
+        # Verify it asked resolve_provider_client for the MAIN provider+model,
+        # not a fallback-chain provider
+        mock_resolve.assert_called_once()
+        assert mock_resolve.call_args.args[0] == "openrouter"
+        assert mock_resolve.call_args.args[1] == "anthropic/claude-sonnet-4.6"
+
+    def test_nous_main_uses_main_model_for_aux(self, monkeypatch):
+        """Nous Portal main user → aux uses their picked Nous model, not free-tier MiMo."""
+        # No OPENROUTER_API_KEY → ensures if main failed we'd fall to chain
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="nous",
+        ), patch(
+            "agent.auxiliary_client._read_main_model",
+            return_value="anthropic/claude-opus-4.6",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client"
+        ) as mock_resolve:
+            mock_client = MagicMock()
+            mock_resolve.return_value = (mock_client, "anthropic/claude-opus-4.6")
+
+            from agent.auxiliary_client import _resolve_auto
+
+            client, model = _resolve_auto()
+
+        assert client is mock_client
+        assert model == "anthropic/claude-opus-4.6"
+        assert mock_resolve.call_args.args[0] == "nous"
+
+    def test_non_aggregator_main_still_uses_main(self, monkeypatch):
+        """Non-aggregator main (DeepSeek) → unchanged behavior, main model used."""
+        monkeypatch.setenv("DEEPSEEK_API_KEY", "ds-test")
+
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="deepseek",
+        ), patch(
+            "agent.auxiliary_client._read_main_model", return_value="deepseek-chat",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client"
+        ) as mock_resolve:
+            mock_client = MagicMock()
+            mock_resolve.return_value = (mock_client, "deepseek-chat")
+
+            from agent.auxiliary_client import _resolve_auto
+
+            client, model = _resolve_auto()
+
+        assert client is mock_client
+        assert model == "deepseek-chat"
+        assert mock_resolve.call_args.args[0] == "deepseek"
+
+    def test_main_unavailable_falls_through_to_chain(self, monkeypatch):
+        """Main provider with no working client → fall back to aux chain."""
+        monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")
+
+        chain_client = MagicMock()
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="anthropic",
+        ), patch(
+            "agent.auxiliary_client._read_main_model", return_value="claude-opus",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client",
+            return_value=(None, None),  # main provider has no client
+        ), patch(
+            "agent.auxiliary_client._try_openrouter",
+            return_value=(chain_client, "google/gemini-3-flash-preview"),
+        ):
+            from agent.auxiliary_client import _resolve_auto
+
+            client, model = _resolve_auto()
+
+        assert client is chain_client
+        assert model == "google/gemini-3-flash-preview"
+
+    def test_no_main_config_uses_chain_directly(self):
+        """No main provider configured → skip step 1, use chain (no regression)."""
+        chain_client = MagicMock()
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="",
+        ), patch(
+            "agent.auxiliary_client._read_main_model", return_value="",
+        ), patch(
+            "agent.auxiliary_client._try_openrouter",
+            return_value=(chain_client, "google/gemini-3-flash-preview"),
+        ):
+            from agent.auxiliary_client import _resolve_auto
+
+            client, model = _resolve_auto()
+
+        assert client is chain_client
+
+    def test_runtime_override_wins_over_config(self, monkeypatch):
+        """main_runtime kwarg overrides config-read main provider/model."""
+        with patch(
+            "agent.auxiliary_client._read_main_provider",
+            return_value="openrouter",
+        ), patch(
+            "agent.auxiliary_client._read_main_model", return_value="config-model",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client"
+        ) as mock_resolve:
+            mock_resolve.return_value = (MagicMock(), "runtime-model")
+
+            from agent.auxiliary_client import _resolve_auto
+
+            _resolve_auto(main_runtime={
+                "provider": "anthropic",
+                "model": "runtime-model",
+                "base_url": "",
+                "api_key": "",
+                "api_mode": "",
+            })
+
+        # Runtime override wins
+        assert mock_resolve.call_args.args[0] == "anthropic"
+        assert mock_resolve.call_args.args[1] == "runtime-model"
+
+
+# ── Vision — resolve_vision_provider_client ─────────────────────────────────
+
+
+class TestResolveVisionMainFirst:
+    """Vision auto-detection prefers main provider + main model first."""
+
+    def test_openrouter_main_vision_uses_main_model(self, monkeypatch):
+        """OpenRouter main with vision-capable model → aux vision uses main model."""
+        monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")
+
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="openrouter",
+        ), patch(
+            "agent.auxiliary_client._read_main_model",
+            return_value="anthropic/claude-sonnet-4.6",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client"
+        ) as mock_resolve, patch(
+            "agent.auxiliary_client._resolve_task_provider_model",
+            return_value=("auto", None, None, None, None),
+        ):
+            mock_client = MagicMock()
+            mock_resolve.return_value = (mock_client, "anthropic/claude-sonnet-4.6")
+
+            from agent.auxiliary_client import resolve_vision_provider_client
+
+            provider, client, model = resolve_vision_provider_client()
+
+        assert provider == "openrouter"
+        assert client is mock_client
+        assert model == "anthropic/claude-sonnet-4.6"
+        # Verify it did NOT call the strict vision backend for OpenRouter
+        # (which would have used a cheap gemini-flash-preview default)
+        mock_resolve.assert_called_once()
+        assert mock_resolve.call_args.args[0] == "openrouter"
+        assert mock_resolve.call_args.args[1] == "anthropic/claude-sonnet-4.6"
+
+    def test_nous_main_vision_uses_main_model(self):
+        """Nous Portal main → aux vision uses main model, not free-tier MiMo-V2-Omni."""
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="nous",
+        ), patch(
+            "agent.auxiliary_client._read_main_model",
+            return_value="openai/gpt-5",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client"
+        ) as mock_resolve, patch(
+            "agent.auxiliary_client._resolve_task_provider_model",
+            return_value=("auto", None, None, None, None),
+        ):
+            mock_client = MagicMock()
+            mock_resolve.return_value = (mock_client, "openai/gpt-5")
+
+            from agent.auxiliary_client import resolve_vision_provider_client
+
+            provider, client, model = resolve_vision_provider_client()
+
+        assert provider == "nous"
+        assert model == "openai/gpt-5"
+
+    def test_exotic_provider_with_vision_override_preserved(self):
+        """xiaomi → mimo-v2-omni override still wins over main_model."""
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="xiaomi",
+        ), patch(
+            "agent.auxiliary_client._read_main_model",
+            return_value="mimo-v2-pro",  # text model
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client"
+        ) as mock_resolve, patch(
+            "agent.auxiliary_client._resolve_task_provider_model",
+            return_value=("auto", None, None, None, None),
+        ):
+            mock_resolve.return_value = (MagicMock(), "mimo-v2-omni")
+
+            from agent.auxiliary_client import resolve_vision_provider_client
+
+            provider, client, model = resolve_vision_provider_client()
+
+        assert provider == "xiaomi"
+        # Should use mimo-v2-omni (vision override), not mimo-v2-pro (text main)
+        assert mock_resolve.call_args.args[1] == "mimo-v2-omni"
+
+    def test_main_unavailable_vision_falls_through_to_aggregators(self):
+        """Main provider fails → fall back to OpenRouter/Nous strict backends."""
+        fallback_client = MagicMock()
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="deepseek",
+        ), patch(
+            "agent.auxiliary_client._read_main_model", return_value="deepseek-chat",
+        ), patch(
+            "agent.auxiliary_client.resolve_provider_client",
+            return_value=(None, None),
+        ), patch(
+            "agent.auxiliary_client._resolve_strict_vision_backend",
+            return_value=(fallback_client, "google/gemini-3-flash-preview"),
+        ), patch(
+            "agent.auxiliary_client._resolve_task_provider_model",
+            return_value=("auto", None, None, None, None),
+        ):
+            from agent.auxiliary_client import resolve_vision_provider_client
+
+            provider, client, model = resolve_vision_provider_client()
+
+        assert client is fallback_client
+        assert provider in ("openrouter", "nous")
+
+    def test_explicit_provider_override_still_wins(self):
+        """Explicit config override bypasses main-first policy."""
+        with patch(
+            "agent.auxiliary_client._read_main_provider", return_value="openrouter",
+        ), patch(
+            "agent.auxiliary_client._read_main_model",
+            return_value="anthropic/claude-opus-4.6",
+        ), patch(
+            "agent.auxiliary_client._resolve_task_provider_model",
+            return_value=("nous", None, None, None, None),  # explicit override
+        ), patch(
+            "agent.auxiliary_client._resolve_strict_vision_backend"
+        ) as mock_strict:
+            mock_strict.return_value = (MagicMock(), "nous-default-model")
+
+            from agent.auxiliary_client import resolve_vision_provider_client
+
+            provider, client, model = resolve_vision_provider_client()
+
+        # Explicit "nous" override → uses strict backend, NOT main model path
+        assert provider == "nous"
+        mock_strict.assert_called_once_with("nous")
+
+
+# ── Constant cleanup ────────────────────────────────────────────────────────
+
+
+def test_aggregator_providers_constant_removed():
+    """The dead _AGGREGATOR_PROVIDERS constant should no longer live in the module.
+
+    Removed when the main-first policy made the aggregator-skip guard obsolete.
+    """
+    import agent.auxiliary_client as aux_mod
+
+    assert not hasattr(aux_mod, "_AGGREGATOR_PROVIDERS"), (
+        "_AGGREGATOR_PROVIDERS was removed when _resolve_auto stopped "
+        "treating aggregators specially. If you re-added it, the main-first "
+        "policy may have regressed."
+    )

tests/agent/test_gemini_cloudcode.py

@@ -826,6 +826,160 @@ class TestGeminiCloudCodeClient:
         finally:
             client.close()
 
+
+class TestGeminiHttpErrorParsing:
+    """Regression coverage for _gemini_http_error Google-envelope parsing.
+
+    These are the paths that users actually hit during Google-side throttling
+    (April 2026: gemini-2.5-pro MODEL_CAPACITY_EXHAUSTED, gemma-4-26b-it
+    returning 404).  The error needs to carry status_code + response so the
+    main loop's error_classifier and Retry-After logic work.
+    """
+
+    @staticmethod
+    def _fake_response(status: int, body: dict | str = "", headers=None):
+        """Minimal httpx.Response stand-in (duck-typed for _gemini_http_error)."""
+        class _FakeResponse:
+            def __init__(self):
+                self.status_code = status
+                if isinstance(body, dict):
+                    self.text = json.dumps(body)
+                else:
+                    self.text = body
+                self.headers = headers or {}
+        return _FakeResponse()
+
+    def test_model_capacity_exhausted_produces_friendly_message(self):
+        from agent.gemini_cloudcode_adapter import _gemini_http_error
+
+        body = {
+            "error": {
+                "code": 429,
+                "message": "Resource has been exhausted (e.g. check quota).",
+                "status": "RESOURCE_EXHAUSTED",
+                "details": [
+                    {
+                        "@type": "type.googleapis.com/google.rpc.ErrorInfo",
+                        "reason": "MODEL_CAPACITY_EXHAUSTED",
+                        "domain": "googleapis.com",
+                        "metadata": {"model": "gemini-2.5-pro"},
+                    },
+                    {
+                        "@type": "type.googleapis.com/google.rpc.RetryInfo",
+                        "retryDelay": "30s",
+                    },
+                ],
+            }
+        }
+        err = _gemini_http_error(self._fake_response(429, body))
+        assert err.status_code == 429
+        assert err.code == "code_assist_capacity_exhausted"
+        assert err.retry_after == 30.0
+        assert err.details["reason"] == "MODEL_CAPACITY_EXHAUSTED"
+        # Message must be user-friendly, not a raw JSON dump.
+        message = str(err)
+        assert "gemini-2.5-pro" in message
+        assert "capacity exhausted" in message.lower()
+        assert "30s" in message
+        # response attr is preserved for run_agent's Retry-After header path.
+        assert err.response is not None
+
+    def test_resource_exhausted_without_reason(self):
+        from agent.gemini_cloudcode_adapter import _gemini_http_error
+
+        body = {
+            "error": {
+                "code": 429,
+                "message": "Quota exceeded for requests per minute.",
+                "status": "RESOURCE_EXHAUSTED",
+            }
+        }
+        err = _gemini_http_error(self._fake_response(429, body))
+        assert err.status_code == 429
+        assert err.code == "code_assist_rate_limited"
+        message = str(err)
+        assert "quota" in message.lower()
+
+    def test_404_model_not_found_produces_model_retired_message(self):
+        from agent.gemini_cloudcode_adapter import _gemini_http_error
+
+        body = {
+            "error": {
+                "code": 404,
+                "message": "models/gemma-4-26b-it is not found for API version v1internal",
+                "status": "NOT_FOUND",
+            }
+        }
+        err = _gemini_http_error(self._fake_response(404, body))
+        assert err.status_code == 404
+        message = str(err)
+        assert "not available" in message.lower() or "retired" in message.lower()
+        # Error message should reference the actual model text from Google.
+        assert "gemma-4-26b-it" in message
+
+    def test_unauthorized_preserves_status_code(self):
+        from agent.gemini_cloudcode_adapter import _gemini_http_error
+
+        err = _gemini_http_error(self._fake_response(
+            401, {"error": {"code": 401, "message": "Invalid token", "status": "UNAUTHENTICATED"}},
+        ))
+        assert err.status_code == 401
+        assert err.code == "code_assist_unauthorized"
+
+    def test_retry_after_header_fallback(self):
+        """If the body has no RetryInfo detail, fall back to Retry-After header."""
+        from agent.gemini_cloudcode_adapter import _gemini_http_error
+
+        resp = self._fake_response(
+            429,
+            {"error": {"code": 429, "message": "Rate limited", "status": "RESOURCE_EXHAUSTED"}},
+            headers={"Retry-After": "45"},
+        )
+        err = _gemini_http_error(resp)
+        assert err.retry_after == 45.0
+
+    def test_malformed_body_still_produces_structured_error(self):
+        """Non-JSON body must not swallow status_code — we still want the classifier path."""
+        from agent.gemini_cloudcode_adapter import _gemini_http_error
+
+        err = _gemini_http_error(self._fake_response(500, "<html>internal error</html>"))
+        assert err.status_code == 500
+        # Raw body snippet must still be there for debugging.
+        assert "500" in str(err)
+
+    def test_status_code_flows_through_error_classifier(self):
+        """End-to-end: CodeAssistError from a 429 must classify as rate_limit.
+
+        This is the whole point of adding status_code to CodeAssistError —
+        _extract_status_code must see it and FailoverReason.rate_limit must
+        fire, so the main loop triggers fallback_providers.
+        """
+        from agent.gemini_cloudcode_adapter import _gemini_http_error
+        from agent.error_classifier import classify_api_error, FailoverReason
+
+        body = {
+            "error": {
+                "code": 429,
+                "message": "Resource has been exhausted",
+                "status": "RESOURCE_EXHAUSTED",
+                "details": [
+                    {
+                        "@type": "type.googleapis.com/google.rpc.ErrorInfo",
+                        "reason": "MODEL_CAPACITY_EXHAUSTED",
+                        "metadata": {"model": "gemini-2.5-pro"},
+                    }
+                ],
+            }
+        }
+        err = _gemini_http_error(self._fake_response(429, body))
+
+        classified = classify_api_error(
+            err, provider="google-gemini-cli", model="gemini-2.5-pro",
+        )
+        assert classified.status_code == 429
+        assert classified.reason == FailoverReason.rate_limit
+
+
 # =============================================================================
 # Provider registration
 # =============================================================================