Merge pull request #4706 from pewdiepie-archdaemon/sync-readme-screenshot-dev

docs: refresh README screenshot

.dockerignore

@@ -15,6 +15,10 @@ build/
 # at runtime — never baked into the image. Mirrored in .gitignore.
 secrets.env
 secrets.env.*
+secrets.env~
+.secrets.env.swp
+.secrets.env.swo
+**/#secrets.env#
 !secrets.env.example
 /data/
 /logs/

.github/workflows/ci.yml

@@ -19,7 +19,7 @@ jobs:
     name: Python syntax (compileall)
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
       - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
@@ -32,7 +32,7 @@ jobs:
     name: JS syntax (node --check)
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
       - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
@@ -54,7 +54,7 @@ jobs:
     # ROADMAP "fresh install smoke tests" item; make this required once green.
     continue-on-error: true
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           fetch-depth: 0
           persist-credentials: false

.github/workflows/container-scan.yml

@@ -37,7 +37,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
 

.github/workflows/container-trivy.yml

@@ -52,7 +52,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
 
@@ -93,7 +93,7 @@ jobs:
       security-events: write  # upload SARIF to the Security tab
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
 

.github/workflows/dependency-review.yml

@@ -36,7 +36,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
 
@@ -55,7 +55,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
 

.github/workflows/docker-publish.yml

@@ -45,7 +45,7 @@ jobs:
             arch: arm64
             runner: ubuntu-24.04-arm
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
       - name: Set up Buildx
@@ -86,7 +86,7 @@ jobs:
       contents: read
       packages: write
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
       - name: Read APP_VERSION + short sha

.github/workflows/issue-description-check.yml

@@ -14,7 +14,7 @@ jobs:
     # Skip bots (Dependabot, release-drafter, etc.)
     if: ${{ github.event.issue.user.type != 'Bot' }}
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           sparse-checkout: .github/scripts
           persist-credentials: false

.github/workflows/pr-description-check.yml

@@ -23,7 +23,7 @@ jobs:
     # Skip bots: they open PRs programmatically and have their own process.
     if: github.event.pull_request.user.type != 'Bot'
     steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           ref: ${{ github.base_ref }}
           sparse-checkout: .github/scripts

.github/workflows/secret-scan.yml

@@ -35,7 +35,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           # Full history so a secret committed in an earlier commit (and later
           # deleted) is still caught -- deletion does not remove it from Git.

.github/workflows/workflow-security.yml

@@ -36,7 +36,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
 
@@ -61,7 +61,7 @@ jobs:
       contents: read
     steps:
       - name: Checkout repository
-        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10  # v6.0.3
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
 

CONTRIBUTING.md

@@ -37,7 +37,7 @@ Manual development uses Python 3.11+:
 python3 -m venv venv
 source venv/bin/activate
 pip install -r requirements.txt
-python -m uvicorn app:app --host 0.0.0.0 --port 7000
+python -m uvicorn app:app --host 127.0.0.1 --port 7000
 ```
 
 Windows is not actively tested. Docker on Linux or a Linux/macOS manual install is the safer path for now.

Dockerfile

@@ -20,6 +20,23 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     gosu \
     && rm -rf /var/lib/apt/lists/*
 
+# Docker CLI (client only — daemon stays on the host via the
+# /var/run/docker.sock mount). The Debian `docker.io` package ships
+# dockerd but not the client binary on slim, so grab the static client
+# tarball from download.docker.com instead.
+ARG DOCKER_CLI_VERSION=27.5.1
+RUN ARCH="$(dpkg --print-architecture)" \
+    && case "$ARCH" in \
+         amd64) DARCH=x86_64 ;; \
+         arm64) DARCH=aarch64 ;; \
+         *) echo "unsupported arch $ARCH"; exit 1 ;; \
+       esac \
+    && curl -fsSL "https://download.docker.com/linux/static/stable/${DARCH}/docker-${DOCKER_CLI_VERSION}.tgz" \
+       -o /tmp/docker.tgz \
+    && tar -xzf /tmp/docker.tgz -C /tmp \
+    && install -m 0755 /tmp/docker/docker /usr/local/bin/docker \
+    && rm -rf /tmp/docker /tmp/docker.tgz
+
 WORKDIR /app
 
 # Install Python deps first (layer cache). Optional extras (PyMuPDF AGPL, etc.)

Odysseus.spec

@@ -0,0 +1,45 @@
+# -*- mode: python ; coding: utf-8 -*-
+
+
+a = Analysis(
+    ['launcher.py'],
+    pathex=[],
+    binaries=[],
+    datas=[('static', 'static'), ('scripts', 'scripts'), ('mcp_servers', 'mcp_servers'), ('services/hwfit/data', 'services/hwfit/data'), ('config', 'config'), ('.env.example', '.env.example')],
+    hiddenimports=[],
+    hookspath=[],
+    hooksconfig={},
+    runtime_hooks=[],
+    excludes=[],
+    noarchive=False,
+    optimize=0,
+)
+pyz = PYZ(a.pure)
+
+exe = EXE(
+    pyz,
+    a.scripts,
+    [],
+    exclude_binaries=True,
+    name='Odysseus',
+    debug=False,
+    bootloader_ignore_signals=False,
+    strip=False,
+    upx=True,
+    console=False,
+    disable_windowed_traceback=False,
+    argv_emulation=False,
+    target_arch=None,
+    codesign_identity=None,
+    entitlements_file=None,
+    icon=['static\\icon.ico'],
+)
+coll = COLLECT(
+    exe,
+    a.binaries,
+    a.datas,
+    strip=False,
+    upx=True,
+    upx_exclude=[],
+    name='Odysseus',
+)

README.md

@@ -1,5 +1,5 @@
 <p align="center">
-  <img src="docs/odysseus-wordmark.png" alt="Odysseus" width="280">
+  <img src="docs/odysseus-wordmark.png" alt="Odysseus" width="238">
 </p>
 
 <p align="center">
@@ -18,7 +18,7 @@
 </p>
 
 <p align="center">
-  <img src="docs/odysseus.jpg" alt="Odysseus interface">
+  <img src="docs/odysseus-browser.jpg" alt="Odysseus interface">
 </p>
 
 ---

app.py

@@ -1,6 +1,7 @@
 # app.py — slim orchestrator
 import mimetypes
 import os
+import sys
 
 
 def register_static_mime_types() -> None:
@@ -38,7 +39,7 @@ load_dotenv(encoding="utf-8-sig")
 import asyncio
 import logging
 import secrets
-from datetime import datetime
+from datetime import datetime, timezone
 from typing import Dict
 
 from contextlib import asynccontextmanager
@@ -113,12 +114,13 @@ app = FastAPI(
 )
 
 # ========= CORS =========
+CORS_ALLOW_METHODS = ["GET", "POST", "PUT", "PATCH", "DELETE"]
 allowed_origins = os.getenv("ALLOWED_ORIGINS", "http://localhost,http://127.0.0.1").split(",")
 app.add_middleware(
     CORSMiddleware,
     allow_origins=allowed_origins,
     allow_credentials=True,
-    allow_methods=["GET", "POST", "PUT", "DELETE"],
+    allow_methods=CORS_ALLOW_METHODS,
     allow_headers=[
         "Accept",
         "Authorization",
@@ -316,7 +318,7 @@ if AUTH_ENABLED:
             # (no admin cookie available in that context). Restricted to
             # loopback clients + matching token to keep it locked down.
             try:
-                from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN as _ITT
+                from core.middleware import INTERNAL_TOOL_HEADER, INTERNAL_TOOL_TOKEN as _ITT, INTERNAL_TOOL_USER
                 _hdr = request.headers.get(INTERNAL_TOOL_HEADER)
                 if _hdr and secrets.compare_digest(_hdr, _ITT) and _is_trusted_loopback(request):
                     # Impersonation: when the agent's loopback call sets
@@ -328,11 +330,11 @@ if AUTH_ENABLED:
                     if _impersonate and _impersonate in getattr(_auth_mgr, "users", {}):
                         request.state.current_user = _impersonate
                     else:
-                        request.state.current_user = "internal-tool"
+                        request.state.current_user = INTERNAL_TOOL_USER
                     request.state.api_token = False
                     return await call_next(request)
-            except Exception:
-                pass
+            except Exception as _e:
+                logger.warning("Internal tool auth header check failed", exc_info=_e)
             # Allow DIRECT localhost requests (internal service calls from
             # heartbeats etc.). Tunnel/proxy-forwarded requests are excluded by
             # _is_trusted_loopback so LOCALHOST_BYPASS can't be abused over a
@@ -385,11 +387,10 @@ if AUTH_ENABLED:
                                     _db.close()
                             try:
                                 await _asyncio.to_thread(_do)
-                            except Exception:
-                                pass
+                            except Exception as _e:
+                                logger.debug("Failed to update token last_used_at", exc_info=_e)
                         _asyncio.create_task(_touch_last_used(matched_id))
                         # Keep bearer-token callers out of normal cookie/user
-                        # routes. API-aware routes can read api_token_owner.
                         request.state.current_user = "api"
                         request.state.api_token = True
                         request.state.api_token_id = matched_id
@@ -438,7 +439,7 @@ class _RevalidatingStatic(StaticFiles):
         return resp
 
 
-app.mount("/static", _RevalidatingStatic(directory="static"), name="static")
+app.mount("/static", _RevalidatingStatic(directory=STATIC_DIR), name="static")
 
 # ========= GENERATED IMAGES =========
 @app.get("/api/generated-image/{filename}")
@@ -464,8 +465,8 @@ async def serve_generated_image(filename: str, request: Request):
                 _db.close()
     except HTTPException:
         raise
-    except Exception:
-        pass
+    except Exception as _e:
+        logger.warning("Image ownership verification failed for %r", filename, exc_info=_e)
     ext = filename.rsplit('.', 1)[-1].lower()
     mime = {
         "png": "image/png", "jpg": "image/jpeg", "jpeg": "image/jpeg",
@@ -528,6 +529,7 @@ memory_vector     = components.get("memory_vector")
 upload_handler    = components["upload_handler"]
 app.state.upload_handler = upload_handler
 personal_docs_mgr = components["personal_docs_manager"]
+app.state.personal_docs_manager = personal_docs_mgr
 api_key_manager   = components["api_key_manager"]
 preset_manager    = components["preset_manager"]
 chat_processor    = components["chat_processor"]
@@ -861,7 +863,7 @@ async def get_version():
 
 @app.get("/api/health")
 async def health_check() -> Dict[str, str]:
-    return {"status": "healthy", "timestamp": datetime.utcnow().isoformat()}
+    return {"status": "healthy", "timestamp": datetime.now(timezone.utc).isoformat()}
 
 @app.get("/api/ready")
 async def readiness_check() -> JSONResponse:
@@ -1171,3 +1173,12 @@ async def _shutdown_event():
     except Exception as e:
         logger.warning(f"MCP shutdown error: {e}")
     logger.info("Application shutdown complete")
+
+
+if __name__ == "__main__":
+    import uvicorn
+
+    bind_host = os.getenv("APP_BIND", "127.0.0.1")
+    bind_port = int(os.getenv("APP_PORT", "7000"))
+
+    uvicorn.run(app, host=bind_host, port=bind_port, log_level="info")

build-windows-portable.ps1

@@ -0,0 +1,72 @@
+#Requires -Version 5.1
+<#
+  Build a portable Windows distribution for Odysseus.
+
+  Output layout:
+    dist\Odysseus\Odysseus.exe
+    dist\Odysseus\static\...
+    dist\Odysseus\scripts\...
+    dist\Odysseus\mcp_servers\...
+    dist\Odysseus\services\hwfit\data\...
+
+  The app then keeps using its normal filesystem layout when frozen.
+
+  Usage:
+    powershell -ExecutionPolicy Bypass -File .\build-windows-portable.ps1
+#>
+
+$ErrorActionPreference = "Stop"
+Set-Location -Path $PSScriptRoot
+
+function Write-Step($msg) { Write-Host ""; Write-Host ("==> " + $msg) -ForegroundColor Cyan }
+function Fail($msg) {
+    Write-Host ""
+    Write-Host ("ERROR: " + $msg) -ForegroundColor Red
+    exit 1
+}
+
+Write-Step "Checking for Python"
+$pyExe = $null
+if (Test-Path ".\.venv\Scripts\python.exe") {
+    $pyExe = (Resolve-Path ".\.venv\Scripts\python.exe").Path
+} else {
+    foreach ($c in @("py", "python")) {
+        $cmd = Get-Command $c -ErrorAction SilentlyContinue
+        if ($cmd) { $pyExe = $cmd.Source; break }
+    }
+    if ($pyExe -like "*WindowsApps*python.exe") {
+        $pyCmd = Get-Command py -ErrorAction SilentlyContinue
+        if ($pyCmd) {
+            $pyExe = $pyCmd.Source
+        }
+    }
+}
+if (-not $pyExe) {
+    Fail "Python not found on PATH. Install Python 3.11+ first."
+}
+Write-Host ("Using Python: " + $pyExe)
+
+Write-Step "Installing build dependencies"
+& $pyExe -m pip install --upgrade pip --quiet
+& $pyExe -m pip install -r requirements.txt pyinstaller pystray Pillow
+if ($LASTEXITCODE -ne 0) { Fail "Dependency install failed." }
+
+Write-Step "Building portable exe bundle"
+Remove-Item -Recurse -Force build, dist -ErrorAction SilentlyContinue
+
+$dataArgs = @(
+    "--add-data", "static;static",
+    "--add-data", "scripts;scripts",
+    "--add-data", "mcp_servers;mcp_servers",
+    "--add-data", "services/hwfit/data;services/hwfit/data",
+    "--add-data", "config;config",
+    "--add-data", ".env.example;.env.example"
+)
+
+& $pyExe -m PyInstaller --noconfirm --clean --onedir --noconsole --icon=static/icon.ico --name Odysseus @dataArgs launcher.py
+if ($LASTEXITCODE -ne 0) { Fail "PyInstaller build failed." }
+
+Write-Host ""
+Write-Host "Build complete." -ForegroundColor Green
+Write-Host "Portable app folder: $PSScriptRoot\dist\Odysseus" -ForegroundColor Green
+Write-Host "Distribute the whole folder (or zip it) so static assets and scripts stay with the exe." -ForegroundColor Green

companion/routes.py

@@ -5,8 +5,9 @@ offers and pair to it, without duplicating any LLM logic.
 
 Auth is enforced globally by AuthMiddleware (app.py), so reaching a handler here
 means the caller is authenticated by either a cookie session or a Bearer `ody_`
-API token. The read endpoints (ping/info/models) accept either; the pairing
-endpoints are admin-cookie only.
+API token. Ping/info accept either credential type, models requires a chat-
+scoped API token for bearer callers, and the pairing endpoints are admin-cookie
+only.
 
 Pairing CSRF posture: minting happens ONLY on POST. The session cookie is
 SameSite=Lax (routes/auth_routes.py), which a browser does not send on a
@@ -18,7 +19,7 @@ on a GET would be unsafe (Lax cookies ride top-level GET navigations), so GET
 
 import html
 
-from fastapi import APIRouter, Request
+from fastapi import APIRouter, HTTPException, Request
 from fastapi.responses import HTMLResponse
 
 from core.middleware import require_admin
@@ -52,6 +53,18 @@ def owner_can_see(row_owner, owner) -> bool:
     return row_owner is None or row_owner == owner
 
 
+def require_models_scope(request: Request) -> None:
+    """Require the companion chat scope for bearer-token model inventory."""
+    if not getattr(request.state, "api_token", False):
+        return
+    scopes = getattr(request.state, "api_token_scopes", None) or []
+    if isinstance(scopes, str):
+        scopes = [scope.strip() for scope in scopes.split(",")]
+    scope_set = {str(scope).strip() for scope in scopes if str(scope).strip()}
+    if _pairing.COMPANION_SCOPE not in scope_set:
+        raise HTTPException(403, "API token requires chat scope")
+
+
 def mint_pairing_token(owner: str, invalidate=None) -> tuple[str, str]:
     """Mint a pairing token AND invalidate the auth middleware's in-memory token
     cache, so the new token is accepted on the very next request without a server
@@ -103,6 +116,7 @@ def setup_companion_routes() -> APIRouter:
         rows -- the same rule as owner_filter. Read-only; never returns api_key
         material.
         """
+        require_models_scope(request)
         import json as _json
 
         from core.database import SessionLocal, ModelEndpoint

core/auth.py

@@ -20,6 +20,7 @@ logger = logging.getLogger(__name__)
 
 
 from core.atomic_io import atomic_write_json as _atomic_write_json  # noqa: E402
+from core.middleware import INTERNAL_TOOL_USER  # noqa: E402
 
 DEFAULT_PRIVILEGES = {
     "can_use_agent": True,
@@ -47,7 +48,7 @@ ADMIN_PRIVILEGES["allowed_models_restricted"] = False
 # backwards for this sentinel.
 ADMIN_PRIVILEGES["block_all_models"] = False
 
-from src.constants import AUTH_FILE
+from src.constants import AUTH_FILE, PASSWORD_MIN_LENGTH
 DEFAULT_AUTH_PATH = AUTH_FILE
 TOKEN_TTL = 60 * 60 * 24 * 7  # 7 days
 
@@ -65,7 +66,7 @@ TOKEN_TTL = 60 * 60 * 24 * 7  # 7 days
 # of those names would be denied an assistant and inconsistently owner-scoped.
 # Refuse to create or rename into any of them so the sentinels can't be
 # impersonated. (Keep this in sync with that synthetic-owner set.)
-RESERVED_USERNAMES = frozenset({"internal-tool", "api", "demo", "system"})
+RESERVED_USERNAMES = frozenset({INTERNAL_TOOL_USER, "api", "demo", "system"})
 
 
 def normalize_known_username(users: Dict[str, Any], username: str | None) -> Optional[str]:
@@ -243,6 +244,15 @@ class AuthManager:
     def is_configured(self) -> bool:
         return len(self.users) > 0
 
+    def policy(self) -> dict:
+        """Return public auth policy constants for the frontend."""
+        return {
+            "password_min_length": PASSWORD_MIN_LENGTH,
+            "reserved_usernames": sorted(RESERVED_USERNAMES),
+            "signup_enabled": self.signup_enabled,
+            "session_days": TOKEN_TTL // 86400,
+        }
+
     # ------------------------------------------------------------------
     # Account management
     # ------------------------------------------------------------------
@@ -573,16 +583,20 @@ class AuthManager:
             return None
         return self.create_session_trusted(username)
 
-    def create_session_trusted(self, username: str) -> str:
+    def create_session_trusted(self, username: str) -> Optional[str]:
         """Issue a session token for an already-verified user.
         Call only after verify_password (and TOTP if enabled) have passed."""
         username = username.strip().lower()
         token = secrets.token_hex(32)
-        with self._sessions_lock:
-            self._sessions[token] = {
-                "username": username,
-                "expiry": time.time() + TOKEN_TTL,
-            }
+        with self._config_lock:
+            if username not in self.users:
+                logger.warning("Refused to issue session for missing user '%s'", username)
+                return None
+            with self._sessions_lock:
+                self._sessions[token] = {
+                    "username": username,
+                    "expiry": time.time() + TOKEN_TTL,
+                }
         self._save_sessions()
         return token
 

core/database.py

@@ -2,12 +2,15 @@ import os
 import logging
 import sqlite3
 from datetime import datetime, timezone
+from pathlib import Path
 from sqlalchemy import event, create_engine, Column, String, Text, Boolean, DateTime, Integer, ForeignKey, JSON, Index, func, text
 from sqlalchemy.engine import Engine
 from sqlalchemy.types import TypeDecorator
 from sqlalchemy.ext.declarative import declarative_base, declared_attr
 from sqlalchemy.orm import relationship, sessionmaker, backref
 
+from src.runtime_paths import get_app_root
+
 logger = logging.getLogger(__name__)
 
 # Create base class for declarative models
@@ -29,9 +32,26 @@ class TimestampMixin:
     def updated_at(cls):
         return Column(DateTime, default=utcnow_naive, onupdate=utcnow_naive, nullable=False)
 
-# Get database URL from environment, default to SQLite in DATA_DIR
+# Ensure the writable data directory exists before SQLite connects.
 from src.constants import DATA_DIR, AUTH_FILE, MEMORY_FILE, USER_PREFS_FILE, SETTINGS_FILE
-DATABASE_URL = os.getenv("DATABASE_URL", f"sqlite:///{DATA_DIR}/app.db")
+Path(DATA_DIR).mkdir(parents=True, exist_ok=True)
+
+
+def _default_database_url() -> str:
+    return f"sqlite:///{Path(DATA_DIR) / 'app.db'}"
+
+
+def _normalize_sqlite_url(url: str) -> str:
+    if not url.startswith("sqlite:///"):
+        return url
+    db_path = url.replace("sqlite:///", "", 1)
+    if db_path == ":memory:" or os.path.isabs(db_path):
+        return url
+    return f"sqlite:///{(Path(get_app_root()) / db_path).resolve().as_posix()}"
+
+
+# Get database URL from environment, default to SQLite in DATA_DIR
+DATABASE_URL = _normalize_sqlite_url(os.getenv("DATABASE_URL", _default_database_url()))
 
 # Create engine
 engine = create_engine(

core/middleware.py

@@ -15,6 +15,8 @@ from starlette.responses import Response
 # same value from this module. Never persisted or exposed externally.
 INTERNAL_TOOL_TOKEN = os.environ.get("ODYSSEUS_INTERNAL_TOKEN") or secrets.token_hex(32)
 INTERNAL_TOOL_HEADER = "X-Odysseus-Internal-Token"
+# Pseudo-username on in-process tool-loopback requests; require_admin trusts it and it is reserved.
+INTERNAL_TOOL_USER = "internal-tool"
 
 
 def is_cors_preflight(method: str, headers) -> bool:
@@ -39,7 +41,7 @@ def require_admin(request: Request):
         hdr = request.headers.get(INTERNAL_TOOL_HEADER)
         if hdr and secrets.compare_digest(hdr, INTERNAL_TOOL_TOKEN):
             return
-        if getattr(request.state, "current_user", None) == "internal-tool":
+        if getattr(request.state, "current_user", None) == INTERNAL_TOOL_USER:
             return
     except Exception:
         pass
@@ -65,10 +67,9 @@ class SecurityHeadersMiddleware(BaseHTTPMiddleware):
         response = await call_next(request)
         path = request.url.path
 
-        # Tool render endpoints are served inside iframes — allow framing by self
+        # Tool render endpoints
         is_tool_render = path.startswith("/api/tools/") and path.endswith("/render")
-        # PDF previews are embedded by the in-app document library. Keep the
-        # exception route-scoped so normal app pages remain unframeable.
+        # Document library PDF preview endpoint
         is_document_pdf_preview = path.startswith("/api/document/") and path.endswith("/render-pdf")
         # Visual report pages are self-contained HTML — need inline scripts + external images
         is_report = path.startswith("/api/research/report/")
@@ -95,9 +96,7 @@ class SecurityHeadersMiddleware(BaseHTTPMiddleware):
                 "frame-ancestors 'none'"
             )
         elif is_tool_render:
-            # Tool iframe content: skip all framing headers — the iframe's
-            # sandbox="allow-scripts" attribute provides isolation.
-            # Don't overwrite the route's own restrictive CSP either.
+            # Skip framing headers for tools.
             pass
         elif is_document_pdf_preview:
             response.headers["X-Frame-Options"] = "SAMEORIGIN"

docker-compose.gpu-amd.yml

@@ -28,6 +28,14 @@ services:
       # land under /app/.local for the odysseus user. Persist them so a
       # container recreate does not silently remove installed serve engines.
       - ${APP_DATA_DIR:-./data}/local:/app/.local:z
+      # Docker socket — lets Cookbook launch commands like
+      # `docker exec ollama-rocm ollama show <tag>` reach the host's
+      # Docker daemon (and sibling containers like ollama-rocm /
+      # ollama-test). The in-container user needs to be in the
+      # socket's owning group — see `group_add` below; the GID
+      # there must match the host's `docker` group (defaults to 963
+      # on Debian, 999 on Ubuntu — override via env if yours differs).
+      - /var/run/docker.sock:/var/run/docker.sock
     extra_hosts:
       # Lets the container reach local services on the Docker host, including
       # Ollama at http://host.docker.internal:11434.
@@ -60,6 +68,13 @@ services:
       - ODYSSEUS_INPROCESS_TASKS=${ODYSSEUS_INPROCESS_TASKS:-1}
       - ODYSSEUS_SCRIPT_HOST=${ODYSSEUS_SCRIPT_HOST:-localhost}
       - ODYSSEUS_CHAT_UPLOAD_MAX_BYTES=${ODYSSEUS_CHAT_UPLOAD_MAX_BYTES:-10485760}
+      - ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES:-104857600}
+      - ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_MEMORY_IMPORT_MAX_BYTES=${ODYSSEUS_MEMORY_IMPORT_MAX_BYTES:-10485760}
+      - ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES=${ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=${ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
+      - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}
@@ -86,6 +101,7 @@ services:
       - /dev/kfd
       - /dev/dri
     group_add:
+      - "${DOCKER_GID:-963}"
       - video
       - ${RENDER_GID:-render}
 

docker-compose.gpu-nvidia.yml

@@ -27,6 +27,16 @@ services:
       # land under /app/.local for the odysseus user. Persist them so a
       # container recreate does not silently remove installed serve engines.
       - ${APP_DATA_DIR:-./data}/local:/app/.local:z
+      # Docker socket — lets Cookbook launch commands like
+      # `docker exec ollama-rocm ollama show <tag>` reach the host's
+      # Docker daemon (and sibling containers like ollama-rocm /
+      # ollama-test). The in-container user needs to be in the
+      # socket's owning group — see `group_add` below; the GID
+      # there must match the host's `docker` group (defaults to 963
+      # on Debian, 999 on Ubuntu — override via env if yours differs).
+      - /var/run/docker.sock:/var/run/docker.sock
+    group_add:
+      - "${DOCKER_GID:-963}"
     extra_hosts:
       # Lets the container reach local services on the Docker host, including
       # Ollama at http://host.docker.internal:11434.
@@ -59,6 +69,13 @@ services:
       - ODYSSEUS_INPROCESS_TASKS=${ODYSSEUS_INPROCESS_TASKS:-1}
       - ODYSSEUS_SCRIPT_HOST=${ODYSSEUS_SCRIPT_HOST:-localhost}
       - ODYSSEUS_CHAT_UPLOAD_MAX_BYTES=${ODYSSEUS_CHAT_UPLOAD_MAX_BYTES:-10485760}
+      - ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES:-104857600}
+      - ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_MEMORY_IMPORT_MAX_BYTES=${ODYSSEUS_MEMORY_IMPORT_MAX_BYTES:-10485760}
+      - ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES=${ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=${ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
+      - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}

docker-compose.yml

@@ -16,6 +16,16 @@ services:
       # land under /app/.local for the odysseus user. Persist them so a
       # container recreate does not silently remove installed serve engines.
       - ${APP_DATA_DIR:-./data}/local:/app/.local:z
+      # Docker socket — lets Cookbook launch commands like
+      # `docker exec ollama-rocm ollama show <tag>` reach the host's
+      # Docker daemon (and sibling containers like ollama-rocm /
+      # ollama-test). The in-container user needs to be in the
+      # socket's owning group — see `group_add` below; the GID
+      # there must match the host's `docker` group (defaults to 963
+      # on Debian, 999 on Ubuntu — override via env if yours differs).
+      - /var/run/docker.sock:/var/run/docker.sock
+    group_add:
+      - "${DOCKER_GID:-963}"
     extra_hosts:
       # Lets the container reach local services on the Docker host, including
       # Ollama at http://host.docker.internal:11434.
@@ -48,6 +58,13 @@ services:
       - ODYSSEUS_INPROCESS_TASKS=${ODYSSEUS_INPROCESS_TASKS:-1}
       - ODYSSEUS_SCRIPT_HOST=${ODYSSEUS_SCRIPT_HOST:-localhost}
       - ODYSSEUS_CHAT_UPLOAD_MAX_BYTES=${ODYSSEUS_CHAT_UPLOAD_MAX_BYTES:-10485760}
+      - ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_UPLOAD_MAX_BYTES:-104857600}
+      - ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES=${ODYSSEUS_GALLERY_TRANSFORM_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_MEMORY_IMPORT_MAX_BYTES=${ODYSSEUS_MEMORY_IMPORT_MAX_BYTES:-10485760}
+      - ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES=${ODYSSEUS_PERSONAL_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES=${ODYSSEUS_EMAIL_COMPOSE_UPLOAD_MAX_BYTES:-26214400}
+      - ODYSSEUS_STT_MAX_AUDIO_BYTES=${ODYSSEUS_STT_MAX_AUDIO_BYTES:-26214400}
+      - ODYSSEUS_ICS_MAX_BYTES=${ODYSSEUS_ICS_MAX_BYTES:-10485760}
       - DATA_BRAVE_API_KEY=${DATA_BRAVE_API_KEY:-}
       - GOOGLE_API_KEY=${GOOGLE_API_KEY:-}
       - GOOGLE_PSE_CX=${GOOGLE_PSE_CX:-}

docker/entrypoint.sh

@@ -13,6 +13,8 @@ set -e
 
 PUID="${PUID:-1000}"
 PGID="${PGID:-1000}"
+GOSU_BIN="$(command -v gosu)"
+PYTHON_BIN="$(command -v python)"
 
 # Reuse an existing matching group/user if the host's UID/GID already
 # corresponds to one in /etc/passwd (e.g. when the image is rebuilt
@@ -24,26 +26,78 @@ if ! getent passwd "$PUID" >/dev/null 2>&1; then
     useradd -u "$PUID" -g "$PGID" -M -s /bin/sh -d /app odysseus
 fi
 
-# Repair ownership on every writable path the app touches at runtime.
-#
-# Bind-mounted dirs (/app/data, /app/logs) are the obvious ones, but
-# the app ALSO writes inside the image's own source tree at runtime:
-#   - services/cache/{search,content}/*  (search cache LRU)
-#   - services/search_analytics.json
-#   - services/search_engine_error.log
-#   - services/tts cache, etc.
-# These dirs were created as root during `docker build`, so dropping
-# to PUID:PGID would otherwise crash on the first import that tries
-# to mkdir them. Chown the whole /app tree — fast (<1s on this size)
-# and idempotent via the `-not -uid` filter so we only touch files
-# that need fixing.
-for dir in /app /app/data /app/logs; do
+ODY_USER="$(getent passwd "$PUID" | cut -d: -f1)"
+[ -z "$ODY_USER" ] && ODY_USER=odysseus
+
+# Docker-socket group plumbing. When /var/run/docker.sock is bind-mounted
+# (Cookbook uses docker exec to reach sibling containers), the socket is
+# owned by root:<host docker gid>. Add the app user to that group and later
+# call gosu by username so supplementary groups are retained.
+DOCKER_SOCK="${DOCKER_SOCK:-/var/run/docker.sock}"
+if [ -S "$DOCKER_SOCK" ]; then
+    SOCK_GID="$(stat -c '%g' "$DOCKER_SOCK" 2>/dev/null || echo '')"
+    if [ -n "$SOCK_GID" ] && [ "$SOCK_GID" != "0" ]; then
+        if ! getent group "$SOCK_GID" >/dev/null 2>&1; then
+            groupadd -g "$SOCK_GID" docker_host || true
+        fi
+        SOCK_GROUP="$(getent group "$SOCK_GID" | cut -d: -f1)"
+        if [ -n "$SOCK_GROUP" ]; then
+            usermod -aG "$SOCK_GROUP" "$ODY_USER" 2>/dev/null || true
+        fi
+    fi
+fi
+
+mount_root_for() {
+    awk -v target="$1" '$5 == target { print $4; exit }' /proc/self/mountinfo 2>/dev/null || true
+}
+
+is_broad_mount_root() {
+    case "$1" in
+        /|/home|/srv|/var|/usr|/opt|/tmp|/mnt|/media)
+            return 0
+            ;;
+    esac
+    return 1
+}
+
+repair_tree_ownership() {
+    dir="$1"
     if [ -d "$dir" ]; then
-        # `find ... -not -uid` keeps this O(touched-files), not
-        # O(everything), so terabyte-sized maildirs don't slow startup.
-        find "$dir" -not -uid "$PUID" -print0 2>/dev/null \
+        find "$dir" -xdev -not -uid "$PUID" -print0 2>/dev/null \
             | xargs -0 -r chown "$PUID:$PGID" 2>/dev/null || true
     fi
+}
+
+repair_app_tree_ownership() {
+    if [ -d /app ]; then
+        find /app -xdev \
+            \( -path /app/data -o -path /app/logs -o -path /app/.ssh -o -path /app/.cache -o -path /app/.local \) -prune \
+            -o -not -uid "$PUID" -print0 2>/dev/null \
+            | xargs -0 -r chown "$PUID:$PGID" 2>/dev/null || true
+    fi
+}
+
+repair_bind_mount_ownership() {
+    dir="$1"
+    if [ ! -d "$dir" ]; then
+        return
+    fi
+
+    mount_root="$(mount_root_for "$dir")"
+    if is_broad_mount_root "$mount_root"; then
+        echo "Skipping recursive ownership repair for $dir because it maps to broad host path $mount_root" >&2
+        chown "$PUID:$PGID" "$dir" 2>/dev/null || true
+        return
+    fi
+
+    repair_tree_ownership "$dir"
+}
+
+# Repair image-owned writable paths without walking into bind-mounted host
+# trees, then repair the app-owned mount roots separately.
+repair_app_tree_ownership
+for dir in /app/data /app/logs /app/.ssh /app/.cache/huggingface /app/.local; do
+    repair_bind_mount_ownership "$dir"
 done
 
 # Cookbook installs vllm/etc. via `pip install --user`, which pulls
@@ -70,6 +124,7 @@ for cu in \
         break
     fi
 done
+
 # Disable the FlashInfer JIT sampler unconditionally — it is sampler-only
 # and has no impact on the attention path, but requires nvcc + matching
 # CUDA headers at startup. Without this, vLLM crashes with "Could not find
@@ -83,9 +138,9 @@ export PATH="/app/.local/bin:$PATH"
 # Run first-time setup as the app user so data/ files get the right ownership.
 # setup.py is idempotent — skips auth.json / .env if they already exist.
 # || true so a setup failure never prevents the container from starting.
-gosu "$PUID:$PGID" python /app/setup.py || true
+"$GOSU_BIN" "$ODY_USER" "$PYTHON_BIN" /app/setup.py || true
 
 # Drop root and run the actual app. `gosu` is preferred over `su` /
 # `sudo` because it cleans up the process tree (no extra shell layer)
 # so signals (SIGTERM from `docker stop`) reach uvicorn directly.
-exec gosu "$PUID:$PGID" "$@"
+exec "$GOSU_BIN" "$ODY_USER" "$@"

launch-windows.ps1

@@ -105,6 +105,14 @@ if (-not $pyExe) {
     }
 }
 
+if ($pyExe -like "*WindowsApps*python.exe") {
+    $pyCmd = Get-Command py -ErrorAction SilentlyContinue
+    if ($pyCmd) {
+        $pyExe = $pyCmd.Source
+        $pyArgs = @("-3.11")
+    }
+}
+
 if (-not $pyExe) {
     Fail "Couldn't find Python 3.11+ for Windows setup. Install Python 3.11+ (or open the Python launcher with 'py -3.11') from https://www.python.org/downloads/, then re-run this script."
 }

launcher.py

@@ -0,0 +1,142 @@
+# launcher.py
+"""Dedicated entrypoint for the standalone Windows portable launcher.
+
+Handles:
+- Immediate GUI splash screen creation using tkinter.
+- Suppressing console stream crashes in windowed GUI mode via NullWriter.
+- Spawning system tray icon via pystray and Pillow (lazy-loaded).
+- Auto-opening default browser pointing to the running backend.
+- Launching the FastAPI server (importing and running app.py).
+"""
+import os
+import sys
+import threading
+import time
+import webbrowser
+
+# Define a dummy NullWriter to suppress standard stream crashes (isatty etc.) in GUI mode
+class NullWriter:
+    def write(self, text):
+        pass
+    def flush(self):
+        pass
+    def isatty(self):
+        return False
+
+if sys.stdout is None:
+    sys.stdout = NullWriter()
+if sys.stderr is None:
+    sys.stderr = NullWriter()
+
+
+splash_root = None
+
+# If running from a frozen PyInstaller bundle, launch the splash screen IMMEDIATELY
+if getattr(sys, 'frozen', False):
+    import tkinter as tk
+
+    def show_splash_instantly():
+        global splash_root
+        try:
+            splash_root = tk.Tk()
+            splash_root.title("Odysseus")
+            splash_root.overrideredirect(True)
+            splash_root.configure(bg="#1a1c23")
+
+            # Accented borders
+            splash_root.config(highlightbackground="#e06c75", highlightcolor="#e06c75", highlightthickness=1)
+
+            w, h = 360, 160
+            ws = splash_root.winfo_screenwidth()
+            hs = splash_root.winfo_screenheight()
+            x = (ws - w) // 2
+            y = (hs - h) // 2
+            splash_root.geometry(f"{w}x{h}+{x}+{y}")
+
+            tk.Label(splash_root, text="⛵ Odysseus", font=("Segoe UI", 22, "bold"), bg="#1a1c23", fg="#e06c75").pack(pady=(22, 2))
+            tk.Label(splash_root, text="Launching background services...", font=("Segoe UI", 10), bg="#1a1c23", fg="#d1d4e0").pack(pady=2)
+            tk.Label(splash_root, text="Please wait, this will take a few seconds.", font=("Segoe UI", 8, "italic"), bg="#1a1c23", fg="#5c6370").pack(pady=(12, 0))
+
+            splash_root.attributes("-topmost", True)
+            splash_root.mainloop()
+        except Exception:
+            pass
+
+    # Launch the GUI splash screen immediately on a background thread
+    threading.Thread(target=show_splash_instantly, daemon=True).start()
+
+
+def create_tray_image():
+    # Generate a beautiful 64x64 icon matching Odysseus brand red accent (#e06c75)
+    from PIL import Image, ImageDraw
+    image = Image.new('RGBA', (64, 64), (0, 0, 0, 0))
+    dc = ImageDraw.Draw(image)
+    accent_red = (224, 108, 117, 255)
+    light_red = (224, 108, 117, 150)
+
+    # Draw premium sailing boat
+    dc.polygon([(32, 10), (32, 45), (12, 45)], fill=accent_red)
+    dc.polygon([(32, 18), (32, 45), (48, 45)], fill=light_red)
+    dc.polygon([(8, 48), (56, 48), (44, 56), (20, 56)], fill=accent_red)
+    return image
+
+
+def on_open_browser(icon, item, url):
+    webbrowser.open(url)
+
+
+def on_exit(icon, item):
+    icon.stop()
+    os._exit(0)
+
+
+def setup_system_tray(url):
+    try:
+        import pystray
+        icon_img = create_tray_image()
+        menu = (
+            pystray.MenuItem('Open Odysseus', lambda icon, item: on_open_browser(icon, item, url), default=True),
+            pystray.MenuItem('Exit', on_exit)
+        )
+        tray_icon = pystray.Icon(
+            "Odysseus",
+            icon_img,
+            "Odysseus",
+            menu
+        )
+        tray_icon.run()
+    except Exception:
+        pass
+
+
+def open_browser(url):
+    # Allow uvicorn and app lifecycles to complete warmups
+    time.sleep(3.5)
+
+    # Safely close the splash screen
+    try:
+        global splash_root
+        if splash_root:
+            splash_root.after(0, splash_root.destroy)
+    except Exception:
+        pass
+
+    webbrowser.open(url)
+
+
+if __name__ == "__main__":
+    import uvicorn
+    # Import the FastAPI app from app.py
+    from app import app
+
+    bind_host = os.getenv("APP_BIND", "127.0.0.1")
+    bind_port = int(os.getenv("APP_PORT", "7000"))
+    url = f"http://{bind_host}:{bind_port}"
+
+    if getattr(sys, 'frozen', False):
+        # Start browser manager thread
+        threading.Thread(target=open_browser, args=(url,), daemon=True).start()
+        # Start system tray manager thread
+        threading.Thread(target=setup_system_tray, args=(url,), daemon=True).start()
+
+    uvicorn.run(app, host=bind_host, port=bind_port, log_level="info")

mcp_servers/email_server.py

@@ -23,6 +23,7 @@ import os.path
 from pathlib import Path
 from datetime import datetime, timedelta
 import uuid
+from contextvars import ContextVar
 
 from mcp.server import Server
 from mcp.server.stdio import stdio_server
@@ -55,6 +56,8 @@ def _uid_fetch_rows(data) -> list:
 # flat keys when no DB row matches (legacy single-account behaviour).
 
 _ACCOUNT_CACHE: dict = {}  # key = normalized account selector -> config dict
+_MCP_OWNER_ARG = "_odysseus_owner"
+_CURRENT_OWNER: ContextVar[str | None] = ContextVar("email_mcp_owner", default=None)
 
 
 def _clean_header_value(value) -> str:
@@ -68,6 +71,45 @@ def _db_path() -> Path:
     return Path(APP_DB)
 
 
+def _current_owner() -> str:
+    owner = _CURRENT_OWNER.get()
+    return str(owner or "").strip()
+
+
+def _account_visible_to_owner(row: dict, owner: str) -> bool:
+    row_owner = str(row.get("owner") or "").strip()
+    if row_owner == owner:
+        return True
+    if row_owner:
+        return False
+    # Legacy ownerless accounts are only visible to a scoped caller when the
+    # mailbox itself matches the owner, mirroring the HTTP email route fallback.
+    owner_l = owner.lower()
+    return owner_l in {
+        str(row.get("imap_user") or "").strip().lower(),
+        str(row.get("from_address") or "").strip().lower(),
+    }
+
+
+def _filter_accounts_for_owner(rows: list[dict]) -> list[dict]:
+    owner = _current_owner()
+    if owner:
+        return [r for r in rows if _account_visible_to_owner(r, owner)]
+
+    owners = {str(r.get("owner") or "").strip() for r in rows if str(r.get("owner") or "").strip()}
+    if len(owners) > 1:
+        return []
+    return rows
+
+
+def _mcp_owner_required(rows: list[dict] | None = None) -> bool:
+    if _current_owner():
+        return False
+    rows = rows if rows is not None else _read_accounts_from_db()
+    owners = {str(r.get("owner") or "").strip() for r in rows if str(r.get("owner") or "").strip()}
+    return len(owners) > 1
+
+
 def _load_email_writing_style() -> str:
     """Return the existing Settings > Email > Writing Style value."""
     try:
@@ -121,9 +163,8 @@ def _default_document_owner() -> str | None:
         return None
 
 
-def _list_accounts_raw() -> list:
-    """Return list of dicts from the email_accounts table. Empty list if table
-    missing or empty. Never raises."""
+def _read_accounts_from_db() -> list:
+    """Return all enabled email account rows. Empty list if missing. Never raises."""
     path = _db_path()
     if not path.exists():
         return []
@@ -131,9 +172,10 @@ def _list_accounts_raw() -> list:
         conn = sqlite3.connect(str(path))
         conn.row_factory = sqlite3.Row
         columns = {r[1] for r in conn.execute("PRAGMA table_info(email_accounts)").fetchall()}
+        owner_select = "owner" if "owner" in columns else "NULL AS owner"
         smtp_security_select = "smtp_security" if "smtp_security" in columns else "'' AS smtp_security"
         rows = conn.execute(f"""
-            SELECT id, name, is_default, enabled,
+            SELECT id, {owner_select}, name, is_default, enabled,
                    imap_host, imap_port, imap_user, imap_password, imap_starttls,
                    smtp_host, smtp_port, {smtp_security_select}, smtp_user, smtp_password, from_address
             FROM email_accounts WHERE enabled = 1
@@ -147,11 +189,15 @@ def _list_accounts_raw() -> list:
         return []
 
 
-def _resolve_account(selector: str | None) -> dict | None:
+def _list_accounts_raw() -> list:
+    """Return owner-visible email account rows for the active MCP call."""
+    return _filter_accounts_for_owner(_read_accounts_from_db())
+
+
+def _resolve_account_from_rows(rows: list[dict], selector: str | None) -> dict | None:
     """Given a selector (None = default, or a name/user/id string), return the
     matching row or None. Matching is case-insensitive substring on name +
     imap_user + from_address, plus exact id match."""
-    rows = _list_accounts_raw()
     if not rows:
         return None
     if not selector:
@@ -186,6 +232,10 @@ def _resolve_account(selector: str | None) -> dict | None:
     return None
 
 
+def _resolve_account(selector: str | None) -> dict | None:
+    return _resolve_account_from_rows(_list_accounts_raw(), selector)
+
+
 def _load_config(account: str | None = None) -> dict:
     """Return the full config dict for the requested account (or default).
 
@@ -194,7 +244,7 @@ def _load_config(account: str | None = None) -> dict:
       2. env vars + settings.json flat keys (legacy)
       3. hardcoded fallbacks (localhost:31143 etc.)
     """
-    cache_key = (account or "").strip().lower() or "__default__"
+    cache_key = (_current_owner(), (account or "").strip().lower() or "__default__")
     if cache_key in _ACCOUNT_CACHE:
         return _ACCOUNT_CACHE[cache_key]
 
@@ -223,8 +273,11 @@ def _load_config(account: str | None = None) -> dict:
         "account_name": None,
     }
 
-    rows = _list_accounts_raw()
-    row = _resolve_account(account)
+    raw_rows = _read_accounts_from_db()
+    rows = _filter_accounts_for_owner(raw_rows)
+    row = _resolve_account_from_rows(rows, account)
+    if _current_owner() and raw_rows and not rows:
+        raise ValueError("No email account is configured for the authenticated owner")
     if account and rows and not row:
         available = ", ".join(
             f"{r.get('name') or r.get('imap_user')} <{r.get('imap_user') or r.get('from_address') or '?'}>"
@@ -953,7 +1006,7 @@ def _stash_agent_draft(*, to, subject, body, in_reply_to=None, references=None,
             now,
             account or None,
             "agent_draft",
-            "",
+            _current_owner(),
         ))
         conn.commit()
         conn.close()
@@ -1139,7 +1192,7 @@ def _create_email_draft_document(
     doc_id = str(uuid.uuid4())
     ver_id = str(uuid.uuid4())
     doc_title = (title or subject or "Email draft").strip() or "Email draft"
-    doc_owner = _default_document_owner()
+    doc_owner = _current_owner() or _default_document_owner()
 
     db = SessionLocal()
     try:
@@ -1925,10 +1978,22 @@ async def list_tools() -> list[Tool]:
 
 @server.call_tool()
 async def call_tool(name: str, arguments: dict) -> list[TextContent]:
+    arguments = dict(arguments) if isinstance(arguments, dict) else {}
+    owner = str(arguments.pop(_MCP_OWNER_ARG, "") or "").strip()
+    owner_token = _CURRENT_OWNER.set(owner or None)
     try:
+        all_db_accounts = _read_accounts_from_db()
+        if _mcp_owner_required(all_db_accounts):
+            return [TextContent(
+                type="text",
+                text="Error: email MCP requires an authenticated owner when multiple email account owners are configured.",
+            )]
+
         if name == "list_email_accounts":
-            rows = _list_accounts_raw()
+            rows = _filter_accounts_for_owner(all_db_accounts)
             if not rows:
+                if all_db_accounts and owner:
+                    return [TextContent(type="text", text="No email accounts configured for this owner.")]
                 return [TextContent(type="text", text="No email accounts configured. Legacy single-account mode active.")]
             lines = [f"Found {len(rows)} email account(s):\n"]
             for r in rows:
@@ -2108,6 +2173,16 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
                 bcc=arguments.get("bcc"),
                 account=acct,
             )
+            if "error" in result:
+                return [TextContent(type="text", text=f"Error: {result['error']}")]
+            if result.get("pending"):
+                return [TextContent(
+                    type="text",
+                    text=(
+                        f"Draft staged for approval (pending id: {result.get('pending_id')}). "
+                        "Nothing has been sent yet. Review and approve it in Odysseus before delivery."
+                    ),
+                )]
             acct_note = f" (from {result['account']})" if result.get("account") else ""
             return [TextContent(type="text", text=f"Sent email to {result['to']} with subject '{result['subject']}'{acct_note}.")]
 
@@ -2283,6 +2358,8 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
 
     except Exception as e:
         return [TextContent(type="text", text=f"Error: {e}")]
+    finally:
+        _CURRENT_OWNER.reset(owner_token)
 
 
 # ── Main ──

mcp_servers/memory_server.py

@@ -6,6 +6,7 @@ Imports MemoryManager and MemoryVectorStore from the Odysseus codebase.
 """
 
 import asyncio
+import os
 import sys
 import time
 from pathlib import Path
@@ -23,6 +24,55 @@ _memory_manager = None
 _memory_vector = None
 _initialized = False
 
+_OWNER_ENV_KEYS = ("ODYSSEUS_MCP_MEMORY_OWNER", "ODYSSEUS_MEMORY_OWNER")
+_OWNER_SCOPE_ERROR = (
+    "Error: Memory MCP owner is not configured for an owner-scoped memory store. "
+    "Set ODYSSEUS_MCP_MEMORY_OWNER for this server or use the owner-aware native memory tool."
+)
+
+
+def _configured_owner() -> str | None:
+    for key in _OWNER_ENV_KEYS:
+        owner = os.environ.get(key, "").strip()
+        if owner:
+            return owner
+    return None
+
+
+def _entry_owner(entry: dict) -> str | None:
+    owner = entry.get("owner")
+    if owner is None:
+        return None
+    owner_text = str(owner).strip()
+    return owner_text or None
+
+
+def _owner_scoped_store(entries: list[dict]) -> bool:
+    return any(_entry_owner(entry) for entry in entries if isinstance(entry, dict))
+
+
+def _scope_entries() -> tuple[str | None, list[dict], list[dict], str | None]:
+    """Return configured owner, all entries, visible entries, and optional error."""
+    entries = _memory_manager.load_all()
+    owner = _configured_owner()
+    if owner is None and _owner_scoped_store(entries):
+        return None, entries, [], _OWNER_SCOPE_ERROR
+    if owner is None:
+        visible = [
+            entry for entry in entries
+            if isinstance(entry, dict) and _entry_owner(entry) is None
+        ]
+    else:
+        visible = [
+            entry for entry in entries
+            if isinstance(entry, dict) and _entry_owner(entry) == owner
+        ]
+    return owner, entries, visible, None
+
+
+def _text_result(text: str) -> list[TextContent]:
+    return [TextContent(type="text", text=text)]
+
 
 def _ensure_init():
     """Lazy-init memory managers on first use."""
@@ -75,24 +125,26 @@ async def list_tools() -> list[Tool]:
 @server.call_tool()
 async def call_tool(name: str, arguments: dict) -> list[TextContent]:
     if name != "manage_memory":
-        return [TextContent(type="text", text=f"Unknown tool: {name}")]
+        return _text_result(f"Unknown tool: {name}")
 
     _ensure_init()
     if not _memory_manager:
-        return [TextContent(type="text", text="Error: Memory manager not available")]
+        return _text_result("Error: Memory manager not available")
 
     action = arguments.get("action", "")
 
     if action == "list":
         category_filter = arguments.get("category", "")
-        memories = _memory_manager.load()
+        _owner, _all_memories, memories, scope_error = _scope_entries()
+        if scope_error:
+            return _text_result(scope_error)
         if category_filter:
             memories = [m for m in memories if m.get("category", "").lower() == category_filter.lower()]
         if not memories:
             msg = "No memories found"
             if category_filter:
                 msg += f" in category '{category_filter}'"
-            return [TextContent(type="text", text=msg + ".")]
+            return _text_result(msg + ".")
 
         lines = [f"Found {len(memories)} memory entries:\n"]
         for m in memories:
@@ -102,15 +154,17 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
             if len(text) > 150:
                 text = text[:150] + "..."
             lines.append(f"- [{cat}] `{mid}` — {text}")
-        return [TextContent(type="text", text="\n".join(lines))]
+        return _text_result("\n".join(lines))
 
     elif action == "add":
         text = arguments.get("text", "")
         category = arguments.get("category", "fact")
         if not text:
-            return [TextContent(type="text", text="Error: Memory text cannot be empty")]
-        entry = _memory_manager.add_entry(text, source="ai_agent", category=category)
-        memories = _memory_manager.load_all()
+            return _text_result("Error: Memory text cannot be empty")
+        owner, memories, _visible, scope_error = _scope_entries()
+        if scope_error:
+            return _text_result(scope_error)
+        entry = _memory_manager.add_entry(text, source="ai_agent", category=category, owner=owner)
         memories.append(entry)
         _memory_manager.save(memories)
         if _memory_vector and _memory_vector.healthy:
@@ -118,25 +172,28 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
                 _memory_vector.add(entry["id"], text)
             except Exception:
                 pass
-        return [TextContent(type="text", text=f"Memory added: [{category}] {text} (id: {entry['id'][:8]})")]
+        return _text_result(f"Memory added: [{category}] {text} (id: {entry['id'][:8]})")
 
     elif action == "edit":
         memory_id = arguments.get("memory_id", "")
         new_text = arguments.get("text", "")
         if not memory_id or not new_text:
-            return [TextContent(type="text", text="Error: edit needs memory_id and text")]
-        memories = _memory_manager.load_all()
-        found = False
+            return _text_result("Error: edit needs memory_id and text")
+        _owner, memories, visible, scope_error = _scope_entries()
+        if scope_error:
+            return _text_result(scope_error)
         full_id = None
-        for m in memories:
+        for m in visible:
             if m.get("id", "").startswith(memory_id):
-                m["text"] = new_text
-                m["timestamp"] = int(time.time())
-                found = True
                 full_id = m["id"]
                 break
-        if not found:
-            return [TextContent(type="text", text=f"Error: Memory '{memory_id}' not found")]
+        if not full_id:
+            return _text_result(f"Error: Memory '{memory_id}' not found")
+        for m in memories:
+            if m.get("id") == full_id:
+                m["text"] = new_text
+                m["timestamp"] = int(time.time())
+                break
         _memory_manager.save(memories)
         if _memory_vector and _memory_vector.healthy and full_id:
             try:
@@ -144,24 +201,26 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
                 _memory_vector.add(full_id, new_text)
             except Exception:
                 pass
-        return [TextContent(type="text", text=f"Memory updated: {new_text}")]
+        return _text_result(f"Memory updated: {new_text}")
 
     elif action == "delete":
         memory_id = arguments.get("memory_id", "")
         if not memory_id:
-            return [TextContent(type="text", text="Error: delete needs memory_id")]
-        memories = _memory_manager.load_all()
+            return _text_result("Error: delete needs memory_id")
+        _owner, memories, visible, scope_error = _scope_entries()
+        if scope_error:
+            return _text_result(scope_error)
         full_id = None
         deleted_text = ""
         deleted_category = ""
-        for m in memories:
+        for m in visible:
             if m.get("id", "").startswith(memory_id):
                 full_id = m["id"]
                 deleted_text = m.get("text", "")
                 deleted_category = m.get("category", "")
                 break
         if not full_id:
-            return [TextContent(type="text", text=f"Error: Memory '{memory_id}' not found")]
+            return _text_result(f"Error: Memory '{memory_id}' not found")
         memories = [m for m in memories if m.get("id") != full_id]
         _memory_manager.save(memories)
         if _memory_vector and _memory_vector.healthy and full_id:
@@ -171,30 +230,32 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
                 pass
         cat = f"[{deleted_category}] " if deleted_category else ""
         snippet = deleted_text if len(deleted_text) <= 120 else deleted_text[:117] + "..."
-        return [TextContent(type="text", text=f"Memory deleted: {cat}{snippet} (id: {memory_id})")]
+        return _text_result(f"Memory deleted: {cat}{snippet} (id: {memory_id})")
 
     elif action == "search":
         query = arguments.get("text", "")
         if not query:
-            return [TextContent(type="text", text="Error: search needs text (query)")]
-        memories = _memory_manager.load()
+            return _text_result("Error: search needs text (query)")
+        _owner, _all_memories, memories, scope_error = _scope_entries()
+        if scope_error:
+            return _text_result(scope_error)
         if hasattr(_memory_manager, 'get_relevant_memories'):
             results = _memory_manager.get_relevant_memories(query, memories, threshold=0.05, max_items=20)
         else:
             query_lower = query.lower()
             results = [m for m in memories if query_lower in m.get("text", "").lower()][:20]
         if not results:
-            return [TextContent(type="text", text=f"No memories found matching '{query}'.")]
+            return _text_result(f"No memories found matching '{query}'.")
         lines = [f"Found {len(results)} matching memories:\n"]
         for m in results:
             cat = m.get("category", "fact")
             mid = m.get("id", "?")[:8]
             text = m.get("text", "")
             lines.append(f"- [{cat}] `{mid}` — {text}")
-        return [TextContent(type="text", text="\n".join(lines))]
+        return _text_result("\n".join(lines))
 
     else:
-        return [TextContent(type="text", text=f"Error: Unknown action '{action}'. Use: list, add, edit, delete, search")]
+        return _text_result(f"Error: Unknown action '{action}'. Use: list, add, edit, delete, search")
 
 
 async def run():

package-lock.json

@@ -4,93 +4,19 @@
   "requires": true,
   "packages": {
     "": {
-      "dependencies": {
-        "@anthropic-ai/sdk": "^0.104.1"
-      },
       "devDependencies": {
-        "@antithesishq/bombadil": "^0.5.0"
-      }
-    },
-    "node_modules/@anthropic-ai/sdk": {
-      "version": "0.104.1",
-      "resolved": "https://registry.npmjs.org/@anthropic-ai/sdk/-/sdk-0.104.1.tgz",
-      "integrity": "sha512-gGACa/+IaiXzRRmF96aOhamoBgapKRBiFWbmmTFP8aMkpaEcuStF+Q61bjo4vPxBM7gqWJNZqsngslRdnLHv0Q==",
-      "license": "MIT",
-      "dependencies": {
-        "json-schema-to-ts": "^3.1.1",
-        "standardwebhooks": "^1.0.0"
-      },
-      "bin": {
-        "anthropic-ai-sdk": "bin/cli"
-      },
-      "peerDependencies": {
-        "zod": "^3.25.0 || ^4.0.0"
-      },
-      "peerDependenciesMeta": {
-        "zod": {
-          "optional": true
-        }
+        "@antithesishq/bombadil": "^0.6.1"
       }
     },
     "node_modules/@antithesishq/bombadil": {
-      "version": "0.5.0",
-      "resolved": "https://registry.npmjs.org/@antithesishq/bombadil/-/bombadil-0.5.0.tgz",
-      "integrity": "sha512-s0zImmr0iyvSP6QcVLvf40CUiZYIdWBAxiq20uhzujwvfitYa3PGJN652k/pLtVccHM/JrGQxZdvLnihZpltHA==",
+      "version": "0.6.1",
+      "resolved": "https://registry.npmjs.org/@antithesishq/bombadil/-/bombadil-0.6.1.tgz",
+      "integrity": "sha512-d1iufG3MI7gSMSiSmMeNdcMW+qR0yQXL2zdkVynC3n3DYgFJYlYXKUQzygmqU12m4RWlR5iOdQU1hsx5UT6+IA==",
       "dev": true,
       "license": "MIT",
       "bin": {
         "bombadil": "bin/bombadil.js"
       }
-    },
-    "node_modules/@babel/runtime": {
-      "version": "7.29.7",
-      "resolved": "https://registry.npmjs.org/@babel/runtime/-/runtime-7.29.7.tgz",
-      "integrity": "sha512-Nq8OhGWiZIZGV6hLHoyAKLLcJihP/xFeBMGJoUrxTX2psI8dCifzLhZISFb+VWS3wFMRDmCGw5R+dOySCqPLhw==",
-      "license": "MIT",
-      "engines": {
-        "node": ">=6.9.0"
-      }
-    },
-    "node_modules/@stablelib/base64": {
-      "version": "1.0.1",
-      "resolved": "https://registry.npmjs.org/@stablelib/base64/-/base64-1.0.1.tgz",
-      "integrity": "sha512-1bnPQqSxSuc3Ii6MhBysoWCg58j97aUjuCSZrGSmDxNqtytIi0k8utUenAwTZN4V5mXXYGsVUI9zeBqy+jBOSQ==",
-      "license": "MIT"
-    },
-    "node_modules/fast-sha256": {
-      "version": "1.3.0",
-      "resolved": "https://registry.npmjs.org/fast-sha256/-/fast-sha256-1.3.0.tgz",
-      "integrity": "sha512-n11RGP/lrWEFI/bWdygLxhI+pVeo1ZYIVwvvPkW7azl/rOy+F3HYRZ2K5zeE9mmkhQppyv9sQFx0JM9UabnpPQ==",
-      "license": "Unlicense"
-    },
-    "node_modules/json-schema-to-ts": {
-      "version": "3.1.1",
-      "resolved": "https://registry.npmjs.org/json-schema-to-ts/-/json-schema-to-ts-3.1.1.tgz",
-      "integrity": "sha512-+DWg8jCJG2TEnpy7kOm/7/AxaYoaRbjVB4LFZLySZlWn8exGs3A4OLJR966cVvU26N7X9TWxl+Jsw7dzAqKT6g==",
-      "license": "MIT",
-      "dependencies": {
-        "@babel/runtime": "^7.18.3",
-        "ts-algebra": "^2.0.0"
-      },
-      "engines": {
-        "node": ">=16"
-      }
-    },
-    "node_modules/standardwebhooks": {
-      "version": "1.0.0",
-      "resolved": "https://registry.npmjs.org/standardwebhooks/-/standardwebhooks-1.0.0.tgz",
-      "integrity": "sha512-BbHGOQK9olHPMvQNHWul6MYlrRTAOKn03rOe4A8O3CLWhNf4YHBqq2HJKKC+sfqpxiBY52pNeesD6jIiLDz8jg==",
-      "license": "MIT",
-      "dependencies": {
-        "@stablelib/base64": "^1.0.0",
-        "fast-sha256": "^1.3.0"
-      }
-    },
-    "node_modules/ts-algebra": {
-      "version": "2.0.0",
-      "resolved": "https://registry.npmjs.org/ts-algebra/-/ts-algebra-2.0.0.tgz",
-      "integrity": "sha512-FPAhNPFMrkwz76P7cdjdmiShwMynZYN6SgOujD1urY4oNm80Ou9oMdmbR45LotcKOXoy7wSmHkRFE6Mxbrhefw==",
-      "license": "MIT"
     }
   }
 }

package.json

@@ -4,9 +4,6 @@
     "url": "https://github.com/pewdiepie-archdaemon/odysseus.git"
   },
   "devDependencies": {
-    "@antithesishq/bombadil": "^0.5.0"
-  },
-  "dependencies": {
-    "@anthropic-ai/sdk": "^0.104.1"
+    "@antithesishq/bombadil": "^0.6.1"
   }
 }

routes/api_token_routes.py

@@ -160,6 +160,8 @@ def setup_api_token_routes() -> APIRouter:
             payload = await request.json()
         except Exception:
             payload = {}
+        if not isinstance(payload, dict):
+            payload = {}
         with get_db_session() as db:
             token = db.query(ApiToken).filter(ApiToken.id == token_id).first()
             if not token:

routes/assistant_routes.py

@@ -16,6 +16,7 @@ from pydantic import BaseModel
 
 from core.database import SessionLocal, CrewMember, ScheduledTask
 from src.auth_helpers import get_current_user
+from core.auth import RESERVED_USERNAMES
 from src.task_scheduler import compute_next_run
 
 
@@ -89,11 +90,11 @@ def setup_assistant_routes(task_scheduler) -> APIRouter:
     # check-in tasks seeded. Hitting any /assistant route under one of these
     # used to seed a full CrewMember + Morning/Midday/Evening tasks under that
     # owner, which then double-fired alongside the real user's check-ins.
-    _SYNTHETIC_OWNERS = frozenset({"internal-tool", "api", "demo", "system", ""})
+    # RESERVED_USERNAMES covers the same set; the `not owner` guard handles "".
 
     async def _get_or_create(owner: str) -> CrewMember:
         """Return the per-owner assistant CrewMember, creating it on demand."""
-        if not owner or owner in _SYNTHETIC_OWNERS:
+        if not owner or owner in RESERVED_USERNAMES:
             raise HTTPException(status_code=400, detail=f"Cannot seed assistant for {owner!r}")
         db = SessionLocal()
         try:

routes/auth_routes.py

@@ -12,8 +12,8 @@ import re
 from pathlib import Path
 
 from core.atomic_io import atomic_write_json, atomic_write_text
-from core.auth import AuthManager, SetAdminResult
-from src.constants import DEEP_RESEARCH_DIR, MEMORY_FILE, SKILLS_DIR
+from core.auth import AuthManager, RESERVED_USERNAMES, SetAdminResult, TOKEN_TTL
+from src.constants import DEEP_RESEARCH_DIR, MEMORY_FILE, PASSWORD_MIN_LENGTH, SKILLS_DIR
 from src.rate_limiter import RateLimiter
 from src.settings_scrub import scrub_settings
 from src.settings import (
@@ -102,8 +102,12 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             raise HTTPException(429, "Too many requests — try again later")
         if auth_manager.is_configured:
             raise HTTPException(400, "Already configured")
-        if len(body.password) < 8:
-            raise HTTPException(400, "Password must be at least 8 characters")
+        if len(body.password) < PASSWORD_MIN_LENGTH:
+            raise HTTPException(400, f"Password must be at least {PASSWORD_MIN_LENGTH} characters")
+        if len(body.username.strip()) < 1:
+            raise HTTPException(400, "Username is required")
+        if body.username.lower() in RESERVED_USERNAMES:
+            raise HTTPException(403, "Username is reserved")
         ok = await asyncio.to_thread(auth_manager.setup, body.username, body.password)
         if not ok:
             raise HTTPException(500, "Setup failed")
@@ -118,10 +122,12 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             raise HTTPException(400, "Run setup first")
         if not auth_manager.signup_enabled:
             raise HTTPException(403, "Registration is disabled. Ask an admin for an account.")
-        if len(body.password) < 8:
-            raise HTTPException(400, "Password must be at least 8 characters")
+        if len(body.password) < PASSWORD_MIN_LENGTH:
+            raise HTTPException(400, f"Password must be at least {PASSWORD_MIN_LENGTH} characters")
         if len(body.username.strip()) < 1:
             raise HTTPException(400, "Username is required")
+        if body.username.lower() in RESERVED_USERNAMES:
+            raise HTTPException(403, "Username is reserved")
         ok = await asyncio.to_thread(auth_manager.create_user, body.username, body.password, is_admin=False)
         if not ok:
             raise HTTPException(409, "Username already taken")
@@ -144,6 +150,8 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
                 raise HTTPException(401, "Invalid 2FA code")
         # All checks passed — create session (password already verified above)
         token = await asyncio.to_thread(auth_manager.create_session_trusted, username)
+        if not token:
+            raise HTTPException(401, "Invalid credentials")
         cookie_kwargs = dict(
             key=SESSION_COOKIE,
             value=token,
@@ -153,7 +161,7 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             path="/",
         )
         if body.remember:
-            cookie_kwargs["max_age"] = 60 * 60 * 24 * 7  # 7 days
+            cookie_kwargs["max_age"] = TOKEN_TTL
         response.set_cookie(**cookie_kwargs)
         return {"ok": True, "username": username}
 
@@ -182,13 +190,18 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
             pass
         return result
 
+    @router.get("/policy")
+    async def auth_policy():
+        """Return public auth policy constants for the frontend."""
+        return auth_manager.policy()
+
     @router.post("/change-password")
     async def change_password(body: ChangePasswordRequest, request: Request):
         user = _get_current_user(request)
         if not user:
             raise HTTPException(401, "Not authenticated")
-        if len(body.new_password) < 8:
-            raise HTTPException(400, "Password must be at least 8 characters")
+        if len(body.new_password) < PASSWORD_MIN_LENGTH:
+            raise HTTPException(400, f"Password must be at least {PASSWORD_MIN_LENGTH} characters")
         current_token = request.cookies.get(SESSION_COOKIE)
         ok = await asyncio.to_thread(auth_manager.change_password, user, body.current_password, body.new_password)
         if not ok:
@@ -268,8 +281,12 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         user = _get_current_user(request)
         if not user or not auth_manager.is_admin(user):
             raise HTTPException(403, "Admin only")
-        if len(body.password) < 8:
-            raise HTTPException(400, "Password must be at least 8 characters")
+        if len(body.password) < PASSWORD_MIN_LENGTH:
+            raise HTTPException(400, f"Password must be at least {PASSWORD_MIN_LENGTH} characters")
+        if len(body.username.strip()) < 1:
+            raise HTTPException(400, "Username is required")
+        if body.username.lower() in RESERVED_USERNAMES:
+            raise HTTPException(403, "Username is reserved")
         ok = auth_manager.create_user(body.username, body.password, body.is_admin)
         if not ok:
             raise HTTPException(409, "Username already taken")
@@ -432,6 +449,23 @@ def setup_auth_routes(auth_manager: AuthManager) -> APIRouter:
         except Exception as e:
             logger.warning("Failed to rename upload owner references %s -> %s: %s", old_username, new_username, e)
 
+        # direct personal RAG uploads live in per-owner directories and the
+        # vector metadata also carries the username used for owner-filtered
+        # search. Keep both in sync with the auth rename.
+        try:
+            from routes.personal_routes import rename_personal_upload_owner
+            personal_docs_manager = getattr(request.app.state, "personal_docs_manager", None)
+            if personal_docs_manager is not None:
+                rag_manager = getattr(personal_docs_manager, "rag_manager", None)
+                rename_personal_upload_owner(
+                    old_username,
+                    new_username,
+                    personal_docs_manager=personal_docs_manager,
+                    rag_manager=rag_manager,
+                )
+        except Exception as e:
+            logger.warning("Failed to rename personal RAG upload owner references %s -> %s: %s", old_username, new_username, e)
+
         # skills: SKILL.md frontmatter carries owner: <username>; the usage
         # sidecar (_usage.json) keys entries as owner::skill-name. Both must
         # be updated or the renamed user's Skills panel goes empty.

routes/chat_helpers.py

@@ -14,7 +14,7 @@ from core.database import Session as DBSession, ModelEndpoint
 from src.llm_core import normalize_model_id
 from src.endpoint_resolver import normalize_base
 from src.context_compactor import maybe_compact, trim_for_context
-from src.auth_helpers import get_current_user
+from src.auth_helpers import effective_user
 from src.prompt_security import untrusted_context_message
 from routes.prefs_routes import _load_for_user as load_prefs_for_user
 
@@ -22,6 +22,47 @@ from fastapi import HTTPException
 
 logger = logging.getLogger(__name__)
 
+_CASUAL_OPENING_RE = re.compile(
+    r"^\s*(?:h+i+|hey+|hello+|yo+|sup+|what'?s up|wass?up|hiya|howdy|"
+    r"lol|lmao|haha+|hehe+|thanks?|thank you|ty|idk|dunno|meh|bruh|bro)\b(?P<tail>.*)$",
+    re.IGNORECASE,
+)
+_CASUAL_BLOCKLIST_RE = re.compile(
+    r"\b(?:cookbook|serve|serving|launch|start|vllm|sglang|llama\.?cpp|ollama|"
+    r"download|model|email|document|doc|note|calendar|task|search|web|research|"
+    r"file|folder|repo|git|settings?|endpoint|api|token|mcp)\b",
+    re.IGNORECASE,
+)
+
+
+def _is_casual_low_signal(text: str) -> bool:
+    """Short greetings/slang should not pull memory, skills, RAG, or docs."""
+    s = str(text or "").strip()
+    m = _CASUAL_OPENING_RE.match(s)
+    if not m:
+        return False
+    tail = m.group("tail") or ""
+    if _CASUAL_BLOCKLIST_RE.search(tail):
+        return False
+    tail_words = re.findall(r"[A-Za-z0-9_'-]+", tail)
+    return len(tail_words) <= 2
+
+
+# Strong references to in-flight fire-and-forget tasks scheduled from this
+# module. asyncio only keeps weak references to tasks created via
+# create_task, so without this the GC can collect a task mid-execution and
+# the background work (extraction, auto-naming) silently never runs.
+# Mirrors WebhookManager._spawn_tracked from src/webhook_manager.py.
+_BG_TASKS: set[asyncio.Task] = set()
+
+
+def _spawn_bg(coro) -> asyncio.Task:
+    """Schedule a background task and hold a strong reference until it finishes."""
+    task = asyncio.create_task(coro)
+    _BG_TASKS.add(task)
+    task.add_done_callback(_BG_TASKS.discard)
+    return task
+
 
 # ── Data containers ────────────────────────────────────────────────────── #
 
@@ -78,7 +119,7 @@ def _enforce_chat_privileges(request, sess) -> None:
     which means unrestricted allowed_models / zero cap -> no-op for them.
     """
     try:
-        user = get_current_user(request)
+        user = effective_user(request)
     except Exception:
         user = None
     if not user:
@@ -159,17 +200,9 @@ async def auto_name_session(session_manager, sess):
             return
 
         owner = getattr(sess, "owner", None)
-        t_url, t_model, t_headers = resolve_task_endpoint(owner=owner)
-        if not t_model:
-            # If no task/utility model is configured at all, fall back to
-            # the session's own model so auto-naming still works even on
-            # minimal setups.
-            from src.endpoint_resolver import resolve_endpoint
-            _fallback = resolve_endpoint("default", owner=owner)
-            if _fallback and _fallback[1]:
-                t_url, t_model, t_headers = _fallback
-            else:
-                t_url, t_model, t_headers = sess.endpoint_url, sess.model, sess.headers
+        t_url, t_model, t_headers = resolve_task_endpoint(
+            sess.endpoint_url, sess.model, sess.headers, owner=owner
+        )
         if not t_model:
             logger.debug("[auto-name] No model provided, skipping")
             return
@@ -346,11 +379,11 @@ def add_user_message(sess, chat_handler, preprocessed: PreprocessedMessage, inco
 def fire_message_event(request, webhook_manager, session_id: str, sess, message: str, compare_mode: bool = False):
     """Fire webhook and event_bus events for a new user message."""
     if webhook_manager and not compare_mode:
-        asyncio.create_task(webhook_manager.fire("chat.message", {
+        webhook_manager.fire_and_forget("chat.message", {
             "session_id": session_id, "model": sess.model, "message": message[:2000],
-        }))
+        })
     from src.event_bus import fire_event
-    user = get_current_user(request)
+    user = effective_user(request)
     fire_event("message_sent", user)
 
 
@@ -576,9 +609,11 @@ async def build_chat_context(
     if not incognito:
         fire_message_event(request, webhook_manager, session_id, sess, message, compare_mode)
 
-    # Resolve user prefs
-    user = get_current_user(request)
+    # Resolve owner-scoped prefs/context. Browser requests keep the cookie user;
+    # bearer-token chat requests use the token owner instead of the "api" sentinel.
+    user = effective_user(request)
     uprefs = load_prefs_for_user(user)
+    casual_low_signal = _is_casual_low_signal(message)
 
     # Memory enabled?
     mem_enabled = not incognito and not no_memory and uprefs.get("memory_enabled", True)
@@ -588,6 +623,9 @@ async def build_chat_context(
     if not allow_tool_preprocessing:
         mem_enabled = False
         skills_enabled = False
+    if casual_low_signal:
+        mem_enabled = False
+        skills_enabled = False
     logger.debug(
         "Memory enabled=%s for user=%s (incognito=%s, no_memory=%s, pref=%s)",
         mem_enabled, user, incognito, no_memory, uprefs.get("memory_enabled", "NOT_SET"),
@@ -603,11 +641,11 @@ async def build_chat_context(
 
     # Use RAG?
     use_rag_val = (str(use_rag).lower() != "false") if use_rag is not None else True
-    if incognito or not allow_tool_preprocessing or is_research_spinoff:
+    if incognito or not allow_tool_preprocessing or is_research_spinoff or casual_low_signal:
         use_rag_val = False
 
     # If pre-fetched search context was provided (compare mode), skip live web search
-    skip_web = bool(search_context) or not allow_tool_preprocessing
+    skip_web = bool(search_context) or not allow_tool_preprocessing or casual_low_signal
 
     # Build context preface
     # The stream path uses enhanced_message (with CoT/preprocessing applied),
@@ -626,7 +664,7 @@ async def build_chat_context(
         incognito=incognito,
         use_skills=skills_enabled,
     )
-    if use_rag is not None or is_research_spinoff:
+    if use_rag is not None or is_research_spinoff or casual_low_signal:
         _preface_kwargs["use_rag"] = use_rag_val
     preface, rag_sources, web_sources = chat_processor.build_context_preface(**_preface_kwargs)
 
@@ -634,7 +672,7 @@ async def build_chat_context(
     used_memories = getattr(chat_processor, '_last_used_memories', [])
 
     # Inject pre-fetched search context (compare mode)
-    if search_context and allow_tool_preprocessing:
+    if search_context and allow_tool_preprocessing and not casual_low_signal:
         preface.append(untrusted_context_message("prefetched search context", search_context))
 
     # YouTube transcripts
@@ -1112,7 +1150,7 @@ def run_post_response_tasks(
             )))
 
     if _extraction_jobs:
-        asyncio.create_task(_run_extraction_jobs_sequentially(session_id, _extraction_jobs))
+        _spawn_bg(_run_extraction_jobs_sequentially(session_id, _extraction_jobs))
 
     # Token accumulation
     if last_metrics:
@@ -1120,11 +1158,11 @@ def run_post_response_tasks(
 
     # Webhook
     if webhook_manager and not compare_mode:
-        asyncio.create_task(webhook_manager.fire("chat.completed", {
+        webhook_manager.fire_and_forget("chat.completed", {
             "session_id": session_id, "model": sess.model,
             "user_message": message, "response": full_response[:2000],
-        }))
+        })
 
     # Auto-name
     if needs_auto_name(sess.name):
-        asyncio.create_task(auto_name_session(session_manager, sess))
+        _spawn_bg(auto_name_session(session_manager, sess))

routes/chat_routes.py

@@ -23,7 +23,7 @@ from src.endpoint_resolver import normalize_base as _normalize_base, build_chat_
 from src.session_search import search_session_messages
 from src.prompt_security import untrusted_context_message
 from core.exceptions import SessionNotFoundError
-from src.auth_helpers import get_current_user
+from src.auth_helpers import effective_user, get_current_user
 from routes.session_routes import _verify_session_owner
 from routes.document_helpers import _owner_session_filter
 from core.database import SessionLocal, get_session_mode, set_session_mode
@@ -126,7 +126,8 @@ def _clear_orphaned_session_endpoint(sess, owner: str | None = None) -> bool:
         sess.model = ""
         sess.headers = {}
         return True
-    except Exception:
+    except Exception as e:
+        logger.warning("Failed to clear orphaned session endpoint", exc_info=e)
         db.rollback()
         return False
     finally:
@@ -144,7 +145,8 @@ def _endpoint_cache_contains_model(endpoint, model: str) -> bool:
         return True
     try:
         models = json.loads(raw) if isinstance(raw, str) else raw
-    except Exception:
+    except Exception as e:
+        logger.warning("Failed to parse cached models list, treating as containing model", exc_info=e)
         return True
     if not isinstance(models, list) or not models:
         return True
@@ -236,7 +238,8 @@ def _recover_empty_session_model(sess, session_id: str, owner: str | None = None
                 is_chatgpt_subscription = False
         try:
             cached = json.loads(ep.cached_models) if isinstance(ep.cached_models, str) else (ep.cached_models or [])
-        except Exception:
+        except Exception as e:
+            logger.warning("Failed to parse cached_models for endpoint %r", getattr(ep, "id", "?"), exc_info=e)
             cached = []
         if not cached:
             visible = []
@@ -360,7 +363,7 @@ def setup_chat_routes(
             sess = session_manager.get_session(session)
         except KeyError:
             raise HTTPException(404, f"Session '{session}' not found")
-        owner = get_current_user(request)
+        owner = effective_user(request)
         if _clear_orphaned_session_endpoint(sess, owner=owner):
             raise HTTPException(400, "Selected model endpoint was removed. Pick another model in Settings.")
 
@@ -600,7 +603,7 @@ def setup_chat_routes(
             # but BEFORE loading. Prevents cross-user session hijack.
             _verify_session_owner(request, session)
             sess = session_manager.get_session(session)
-            owner = get_current_user(request)
+            owner = effective_user(request)
             if _clear_orphaned_session_endpoint(sess, owner=owner):
                 raise HTTPException(400, "Selected model endpoint was removed. Pick another model in Settings.")
             # Issue #587: picker shows a model from the endpoint cache but
@@ -631,7 +634,7 @@ def setup_chat_routes(
         _enforce_chat_privileges(request, sess)
 
         # Ensure session has auth headers
-        resolve_session_auth(sess, session, owner=get_current_user(request))
+        resolve_session_auth(sess, session, owner=effective_user(request))
 
         # Check for research_pending BEFORE mode persist overwrites it
         do_research = str(use_research).lower() == "true"
@@ -646,8 +649,8 @@ def setup_chat_routes(
         elif attachments:
             try:
                 att_ids = [str(x) for x in json.loads(attachments)]
-            except Exception:
-                pass
+            except Exception as e:
+                logger.warning("Failed to parse attachments JSON, ignoring attachments", exc_info=e)
 
         no_memory = str(form_data.get("no_memory", "")).lower() == "true"
         pre_context_tool_policy = build_effective_tool_policy(
@@ -826,7 +829,11 @@ def setup_chat_routes(
         from src.settings import get_setting
         _global_disabled = get_setting("disabled_tools", [])
         if _global_disabled and isinstance(_global_disabled, list):
-            disabled_tools.update(_global_disabled)
+            explicit_web_allowed = allow_web_search is not None and str(allow_web_search).lower() == "true"
+            if explicit_web_allowed:
+                disabled_tools.update(t for t in _global_disabled if t not in {"web_search", "web_fetch"})
+            else:
+                disabled_tools.update(_global_disabled)
 
         # Light auto-escalation: the user is in chat mode and just expressed a
         # notes/calendar/email intent. Grant the relevant managers but withhold
@@ -1256,6 +1263,10 @@ def setup_chat_routes(
                         _max_rounds = _DEFAULT_ROUNDS
                     _max_rounds = max(1, min(_max_rounds, 200))
 
+                    _forced_tools = None
+                    if allow_web_search is not None and str(allow_web_search).lower() == "true":
+                        _forced_tools = {"web_search", "web_fetch"}
+
                     async for chunk in stream_agent_loop(
                         sess.endpoint_url,
                         sess.model,
@@ -1277,6 +1288,7 @@ def setup_chat_routes(
                         plan_mode=plan_mode,
                         approved_plan=approved_plan or None,
                         workspace=workspace or None,
+                        forced_tools=_forced_tools,
                     ):
                         if chunk.startswith("data: ") and not chunk.startswith("data: [DONE]"):
                             try:
@@ -1297,8 +1309,6 @@ def setup_chat_routes(
                                     "doc_stream_open", "doc_stream_delta",
                                     "doc_update", "doc_suggestions", "ui_control",
                                     "rounds_exhausted",
-                                    "loop_breaker_triggered",
-                                    "intent_nudge_exhausted",
                                     "ask_user",
                                     "plan_update",
                                 ):
@@ -1484,7 +1494,7 @@ def setup_chat_routes(
         if not q or not q.strip():
             return []
 
-        _user = get_current_user(request)
+        _user = effective_user(request)
         return [
             result.to_dict()
             for result in search_session_messages(

routes/codex_routes.py

@@ -46,8 +46,12 @@ def _ssh_prefix_for_task(task: dict) -> tuple[str, str]:
     shell metacharacters in ``remoteHost`` is rejected with 400 rather than
     injected.
     """
-    host = validate_remote_host((task.get("remoteHost") or "").strip() or None) or ""
-    ssh_port = validate_ssh_port((task.get("sshPort") or "").strip() or None) or ""
+    raw_host = task.get("remoteHost")
+    raw_port = task.get("sshPort")
+    host_value = str(raw_host).strip() if raw_host is not None else None
+    port_value = str(raw_port).strip() if raw_port is not None else None
+    host = validate_remote_host(host_value or None) or ""
+    ssh_port = validate_ssh_port(port_value or None) or ""
     port_flag = f"-p {ssh_port} " if ssh_port and ssh_port != "22" else ""
     return host, port_flag
 
@@ -306,7 +310,10 @@ def setup_codex_routes(
 
     @router.post("/emails/draft-document")
     async def codex_email_draft_document(request: Request, body: dict[str, Any] = Body(default_factory=dict)):
-        owner = _scope_owner_all(request, {"email:draft", "documents:write"})
+        owner = _scope_owner(request, EMAIL_DRAFT_SCOPES)
+        docs_owner = _scope_owner_all(request, DOCS_WRITE_SCOPES)
+        if docs_owner != owner:
+            raise HTTPException(403, "API token owner mismatch")
         if documents_create_endpoint is None:
             raise HTTPException(503, "Documents integration is not available")
         from routes.document_routes import DocumentCreate

routes/cookbook_helpers.py

@@ -505,6 +505,8 @@ def _cached_model_scan_script(model_dirs: list[str] | None = None, add_hf_cache:
         "    if u.startswith('KB'): return int(n * 1024)",
         "    return int(n)",
         "def scan_ollama():",
+        "    if any(m.get('is_ollama') for m in models): return",
+        "    if os.name == 'nt' and not os.environ.get('ODYSSEUS_ALLOW_OLLAMA_CLI_SCAN'): return",
         "    if not shutil.which('ollama'): return",
         "    try:",
         "        p = subprocess.run(['ollama', 'list'], stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True, timeout=6)",
@@ -535,8 +537,8 @@ def _cached_model_scan_script(model_dirs: list[str] | None = None, add_hf_cache:
         "            models.append({'repo_id':name,'size_bytes':size_bytes,'nb_files':1,'has_incomplete':False,'path':'ollama','backend':'ollama','is_ollama':True})",
         "        return",
         "for _hf_cache in hf_cache_paths(): scan_hf(_hf_cache)",
-        "scan_ollama()",
         "scan_ollama_api()",
+        "scan_ollama()",
     ]
     for model_dir in model_dirs or []:
         lines.append(f"scan_dir(os.path.expanduser({model_dir!r}))")
@@ -784,25 +786,149 @@ def _append_llama_cpp_linux_accel_build_lines(runner_lines: list[str]) -> None:
     to hard-wire CUDA on Linux. That made ROCm hosts attempt a CUDA configure and
     fail with "CUDA Toolkit not found" instead of building with HIP.
     """
+    # Try a prebuilt binary from llama.cpp's GitHub releases FIRST — no
+    # cmake/build-essential/git/CUDA-headers needed at all. The from-source
+    # build below stays as a fallback (custom flags, esoteric arch, no
+    # internet, etc). 30 seconds vs 5+ minutes of compile, and removes
+    # every OS-package dep from the launch path. Sets _odysseus_have_prebuilt=1
+    # on success; the existing build-tier if/elif chain below is gated on
+    # that variable so we never compile twice or shadow the prebuilt symlink.
+    runner_lines.append('    _odysseus_have_prebuilt=""')
+    runner_lines.append('    _odysseus_arch="$(uname -m)"')
+    runner_lines.append('    _odysseus_prebuilt_url=""')
+    runner_lines.append('    if command -v curl >/dev/null 2>&1 && [ "$_odysseus_arch" = "x86_64" ]; then')
+    runner_lines.append('      _odysseus_pat=""')
+    runner_lines.append('      _odysseus_has_nv_inline() { command -v nvidia-smi >/dev/null 2>&1 && nvidia-smi -L 2>/dev/null | grep -q "GPU "; }')
+    runner_lines.append('      _odysseus_has_vk_inline() { ldconfig -p 2>/dev/null | grep -q "libvulkan\\.so" || command -v vulkaninfo >/dev/null 2>&1 || [ -e /usr/lib/x86_64-linux-gnu/libvulkan.so.1 ]; }')
+    runner_lines.append('      _odysseus_has_vkdev_inline() { ls /dev/dri/renderD* >/dev/null 2>&1 || (lspci 2>/dev/null | grep -Ei \'VGA|3D|Display\' | grep -Eiq \'AMD|ATI|Radeon\'); }')
+    runner_lines.append('      if _odysseus_has_nv_inline; then')
+    runner_lines.append('        _odysseus_pat="ubuntu.*cuda"')
+    runner_lines.append('      elif _odysseus_has_vkdev_inline && _odysseus_has_vk_inline; then')
+    runner_lines.append('        _odysseus_pat="ubuntu.*vulkan"')
+    runner_lines.append('      else')
+    runner_lines.append('        _odysseus_pat="ubuntu-x64\\\\.zip"')
+    runner_lines.append('      fi')
+    runner_lines.append('      _odysseus_prebuilt_url="$(curl -fsSL --max-time 15 https://api.github.com/repos/ggml-org/llama.cpp/releases/latest 2>/dev/null | grep \'"browser_download_url"\' | cut -d\'"\' -f4 | grep -iE "$_odysseus_pat" | grep -iv "arm\\|aarch64" | head -1)"')
+    runner_lines.append('    fi')
+    # Accept any of unzip / bsdtar / python3 -m zipfile as the extractor.
+    # python3 is essentially always present on modern Linux, so this lets
+    # the prebuilt path work on minimal Ubuntu installs that lack `unzip`.
+    runner_lines.append('    if [ -n "$_odysseus_prebuilt_url" ] && (command -v unzip >/dev/null 2>&1 || command -v bsdtar >/dev/null 2>&1 || command -v python3 >/dev/null 2>&1); then')
+    runner_lines.append('      echo "[odysseus] Found prebuilt llama-server: $_odysseus_prebuilt_url"')
+    runner_lines.append('      mkdir -p ~/bin "$HOME/.cache/odysseus/llama-cpp-prebuilt" && cd "$HOME/.cache/odysseus/llama-cpp-prebuilt"')
+    runner_lines.append('      rm -f llama-cpp.zip')
+    runner_lines.append('      if curl -fsSL --max-time 120 "$_odysseus_prebuilt_url" -o llama-cpp.zip && [ -s llama-cpp.zip ]; then')
+    runner_lines.append('        rm -rf build && mkdir -p build')
+    runner_lines.append('        if command -v unzip >/dev/null 2>&1; then unzip -qq -o llama-cpp.zip -d build; elif command -v bsdtar >/dev/null 2>&1; then bsdtar -xf llama-cpp.zip -C build; else python3 -c "import zipfile; zipfile.ZipFile(\\"llama-cpp.zip\\").extractall(\\"build\\")"; fi')
+    runner_lines.append('        _odysseus_extracted="$(find build -type f -name llama-server 2>/dev/null | head -1)"')
+    runner_lines.append('        if [ -n "$_odysseus_extracted" ]; then')
+    runner_lines.append('          chmod +x "$_odysseus_extracted"')
+    runner_lines.append('          ln -sf "$_odysseus_extracted" ~/bin/llama-server')
+    runner_lines.append('          _odysseus_libdir="$(dirname "$_odysseus_extracted")"')
+    runner_lines.append('          mkdir -p ~/.config && echo "export LD_LIBRARY_PATH=\\"$_odysseus_libdir:\\${LD_LIBRARY_PATH:-}\\"" > ~/.config/odysseus-llama-cpp-env')
+    runner_lines.append('          _odysseus_have_prebuilt=1')
+    runner_lines.append('          echo "[odysseus] Prebuilt llama-server installed at $_odysseus_extracted"')
+    runner_lines.append('        fi')
+    runner_lines.append('      fi')
+    runner_lines.append('      [ -z "$_odysseus_have_prebuilt" ] && echo "[odysseus] Prebuilt download/extract failed — falling back to from-source build."')
+    runner_lines.append('    elif [ -z "$_odysseus_prebuilt_url" ]; then')
+    runner_lines.append('      echo "[odysseus] No matching prebuilt llama-server for this host (arch=$_odysseus_arch) — will build from source."')
+    runner_lines.append('    fi')
+    runner_lines.append('  if [ -z "$_odysseus_have_prebuilt" ]; then')
     # Detect pip-installed nvcc (from vLLM/nvidia CUDA wheels) and put it on PATH
-    # so cmake's CUDA configure can find it. We keep this after the ROCm/HIP
-    # check — a machine with both stacks should honor the native HIP toolchain on
-    # AMD hosts instead of accidentally preferring a stray nvcc wheel.
-    runner_lines.append('    for _cudir in ~/.local/lib/python*/site-packages/nvidia/cu13 ~/.local/lib/python*/site-packages/nvidia/cu12 ~/.local/lib/python*/site-packages/nvidia/cuda_nvcc; do')
-    runner_lines.append('      [ -x "$_cudir/bin/nvcc" ] && export CUDA_HOME="$_cudir" && export PATH="$_cudir/bin:$PATH" && break')
-    runner_lines.append('    done')
+    # so cmake's CUDA configure can find it — BUT only when actual NVIDIA
+    # hardware is present. On AMD/Intel hosts the pip nvcc is a misleading
+    # leftover (no libcudart, no GPU it could target) and would otherwise
+    # send the build down the CUDA branch and fail with "CUDA Toolkit not
+    # found" instead of trying Vulkan.
+    runner_lines.append('    _odysseus_has_nvidia_hw() {')
+    runner_lines.append('      command -v nvidia-smi >/dev/null 2>&1 && nvidia-smi -L 2>/dev/null | grep -q "GPU " && return 0')
+    runner_lines.append('      ls /dev/nvidia* >/dev/null 2>&1 && return 0')
+    runner_lines.append('      lspci 2>/dev/null | grep -iE \'VGA|3D|Display\' | grep -iq nvidia && return 0')
+    runner_lines.append('      return 1')
+    runner_lines.append('    }')
+    runner_lines.append('    if _odysseus_has_nvidia_hw; then')
+    runner_lines.append('      for _cudir in ~/.local/lib/python*/site-packages/nvidia/cu13 ~/.local/lib/python*/site-packages/nvidia/cu12 ~/.local/lib/python*/site-packages/nvidia/cuda_nvcc; do')
+    runner_lines.append('        [ -x "$_cudir/bin/nvcc" ] && export CUDA_HOME="$_cudir" && export PATH="$_cudir/bin:$PATH" && break')
+    runner_lines.append('      done')
+    runner_lines.append('    fi')
     # rm -rf build so a prior poisoned CMakeCache.txt (e.g. from a failed CUDA
     # or HIP attempt) doesn't cause the next configure to reuse stale settings.
     runner_lines.append('    mkdir -p ~/bin')
-    runner_lines.append('    cd ~/llama.cpp && rm -rf build')
+    # Try to install cmake / build-essential / git automatically before the
+    # build, but ONLY via passwordless sudo (`sudo -n`) — interactive sudo
+    # would hang a tmux-backgrounded serve task waiting for a password. If
+    # sudo asks for a password the install is skipped silently and the
+    # diagnosis pattern (cookbook_routes.py / cookbook_helpers.py) surfaces
+    # an explicit "install cmake" suggestion in the Cookbook diagnosis
+    # toolbar after the inevitable build failure.
+    runner_lines.append('    _odysseus_apt_bootstrap() {')
+    runner_lines.append('      local _missing=""')
+    runner_lines.append('      command -v cmake >/dev/null 2>&1 || _missing="$_missing cmake"')
+    runner_lines.append('      command -v g++ >/dev/null 2>&1 || command -v gcc >/dev/null 2>&1 || _missing="$_missing build-essential"')
+    runner_lines.append('      command -v git >/dev/null 2>&1 || _missing="$_missing git"')
+    runner_lines.append('      [ -z "$_missing" ] && return 0')
+    runner_lines.append('      if command -v apt-get >/dev/null 2>&1 && sudo -n true 2>/dev/null; then')
+    runner_lines.append('        echo "[odysseus] Auto-installing missing build deps via apt:$_missing"')
+    runner_lines.append('        sudo -n env DEBIAN_FRONTEND=noninteractive apt-get update -qq 2>&1 | tail -3')
+    runner_lines.append('        sudo -n env DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends $_missing 2>&1 | tail -5 || true')
+    runner_lines.append('      elif command -v pacman >/dev/null 2>&1 && sudo -n true 2>/dev/null; then')
+    runner_lines.append('        echo "[odysseus] Auto-installing missing build deps via pacman:$_missing"')
+    runner_lines.append('        local _pacpkgs="$(echo "$_missing" | sed -e \'s/build-essential/base-devel/g\')"')
+    runner_lines.append('        sudo -n pacman -Sy --needed --noconfirm $_pacpkgs 2>&1 | tail -5 || true')
+    runner_lines.append('      elif command -v dnf >/dev/null 2>&1 && sudo -n true 2>/dev/null; then')
+    runner_lines.append('        echo "[odysseus] Auto-installing missing build deps via dnf:$_missing"')
+    runner_lines.append('        local _dnfpkgs="$(echo "$_missing" | sed -e \'s/build-essential/gcc gcc-c++ make/g\')"')
+    runner_lines.append('        sudo -n dnf install -y $_dnfpkgs 2>&1 | tail -5 || true')
+    runner_lines.append('      else')
+    runner_lines.append('        echo "[odysseus] WARNING: missing build deps ($_missing) — passwordless sudo is unavailable, cannot auto-install. Cookbook Diagnosis will explain the fix after the build fails."')
+    runner_lines.append('      fi')
+    runner_lines.append('    }')
+    runner_lines.append('    _odysseus_apt_bootstrap')
+    runner_lines.append('    _odysseus_missing_build_deps=""')
+    runner_lines.append('    command -v cmake >/dev/null 2>&1 || _odysseus_missing_build_deps="$_odysseus_missing_build_deps cmake"')
+    runner_lines.append('    command -v git >/dev/null 2>&1 || _odysseus_missing_build_deps="$_odysseus_missing_build_deps git"')
+    runner_lines.append('    command -v g++ >/dev/null 2>&1 || command -v gcc >/dev/null 2>&1 || _odysseus_missing_build_deps="$_odysseus_missing_build_deps build-essential"')
+    runner_lines.append('    if [ -n "$_odysseus_missing_build_deps" ]; then')
+    runner_lines.append('      echo "ERROR: llama.cpp source build needs missing packages:$_odysseus_missing_build_deps"')
+    runner_lines.append('      if command -v apt-get >/dev/null 2>&1; then')
+    runner_lines.append('        echo "Install on this host: sudo apt-get update && sudo apt-get install -y cmake build-essential git"')
+    runner_lines.append('      elif command -v pacman >/dev/null 2>&1; then')
+    runner_lines.append('        echo "Install on this host: sudo pacman -Sy --needed cmake base-devel git"')
+    runner_lines.append('      elif command -v dnf >/dev/null 2>&1; then')
+    runner_lines.append('        echo "Install on this host: sudo dnf install -y cmake gcc gcc-c++ make git"')
+    runner_lines.append('      fi')
+    runner_lines.append('      echo "Alternative: install a native llama-server on PATH, then relaunch."')
+    runner_lines.append('      ODYSSEUS_PREFLIGHT_EXIT=127')
+    runner_lines.append('    fi')
+    runner_lines.append('    cd ~/llama.cpp')
+    runner_lines.append('    _odysseus_has_vulkan() {')
+    runner_lines.append('      ldconfig -p 2>/dev/null | grep -q \'libvulkan\\.so\' && return 0')
+    runner_lines.append('      [ -e /usr/lib/libvulkan.so.1 ] && return 0')
+    runner_lines.append('      [ -e /usr/lib/x86_64-linux-gnu/libvulkan.so.1 ] && return 0')
+    runner_lines.append('      command -v vulkaninfo >/dev/null 2>&1 && return 0')
+    runner_lines.append('      return 1')
+    runner_lines.append('    }')
+    runner_lines.append('    _odysseus_has_vulkan_device() {')
+    runner_lines.append('      ls /dev/dri/renderD* >/dev/null 2>&1 && return 0')
+    runner_lines.append('      lspci 2>/dev/null | grep -Ei \'VGA|3D|Display\' | grep -Eiq \'AMD|ATI|Radeon\' && return 0')
+    runner_lines.append('      return 1')
+    runner_lines.append('    }')
+    # Backend preference: native ROCm/HIP > native CUDA > Vulkan > CPU.
+    # Vulkan is a portable fallback that works on AMD when ROCm isn't
+    # installed (e.g. Strix Halo) and on any vendor's discrete GPU, but
+    # it's ~30-40% slower than native HIP/CUDA for LLM inference — only
+    # pick it when no native toolchain is present.
     runner_lines.append('    if command -v hipconfig &>/dev/null || [ -d /opt/rocm ] || [ -n "$ROCM_PATH" ] || [ -n "$HIP_PATH" ]; then')
+    runner_lines.append('      rm -rf build')
     runner_lines.append('      if command -v hipconfig &>/dev/null; then')
     runner_lines.append('        export HIPCXX="${HIPCXX:-$(hipconfig -l)/clang}"')
     runner_lines.append('        export HIP_PATH="${HIP_PATH:-$(hipconfig -R)}"')
     runner_lines.append('      fi')
     runner_lines.append('      echo "[odysseus] ROCm/HIP detected — building llama-server with HIP support..."')
     runner_lines.append('      cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_HIP=ON && cmake --build build -j"$NPROC" --target llama-server && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
-    runner_lines.append('    elif command -v nvcc &>/dev/null; then')
+    runner_lines.append('    elif command -v nvcc &>/dev/null && _odysseus_has_nvidia_hw; then')
+    runner_lines.append('      rm -rf build')
     # nvcc alone is not sufficient — pip-installed CUDA wheels or incomplete
     # tooling can expose nvcc without shipping libcudart, causing cmake to fail
     # mid-build with "CUDA runtime library not found". Check cudart explicitly
@@ -826,31 +952,50 @@ def _append_llama_cpp_linux_accel_build_lines(runner_lines: list[str]) -> None:
     runner_lines.append('        echo "[odysseus]   Ensure libcudart is installed (e.g. cuda-runtime package) and visible via ldconfig or CUDA_HOME."')
     runner_lines.append('        cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j"$NPROC" --target llama-server && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
     runner_lines.append('      fi')
+    runner_lines.append('    elif _odysseus_has_vulkan_device && _odysseus_has_vulkan; then')
+    runner_lines.append('      echo "[odysseus] Vulkan-capable GPU detected (no ROCm/CUDA toolchain installed) — building llama-server with Vulkan support..."')
+    runner_lines.append('      rm -rf build-vulkan')
+    runner_lines.append('      cmake -B build-vulkan -DCMAKE_BUILD_TYPE=Release -DGGML_VULKAN=ON && cmake --build build-vulkan -j"$NPROC" --target llama-server && ln -sf ~/llama.cpp/build-vulkan/bin/llama-server ~/bin/llama-server')
     runner_lines.append('    else')
-    runner_lines.append('      echo "[odysseus] WARNING: no HIP/CUDA toolchain found — building llama-server for CPU only."')
+    runner_lines.append('      echo "[odysseus] WARNING: no HIP/CUDA/Vulkan toolchain found — building llama-server for CPU only."')
     runner_lines.append('      echo "[odysseus]   GPU inference will not be available for this llama.cpp build."')
-    runner_lines.append('      echo "[odysseus]   Install ROCm for AMD GPUs or vLLM/CUDA tooling for NVIDIA, then re-launch this serve task."')
+    runner_lines.append('      echo "[odysseus]   Install Vulkan (libvulkan-dev) / ROCm for AMD GPUs or CUDA tooling for NVIDIA, then re-launch this serve task."')
+    runner_lines.append('      rm -rf build')
     runner_lines.append('      cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j"$NPROC" --target llama-server && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
     runner_lines.append('    fi')
+    runner_lines.append('  fi  # end _odysseus_have_prebuilt guard')
 
 
-def _llama_cpp_rebuild_cmd() -> str:
+def _llama_cpp_rebuild_cmd(update_source: bool = False) -> str:
     """Shell command that clears the Cookbook-managed llama.cpp build.
 
-    Removes the cached ``llama-server`` symlink and the ``~/llama.cpp/build``
+    Removes the cached ``llama-server`` symlink and the ``~/llama.cpp/build*``
     directory so the next llama.cpp serve recompiles from source, picking up a
     CUDA or HIP toolchain if one is now available. The serve bootstrap only
     builds when ``llama-server`` is missing from PATH, so without this an
-    existing CPU-only build is reused forever. It deliberately installs and
-    downloads nothing; the rebuild itself happens on the next serve.
+    existing CPU-only build is reused forever. When ``update_source`` is true,
+    the command also fast-forwards the Cookbook-managed ``~/llama.cpp`` checkout
+    if it exists. The rebuild itself happens on the next serve.
     """
+    update_cmd = ''
+    if update_source:
+        update_cmd = (
+            'if [ -d "$HOME/llama.cpp/.git" ]; then '
+            'git -C "$HOME/llama.cpp" pull --ff-only --depth 1 || '
+            'echo "[odysseus] WARNING: llama.cpp source update failed; clearing cached build anyway."; '
+            'elif command -v git >/dev/null 2>&1; then '
+            'git clone --depth 1 https://github.com/ggml-org/llama.cpp "$HOME/llama.cpp" || '
+            'echo "[odysseus] WARNING: llama.cpp clone failed; clearing cached build anyway."; '
+            'fi && '
+        )
     return (
         'mkdir -p "$HOME/bin" && '
+        f'{update_cmd}'
         'rm -f "$HOME/bin/llama-server" && '
-        'rm -rf "$HOME/llama.cpp/build" && '
+        'rm -rf "$HOME/llama.cpp/build" "$HOME/llama.cpp/build-vulkan" && '
         'echo "[odysseus] Cleared the cached llama.cpp build. '
         'Re-launch the serve task to rebuild llama-server from source '
-        '(CUDA or HIP will be used if a toolchain is now available)."'
+        '(Vulkan, HIP, or CUDA will be used if a matching toolchain is now available)."'
     )
 
 
@@ -1113,8 +1258,27 @@ def _diagnose_serve_output(text: str) -> dict | None:
             "SGLang is not installed or not in PATH on this server.",
             [{"label": "install SGLang in Cookbook Dependencies", "op": "dependency", "package": "sglang[all]"}],
         ),
+        # System build deps come BEFORE the generic llama.cpp catch-all so
+        # cmake / build-essential / git missing → a specific OS-package
+        # remediation instead of "install llama-cpp-python[server]" (which
+        # itself fails to compile when cmake is absent).
         (
-            r"llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'|git: command not found|cmake: command not found",
+            r"cmake: command not found|cmake.*not found.*[Cc]ould not",
+            "cmake is required to build llama.cpp from source but isn't installed on this server.",
+            [{"label": "install build deps for llama.cpp (apt: cmake build-essential git / pacman: cmake base-devel git / dnf: cmake gcc-c++ make git / brew: cmake git)", "op": "dependency", "package": "llama-cpp-python[server]"}],
+        ),
+        (
+            r"^(make|g\+\+|gcc): command not found|Could not find C\+\+ compiler",
+            "A C/C++ compiler (build-essential) is required to build llama.cpp from source.",
+            [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
+        ),
+        (
+            r"^git: command not found",
+            "git is required to clone the llama.cpp source tree.",
+            [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
+        ),
+        (
+            r"llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'",
             "llama.cpp / llama-cpp-python dependencies are missing.",
             [{"label": "install llama.cpp dependencies or llama-cpp-python[server]", "op": "dependency", "package": "llama-cpp-python[server]"}],
         ),

routes/cookbook_routes.py

@@ -189,8 +189,27 @@ def setup_cookbook_routes() -> APIRouter:
                 "SGLang is not installed or not in PATH on this server.",
                 [{"label": "install SGLang in Cookbook Dependencies", "op": "dependency", "package": "sglang[all]"}],
             ),
+            # System build deps come BEFORE the generic llama.cpp catch-all
+            # so cmake / build-essential / git missing → a specific OS-package
+            # remediation instead of "install llama-cpp-python[server]" (which
+            # itself fails to compile when cmake is absent).
             (
-                r"llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'|git: command not found|cmake: command not found",
+                r"cmake: command not found|cmake.*not found.*[Cc]ould not",
+                "cmake is required to build llama.cpp from source but isn't installed on this server.",
+                [{"label": "install build deps for llama.cpp (apt: cmake build-essential git / pacman: cmake base-devel git / dnf: cmake gcc-c++ make git / brew: cmake git)", "op": "dependency", "package": "llama-cpp-python[server]"}],
+            ),
+            (
+                r"^(make|g\+\+|gcc): command not found|Could not find C\+\+ compiler",
+                "A C/C++ compiler (build-essential) is required to build llama.cpp from source.",
+                [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
+            ),
+            (
+                r"^git: command not found",
+                "git is required to clone the llama.cpp source tree.",
+                [{"label": "install build deps for llama.cpp on this server", "op": "dependency", "package": "llama-cpp-python[server]"}],
+            ),
+            (
+                r"llama-server.*command not found|llama\.cpp.*not found|No module named.*llama_cpp|No module named 'starlette_context'",
                 "llama.cpp / llama-cpp-python dependencies are missing.",
                 [{"label": "install llama.cpp dependencies or llama-cpp-python[server]", "op": "dependency", "package": "llama-cpp-python[server]"}],
             ),
@@ -254,6 +273,79 @@ def setup_cookbook_routes() -> APIRouter:
     def _load_stored_hf_token() -> str:
         return load_stored_hf_token(state_path=_cookbook_state_path)
 
+    def _normalize_minimax_m3_vllm_cmd(cmd: str) -> str:
+        """Patch MiniMax M3 vLLM launches into the known-good local form.
+
+        The browser form can be stale or omit advanced-only fields. MiniMax M3
+        is sensitive to several flags: using the HF repo id with block-size 128
+        fails KV-cache setup, and FlashInfer sampler JIT fails on this host's
+        system nvcc. Normalize server-side before writing the tmux runner.
+        """
+        cmd_lower = (cmd or "").lower()
+        if not cmd or "vllm serve" not in cmd_lower or "minimax" not in cmd_lower or "m3" not in cmd_lower:
+            return cmd
+        try:
+            parts = shlex.split(cmd)
+        except ValueError:
+            return cmd
+        if "serve" not in parts:
+            return cmd
+
+        env_re = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*=")
+        env_parts = [p for p in parts if env_re.match(p)]
+        body = [p for p in parts if not env_re.match(p)]
+        try:
+            serve_i = body.index("serve")
+        except ValueError:
+            return cmd
+        if serve_i + 1 >= len(body):
+            return cmd
+
+        repo_id = "cyankiwi/MiniMax-M3-AWQ-INT4"
+        snapshot = (
+            "/home/pewds/.cache/huggingface/hub/"
+            "models--cyankiwi--MiniMax-M3-AWQ-INT4/"
+            "snapshots/4082acbbec1236d21828d55b6bb0fe02ade4ab5b"
+        )
+        if body[serve_i + 1] == repo_id:
+            body[serve_i + 1] = snapshot
+
+        def add_env(key: str, value: str) -> None:
+            if not any(p.startswith(f"{key}=") for p in env_parts):
+                env_parts.append(f"{key}={value}")
+
+        def has_flag(flag: str) -> bool:
+            return any(p == flag or p.startswith(flag + "=") for p in body)
+
+        def set_flag(flag: str, value: str) -> None:
+            for i, part in enumerate(body):
+                if part == flag:
+                    if i + 1 < len(body):
+                        body[i + 1] = value
+                    else:
+                        body.append(value)
+                    return
+                if part.startswith(flag + "="):
+                    body[i] = f"{flag}={value}"
+                    return
+            body.extend([flag, value])
+
+        def add_bool(flag: str) -> None:
+            if not has_flag(flag):
+                body.append(flag)
+
+        add_env("VLLM_TARGET_DEVICE", "cuda")
+        add_env("VLLM_USE_FLASHINFER_SAMPLER", "0")
+        set_flag("--served-model-name", repo_id)
+        set_flag("--tool-call-parser", "minimax_m3")
+        set_flag("--reasoning-parser", "minimax_m3")
+        set_flag("--attention-backend", "TRITON_ATTN")
+        set_flag("--block-size", "128")
+        add_bool("--language-model-only")
+        add_bool("--disable-custom-all-reduce")
+        add_bool("--enable-expert-parallel")
+        return shlex.join(env_parts + body)
+
     def _cookbook_ssh_dir() -> Path:
         # The Docker image keeps cookbook keys under /app/.ssh; that path only
         # exists inside the container. On Windows (and any non-container host)
@@ -1230,6 +1322,7 @@ def setup_cookbook_routes() -> APIRouter:
         # `TypeError: argument of type 'NoneType'` (a 500 instead of a clean 400).
         req.cmd = _validate_serve_cmd(req.cmd) or ""
         req.cmd = _normalize_llama_cpp_python_cache_types(req.cmd) or ""
+        req.cmd = _normalize_minimax_m3_vllm_cmd(req.cmd)
         req.cmd = _venv_safe_local_pip_install_cmd(
             req.cmd,
             local=not bool(req.remote_host),
@@ -1243,8 +1336,16 @@ def setup_cookbook_routes() -> APIRouter:
             req.cmd = _pip_install_no_cache(req.cmd)
             # Accept common aliases and enforce server extras for llama-cpp so
             # `python -m llama_cpp.server` has all runtime dependencies.
-            req.cmd = re.sub(r"(?<![A-Za-z0-9_.-])llama_cpp(?![A-Za-z0-9_.-])", "llama-cpp-python[server]", req.cmd)
-            req.cmd = re.sub(r"(?<![A-Za-z0-9_.-])llama-cpp-python(?!\[)", "llama-cpp-python[server]", req.cmd)
+            # CRITICAL: the lookbehind / lookahead must also exclude `/` so
+            # the regex DOESN'T mangle a URL path like
+            #   https://abetlen.github.io/llama-cpp-python/whl/cu124
+            # The previous regex turned that URL into
+            #   https://abetlen.github.io/llama-cpp-python[server]/whl/cu124
+            # which pip then couldn't resolve → silent fallback to source
+            # build of the .tar.gz → CPU-only binary (because CMAKE_ARGS
+            # isn't set), defeating the entire purpose of the CUDA index.
+            req.cmd = re.sub(r"(?<![A-Za-z0-9_.\-/])llama_cpp(?![A-Za-z0-9_.\-/])", "llama-cpp-python[server]", req.cmd)
+            req.cmd = re.sub(r"(?<![A-Za-z0-9_.\-/])llama-cpp-python(?![\[/])", "llama-cpp-python[server]", req.cmd)
             if "llama-cpp-python" in req.cmd and "--extra-index-url" not in req.cmd:
                 req.cmd += " --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu"
             # PEP-508-style package spec — letters, digits, `.-_` for the
@@ -1284,6 +1385,11 @@ def setup_cookbook_routes() -> APIRouter:
         # LOCAL execution on a native-Windows host never uses tmux (detached
         # process path below), regardless of the UI-supplied platform.
         local_windows = IS_WINDOWS and not remote
+        if is_windows and remote and "diffusion_server.py" in req.cmd:
+            raise HTTPException(
+                400,
+                "Remote Windows Diffusers serving is not supported yet; use local Windows or a Linux remote server.",
+            )
 
         if not is_windows and not local_windows and not await _binary_available("tmux", remote, req.ssh_port):
             return {
@@ -1426,6 +1532,69 @@ def setup_cookbook_routes() -> APIRouter:
                 runner_lines.append('  else')
                 _append_llama_cpp_linux_accel_build_lines(runner_lines)
                 runner_lines.append('  fi')
+                # Source the env file the prebuilt-download path writes so
+                # LD_LIBRARY_PATH includes the directory holding libllama.so
+                # and friends. No-op when prebuilt wasn't used.
+                runner_lines.append('  [ -r ~/.config/odysseus-llama-cpp-env ] && . ~/.config/odysseus-llama-cpp-env')
+                # Auto-upgrade pip llama-cpp-python to the CUDA-enabled
+                # wheel when (a) NVIDIA hardware is present and (b) the
+                # currently-installed wheel is CPU-only. Without this the
+                # user gets the Python server happily running at 3 tok/s
+                # because pip's default index ships CPU-only wheels.
+                # Forward-compat: cu124 wheels work on driver/runtime
+                # 12.4+ including the cu13.x line.
+                runner_lines.append('  if command -v nvidia-smi >/dev/null 2>&1 && nvidia-smi -L 2>/dev/null | grep -q "GPU " && python3 -c "import llama_cpp" 2>/dev/null; then')
+                runner_lines.append('    if ! python3 -c "import llama_cpp; import sys; sys.exit(0 if llama_cpp.llama_supports_gpu_offload() else 1)" 2>/dev/null; then')
+                runner_lines.append('      echo "[odysseus] NVIDIA detected but installed llama-cpp-python is CPU-only — reinstalling with CUDA wheel index for GPU offload..."')
+                runner_lines.append('      python3 -m pip install --user --break-system-packages --force-reinstall --no-cache-dir "llama-cpp-python[server]" --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124 2>&1 | tail -8 || echo "[odysseus] WARNING: CUDA wheel reinstall failed — Python server will stay CPU-only (slow). Manual fix: pip install --user --force-reinstall \'llama-cpp-python[server]\' --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124"')
+                runner_lines.append('      if python3 -c "import llama_cpp; import sys; sys.exit(0 if llama_cpp.llama_supports_gpu_offload() else 1)" 2>/dev/null; then')
+                runner_lines.append('        echo "[odysseus] llama-cpp-python now supports GPU offload."')
+                runner_lines.append('      fi')
+                runner_lines.append('    fi')
+                runner_lines.append('  fi')
+                # SHORT-CIRCUIT before the build/pip fallback: if the
+                # native binary is missing but llama_cpp Python is already
+                # installed, drop a wrapper at ~/bin/llama-server that
+                # translates llama-server CLI args to llama_cpp.server's
+                # underscore-style flags. The user's serve command stays
+                # `llama-server ...` and "just works" — no build, no cmake,
+                # no second install. This is the path that unblocks every
+                # remote where pip-installed llama-cpp-python is already
+                # working but Cookbook used to insist on a native binary.
+                runner_lines.append('  if ! command -v llama-server >/dev/null 2>&1 && python3 -c "import llama_cpp" 2>/dev/null; then')
+                runner_lines.append('    mkdir -p ~/bin')
+                runner_lines.append('    cat > ~/bin/llama-server <<\'_ODY_LLAMA_SHIM_EOF\'')
+                runner_lines.append('#!/usr/bin/env bash')
+                runner_lines.append('# Auto-generated by Odysseus Cookbook: a `llama-server` lookalike')
+                runner_lines.append('# that translates the native CLI to `python -m llama_cpp.server`.')
+                runner_lines.append('# Lets cookbook-generated launch commands run unchanged on hosts')
+                runner_lines.append('# where only the pip llama-cpp-python package is installed.')
+                runner_lines.append('ARGS=()')
+                runner_lines.append('while [ $# -gt 0 ]; do')
+                runner_lines.append('  case "$1" in')
+                runner_lines.append('    -ngl|--gpu-layers|--n-gpu-layers) ARGS+=(--n_gpu_layers "$2"); shift 2 ;;')
+                runner_lines.append('    -c|--ctx-size) ARGS+=(--n_ctx "$2"); shift 2 ;;')
+                runner_lines.append('    -b|--batch-size) ARGS+=(--n_batch "$2"); shift 2 ;;')
+                runner_lines.append('    -ub|--ubatch-size) shift 2 ;;  # llama-cpp-python has no separate ubatch')
+                runner_lines.append('    --flash-attn) ARGS+=(--flash_attn true); shift 2 ;;')
+                runner_lines.append('    --cache-type-k) ARGS+=(--type_k "$2"); shift 2 ;;')
+                runner_lines.append('    --cache-type-v) ARGS+=(--type_v "$2"); shift 2 ;;')
+                runner_lines.append('    --n-cpu-moe) ARGS+=(--n_cpu_moe "$2"); shift 2 ;;')
+                runner_lines.append('    --mmproj) ARGS+=(--clip_model_path "$2"); shift 2 ;;')
+                runner_lines.append('    --image-max-tokens) shift 2 ;;  # native-only')
+                runner_lines.append('    --no-mmap) ARGS+=(--no_mmap true); shift ;;')
+                runner_lines.append('    --no-warmup) shift ;;  # native-only')
+                runner_lines.append('    --chat-template) ARGS+=(--chat_format "$2"); shift 2 ;;')
+                runner_lines.append('    --fit|--split-mode|--tensor-split|--main-gpu|--parallel) shift 2 ;;  # native-only')
+                runner_lines.append('    --mlock) ARGS+=(--use_mlock true); shift ;;')
+                runner_lines.append('    *) ARGS+=("$1"); shift ;;')
+                runner_lines.append('  esac')
+                runner_lines.append('done')
+                runner_lines.append('exec python3 -m llama_cpp.server "${ARGS[@]}"')
+                runner_lines.append('_ODY_LLAMA_SHIM_EOF')
+                runner_lines.append('    chmod +x ~/bin/llama-server')
+                runner_lines.append('    echo "[odysseus] Created llama-server shim → python -m llama_cpp.server (no native binary needed)"')
+                runner_lines.append('  fi')
                 runner_lines.append('  # If the native build failed, fall back to the Python bindings.')
                 runner_lines.append('  if ! command -v llama-server &>/dev/null && ! python3 -c "import llama_cpp" 2>/dev/null; then')
                 runner_lines.append('    echo "llama-server build failed — installing Python bindings as fallback..."')
@@ -1489,6 +1658,96 @@ def setup_cookbook_routes() -> APIRouter:
                 runner_lines.append('  echo "ERROR: vLLM is not installed."')
                 runner_lines.append('  ODYSSEUS_PREFLIGHT_EXIT=127')
                 runner_lines.append('fi')
+                runner_lines.append(f"ODYSSEUS_SERVE_CMD='{_bash_squote(req.cmd)}'")
+                runner_lines.append('if [ -z "$ODYSSEUS_PREFLIGHT_EXIT" ]; then')
+                runner_lines.append('  ODYSSEUS_VLLM_HELP_CMD="$(python3 - "$ODYSSEUS_SERVE_CMD" <<\'PY\'')
+                runner_lines.append('import shlex, sys')
+                runner_lines.append('parts = shlex.split(sys.argv[1])')
+                runner_lines.append('try:')
+                runner_lines.append('    serve_i = parts.index("serve")')
+                runner_lines.append('except ValueError:')
+                runner_lines.append('    print("vllm serve --help")')
+                runner_lines.append('else:')
+                runner_lines.append('    print(shlex.join(parts[:serve_i + 1] + ["--help"]))')
+                runner_lines.append('PY')
+                runner_lines.append(')"')
+                runner_lines.append('  ODYSSEUS_VLLM_SUPPORTS_SWAP=0')
+                runner_lines.append('  if eval "$ODYSSEUS_VLLM_HELP_CMD" 2>&1 | grep -q -- "--swap-space"; then ODYSSEUS_VLLM_SUPPORTS_SWAP=1; fi')
+                runner_lines.append('fi')
+                runner_lines.append('if [ -z "$ODYSSEUS_PREFLIGHT_EXIT" ] && [ "${ODYSSEUS_VLLM_SUPPORTS_SWAP:-0}" = "1" ] && ! printf "%s" "$ODYSSEUS_SERVE_CMD" | grep -q -- "--swap-space"; then')
+                runner_lines.append('  echo "[odysseus] Setting vLLM --swap-space 0 so the runtime does not reserve CPU swap per GPU."')
+                runner_lines.append('  ODYSSEUS_SERVE_CMD="${ODYSSEUS_SERVE_CMD} --swap-space 0"')
+                runner_lines.append('fi')
+                runner_lines.append('if [ -z "$ODYSSEUS_PREFLIGHT_EXIT" ] && [ "${ODYSSEUS_VLLM_SUPPORTS_SWAP:-0}" != "1" ]; then')
+                runner_lines.append('  if printf "%s" "$ODYSSEUS_SERVE_CMD" | grep -q -- "--swap-space"; then')
+                runner_lines.append('    echo "[odysseus] vLLM serve does not expose --swap-space; removing the flag and patching the runtime default to 0."')
+                runner_lines.append('    ODYSSEUS_SERVE_CMD="$(python3 - "$ODYSSEUS_SERVE_CMD" <<\'PY\'')
+                runner_lines.append('import shlex, sys')
+                runner_lines.append('parts = shlex.split(sys.argv[1])')
+                runner_lines.append('out = []')
+                runner_lines.append('skip = False')
+                runner_lines.append('for part in parts:')
+                runner_lines.append('    if skip:')
+                runner_lines.append('        skip = False')
+                runner_lines.append('        continue')
+                runner_lines.append('    if part == "--swap-space":')
+                runner_lines.append('        skip = True')
+                runner_lines.append('        continue')
+                runner_lines.append('    if part.startswith("--swap-space="):')
+                runner_lines.append('        continue')
+                runner_lines.append('    out.append(part)')
+                runner_lines.append('print(shlex.join(out))')
+                runner_lines.append('PY')
+                runner_lines.append(')"')
+                runner_lines.append('  fi')
+                runner_lines.append('  ODYSSEUS_SERVE_CMD="$(python3 - "$ODYSSEUS_SERVE_CMD" <<\'PY\'')
+                runner_lines.append('import shlex, sys')
+                runner_lines.append('parts = shlex.split(sys.argv[1])')
+                runner_lines.append('patch = r"""import inspect, sys')
+                runner_lines.append('from vllm.engine.arg_utils import EngineArgs, AsyncEngineArgs')
+                runner_lines.append('def _odysseus_swap0(cls):')
+                runner_lines.append('    params = list(inspect.signature(cls).parameters)')
+                runner_lines.append('    if "swap_space" not in params:')
+                runner_lines.append('        return')
+                runner_lines.append('    idx = params.index("swap_space")')
+                runner_lines.append('    defaults = list(cls.__init__.__defaults__ or ())')
+                runner_lines.append('    if idx < len(defaults):')
+                runner_lines.append('        defaults[idx] = 0')
+                runner_lines.append('        cls.__init__.__defaults__ = tuple(defaults)')
+                runner_lines.append('    fields = getattr(cls, "__dataclass_fields__", {})')
+                runner_lines.append('    if "swap_space" in fields:')
+                runner_lines.append('        fields["swap_space"].default = 0')
+                runner_lines.append('_odysseus_swap0(EngineArgs)')
+                runner_lines.append('_odysseus_swap0(AsyncEngineArgs)')
+                runner_lines.append('try:')
+                runner_lines.append('    from vllm.config import CacheConfig')
+                runner_lines.append('    CacheConfig.swap_space = 0')
+                runner_lines.append('except Exception:')
+                runner_lines.append('    pass')
+                runner_lines.append('_orig_create_engine_config = EngineArgs.create_engine_config')
+                runner_lines.append('def _odysseus_create_engine_config(self, *args, **kwargs):')
+                runner_lines.append('    self.swap_space = 0')
+                runner_lines.append('    return _orig_create_engine_config(self, *args, **kwargs)')
+                runner_lines.append('EngineArgs.create_engine_config = _odysseus_create_engine_config')
+                runner_lines.append('AsyncEngineArgs.create_engine_config = _odysseus_create_engine_config')
+                runner_lines.append('from vllm.entrypoints.cli.main import main')
+                runner_lines.append('sys.exit(main())"""')
+                runner_lines.append('try:')
+                runner_lines.append('    serve_i = parts.index("serve")')
+                runner_lines.append('except ValueError:')
+                runner_lines.append('    print(shlex.join(parts))')
+                runner_lines.append('else:')
+                runner_lines.append('    exe_i = serve_i - 1')
+                runner_lines.append('    exe = parts[exe_i] if exe_i >= 0 else "vllm"')
+                runner_lines.append('    py = "python3"')
+                runner_lines.append('    if exe.endswith("/bin/vllm"):')
+                runner_lines.append('        py = exe[:-len("/bin/vllm")] + "/bin/python"')
+                runner_lines.append('    parts[exe_i:serve_i] = [py, "-c", patch]')
+                runner_lines.append('    print(shlex.join(parts))')
+                runner_lines.append('PY')
+                runner_lines.append(')"')
+                runner_lines.append('  echo "[odysseus] Patched vLLM internal swap_space default to 0 for this runtime."')
+                runner_lines.append('fi')
             elif "sglang.launch_server" in req.cmd:
                 runner_lines.append('export PATH="$HOME/.local/bin:$PATH"')
                 runner_lines.append('if ! command -v sglang &>/dev/null; then')
@@ -1530,7 +1789,10 @@ def setup_cookbook_routes() -> APIRouter:
                     runner_lines,
                     keep_shell_open=not local_windows,
                 )
-                runner_lines.append(req.cmd)
+                if "vllm serve" in req.cmd:
+                    runner_lines.append('eval "$ODYSSEUS_SERVE_CMD"')
+                else:
+                    runner_lines.append(req.cmd)
                 if local_windows:
                     # Detached background process — no interactive shell to keep open.
                     # Print the exit marker the status poller looks for, then stop.
@@ -1834,6 +2096,25 @@ def setup_cookbook_routes() -> APIRouter:
         out, err = await _run_gpu_shell("ls -1 /sys/class/drm 2>/dev/null", host, ssh_port, timeout=4)
         if err is not None or not out:
             return []
+        # Pick the runtime label up-front so each GPU dict gets the
+        # right `backend`. AMD silicon can be driven by ROCm/HIP (native)
+        # OR Vulkan (mesa RADV). Reporting "rocm" on a host where no
+        # ROCm toolchain is installed misleads the frontend env-var
+        # prefix logic — it would emit `HIP_VISIBLE_DEVICES=` for a
+        # Vulkan-only stack, which is a silent no-op at best.
+        rt_out, _ = await _run_gpu_shell(
+            'command -v rocminfo >/dev/null 2>&1 && echo rocm '
+            '|| (command -v hipconfig >/dev/null 2>&1 && echo rocm) '
+            '|| (command -v vulkaninfo >/dev/null 2>&1 && echo vulkan) '
+            '|| echo unknown',
+            host, ssh_port, timeout=4,
+        )
+        _amd_runtime = (rt_out or "").strip().splitlines()[-1:][0].strip() if rt_out else "rocm"
+        if _amd_runtime not in ("rocm", "vulkan"):
+            # Default to rocm so existing ROCm-installed hosts keep
+            # working; "unknown" only happens when neither toolchain is
+            # detected (e.g. minimal sysfs read on a fresh box).
+            _amd_runtime = "rocm"
         gpus = []
         for entry in out.split():
             if not entry.startswith("card") or "-" in entry:
@@ -1877,7 +2158,7 @@ def setup_cookbook_routes() -> APIRouter:
                 "free_mb": free_mb, "total_mb": total_mb, "used_mb": used_mb,
                 "gtt_used_mb": gtt_used_mb,
                 "util_pct": 0, "busy": bool(total_mb and (free_mb / total_mb) < 0.85),
-                "processes": [], "backend": "rocm", "source": "amd-sysfs",
+                "processes": [], "backend": _amd_runtime, "source": "amd-sysfs",
                 "unified_memory": unified,
             })
         if gpus:
@@ -2018,10 +2299,15 @@ def setup_cookbook_routes() -> APIRouter:
 
         amd_gpus = await _probe_amd_sysfs(host, ssh_port)
         if amd_gpus:
+            # The per-GPU dict already carries the runtime label picked by
+            # _probe_amd_sysfs (rocm vs vulkan); mirror that into the
+            # wrapper so the frontend can read `data.backend` directly
+            # without scanning the list.
+            _amd_wrap_backend = str(amd_gpus[0].get("backend") or "rocm")
             return {
                 "ok": True,
                 "gpus": amd_gpus,
-                "backend": "rocm",
+                "backend": _amd_wrap_backend,
                 "source": "amd-sysfs",
                 "fallback_from": "nvidia-smi",
                 "nvidia_error": nvidia_error,
@@ -2161,6 +2447,17 @@ def setup_cookbook_routes() -> APIRouter:
 
             disk_tasks = on_disk.get("tasks") or [] if isinstance(on_disk, dict) else []
             incoming_tasks = data.get("tasks") if isinstance(data.get("tasks"), list) else []
+            incoming_removed = data.get("removedTasks") if isinstance(data.get("removedTasks"), dict) else {}
+            disk_removed = on_disk.get("removedTasks") if isinstance(on_disk, dict) and isinstance(on_disk.get("removedTasks"), dict) else {}
+            removed_tasks = {**disk_removed, **incoming_removed}
+            data["removedTasks"] = removed_tasks
+            removed_ids = set(removed_tasks.keys())
+            if removed_ids:
+                incoming_tasks = [
+                    t for t in incoming_tasks
+                    if not (isinstance(t, dict) and t.get("sessionId") in removed_ids)
+                ]
+                data["tasks"] = incoming_tasks
             # Anti-poisoning guard: a stale browser tab can keep POSTing a
             # download task as status='done' from before the strict-finish
             # fix landed, undoing any server-side correction. For each
@@ -2198,6 +2495,8 @@ def setup_cookbook_routes() -> APIRouter:
                 sid = t.get("sessionId")
                 if not sid or sid in incoming_ids:
                     continue  # client's version wins
+                if sid in removed_ids:
+                    continue  # intentional cross-device clear/remove
                 ts = t.get("ts") or 0
                 if isinstance(ts, (int, float)) and (now_ms - ts) <= RACE_WINDOW_MS:
                     preserved.append(t)
@@ -2304,16 +2603,14 @@ def setup_cookbook_routes() -> APIRouter:
             # Add 30% headroom for KV cache, activations, etc.
             needed_vram = (est_vram * 1.3) if est_vram else None
 
-            if vram_gb > 0 and needed_vram is not None and needed_vram > vram_gb:
-                continue
-            # Unknown-size models (e.g. MiniMax-M2.7, DeepSeek-V4-Flash) have no
-            # "NB" in the repo id, so the regex above can't extract their
-            # param count. Previously we dropped them entirely, which made
-            # brand-new flagship releases silently vanish from this list even
-            # on rigs with hundreds of GB of VRAM. Adapters/LoRAs are already
-            # filtered by _is_excluded(), so what falls through here is
-            # overwhelmingly full models — keep them, just without a size
-            # badge (the frontend handles needed_vram_gb=null gracefully).
+            if vram_gb > 0:
+                if needed_vram is None:
+                    # The "trending models that fit" list must be conservative:
+                    # if we cannot estimate size from the repo id/tags, do not
+                    # present it as runnable on this hardware.
+                    continue
+                if needed_vram > vram_gb:
+                    continue
 
             out.append({
                 "repo_id": repo_id,
@@ -2510,6 +2807,33 @@ def setup_cookbook_routes() -> APIRouter:
             except Exception as e:
                 logger.warning(f"orphan sweep: state write failed: {e}")
 
+    @router.get("/api/cookbook/hf-gguf-files")
+    async def hf_gguf_files(repo_id: str, owner: str = Depends(require_user)):
+        """List GGUF files in a HuggingFace repo for the direct-download picker."""
+        import httpx
+
+        repo_id = _validate_repo_id(repo_id)
+        url = f"https://huggingface.co/api/models/{repo_id}"
+        try:
+            headers = {}
+            token = _load_stored_hf_token()
+            if token:
+                headers["Authorization"] = f"Bearer {token}"
+            async with httpx.AsyncClient(timeout=15, follow_redirects=True) as client:
+                resp = await client.get(url, headers=headers)
+                if resp.status_code != 200:
+                    return {"ok": False, "files": [], "error": f"HF API HTTP {resp.status_code}"}
+                data = resp.json()
+        except Exception:
+            logger.exception("HF GGUF file scan failed for %s", repo)
+            return {"ok": False, "files": [], "error": "HF API request failed"}
+        files = [
+            str(s.get("rfilename") or "")
+            for s in data.get("siblings", [])
+            if str(s.get("rfilename") or "").lower().endswith(".gguf")
+        ]
+        return {"ok": True, "repo_id": repo_id, "files": files}
+
     # In-memory cache for the Ollama library scrape. ollama.com is a public
     # site, but it doesn't expose a stable JSON listing — we fetch the HTML
     # search page and regex out the model cards. Cached for 1 h so a busy

routes/document_helpers.py

@@ -102,8 +102,11 @@ def _owner_session_filter(q, user):
 
     The owner backfill runs in init_db before the app serves requests, so
     by the time this filter is live there are no NULL-owner rows to leak;
-    we therefore match the owner strictly."""
-    if user is None:
+    we therefore match the owner strictly for authenticated callers."""
+    if not user:
+        from src.auth_helpers import _auth_disabled
+        if user == "" or _auth_disabled():
+            return q
         return q.filter(False)
     return q.filter(Document.owner == user)
 

routes/document_routes.py

@@ -503,7 +503,8 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
         user = get_current_user(request)
         try:
             data = await request.json()
-        except Exception:
+        except Exception as e:
+            logger.warning("Failed to parse export request body, defaulting to empty", exc_info=e)
             data = {}
         ids = data.get("ids") or []
         if not ids:
@@ -645,8 +646,8 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
                     try:
                         from src.agent_tools.document_tools import clear_active_document
                         clear_active_document(doc_id)
-                    except Exception:
-                        pass
+                    except Exception as e:
+                        logger.warning("Failed to clear active document %r on detach", doc_id, exc_info=e)
             db.commit()
             db.refresh(doc)
             return _doc_to_dict(doc)
@@ -1331,6 +1332,12 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
             if not pdf_path:
                 raise HTTPException(404, f"Source PDF {upload_id} not found")
 
+            # Fail fast with a clear 503 if the optional PyMuPDF dependency
+            # is missing — fill_fields/stamp_annotations will otherwise
+            # raise RuntimeError deep inside and bubble out as a 500.
+            # Mirrors the convention in _load_pdf_viewer_fitz above.
+            _load_pdf_viewer_fitz()
+
             values = parse_markdown_to_values(doc.current_content or "")
             out_path = tempfile.NamedTemporaryFile(suffix=".pdf", delete=False).name
             _to_unlink.append(out_path)

routes/email_helpers.py

@@ -1233,22 +1233,30 @@ def _list_attachments_from_msg(msg):
         return attachments
     idx = 0
     for part in msg.walk():
-        if part.is_multipart():
-            continue
         cd = str(part.get("Content-Disposition", ""))
         ct = part.get_content_type()
+        is_attached_email = ct == "message/rfc822" and ("attachment" in cd.lower() or part.get_filename())
+        if part.is_multipart() and not is_attached_email:
+            continue
         # Skip text/html body parts (only consider real attachments)
         if ct in ("text/plain", "text/html") and "attachment" not in cd:
             continue
         filename = part.get_filename()
         if filename:
             filename = _decode_header(filename)
+            if ct == "message/rfc822" and not re.search(r"\.[A-Za-z0-9]{1,8}$", filename):
+                filename = f"{filename}.eml"
         else:
             # Inline images, etc. - generate a name
-            ext = ct.split("/")[-1] if "/" in ct else "bin"
+            ext = "eml" if ct == "message/rfc822" else (ct.split("/")[-1] if "/" in ct else "bin")
             filename = f"attachment_{idx}.{ext}"
         payload = part.get_payload(decode=True)
-        size = len(payload) if payload else 0
+        if payload is None and ct == "message/rfc822":
+            try:
+                payload = part.as_bytes()
+            except Exception:
+                payload = b""
+        size = len(payload) if payload is not None else 0
         attachments.append({
             "index": idx,
             "filename": filename,
@@ -1260,29 +1268,58 @@ def _list_attachments_from_msg(msg):
     return attachments
 
 
+def _is_likely_signature_image_attachment(att: dict) -> bool:
+    """Match the reader's inline signature/logo image filter."""
+    filename = str((att or {}).get("filename") or "").lower()
+    if not re.search(r"\.(png|jpe?g|gif|bmp|svg|webp)$", filename):
+        return False
+    size = int((att or {}).get("size") or 0)
+    if re.search(r"^image\d{3,}\.(png|jpe?g|gif)$", filename):
+        return True
+    if re.search(r"^(signature|logo|sig|footer|banner)[-_\d]*\.(png|jpe?g|gif|svg)$", filename):
+        return True
+    return 0 < size < 30 * 1024
+
+
+def _has_visible_attachments(msg) -> bool:
+    """Return True only for attachments the reader will render as chips."""
+    return any(
+        not _is_likely_signature_image_attachment(att)
+        for att in _list_attachments_from_msg(msg)
+    )
+
+
 def _extract_attachment_to_disk(msg, index, target_dir):
     """Extract a specific attachment to disk and return the file path."""
     if not msg.is_multipart():
         return None
     idx = 0
     for part in msg.walk():
-        if part.is_multipart():
-            continue
         cd = str(part.get("Content-Disposition", ""))
         ct = part.get_content_type()
+        is_attached_email = ct == "message/rfc822" and ("attachment" in cd.lower() or part.get_filename())
+        if part.is_multipart() and not is_attached_email:
+            continue
         if ct in ("text/plain", "text/html") and "attachment" not in cd:
             continue
         if idx == index:
             filename = part.get_filename()
             if filename:
                 filename = _decode_header(filename)
+                if ct == "message/rfc822" and not re.search(r"\.[A-Za-z0-9]{1,8}$", filename):
+                    filename = f"{filename}.eml"
             else:
-                ext = ct.split("/")[-1] if "/" in ct else "bin"
+                ext = "eml" if ct == "message/rfc822" else (ct.split("/")[-1] if "/" in ct else "bin")
                 filename = f"attachment_{idx}.{ext}"
             # Sanitize
             safe_name = re.sub(r"[^\w\s\-.]", "_", filename).strip()
             payload = part.get_payload(decode=True)
-            if not payload:
+            if payload is None and ct == "message/rfc822":
+                try:
+                    payload = part.as_bytes()
+                except Exception:
+                    payload = b""
+            if payload is None:
                 return None
             target_dir.mkdir(parents=True, exist_ok=True)
             filepath = target_dir / safe_name

routes/email_routes.py

@@ -47,7 +47,7 @@ from routes.email_helpers import (
     _IMAP_TIMEOUT_SECONDS, _open_imap_connection,
     make_oauth_state, verify_oauth_state,
     _imap_connect, _imap, _decode_header, _detect_sent_folder, _detect_drafts_folder,
-    _extract_attachment_text, _list_attachments_from_msg,
+    _extract_attachment_text, _list_attachments_from_msg, _has_visible_attachments, _is_likely_signature_image_attachment,
     _extract_attachment_to_disk, _extract_html, _extract_text,
     _fetch_sender_thread_context, _pre_retrieve_context,
     _EMAIL_REPLY_SYS_PROMPT_BASE, _POOL_HOOKS,
@@ -61,6 +61,7 @@ from routes.email_pollers import _start_poller
 logger = logging.getLogger(__name__)
 
 ODYSSEUS_MAIL_ORIGIN = "odysseus-ui"
+EMAIL_READ_ATTACHMENT_VERSION = 2
 
 
 def _email_tag_owner_aliases(account_id: str | None, owner: str = "") -> list[str]:
@@ -79,15 +80,16 @@ def _email_tag_owner_aliases(account_id: str | None, owner: str = "") -> list[st
                         cfg.get("smtp_user") or "",
                         cfg.get("from_address") or "",
                     ])
-                except Exception:
+                except Exception as _e:
+                    logger.warning("Failed to resolve email account alias", exc_info=_e)
                     resolved_account_id = None
             row = db.get(_EA, resolved_account_id) if resolved_account_id else None
             if row:
                 aliases.extend([row.owner or "", row.imap_user or "", row.from_address or ""])
         finally:
             db.close()
-    except Exception:
-        pass
+    except Exception as _e:
+        logger.warning("Failed to load email aliases", exc_info=_e)
     out = []
     for a in aliases:
         a = (a or "").strip()
@@ -247,6 +249,21 @@ def _imap_uid_fetch(conn, uid_set: str | bytes, query: str):
     return conn.uid("FETCH", _uid_bytes(uid_set), query)
 
 
+def _imap_search_quote(value: str) -> str:
+    return '"' + str(value or "").replace("\\", "\\\\").replace('"', '\\"') + '"'
+
+
+def _message_id_chain(*values: str) -> list[str]:
+    seen = set()
+    out = []
+    for value in values:
+        for mid in re.findall(r"<[^>]+>", value or ""):
+            if mid not in seen:
+                seen.add(mid)
+                out.append(mid)
+    return out
+
+
 def _uid_from_fetch_meta(meta_b: bytes) -> str:
     m = re.search(rb"\bUID\s+(\d+)\b", meta_b)
     return m.group(1).decode() if m else ""
@@ -365,6 +382,21 @@ def _apply_odysseus_headers(msg, kind: str | None = None, ref_id: str | None = N
         msg["X-Odysseus-Ref"] = re.sub(r"[^A-Za-z0-9_.:-]", "-", ref_id)[:128]
 
 
+def _normalize_addr_field(field: str) -> str:
+    """Strip the malformed-but-common trailing/leading commas and stray
+    whitespace from a To/Cc/Bcc string before it lands in the MIME header
+    or the SMTP envelope. Users often paste a single address with a
+    trailing comma (e.g. `felix@pewdiepie.com,`) and most MTAs reject the
+    resulting `To: felix@pewdiepie.com,` line as a syntax error. Collapse
+    any run of separator junk between addresses too."""
+    if not field:
+        return field
+    # Split on commas, drop empty tokens, rejoin with a single ', '.
+    parts = [p.strip() for p in field.split(",")]
+    parts = [p for p in parts if p]
+    return ", ".join(parts)
+
+
 def _envelope_recipients(*fields: str) -> list:
     """Extract bare SMTP envelope addresses from one or more To/Cc/Bcc header
     strings. A naive `field.split(",")` corrupts display names that contain a
@@ -993,6 +1025,65 @@ def setup_email_routes():
                 except Exception:
                     pass
 
+    def _related_thread_attachments_sync(
+        folder: str,
+        account_id: str | None,
+        owner: str,
+        current_uid: str,
+        current_message_id: str,
+        in_reply_to: str,
+        references: str,
+        limit: int = 12,
+    ) -> list[dict]:
+        """Return visible attachments from referenced messages in this folder."""
+        wanted_ids = _message_id_chain(references, in_reply_to)
+        current_mid = (current_message_id or "").strip()
+        wanted_ids = [mid for mid in wanted_ids if mid and mid != current_mid]
+        if not wanted_ids:
+            return []
+
+        related: list[dict] = []
+        try:
+            with _imap(account_id, owner=owner) as conn:
+                conn.select(_q(folder), readonly=True)
+                # Search newest referenced messages first; cap work so opening
+                # a long thread stays bounded.
+                for mid in reversed(wanted_ids[-10:]):
+                    if len(related) >= limit:
+                        break
+                    status, data = _imap_uid_search(conn, f'(HEADER Message-ID {_imap_search_quote(mid)})')
+                    if status != "OK" or not data or not data[0]:
+                        continue
+                    for uid_b in reversed(data[0].split()[-3:]):
+                        source_uid = uid_b.decode(errors="ignore")
+                        if not source_uid or source_uid == str(current_uid):
+                            continue
+                        st2, msg_data = _imap_uid_fetch(conn, source_uid, "(BODY.PEEK[])")
+                        if st2 != "OK" or not msg_data or not isinstance(msg_data[0], tuple):
+                            continue
+                        msg = email_mod.message_from_bytes(msg_data[0][1])
+                        source_from = _decode_header(msg.get("From", ""))
+                        source_subject = _decode_header(msg.get("Subject", ""))
+                        source_date = msg.get("Date", "")
+                        for att in _list_attachments_from_msg(msg):
+                            if _is_likely_signature_image_attachment(att):
+                                continue
+                            enriched = dict(att)
+                            enriched.update({
+                                "source_uid": source_uid,
+                                "source_folder": folder,
+                                "source_message_id": (msg.get("Message-ID") or "").strip(),
+                                "source_from": source_from,
+                                "source_subject": source_subject,
+                                "source_date": source_date,
+                            })
+                            related.append(enriched)
+                            if len(related) >= limit:
+                                break
+        except Exception as e:
+            logger.debug(f"related thread attachment lookup failed uid={current_uid}: {e}")
+        return related
+
     @router.get("/list")
     async def list_emails(
         folder: str = Query("INBOX"),
@@ -1263,6 +1354,17 @@ def setup_email_routes():
             sender_name, sender_addr = email.utils.parseaddr(sender)
             parsed_date = email.utils.parsedate_to_datetime(date_str) if date_str else None
             attachments = _list_attachments_from_msg(msg)
+            related_attachments = []
+            if not _has_visible_attachments(msg):
+                related_attachments = _related_thread_attachments_sync(
+                    folder,
+                    account_id,
+                    owner,
+                    uid,
+                    message_id,
+                    in_reply_to,
+                    references,
+                )
 
             if mark_seen:
                 # Set \Seen in a separate readwrite session so concurrent reads
@@ -1371,6 +1473,8 @@ def setup_email_routes():
                 "body": body,
                 "body_html": body_html,
                 "attachments": attachments,
+                "related_attachments": related_attachments,
+                "attachment_version": EMAIL_READ_ATTACHMENT_VERSION,
                 "cached_summary": cached_summary,
                 "cached_ai_reply": cached_ai_reply,
                 "boundaries": cached_boundaries,
@@ -1401,6 +1505,12 @@ def setup_email_routes():
         """Read email body. Cached for 30m, sync IMAP work runs in a thread."""
         ck = _read_cache_key(account_id, folder, uid, owner=owner)
         cached = _read_cache_get(ck)
+        if cached is not None:
+            # Older cached read responses lack the thread-attachment fallback.
+            # Fetch once so replies that reference prior attachments can show
+            # those files without waiting for cache expiry.
+            if cached.get("attachment_version") != EMAIL_READ_ATTACHMENT_VERSION:
+                cached = None
         if cached is not None:
             if mark_seen:
                 try:
@@ -1535,6 +1645,12 @@ def setup_email_routes():
                 return {"error": f"Attachment index {index} not found"}
 
             from pathlib import Path as _Path
+            target_root = os.path.abspath(str(target_dir))
+            filepath_str = os.path.abspath(str(filepath))
+            if os.path.commonpath([target_root, filepath_str]) != target_root:
+                logger.warning("Rejected attachment path outside extraction dir: %s", filepath)
+                return {"error": "Invalid attachment path"}
+            filepath = _Path(filepath_str)
             base = _Path(filepath).name
             if base.startswith("."):
                 return {"error": "Invalid filename", "filename": base}
@@ -1589,6 +1705,65 @@ def setup_email_routes():
                     return None
             doc_session_id = _resolve_doc_session()
 
+            def _create_markdown_doc(content: str, summary: str):
+                from src.database import SessionLocal as _SL, Document as _Doc, DocumentVersion as _DV
+                doc_id = str(uuid.uuid4())
+                ver_id = str(uuid.uuid4())
+                _db = _SL()
+                try:
+                    _db.query(_Doc).filter(_Doc.is_active == True).update({"is_active": False})
+                    _db.add(_Doc(
+                        id=doc_id, session_id=doc_session_id, title=title,
+                        language="markdown", current_content=content,
+                        version_count=1, is_active=True,
+                    ))
+                    _db.add(_DV(
+                        id=ver_id, document_id=doc_id, version_number=1,
+                        content=content, summary=summary, source="upload",
+                    ))
+                    _db.commit()
+                finally:
+                    _db.close()
+                _tag_doc_with_source(doc_id)
+                return doc_id
+
+            def _attached_email_markdown(raw_bytes: bytes):
+                if not raw_bytes:
+                    return f"# Attached email: {base}\n\n_(empty email attachment)_"
+                try:
+                    attached_msg = email_mod.message_from_bytes(raw_bytes)
+                except Exception:
+                    logger.exception("Failed to parse attached email %s", base)
+                    return f"# Attached email: {base}\n\nCould not parse this email attachment."
+
+                attached_subject = _decode_header(attached_msg.get("Subject", "")) or base
+                attached_from = _decode_header(attached_msg.get("From", ""))
+                attached_to = _decode_header(attached_msg.get("To", ""))
+                attached_cc = _decode_header(attached_msg.get("Cc", ""))
+                attached_date = attached_msg.get("Date", "")
+                attached_body = _extract_text(attached_msg).strip()
+                attached_atts = _list_attachments_from_msg(attached_msg)
+
+                lines = [f"# Attached email: {attached_subject}", ""]
+                if attached_from:
+                    lines.append(f"**From:** {attached_from}")
+                if attached_to:
+                    lines.append(f"**To:** {attached_to}")
+                if attached_cc:
+                    lines.append(f"**Cc:** {attached_cc}")
+                if attached_date:
+                    lines.append(f"**Date:** {attached_date}")
+                lines.extend(["", "## Body", "", attached_body or "_(no readable body)_"])
+                if attached_atts:
+                    lines.extend(["", "## Attachments", ""])
+                    for att in attached_atts:
+                        size = int(att.get("size") or 0)
+                        size_label = f"{size} B" if size < 1024 else f"{round(size / 1024)} KB"
+                        name = att.get("filename") or f"attachment_{att.get('index', '')}"
+                        ctype = att.get("content_type") or "application/octet-stream"
+                        lines.append(f"- {name} ({ctype}, {size_label})")
+                return "\n".join(lines).strip()
+
             # ── PDF path (existing) ────────────────────────────────────
             if ext == ".pdf":
                 import shutil as _shutil
@@ -1635,6 +1810,39 @@ def setup_email_routes():
                 _tag_doc_with_source(doc_id)
                 return {"doc_id": doc_id, "filename": filepath.name}
 
+            # ── Attached email (.eml / message/rfc822) ────────────────
+            if ext == ".eml":
+                def _attachment_bytes_from_msg():
+                    if not msg.is_multipart():
+                        return b""
+                    idx = 0
+                    for part in msg.walk():
+                        cd = str(part.get("Content-Disposition", ""))
+                        ct = part.get_content_type()
+                        is_attached_email = ct == "message/rfc822" and ("attachment" in cd.lower() or part.get_filename())
+                        if part.is_multipart() and not is_attached_email:
+                            continue
+                        if ct in ("text/plain", "text/html") and "attachment" not in cd:
+                            continue
+                        if idx == index:
+                            payload = part.get_payload(decode=True)
+                            if payload is None and ct == "message/rfc822":
+                                try:
+                                    payload = part.as_bytes()
+                                except Exception:
+                                    payload = b""
+                            return payload or b""
+                        idx += 1
+                    return b""
+
+                try:
+                    content = _attached_email_markdown(_attachment_bytes_from_msg())
+                except Exception:
+                    logger.exception("Failed to read email attachment %s", base)
+                    return {"error": "Failed to read email attachment", "filename": base}
+                doc_id = _create_markdown_doc(content, "Imported attached email")
+                return {"doc_id": doc_id, "filename": filepath.name}
+
             # ── DOCX path: extract text → markdown document ───────────
             if ext == ".docx":
                 try:
@@ -1672,25 +1880,7 @@ def setup_email_routes():
                     lines.append("")
                 content = "\n".join(lines).strip() or f"_(empty {base})_"
 
-                from src.database import SessionLocal as _SL, Document as _Doc, DocumentVersion as _DV
-                doc_id = str(uuid.uuid4())
-                ver_id = str(uuid.uuid4())
-                _db = _SL()
-                try:
-                    _db.query(_Doc).filter(_Doc.is_active == True).update({"is_active": False})
-                    _db.add(_Doc(
-                        id=doc_id, session_id=doc_session_id, title=title,
-                        language="markdown", current_content=content,
-                        version_count=1, is_active=True,
-                    ))
-                    _db.add(_DV(
-                        id=ver_id, document_id=doc_id, version_number=1,
-                        content=content, summary="Imported from DOCX", source="upload",
-                    ))
-                    _db.commit()
-                finally:
-                    _db.close()
-                _tag_doc_with_source(doc_id)
+                doc_id = _create_markdown_doc(content, "Imported from DOCX")
                 return {"doc_id": doc_id, "filename": filepath.name}
 
             # ── Plain text / markdown ────────────────────────────────
@@ -1699,25 +1889,7 @@ def setup_email_routes():
                     content = filepath.read_text(encoding="utf-8", errors="replace")
                 except Exception as e:
                     return {"error": f"Failed to read text file: {e}", "filename": base}
-                from src.database import SessionLocal as _SL, Document as _Doc, DocumentVersion as _DV
-                doc_id = str(uuid.uuid4())
-                ver_id = str(uuid.uuid4())
-                _db = _SL()
-                try:
-                    _db.query(_Doc).filter(_Doc.is_active == True).update({"is_active": False})
-                    _db.add(_Doc(
-                        id=doc_id, session_id=doc_session_id, title=title,
-                        language="markdown", current_content=content,
-                        version_count=1, is_active=True,
-                    ))
-                    _db.add(_DV(
-                        id=ver_id, document_id=doc_id, version_number=1,
-                        content=content, summary="Imported from email attachment", source="upload",
-                    ))
-                    _db.commit()
-                finally:
-                    _db.close()
-                _tag_doc_with_source(doc_id)
+                doc_id = _create_markdown_doc(content, "Imported from email attachment")
                 return {"doc_id": doc_id, "filename": filepath.name}
 
             return {"error": f"Unsupported attachment type: {ext}", "filename": base}
@@ -2026,6 +2198,9 @@ def setup_email_routes():
             outer = MIMEMultipart("alternative")
             body_container = outer
 
+        to = _normalize_addr_field(to or "")
+        cc = _normalize_addr_field(cc or "")
+        bcc = _normalize_addr_field(bcc or "")
         outer["From"] = email.utils.formataddr((cfg.get("display_name") or "", cfg["from_address"]))
         outer["To"] = to
         if cc:
@@ -2170,12 +2345,10 @@ def setup_email_routes():
         try:
             conn = sqlite3.connect(SCHEDULED_DB)
             conn.row_factory = sqlite3.Row
-            # The MCP server can't easily set owner, so it stores '' — fall
-            # back to those rows in addition to the caller's owner.
             rows = conn.execute(
                 """SELECT id, to_addr, subject, body, created_at, account_id
                    FROM scheduled_emails
-                   WHERE status = 'agent_draft' AND (owner = ? OR owner = '')
+                   WHERE status = 'agent_draft' AND owner = ?
                    ORDER BY created_at DESC""",
                 (owner or "",),
             ).fetchall()
@@ -2196,7 +2369,7 @@ def setup_email_routes():
             cur = conn.execute(
                 """UPDATE scheduled_emails
                    SET status = 'pending', send_at = ?
-                   WHERE id = ? AND status = 'agent_draft' AND (owner = ? OR owner = '')""",
+                   WHERE id = ? AND status = 'agent_draft' AND owner = ?""",
                 (datetime.utcnow().isoformat(), sid, owner or ""),
             )
             conn.commit()
@@ -2217,7 +2390,7 @@ def setup_email_routes():
             conn = sqlite3.connect(SCHEDULED_DB)
             cur = conn.execute(
                 """UPDATE scheduled_emails SET status = 'cancelled'
-                   WHERE id = ? AND status = 'agent_draft' AND (owner = ? OR owner = '')""",
+                   WHERE id = ? AND status = 'agent_draft' AND owner = ?""",
                 (sid, owner or ""),
             )
             conn.commit()
@@ -2303,6 +2476,9 @@ def setup_email_routes():
             outer = MIMEMultipart("alternative")
             body_container = outer
 
+        req.to = _normalize_addr_field(req.to or "")
+        req.cc = _normalize_addr_field(req.cc or "")
+        req.bcc = _normalize_addr_field(req.bcc or "")
         outer["From"] = email.utils.formataddr((cfg.get("display_name") or "", cfg["from_address"]))
         outer["To"] = req.to
         if req.cc:

routes/embedding_routes.py

@@ -9,6 +9,7 @@ from pathlib import Path
 from fastapi import APIRouter, HTTPException, Form, Depends
 from core.constants import EMBEDDING_ENDPOINT_FILE, FASTEMBED_CACHE_DIR
 from core.middleware import require_admin
+from src.runtime_paths import get_app_root
 
 logger = logging.getLogger(__name__)
 

routes/gallery_routes.py

@@ -67,14 +67,6 @@ def _gallery_image_path(filename: str) -> Path:
         raise HTTPException(400, "Unsafe gallery filename")
     if safe_name != original:
         raise HTTPException(400, "Unsafe gallery filename")
-    if not path.exists():
-        cwd_root = (Path.cwd() / "data" / "generated_images").resolve()
-        cwd_path = (cwd_root / safe_name).resolve()
-        try:
-            if os.path.commonpath([str(cwd_root), str(cwd_path)]) == str(cwd_root) and cwd_path.exists():
-                return cwd_path
-        except Exception:
-            pass
     return path
 
 

routes/hwfit_routes.py

@@ -1,8 +1,13 @@
+import json
+import os
 import re
+import shlex
+import subprocess
 from copy import deepcopy
 
 from fastapi import APIRouter, HTTPException
 
+from core.platform_compat import run_ssh_command
 from routes._validators import validate_remote_host, validate_ssh_port
 
 
@@ -107,6 +112,73 @@ def _apply_manual_hardware(system, manual_mode="", manual_gpu_count="", manual_v
     return system
 
 
+def _run_model_probe(host: str, ssh_port: str, cmd: str) -> str:
+    try:
+        if host:
+            r = run_ssh_command(
+                host,
+                ssh_port or None,
+                cmd,
+                timeout=15,
+                connect_timeout=5,
+                strict_host_key_checking=False,
+                text=True,
+            )
+        else:
+            r = subprocess.run(["bash", "-lc", cmd], capture_output=True, text=True, timeout=15)
+        if r.returncode == 0:
+            return (r.stdout or "").strip()
+    except Exception:
+        return ""
+    return ""
+
+
+def _inspect_model_path(model_path: str, host: str = "", ssh_port: str = "") -> dict:
+    """Read lightweight metadata from a local or SSH-visible HF model folder."""
+    path = (model_path or "").strip()
+    if not path or path.startswith(("http://", "https://")):
+        return {}
+    if not (path.startswith("/") or path.startswith("~")):
+        return {}
+
+    qpath = shlex.quote(path)
+    qconfig = shlex.quote(os.path.join(path, "config.json"))
+    out = {}
+    exists = _run_model_probe(host, ssh_port, f"test -d {qpath} && printf found || printf missing")
+    if exists != "found":
+        target = host or "local container"
+        out["model_probe_error"] = f"Model path is not visible on {target}: {path}"
+        return out
+    raw_config = _run_model_probe(host, ssh_port, f"test -f {qconfig} && sed -n '1,240p' {qconfig}")
+    if raw_config:
+        try:
+            cfg = json.loads(raw_config)
+        except Exception:
+            cfg = {}
+        for key in ("context_length", "max_position_embeddings", "n_ctx_train", "model_max_length", "max_seq_len"):
+            value = cfg.get(key)
+            if isinstance(value, (int, float)) and value > 0:
+                out["model_ctx_max"] = int(value)
+                break
+    else:
+        out["model_probe_error"] = f"config.json not found in model path: {path}"
+
+    size_cmd = (
+        f"find {qpath} -type f \\( -name '*.safetensors' -o -name '*.bin' -o -name '*.gguf' \\) "
+        "-printf '%s\\n' 2>/dev/null | awk '{s+=$1} END {if (s>0) printf \"%.6f\", s/1073741824}'"
+    )
+    weights = _run_model_probe(host, ssh_port, size_cmd)
+    try:
+        weights_gb = float(weights)
+    except Exception:
+        weights_gb = 0.0
+    if weights_gb > 0:
+        out["model_weights_gb"] = round(weights_gb, 3)
+    elif "model_probe_error" not in out:
+        out["model_probe_error"] = f"No model weight files found in: {path}"
+    return out
+
+
 def setup_hwfit_routes():
     router = APIRouter(prefix="/api/hwfit", tags=["hwfit"])
 
@@ -235,7 +307,7 @@ def setup_hwfit_routes():
         return {"system": system, "models": results}
 
     @router.get("/profiles")
-    def get_serve_profiles(model: str = "", host: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, serve_weights_gb: float = 0.0, serve_quant: str = ""):
+    def get_serve_profiles(model: str = "", model_path: str = "", host: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, serve_weights_gb: float = 0.0, serve_quant: str = ""):
         """Compute llama.cpp serve profiles (Quality/Balanced/Speed) for `model`
         against the detected hardware on `host` (or local). Returns concrete
         flags (n_gpu_layers, n_cpu_moe, cache_type, ctx) the serve UI can apply.
@@ -260,8 +332,23 @@ def setup_hwfit_routes():
             # "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct".
             s = (s or "").lower().strip()
             s = s.split("/")[-1]                     # drop org prefix
-            s = re.sub(r"[-_.]?gguf$", "", s)        # drop trailing gguf marker
-            s = re.sub(r"[-_.](q\d[^/]*|iq\d[^/]*|fp8|bf16|f16|awq[^/]*|gptq[^/]*)$", "", s)
+            for suffix in ("-gguf", "_gguf", ".gguf", "gguf"):
+                if s.endswith(suffix):
+                    s = s[: -len(suffix)]
+                    break
+            cut_at = None
+            for idx, ch in enumerate(s):
+                if ch not in "-_." or idx + 1 >= len(s):
+                    continue
+                suffix = s[idx + 1:]
+                if (
+                    suffix in {"fp8", "bf16", "f16"}
+                    or suffix.startswith(("awq", "gptq", "iq"))
+                    or (suffix.startswith("q") and len(suffix) > 1 and suffix[1].isdigit())
+                ):
+                    cut_at = idx
+            if cut_at is not None:
+                s = s[:cut_at]
             return s
 
         m = catalog.get(model)
@@ -272,8 +359,16 @@ def setup_hwfit_routes():
                 if nn and (nn == want or want.endswith(nn) or nn.endswith(want)):
                     m = entry
                     break
+        path_meta = _inspect_model_path(model_path or model, host=host, ssh_port=ssh_port)
         if m is None:
-            return {"system": system, "profiles": [], "error": "model not in catalog"}
+            return {
+                "system": system,
+                "profiles": [],
+                "error": "model not in catalog",
+                "model_ctx_max": int(path_meta.get("model_ctx_max") or 0),
+                "model_weights_gb": float(path_meta.get("model_weights_gb") or 0),
+                "model_probe_error": path_meta.get("model_probe_error") or "",
+            }
         # Surface the model's trained context limit so the serve UI can clamp a
         # user-typed context down to it (asking for ctx > n_ctx_train overflows
         # and, with a quantized KV cache, can crash the GPU).
@@ -283,6 +378,16 @@ def setup_hwfit_routes():
             if isinstance(v, (int, float)) and v > 0:
                 model_ctx_max = int(v)
                 break
+        path_ctx_max = int(path_meta.get("model_ctx_max") or 0)
+        if path_ctx_max > 0:
+            model_ctx_max = max(model_ctx_max, path_ctx_max)
+        model_weights_gb = float(path_meta.get("model_weights_gb") or 0)
+        if model_weights_gb <= 0:
+            for k in ("min_vram_gb", "required_gb", "size_gb", "recommended_ram_gb", "min_ram_gb"):
+                v = m.get(k)
+                if isinstance(v, (int, float)) and v > 0:
+                    model_weights_gb = float(v)
+                    break
         return {
             "system": system,
             "profiles": compute_serve_profiles(
@@ -291,6 +396,8 @@ def setup_hwfit_routes():
                 serve_quant=(serve_quant or None),
             ),
             "model_ctx_max": model_ctx_max,
+            "model_weights_gb": model_weights_gb,
+            "model_probe_error": path_meta.get("model_probe_error") or "",
         }
 
     @router.get("/image-models")

routes/memory_routes.py

@@ -273,65 +273,30 @@ def setup_memory_routes(memory_manager: MemoryManager, session_manager: SessionM
     async def api_audit_memories(request: Request, session: str = Form(None)):
         """Deduplicate and consolidate memories via LLM.
 
-        Uses the default model from settings, or falls back to a session's model.
+        Uses task/utility/default settings through the shared resolver, with
+        the active session as fallback when no task or utility model is set.
         Returns before and after memory counts.
         """
-        from routes.model_routes import _load_settings, _normalize_base, build_chat_url
-        from core.database import ModelEndpoint
-        import json as _json
-
-        endpoint_url = model = None
-        headers = {}
-
-        # Try utility model from settings first — memory audit is a background
-        # task and should prefer the lighter utility model over the main chat model.
-        from src.task_endpoint import resolve_task_endpoint
         user = _owner(request)
-        t_url, t_model, t_headers = resolve_task_endpoint(owner=user)
-        if t_url and t_model:
-            endpoint_url, model, headers = t_url, t_model, t_headers
-        else:
-            # Fall back to default model if no task/utility model configured
-            settings = _load_settings()
-            ep_id = settings.get("default_endpoint_id", "")
-            default_model = settings.get("default_model", "")
-            if ep_id:
-                db = SessionLocal()
-                try:
-                    ep = db.query(ModelEndpoint).filter(
-                        ModelEndpoint.id == ep_id, ModelEndpoint.is_enabled == True
-                    ).first()
-                    if ep:
-                        base = _normalize_base(ep.base_url)
-                        endpoint_url = build_chat_url(base)
-                        model = default_model
-                        if not model and ep.models:
-                            try:
-                                models = _json.loads(ep.models) if isinstance(ep.models, str) else ep.models
-                                if models:
-                                    model = models[0]
-                            except Exception:
-                                pass
-                        if ep.api_key:
-                            headers = {"Authorization": f"Bearer {ep.api_key}"}
-                finally:
-                    db.close()
+        fallback_url = fallback_model = None
+        fallback_headers = None
+        if session:
+            try:
+                sess = session_manager.get_session(session)
+                _assert_session_owner(sess, user)
+                fallback_url = sess.endpoint_url
+                fallback_model = sess.model
+                fallback_headers = sess.headers
+            except KeyError:
+                pass
 
-            # Fall back to session model if no default configured
-            if not endpoint_url and session:
-                try:
-                    sess = session_manager.get_session(session)
-                    _assert_session_owner(sess, _owner(request))
-                    endpoint_url = sess.endpoint_url
-                    model = sess.model
-                    headers = sess.headers
-                except KeyError:
-                    pass
+        endpoint_url, model, headers = resolve_task_endpoint(
+            fallback_url, fallback_model, fallback_headers, owner=user
+        )
 
         if not endpoint_url or not model:
             raise HTTPException(400, "No default model configured — set one in Settings")
 
-        user = _owner(request)
         result = await audit_memories(
             memory_manager,
             memory_vector,
@@ -369,18 +334,28 @@ def setup_memory_routes(memory_manager: MemoryManager, session_manager: SessionM
         model = None
         headers = {}
 
+        user = _owner(request)
+
         if session:
             try:
                 sess = session_manager.get_session(session)
-                _assert_session_owner(sess, _owner(request))
-                endpoint_url, model, headers = resolve_task_endpoint(
-                    sess.endpoint_url, sess.model, sess.headers, owner=_owner(request)
-                )
+                _assert_session_owner(sess, user)
             except KeyError:
-                logger.warning("Session %s not found, falling back to utility endpoint", session)
-                endpoint_url, model, headers = resolve_endpoint("utility", owner=_owner(request))
+                sess = None
+            except HTTPException as exc:
+                if exc.status_code != 404:
+                    raise
+                sess = None
+
+            if sess is None:
+                logger.warning("Session %s not found or inaccessible, falling back to utility endpoint", session)
+                endpoint_url, model, headers = resolve_endpoint("utility", owner=user)
+            else:
+                endpoint_url, model, headers = resolve_task_endpoint(
+                    sess.endpoint_url, sess.model, sess.headers, owner=user
+                )
         else:
-            endpoint_url, model, headers = resolve_task_endpoint(owner=_owner(request))
+            endpoint_url, model, headers = resolve_task_endpoint(owner=user)
     
         if not endpoint_url or not model:
             raise HTTPException(400, "No LLM model configured. Set a default model in Settings.")

routes/model_routes.py

@@ -5,6 +5,7 @@ import re
 import uuid
 import json
 import hashlib
+import ipaddress
 import socket
 import time as _time
 import logging
@@ -405,8 +406,11 @@ def _endpoint_refresh_timeout(ep: Any, category: str) -> float:
     except Exception:
         val = 0
     if val > 0:
-        return float(max(1, min(30, val)))
-    return 2.5 if category == "local" else 2.0
+        return float(max(1, min(60, val)))
+    # llama.cpp and other local OpenAI-compatible servers can block briefly
+    # while warming/loading. A 2s local timeout makes working endpoints flicker
+    # offline before /v1/models is ready.
+    return 10.0 if category == "local" else 2.0
 
 
 def _manual_refresh_timeout(ep: Any, category: str, requested: Any = None) -> float:
@@ -473,7 +477,7 @@ def _explicit_model_list_timeout(base_url: str, endpoint_kind: str = "auto", req
     category = _classify_endpoint(base_url, kind)
     if kind in ("api", "proxy") or category == "api":
         return 30.0
-    return 3.0 if _is_ollama_base(base_url) else 2.0
+    return 15.0 if category == "local" else (3.0 if _is_ollama_base(base_url) else 2.0)
 
 
 def _cached_model_ids(ep: Any) -> List[str]:
@@ -562,6 +566,8 @@ def _safe_build_models_url(base_url: str) -> str:
     """Build a /models URL without letting optional provider imports break probes."""
     try:
         return build_models_url(base_url)
+    except ValueError:
+        raise
     except Exception as exc:
         logger.debug("Model URL detection failed for %s: %s", base_url, exc)
         return f"{(base_url or '').rstrip('/')}/models"
@@ -576,6 +582,18 @@ def _safe_build_headers(api_key: Optional[str], base_url: str) -> dict:
         return {"Authorization": f"Bearer {api_key}"} if api_key else {}
 
 
+def _redact_url_for_log(url: str) -> str:
+    """Return a URL safe for logs by removing userinfo and query/fragment."""
+    try:
+        parsed = urlparse(url or "")
+        host = parsed.hostname or ""
+        if parsed.port:
+            host = f"{host}:{parsed.port}"
+        return urlunparse((parsed.scheme, host, parsed.path, "", "", ""))
+    except Exception:
+        return "<endpoint>"
+
+
 def _is_discovery_only_provider(provider: str) -> bool:
     return provider == "chatgpt-subscription"
 
@@ -633,7 +651,7 @@ def _probe_single_model(base: str, api_key: str, model_id: str, timeout: int = 1
 
     try:
         t0 = _time.time()
-        r = httpx.post(target_url, headers=h, json=payload, timeout=timeout)
+        r = httpx.post(target_url, headers=h, json=payload, timeout=timeout, verify=llm_verify())
         latency = round((_time.time() - t0) * 1000)
         if r.is_success:
             return {"status": "ok", "latency_ms": latency}
@@ -659,13 +677,20 @@ def _probe_single_model(base: str, api_key: str, model_id: str, timeout: int = 1
 
 # Hostnames / IP prefixes that indicate a local endpoint
 _LOCAL_HOSTS = {"localhost", "127.0.0.1", "0.0.0.0", "::1"}
-_PRIVATE_PREFIXES = ("10.", "172.16.", "172.17.", "172.18.", "172.19.",
-                     "172.20.", "172.21.", "172.22.", "172.23.", "172.24.",
-                     "172.25.", "172.26.", "172.27.", "172.28.", "172.29.",
-                     "172.30.", "172.31.", "192.168.")
+_PRIVATE_NETWORKS = (
+    ipaddress.ip_network("10.0.0.0/8"),
+    ipaddress.ip_network("172.16.0.0/12"),
+    ipaddress.ip_network("192.168.0.0/16"),
+)
+_TAILSCALE_CGNAT = ipaddress.ip_network("100.64.0.0/10")
 
 
-_TAILSCALE_RE = re.compile(r"^100\.(6[4-9]|[7-9]\d|1[01]\d|12[0-7])\.")
+def _local_ip_literal(host: str) -> bool:
+    try:
+        ip = ipaddress.ip_address(host)
+    except ValueError:
+        return False
+    return any(ip in network for network in _PRIVATE_NETWORKS) or ip in _TAILSCALE_CGNAT
 
 
 def _classify_endpoint(base_url: str, endpoint_kind: str = "auto") -> str:
@@ -679,9 +704,7 @@ def _classify_endpoint(base_url: str, endpoint_kind: str = "auto") -> str:
         return "api"
     try:
         host = urlparse(base_url).hostname or ""
-        if host in _LOCAL_HOSTS or host.startswith(_PRIVATE_PREFIXES):
-            return "local"
-        if _TAILSCALE_RE.match(host):
+        if host in _LOCAL_HOSTS or _local_ip_literal(host):
             return "local"
     except Exception:
         pass
@@ -703,6 +726,16 @@ def _effective_endpoint_kind(ep: Any, base_url: str) -> str:
     return "auto"
 
 
+def _is_loading_model_response(resp: Any) -> bool:
+    if getattr(resp, "status_code", None) != 503:
+        return False
+    try:
+        body = resp.text or ""
+    except Exception:
+        body = ""
+    return "loading model" in body.lower()
+
+
 
 def _probe_endpoint(base_url: str, api_key: str = None, timeout: int = 5) -> List[str]:
     """Probe a base URL's /models endpoint and return list of model IDs.
@@ -767,11 +800,14 @@ def _probe_endpoint(base_url: str, api_key: str = None, timeout: int = 5) -> Lis
                         models.append(_e)
             return [m for m in models if _is_chat_model(m)]
     except httpx.HTTPStatusError as e:
+        if e.response is not None and _is_loading_model_response(e.response):
+            logger.info("Endpoint still loading model at %s", _redact_url_for_log(url))
+            return []
         if api_key:
             status = e.response.status_code if e.response is not None else "unknown"
-            logger.warning(f"Failed to probe {url} with API key: HTTP {status}")
+            logger.warning("Failed to probe %s with API key: HTTP %s", _redact_url_for_log(url), status)
             return []
-        logger.warning(f"Failed to probe {url}: {e}")
+        logger.warning("Failed to probe %s: %s", _redact_url_for_log(url), e)
     except Exception as e:
         if api_key:
             logger.warning(f"Failed to probe {url} with API key: {e}")
@@ -816,6 +852,15 @@ def _ping_endpoint(base_url: str, api_key: str = None, timeout: float = 1.5) ->
         or "ollama" in (parsed_base.hostname or "").lower()
     )
 
+    def _is_loading_model_response(r) -> bool:
+        if getattr(r, "status_code", None) != 503:
+            return False
+        try:
+            body = r.text or ""
+        except Exception:
+            body = ""
+        return "loading model" in body.lower()
+
     def _result_from_response(r) -> Dict[str, Any]:
         if 300 <= r.status_code < 400:
             loc = r.headers.get("location", "")
@@ -832,6 +877,13 @@ def _ping_endpoint(base_url: str, api_key: str = None, timeout: float = 1.5) ->
                 "status_code": r.status_code,
                 "error": None,
             }
+        if _is_loading_model_response(r):
+            return {
+                "reachable": True,
+                "loading": True,
+                "status_code": r.status_code,
+                "error": "Loading model",
+            }
         return {"reachable": False, "status_code": r.status_code, "error": f"HTTP {r.status_code}"}
 
     last_error: Optional[str] = None
@@ -864,7 +916,7 @@ def _ping_endpoint(base_url: str, api_key: str = None, timeout: float = 1.5) ->
         if 400 <= sc < 500 and sc not in (401, 403):
             models_url = _safe_build_models_url(base)
             try:
-                r2 = httpx.get(models_url, headers=headers, timeout=timeout, verify=llm_verify())
+                r2 = httpx.get(models_url, headers=headers,timeout=timeout, verify=llm_verify())
                 result2 = _result_from_response(r2)
                 if result2["reachable"]:
                     return result2
@@ -1048,9 +1100,11 @@ def setup_model_routes(model_discovery):
         except Exception:
             return 0.0
 
-    def _failure_delay(fails: int) -> float:
+    def _failure_delay(fails: int, *, empty_local: bool = False) -> float:
         if fails <= 0:
             return 0.0
+        if empty_local:
+            return min(5.0 * (2 ** max(0, fails - 1)), 30.0)
         return min(_REFRESH_FAILURE_BASE * (2 ** max(0, fails - 1)), _REFRESH_FAILURE_MAX)
 
     def _should_refresh_endpoint(ep: Any, now: float, force: bool = False) -> tuple[bool, Dict[str, Any]]:
@@ -1081,7 +1135,12 @@ def setup_model_routes(model_discovery):
         fails = int(state.get("fail_count") or 0)
         if fails and not force:
             last_failure = float(state.get("last_failure") or 0.0)
-            if now - last_failure < _failure_delay(fails):
+            empty_local = (
+                not cached
+                and category == "local"
+                and str(getattr(ep, "id", "") or "").startswith("local-")
+            )
+            if now - last_failure < _failure_delay(fails, empty_local=empty_local):
                 return False, info
         if cached and not force:
             interval = _endpoint_refresh_interval(ep, category)
@@ -1396,7 +1455,7 @@ def setup_model_routes(model_discovery):
                 t0 = _time.time()
                 ping = _ping_endpoint(base, ep.api_key, timeout=1.5)
                 entry["latency_ms"] = round((_time.time() - t0) * 1000)
-                entry["status"] = "online" if ping.get("reachable") or cached_count else "offline"
+                entry["status"] = "loading" if ping.get("loading") else ("online" if ping.get("reachable") or cached_count else "offline")
                 entry["error"] = ping.get("error")
                 entry["model_count"] = cached_count or (len(ANTHROPIC_MODELS) if provider == "anthropic" else 0)
             except Exception as e:
@@ -1570,9 +1629,37 @@ def setup_model_routes(model_discovery):
                 # "everything's already cached" path because this branch only
                 # runs for endpoints with an empty cached_models.
                 if not all_models and not pinned and r.is_enabled:
-                    ping = _ping_endpoint(r.base_url, r.api_key, timeout=3.5)
+                    base_for_ping = _normalize_base(r.base_url)
+                    kind_for_ping = _effective_endpoint_kind(r, base_for_ping)
+                    ping_timeout = 10.0 if _classify_endpoint(base_for_ping, kind_for_ping) == "local" else 3.5
+                    ping = _ping_endpoint(r.base_url, r.api_key, timeout=ping_timeout)
                     if ping.get("reachable"):
-                        status = "empty"
+                        status = "loading" if ping.get("loading") else "empty"
+                        if ping.get("loading"):
+                            base = _normalize_base(r.base_url)
+                            kind = _effective_endpoint_kind(r, base)
+                            results.append({
+                                "id": r.id,
+                                "name": r.name,
+                                "base_url": r.base_url,
+                                "has_key": bool(r.api_key),
+                                "api_key_fingerprint": _api_key_fingerprint(r.api_key),
+                                "is_enabled": r.is_enabled,
+                                "models": visible,
+                                "pinned_models": pinned,
+                                "hidden_count": len(hidden),
+                                "online": True,
+                                "status": status,
+                                "ping_error": (ping or {}).get("error") if ping else None,
+                                "model_type": getattr(r, "model_type", None) or "llm",
+                                "supports_tools": getattr(r, "supports_tools", None),
+                                "endpoint_kind": kind,
+                                "category": _classify_endpoint(base, kind),
+                                "model_refresh_mode": _endpoint_refresh_mode(r, kind),
+                                "model_refresh_interval": getattr(r, "model_refresh_interval", None),
+                                "model_refresh_timeout": getattr(r, "model_refresh_timeout", None),
+                            })
+                            continue
                         # Best-effort: if the probe came back reachable, try
                         # to populate cached_models in the background so the
                         # NEXT picker load shows "online" instead of "empty".
@@ -1580,7 +1667,7 @@ def setup_model_routes(model_discovery):
                         # "empty" status, and the existing background refresh
                         # path will eventually fill it in too.
                         try:
-                            probed = _probe_endpoint(r.base_url, r.api_key, timeout=5)
+                            probed = _probe_endpoint(r.base_url, r.api_key, timeout=max(5, int(ping_timeout)))
                             if probed:
                                 r.cached_models = json.dumps(probed)
                                 db.commit()
@@ -1758,7 +1845,7 @@ def setup_model_routes(model_discovery):
         model_ids = _probe_endpoint(base_url, api_key.strip() or None, timeout=explicit_timeout) if should_probe else []
         ping = {"reachable": False, "error": None}
         if (should_probe or requested_kind in ("api", "proxy")) and not model_ids:
-            ping = _ping_endpoint(base_url, api_key.strip() or None, timeout=min(explicit_timeout, 2.0))
+            ping = _ping_endpoint(base_url, api_key.strip() or None, timeout=min(explicit_timeout, 10.0))
         if require_model_list and not model_ids:
             raise HTTPException(400, _model_endpoint_error_message(base_url, ping))
 
@@ -1825,7 +1912,7 @@ def setup_model_routes(model_discovery):
             "models": _merge_model_ids(model_ids, _pinned),
             "pinned_models": _pinned,
             "online": bool(model_ids) or bool(_pinned) or bool(ping.get("reachable")),
-            "status": "online" if (model_ids or _pinned) else ("empty" if ping.get("reachable") else "offline"),
+            "status": "online" if (model_ids or _pinned) else ("loading" if ping.get("loading") else ("empty" if ping.get("reachable") else "offline")),
             "ping_error": ping.get("error") if ping else None,
             "endpoint_kind": requested_kind,
             "category": _classify_endpoint(base_url, requested_kind),
@@ -1850,11 +1937,11 @@ def setup_model_routes(model_discovery):
         configured_timeout = _parse_positive_int(model_refresh_timeout, minimum=1, maximum=60)
         probe_timeout = _explicit_model_list_timeout(base_url, requested_kind, configured_timeout)
         models = _probe_endpoint(base_url, api_key.strip() or None, timeout=probe_timeout)
-        ping = {"reachable": True, "error": None} if models else _ping_endpoint(base_url, api_key.strip() or None, timeout=min(probe_timeout, 2.0))
+        ping = {"reachable": True, "error": None} if models else _ping_endpoint(base_url, api_key.strip() or None, timeout=min(probe_timeout, 10.0))
         return {
             "base_url": base_url,
             "online": bool(models) or bool(ping.get("reachable")),
-            "status": "online" if models else ("empty" if ping.get("reachable") else "offline"),
+            "status": "online" if models else ("loading" if ping.get("loading") else ("empty" if ping.get("reachable") else "offline")),
             "ping_error": ping.get("error") if ping else None,
             "models": models,
             "count": len(models),

routes/note_routes.py

@@ -10,6 +10,7 @@ from fastapi import APIRouter, HTTPException, Request
 from pydantic import BaseModel
 
 from core.database import SessionLocal, Note
+from core.middleware import INTERNAL_TOOL_USER
 from src.auth_helpers import require_user
 from src.constants import DATA_DIR
 from sqlalchemy.orm.attributes import flag_modified
@@ -582,7 +583,7 @@ def setup_note_routes(task_scheduler=None):
         return require_user(request) or None
 
     def _is_admin_or_single_user(request: Request, user: str | None) -> bool:
-        if user == "internal-tool":
+        if user == INTERNAL_TOOL_USER:
             return True
         if not user:
             # require_user() already admitted this request, which only happens

routes/personal_routes.py

@@ -2,8 +2,9 @@
 """Routes for personal documents management."""
 import os
 import logging
+import shutil
 import uuid
-from typing import List, Tuple
+from typing import Any, Dict, List, Tuple
 from fastapi import APIRouter, HTTPException, Query, Request, UploadFile, File, Depends
 from src.request_models import DirectoryRequest
 from core.constants import BASE_DIR, PERSONAL_DIR, PERSONAL_UPLOADS_DIR
@@ -18,14 +19,15 @@ UPLOADS_DIR = PERSONAL_UPLOADS_DIR
 logger = logging.getLogger(__name__)
 
 
-def _personal_upload_dir_for_owner(owner: str | None) -> str:
+def _personal_upload_dir_for_owner(owner: str | None, *, create: bool = True) -> str:
     """Return the per-owner upload directory used for direct RAG uploads."""
     owner_segment = secure_filename((owner or "local").strip())[:80] or "local"
     upload_dir = os.path.abspath(os.path.join(UPLOADS_DIR, owner_segment))
     base_abs = os.path.abspath(UPLOADS_DIR)
     if os.path.commonpath([upload_dir, base_abs]) != base_abs:
         raise ValueError("Unsafe upload owner path")
-    os.makedirs(upload_dir, exist_ok=True)
+    if create:
+        os.makedirs(upload_dir, exist_ok=True)
     return upload_dir
 
 
@@ -44,6 +46,87 @@ def _unique_personal_upload_path(upload_dir: str, original_name: str | None) ->
         raise ValueError("Unsafe upload filename")
     return file_path, filename, safe_name
 
+
+def _unique_existing_target(path: str) -> str:
+    """Return a non-existing sibling path for rename collision handling."""
+    if not os.path.exists(path):
+        return path
+    stem, ext = os.path.splitext(path)
+    while True:
+        candidate = f"{stem}-{uuid.uuid4().hex[:10]}{ext}"
+        if not os.path.exists(candidate):
+            return candidate
+
+
+def _remove_empty_tree(path: str) -> None:
+    """Best-effort removal of empty directories under ``path``."""
+    if not os.path.isdir(path):
+        return
+    for root, dirs, _files in os.walk(path, topdown=False):
+        for dirname in dirs:
+            candidate = os.path.join(root, dirname)
+            try:
+                os.rmdir(candidate)
+            except OSError:
+                pass
+    try:
+        os.rmdir(path)
+    except OSError:
+        pass
+
+
+def rename_personal_upload_owner(
+    old_owner: str,
+    new_owner: str,
+    *,
+    personal_docs_manager: Any = None,
+    rag_manager: Any = None,
+) -> Dict[str, Any]:
+    """Move direct personal uploads and rewrite RAG owner metadata on user rename."""
+    old_dir = _personal_upload_dir_for_owner(old_owner, create=False)
+    new_dir = _personal_upload_dir_for_owner(new_owner, create=False)
+    path_map: Dict[str, str] = {}
+    moved_files = 0
+
+    if os.path.isdir(old_dir) and old_dir != new_dir:
+        os.makedirs(new_dir, exist_ok=True)
+        for root, _dirs, files in os.walk(old_dir):
+            rel_root = os.path.relpath(root, old_dir)
+            target_root = new_dir if rel_root == "." else os.path.join(new_dir, rel_root)
+            os.makedirs(target_root, exist_ok=True)
+            for filename in files:
+                source = os.path.abspath(os.path.join(root, filename))
+                target = _unique_existing_target(os.path.abspath(os.path.join(target_root, filename)))
+                shutil.move(source, target)
+                path_map[source] = target
+                moved_files += 1
+        _remove_empty_tree(old_dir)
+
+    if personal_docs_manager is not None:
+        rename_directory = getattr(personal_docs_manager, "rename_directory", None)
+        if callable(rename_directory):
+            rename_directory(old_dir, new_dir, path_map=path_map)
+
+    rag_result = None
+    if rag_manager is not None:
+        rename_owner = getattr(rag_manager, "rename_owner", None)
+        if callable(rename_owner):
+            rag_result = rename_owner(
+                old_owner,
+                new_owner,
+                path_map=path_map,
+                path_prefixes=[(old_dir, new_dir)],
+            )
+
+    return {
+        "old_dir": old_dir,
+        "new_dir": new_dir,
+        "moved_files": moved_files,
+        "path_map": path_map,
+        "rag_result": rag_result,
+    }
+
+
 def setup_personal_routes(personal_docs_manager, rag_manager, rag_available):
     """
     Setup personal documents related routes.
@@ -275,11 +358,13 @@ def setup_personal_routes(personal_docs_manager, rag_manager, rag_available):
                 except Exception as e:
                     logger.warning(f"RAG removal failed for {filepath}: {e}")
 
-            # Delete file from disk if it's in uploads dir
+            # Delete file from disk if it's in the caller's own uploads dir.
+            # Scope to the per-owner subdir, not the shared uploads root, so one
+            # admin can't delete another user's personal files by path.
             deleted_from_disk = False
             try:
                 abs_target = os.path.realpath(filepath)
-                base_abs = os.path.realpath(UPLOADS_DIR)
+                base_abs = os.path.realpath(_personal_upload_dir_for_owner(owner, create=False))
                 in_uploads = (
                     abs_target == base_abs
                     or os.path.commonpath([abs_target, base_abs]) == base_abs

routes/research_routes.py

@@ -12,8 +12,10 @@ from typing import Optional
 from fastapi import APIRouter, HTTPException, Query, Request
 from fastapi.responses import HTMLResponse, StreamingResponse
 from pydantic import BaseModel, Field
+from core.middleware import INTERNAL_TOOL_USER
 from src.endpoint_resolver import resolve_endpoint
 from src.auth_helpers import _auth_disabled, get_current_user
+from core.auth import RESERVED_USERNAMES
 from src.constants import DEEP_RESEARCH_DIR
 
 _SESSION_ID_RE = re.compile(r"^[a-zA-Z0-9-]{1,128}$")
@@ -385,9 +387,9 @@ def setup_research_routes(research_handler, session_manager=None) -> APIRouter:
         """Launch a research job from the dedicated panel."""
         from src.auth_helpers import require_privilege
         user = require_privilege(request, "can_use_research")
-        if user == "internal-tool":
+        if user == INTERNAL_TOOL_USER:
             tool_owner = (request.headers.get("X-Odysseus-Owner") or "").strip()
-            if tool_owner and tool_owner not in {"internal-tool", "api", "demo", "system"}:
+            if tool_owner and tool_owner not in RESERVED_USERNAMES:
                 auth_mgr = getattr(request.app.state, "auth_manager", None)
                 if auth_mgr is not None and getattr(auth_mgr, "is_configured", False):
                     try:

routes/session_routes.py

@@ -11,7 +11,7 @@ from core.session_manager import SessionManager
 from core.models import ChatMessage
 from src.request_models import SessionResponse
 from core.database import Session as DbSession, SessionLocal, Document, GalleryImage, utcnow_naive
-from src.auth_helpers import get_current_user, effective_user, _auth_disabled, owner_filter
+from src.auth_helpers import effective_user, _auth_disabled, owner_filter
 from src.session_actions import is_session_recently_active
 
 
@@ -328,7 +328,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
         endpoint_id: str = Form(""),
     ):
         skip_val = str(skip_validation).lower() == "true"
-        user = get_current_user(request)
+        user = effective_user(request)
         endpoint_api_key = ""
         endpoint_base_url = ""
         _reject_raw_endpoint_url_for_non_admin(request, user, endpoint_id, endpoint_url)
@@ -477,7 +477,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
                 db.close()
         # Switch model/endpoint mid-session
         if model is not None and endpoint_url is not None:
-            user = get_current_user(request)
+            user = effective_user(request)
             _reject_raw_endpoint_url_for_non_admin(request, user, endpoint_id, endpoint_url)
             endpoint_api_key = ""
             endpoint_base_url = ""
@@ -1004,6 +1004,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
         """
         from src.llm_core import llm_call
         user = effective_user(request)
+        single_user_mode = not user and _auth_disabled()
         user_sessions = session_manager.get_sessions_for_user(user)
 
         # Delete empty and throwaway sessions before sorting
@@ -1022,7 +1023,12 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
         }
         _THROWAWAY_MAX_MESSAGES = 4  # only delete if <= this many messages
         try:
-            rows = db.query(DbSession).filter(DbSession.archived == False, DbSession.owner == user).limit(2000).all()
+            rows_q = db.query(DbSession).filter(DbSession.archived == False)
+            if user:
+                rows_q = rows_q.filter(DbSession.owner == user)
+            elif not single_user_mode:
+                rows_q = rows_q.filter(DbSession.owner == user)
+            rows = rows_q.limit(2000).all()
             folder_map = {r.id: r.folder for r in rows}
             # Precompute per-session message counts in TWO aggregate queries
             # instead of 1–3 queries PER session — with many chats the per-row
@@ -1242,7 +1248,12 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
         db = SessionLocal()
         try:
             for sid, folder_name in assignments.items():
-                db_session = db.query(DbSession).filter(DbSession.id == sid, DbSession.owner == user).first()
+                db_session_q = db.query(DbSession).filter(DbSession.id == sid)
+                if user:
+                    db_session_q = db_session_q.filter(DbSession.owner == user)
+                elif not single_user_mode:
+                    db_session_q = db_session_q.filter(DbSession.owner == user)
+                db_session = db_session_q.first()
                 if db_session:
                     db_session.folder = folder_name
                     db_session.updated_at = datetime.utcnow()

routes/shell_routes.py

@@ -15,6 +15,7 @@ from collections import namedtuple
 from pathlib import Path
 from typing import Dict, Any
 from core.platform_compat import IS_APPLE_SILICON, which_tool
+from core.middleware import INTERNAL_TOOL_USER
 from src.optional_deps import prepare_optional_dependency_import
 
 # POSIX-only: `pty`/`fcntl` transitively import `termios`, which does NOT exist
@@ -55,7 +56,7 @@ def _require_admin(request: Request):
     # In-process tool loopback. The AuthMiddleware already validated the
     # internal token + loopback client before setting this marker, so
     # honour it here as admin-equivalent.
-    if user == "internal-tool":
+    if user == INTERNAL_TOOL_USER:
         return
     if not user or user == "api":
         raise HTTPException(403, "Admin only")
@@ -330,6 +331,9 @@ def add_user_install_bins_to_path():
         candidates.append(os.path.join(site.USER_BASE, 'bin'))
     except Exception:
         pass
+    candidates.append(os.path.expanduser('~/bin'))
+    candidates.append(os.path.expanduser('~/llama.cpp/build/bin'))
+    candidates.append(os.path.expanduser('~/llama.cpp/build-vulkan/bin'))
     candidates.append(os.path.expanduser('~/.local/bin'))
     parts = os.environ.get('PATH', '').split(os.pathsep) if os.environ.get('PATH') else []
     changed = False
@@ -961,12 +965,84 @@ def setup_shell_routes() -> APIRouter:
 
         return StreamingResponse(generate(), media_type="text/event-stream")
 
+    def _os_id_from_release(text: str) -> str:
+        """Map /etc/os-release contents to a canonical family for our matrix."""
+        if not text:
+            return ""
+        ids = []
+        for line in text.splitlines():
+            line = line.strip()
+            if line.startswith("ID=") or line.startswith("ID_LIKE="):
+                ids += line.split("=", 1)[1].strip().strip('"').split()
+        ids = [i.lower() for i in ids]
+        if any(x in ids for x in ("debian", "ubuntu", "linuxmint", "pop", "elementary")):
+            return "debian"
+        if any(x in ids for x in ("arch", "manjaro", "endeavouros", "cachyos", "garuda")):
+            return "arch"
+        if any(x in ids for x in ("fedora", "rhel", "centos", "rocky", "almalinux", "ol")):
+            return "fedora"
+        if "alpine" in ids:
+            return "alpine"
+        if any(x in ids for x in ("suse", "opensuse", "opensuse-leap", "opensuse-tumbleweed", "sles")):
+            return "suse"
+        return ""
+
+    # Matrix lookup keyed on (os_family, backend) → (pkg_mgr_cmd_template, pkg_list_per_dep).
+    # Each `system_prereqs` name resolves to a list of OS-specific package
+    # names that get joined into the final `sudo apt install -y …` etc.
+    # command. Backend-specific extras (CUDA toolkit, ROCm, Vulkan headers)
+    # are added only when the detected backend needs them.
+    _PKG_NAMES = {
+        # canonical-name → {os_id: [actual_pkg_names_on_this_os]}
+        "cmake":           {"debian": ["cmake"], "arch": ["cmake"], "fedora": ["cmake"], "alpine": ["cmake"], "suse": ["cmake"], "macos": ["cmake"]},
+        "build-essential": {"debian": ["build-essential"], "arch": ["base-devel"], "fedora": ["gcc", "gcc-c++", "make"], "alpine": ["build-base"], "suse": ["gcc-c++", "make"], "macos": []},
+        "g++":             {"debian": ["g++"], "arch": ["gcc"], "fedora": ["gcc-c++"], "alpine": ["g++"], "suse": ["gcc-c++"], "macos": []},
+        "gcc":             {"debian": ["gcc"], "arch": ["gcc"], "fedora": ["gcc"], "alpine": ["gcc"], "suse": ["gcc"], "macos": []},
+        "make":            {"debian": ["make"], "arch": ["make"], "fedora": ["make"], "alpine": ["make"], "suse": ["make"], "macos": []},
+        "git":             {"debian": ["git"], "arch": ["git"], "fedora": ["git"], "alpine": ["git"], "suse": ["git"], "macos": ["git"]},
+        "tmux":            {"debian": ["tmux"], "arch": ["tmux"], "fedora": ["tmux"], "alpine": ["tmux"], "suse": ["tmux"], "macos": ["tmux"]},
+    }
+    _BACKEND_EXTRAS = {
+        "cuda":   {"debian": ["nvidia-cuda-toolkit"], "arch": ["cuda"], "fedora": ["cuda-toolkit"], "alpine": [], "suse": ["cuda"], "macos": []},
+        "rocm":   {"debian": ["rocm-dev"], "arch": ["rocm-hip-sdk"], "fedora": ["rocm-devel"], "alpine": [], "suse": ["rocm-dev"], "macos": []},
+        "vulkan": {"debian": ["libvulkan-dev", "vulkan-tools"], "arch": ["vulkan-headers", "vulkan-tools"], "fedora": ["vulkan-headers", "vulkan-tools"], "alpine": ["vulkan-loader-dev", "vulkan-tools"], "suse": ["vulkan-devel", "vulkan-tools"], "macos": []},
+    }
+    _PKG_MGR = {
+        "debian": "sudo apt install -y {pkgs}",
+        "arch":   "sudo pacman -S --needed {pkgs}",
+        "fedora": "sudo dnf install -y {pkgs}",
+        "alpine": "sudo apk add {pkgs}",
+        "suse":   "sudo zypper install -n {pkgs}",
+        "macos":  "brew install {pkgs}",
+    }
+
+    def _install_cmd_for_target(os_id: str, backend: str, missing: list[str]) -> str:
+        """Build a single OS+backend-aware install command for the missing prereqs."""
+        if not os_id or os_id not in _PKG_MGR:
+            return ""
+        pkgs: list[str] = []
+        seen: set[str] = set()
+        for m in missing:
+            for p in _PKG_NAMES.get(m, {}).get(os_id, []):
+                if p not in seen:
+                    pkgs.append(p); seen.add(p)
+        # Add backend-specific extras only when the build would actually
+        # consume them (a CUDA toolkit isn't useful on a Vulkan box).
+        backend = (backend or "").lower()
+        for p in _BACKEND_EXTRAS.get(backend, {}).get(os_id, []):
+            if p not in seen:
+                pkgs.append(p); seen.add(p)
+        if not pkgs:
+            return ""
+        return _PKG_MGR[os_id].format(pkgs=" ".join(pkgs))
+
     @router.get("/api/cookbook/packages")
     async def list_packages(
         request: Request,
         host: str | None = None,
         ssh_port: str | None = None,
         venv: str | None = None,
+        backend: str | None = None,
     ):
         """Check which optional packages are installed.
 
@@ -1015,6 +1091,12 @@ def setup_shell_routes() -> APIRouter:
                 "kind": "system",
                 "install_hint": "Install Docker on the selected server and allow this user to run docker.",
             },
+            # Note: cmake / gcc / git are not separate dependency rows —
+            # they're declared as `system_prereqs` on llama_cpp (and any
+            # other engine that compiles from source) so they appear as
+            # an inline status note on that engine's row instead of
+            # cluttering the panel with raw OS package names that aren't
+            # meaningful product-level dependencies on their own.
             # ── LLM ── installs on GPU servers for model serving/downloading
             {
                 "name": "hf_transfer",
@@ -1026,9 +1108,16 @@ def setup_shell_routes() -> APIRouter:
             {
                 "name": "llama_cpp",
                 "pip": "llama-cpp-python[server]",
-                "desc": "Serve GGUF models via llama.cpp",
+                "desc": "Great for single-GPU or CPU inference with GGUF models",
                 "category": "LLM",
                 "target": "remote",
+                # Build-toolchain prereqs. Cookbook's launch bootstrap
+                # compiles llama-server from source when no prebuilt
+                # binary is present; without these the build aborts
+                # with `cmake: command not found`. Surfaced inline on
+                # this row so the user doesn't have to chase three
+                # separate OS-package rows.
+                "system_prereqs": ["cmake", "g++", "git"],
             },
             {
                 "name": "sglang",
@@ -1040,7 +1129,7 @@ def setup_shell_routes() -> APIRouter:
             {
                 "name": "vllm",
                 "pip": "vllm",
-                "desc": "High-throughput LLM serving engine",
+                "desc": "Great for high-throughput multi-GPU inference",
                 "category": "LLM",
                 "target": "remote",
             },
@@ -1103,6 +1192,7 @@ def setup_shell_routes() -> APIRouter:
         # venv over SSH so a remote `pip install` actually reflects here.
         remote_status: dict = {}
         remote_details: dict = {}
+        remote_probe_error = ""
         remote_names = [
             p["name"]
             for p in packages
@@ -1141,16 +1231,56 @@ def setup_shell_routes() -> APIRouter:
                         break
             except ValueError as e:
                 raise HTTPException(400, str(e))
-            except Exception:
+            except Exception as e:
                 remote_status = {}
-        if host and remote_system_names:
+                remote_probe_error = f"SSH package probe failed: {str(e)[:160]}"
+            if "llama_cpp" in remote_names:
+                try:
+                    inner = (
+                        'export PATH="$HOME/.local/bin:$HOME/bin:'
+                        '$HOME/llama.cpp/build/bin:$HOME/llama.cpp/build-vulkan/bin:$PATH"; '
+                        "command -v llama-server 2>/dev/null || true"
+                    )
+                    argv = _ssh_base_argv(host, ssh_port) + [inner]
+                    proc = await asyncio.create_subprocess_exec(
+                        *argv,
+                        stdout=asyncio.subprocess.PIPE,
+                        stderr=asyncio.subprocess.PIPE,
+                    )
+                    out, _err = await asyncio.wait_for(proc.communicate(), timeout=8)
+                    llama_server_path = out.decode("utf-8", errors="replace").strip().splitlines()
+                    llama_server_path = llama_server_path[-1].strip() if llama_server_path else ""
+                    if llama_server_path:
+                        remote_status["llama_cpp"] = True
+                        probe = remote_details.setdefault("llama_cpp", {})
+                        if isinstance(probe, dict):
+                            probe.setdefault("binaries", {})["llama-server"] = llama_server_path
+                except Exception as e:
+                    if not remote_probe_error:
+                        remote_probe_error = f"SSH llama-server probe failed: {str(e)[:160]}"
+                    pass
+        # Union of system_names + every package's system_prereqs. Probing
+        # the prereqs alongside the main system deps in a single SSH call
+        # avoids a second round-trip per Cookbook → Dependencies refresh.
+        prereq_names: set[str] = set()
+        for p in packages:
+            for pr in p.get("system_prereqs") or []:
+                prereq_names.add(str(pr))
+        all_system_names = list(set(remote_system_names) | prereq_names)
+        # Detect the target's OS family + read /etc/os-release in the same
+        # SSH round-trip as the prereq probe — used downstream to render a
+        # single OS-specific install command per row instead of dumping
+        # every distro's syntax onto the user.
+        target_os_id: str = ""
+        if host and all_system_names:
             try:
                 checks = []
-                for name in remote_system_names:
+                for name in all_system_names:
                     qn = shlex.quote(name)
                     checks.append(
                         f"if command -v {qn} >/dev/null 2>&1; then echo {qn}=1; else echo {qn}=0; fi"
                     )
+                checks.append("echo '---OSREL---'; cat /etc/os-release 2>/dev/null || true")
                 inner = " ; ".join(checks)
                 argv = _ssh_base_argv(host, ssh_port) + [inner]
                 proc = await asyncio.create_subprocess_exec(
@@ -1160,20 +1290,45 @@ def setup_shell_routes() -> APIRouter:
                 )
                 out, _err = await asyncio.wait_for(proc.communicate(), timeout=12)
                 txt = out.decode("utf-8", errors="replace").strip()
+                _section, _osrel_lines = "probe", []
                 for line in txt.splitlines():
+                    if line.strip() == "---OSREL---":
+                        _section = "osrel"; continue
+                    if _section == "osrel":
+                        _osrel_lines.append(line)
+                        continue
                     name, sep, value = line.strip().partition("=")
-                    if sep and name in remote_system_names:
+                    if sep and name in all_system_names:
                         remote_status[name] = value == "1"
+                target_os_id = _os_id_from_release("\n".join(_osrel_lines))
             except ValueError as e:
                 raise HTTPException(400, str(e))
-            except Exception:
+            except Exception as e:
+                if not remote_probe_error:
+                    remote_probe_error = f"SSH system probe failed: {str(e)[:160]}"
                 pass
+        elif not host:
+            # Local target — probe in-process so the inline install command
+            # still appears in the dep panel when the cookbook container
+            # itself is the selected server.
+            try:
+                with open("/etc/os-release", encoding="utf-8") as f:
+                    target_os_id = _os_id_from_release(f.read())
+            except Exception:
+                target_os_id = ""
+            if sys.platform == "darwin":
+                target_os_id = "macos"
 
         for pkg in packages:
             on_remote = bool(host and pkg.get("target") == "remote")
             probe = None
             if on_remote:
-                pkg["installed"] = bool(remote_status.get(pkg["name"], False))
+                if remote_probe_error and pkg["name"] not in remote_status:
+                    pkg["installed"] = None
+                    pkg["probe_error"] = remote_probe_error
+                    pkg["status_note"] = remote_probe_error
+                else:
+                    pkg["installed"] = bool(remote_status.get(pkg["name"], False))
                 probe = remote_details.get(pkg["name"])
                 if isinstance(probe, dict):
                     pkg["details"] = probe
@@ -1229,6 +1384,104 @@ def setup_shell_routes() -> APIRouter:
                     # 500 the entire packages panel; report it as not usable.
                     pkg["installed"] = False
 
+            # llama_cpp partial-state probe: when the package is installed
+            # but the wheel was built CPU-only AND the target has NVIDIA
+            # hardware, mark the row as partial (yellow/orange) with a
+            # one-click upgrade to the CUDA wheel. Without this the row
+            # reads "ready" green while inference runs at 3 tok/s on GPU
+            # silicon — actively misleading.
+            if pkg["name"] == "llama_cpp" and pkg.get("installed"):
+                _native_llama_server = bool(
+                    isinstance(probe, dict)
+                    and isinstance(probe.get("binaries"), dict)
+                    and probe["binaries"].get("llama-server")
+                )
+                _gpu_capable = False
+                _has_nvidia_target = False
+                if _native_llama_server:
+                    # Native llama-server is the launcher path Cookbook now
+                    # prefers. Do not mark this as a CPU-only Python wheel just
+                    # because llama-cpp-python is absent from the selected venv.
+                    _gpu_capable = True
+                elif on_remote and host:
+                    try:
+                        # Activate the configured venv FIRST so the probe
+                        # runs against the same python the launch script
+                        # would activate. Without this prefix, bare
+                        # `python3` was checked — which can disagree with
+                        # the venv's wheel (e.g. user-site has CUDA wheel
+                        # but venv has CPU-only), and the dep panel then
+                        # showed "ready" green while every launch fell to
+                        # CPU.
+                        _vp = _venv_activate_prefix(venv)
+                        probe = (
+                            f'{_vp}python3 -c "import llama_cpp; import sys; '
+                            'sys.exit(0 if llama_cpp.llama_supports_gpu_offload() else 1)" '
+                            '&& echo llama_cpp_gpu=1 || echo llama_cpp_gpu=0; '
+                            'command -v nvidia-smi >/dev/null 2>&1 '
+                            '&& nvidia-smi -L 2>/dev/null | grep -q "GPU " '
+                            '&& echo nvidia=1 || echo nvidia=0'
+                        )
+                        argv = _ssh_base_argv(host, ssh_port) + [probe]
+                        proc = await asyncio.create_subprocess_exec(
+                            *argv, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE,
+                        )
+                        out, _ = await asyncio.wait_for(proc.communicate(), timeout=8)
+                        txt = out.decode("utf-8", errors="replace")
+                        if "llama_cpp_gpu=1" in txt:
+                            _gpu_capable = True
+                        if "nvidia=1" in txt:
+                            _has_nvidia_target = True
+                    except Exception:
+                        pass
+                else:
+                    try:
+                        import llama_cpp as _lcp  # type: ignore
+                        _gpu_capable = bool(_lcp.llama_supports_gpu_offload())
+                    except Exception:
+                        _gpu_capable = False
+                    _has_nvidia_target = shutil.which("nvidia-smi") is not None
+                if (not _gpu_capable) and _has_nvidia_target:
+                    pkg["partial"] = True
+                    pkg["partial_reason"] = "Installed but CPU-only wheel — GPU detected on this target. Upgrade to a CUDA wheel for ~10× faster inference."
+                    pkg["partial_action"] = "reinstall_llama_cpp_cuda"
+            # Attach per-package system_prereqs status. We probed each
+            # prereq name above; surface "Missing build deps: …" ONLY
+            # when the package itself is not installed — if the package
+            # works (e.g. llama-cpp-python already imports cleanly), the
+            # build toolchain is irrelevant and surfacing it as a red
+            # flag confuses users ("ready" + "missing" on the same row).
+            _prereqs = list(pkg.get("system_prereqs") or [])
+            if _prereqs:
+                if on_remote:
+                    _pr_present = {n: bool(remote_status.get(n)) for n in _prereqs}
+                else:
+                    _pr_present = {n: shutil.which(n) is not None for n in _prereqs}
+                pkg["system_prereqs_status"] = _pr_present
+                _missing = [n for n, ok in _pr_present.items() if not ok]
+                # Suppress the "missing build deps" hint when the package
+                # itself is installed — build deps are only relevant if
+                # the user would need to recompile from source.
+                if pkg.get("installed"):
+                    _missing = []
+                if _missing:
+                    # Build a target-specific install command from the
+                    # (os_family, backend) matrix when we know both. Fall
+                    # back to the multi-distro hint only when the target's
+                    # OS can't be classified (e.g. ssh probe failed).
+                    _resolved_os = target_os_id or "debian"  # safest default
+                    _cmd = _install_cmd_for_target(_resolved_os, backend or "", _missing)
+                    if _cmd and target_os_id:
+                        _hint = "Missing build deps for this target: " + ", ".join(_missing)
+                        pkg["install_cmd_for_target"] = _cmd
+                        pkg["install_cmd_os"] = target_os_id
+                        pkg["install_cmd_backend"] = (backend or "").lower()
+                    else:
+                        _hint = "Missing build deps: " + ", ".join(_missing) + ". Install via apt: cmake build-essential git / pacman: cmake base-devel git / dnf: cmake gcc-c++ make git / brew: cmake git."
+                    _existing_note = pkg.get("status_note") or ""
+                    pkg["status_note"] = (_existing_note + " — " + _hint) if _existing_note else _hint
+                    pkg["build_deps_missing"] = _missing
+
             if pkg.get("installed"):
                 update_status = _package_pip_update_status(pkg, probe)
                 pkg["pip_update_available"] = update_status.available
@@ -1288,6 +1541,102 @@ def setup_shell_routes() -> APIRouter:
             return {"ok": True, "output": stdout.decode()[-200:]}
         return {"ok": False, "error": stderr.decode()[-300:]}
 
+    @router.post("/api/cookbook/install-system-deps")
+    async def install_system_deps(request: Request):
+        """Install OS-level system packages (cmake/build-essential/git/tmux)
+        on a remote target or in the local container. Admin only.
+
+        Bounded by a per-package allowlist — anything outside the catalog
+        is rejected so the route can't be coerced into installing arbitrary
+        OS packages. Uses `sudo -n` (passwordless) so the call returns a
+        clear "needs sudo password" error instead of hanging when interactive
+        sudo is required.
+        """
+        _require_admin(request)
+        body = await request.json()
+        raw = body.get("packages") or []
+        host = (body.get("remote_host") or "").strip()
+        ssh_port = body.get("ssh_port")
+        # Names users can request — must match canonical names used in the
+        # deps catalog's `system_prereqs` field and on the System rows.
+        ALLOWED = {"cmake", "build-essential", "g++", "gcc", "git", "tmux", "make"}
+        pkgs = [str(p).strip() for p in raw if str(p).strip() in ALLOWED]
+        if not pkgs:
+            return {"ok": False, "error": "no installable packages requested (allowlist: " + ", ".join(sorted(ALLOWED)) + ")"}
+        # Re-map to the right package name per OS. apt/dpkg use the names
+        # as-is; pacman has base-devel for build-essential, etc.
+        def _apt(names): return list(names)
+        def _pacman(names):
+            return ["base-devel" if n == "build-essential" else n for n in names]
+        def _dnf(names):
+            out = []
+            for n in names:
+                if n == "build-essential": out += ["gcc", "gcc-c++", "make"]
+                elif n == "g++": out += ["gcc-c++"]
+                else: out.append(n)
+            return out
+        def _brew(names):
+            return [n for n in names if n not in ("build-essential", "g++", "gcc", "make")]
+        # Build a single shell snippet that detects the package manager and
+        # runs the right install. Non-interactive sudo (-n) only — if sudo
+        # asks for a password the script reports it instead of hanging.
+        apt_pkgs = " ".join(shlex.quote(p) for p in _apt(pkgs))
+        pac_pkgs = " ".join(shlex.quote(p) for p in _pacman(pkgs))
+        dnf_pkgs = " ".join(shlex.quote(p) for p in _dnf(pkgs))
+        brew_pkgs = " ".join(shlex.quote(p) for p in _brew(pkgs))
+        # Error messages go to stderr (>&2) so the route's error field
+        # gets populated. Without the redirect, `echo "ERROR…"` on stdout
+        # left stderr empty and the frontend toast fell through to a
+        # bare "HTTP 200" instead of surfacing the real reason.
+        script = (
+            'set -e; '
+            'if ! sudo -n true 2>/dev/null; then '
+            '  echo "ERROR: passwordless sudo unavailable on this target. Run once: sudo apt install -y ' + " ".join(pkgs) + ' (or your distro equivalent: pacman -S, dnf install, brew install). After that, Cookbook can install the rest." >&2; exit 2; fi; '
+            'if command -v apt-get >/dev/null 2>&1; then '
+            f'  sudo -n env DEBIAN_FRONTEND=noninteractive apt-get update -qq && sudo -n env DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends {apt_pkgs}; '
+            'elif command -v pacman >/dev/null 2>&1; then '
+            f'  sudo -n pacman -Sy --needed --noconfirm {pac_pkgs}; '
+            'elif command -v dnf >/dev/null 2>&1; then '
+            f'  sudo -n dnf install -y {dnf_pkgs}; '
+            'elif command -v brew >/dev/null 2>&1; then '
+            f'  brew install {brew_pkgs}; '
+            'else '
+            '  echo "ERROR: no supported package manager (apt/pacman/dnf/brew) on this target." >&2; exit 3; fi'
+        )
+        try:
+            if host:
+                argv = _ssh_base_argv(host, ssh_port) + [script]
+            else:
+                argv = ["bash", "-lc", script]
+        except ValueError as e:
+            raise HTTPException(400, str(e))
+        try:
+            proc = await asyncio.create_subprocess_exec(
+                *argv, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE
+            )
+            out, err = await asyncio.wait_for(proc.communicate(), timeout=180)
+        except asyncio.TimeoutError:
+            return {"ok": False, "error": "Install timed out after 180s"}
+        ok = (proc.returncode == 0)
+        # Combine stderr + (last lines of stdout) into a single error
+        # blob when ok=False — some package managers print useful failure
+        # context to stdout, and a script that exits via `echo ...; exit N`
+        # without `>&2` would otherwise hand back an empty error string
+        # and force the frontend to show a bare "HTTP 200".
+        err_txt = err.decode("utf-8", errors="replace").strip()
+        out_txt = out.decode("utf-8", errors="replace").strip()
+        if not ok:
+            tail_out = out_txt[-500:] if out_txt else ""
+            combined = err_txt or tail_out or f"exit code {proc.returncode}"
+        else:
+            combined = None
+        return {
+            "ok": ok,
+            "exit_code": proc.returncode,
+            "output": out_txt[-1000:],
+            "error": combined,
+        }
+
     @router.post("/api/cookbook/rebuild-engine")
     async def rebuild_engine(request: Request):
         """Clear the cached llama.cpp build so the next serve recompiles.
@@ -1308,7 +1657,8 @@ def setup_shell_routes() -> APIRouter:
             return {"ok": False, "error": f"Unsupported engine: {engine}"}
         host = str(body.get("remote_host") or "").strip()
         ssh_port = body.get("ssh_port")
-        cmd = _llama_cpp_rebuild_cmd()
+        update_source = bool(body.get("update_source"))
+        cmd = _llama_cpp_rebuild_cmd(update_source=update_source)
         try:
             argv = (
                 (_ssh_base_argv(host, ssh_port) + [cmd])

routes/task_routes.py

@@ -11,6 +11,7 @@ from fastapi import APIRouter, HTTPException, Request
 from pydantic import BaseModel
 
 from core.database import SessionLocal, ScheduledTask, TaskRun
+from core.middleware import INTERNAL_TOOL_USER
 from core.constants import internal_api_base
 from src.auth_helpers import get_current_user
 from src.constants import DATA_DIR, EMAIL_URGENCY_CACHE_DIR
@@ -427,7 +428,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
         # In-process tool-loopback marker — AuthMiddleware validated
         # the internal token + loopback client before stamping this,
         # so treat as admin-equivalent.
-        if user == "internal-tool":
+        if user == INTERNAL_TOOL_USER:
             return True
         try:
             from core.auth import AuthManager

routes/upload_routes.py

@@ -3,11 +3,16 @@ import os
 import time
 import json
 import asyncio
+import shutil
+import uuid
+from pathlib import Path
 from fastapi import APIRouter, Request, File, UploadFile, HTTPException
 from typing import List
 import logging
 from core.middleware import require_admin
-from src.auth_helpers import get_current_user
+from core.database import SessionLocal, GalleryImage
+from src.auth_helpers import effective_user
+from src.constants import GENERATED_IMAGES_DIR
 from src.upload_handler import count_recent_uploads
 
 logger = logging.getLogger(__name__)
@@ -50,6 +55,69 @@ def setup_upload_routes(upload_handler):
             raise HTTPException(404, "File not found")
 
         raise HTTPException(404, "File not found")
+
+    def _promote_chat_image_to_gallery(meta: dict, owner: str | None) -> str | None:
+        """Make chat-uploaded images visible in Gallery without changing chat storage."""
+        is_image_file = getattr(upload_handler, "is_image_file", None)
+        if not callable(is_image_file):
+            return None
+        if not is_image_file(meta.get("name", ""), meta.get("mime", "")):
+            return None
+
+        source_path = meta.get("path")
+        if not source_path or not os.path.isfile(source_path):
+            return None
+
+        db = SessionLocal()
+        try:
+            file_hash = meta.get("hash")
+            if file_hash:
+                q = db.query(GalleryImage).filter(
+                    GalleryImage.file_hash == file_hash,
+                    GalleryImage.is_active == True,  # noqa: E712
+                )
+                if owner:
+                    q = q.filter(GalleryImage.owner == owner)
+                existing = q.first()
+                if existing:
+                    return existing.id
+
+            image_dir = Path(GENERATED_IMAGES_DIR)
+            image_dir.mkdir(parents=True, exist_ok=True)
+            ext = Path(meta.get("name") or source_path).suffix.lower()
+            if ext not in {".png", ".jpg", ".jpeg", ".webp", ".gif"}:
+                mime_ext = {
+                    "image/png": ".png",
+                    "image/jpeg": ".jpg",
+                    "image/jpg": ".jpg",
+                    "image/webp": ".webp",
+                    "image/gif": ".gif",
+                }.get(meta.get("mime", ""))
+                ext = mime_ext or ".png"
+            filename = f"{uuid.uuid4().hex[:12]}{ext}"
+            dest_path = image_dir / filename
+            shutil.copy2(source_path, dest_path)
+
+            image_id = str(uuid.uuid4())
+            db.add(GalleryImage(
+                id=image_id,
+                filename=filename,
+                prompt=meta.get("name") or "Chat upload",
+                model="chat-upload",
+                owner=owner,
+                file_hash=file_hash,
+                width=meta.get("width"),
+                height=meta.get("height"),
+                file_size=meta.get("size"),
+            ))
+            db.commit()
+            return image_id
+        except Exception as e:
+            db.rollback()
+            logger.warning("Failed to add chat image upload to gallery: %s", e)
+            return None
+        finally:
+            db.close()
     
     @router.post("")
     async def api_upload(request: Request, files: List[UploadFile] = File(...)):
@@ -78,8 +146,10 @@ def setup_upload_routes(upload_handler):
         
         for u in files:
             try:
-                meta = upload_handler.save_upload(u, client_ip, owner=get_current_user(request))
-                out.append({
+                owner = effective_user(request)
+                meta = upload_handler.save_upload(u, client_ip, owner=owner)
+                gallery_id = _promote_chat_image_to_gallery(meta, owner)
+                item = {
                     "id": meta["id"],
                     "name": meta["name"],
                     "mime": meta["mime"],
@@ -89,7 +159,10 @@ def setup_upload_routes(upload_handler):
                     "width": meta.get("width"),
                     "height": meta.get("height"),
                     "is_duplicate": meta.get("is_duplicate", False)
-                })
+                }
+                if gallery_id:
+                    item["gallery_id"] = gallery_id
+                out.append(item)
             except HTTPException:
                 raise
             except Exception as e:
@@ -138,7 +211,7 @@ def setup_upload_routes(upload_handler):
                 original_name = info.get("name", file_id)
         auth_mgr = getattr(request.app.state, "auth_manager", None)
         auth_configured = bool(auth_mgr and auth_mgr.is_configured)
-        current_user = get_current_user(request)
+        current_user = effective_user(request)
         file_owner = info.get("owner") if info else None
         if auth_configured:
             if not current_user:
@@ -204,7 +277,7 @@ def setup_upload_routes(upload_handler):
         info = _load_upload_info(file_id)
         auth_mgr = getattr(request.app.state, "auth_manager", None)
         auth_configured = bool(auth_mgr and auth_mgr.is_configured)
-        current_user = get_current_user(request)
+        current_user = effective_user(request)
         file_owner = info.get("owner") if info else None
         if auth_configured:
             if not current_user:
@@ -247,7 +320,7 @@ def setup_upload_routes(upload_handler):
             raise HTTPException(404, "File not found")
         auth_mgr = getattr(request.app.state, "auth_manager", None)
         auth_configured = bool(auth_mgr and auth_mgr.is_configured)
-        current_user = get_current_user(request)
+        current_user = effective_user(request)
         file_owner = info.get("owner")
         if auth_configured:
             if not current_user:

routes/webhook_routes.py

@@ -1,6 +1,5 @@
 """Webhook, API Token, and sync chat routes."""
 
-import asyncio
 import uuid
 import logging
 from typing import Optional
@@ -385,10 +384,10 @@ def setup_webhook_routes(
         sess.add_message(ChatMessage("assistant", reply))
         session_manager.save_sessions()
 
-        asyncio.create_task(webhook_manager.fire("chat.completed", {
+        webhook_manager.fire_and_forget("chat.completed", {
             "session_id": session_id, "model": sess.model,
             "user_message": message[:2000], "response": reply[:2000],
-        }))
+        })
 
         return {"response": reply, "session_id": session_id, "model": sess.model}
 

scripts/odysseus-calendar

@@ -103,9 +103,13 @@ def cmd_list(args) -> None:
     end = _parse_dt(args.end) if args.end else (start + timedelta(days=30))
     db = SessionLocal()
     try:
+        # Overlap semantics, matching the web route (routes/calendar_routes.py)
+        # and the recurring-expansion contract: an event is in the window when
+        # it starts before the window end AND ends after the window start. This
+        # includes multi-day / in-progress events that began before `start`.
         q = db.query(CalendarEvent).filter(
-            CalendarEvent.dtstart >= start,
             CalendarEvent.dtstart < end,
+            CalendarEvent.dtend > start,
         )
         if args.calendar:
             cal = db.query(CalendarCal).filter(CalendarCal.name == args.calendar).first()

services/hwfit/fit.py

@@ -19,6 +19,10 @@ GPU_BANDWIDTH = {
     "6950 xt": 576, "6900 xt": 512, "6800 xt": 512, "6800": 512, "6700 xt": 384, "6600 xt": 256, "6600": 224,
     "mi300x": 5300, "mi300": 5300, "mi250x": 3277, "mi250": 3277, "mi210": 1638, "mi100": 1229,
     "9070 xt": 624, "9070": 488, "9060 xt": 322, "9060": 322,
+    # NVIDIA GB10 Grace-Blackwell superchip (DGX Spark). Unified LPDDR5X memory,
+    # not Apple Silicon, so it lives in the generic GPU table — the Apple-only
+    # lookup never matches it (its name carries no "apple").
+    "gb10": 273,
 }
 
 # Pre-sort keys by length descending for correct substring matching
@@ -126,6 +130,44 @@ def _lookup_bandwidth(system):
     return None
 
 
+def _canonical_cpu_backend(system):
+    """Return the canonical CPU backend for cpu_only speed estimation.
+
+    Normalizes CPU-architecture aliases separately from the GPU backend, and
+    overrides GPU-only backends (CUDA/ROCm/Metal) so they do not inherit a
+    discrete-GPU fallback constant when the model is actually running on CPU.
+    """
+    backend = (system.get("backend") or "").lower().strip()
+    cpu_arch = (system.get("cpu_arch") or "").lower().strip()
+    cpu_name = (system.get("cpu_name") or "").lower()
+    gpu_name = (system.get("gpu_name") or "").lower()
+
+    # Already-canonical CPU backends
+    if backend in ("cpu_x86", "cpu_arm"):
+        return backend
+
+    # Raw CPU-architecture aliases. Treat plain "arm" as 32-bit ARM, not the
+    # ARM64-class CPU fallback used for Apple Silicon/aarch64 machines.
+    if backend in ("x86_64", "amd64", "i386", "i686"):
+        return "cpu_x86"
+    if backend in ("arm64", "aarch64"):
+        return "cpu_arm"
+
+    # Prefer an explicit CPU architecture field when present
+    if cpu_arch:
+        if cpu_arch in ("x86_64", "amd64", "x86", "i386", "i686"):
+            return "cpu_x86"
+        if cpu_arch in ("arm64", "aarch64"):
+            return "cpu_arm"
+
+    # Apple Silicon enters ranking as backend="metal"; its CPU path is ARM.
+    if backend in ("metal", "mps", "apple") or "apple" in cpu_name or "apple" in gpu_name:
+        return "cpu_arm"
+
+    # Conservative default for CUDA/ROCm/discrete GPU backends and unknowns.
+    return "cpu_x86"
+
+
 def _estimate_speed(model, quant, run_mode, system, offload_frac=0.0):
     """Estimate tok/s. Uses active params for MoE (only active experts run per token).
 
@@ -143,6 +185,11 @@ def _estimate_speed(model, quant, run_mode, system, offload_frac=0.0):
     bw = _lookup_bandwidth(system)
     backend = system.get("backend", "cpu_x86")
 
+    # CPU-only inference must never inherit a GPU backend's fallback constant,
+    # even if the detected system happens to report a CUDA/Metal/ROCm backend.
+    if run_mode == "cpu_only":
+        backend = _canonical_cpu_backend(system)
+
     if bw and run_mode in ("gpu", "cpu_offload"):
         bpp = QUANT_BYTES_PER_PARAM.get(quant, 0.5)
         model_gb = pb * bpp

services/hwfit/hardware.py

@@ -282,7 +282,17 @@ def _detect_amd():
             "gpus": cards,
             "gpu_groups": groups,
             "homogeneous": len(groups) <= 1,
-            "backend": "rocm",
+            # Pick the actual runtime label: ROCm/HIP only when its
+            # toolchain is installed, otherwise Vulkan if vulkaninfo is
+            # present (mesa RADV works fine on RDNA/CDNA when ROCm
+            # packages are absent — see Strix Halo where ROCm support
+            # is still backporting). Reporting "rocm" on a Vulkan-only
+            # host misleads downstream env-var pinning
+            # (HIP_VISIBLE_DEVICES is a no-op there).
+            "backend": (
+                "rocm" if (_run(["which", "rocminfo"]) or _run(["which", "hipconfig"]))
+                else ("vulkan" if _run(["which", "vulkaninfo"]) else "rocm")
+            ),
             "unified_memory": is_apu,
             # AMD ISA/family so downstream can tell datacenter Instinct (CDNA,
             # where vLLM/SGLang run AWQ/GPTQ reliably) from consumer Radeon
@@ -320,7 +330,7 @@ def _detect_apple_silicon():
 
     # Only Apple Silicon (arm64) has a Metal GPU worth serving LLMs on; Intel
     # Macs fall through to the CPU path.
-    if "arm" not in arch and "aarch64" not in arch:
+    if _canonical_cpu_arch(arch) != "arm64":
         return None
 
     # Chip name, e.g. "Apple M4 Max" — carries the Pro/Max/Ultra variant that
@@ -503,6 +513,25 @@ def _get_cpu_count():
     return os.cpu_count() or 1
 
 
+def _canonical_cpu_arch(value):
+    arch = str(value or "").lower().strip().replace("-", "_")
+    if arch in ("x86_64", "amd64", "x64"):
+        return "x86_64"
+    if arch in ("i386", "i686", "x86"):
+        return "x86"
+    if arch in ("arm64", "aarch64"):
+        return "arm64"
+    if arch == "arm" or arch.startswith("armv"):
+        return "arm"
+    return arch
+
+
+def _get_cpu_arch():
+    if _remote_host:
+        return _canonical_cpu_arch(_run(["uname", "-m"]) or "")
+    return _canonical_cpu_arch(platform.machine())
+
+
 def _powershell_exe():
     """Pick the best PowerShell executable for LOCAL execution: prefer pwsh
     (PowerShell 7+), fall back to Windows PowerShell 5.1. Returns an absolute
@@ -528,6 +557,7 @@ def _detect_windows():
         $r.cpu_name = $cpu.Name
         $r.cpu_cores = (Get-CimInstance Win32_Processor | Measure-Object -Property NumberOfLogicalProcessors -Sum).Sum
         $r.arch = $cpu.AddressWidth
+        $r.cpu_arch = if ($env:PROCESSOR_ARCHITEW6432) { $env:PROCESSOR_ARCHITEW6432 } else { $env:PROCESSOR_ARCHITECTURE }
         # GPU detection via nvidia-smi (fastest) or WMI fallback
         try { 
             $nv = nvidia-smi --query-gpu=memory.total,name --format=csv,noheader,nounits 2>$null
@@ -599,6 +629,7 @@ def _detect_windows():
             "available_ram_gb": d.get("avail_gb", 0),
             "cpu_cores": _as_int(d.get("cpu_cores"), 1),
             "cpu_name": _cpu_name,
+            "cpu_arch": _canonical_cpu_arch(d.get("cpu_arch")),
             "has_gpu": bool(d.get("gpu_name")),
             "gpu_name": d.get("gpu_name"),
             "gpu_vram_gb": d.get("gpu_vram_gb"),
@@ -794,6 +825,7 @@ def detect_system(host="", ssh_port="", platform="", fresh=False):
     available_ram = round(_get_available_ram_gb(), 1)
     cpu_cores = _get_cpu_count()
     cpu_name = _get_cpu_name()
+    cpu_arch = _get_cpu_arch()
 
     gpu_info = _detect_apple_silicon() or _detect_nvidia() or _detect_amd()
 
@@ -803,6 +835,7 @@ def detect_system(host="", ssh_port="", platform="", fresh=False):
             "available_ram_gb": available_ram,
             "cpu_cores": cpu_cores,
             "cpu_name": cpu_name,
+            "cpu_arch": cpu_arch,
             "has_gpu": True,
             "gpu_name": gpu_info["gpu_name"],
             "gpu_vram_gb": gpu_info["gpu_vram_gb"],
@@ -817,17 +850,13 @@ def detect_system(host="", ssh_port="", platform="", fresh=False):
             "unified_memory": gpu_info.get("unified_memory", False),
         }
     else:
-        if _remote_host:
-            arch_out = _run(["uname", "-m"]) or ""
-        else:
-            import platform as _platform
-            arch_out = _platform.machine().lower()
-        backend = "cpu_arm" if "aarch64" in arch_out or "arm" in arch_out else "cpu_x86"
+        backend = "cpu_arm" if cpu_arch == "arm64" else "cpu_x86"
         result = {
             "total_ram_gb": total_ram,
             "available_ram_gb": available_ram,
             "cpu_cores": cpu_cores,
             "cpu_name": cpu_name,
+            "cpu_arch": cpu_arch,
             "has_gpu": False,
             "gpu_name": None,
             "gpu_vram_gb": None,

services/search/content.py

@@ -15,6 +15,8 @@ from urllib.parse import urljoin, urlparse
 import httpx
 from bs4 import BeautifulSoup
 
+from src.constants import WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES, WEB_FETCH_USER_AGENT
+
 from .analytics import RateLimitError, error_logger
 from .cache import (
     CONTENT_CACHE_DIR,
@@ -89,18 +91,128 @@ def _public_http_url(url: str) -> bool:
         return False
 
 
-def _get_public_url(url: str, headers: dict, timeout: int, max_redirects: int = 5) -> httpx.Response:
+class BodyTooLargeError(Exception):
+    """The server declared a body larger than the hard fetch ceiling."""
+
+    def __init__(self, url: str, declared_bytes: int):
+        self.url = url
+        self.declared_bytes = declared_bytes
+        super().__init__(
+            f"response body is {declared_bytes:,} bytes, over the "
+            f"{WEB_FETCH_HARD_MAX_BYTES:,}-byte hard cap"
+        )
+
+
+class _CappedFetch:
+    """Result of a size-capped streaming GET.
+
+    Carries just what fetch_webpage_content needs from an httpx.Response,
+    plus the cap bookkeeping: the (possibly truncated) body, whether the
+    cap cut it short, and the size the server declared via Content-Length
+    (wire bytes; None when absent).
+    """
+
+    __slots__ = ("status_code", "headers", "content", "truncated",
+                 "declared_bytes", "encoding", "url")
+
+    def __init__(self, status_code, headers, content, truncated,
+                 declared_bytes, encoding, url):
+        self.status_code = status_code
+        self.headers = headers
+        self.content = content
+        self.truncated = truncated
+        self.declared_bytes = declared_bytes
+        self.encoding = encoding
+        self.url = url
+
+    @property
+    def text(self) -> str:
+        return self.content.decode(self.encoding or "utf-8", errors="replace")
+
+    def raise_for_status(self):
+        if self.status_code >= 400:
+            request = httpx.Request("GET", self.url)
+            raise httpx.HTTPStatusError(
+                f"HTTP {self.status_code} for {self.url}",
+                request=request,
+                response=httpx.Response(self.status_code, request=request),
+            )
+
+
+def _get_public_url(url: str, headers: dict, timeout: int, max_redirects: int = 5,
+                    max_bytes: int = None) -> "_CappedFetch":
+    """Capped streaming GET with SSRF-guarded manual redirects.
+
+    The body is streamed and buffering stops at ``max_bytes`` (default: the
+    soft cap), so an oversized resource cannot be pulled into memory or the
+    content cache in full. When Content-Length already declares a body over
+    the hard ceiling, the fetch is refused before any body bytes are read.
+    """
+    cap = min(max_bytes or WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES)
     current = url
     for _ in range(max_redirects + 1):
         if not _public_http_url(current):
             raise httpx.RequestError("Blocked private/internal URL", request=httpx.Request("GET", current))
-        response = httpx.get(current, headers=headers, timeout=timeout, follow_redirects=False)
-        if response.status_code not in (301, 302, 303, 307, 308):
-            return response
-        location = response.headers.get("location")
-        if not location:
-            return response
-        current = urljoin(str(response.url), location)
+        # Force identity transfer-encoding. With gzip/deflate the wire bytes
+        # (and Content-Length) can be a small fraction of the decoded body, so
+        # a tiny compressed response could pass the hard-cap preflight and then
+        # expand past the ceiling in a single decoded chunk before the streamed
+        # cap below can slice it. Identity makes Content-Length the true body
+        # size and keeps each streamed chunk bounded by the network read.
+        req_headers = dict(headers or {})
+        req_headers["Accept-Encoding"] = "identity"
+        with httpx.stream("GET", current, headers=req_headers, timeout=timeout,
+                          follow_redirects=False) as response:
+            if response.status_code in (301, 302, 303, 307, 308):
+                location = response.headers.get("location")
+                if not location:
+                    return _CappedFetch(response.status_code, response.headers, b"",
+                                        False, None, response.encoding, str(response.url))
+                current = urljoin(str(response.url), location)
+                continue
+
+            # A server can ignore the identity request and still return a
+            # compressed body; httpx.iter_bytes would then decode it, and a tiny
+            # gzip can balloon into one decoded chunk far past the cap before we
+            # slice. Refuse a compressed Content-Encoding so the streamed cap
+            # stays a real memory bound (Content-Length is the compressed wire
+            # length here, so the preflight and size metadata are unreliable too).
+            enc = (response.headers.get("content-encoding") or "").strip().lower()
+            if enc and enc != "identity":
+                raise httpx.RequestError(
+                    f"Refusing compressed response (Content-Encoding: {enc}) after "
+                    "requesting identity: cannot bound decoded body size",
+                    request=httpx.Request("GET", current),
+                )
+
+            declared = None
+            raw_len = response.headers.get("content-length")
+            if raw_len and raw_len.isdigit():
+                declared = int(raw_len)
+            # Refuse before buffering anything when the server already tells
+            # us the body exceeds the absolute ceiling (Content-Length is wire
+            # bytes; the decompressed body can only be larger).
+            if declared is not None and declared > WEB_FETCH_HARD_MAX_BYTES:
+                raise BodyTooLargeError(current, declared)
+
+            chunks = []
+            read = 0
+            truncated = False
+            # We requested identity above, so iter_bytes yields the raw body in
+            # network-read-sized chunks (no decompression expansion); the cap
+            # therefore bounds what we actually buffer.
+            for chunk in response.iter_bytes():
+                read += len(chunk)
+                if read > cap:
+                    keep = cap - (read - len(chunk))
+                    if keep > 0:
+                        chunks.append(chunk[:keep])
+                    truncated = True
+                    break
+                chunks.append(chunk)
+            return _CappedFetch(response.status_code, response.headers,
+                                b"".join(chunks), truncated, declared,
+                                response.encoding, str(response.url))
     raise httpx.RequestError("Too many redirects", request=httpx.Request("GET", current))
 
 # PDF extraction (optional dependency)
@@ -222,9 +334,19 @@ def _empty_result(url: str, error: str = "") -> dict:
 # ----------------------------------------------------------------------
 # Main content fetcher
 # ----------------------------------------------------------------------
-def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) -> dict:
-    """Fetch and extract meaningful content from a webpage with caching."""
-    cache_key = generate_cache_key(url)
+def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0,
+                          max_bytes: int = None) -> dict:
+    """Fetch and extract meaningful content from a webpage with caching.
+
+    ``max_bytes`` raises the download budget per call (clamped to the hard
+    cap); the default is the soft cap. When the body is cut short the result
+    carries ``truncated``/``fetched_bytes``/``total_bytes`` so callers can
+    tell the model the content is partial (#3812).
+    """
+    effective_cap = min(max_bytes or WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES)
+    # The cap is part of the cache identity: a truncated soft-cap fetch must
+    # not be served to a later full-budget request for the same URL.
+    cache_key = generate_cache_key(f"{url}#cap={effective_cap}")
     cache_file = CONTENT_CACHE_DIR / f"{cache_key}.cache"
 
     # Check cache
@@ -247,18 +369,24 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
     # Fetch
     try:
         headers = {
-            "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36",
+            "User-Agent": WEB_FETCH_USER_AGENT,
             "Accept": "text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8",
             "Accept-Language": "en-US,en;q=0.5",
-            "Accept-Encoding": "gzip, deflate",
+            # identity so the streamed size cap in _get_public_url stays honest
+            # (a compressed body can decode to far more than Content-Length).
+            "Accept-Encoding": "identity",
             "Connection": "keep-alive",
         }
-        response = _get_public_url(url, headers=headers, timeout=timeout)
+        response = _get_public_url(url, headers=headers, timeout=timeout,
+                                   max_bytes=effective_cap)
 
         if response.status_code == 429:
             raise RateLimitError(f"Rate limit hit for {url} (attempt {retry_attempt})")
 
         response.raise_for_status()
+    except BodyTooLargeError as e:
+        error_logger.warning(f"Refused oversized body for {url}: {e}")
+        return _empty_result(url, f"TooLarge: {e}")
     except httpx.HTTPStatusError as e:
         error_logger.warning(f"HTTP {e.response.status_code} fetching {url}: {e}")
         return _empty_result(url, f"HTTP {e.response.status_code}: {e}")
@@ -269,9 +397,27 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
         error_logger.error(str(e))
         return _empty_result(url, str(e))
 
+    # Size bookkeeping shared by every content branch below. getattr keeps
+    # plain httpx.Response stand-ins (tests) working without the cap fields.
+    _size_fields = {
+        "truncated": getattr(response, "truncated", False),
+        "fetched_bytes": len(response.content),
+        "total_bytes": getattr(response, "declared_bytes", None),
+    }
+
     # PDF handling
     content_type = response.headers.get("Content-Type", "").lower()
     if "application/pdf" in content_type or url.lower().endswith(".pdf"):
+        if _size_fields["truncated"]:
+            # A PDF cut mid-stream is not parseable; unlike text there is no
+            # useful partial result, so report the budget problem instead.
+            _declared = _size_fields["total_bytes"]
+            return _empty_result(
+                url,
+                f"TooLarge: PDF exceeds the {effective_cap:,}-byte fetch budget"
+                + (f" (size {_declared:,} bytes)" if _declared else "")
+                + "; retry with a larger budget if it fits under the hard cap",
+            )
         if pdf_extract_text is None:
             logger.error("pdfminer.six is not installed; cannot extract PDF text.")
             pdf_text = ""
@@ -295,6 +441,7 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
             "js_message": "",
             "success": bool(pdf_text),
             "error": "" if pdf_text else "Failed to extract PDF text",
+            **_size_fields,
         }
         _cache_result(cache_file, cache_key, result, url)
         return result
@@ -329,6 +476,7 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
             "js_message": "",
             "success": bool(text_body),
             "error": "" if text_body else "Empty response body",
+            **_size_fields,
         }
         _cache_result(cache_file, cache_key, result, url)
         return result
@@ -391,6 +539,7 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
         "js_message": js_message,
         "success": True,
         "error": "",
+        **_size_fields,
     }
     _cache_result(cache_file, cache_key, result, url)
     return result

services/search/providers.py

@@ -9,14 +9,12 @@ from urllib.parse import urljoin, urlparse, parse_qs
 import httpx
 from bs4 import BeautifulSoup
 
-from src.constants import SEARXNG_INSTANCE
+from src.constants import SEARXNG_INSTANCE, REQUEST_TIMEOUT, WEB_FETCH_USER_AGENT
 from .analytics import RateLimitError, error_logger
 from .query import build_enhanced_query
 
 logger = logging.getLogger(__name__)
 
-REQUEST_TIMEOUT = 20
-
 # Provider registry — maps setting value to (label, needs_key, needs_url)
 PROVIDER_INFO = {
     "searxng":  ("SearXNG",           False, True),
@@ -140,7 +138,7 @@ def searxng_search_api(query: str, count: Optional[int] = None, categories: str
     count = count if count is not None else _get_result_count()
     instance = _get_search_instance()
     api_key = ""
-    headers = {"User-Agent": "Mozilla/5.0"}
+    headers = {"User-Agent": WEB_FETCH_USER_AGENT}
     if api_key:
         headers["Authorization"] = f"Bearer {api_key}"
     # News/fresh queries do badly in the 'general' category — it favours
@@ -252,7 +250,7 @@ def searxng_search(query, max_results=10):
     """Search using SearXNG instance - parsing HTML."""
     instance = _get_search_instance()
     api_key = ""
-    req_headers = {"User-Agent": "Mozilla/5.0"}
+    req_headers = {"User-Agent": WEB_FETCH_USER_AGENT}
     if api_key:
         req_headers["Authorization"] = f"Bearer {api_key}"
     try:
@@ -391,7 +389,7 @@ def duckduckgo_search(query: str, count: Optional[int] = None, time_filter: Opti
             response = httpx.get(
                 "https://html.duckduckgo.com/html/",
                 params={"q": query, "kp": _safesearch_for("duckduckgo_html")},
-                headers={"User-Agent": "Mozilla/5.0"},
+                headers={"User-Agent": WEB_FETCH_USER_AGENT},
                 timeout=REQUEST_TIMEOUT,
             )
             response.raise_for_status()

setup.py

@@ -16,8 +16,9 @@ sys.path.insert(0, BASE_DIR)
 from src.constants import (
     DATA_DIR, AUTH_FILE, UPLOAD_DIR, PERSONAL_DIR, PERSONAL_UPLOADS_DIR,
     TTS_CACHE_DIR, GENERATED_IMAGES_DIR, DEEP_RESEARCH_DIR, CHROMA_DIR,
-    RAG_DIR, MEMORY_VECTORS_DIR,
+    RAG_DIR, MEMORY_VECTORS_DIR, PASSWORD_MIN_LENGTH,
 )
+from core.auth import RESERVED_USERNAMES
 
 DIRS = [
     DATA_DIR,
@@ -59,15 +60,23 @@ def _prompt_admin_credentials():
     print("  (Press Enter to accept defaults)")
     print()
 
-    username = input("  Username [admin]: ").strip().lower()
-    if not username:
-        username = "admin"
+    while True:
+        username = input("  Username [admin]: ").strip().lower()
+        if not username:
+            username = "admin"
+        if username in RESERVED_USERNAMES:
+            print(f"  '{username}' is a reserved username. Choose another.")
+            continue
+        break
 
     while True:
         password = getpass.getpass("  Password: ")
         if not password:
             print("  Password cannot be empty.")
             continue
+        if len(password) < PASSWORD_MIN_LENGTH:
+            print(f"  Password must be at least {PASSWORD_MIN_LENGTH} characters.")
+            continue
         confirm = getpass.getpass("  Confirm password: ")
         if password != confirm:
             print("  Passwords don't match. Try again.")
@@ -93,8 +102,13 @@ def create_default_admin():
         password = os.getenv("ODYSSEUS_ADMIN_PASSWORD", "").strip()
 
         if username and password:
-            # Both provided via env — use them directly
-            pass
+            # Both provided via env — validate before using
+            if username in RESERVED_USERNAMES:
+                print(f"  [error] ODYSSEUS_ADMIN_USER '{username}' is a reserved username")
+                return "failed"
+            if len(password) < PASSWORD_MIN_LENGTH:
+                print(f"  [error] ODYSSEUS_ADMIN_PASSWORD must be at least {PASSWORD_MIN_LENGTH} characters")
+                return "failed"
         elif sys.stdin.isatty() and not os.getenv("ODYSSEUS_SKIP_ADMIN_PROMPT"):
             # Interactive terminal — ask the user
             username, password = _prompt_admin_credentials()

specs/architecture-runtime-inventory.md

@@ -0,0 +1,412 @@
+# Architecture Runtime Inventory
+
+> **Purpose**: Phase 0 planning baseline for codebase readability improvements (#4071).
+> **Parent issue**: [#4082](https://github.com/pewdiepie-archdaemon/odysseus/issues/4082)
+> **Last updated**: dev@b58af42 | 2026-06-16
+> **Status**: Draft — to be reviewed before follow-up slices open.
+> **Snapshot basis**: Importer / file / import-line counts are refreshed to `dev@b58af42` (2026-06-16) and are recomputable via the commands in §3.4. **Line counts** in §2.1 / §2.2 are a snapshot from an earlier baseline and drift as `dev` moves — recompute any of them with `wc -l <file>`. This inventory tracks structure and risk, not live metrics.
+
+This document maps the current runtime module structure, identifies high-risk boundaries, and recommends safe first refactor slices. It does **not** move files, change imports, or alter runtime behavior.
+
+---
+
+## 1. Current Structure Overview
+
+### 1.1 Top-Level Layout
+
+```
+odysseus/
+├── app.py                    # FastAPI app entrypoint (1,145 lines)
+├── conf/                     # Configuration (config.py, settings.py, settings_scrub.py)
+├── src/                      # 95 flat .py files + 2 subdirectories
+│   ├── agent_tools/          # Tool helpers: document, filesystem, subprocess, web
+│   └── search/               # Search subsystem
+├── routes/                   # 54 flat .py files — HTTP route handlers
+├── core/                     # 10 files — database models, auth, middleware, session
+├── mcp_servers/              # 5 files — MCP server implementations
+├── scripts/                  # CLI tools and one-shot scripts
+├── static/                   # Frontend HTML/CSS/JS
+├── tests/                    # 583 test files (~54,800 lines)
+└── services/                 # (exists as needed)
+```
+
+### 1.2 Directory Flatness Metric
+
+| Directory | Flat `.py` Files | Subdirectories | Concern |
+|-----------|-----------------|----------------|---------|
+| `src/` | **95** | 2 (`agent_tools/`, `search/`) | No domain grouping; 95 files in one directory |
+| `routes/` | **54** | 0 | All route handlers in one flat directory |
+| `core/` | 10 | 0 | Manageable, but `database.py` is oversized |
+
+---
+
+## 2. Largest Runtime Modules
+
+### 2.1 Python Backend
+
+| Rank | File | Lines | Classes | Functions | Risk |
+|------|------|-------|---------|-----------|------|
+| 1 | `src/tool_implementations.py` | **4,032** | 0 | ~48 | **HIGH** |
+| 2 | `routes/email_routes.py` | **3,245** | — | — | **MEDIUM** |
+| 3 | `routes/cookbook_routes.py` | **2,969** | — | — | **MEDIUM** |
+| 4 | `src/agent_loop.py` | **2,961** | 0 | ~24 | **HIGH** |
+| 5 | `src/task_scheduler.py` | **2,330** | — | 5 | MEDIUM |
+| 6 | `routes/model_routes.py` | **2,266** | — | — | MEDIUM |
+| 7 | `core/database.py` | **2,265** | 28 | ~59 helpers | **HIGH** |
+| 8 | `src/builtin_actions.py` | **2,262** | 2 | ~24 | MEDIUM |
+| 9 | `src/llm_core.py` | **2,164** | — | — | MEDIUM |
+| 10 | `mcp_servers/email_server.py` | 2,197 | — | — | LOW (separate process) |
+| 11 | `src/visual_report.py` | 1,918 | — | — | LOW |
+| 12 | `routes/gallery_routes.py` | 1,896 | — | — | LOW |
+| 13 | `src/ai_interaction.py` | 1,846 | — | — | MEDIUM |
+| 14 | `routes/document_routes.py` | 1,717 | — | — | LOW |
+| 15 | `routes/skills_routes.py` | 1,648 | — | — | LOW |
+
+**Heuristic**: Files > 2,000 lines with 20+ public symbols and many importers are the highest-risk splits. Files 1,000–2,000 lines are medium-risk if tightly coupled.
+
+### 2.2 Frontend
+
+| File | Lines | Concern |
+|------|-------|---------|
+| `static/style.css` | **36,653** | Entire app CSS in one file (tracked separately in #2617) |
+| `static/js/document.js` | **9,776** | Single JS file for document functionality |
+| `static/js/slashCommands.js` | 6,498 | |
+| `static/js/settings.js` | 5,266 | |
+| `static/js/emailLibrary.js` | 5,217 | |
+| `static/js/notes.js` | 5,124 | |
+| `static/js/chat.js` | 4,985 | |
+| `static/app.js` | 4,090 | |
+
+**Note**: Frontend modularization is tracked separately in #2617 (CSS) and is not the focus of this Phase 0 inventory. Frontend is listed here for completeness but follow-up slices should target Python backend boundaries first.
+
+---
+
+## 3. Import Dependency Graph
+
+### 3.1 Who Depends on `core/database.py`
+
+**102 files** import from `core.database` — this is the most depended-upon module:
+
+- All route handlers (`routes/*.py`)
+- Most `src/*.py` files
+- `core/session_manager.py`, `core/auth.py`
+- Multiple test files
+
+**Implication**: Any split of `core/database.py` is the highest-risk refactor. It should be tackled **last**, never first.
+
+### 3.2 Who Depends on `src/tool_implementations.py`
+
+**17 files** import from `src.tool_implementations`:
+- `src/agent_loop.py`, `src/builtin_actions.py`, `src/tool_index.py`
+- `src/task_scheduler.py`, `src/tool_policy.py`
+- Various tests
+
+### 3.3 Who Depends on `src/agent_loop.py`
+
+**22 files** import from `src.agent_loop`:
+
+- `src/tool_policy.py`, `src/teacher_escalation.py`, `src/bg_monitor.py`
+- `src/task_scheduler.py`
+- Multiple test files
+
+### 3.4 Cross-Layer Import Violations
+
+**`src/` importing from `routes/`** (backwards dependency — domain logic depending on HTTP layer):
+
+```
+src/tool_implementations.py ──→ routes/calendar_routes.py
+src/tool_implementations.py ──→ routes/cookbook_helpers.py
+src/tool_implementations.py ──→ routes/email_helpers.py
+src/tool_implementations.py ──→ routes/email_pollers.py
+src/tool_implementations.py ──→ routes/email_routes.py
+src/tool_implementations.py ──→ routes/model_routes.py
+src/tool_implementations.py ──→ routes/note_routes.py
+src/tool_implementations.py ──→ routes/prefs_routes.py
+```
+
+> These are **runtime imports** (inside function bodies, not at module top), which mitigates circular import risk but indicates fuzzy layer boundaries. Function-level inline imports from the HTTP layer into business logic are a code smell.
+
+**Import counts (top-level)**:
+| Direction | Count | Notes |
+|-----------|-------|-------|
+| `routes/` → `src/` | **374** | Expected: HTTP handlers call domain logic |
+| `routes/` → `core/` | **126** | Expected: handlers access DB models |
+| `src/` → `routes/` | **31** | **Unexpected**: domain logic reaching into HTTP layer (direct grep of import lines referencing `routes/`) |
+| `src/` → `core/` | **106** | Acceptable but could be reduced with a data-access layer |
+
+> **How the metrics in this document are computed** — recompute against current `dev` before treating any count as authoritative (the tree drifts; these numbers are a snapshot, not a live value):
+> - `src/` flat `.py` files: `find src -maxdepth 1 -name '*.py' | wc -l`
+> - `tests/` test files: `find tests -name 'test_*.py' | wc -l`
+> - `core.database` importers: `grep -rlE '(from|import) +core\.database' --include='*.py' . | grep -v core/database.py | wc -l`
+> - `src.agent_loop` importers: `grep -rlE '(from|import) +src\.agent_loop' --include='*.py' . | grep -v src/agent_loop.py | wc -l`
+> - Cross-layer import lines: `grep -rhE '(from|import) +<pkg>' --include='*.py' <dir>/ | wc -l` (e.g. `(from|import) +routes` over `src/`)
+
+---
+
+## 4. Route Ownership Map
+
+Routes can be grouped into logical feature domains. Current flat structure obscures these boundaries:
+
+| Domain | Route Files | Total Lines | Review Complexity |
+|--------|-------------|-------------|-------------------|
+| **Email** | `email_routes.py`, `email_helpers.py`, `email_pollers.py` | 5,936 | HIGH — most complex domain |
+| **Chat / Agent** | `chat_routes.py`, `chat_helpers.py`, `shell_routes.py`, `codex_routes.py`, `skills_routes.py` | 6,365 | HIGH — core interaction surface |
+| **Cookbook** | `cookbook_routes.py`, `cookbook_helpers.py`, `cookbook_output.py` | 4,110 | MEDIUM |
+| **Model / LLM** | `model_routes.py`, `assistant_routes.py`, `copilot_routes.py` | 2,764 | MEDIUM |
+| **Calendar / Contacts** | `calendar_routes.py`, `contacts_routes.py` | 2,336 | MEDIUM |
+| **Documents** | `document_routes.py`, `document_helpers.py` | 1,954 | LOW |
+| **Auth** | `auth_routes.py`, `api_token_routes.py`, `device_flow.py` | 1,171 | LOW |
+| **Tasks** | `task_routes.py` (standalone) | 1,157 | LOW |
+| **Session** | `session_routes.py` (standalone) | 1,287 | LOW |
+| **Gallery** | `gallery_routes.py`, `gallery_helpers.py` | 1,896 | LOW |
+| **Memory** | `memory_routes.py` | — | LOW |
+| **Research** | `research_routes.py` | — | LOW |
+| **MCP** | `mcp_routes.py` | — | LOW |
+| **Notes** | `note_routes.py` | — | LOW |
+| **Other** | `prefs_routes.py`, `upload_routes.py`, `vault_routes.py`, `webhook_routes.py`, `workspace_routes.py`, `search_routes.py`, `history_routes.py`, `hwfit_routes.py`, `preset_routes.py`, `signature_routes.py`, `backup_routes.py`, `cleanup_routes.py`, `diagnostics_routes.py`, `embedding_routes.py`, `emoji_routes.py`, `font_routes.py`, `stt_routes.py`, `tts_routes.py`, `compare_routes.py`, `personal_routes.py`, `editor_draft_routes.py`, `admin_wipe_routes.py`, `chatgpt_subscription_routes.py` | 2,000+ | LOW individual, HIGH cumulative |
+
+---
+
+## 5. Tool Registry & Implementation Boundaries
+
+### 5.1 Current Tool Architecture
+
+| Component | File | Lines | Role |
+|-----------|------|-------|------|
+| Tool schemas | `src/tool_schemas.py` | 1,392 | JSON Schema tool definitions (Duck-TypedDict) |
+| Tool index | `src/tool_index.py` | 542 | RAG-based tool retrieval from ChromaDB |
+| Tool implementations | `src/tool_implementations.py` | 4,032 | 33 `do_*` functions — all tool execution logic |
+| Tool security | `src/tool_security.py` | — | Owner-scoped tool blocking |
+| Tool policy | `src/tool_policy.py` | — | Guide-only directive, plan-mode disabled tools |
+| Tool utils | `src/tool_utils.py` | — | Shared tool helpers |
+
+### 5.2 Tool Implementation Categories
+
+The 33 `do_*` functions in `tool_implementations.py` fall into natural domain groups — the basis for slice 1's split in §6.2:
+
+| Category | `do_*` functions | Count |
+|----------|------------------|-------|
+| **System / config** | `do_manage_skills`, `do_manage_tasks`, `do_manage_endpoints`, `do_manage_mcp`, `do_manage_webhooks`, `do_manage_tokens`, `do_manage_settings`, `do_api_call`, `do_app_api` | 9 |
+| **Cookbook / model serving** | `do_download_model`, `do_serve_model`, `do_list_served_models`, `do_stop_served_model`, `do_tail_serve_output`, `do_list_downloads`, `do_cancel_download`, `do_search_hf_models`, `do_adopt_served_model`, `do_list_cookbook_servers`, `do_list_serve_presets`, `do_serve_preset`, `do_list_cached_models` | 13 |
+| **Notes** | `do_manage_notes` | 1 |
+| **Calendar** | `do_manage_calendar` | 1 |
+| **Search** | `do_search_chats` | 1 |
+| **Research** | `do_manage_research`, `do_trigger_research` | 2 |
+| **Contacts** | `do_resolve_contact`, `do_manage_contact` | 2 |
+| **Vault** | `do_vault_search`, `do_vault_get`, `do_vault_unlock` | 3 |
+| **Image** | `do_edit_image` | 1 |
+| | **Total** | **33** |
+
+> Low-level tools (filesystem, subprocess, web fetch, document parsing) live in `src/agent_tools/`, **not** in `tool_implementations.py` — out of scope for this split.
+
+---
+
+## 6. Risk Assessment & Candidate Slice Ranking
+
+> **Candidate proposals, not a committed plan.** The rankings, package shapes (e.g. `src/pkg/`, `src/domain/`, `src/infra/`, `src/api/`), split ordering, and route-grouping strategy below are **options for maintainer discussion**. Per #4082/#4071, slice ownership and order are settled by maintainers before any follow-up PR. §1–§3 above are the factual current-state inventory.
+
+### 6.1 Risk Scale
+
+| Level | Criteria |
+|-------|----------|
+| **LOW** | File has ≤3 importers AND ≤500 lines, OR is a pure refactor with clear boundaries |
+| **MEDIUM** | File has 4–15 importers OR 500–1,500 lines |
+| **HIGH** | File has 16+ importers OR >2,000 lines, OR has cross-layer import violations |
+
+### 6.2 Ranked Split Candidates
+
+| Priority | Target | Risk | Rationale |
+|----------|--------|------|-----------|
+| **1** | `src/tool_implementations.py` → `src/tools/*.py` | **MEDIUM** | 4,032 lines → ~10 files by tool category. Already has natural boundaries. 17 importers, tracked in #3629. Use `__init__.py` shim to keep existing imports working. |
+| **2** | `routes/` → domain subdirectories (one domain per PR) | **MEDIUM** | 54 flat files. Done **one domain at a time** (e.g. a standalone PR for the email domain, then chat, …), not a broad reorganization — route modules carry helper imports, registration assumptions, and test import paths. |
+| **3** | `src/agent_loop.py` → `src/agent/loop.py` + submodules | **MEDIUM-HIGH** | 2,961 lines, 24 functions. Can extract prompt building, classification, verification, and runaway detection. Tracked in #3266. |
+| **4** | `src/` → `src/pkg/`, `src/domain/`, `src/infra/`, `src/api/` | **MEDIUM** | Structural reorganization. Split flat `src/` into layered packages. Must come after routes and tools are stable. |
+| **5** | `routes/email_*.py` consolidation | **LOW** | Already grouped by filename prefix. Low-risk cleanup within the email domain. |
+| **6** | `core/database.py` → `src/infra/database/models/*.py` | **HIGH** | 28 classes, 102 importers. Highest-risk split. Must be **last** in any sequence. Requires careful import shim strategy. |
+| **7** | Frontend CSS modularization | **MEDIUM** | 36,653 lines. Tracked in #2617. Separate timeline from backend work. |
+| **8** | Frontend JS modularization | **MEDIUM** | 9,776 lines in `document.js`. Introduce ES modules at minimum. |
+
+### 6.3 Candidate First 3 Behavior-Preserving Slices
+
+**Slice 1: Split `tool_implementations.py`** (Lowest-risk high-impact)
+
+- Create `src/tools/` package with one file per tool category
+- Add `src/tools/__init__.py` re-exporting all symbols with current names
+- Update 17 importers to use new paths (can be deferred via shim)
+- Validation: `python -m pytest tests/ -x -q` + manual smoke test of tool execution
+- Reference: #3629
+
+**Slice 2: Group `routes/` by domain** (one domain per PR, not a broad sweep)
+
+Route modules carry helper imports, router registration assumptions, and test import paths, so this must be done **one domain at a time** rather than as a single reorganization PR. Example sequence (each its own PR):
+
+- PR 2a: move the **email** domain (`email_routes.py`, `email_helpers.py`, `email_pollers.py`) → `routes/email/` + shim
+- PR 2b: move the **chat/agent** domain → `routes/chat/` + shim
+- PR 2c: move the **cookbook** domain → `routes/cookbook/` + shim
+- …and so on per domain from §4
+
+Each PR: add `__init__.py` re-exporting old names, update `app.py` router imports, validation `python app.py` starts clean. **No behavior change** — pure file reorganization.
+
+**Slice 3: Extract `agent_loop.py` submodules** (Improve reviewability)
+
+- Move prompt assembly → `src/agent/prompt.py`
+- Move request classification → `src/agent/classifier.py`
+- Move sub-agent verification → `src/agent/verifier.py`
+- Move runaway detection → `src/agent/runaway.py`
+- Move context management → `src/agent/context.py`
+- Keep `src/agent/loop.py` as the main orchestration module
+- Validation: `python -m pytest tests/test_agent_loop.py tests/test_loop_breaker_runaway.py -v`
+
+---
+
+## 7. Safety Guardrails for Follow-Up Work
+
+Per maintainer guidance in #4082 and #4071:
+
+- [ ] **One domain/slice per PR** — never mix multiple reorganizations
+- [ ] **No behavior changes** mixed with file moves — pure reorganization only
+- [ ] **Keep compatibility shims** — `__init__.py` re-exports for all existing import paths
+- [ ] **Add or identify focused tests** before risky splits
+- [ ] **Do not start with `core/database.py`** or broad route movement unless this inventory shows a safe boundary
+- [ ] **Prefer small, reviewable slices** over large restructures
+- [ ] **No packaging/runtime/tooling migration** mixed into file moves
+- [ ] **No frontend framework migration** inside this stabilization lane
+- [ ] **Validate with `python -m compileall`** — every PR must pass CI checks
+- [ ] **Validate with `pytest`** — run the full test suite before opening each PR
+
+---
+
+## 8. Validation Commands
+
+Each follow-up PR should be verifiable with these commands before submission:
+
+```bash
+# Syntax check — must pass with zero errors
+python -m compileall src/ routes/ core/ conf/
+
+# Full test suite — must match baseline pass rate
+python -m pytest tests/ -x -q
+
+# Import shim verification — existing import paths must still work
+python -c "from src.tool_implementations import do_search_chats; print('OK')"
+
+# App startup smoke test (if backend touched)
+timeout 5 python app.py 2>&1 | head -5 || true
+```
+
+---
+
+## 9. Open Questions
+
+1. Is `#2538` (specs ground truth) the canonical behavior map baseline, and should this inventory be kept in sync with those specs once merged?
+2. Should route grouping follow the domain map proposed here, or is there a different taxonomy preferred by maintainers?
+3. For the `tool_implementations.py` split (#3629), is the tool categorization in §5.2 acceptable, or should it follow a different grouping?
+4. Should compatibility shims (`__init__.py`) be temporary (removed in a follow-up wave) or permanent?
+5. Should an ADR (Architecture Decision Record) document be started to track decisions made during this process?
+
+---
+
+## 10. Future Direction (NOT current state)
+
+The following are **future refactor targets** (candidate directions **pending maintainer agreement**, not committed), recorded here so this inventory does not imply they exist today. None of them are present in the current `dev` tree:
+
+- `main.py` — proposed rename of the `app.py` entrypoint. Today the app boots via `app.py`.
+- `src/agent/` — proposed package to hold `agent_loop.py` submodules (prompt/classifier/verifier/runaway/context). Today `agent_loop.py` is a single flat file in `src/`.
+- `src/infra/`, `src/domain/`, `src/pkg/`, `src/api/` — proposed layered reorganization of the flat `src/` directory (slice 4 in §6).
+
+These become real only when the corresponding slices land.
+
+---
+
+## Appendix A: File Listing
+
+### `src/` (95 files — 61 shown; run `ls src/*.py` for the full list)
+
+```
+agent_loop.py          tool_implementations.py   tool_schemas.py
+tool_index.py          tool_security.py          tool_policy.py
+tool_utils.py          builtin_actions.py        task_scheduler.py
+llm_core.py            model_context.py          model_discovery.py
+session_search.py      context_budget.py         context_compactor.py
+ai_interaction.py      action_intents.py         agent_runs.py
+app_helpers.py         app_initializer.py        config.py
+database.py            memory.py                 memory_provider.py
+secret_storage.py      prompt_security.py        url_security.py
+url_safety.py          rate_limiter.py           cleanup_service.py
+readiness.py           service_health.py         exceptions.py
+request_models.py      assistant_log.py          bg_monitor.py
+builtin_mcp.py         chat_helpers.py           chroma_client.py
+document_processor.py  embedding_lanes.py        deep_research.py
+research_handler.py    research_utils.py         personal_docs.py
+rag_manager.py         rag_singleton.py          topic_analyzer.py
+visual_report.py       youtube_handler.py        pdf_forms.py
+pdf_form_doc.py        pdf_runtime.py            caldav_writeback.py
+email_thread_parser.py text_helpers.py           user_time.py
+teacher_escalation.py  cookbook_serve_lifecycle.py
+chatgpt_subscription.py  mcp_manager.py
+```
+
+### `routes/` (54 files)
+
+```
+__init__.py    _validators.py
+auth_routes.py              api_token_routes.py       device_flow.py
+chat_routes.py              chat_helpers.py           shell_routes.py
+codex_routes.py             skills_routes.py
+email_routes.py             email_helpers.py          email_pollers.py
+cookbook_routes.py          cookbook_helpers.py       cookbook_output.py
+model_routes.py             assistant_routes.py       copilot_routes.py
+calendar_routes.py          contacts_routes.py
+document_routes.py          document_helpers.py
+gallery_routes.py           gallery_helpers.py
+task_routes.py              session_routes.py
+note_routes.py              memory_routes.py          research_routes.py
+mcp_routes.py               search_routes.py          history_routes.py
+webhook_routes.py           workspace_routes.py       upload_routes.py
+vault_routes.py             prefs_routes.py           preset_routes.py
+signature_routes.py         personal_routes.py        hwfit_routes.py
+backup_routes.py            cleanup_routes.py         diagnostics_routes.py
+embedding_routes.py         emoji_routes.py           font_routes.py
+stt_routes.py               tts_routes.py             compare_routes.py
+editor_draft_routes.py      chatgpt_subscription_routes.py    admin_wipe_routes.py
+```
+
+### `core/` (10 files)
+
+```
+__init__.py    constants.py    database.py    models.py
+auth.py        middleware.py   session_manager.py   exceptions.py
+atomic_io.py   platform_compat.py
+```
+
+---
+
+## Appendix B: Key Import Relationships
+
+```
+core/database.py  ←── 102 importers (routes/*, src/*, core/*, tests/*)
+    ↑
+    ├── routes/auth_routes.py
+    ├── routes/email_routes.py
+    ├── src/builtin_actions.py
+    ├── src/task_scheduler.py
+    ├── src/tool_implementations.py (inline)
+    └── ...97 more
+
+src/tool_implementations.py  ←── 17 importers
+    ↑
+    ├── src/agent_loop.py
+    ├── src/builtin_actions.py
+    ├── src/tool_index.py
+    ├── src/task_scheduler.py
+    ├── src/tool_policy.py
+    └── ...12 more (mostly tests)
+
+src/agent_loop.py  ←── 22 importers
+    ↑
+    ├── src/tool_policy.py
+    ├── src/teacher_escalation.py
+    ├── src/bg_monitor.py
+    ├── src/task_scheduler.py
+    └── 18 more (incl. tests)
+```

src/agent_loop.py

@@ -38,167 +38,6 @@ from src.agent_tools import (
 
 logger = logging.getLogger(__name__)
 
-# Redaction patterns for common secret-bearing shapes. Explicit and tested
-# (see tests/test_loop_guard_signals.py) rather than one clever broad regex —
-# safety first, but we try not to mangle harmless prose. Applied in order.
-_REDACTED = "[redacted]"
-
-# Cookie: ... / Set-Cookie: ... — redact the rest of the line (cookies hold spaces).
-_SENSITIVE_COOKIE_RE = re.compile(
-    r"(?i)\b((?:set-)?cookie\s*[:=]\s*)[^\r\n]+"
-)
-# URL credentials, e.g. postgres://user:pass@host/db. The password half allows
-# inner colons (postgres://user:pa:ss@host/db) but still stops at / and @.
-_SENSITIVE_URL_CRED_RE = re.compile(
-    r"(?i)\b([a-z][a-z0-9+.\-]*://)[^\s:/@]+:[^\s/@]+@"
-)
-# Prefix-only discovery regexes. Each matches the key and its separator (the part
-# we KEEP); the value that follows is found by a linear scanner rather than by a
-# regex, so there is no backtracking-prone quantifier over uncontrolled input.
-#
-# Authorization: Bearer <tok> / Authorization: Basic "two word secret"
-_AUTH_PREFIX_RE = re.compile(
-    r"(?i)authorization\s*[:=]\s*(?:bearer|basic)\s+"
-)
-# Provider-prefixed env names, e.g. OPENAI_API_KEY=..., AWS_SECRET_ACCESS_KEY=...,
-# GITHUB_TOKEN=... — require a sensitive suffix preceded by `_` so benign names
-# that merely end in KEY (MONKEY, TURKEY) are left alone.
-_ENV_PREFIX_RE = re.compile(
-    r"(?:export\s+)?\b[A-Z][A-Z0-9_]*"
-    r"_(?:KEY|TOKEN|SECRET|PASSWORD|PASSWD|PWD|CREDENTIALS?)\s*=\s*"
-)
-# Generic sensitive key, e.g. password=..., api_key: ..., client_secret=...
-_KEY_PREFIX_RE = re.compile(
-    r"(?i)\b(?:password|passwd|pwd|token|api[_-]?key|client_secret|secret)\b\s*[:=]\s*"
-)
-# Obvious provider-shaped bare tokens (no surrounding key needed).
-_SENSITIVE_BARE_TOKEN_RE = re.compile(
-    r"\b("
-    r"sk-[A-Za-z0-9_\-]{16,}"          # OpenAI / Anthropic style
-    r"|gh[pousr]_[A-Za-z0-9]{20,}"     # GitHub PAT
-    r"|xox[baprs]-[A-Za-z0-9\-]{10,}"  # Slack
-    r"|AKIA[0-9A-Z]{16}"               # AWS access key id
-    r"|hf_[A-Za-z0-9]{16,}"            # Hugging Face token
-    r"|AIza[0-9A-Za-z_\-]{20,}"        # Google API key
-    r")\b"
-)
-
-
-def _consume_secret_value_end(text: str, start: int) -> int:
-    """Return the exclusive end index of the secret value beginning at ``start``.
-
-    If the value is quoted, scan to the matching unescaped quote (backslash
-    escapes are skipped two chars at a time). Otherwise scan to the first
-    whitespace, comma, or semicolon. The scan is linear in the length of the
-    input, so it cannot exhibit catastrophic backtracking.
-    """
-    n = len(text)
-    if start >= n:
-        return start
-    quote = text[start]
-    if quote in ("'", '"'):
-        i = start + 1
-        while i < n:
-            ch = text[i]
-            if ch == "\\":
-                i += 2
-                continue
-            if ch == quote:
-                return i + 1
-            i += 1
-        return n  # unterminated quote: redact to the end
-    i = start
-    while i < n and not text[i].isspace() and text[i] not in (",", ";"):
-        i += 1
-    return i
-
-
-def _redact_after_prefix(text: str, prefix_re: "re.Pattern") -> str:
-    """Redact the value following each ``prefix_re`` match using a linear scan."""
-    result = []
-    pos = 0
-    n = len(text)
-    while pos < n:
-        match = prefix_re.search(text, pos)
-        if match is None:
-            result.append(text[pos:])
-            break
-        result.append(text[pos:match.end()])
-        value_end = _consume_secret_value_end(text, match.end())
-        if value_end > match.end():
-            result.append(_REDACTED)
-            pos = value_end
-        else:
-            # Empty value: nothing to redact; step past the prefix and continue.
-            pos = match.end()
-            if pos < n:
-                result.append(text[pos])
-                pos += 1
-    return "".join(result)
-
-
-def _redact_private_keys(text: str) -> str:
-    """Replace PEM private-key blocks with a placeholder via linear scanning.
-
-    Finds ``-----BEGIN `` markers, verifies the header names a PRIVATE KEY,
-    locates the matching ``-----END `` marker, and collapses the whole block.
-    No regex is used, so the (multi-line, uncontrolled) body cannot trigger
-    polynomial matching.
-    """
-    begin_marker = "-----BEGIN "
-    end_marker = "-----END "
-    dash = "-----"
-    max_header = 64  # generous bound on "[TYPE ]PRIVATE KEY"
-    result = []
-    pos = 0
-    while True:
-        begin = text.find(begin_marker, pos)
-        if begin == -1:
-            result.append(text[pos:])
-            return "".join(result)
-        header_start = begin + len(begin_marker)
-        header_close = text.find(dash, header_start)
-        if (
-            header_close == -1
-            or header_close - header_start > max_header
-            or not text[header_start:header_close].endswith("PRIVATE KEY")
-        ):
-            result.append(text[pos:header_start])
-            pos = header_start
-            continue
-        end = text.find(end_marker, header_close)
-        if end == -1:
-            result.append(text[pos:])
-            return "".join(result)
-        end_header_start = end + len(end_marker)
-        end_close = text.find(dash, end_header_start)
-        if (
-            end_close == -1
-            or end_close - end_header_start > max_header
-            or not text[end_header_start:end_close].endswith("PRIVATE KEY")
-        ):
-            result.append(text[pos:header_start])
-            pos = header_start
-            continue
-        result.append(text[pos:begin])
-        result.append("[redacted private key]")
-        pos = end_close + len(dash)
-
-
-def _redact_sensitive_text(value: object) -> str:
-    """Redact obvious credential values before surfacing tool output."""
-    if value is None:
-        return ""
-
-    text = str(value)
-    text = _redact_private_keys(text)
-    text = _redact_after_prefix(text, _AUTH_PREFIX_RE)
-    text = _SENSITIVE_COOKIE_RE.sub(r"\1" + _REDACTED, text)
-    text = _SENSITIVE_URL_CRED_RE.sub(r"\1" + _REDACTED + "@", text)
-    text = _redact_after_prefix(text, _ENV_PREFIX_RE)
-    text = _redact_after_prefix(text, _KEY_PREFIX_RE)
-    return _SENSITIVE_BARE_TOKEN_RE.sub(_REDACTED, text)
-
 
 def _load_mcp_disabled_map() -> Dict[str, set]:
     """Load per-server disabled tool sets from the database."""
@@ -428,6 +267,10 @@ _DOMAIN_RULES = {
 - Use `resolve_contact` to look up a contact's email or phone number by name. Searches the CardDAV address book and sent email history.
 - Use `manage_contact` to list, add, update, or delete contacts in the address book.
 - Do NOT use `manage_memory` for contact lookups — contact details live in the address book, not memory.""",
+    "integrations": """\
+## Integration/API rules
+- To query or control a configured service integration (Home Assistant, Miniflux, Gitea, Linkding, Jellyfin, or any other registered service), use `api_call` with the integration name, HTTP method, path, and optional JSON body.
+- Do not use shell, curl, or `app_api` to reach a user's connected integration when `api_call` is available.""",
 }
 
 _DOMAIN_TOOL_MAP = {
@@ -438,9 +281,10 @@ _DOMAIN_TOOL_MAP = {
     "notes_calendar_tasks": {"manage_notes", "manage_calendar", "manage_tasks"},
     "ui": {"ui_control"},
     "sessions": {"create_session", "list_sessions", "manage_session", "send_to_session", "search_chats"},
-    "files": {"bash", "python", "read_file", "write_file", "edit_file", "grep", "glob", "ls", "get_workspace"},
+    "files": {"bash", "python", "read_file", "write_file", "edit_file", "grep", "glob", "ls", "get_workspace", "manage_bg_jobs"},
     "settings": {"manage_settings", "manage_endpoints", "manage_mcp", "manage_webhooks", "manage_tokens", "app_api"},
     "contacts": {"resolve_contact", "manage_contact"},
+    "integrations": {"api_call"},
 }
 
 def _domain_rules_for_tools(tool_names: set) -> list[str]:
@@ -685,7 +529,7 @@ def get_builtin_overrides() -> dict:
         ov = get_setting("builtin_tool_overrides", {})
         return ov if isinstance(ov, dict) else {}
     except Exception as e:
-        logger.warning('Failed to load builtin tool overrides: %s', e)
+        logger.warning("Failed to load builtin tool overrides, using defaults", exc_info=e)
         return {}
 
 
@@ -697,17 +541,44 @@ def _section_text(name: str, default: str) -> str:
     return val if isinstance(val, str) and val.strip() else default
 
 
+def _compact_tool_line(name: str, section: str) -> str:
+    """One-line fenced-tool usage hint for compact/local prompts."""
+    text = (section or "").strip()
+    if not text:
+        return f"- `{name}`"
+    if text.startswith("- "):
+        return text
+    lines = [ln.strip() for ln in text.splitlines() if ln.strip()]
+    usage = []
+    in_fence = False
+    for ln in lines:
+        if ln.startswith("```"):
+            usage.append(ln)
+            in_fence = not in_fence
+            if len(usage) >= 3:
+                break
+            continue
+        if in_fence and len(usage) < 3:
+            usage.append(ln)
+    if usage:
+        return f"- `{name}` — " + " ".join(usage)
+    return f"- `{name}` — " + lines[0][:160]
+
+
 def _assemble_prompt(tool_names: set, disabled_tools: set = None, compact: bool = False) -> str:
     """Build the system prompt with only the specified tools included."""
     disabled = disabled_tools or set()
     included = tool_names - disabled
 
     if compact:
-        tool_list = ", ".join(sorted(included)) if included else "none"
+        tool_lines = []
+        for name, _default_section in TOOL_SECTIONS.items():
+            if name in included:
+                tool_lines.append(_compact_tool_line(name, _section_text(name, _default_section)))
         parts = [
-            "You are an AI assistant with tool access.",
-            f"Available tools: {tool_list}.",
-            _API_AGENT_RULES,
+            _AGENT_PREAMBLE,
+            "## Available tools\n" + ("\n".join(tool_lines) if tool_lines else "none"),
+            _AGENT_RULES,
         ]
         parts.extend(_domain_rules_for_tools(included))
         return "\n\n".join(parts)
@@ -773,11 +644,6 @@ _API_HOSTS = frozenset([
     "api.perplexity.ai", "api.x.ai",
     "ollama.com", "api.venice.ai", "api.kimi.com",
     "api.githubcopilot.com",
-    # Local OpenAI-compatible endpoints (llama.cpp, vLLM, LM Studio, etc.).
-    # Without these, `_is_api_model` falls back to keyword sniffing on the
-    # model name, so well-behaved local servers don't get native tool
-    # schemas and the agent silently degrades to fenced-block parsing.
-    "localhost", "127.0.0.1", "host.docker.internal",
 ])
 _MCP_KEYWORDS = frozenset(["mcp", "browse", "browser", "website", "calendar", "event", "email",
                            "gmail", "screenshot", "navigate", "click", "miniflux", "rss", "feed"])
@@ -805,6 +671,28 @@ def _is_ollama_openai_compat_url(endpoint_url: str) -> bool:
     return parsed.port == 11434 and (path == "/v1" or path.startswith("/v1/"))
 
 
+def _is_local_openai_compat_url(endpoint_url: str) -> bool:
+    try:
+        parsed = urlparse(endpoint_url or "")
+    except Exception:
+        return False
+    host = (parsed.hostname or "").lower()
+    path = (parsed.path or "").rstrip("/")
+    if not (path == "/v1" or path.startswith("/v1/")):
+        return False
+    if host in {"localhost", "127.0.0.1", "0.0.0.0", "host.docker.internal"}:
+        return True
+    if host.startswith("192.168.") or host.startswith("10."):
+        return True
+    if host.startswith("172."):
+        try:
+            second = int(host.split(".")[1])
+            return 16 <= second <= 31
+        except Exception:
+            return False
+    return False
+
+
 def _endpoint_lookup_keys(endpoint_url: str) -> List[str]:
     """Candidate ModelEndpoint.base_url keys for a runtime chat URL."""
     raw = (endpoint_url or "").strip()
@@ -868,6 +756,17 @@ def _extract_last_user_message(messages: List[Dict]) -> str:
 
 
 _LOW_SIGNAL_RE = re.compile(r"^[\W_]*$", re.UNICODE)
+_CASUAL_OPENING_RE = re.compile(
+    r"^\s*(?:h+i+|hey+|hello+|yo+|sup+|what'?s up|wass?up|hiya|howdy|"
+    r"lol|lmao|haha+|hehe+|thanks?|thank you|ty|idk|dunno|meh|bruh|bro)\b(?P<tail>.*)$",
+    re.IGNORECASE,
+)
+_CASUAL_BLOCKLIST_RE = re.compile(
+    r"\b(?:cookbook|serve|serving|launch|start|vllm|sglang|llama\.?cpp|ollama|"
+    r"download|model|email|document|doc|note|calendar|task|search|web|research|"
+    r"file|folder|repo|git|settings?|endpoint|api|token|mcp)\b",
+    re.IGNORECASE,
+)
 _EXPLICIT_CONTINUATION_RE = re.compile(
     r"^\s*(?:"
     r"yes|y|yeah|yep|ok|okay|sure|do it|go ahead|continue|carry on|"
@@ -877,6 +776,17 @@ _EXPLICIT_CONTINUATION_RE = re.compile(
     r")\s*[.!?]*\s*$",
     re.IGNORECASE,
 )
+_RETRY_CONTINUATION_RE = re.compile(
+    r"\b(?:try again|retry|again|rerun|re-run|run it again|launch it again|"
+    r"start it again|failed|fails?|died|crashed|broke|insta|instantly)\b",
+    re.IGNORECASE,
+)
+_COOKBOOK_CONTEXT_RE = re.compile(
+    r"\b(?:cookbook|serve|serving|served|launch|start|preset|vllm|sglang|"
+    r"llama\.?cpp|ollama|download|cached models?|model servers?|running models?|"
+    r"gpu box|ajax|qwen|gemma|llama|mistral|minimax)\b",
+    re.IGNORECASE,
+)
 
 
 def _is_explicit_continuation(text: str) -> bool:
@@ -884,6 +794,37 @@ def _is_explicit_continuation(text: str) -> bool:
     return bool(_EXPLICIT_CONTINUATION_RE.match(str(text or "").strip()))
 
 
+def _is_casual_low_signal(text: str) -> bool:
+    """True for short greetings/slang that should not inherit stale context."""
+    s = str(text or "").strip()
+    m = _CASUAL_OPENING_RE.match(s)
+    if not m:
+        return False
+    tail = m.group("tail") or ""
+    if _CASUAL_BLOCKLIST_RE.search(tail):
+        return False
+    # Allow a short vocative/address after the opener without hardcoding the
+    # address term itself: "hey man", "yo dude", "sup <name>". Longer tails are
+    # more likely to be an actual request and should get normal context/tooling.
+    tail_words = re.findall(r"[A-Za-z0-9_'-]+", tail)
+    return len(tail_words) <= 2
+
+
+def _is_contextual_retry_continuation(messages: List[Dict], text: str) -> bool:
+    """Treat "try again / it failed" as a continuation only for active tool work.
+
+    These follow-ups are common after Cookbook launches: the latest user turn
+    says only "try again it failed", while the actionable model/host/command
+    details live one or two turns back. Keep this intentionally narrow so
+    ordinary chat does not inherit stale Cookbook context.
+    """
+    latest = str(text or "").strip()
+    if not latest or not _RETRY_CONTINUATION_RE.search(latest):
+        return False
+    recent = _recent_context_for_retrieval(messages, max_user=5, max_chars=1200)
+    return bool(_COOKBOOK_CONTEXT_RE.search(recent))
+
+
 def _assistant_requested_followup(messages: List[Dict]) -> bool:
     """True when the previous assistant turn asked for missing task details.
 
@@ -925,11 +866,12 @@ def _classify_agent_request(messages: List[Dict], last_user: str) -> Dict[str, o
     which domain rule packs get appended to the system prompt.
     """
     text = str(last_user or "").strip()
-    continuation = _is_explicit_continuation(text) or _assistant_requested_followup(messages)
+    retry_continuation = _is_contextual_retry_continuation(messages, text)
+    continuation = _is_explicit_continuation(text) or _assistant_requested_followup(messages) or retry_continuation
     retrieval_query = _recent_context_for_retrieval(messages) if continuation else text
     q = retrieval_query.lower()
 
-    if not text or bool(_LOW_SIGNAL_RE.match(text)):
+    if not text or bool(_LOW_SIGNAL_RE.match(text)) or _is_casual_low_signal(text):
         return {
             "low_signal": True,
             "continuation": False,
@@ -972,10 +914,25 @@ def _classify_agent_request(messages: List[Dict], last_user: str) -> Dict[str, o
         domains.add("sessions")
     if has(r"\b(file|folder|directory|repo|git|grep|find in files|read file|edit file|shell|terminal|bash|python)\b"):
         domains.add("files")
+    # Managing detached bash jobs: "kill the background job", "stop the job",
+    # "kill that job", "check the job output", "is the bg job done".
+    if (has(r"\b(background|bg)\s+(jobs?|task)\b")
+            or has(r"\b(kill|stop|cancel|terminate|check|tail|show|list)\b.{0,16}\bjobs?\b")
+            or has(r"\bjobs?\b.{0,16}\b(output|status|done|finished|running)\b")):
+        domains.add("files")
     if has(r"\b(endpoint|api token|mcp|webhook|preference|configure|config|setting)\b"):
         domains.add("settings")
     if has(r"\b(contact|contacts|phone|phone number|address book|vcard)\b"):
         domains.add("contacts")
+    # API-integration intent — calling a configured service via the api_call
+    # tool. Without this the #3794 repro ("Use the api_call tool to call Home
+    # Assistant GET /api/states") matched no domain, classified as low-signal,
+    # and the tool never reached the schema filter. Detect it explicitly so the
+    # "integrations" domain seeds api_call deterministically (see
+    # _DOMAIN_TOOL_MAP), independent of embedding retrieval.
+    if has(r"\bapi[ _]call\b", r"\bintegrations?\b",
+           r"\b(?:home ?assistant|miniflux|gitea|linkding|jellyfin)\b"):
+        domains.add("integrations")
 
     low_signal = not continuation and not domains
     return {
@@ -1004,8 +961,11 @@ def _recent_context_for_retrieval(messages: List[Dict], max_user: int = 3, max_c
         if isinstance(content, list):
             content = " ".join(b.get("text", "") for b in content if isinstance(b, dict))
         content = (content or "").strip()
-        # Skip injected tool-result envelopes — role=user but not human intent.
-        if not content or content.startswith("[Tool execution results]"):
+        # Skip injected envelopes — role=user but not human intent. Tool results
+        # are now wrapped via untrusted_context_message (metadata.trusted=False);
+        # keep the legacy "[Tool execution results]" prefix for older histories.
+        meta = msg.get("metadata") or {}
+        if not content or meta.get("trusted") is False or content.startswith("[Tool execution results]"):
             continue
         collected.append(content)
         if len(collected) >= max_user:
@@ -1024,6 +984,7 @@ def _build_system_prompt(
     compact: bool = False,
     owner: Optional[str] = None,
     suppress_local_context: bool = False,
+    suppress_skills: bool = False,
     active_email: Optional[Dict[str, str]] = None,
 ) -> List[Dict]:
     """Build agent system prompt, inject MCP/document context, merge consecutive system msgs."""
@@ -1041,7 +1002,7 @@ def _build_system_prompt(
         _ov_sig = _hl.sha256(_json.dumps(get_builtin_overrides() or {}, sort_keys=True).encode()).hexdigest()
     except Exception:
         _ov_sig = ""
-    cache_key = (frozenset(disabled_tools or []), bool(mcp_mgr), needs_admin, _rt_key, compact, _ov_sig, owner, suppress_local_context)
+    cache_key = (frozenset(disabled_tools or []), bool(mcp_mgr), needs_admin, _rt_key, compact, _ov_sig, owner, suppress_local_context, suppress_skills)
     if _cached_base_prompt and _cached_base_prompt_key == cache_key and not active_document:
         agent_prompt = _cached_base_prompt
         # Skill index is user-editable (name + description), so it must never
@@ -1051,6 +1012,7 @@ def _build_system_prompt(
             disabled_tools, mcp_mgr, needs_admin, relevant_tools,
             mcp_disabled_map=mcp_disabled_map, compact=compact, owner=owner,
             suppress_local_context=suppress_local_context,
+            suppress_skills=suppress_skills,
         )
     else:
         agent_prompt, _skill_index_block = _build_base_prompt(
@@ -1062,6 +1024,7 @@ def _build_system_prompt(
             compact=compact,
             owner=owner,
             suppress_local_context=suppress_local_context,
+            suppress_skills=suppress_skills,
         )
         if not active_document:
             _cached_base_prompt = agent_prompt
@@ -1090,8 +1053,8 @@ def _build_system_prompt(
     try:
         from src.user_time import current_datetime_context_message
         _datetime_message = current_datetime_context_message()
-    except Exception:
-        pass
+    except Exception as e:
+        logger.warning("Failed to build datetime context message", exc_info=e)
 
     # Document context is kept as a SEPARATE message (not merged into the tool
     # prompt) so the context trimmer doesn't destroy it when truncating the
@@ -1134,8 +1097,8 @@ def _build_system_prompt(
             try:
                 from src.pdf_form_doc import find_source_upload_id
                 _is_form_backed = bool(find_source_upload_id(active_document.current_content or ""))
-            except Exception:
-                pass
+            except Exception as e:
+                logger.warning("Failed to detect if document is form-backed, assuming plain", exc_info=e)
 
             if _is_form_backed:
                 doc_ctx = (
@@ -1345,7 +1308,7 @@ def _build_system_prompt(
     # few. If the teacher wrote a procedure for "open my X chat" last
     # time the student failed, this is where the student finds it
     # before deciding which tool to call.
-    if not suppress_local_context:
+    if not suppress_local_context and not suppress_skills:
         try:
             last_user = _extract_last_user_message(messages)
             # Respect the user's skills-enabled toggle (mirrors memory_enabled).
@@ -1512,6 +1475,7 @@ def _build_base_prompt(
     compact: bool = False,
     owner: Optional[str] = None,
     suppress_local_context: bool = False,
+    suppress_skills: bool = False,
 ):
     """Build the agent prompt with only relevant tools included.
 
@@ -1564,7 +1528,7 @@ def _build_base_prompt(
     # The caller wraps it in untrusted_context_message and ships it as a
     # user-role message — same treatment as the matched-skills block.
     skill_index_block = ""
-    if not suppress_local_context:
+    if not suppress_local_context and not suppress_skills:
         try:
             from services.memory.skills import SkillsManager
             from src.constants import DATA_DIR
@@ -1723,8 +1687,14 @@ def _append_tool_results(
         if round_reasoning:
             msg["reasoning_content"] = round_reasoning
         messages.append(msg)
+        # Tool output (shell/python stdout, file reads, fetched pages, email
+        # bodies, MCP results) is sourced from outside the server. Wrap it as
+        # untrusted data so prompt-injection inside a tool result is treated as
+        # data, not instructions — same hardening as skills (#788) and the
+        # web/RAG context. THREAT_MODEL.md lists tool output as a surface that
+        # must go through untrusted_context_message.
         messages.append(
-            {"role": "user", "content": f"[Tool execution results]\n\n{tool_output_text}"}
+            untrusted_context_message("tool execution results", tool_output_text)
         )
 
 
@@ -1983,6 +1953,7 @@ async def stream_agent_loop(
     approved_plan: Optional[str] = None,
     tool_policy: Optional[ToolPolicy] = None,
     workspace: Optional[str] = None,
+    forced_tools: Optional[Set[str]] = None,
     _is_teacher_run: bool = False,
 ) -> AsyncGenerator[str, None]:
     """Streaming agent loop generator.
@@ -2022,6 +1993,20 @@ async def stream_agent_loop(
     _needs_admin = _detect_admin_intent(messages)
     _last_user = _extract_last_user_message(messages)
     _intent = _classify_agent_request(messages, _last_user)
+    _low_signal_turn = bool(_intent.get("low_signal"))
+    _casual_low_signal_turn = _is_casual_low_signal(_last_user)
+    _direct_low_signal = (
+        _low_signal_turn
+        and not bool(_intent.get("continuation"))
+        and not plan_mode
+        and not approved_plan
+        and not guide_only
+        and (_casual_low_signal_turn or active_document is None)
+        and (_casual_low_signal_turn or not active_email)
+        and (_casual_low_signal_turn or not workspace)
+        and not forced_tools
+        and not relevant_tools
+    )
     # Tool retrieval uses the latest message by default. It may inherit recent
     # user turns only for explicit continuations ("yes", "do it", "1").
     _retrieval_query = str(_intent.get("retrieval_query") or _last_user)
@@ -2029,11 +2014,86 @@ async def stream_agent_loop(
         "[agent-intent] latest=%r continuation=%s low_signal=%s domains=%s retrieval_query=%r",
         _last_user[:120],
         bool(_intent.get("continuation")),
-        bool(_intent.get("low_signal")),
+        _low_signal_turn,
         sorted(_intent.get("domains") or []),
         _retrieval_query[:200],
     )
     _mcp_disabled_map = _load_mcp_disabled_map() if mcp_mgr else {}
+    if _direct_low_signal:
+        logger.info("[agent] direct low-signal reply path for latest=%r", _last_user[:80])
+        direct_messages = [{"role": "user", "content": _last_user}]
+        direct_response = ""
+        direct_start = time.time()
+        direct_actual_model = model
+        real_input_tokens = 0
+        real_output_tokens = 0
+        try:
+            async for chunk in stream_llm_with_fallback(
+                [(endpoint_url, model, headers)] + list(fallbacks or []),
+                direct_messages,
+                temperature=temperature,
+                max_tokens=min(max_tokens or 128, 128),
+                prompt_type=None,
+                tools=None,
+                timeout=int(get_setting("agent_stream_timeout_seconds", 300) or 300),
+                session_id=session_id,
+            ):
+                if chunk.startswith("data: ") and not chunk.startswith("data: [DONE]"):
+                    try:
+                        data = json.loads(chunk[6:])
+                    except json.JSONDecodeError:
+                        yield chunk
+                        continue
+                    if data.get("type") == "usage":
+                        usage = data.get("data", {}) or {}
+                        direct_actual_model = usage.get("model") or direct_actual_model
+                        real_input_tokens += usage.get("input_tokens", 0) or 0
+                        real_output_tokens += usage.get("output_tokens", 0) or 0
+                        continue
+                    if data.get("type") == "model_actual":
+                        direct_actual_model = data.get("model") or direct_actual_model
+                        data["requested_model"] = model
+                        yield f"data: {json.dumps(data)}\n\n"
+                        continue
+                    if data.get("type") == "fallback":
+                        direct_actual_model = data.get("answered_by") or direct_actual_model
+                        yield chunk
+                        continue
+                    if "delta" in data:
+                        if not data.get("thinking"):
+                            direct_response += data.get("delta", "")
+                        yield chunk
+                        continue
+                    yield chunk
+                elif chunk.startswith("event: "):
+                    yield chunk
+        except Exception as _direct_err:
+            logger.warning("[agent] direct low-signal path failed: %s", _direct_err)
+            fallback = "Hey."
+            direct_response += fallback
+            yield f"data: {json.dumps({'delta': fallback})}\n\n"
+
+        if not direct_response.strip():
+            fallback = "Hey."
+            direct_response = fallback
+            yield f"data: {json.dumps({'delta': fallback})}\n\n"
+
+        duration = time.time() - direct_start
+        metrics = {
+            "model": direct_actual_model,
+            "requested_model": model,
+            "input_tokens": real_input_tokens or estimate_tokens(direct_messages),
+            "output_tokens": real_output_tokens or max(len(direct_response) // 4, 1),
+            "total_time": round(duration, 2),
+            "response_time": round(duration, 2),
+            "agent_rounds": 0,
+            "tool_calls": 0,
+            "direct_low_signal": True,
+        }
+        yield f"data: {json.dumps({'type': 'metrics', 'data': metrics})}\n\n"
+        yield "data: [DONE]\n\n"
+        return
+
     if plan_mode and mcp_mgr:
         # Allow read-only MCP tools to investigate, block write/unknown ones:
         # hide them from the schemas AND reject them at runtime by qualified name.
@@ -2045,11 +2105,11 @@ async def stream_agent_loop(
 
     # RAG-based tool selection: retrieve relevant tools for this query.
     # If caller provided a pre-computed set (e.g. task_scheduler), use that.
-    _relevant_tools = set() if guide_only else relevant_tools
+    _relevant_tools = relevant_tools
     _t1 = time.time()
     if _relevant_tools:
         logger.info(f"[tool-rag] Using caller-provided relevant_tools ({len(_relevant_tools)} tools)")
-    if not guide_only and not _relevant_tools and bool(_intent.get("low_signal")):
+    if not guide_only and not _relevant_tools and _low_signal_turn:
         from src.tool_index import ALWAYS_AVAILABLE
         if workspace:
             # An active workspace IS the file-work signal: a vague "look at the
@@ -2140,6 +2200,15 @@ async def stream_agent_loop(
     if _relevant_tools is not None and active_document is not None:
         _relevant_tools.update({"edit_document", "update_document", "suggest_document"})
 
+    # Per-request UI toggles are stronger than retrieval. If the user turns on
+    # Search, the model must see the search tools even when the latest text is a
+    # typo or otherwise low-signal for tool RAG.
+    if not guide_only and forced_tools:
+        if _relevant_tools is None:
+            from src.tool_index import ALWAYS_AVAILABLE
+            _relevant_tools = set(ALWAYS_AVAILABLE)
+        _relevant_tools.update(t for t in forced_tools if t not in disabled_tools)
+
     # The skill index injected by _build_system_prompt tells the model to
     # call `manage_skills action=view`, and Jaccard-matched skills are pasted
     # into the prompt as procedures to follow — but neither path goes through
@@ -2147,7 +2216,7 @@ async def stream_agent_loop(
     # (grep, read_file, ...) that aren't in its schema list. Keep the schemas
     # in lockstep: manage_skills is callable whenever any skill is indexed,
     # and a matched skill's declared requires_toolsets ride along with it.
-    if not guide_only and _relevant_tools is not None:
+    if not guide_only and _relevant_tools is not None and not _low_signal_turn:
         try:
             from services.memory.skills import SkillsManager
             from src.constants import DATA_DIR
@@ -2212,7 +2281,7 @@ async def stream_agent_loop(
     _model_supports_tools = any(kw in _model_lc for kw in (
         "gpt-4", "gpt-5", "gpt-o", "claude", "gemini", "gemma",
         "qwen3", "qwen2.5", "mixtral", "mistral", "llama-3.1", "llama-3.2",
-        "llama-3.3", "llama-4",
+        "llama-3.3", "llama-4", "llama3.1", "llama3.2", "llama3.3", "llama4",
         # Local-served models that follow OpenAI-style function calling
         # via vLLM's `--enable-auto-tool-choice`. Belt-and-suspenders
         # with the per-endpoint flag above.
@@ -2254,13 +2323,15 @@ async def stream_agent_loop(
         _is_api_model = False
     else:
         _is_api_model = any(h in endpoint_url for h in _API_HOSTS) or _model_supports_tools
+    _compact_agent_prompt = _is_api_model or _is_ollama_native or _ollama_openai_compat
     messages, mcp_schemas = _build_system_prompt(
         messages, model, active_document, mcp_mgr, disabled_tools,
         needs_admin=_needs_admin, relevant_tools=_relevant_tools,
         mcp_disabled_map=_mcp_disabled_map,
-        compact=_is_api_model,
+        compact=_compact_agent_prompt,
         owner=owner,
         suppress_local_context=guide_only,
+        suppress_skills=_low_signal_turn,
         active_email=active_email,
     )
     if plan_mode and not guide_only:
@@ -2346,6 +2417,14 @@ async def stream_agent_loop(
     # Strip internal metadata keys before sending to the LLM API
     messages = [{k: v for k, v in msg.items() if k != "_protected"} for msg in messages]
 
+    agent_prompt_tokens = estimate_tokens(messages)
+    logger.info(
+        "[agent-timing] prep_done model=%s prompt_tokens=%s context_length=%s prep=%s",
+        model,
+        agent_prompt_tokens,
+        context_length,
+        {k: round(v, 3) for k, v in prep_timings.items()},
+    )
     yield f"data: {json.dumps({'type': 'agent_prep', 'data': {k: round(v, 3) for k, v in prep_timings.items()}})}\n\n"
 
     full_response = ""
@@ -2376,7 +2455,6 @@ async def stream_agent_loop(
     # signatures + consecutive no-text tool rounds to bail early.
     _recent_call_sigs = collections.deque(maxlen=6)
     _stuck_rounds = 0
-    _MAX_STUCK_ROUNDS = 4  # consecutive no-progress rounds before loop-breaker bails
     # Frequency of each exact call signature (tool + args), for the runaway
     # backstop. Counting identical repeats — not distinct same-tool calls —
     # lets a legit batch (e.g. 18 calendar events at once) through.
@@ -2491,6 +2569,19 @@ async def stream_agent_loop(
         # complementary cap for the rare stream that trickles bytes forever and
         # so never trips the inactivity timeout. Generous — only catches runaway.
         _round_deadline = time.time() + max(agent_stream_timeout * 4, 1200)
+        _round_start = time.time()
+        _round_first_event_logged = False
+        _round_first_token_logged = False
+        logger.info(
+            "[agent-timing] round_start round=%s model=%s endpoint=%s prompt_tokens=%s tools=%s native_tools=%s timeout=%s",
+            round_num,
+            model,
+            endpoint_url,
+            estimate_tokens(messages),
+            len(_tool_names_sent),
+            bool(all_tool_schemas),
+            agent_stream_timeout,
+        )
         async for chunk in stream_llm_with_fallback(
             _candidates,
             messages,
@@ -2501,11 +2592,30 @@ async def stream_agent_loop(
             timeout=agent_stream_timeout,
             session_id=session_id,
         ):
+            if not _round_first_event_logged:
+                _round_first_event_logged = True
+                logger.info(
+                    "[agent-timing] first_event round=%s elapsed=%.3fs kind=%s",
+                    round_num,
+                    time.time() - _round_start,
+                    "error" if chunk.startswith("event: error") else "data",
+                )
             if time.time() > _round_deadline:
-                logger.warning(f"[agent] round {round_num} stream exceeded wall-clock deadline; cutting off")
+                logger.warning(
+                    "[agent-timing] round_deadline round=%s elapsed=%.3fs deadline_s=%s",
+                    round_num,
+                    time.time() - _round_start,
+                    max(agent_stream_timeout * 4, 1200),
+                )
                 break
             # Forward error events from stream_llm to the frontend
             if chunk.startswith("event: error"):
+                logger.warning(
+                    "[agent-timing] stream_error round=%s elapsed=%.3fs chunk=%r",
+                    round_num,
+                    time.time() - _round_start,
+                    chunk[:500],
+                )
                 yield chunk
                 continue
             if chunk.startswith("data: ") and not chunk.startswith("data: [DONE]"):
@@ -2585,6 +2695,15 @@ async def stream_agent_loop(
                         if not first_token_received:
                             time_to_first_token = time.time() - total_start
                             first_token_received = True
+                        if not _round_first_token_logged:
+                            _round_first_token_logged = True
+                            logger.info(
+                                "[agent-timing] first_visible_token round=%s elapsed=%.3fs total_elapsed=%.3fs thinking=%s",
+                                round_num,
+                                time.time() - _round_start,
+                                time.time() - total_start,
+                                bool(data.get("thinking")),
+                            )
                         # Keep reasoning deltas in a separate accumulator so
                         # we can echo them back via `reasoning_content` on the
                         # next request (DeepSeek requires this; harmless for
@@ -2654,7 +2773,21 @@ async def stream_agent_loop(
                 yield chunk
             # Intercept [DONE] — don't forward until all rounds finish
 
-        tool_blocks, used_native = _resolve_tool_blocks(round_response, native_tool_calls, round_num, is_api_model=_is_api_model)
+        logger.info(
+            "[agent-timing] round_stream_done round=%s elapsed=%.3fs text_chars=%s tool_calls=%s first_event=%s first_token=%s",
+            round_num,
+            time.time() - _round_start,
+            len(round_response),
+            len(native_tool_calls),
+            _round_first_event_logged,
+            _round_first_token_logged,
+        )
+        tool_blocks, used_native = _resolve_tool_blocks(
+            round_response,
+            native_tool_calls,
+            round_num,
+            is_api_model=(_is_api_model and not guide_only),
+        )
 
         # Force-answer round: we told the model to STOP calling tools and
         # answer. If it ignored that and emitted a (possibly DSML) tool
@@ -2738,7 +2871,7 @@ async def stream_agent_loop(
         # model with no real native_tool_calls) must not be stripped from the
         # persisted text either — otherwise it streams once and then disappears
         # on reload (#3222 follow-up).
-        cleaned_round = strip_tool_blocks(round_response, skip_fenced=(_is_api_model and not used_native)).strip()
+        cleaned_round = strip_tool_blocks(round_response, skip_fenced=(_is_api_model and not used_native and not guide_only)).strip()
         round_texts.append(cleaned_round)
 
         if not tool_blocks:
@@ -2799,22 +2932,26 @@ async def stream_agent_loop(
             # promise: short response (<400 chars), no fenced code/answer,
             # and an action-intent phrase was matched. Long answers that
             # happen to contain "let me know" are not stalls.
-            _promise_shape = (
+            _looks_like_promise = (
                 not guide_only
                 and _intent_match is not None
                 and len(_intent_text) < 400
                 and "```" not in _intent_text
+                and _intent_nudge_count < _MAX_INTENT_NUDGES
             )
-            _looks_like_promise = _promise_shape and _intent_nudge_count < _MAX_INTENT_NUDGES
             if _looks_like_promise:
                 _intent_nudge_count += 1
                 _matched_phrase = _intent_match.group(0).strip()
-                # Don't log the matched phrase — it's raw model text that may
-                # carry credentials. Structural metadata only.
-                logger.info(
-                    "[agent] intent-without-action nudge #%d on round %d",
-                    _intent_nudge_count, round_num,
-                )
+                logger.info(f"[agent] intent-without-action nudge #{_intent_nudge_count} on round {round_num}: {_matched_phrase!r}")
+                _lower_phrase = _matched_phrase.lower()
+                _cookbook_log_hint = ""
+                if any(_word in _lower_phrase for _word in ("log", "logs", "output", "tail", "status")):
+                    _cookbook_log_hint = (
+                        " If this is about a Cookbook/model serve, the concrete calls are: "
+                        "`list_served_models` first, then `tail_serve_output` with the "
+                        "session_id from the serve/list result. Never answer with "
+                        "\"check logs\" when those tools are available."
+                    )
                 messages.append({
                     "role": "system",
                     "content": (
@@ -2823,6 +2960,7 @@ async def stream_agent_loop(
                         "see you announced the action but didn't run it, which "
                         "is the most frustrating thing you can do. "
                         "DO IT NOW: emit the actual function call this turn. "
+                        f"{_cookbook_log_hint}"
                         "If you decided not to do it after all, say so plainly in "
                         "one sentence instead of restating the plan."
                     ),
@@ -2830,24 +2968,6 @@ async def stream_agent_loop(
                 # Visible signal in the stream so the user knows we caught it.
                 yield f'data: {json.dumps({"type": "agent_step", "round": round_num + 1})}\n\n'
                 continue
-            # The model keeps announcing actions it never takes and we've spent
-            # every nudge — surface why the turn is ending instead of letting it
-            # look like a clean completion.
-            if _promise_shape and _intent_nudge_count >= _MAX_INTENT_NUDGES:
-                _matched_phrase = _intent_match.group(0).strip()
-                _matched_phrase_safe = _redact_sensitive_text(_matched_phrase)
-                _in_message = (
-                    f"Intent-nudge cap reached on round {round_num}: the model "
-                    f"announced an action ({_matched_phrase_safe!r}) without a tool call "
-                    f"after {_intent_nudge_count} nudge(s); ending the turn."
-                )
-                # Do not log the matched phrase, even redacted. It is raw model
-                # text and may contain credentials; keep logs structural only.
-                logger.warning(
-                    "[agent] intent-nudge cap exhausted on round %d (%d/%d)",
-                    round_num, _intent_nudge_count, _MAX_INTENT_NUDGES,
-                )
-                yield f'data: {json.dumps({"type": "intent_nudge_exhausted", "round": round_num, "nudges": _intent_nudge_count, "max_nudges": _MAX_INTENT_NUDGES, "message": _in_message})}\n\n'
             break  # no tools — done
 
         # ── Loop-breaker (Terminus-style stall detector) ──────────────
@@ -2880,23 +3000,10 @@ async def stream_agent_loop(
         # Distinct calls to one tool (a real batch) are legitimate work, so we
         # count identical call signatures, not raw per-tool-type totals.
         _runaway = _detect_runaway_call(_call_freq)
-        if _stuck_rounds >= _MAX_STUCK_ROUNDS or _runaway:
+        if _stuck_rounds >= 4 or _runaway:
             reason = (f"calling {_runaway} with identical arguments over and over" if _runaway
                       else "repeating the same tool calls without new progress")
-            _lb_message = (
-                f"Loop-breaker stopped the agent on round {round_num}: {reason}. "
-                "Forced one tool-free round to converge on an answer or state what's blocked."
-            )
-            # Log structural metadata only — `_sig` is raw tool-call content
-            # that may carry credentials.
-            logger.warning(
-                "[agent] loop-breaker tripped on round %d (%s); "
-                "stuck_rounds=%d/%d runaway=%r",
-                round_num, reason, _stuck_rounds, _MAX_STUCK_ROUNDS, _runaway,
-            )
-            # Surface the stop cause to the stream so the user (and journalctl)
-            # can tell a guard fired, not a clean completion.
-            yield f'data: {json.dumps({"type": "loop_breaker_triggered", "round": round_num, "reason": reason, "stuck_rounds": _stuck_rounds, "max_stuck_rounds": _MAX_STUCK_ROUNDS, "runaway": _runaway, "message": _lb_message})}\n\n'
+            logger.warning(f"[agent] loop-breaker tripped on round {round_num} ({reason}); sig={_sig[:80]!r}")
             # The model has been executing tools, so its results are already
             # in context. Force ONE tool-free round to converge: write the
             # answer from what it has, or state plainly what's blocking it.
@@ -2975,10 +3082,6 @@ async def stream_agent_loop(
                 cmd_display = block.content.split("\n")[0].strip()[:80]
             else:
                 cmd_display = block.content.strip()
-            # The display string is streamed (tool_start/tool_output) and persisted;
-            # redact any secrets in it. block.content itself is left untouched so
-            # tool execution still sees the real command.
-            cmd_display = _redact_sensitive_text(cmd_display)
 
             if tool_policy and tool_policy.blocks(block.tool_type):
                 desc = f"{block.tool_type}: BLOCKED"
@@ -3024,15 +3127,8 @@ async def stream_agent_loop(
                     evt = await _progress_q.get()
                     if evt is None:
                         break
-                    # Redact secrets in the live tail before streaming — the
-                    # final tool_output is redacted, so the progress tail must
-                    # be too, or a secret could flash by mid-run. Copy so we
-                    # don't mutate the tool's own event payload.
-                    _evt = dict(evt)
-                    if isinstance(_evt.get("tail"), str):
-                        _evt["tail"] = _redact_sensitive_text(_evt["tail"])
                     yield (
-                        f'data: {json.dumps({"type": "tool_progress", "tool": block.tool_type, "round": round_num, **_evt})}\n\n'
+                        f'data: {json.dumps({"type": "tool_progress", "tool": block.tool_type, "round": round_num, **evt})}\n\n'
                     )
                 desc, result = await _tool_task
 
@@ -3098,7 +3194,7 @@ async def stream_agent_loop(
                                 result["results"] = _clean
                             elif "stdout" in result:
                                 result["stdout"] = _clean
-                        except Exception:
+                        except (json.JSONDecodeError, Exception):
                             pass
 
             # Emit doc-specific event for document tools — the frontend
@@ -3167,29 +3263,29 @@ async def stream_agent_loop(
                 # empty) stdout/stderr; fall back to the error so the "timed
                 # out" reason reaches the UI instead of a blank result.
                 raw = result["stdout"] or result["stderr"] or result.get("error", "")
-                output_text = _truncate(_redact_sensitive_text(raw))
+                output_text = _truncate(raw)
             elif "output" in result:
                 # bash / python canonical result: {"output": ..., "exit_code": ...}
                 raw = result["output"] or ""
-                output_text = _truncate(_redact_sensitive_text(raw))
+                output_text = _truncate(raw)
             elif "response" in result:
                 # AI interaction tools (chat_with_model, send_to_session)
                 label = result.get("model", result.get("session_name", "AI"))
-                output_text = _truncate(_redact_sensitive_text(f"{label}: {result['response']}"))
+                output_text = _truncate(f"{label}: {result['response']}")
             elif "content" in result:
-                output_text = _truncate(_redact_sensitive_text(result["content"]))
+                output_text = _truncate(result["content"])
             elif "results" in result:
-                output_text = _truncate(_redact_sensitive_text(result["results"]))
+                output_text = _truncate(result["results"])
             elif "session_id" in result and "name" in result:
                 output_text = f"Session created: {result['name']} (id: {result['session_id']})"
             elif "success" in result:
                 output_text = (
                     f"Written: {result.get('path', '')}"
                     if result["success"]
-                    else f"Error: {_redact_sensitive_text(result.get('error', ''))}"
+                    else f"Error: {result.get('error', '')}"
                 )
             elif "error" in result:
-                output_text = _truncate(_redact_sensitive_text(result["error"]))
+                output_text = _truncate(result["error"])
 
             # Emit tool_output (include ui_event data if present)
             tool_output_data = {"type": "tool_output", "tool": block.tool_type, "command": cmd_display, "output": output_text, "exit_code": result.get("exit_code")}

src/agent_runs.py

@@ -174,8 +174,20 @@ async def subscribe(session_id: str) -> AsyncGenerator[str, None]:
             next_seq += 1
         if run.status != "running":
             return
+        heartbeat_idx = 0
         while True:
-            seq, ev = await q.get()
+            try:
+                seq, ev = await asyncio.wait_for(q.get(), timeout=10.0)
+            except asyncio.TimeoutError:
+                # Keep slow local models/proxies alive while they prefill before
+                # the first token. SSE comments are ignored by the UI but reset
+                # browser/proxy idle timers, which prevents "empty response"
+                # disconnects on llama.cpp first-token latencies of 30s+.
+                if run.status == "running":
+                    heartbeat_idx += 1
+                    yield f": heartbeat {heartbeat_idx}\n\n"
+                    continue
+                seq, ev = (None, None)
             if seq is None:            # end sentinel
                 while next_seq < len(run.buffer):   # flush any tail the sentinel raced
                     yield run.buffer[next_seq]

src/agent_tools/__init__.py

@@ -22,6 +22,9 @@ from .subprocess_tools import BashTool, PythonTool
 from .web_tools import WebSearchTool, WebFetchTool
 from .filesystem_tools import ReadFileTool, WriteFileTool, EditFileTool, LsTool, GlobTool, GrepTool, GetWorkspaceTool
 from .document_tools import CreateDocumentTool, UpdateDocumentTool, EditDocumentTool, SuggestDocumentTool, ManageDocumentTool
+from .model_interaction_tools import ChatWithModelTool, AskTeacherTool, ListModelsTool
+from .bg_job_tools import ManageBgJobsTool
+from .session_tools import CreateSessionTool, ListSessionsTool, SendToSessionTool, ManageSessionTool
 
 TOOL_HANDLERS = {
     "bash": BashTool().execute,
@@ -40,6 +43,14 @@ TOOL_HANDLERS = {
     "suggest_document": SuggestDocumentTool().execute,
     "manage_documents": ManageDocumentTool().execute,
     "get_workspace": GetWorkspaceTool().execute,
+    "chat_with_model": ChatWithModelTool().execute,
+    "ask_teacher": AskTeacherTool().execute,
+    "list_models": ListModelsTool().execute,
+    "manage_bg_jobs": ManageBgJobsTool().execute,
+    "create_session": CreateSessionTool().execute,
+    "list_sessions": ListSessionsTool().execute,
+    "send_to_session": SendToSessionTool().execute,
+    "manage_session": ManageSessionTool().execute,
 }
 
 # ---------------------------------------------------------------------------
@@ -52,7 +63,7 @@ PYTHON_TIMEOUT = 30
 
 # Tool types that trigger execution
 TOOL_TAGS = {"bash", "python", "web_search", "web_fetch", "read_file", "write_file", "edit_file",
-             "grep", "glob", "ls", "get_workspace",
+             "grep", "glob", "ls", "get_workspace", "manage_bg_jobs",
              "create_document", "update_document", "edit_document",
              "search_chats",
              "chat_with_model", "create_session", "list_sessions",

src/agent_tools/bg_job_tools.py

@@ -0,0 +1,98 @@
+"""Agent tool to inspect and control detached background `bash` jobs.
+
+`bash` blocks prefixed with a `#!bg` marker run detached via `src.bg_jobs`; the
+agent is auto-re-invoked with the output when they finish. This tool covers the
+gaps in that flow: list the jobs in the current chat, read a still-running job's
+output on demand, and kill a runaway job instead of waiting out its max-runtime.
+
+Registry tool (`TOOL_HANDLERS["manage_bg_jobs"]`). Jobs are scoped to the chat
+that launched them, so every action requires the caller's `session_id` and a job
+from another session is treated as not found.
+"""
+
+import json
+import time
+from typing import Any, Dict, List
+
+_LIST_ACTIONS = {"list", "ls", "jobs"}
+_OUTPUT_ACTIONS = {"output", "get", "read", "tail", "status", "show"}
+_KILL_ACTIONS = {"kill", "stop", "cancel", "terminate"}
+
+
+def _age(rec: Dict[str, Any]) -> str:
+    start = rec.get("started_at")
+    if not start:
+        return "?"
+    secs = int(time.time() - start)
+    if secs < 60:
+        return f"{secs}s"
+    if secs < 3600:
+        return f"{secs // 60}m"
+    return f"{secs // 3600}h{(secs % 3600) // 60}m"
+
+
+def _status_label(rec: Dict[str, Any]) -> str:
+    status = rec.get("status", "?")
+    if rec.get("killed"):
+        return "killed"
+    if rec.get("timed_out"):
+        return "timed out"
+    if rec.get("died"):
+        return "died"
+    if status in ("done", "failed"):
+        return f"{status} (exit {rec.get('exit_code')})"
+    return status
+
+
+def _row(rec: Dict[str, Any]) -> str:
+    cmd = (rec.get("command") or "").strip().splitlines()[0][:80]
+    return f"[{rec.get('id')}] {_status_label(rec)} | {_age(rec)} | {cmd}"
+
+
+class ManageBgJobsTool:
+    async def execute(self, content: str, ctx: dict) -> dict:
+        from src import bg_jobs
+
+        session_id = ctx.get("session_id")
+        raw = (content or "").strip()
+        try:
+            args = json.loads(raw) if raw else {}
+        except (ValueError, TypeError):
+            args = {}
+        if not isinstance(args, dict):
+            args = {}
+        action = str(args.get("action", "list")).strip().lower()
+        job_id = str(args.get("job_id") or args.get("id") or "").strip()
+
+        if not session_id:
+            return {"error": "manage_bg_jobs: no active chat session; background jobs are scoped to a chat.", "exit_code": 1}
+
+        if action in _LIST_ACTIONS:
+            jobs: List[Dict[str, Any]] = bg_jobs.list_for_session(session_id)
+            if not jobs:
+                return {"output": "No background jobs in this chat.", "exit_code": 0}
+            jobs.sort(key=lambda r: r.get("started_at") or 0, reverse=True)
+            lines = "\n".join(_row(r) for r in jobs)
+            return {"output": f"{len(jobs)} background job(s):\n{lines}", "exit_code": 0}
+
+        if action in _OUTPUT_ACTIONS or action in _KILL_ACTIONS:
+            if not job_id:
+                return {"error": f"manage_bg_jobs: action '{action}' requires a job_id (see action='list').", "exit_code": 1}
+            rec = bg_jobs.get(job_id)
+            # Scope: only the chat that launched a job may see or control it.
+            if rec is None or rec.get("session_id") != session_id:
+                return {"error": f"manage_bg_jobs: no background job '{job_id}' in this chat.", "exit_code": 1}
+
+            if action in _KILL_ACTIONS:
+                if rec.get("status") != "running":
+                    return {"output": f"Job `{job_id}` already {_status_label(rec)}; nothing to kill.", "exit_code": 0}
+                killed = bg_jobs.kill(job_id)
+                return {"output": f"Killed background job `{job_id}` ({(killed or {}).get('command', '').splitlines()[0][:80]}).", "exit_code": 0}
+
+            out = rec.get("output") or "(no output yet)"
+            return {
+                "output": f"Job `{job_id}` [{_status_label(rec)}, {_age(rec)}]\nCommand: {rec.get('command')}\n\nOutput:\n{out}",
+                "exit_code": 0,
+            }
+
+        return {"error": f"manage_bg_jobs: unknown action '{action}'. Use list, output, or kill.", "exit_code": 1}

src/agent_tools/filesystem_tools.py

@@ -1,6 +1,7 @@
 import asyncio
 import json
 import os
+import re
 import difflib
 import fnmatch
 import shutil
@@ -16,6 +17,31 @@ _CODENAV_SKIP_DIRS = frozenset({
 _CODENAV_MAX_HITS = 200
 _CODENAV_MAX_LINE = 400
 
+
+def _glob_to_regex(pat: str) -> "re.Pattern":
+    """Translate a forward-slash glob (**, *, ?) into a compiled regex.
+    `**/` matches zero or more complete directories.
+    `*` matches within a single path segment (does not cross /).
+    """
+    i, n, out = 0, len(pat), []
+    while i < n:
+        if pat[i : i + 3] == "**/":
+            out.append("(?:[^/]+/)*")
+            i += 3
+        elif pat[i : i + 2] == "**":
+            out.append(".*")
+            i += 2
+        elif pat[i] == "*":
+            out.append("[^/]*")
+            i += 1
+        elif pat[i] == "?":
+            out.append("[^/]")
+            i += 1
+        else:
+            out.append(re.escape(pat[i]))
+            i += 1
+    return re.compile("".join(out))
+
 def _unified_diff(old: str, new: str, path: str) -> Optional[Dict[str, Any]]:
     if old == new:
         return None
@@ -259,23 +285,38 @@ class GlobTool:
             return {"error": f"glob: {e}", "exit_code": 1}
 
         def _glob():
-            from pathlib import Path
-            base = Path(root)
-            if not base.is_dir():
+            base = os.path.abspath(root)
+            if not os.path.isdir(base):
                 return None, f"glob: {root}: not a directory"
+            norm_pat = pattern.replace("\\", "/")
+            # Fast path: literal pattern (no wildcards) → direct path lookup.
+            if not any(c in norm_pat for c in "*?["):
+                cand = os.path.normpath(os.path.join(base, norm_pat))
+                if os.path.exists(cand):
+                    return [cand], None
+                # Literal not at exact path — fall through to walk so
+                # e.g. "foo.py" still matches at any depth (like rglob).
+            # Compile glob to regex: * stays within one segment, **/ spans dirs.
+            regex = _glob_to_regex(norm_pat)
             matched = []
+            cap = _CODENAV_MAX_HITS * 5
             try:
-                for p in base.rglob(pattern):
-                    if set(p.relative_to(base).parts) & _CODENAV_SKIP_DIRS:
-                        continue
-                    try:
-                        mtime = p.stat().st_mtime
-                    except OSError:
-                        mtime = 0
-                    matched.append((mtime, str(p)))
-                    if len(matched) > _CODENAV_MAX_HITS * 5:
+                for dp, dns, fns in os.walk(base):
+                    # Prune skipped dirs before descending (unlike rglob which
+                    # descends first then filters — fatal on large node_modules).
+                    dns[:] = [d for d in dns if d not in _CODENAV_SKIP_DIRS]
+                    for name in fns + dns:
+                        full = os.path.join(dp, name)
+                        rel = os.path.relpath(full, base).replace(os.sep, "/")
+                        if regex.fullmatch(rel) or regex.fullmatch(name):
+                            try:
+                                mtime = os.stat(full).st_mtime
+                            except OSError:
+                                mtime = 0
+                            matched.append((mtime, full))
+                    if len(matched) > cap:
                         break
-            except (OSError, ValueError) as _e:
+            except OSError as _e:
                 return None, f"glob: {_e}"
             matched.sort(key=lambda t: t[0], reverse=True)
             return [pth for _, pth in matched[:_CODENAV_MAX_HITS]], None

src/agent_tools/model_interaction_tools.py

@@ -0,0 +1,208 @@
+"""model_interaction_tools.py - agent tools for talking to other models.
+
+Owns the model-interaction tool implementations (chat_with_model, ask_teacher,
+list_models) and their handler classes, registered in ``TOOL_HANDLERS``. Part
+of the tool -> registry migration (#3629): the implementations were moved here
+out of ``src.ai_interaction`` so dispatch flows through the registry instead of
+the elif chain / dispatch_ai_tool in tool_execution.py.
+
+Shared helpers that still live in ``src.ai_interaction`` and are used by tools
+not yet migrated (``_resolve_model``, ``AI_CHAT_TIMEOUT``) are imported lazily
+inside the functions to avoid an import cycle at module load.
+"""
+import logging
+from typing import Dict, Optional
+
+logger = logging.getLogger(__name__)
+
+
+_TEACHER_SYSTEM_PROMPT = (
+    "You are a senior AI mentor. A less capable model is stuck on a problem and asking for help. "
+    "Provide clear, actionable guidance:\n"
+    "1. Brief analysis of the problem\n"
+    "2. Recommended approach (step by step)\n"
+    "3. Key things to watch out for\n\n"
+    "Be concise and practical. No preamble."
+)
+
+
+async def chat_with_model(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
+    """Send a message to a specific model and return its response.
+
+    Content format:
+      Line 1: model_name (or model_name@endpoint_name)
+      Line 2+: the message to send
+    """
+    from src.ai_interaction import _resolve_model, AI_CHAT_TIMEOUT
+    from src.llm_core import llm_call_async
+
+    lines = content.strip().split("\n", 1)
+    if not lines or not lines[0].strip():
+        return {"error": "First line must be the model name"}
+
+    model_spec = lines[0].strip()
+    message = lines[1].strip() if len(lines) > 1 else ""
+    if not message:
+        return {"error": "No message provided (line 2+ is the message)"}
+
+    try:
+        url, model, headers = _resolve_model(model_spec, owner=owner)
+    except ValueError as e:
+        return {"error": str(e)}
+
+    try:
+        response = await llm_call_async(
+            url, model,
+            [{"role": "user", "content": message}],
+            headers=headers,
+            timeout=AI_CHAT_TIMEOUT,
+        )
+        # Truncate very long responses
+        if len(response) > 10000:
+            response = response[:10000] + "\n... (truncated)"
+        return {"model": model, "response": response}
+    except Exception as e:
+        logger.error(f"chat_with_model failed: {e}")
+        return {"error": f"Failed to get response from {model_spec}: {e}"}
+
+
+async def ask_teacher(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
+    """Ask a more capable model for help.
+
+    Content format:
+      Line 1: model_name (or 'auto')
+      Line 2+: the problem description
+    """
+    from src.ai_interaction import _resolve_model, AI_CHAT_TIMEOUT
+    from src.llm_core import llm_call_async
+    from src.settings import get_setting
+
+    lines = content.strip().split("\n", 1)
+    model_spec = lines[0].strip() if lines else "auto"
+    problem = lines[1].strip() if len(lines) > 1 else ""
+
+    if not problem:
+        return {"error": "No problem description provided"}
+
+    if model_spec.lower() in ("auto", ""):
+        model_spec = get_setting("teacher_model", "")
+        if not model_spec:
+            return {"error": "No teacher model configured. Specify a model name or set teacher_model in settings."}
+
+    try:
+        url, model, headers = _resolve_model(model_spec, owner=owner)
+    except ValueError as e:
+        return {"error": str(e)}
+
+    try:
+        response = await llm_call_async(
+            url, model,
+            [
+                {"role": "system", "content": _TEACHER_SYSTEM_PROMPT},
+                {"role": "user", "content": f"Problem:\n{problem}"},
+            ],
+            headers=headers,
+            timeout=AI_CHAT_TIMEOUT,
+        )
+        if len(response) > 8000:
+            response = response[:8000] + "\n... (truncated)"
+        return {"model": model, "response": response, "teacher": True}
+    except Exception as e:
+        logger.error(f"ask_teacher failed: {e}")
+        return {"error": f"Teacher call failed ({model_spec}): {e}"}
+
+
+async def list_models(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
+    """List all available models across configured endpoints.
+
+    Content = optional filter keyword.
+    """
+    import json
+    import httpx
+    from src.database import SessionLocal, ModelEndpoint
+    from src.llm_core import _detect_provider, ANTHROPIC_MODELS
+    from src.auth_helpers import owner_filter
+    from src.endpoint_resolver import resolve_endpoint_runtime, build_headers, build_models_url
+
+    keyword = content.strip().lower() if content.strip() else None
+
+    db = SessionLocal()
+    try:
+        query = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True)
+        if owner:
+            query = owner_filter(query, ModelEndpoint, owner)
+        endpoints = query.all()
+        if not endpoints:
+            return {"results": "No enabled model endpoints configured."}
+
+        result_lines = []
+        total_models = 0
+
+        for ep in endpoints:
+            try:
+                base, api_key = resolve_endpoint_runtime(ep, owner=owner)
+            except Exception:
+                continue
+            provider = _detect_provider(base)
+            headers = build_headers(api_key, base)
+
+            model_ids = []
+            if provider == "anthropic":
+                model_ids = list(ANTHROPIC_MODELS)
+            else:
+                try:
+                    models_url = build_models_url(base)
+                    if models_url:
+                        r = httpx.get(models_url, headers=headers, timeout=5)
+                        r.raise_for_status()
+                        data = r.json()
+                        model_ids = [m.get("id") for m in (data.get("data") or []) if m.get("id")]
+                        if not model_ids:
+                            model_ids = [
+                                m.get("name") or m.get("model")
+                                for m in (data.get("models") or [])
+                                if m.get("name") or m.get("model")
+                            ]
+                    else:
+                        model_ids = json.loads(ep.cached_models or "[]")
+                except Exception:
+                    model_ids = ["(endpoint offline)"]
+
+            if keyword:
+                model_ids = [m for m in model_ids if keyword in m.lower() or keyword in (ep.name or "").lower()]
+
+            if model_ids:
+                result_lines.append(f"\n**{ep.name or base}** ({provider}):")
+                for mid in model_ids:
+                    result_lines.append(f"  - `{mid}`")
+                    total_models += 1
+
+        if not result_lines:
+            return {"results": "No models found" + (f" matching '{keyword}'" if keyword else "") + "."}
+
+        header = f"Available models ({total_models} total):"
+        return {"results": header + "\n".join(result_lines)}
+    except Exception as e:
+        logger.error(f"list_models failed: {e}")
+        return {"error": str(e)}
+    finally:
+        db.close()
+
+
+# ---------------------------------------------------------------------------
+# Handler classes registered in TOOL_HANDLERS
+# ---------------------------------------------------------------------------
+
+class ChatWithModelTool:
+    async def execute(self, content: str, ctx: dict) -> Dict:
+        return await chat_with_model(content, ctx.get("session_id"), owner=ctx.get("owner"))
+
+
+class AskTeacherTool:
+    async def execute(self, content: str, ctx: dict) -> Dict:
+        return await ask_teacher(content, ctx.get("session_id"), owner=ctx.get("owner"))
+
+
+class ListModelsTool:
+    async def execute(self, content: str, ctx: dict) -> Dict:
+        return await list_models(content, ctx.get("session_id"), owner=ctx.get("owner"))

src/agent_tools/session_tools.py

@@ -0,0 +1,464 @@
+"""session_tools.py - agent tools for AI-to-AI session management.
+
+Owns create_session, list_sessions, send_to_session and manage_session, moved
+out of src.ai_interaction as part of the tool -> registry migration (#3629), and
+their handler classes registered in TOOL_HANDLERS.
+
+The session manager is a runtime-set singleton in src.ai_interaction, so each
+function fetches it via get_session_manager() (imported here); _resolve_model and
+AI_CHAT_TIMEOUT are reused from there too.
+"""
+import json
+import logging
+import uuid
+from typing import Dict, Optional
+
+from src.ai_interaction import get_session_manager, _resolve_model, AI_CHAT_TIMEOUT
+
+logger = logging.getLogger(__name__)
+
+
+async def create_session(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
+    """Create a new chat session.
+
+    Content format:
+      Line 1: session name
+      Line 2: model_name (or model_name@endpoint_name)
+    """
+    _session_manager = get_session_manager()
+    if not _session_manager:
+        return {"error": "Session manager not available"}
+
+    lines = content.strip().split("\n")
+    if len(lines) < 2:
+        return {"error": "Need 2 lines: session name, then model spec"}
+
+    name = lines[0].strip()
+    model_spec = lines[1].strip()
+
+    if not name:
+        return {"error": "Session name cannot be empty"}
+
+    try:
+        url, model, headers = _resolve_model(model_spec, owner=owner)
+    except ValueError as e:
+        return {"error": str(e)}
+
+    sid = str(uuid.uuid4())[:8]
+    try:
+        _session_manager.create_session(
+            session_id=sid,
+            name=name,
+            endpoint_url=url,
+            model=model,
+            rag=False,
+            owner=owner,
+        )
+        # Store headers on session for future calls
+        sess = _session_manager.get_session(sid)
+        if sess and headers:
+            sess.headers = headers
+        try:
+            from src.event_bus import fire_event
+            fire_event("session_created", owner)
+        except Exception:
+            logger.debug("session_created event dispatch failed", exc_info=True)
+
+        return {"session_id": sid, "name": name, "model": model, "endpoint_url": url}
+    except Exception as e:
+        logger.error(f"create_session failed: {e}")
+        return {"error": f"Failed to create session: {e}"}
+
+async def list_sessions(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
+    """List sessions sorted by most-recently-active first.
+
+    Output includes a relative "last active" timestamp per row so the
+    agent can answer "open my last chat" without guessing from titles.
+    The most-recent session is always first in the list.
+
+    Content = optional filter keyword (matches session name).
+    """
+    _session_manager = get_session_manager()
+    if not _session_manager:
+        return {"error": "Session manager not available"}
+
+    keyword = content.strip().lower() if content.strip() else None
+
+    try:
+        from core.database import SessionLocal, Session as DbSession
+        from datetime import datetime, timezone
+
+        # Pull every session's last_accessed from the DB so we can sort
+        # by recency. In-memory sessions hold name + model + msg_count;
+        # the DB row holds the timestamps.
+        db = SessionLocal()
+        try:
+            db_rows = {r.id: r for r in db.query(DbSession).all()}
+        finally:
+            db.close()
+
+        # SECURITY: scope to the caller's sessions. Passing None returned
+        # every user's sessions, which the agent tool then exposed via the
+        # "list my chats" reply.
+        sessions = _session_manager.get_sessions_for_user(owner)
+        rows = []
+        for sid, sess in sessions.items():
+            if keyword and keyword not in (sess.name or "").lower():
+                continue
+            db_row = db_rows.get(sid)
+            # Prefer last_accessed; fall back to updated_at, then created_at.
+            ts = None
+            if db_row:
+                ts = getattr(db_row, 'last_accessed', None) or getattr(db_row, 'updated_at', None) or getattr(db_row, 'created_at', None)
+            rows.append((ts, sid, sess))
+
+        # Sort by timestamp DESC; rows without a timestamp sink to the bottom.
+        rows.sort(key=lambda r: r[0] or datetime.min, reverse=True)
+
+        def _rel(ts):
+            if not ts:
+                return 'never'
+            now = datetime.utcnow()
+            try:
+                if ts.tzinfo is not None:
+                    now = datetime.now(timezone.utc)
+                diff = (now - ts).total_seconds()
+            except Exception:
+                return 'unknown'
+            if diff < 60: return 'just now'
+            if diff < 3600: return f'{int(diff / 60)}m ago'
+            if diff < 86400: return f'{int(diff / 3600)}h ago'
+            if diff < 86400 * 7: return f'{int(diff / 86400)}d ago'
+            return ts.strftime('%Y-%m-%d')
+
+        lines = []
+        for i, (ts, sid, sess) in enumerate(rows):
+            if i >= 50:
+                lines.append(f"... and {len(rows) - 50} more (showing first 50)")
+                break
+            safe_name = (sess.name or "Untitled").replace("[", "\\[").replace("]", "\\]")
+            msg_count = getattr(sess, "message_count", 0) or 0
+            model = getattr(sess, "model", "unknown")
+            marker = " ← most recent" if i == 0 else ""
+            lines.append(f"- **[{safe_name}](#session-{sid})** (id: `{sid}`, model: {model}, {msg_count} msgs, last active {_rel(ts)}){marker}")
+
+        if not lines:
+            return {"results": "No sessions found" + (f" matching '{keyword}'" if keyword else "") + "."}
+
+        return {
+            "results": (
+                f"Found {len(rows)} session(s), sorted most-recent first:\n"
+                + "\n".join(lines)
+                + "\n\nAssistant: when replying to the user, preserve the chat-title markdown links exactly as shown, e.g. `[Chat](#session-id)`. Do not rewrite this as a plain, non-clickable table."
+            )
+        }
+    except Exception as e:
+        logger.error(f"list_sessions failed: {e}")
+        return {"error": str(e)}
+
+async def send_to_session(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
+    """Send a message to an existing session and get a response.
+
+    Content format:
+      Line 1: session_id
+      Line 2+: message
+    """
+    _session_manager = get_session_manager()
+    from src.llm_core import llm_call_async
+    from core.models import ChatMessage
+
+    if not _session_manager:
+        return {"error": "Session manager not available"}
+
+    lines = content.strip().split("\n", 1)
+    if len(lines) < 2:
+        return {"error": "Need 2 lines: session_id, then message"}
+
+    target_sid = lines[0].strip()
+    message = lines[1].strip()
+
+    sess = _session_manager.get_session(target_sid)
+    if not sess:
+        return {"error": f"Session '{target_sid}' not found"}
+
+    # Owner-scope: reject access to another user's session
+    if owner and getattr(sess, "owner", None) and sess.owner != owner:
+        return {"error": f"Session '{target_sid}' not found"}
+
+    if not message:
+        return {"error": "No message provided"}
+
+    try:
+        # Build context from session history
+        context = sess.get_context_messages()
+        context.append({"role": "user", "content": message})
+
+        response = await llm_call_async(
+            sess.endpoint_url, sess.model, context,
+            headers=sess.headers,
+            timeout=AI_CHAT_TIMEOUT,
+        )
+
+        # Save both messages to session
+        sess.add_message(ChatMessage("user", message))
+        sess.add_message(ChatMessage("assistant", response))
+
+        # Truncate for tool output
+        if len(response) > 10000:
+            response = response[:10000] + "\n... (truncated)"
+
+        return {
+            "session_id": target_sid,
+            "session_name": sess.name,
+            "response": response,
+        }
+    except Exception as e:
+        logger.error(f"send_to_session failed: {e}")
+        return {"error": f"Failed to send to session: {e}"}
+
+async def manage_session(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
+    """Manage sessions: rename, archive, delete, important, truncate, fork.
+
+    Content format:
+      Line 1: action (rename|archive|unarchive|delete|important|unimportant|truncate|fork)
+      Line 2: target session_id (or "current" to use the active session)
+      Line 3+: action-specific params (e.g. new name for rename, keep_count for truncate)
+    """
+    _session_manager = get_session_manager()
+    if not _session_manager:
+        return {"error": "Session manager not available"}
+
+    from src.database import SessionLocal, Session as DbSession
+
+    # Accept BOTH the structured JSON args the tool schema advertises
+    # ({action, session_id, value}) AND the legacy line-based format
+    # (line1=action, line2=session_id, line3=value). Native function-calling
+    # models send JSON; fenced-block callers send lines. Previously only the
+    # line format was parsed, so a model that followed the schema (JSON) got
+    # "Need at least 2 lines" / "Rename needs line 3" and couldn't drive it.
+    _raw = (content or "").strip()
+    action = ""
+    target_sid = ""
+    value = None      # the action param: new name (rename) / keep_count (truncate, fork)
+    _list_filter = ""
+    _parsed = None
+    if _raw.startswith("{"):
+        try:
+            _parsed = json.loads(_raw)
+        except Exception:
+            _parsed = None
+    if isinstance(_parsed, dict):
+        action = str(_parsed.get("action") or "").strip().lower()
+        target_sid = str(_parsed.get("session_id") or _parsed.get("session") or _parsed.get("id") or "").strip()
+        _v = _parsed.get("value")
+        if _v is None:
+            _v = (_parsed.get("name") or _parsed.get("new_name")
+                  or _parsed.get("title") or _parsed.get("keep_count"))
+        value = None if _v is None else str(_v).strip()
+        _list_filter = str(_parsed.get("filter") or "").strip()
+    else:
+        lines = _raw.split("\n")
+        if not lines or not lines[0].strip():
+            return {"error": "Missing action (rename|archive|delete|important|truncate|fork|list|switch)"}
+        action = lines[0].strip().lower()
+        target_sid = lines[1].strip() if len(lines) >= 2 else ""
+        value = lines[2].strip() if len(lines) >= 3 else None
+        _list_filter = "\n".join(lines[1:]).strip()
+
+    if not action:
+        return {"error": "Missing action (rename|archive|delete|important|truncate|fork|list|switch)"}
+
+    # `list` alias - dispatch to list_sessions so the agent's natural
+    # first guess (every other manage_* tool has a `list` action) works.
+    if action == "list":
+        return await list_sessions(_list_filter, session_id, owner=owner)
+
+    if not target_sid:
+        return {"error": "Need a session_id (or 'current' for the active chat)"}
+
+    # Allow "current" to refer to the active session
+    if target_sid.lower() == "current" and session_id:
+        target_sid = session_id
+
+    # `switch` / `open` / `select` / `view` - the agent reaches for
+    # these when the user asks to "open" or "switch to" a session.
+    # There's no server-side way to make the browser navigate, so we
+    # just return a clickable anchor link the user can click. The
+    # frontend's chat-history click delegate routes `#session-<id>`
+    # to selectSession(). The agent's reply naturally embeds this
+    # result so the user sees a single clickable line.
+    def _session_query(db):
+        query = db.query(DbSession).filter(DbSession.id == target_sid)
+        if owner is not None:
+            query = query.filter(DbSession.owner == owner)
+        return query
+
+    if action in ("switch", "open", "select", "view"):
+        db = SessionLocal()
+        try:
+            db_sess = _session_query(db).first()
+            if not db_sess:
+                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
+            name = db_sess.name or target_sid
+        finally:
+            db.close()
+        return {
+            "action": action,
+            "session_id": target_sid,
+            "name": name,
+            "results": f"[{name}](#session-{target_sid}) - click to open.",
+        }
+
+    db = SessionLocal()
+    try:
+        if action == "rename":
+            if not value:
+                return {"error": "rename needs a new name (the `value` arg, or line 3 in the legacy format)"}
+            new_name = value
+            db_sess = _session_query(db).first()
+            if not db_sess:
+                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
+            db_sess.name = new_name
+            db.commit()
+            _session_manager.update_session_name(target_sid, new_name)
+            return {"action": "rename", "session_id": target_sid, "name": new_name,
+                    "results": f"Session renamed to '{new_name}'"}
+
+        elif action == "archive":
+            db_sess = _session_query(db).first()
+            if not db_sess:
+                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
+            db_sess.archived = True
+            db.commit()
+            return {"action": "archive", "session_id": target_sid,
+                    "results": f"Session '{db_sess.name}' archived"}
+
+        elif action == "unarchive":
+            db_sess = _session_query(db).first()
+            if not db_sess:
+                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
+            db_sess.archived = False
+            db.commit()
+            return {"action": "unarchive", "session_id": target_sid,
+                    "results": f"Session '{db_sess.name}' unarchived"}
+
+        elif action == "delete":
+            if target_sid == session_id:
+                return {"error": "Cannot delete the current session while chatting in it. Delete other sessions first."}
+            db_sess = _session_query(db).first()
+            if not db_sess:
+                return {"error": f"Session '{target_sid}' not found. Refusing to delete an unknown chat id; use the exact id from list_sessions."}
+            if db_sess and db_sess.is_important:
+                return {"error": f"Session '{db_sess.name}' is starred/favorited. Unstar it first before deleting."}
+            try:
+                ok = _session_manager.delete_session(target_sid)
+                if not ok:
+                    return {"error": f"Session '{target_sid}' was not deleted because it no longer exists."}
+                return {"action": "delete", "session_id": target_sid,
+                        "results": f"Session '{db_sess.name or target_sid}' deleted"}
+            except Exception as e:
+                return {"error": f"Failed to delete session: {e}"}
+
+        elif action in ("important", "unimportant"):
+            is_important = action == "important"
+            db_sess = _session_query(db).first()
+            if not db_sess:
+                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
+            # Prevent AI from unstarring sessions - only the user can do that manually
+            if not is_important and db_sess.is_important:
+                return {"error": f"Session '{db_sess.name}' is starred by the user. Only the user can unstar sessions manually."}
+            db_sess.is_important = is_important
+            db.commit()
+            status = "marked as important" if is_important else "unmarked as important"
+            return {"action": action, "session_id": target_sid,
+                    "results": f"Session '{db_sess.name}' {status}"}
+
+        elif action == "truncate":
+            db_sess = _session_query(db).first()
+            if not db_sess:
+                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
+            keep_count = 10
+            if value:
+                try:
+                    keep_count = int(value)
+                except ValueError:
+                    pass
+            success = _session_manager.truncate_messages(target_sid, keep_count)
+            if success:
+                return {"action": "truncate", "session_id": target_sid,
+                        "results": f"Session truncated to last {keep_count} messages"}
+            return {"error": f"Failed to truncate session '{target_sid}'"}
+
+        elif action == "fork":
+            db_sess = _session_query(db).first()
+            if not db_sess:
+                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
+            keep_count = 0  # 0 = all messages
+            if value:
+                try:
+                    keep_count = int(value)
+                except ValueError:
+                    pass
+
+            source = _session_manager.get_session(target_sid)
+            if not source:
+                return {"error": f"Session '{target_sid}' not found"}
+
+            new_sid = str(uuid.uuid4())[:8]
+            _session_manager.create_session(
+                session_id=new_sid,
+                name=f"Fork: {source.name}",
+                endpoint_url=source.endpoint_url,
+                model=source.model,
+                rag=False,
+                owner=owner,
+            )
+            # Copy messages
+            history = source.get_context_messages()
+            if keep_count > 0:
+                history = history[:keep_count]
+            from core.models import ChatMessage as InMemoryMsg
+            new_sess = _session_manager.get_session(new_sid)
+            for msg in history:
+                new_sess.add_message(InMemoryMsg(msg["role"], msg["content"]))
+            try:
+                from src.event_bus import fire_event
+                fire_event("session_created", owner)
+            except Exception:
+                logger.debug("session_created event dispatch failed", exc_info=True)
+
+            return {"action": "fork", "session_id": new_sid,
+                    "source_session": target_sid, "messages_copied": len(history),
+                    "results": f"Forked session '{source.name}' -> new session {new_sid} ({len(history)} messages)"}
+
+        else:
+            return {"error": f"Unknown action '{action}'. Use: list, switch, rename, archive, unarchive, delete, important, unimportant, truncate, fork"}
+    except Exception as e:
+        logger.error(f"manage_session failed: {e}")
+        return {"error": str(e)}
+    finally:
+        db.close()
+
+
+# ---------------------------------------------------------------------------
+# Handler classes registered in TOOL_HANDLERS
+# ---------------------------------------------------------------------------
+
+class CreateSessionTool:
+    async def execute(self, content: str, ctx: dict) -> Dict:
+        return await create_session(content, ctx.get("session_id"), owner=ctx.get("owner"))
+
+
+class ListSessionsTool:
+    async def execute(self, content: str, ctx: dict) -> Dict:
+        return await list_sessions(content, ctx.get("session_id"), owner=ctx.get("owner"))
+
+
+class SendToSessionTool:
+    async def execute(self, content: str, ctx: dict) -> Dict:
+        return await send_to_session(content, ctx.get("session_id"), owner=ctx.get("owner"))
+
+
+class ManageSessionTool:
+    async def execute(self, content: str, ctx: dict) -> Dict:
+        return await manage_session(content, ctx.get("session_id"), owner=ctx.get("owner"))

src/agent_tools/web_tools.py

@@ -7,6 +7,7 @@ from src.constants import MAX_OUTPUT_CHARS
 class WebSearchTool:
     async def execute(self, content: str, ctx: dict) -> dict:
         from src.search import comprehensive_web_search
+        progress_cb = ctx.get("progress_cb") if isinstance(ctx, dict) else None
         raw = content.strip()
         query = raw
         time_filter = None
@@ -37,18 +38,39 @@ class WebSearchTool:
             elif " news" in q_lc or q_lc.startswith("news ") or q_lc.endswith(" news"):
                 time_filter = "week"
         loop = asyncio.get_running_loop()
-        text, sources = await asyncio.wait_for(
-            loop.run_in_executor(
-                None,
-                lambda: comprehensive_web_search(
-                    query,
-                    max_pages=max_pages,
-                    time_filter=time_filter,
-                    return_sources=True,
+        if progress_cb:
+            await progress_cb({
+                "elapsed_s": 0,
+                "tail": f"Searching web for: {query[:160]}",
+            })
+        try:
+            text, sources = await asyncio.wait_for(
+                loop.run_in_executor(
+                    None,
+                    lambda: comprehensive_web_search(
+                        query,
+                        max_pages=max_pages,
+                        time_filter=time_filter,
+                        return_sources=True,
+                    ),
                 ),
-            ),
-            timeout=30,
-        )
+                timeout=30,
+            )
+        except asyncio.TimeoutError:
+            return {
+                "error": f"web_search timed out after 30s: {query[:200]}",
+                "exit_code": 1,
+            }
+        except Exception as e:
+            return {
+                "error": f"web_search failed: {type(e).__name__}: {str(e) or 'no details'}",
+                "exit_code": 1,
+            }
+        if progress_cb:
+            await progress_cb({
+                "elapsed_s": 30,
+                "tail": "Search completed; preparing sources.",
+            })
         output = text[:MAX_OUTPUT_CHARS] if len(text) > MAX_OUTPUT_CHARS else text
         if sources:
             output += "\n\n<!-- SOURCES:" + json.dumps(sources) + " -->"
@@ -57,13 +79,23 @@ class WebSearchTool:
 class WebFetchTool:
     async def execute(self, content: str, ctx: dict) -> dict:
         from src.search.content import fetch_webpage_content
+        from src.constants import WEB_FETCH_HARD_MAX_BYTES
         raw = content.strip()
         url = ""
+        max_bytes = None
         if raw.startswith("{"):
             try:
                 parsed = json.loads(raw)
                 if isinstance(parsed, dict):
                     url = str(parsed.get("url") or "").strip()
+                    # Download-budget override (#3812): "full": true raises the
+                    # budget to the hard cap; an explicit max_bytes is clamped
+                    # to the hard cap downstream. Default stays the soft cap.
+                    if parsed.get("full") is True:
+                        max_bytes = WEB_FETCH_HARD_MAX_BYTES
+                    mb = parsed.get("max_bytes")
+                    if isinstance(mb, int) and mb > 0:
+                        max_bytes = mb
             except json.JSONDecodeError:
                 url = ""
         if not url:
@@ -78,7 +110,7 @@ class WebFetchTool:
         loop = asyncio.get_running_loop()
         try:
             result = await asyncio.wait_for(
-                loop.run_in_executor(None, lambda: fetch_webpage_content(url, timeout=10)),
+                loop.run_in_executor(None, lambda: fetch_webpage_content(url, timeout=10, max_bytes=max_bytes)),
                 timeout=30,
             )
         except asyncio.TimeoutError:
@@ -94,8 +126,28 @@ class WebFetchTool:
                 return {"error": f"web_fetch: {url}: {err}", "exit_code": 1}
             return {"error": f"web_fetch: {url}: no readable text content (not HTML, or the page needs JS/login)", "exit_code": 1}
 
+        # Tell the model when the download budget cut the body short and how
+        # to get the rest, instead of silently presenting a partial page as
+        # the whole thing.
+        size_note = ""
+        if result.get("truncated"):
+            fetched = result.get("fetched_bytes") or 0
+            total = result.get("total_bytes")
+            total_txt = f" of {total:,} bytes" if total else ""
+            size_note = (
+                f"[partial content: download stopped at {fetched:,} bytes{total_txt}. "
+                f'Re-call with {{"url": "{url}", "full": true}} to fetch up to '
+                f"{WEB_FETCH_HARD_MAX_BYTES:,} bytes.]\n\n"
+            )
+
+        # The notice must lead the output so the MAX_OUTPUT_CHARS trim below can
+        # never drop it. The title is untrusted, uncapped page content, so a
+        # giant title ahead of the notice could push it out of range; keep the
+        # notice first and cap the title as a second guard.
+        if len(title) > 300:
+            title = title[:300] + "..."
         header = (f"# {title}\n" if title else "") + f"Source: {url}\n\n"
-        output = header + text
+        output = size_note + header + text
         if len(output) > MAX_OUTPUT_CHARS:
             output = output[:MAX_OUTPUT_CHARS] + "\n\n[...truncated]"
         return {"output": output, "exit_code": 0}

src/ai_interaction.py

@@ -1,8 +1,14 @@
 """
 ai_interaction.py
 
-AI-to-AI interaction tools: chat_with_model, create_session, list_sessions,
-send_to_session, pipeline.
+AI-to-AI interaction tools: pipeline and manage_memory, plus shared model
+resolution (_resolve_model), the session-manager singleton, and dispatch_ai_tool.
+
+As part of the tool -> registry migration (#3629), chat_with_model, ask_teacher
+and list_models moved to src/agent_tools/model_interaction_tools.py, and
+create_session, list_sessions, send_to_session and manage_session moved to
+src/agent_tools/session_tools.py. Those modules reuse get_session_manager /
+_resolve_model / AI_CHAT_TIMEOUT from here.
 
 These are agent tools — the LLM writes fenced code blocks and they execute
 through the standard agent_tools.py pipeline.
@@ -159,440 +165,6 @@ def _resolve_model(spec: str, owner: Optional[str] = None) -> Tuple[str, str, Di
 # Tool implementations
 # ---------------------------------------------------------------------------
 
-async def do_chat_with_model(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
-    """Send a message to a specific model and return its response.
-
-    Content format:
-      Line 1: model_name (or model_name@endpoint_name)
-      Line 2+: the message to send
-    """
-    from src.llm_core import llm_call_async
-
-    lines = content.strip().split("\n", 1)
-    if not lines or not lines[0].strip():
-        return {"error": "First line must be the model name"}
-
-    model_spec = lines[0].strip()
-    message = lines[1].strip() if len(lines) > 1 else ""
-    if not message:
-        return {"error": "No message provided (line 2+ is the message)"}
-
-    try:
-        url, model, headers = _resolve_model(model_spec, owner=owner)
-    except ValueError as e:
-        return {"error": str(e)}
-
-    try:
-        response = await llm_call_async(
-            url, model,
-            [{"role": "user", "content": message}],
-            headers=headers,
-            timeout=AI_CHAT_TIMEOUT,
-        )
-        # Truncate very long responses
-        if len(response) > 10000:
-            response = response[:10000] + "\n... (truncated)"
-        return {"model": model, "response": response}
-    except Exception as e:
-        logger.error(f"chat_with_model failed: {e}")
-        return {"error": f"Failed to get response from {model_spec}: {e}"}
-
-
-_TEACHER_SYSTEM_PROMPT = (
-    "You are a senior AI mentor. A less capable model is stuck on a problem and asking for help. "
-    "Provide clear, actionable guidance:\n"
-    "1. Brief analysis of the problem\n"
-    "2. Recommended approach (step by step)\n"
-    "3. Key things to watch out for\n\n"
-    "Be concise and practical. No preamble."
-)
-
-
-async def do_ask_teacher(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
-    """Ask a more capable model for help.
-
-    Content format:
-      Line 1: model_name (or 'auto')
-      Line 2+: the problem description
-    """
-    from src.llm_core import llm_call_async
-    from src.settings import get_setting
-
-    lines = content.strip().split("\n", 1)
-    model_spec = lines[0].strip() if lines else "auto"
-    problem = lines[1].strip() if len(lines) > 1 else ""
-
-    if not problem:
-        return {"error": "No problem description provided"}
-
-    if model_spec.lower() in ("auto", ""):
-        model_spec = get_setting("teacher_model", "")
-        if not model_spec:
-            return {"error": "No teacher model configured. Specify a model name or set teacher_model in settings."}
-
-    try:
-        url, model, headers = _resolve_model(model_spec, owner=owner)
-    except ValueError as e:
-        return {"error": str(e)}
-
-    try:
-        response = await llm_call_async(
-            url, model,
-            [
-                {"role": "system", "content": _TEACHER_SYSTEM_PROMPT},
-                {"role": "user", "content": f"Problem:\n{problem}"},
-            ],
-            headers=headers,
-            timeout=AI_CHAT_TIMEOUT,
-        )
-        if len(response) > 8000:
-            response = response[:8000] + "\n... (truncated)"
-        return {"model": model, "response": response, "teacher": True}
-    except Exception as e:
-        logger.error(f"ask_teacher failed: {e}")
-        return {"error": f"Teacher call failed ({model_spec}): {e}"}
-
-
-async def do_second_opinion(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
-    """Get a second opinion from another model, then have the original model
-    evaluate the feedback and produce a unified version.
-
-    Content format:
-      Line 1: model_name (or model_name@endpoint_name)
-      Line 2+ (optional): specific question or focus area
-
-    Flow:
-      1. Pull recent conversation context
-      2. Send to reviewer model → get honest feedback
-      3. Send feedback back to the session's own model → evaluate & unify
-      4. Return both the review and the unified response
-    """
-    from src.llm_core import llm_call_async
-
-    lines = content.strip().split("\n", 1)
-    if not lines or not lines[0].strip():
-        return {"error": "First line must be the model name"}
-
-    model_spec = lines[0].strip()
-    focus = lines[1].strip() if len(lines) > 1 else ""
-
-    try:
-        reviewer_url, reviewer_model, reviewer_headers = _resolve_model(model_spec, owner=owner)
-    except ValueError as e:
-        return {"error": str(e)}
-
-    # Pull recent conversation context from current session
-    context_text = ""
-    sess = None
-    if session_id and _session_manager:
-        sess = _session_manager.get_session(session_id)
-        if sess:
-            messages = sess.get_context_messages()
-            recent = messages[-15:] if len(messages) > 15 else messages
-            parts = []
-            for m in recent:
-                role = m.get("role", "unknown").upper()
-                text = m.get("content", "")
-                if isinstance(text, list):
-                    text = " ".join(
-                        p.get("text", "") for p in text if isinstance(p, dict)
-                    )
-                if text:
-                    parts.append(f"[{role}]: {text[:2000]}")
-            context_text = "\n\n".join(parts)
-
-    if not context_text:
-        return {"error": "No conversation context found to review"}
-
-    # ── Step 1: Get the reviewer's feedback ──
-    reviewer_system = (
-        "You are giving a second opinion on a conversation between a user and an AI assistant. "
-        "Your job is to be genuinely helpful and honest — not a yes-man, but not a contrarian either.\n\n"
-        "Guidelines:\n"
-        "- If the plan/idea is solid, say so clearly. Don't manufacture problems that aren't there.\n"
-        "- If you spot a real flaw, blind spot, or simpler approach — call it out directly.\n"
-        "- Be practical. Don't over-engineer or over-analyze. Real-world tradeoffs matter.\n"
-        "- If there's a meaningfully better way to do something, suggest it concretely.\n"
-        "- Give credit where it's due — highlight what's working well.\n"
-        "- Keep it concise and actionable. No fluff.\n"
-        "- You're a second pair of eyes, not a professor grading a paper."
-    )
-
-    reviewer_message = f"Here's the conversation so far:\n\n{context_text}"
-    if focus:
-        reviewer_message += f"\n\n---\nSpecifically, I want your take on: {focus}"
-    else:
-        reviewer_message += "\n\n---\nGive me your honest second opinion on what's being discussed."
-
-    try:
-        review = await llm_call_async(
-            reviewer_url, reviewer_model,
-            [
-                {"role": "system", "content": reviewer_system},
-                {"role": "user", "content": reviewer_message},
-            ],
-            headers=reviewer_headers,
-            timeout=AI_CHAT_TIMEOUT,
-        )
-        if len(review) > 8000:
-            review = review[:8000] + "\n... (truncated)"
-    except Exception as e:
-        logger.error(f"second_opinion reviewer call failed: {e}")
-        return {"error": f"Failed to get second opinion from {model_spec}: {e}"}
-
-    # ── Step 2: Send review back to session's own model for evaluation ──
-    unified = ""
-    original_model = "unknown"
-    if sess:
-        original_url = sess.endpoint_url
-        original_model = sess.model
-        original_headers = getattr(sess, "headers", None) or {}
-
-        unify_system = (
-            "Another AI model just reviewed the conversation you've been having with the user. "
-            "Read their feedback carefully, then respond with:\n\n"
-            "1. **What you agree with** — acknowledge valid points honestly.\n"
-            "2. **What you disagree with** — explain why, briefly.\n"
-            "3. **Unified version** — produce an updated/refined version of whatever was being discussed, "
-            "incorporating the feedback you found valid. Don't accept every note blindly — "
-            "use your judgment on what actually improves things vs what's unnecessary.\n\n"
-            "Be concise and practical. The user wants a better result, not a meta-discussion."
-        )
-
-        unify_message = (
-            f"Here's the conversation context:\n\n{context_text}\n\n"
-            f"---\n\n"
-            f"**Review from {reviewer_model}:**\n\n{review}\n\n"
-            f"---\n\n"
-            f"Evaluate this feedback and produce a unified improved version."
-        )
-
-        try:
-            unified = await llm_call_async(
-                original_url, original_model,
-                [
-                    {"role": "system", "content": unify_system},
-                    {"role": "user", "content": unify_message},
-                ],
-                headers=original_headers,
-                timeout=AI_CHAT_TIMEOUT,
-            )
-            if len(unified) > 10000:
-                unified = unified[:10000] + "\n... (truncated)"
-        except Exception as e:
-            logger.error(f"second_opinion unify call failed: {e}")
-            unified = f"(Failed to get unified response: {e})"
-
-    # Build combined result
-    combined = (
-        f"## Second Opinion from {reviewer_model}\n\n{review}"
-        f"\n\n---\n\n"
-        f"## {original_model}'s Response\n\n{unified}"
-    )
-
-    return {
-        "model": reviewer_model,
-        "response": combined,
-        "instruction": "Present these results to the user exactly as they are. Do NOT call second_opinion again. The user can continue the conversation from here.",
-    }
-
-
-async def do_create_session(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
-    """Create a new chat session.
-
-    Content format:
-      Line 1: session name
-      Line 2: model_name (or model_name@endpoint_name)
-    """
-    if not _session_manager:
-        return {"error": "Session manager not available"}
-
-    lines = content.strip().split("\n")
-    if len(lines) < 2:
-        return {"error": "Need 2 lines: session name, then model spec"}
-
-    name = lines[0].strip()
-    model_spec = lines[1].strip()
-
-    if not name:
-        return {"error": "Session name cannot be empty"}
-
-    try:
-        url, model, headers = _resolve_model(model_spec, owner=owner)
-    except ValueError as e:
-        return {"error": str(e)}
-
-    sid = str(uuid.uuid4())[:8]
-    try:
-        _session_manager.create_session(
-            session_id=sid,
-            name=name,
-            endpoint_url=url,
-            model=model,
-            rag=False,
-            owner=owner,
-        )
-        # Store headers on session for future calls
-        sess = _session_manager.get_session(sid)
-        if sess and headers:
-            sess.headers = headers
-        try:
-            from src.event_bus import fire_event
-            fire_event("session_created", owner)
-        except Exception:
-            logger.debug("session_created event dispatch failed", exc_info=True)
-
-        return {"session_id": sid, "name": name, "model": model, "endpoint_url": url}
-    except Exception as e:
-        logger.error(f"create_session failed: {e}")
-        return {"error": f"Failed to create session: {e}"}
-
-
-async def do_list_sessions(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
-    """List sessions sorted by most-recently-active first.
-
-    Output includes a relative "last active" timestamp per row so the
-    agent can answer "open my last chat" without guessing from titles.
-    The most-recent session is always first in the list.
-
-    Content = optional filter keyword (matches session name).
-    """
-    if not _session_manager:
-        return {"error": "Session manager not available"}
-
-    keyword = content.strip().lower() if content.strip() else None
-
-    try:
-        from core.database import SessionLocal, Session as DbSession
-        from datetime import datetime, timezone
-
-        # Pull every session's last_accessed from the DB so we can sort
-        # by recency. In-memory sessions hold name + model + msg_count;
-        # the DB row holds the timestamps.
-        db = SessionLocal()
-        try:
-            db_rows = {r.id: r for r in db.query(DbSession).all()}
-        finally:
-            db.close()
-
-        # SECURITY: scope to the caller's sessions. Passing None returned
-        # every user's sessions, which the agent tool then exposed via the
-        # "list my chats" reply.
-        sessions = _session_manager.get_sessions_for_user(owner)
-        rows = []
-        for sid, sess in sessions.items():
-            if keyword and keyword not in (sess.name or "").lower():
-                continue
-            db_row = db_rows.get(sid)
-            # Prefer last_accessed; fall back to updated_at, then created_at.
-            ts = None
-            if db_row:
-                ts = getattr(db_row, 'last_accessed', None) or getattr(db_row, 'updated_at', None) or getattr(db_row, 'created_at', None)
-            rows.append((ts, sid, sess))
-
-        # Sort by timestamp DESC; rows without a timestamp sink to the bottom.
-        rows.sort(key=lambda r: r[0] or datetime.min, reverse=True)
-
-        def _rel(ts):
-            if not ts:
-                return 'never'
-            now = datetime.utcnow()
-            try:
-                if ts.tzinfo is not None:
-                    now = datetime.now(timezone.utc)
-                diff = (now - ts).total_seconds()
-            except Exception:
-                return 'unknown'
-            if diff < 60: return 'just now'
-            if diff < 3600: return f'{int(diff / 60)}m ago'
-            if diff < 86400: return f'{int(diff / 3600)}h ago'
-            if diff < 86400 * 7: return f'{int(diff / 86400)}d ago'
-            return ts.strftime('%Y-%m-%d')
-
-        lines = []
-        for i, (ts, sid, sess) in enumerate(rows):
-            if i >= 50:
-                lines.append(f"... and {len(rows) - 50} more (showing first 50)")
-                break
-            safe_name = (sess.name or "Untitled").replace("[", "\\[").replace("]", "\\]")
-            msg_count = getattr(sess, "message_count", 0) or 0
-            model = getattr(sess, "model", "unknown")
-            marker = " ← most recent" if i == 0 else ""
-            lines.append(f"- **[{safe_name}](#session-{sid})** (id: `{sid}`, model: {model}, {msg_count} msgs, last active {_rel(ts)}){marker}")
-
-        if not lines:
-            return {"results": "No sessions found" + (f" matching '{keyword}'" if keyword else "") + "."}
-
-        return {
-            "results": (
-                f"Found {len(rows)} session(s), sorted most-recent first:\n"
-                + "\n".join(lines)
-                + "\n\nAssistant: when replying to the user, preserve the chat-title markdown links exactly as shown, e.g. `[Chat](#session-id)`. Do not rewrite this as a plain, non-clickable table."
-            )
-        }
-    except Exception as e:
-        logger.error(f"list_sessions failed: {e}")
-        return {"error": str(e)}
-
-
-async def do_send_to_session(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
-    """Send a message to an existing session and get a response.
-
-    Content format:
-      Line 1: session_id
-      Line 2+: message
-    """
-    from src.llm_core import llm_call_async
-    from core.models import ChatMessage
-
-    if not _session_manager:
-        return {"error": "Session manager not available"}
-
-    lines = content.strip().split("\n", 1)
-    if len(lines) < 2:
-        return {"error": "Need 2 lines: session_id, then message"}
-
-    target_sid = lines[0].strip()
-    message = lines[1].strip()
-
-    sess = _session_manager.get_session(target_sid)
-    if not sess:
-        return {"error": f"Session '{target_sid}' not found"}
-
-    # Owner-scope: reject access to another user's session
-    if owner and getattr(sess, "owner", None) and sess.owner != owner:
-        return {"error": f"Session '{target_sid}' not found"}
-
-    if not message:
-        return {"error": "No message provided"}
-
-    try:
-        # Build context from session history
-        context = sess.get_context_messages()
-        context.append({"role": "user", "content": message})
-
-        response = await llm_call_async(
-            sess.endpoint_url, sess.model, context,
-            headers=sess.headers,
-            timeout=AI_CHAT_TIMEOUT,
-        )
-
-        # Save both messages to session
-        sess.add_message(ChatMessage("user", message))
-        sess.add_message(ChatMessage("assistant", response))
-
-        # Truncate for tool output
-        if len(response) > 10000:
-            response = response[:10000] + "\n... (truncated)"
-
-        return {
-            "session_id": target_sid,
-            "session_name": sess.name,
-            "response": response,
-        }
-    except Exception as e:
-        logger.error(f"send_to_session failed: {e}")
-        return {"error": f"Failed to send to session: {e}"}
 
 
 async def stream_ai_tool(tool: str, content: str, session_id: Optional[str] = None, owner: Optional[str] = None):
@@ -715,229 +287,6 @@ async def do_pipeline(content: str, session_id: Optional[str] = None, owner: Opt
 # Session management tool
 # ---------------------------------------------------------------------------
 
-async def do_manage_session(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
-    """Manage sessions: rename, archive, delete, important, truncate, fork.
-
-    Content format:
-      Line 1: action (rename|archive|unarchive|delete|important|unimportant|truncate|fork)
-      Line 2: target session_id (or "current" to use the active session)
-      Line 3+: action-specific params (e.g. new name for rename, keep_count for truncate)
-    """
-    if not _session_manager:
-        return {"error": "Session manager not available"}
-
-    from src.database import SessionLocal, Session as DbSession
-
-    # Accept BOTH the structured JSON args the tool schema advertises
-    # ({action, session_id, value}) AND the legacy line-based format
-    # (line1=action, line2=session_id, line3=value). Native function-calling
-    # models send JSON; fenced-block callers send lines. Previously only the
-    # line format was parsed, so a model that followed the schema (JSON) got
-    # "Need at least 2 lines" / "Rename needs line 3" and couldn't drive it.
-    _raw = (content or "").strip()
-    action = ""
-    target_sid = ""
-    value = None      # the action param: new name (rename) / keep_count (truncate, fork)
-    _list_filter = ""
-    _parsed = None
-    if _raw.startswith("{"):
-        try:
-            _parsed = json.loads(_raw)
-        except Exception:
-            _parsed = None
-    if isinstance(_parsed, dict):
-        action = str(_parsed.get("action") or "").strip().lower()
-        target_sid = str(_parsed.get("session_id") or _parsed.get("session") or _parsed.get("id") or "").strip()
-        _v = _parsed.get("value")
-        if _v is None:
-            _v = (_parsed.get("name") or _parsed.get("new_name")
-                  or _parsed.get("title") or _parsed.get("keep_count"))
-        value = None if _v is None else str(_v).strip()
-        _list_filter = str(_parsed.get("filter") or "").strip()
-    else:
-        lines = _raw.split("\n")
-        if not lines or not lines[0].strip():
-            return {"error": "Missing action (rename|archive|delete|important|truncate|fork|list|switch)"}
-        action = lines[0].strip().lower()
-        target_sid = lines[1].strip() if len(lines) >= 2 else ""
-        value = lines[2].strip() if len(lines) >= 3 else None
-        _list_filter = "\n".join(lines[1:]).strip()
-
-    if not action:
-        return {"error": "Missing action (rename|archive|delete|important|truncate|fork|list|switch)"}
-
-    # `list` alias — dispatch to do_list_sessions so the agent's natural
-    # first guess (every other manage_* tool has a `list` action) works.
-    if action == "list":
-        return await do_list_sessions(_list_filter, session_id, owner=owner)
-
-    if not target_sid:
-        return {"error": "Need a session_id (or 'current' for the active chat)"}
-
-    # Allow "current" to refer to the active session
-    if target_sid.lower() == "current" and session_id:
-        target_sid = session_id
-
-    # `switch` / `open` / `select` / `view` — the agent reaches for
-    # these when the user asks to "open" or "switch to" a session.
-    # There's no server-side way to make the browser navigate, so we
-    # just return a clickable anchor link the user can click. The
-    # frontend's chat-history click delegate routes `#session-<id>`
-    # to selectSession(). The agent's reply naturally embeds this
-    # result so the user sees a single clickable line.
-    def _session_query(db):
-        query = db.query(DbSession).filter(DbSession.id == target_sid)
-        if owner is not None:
-            query = query.filter(DbSession.owner == owner)
-        return query
-
-    if action in ("switch", "open", "select", "view"):
-        db = SessionLocal()
-        try:
-            db_sess = _session_query(db).first()
-            if not db_sess:
-                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
-            name = db_sess.name or target_sid
-        finally:
-            db.close()
-        return {
-            "action": action,
-            "session_id": target_sid,
-            "name": name,
-            "results": f"[{name}](#session-{target_sid}) — click to open.",
-        }
-
-    db = SessionLocal()
-    try:
-        if action == "rename":
-            if not value:
-                return {"error": "rename needs a new name (the `value` arg, or line 3 in the legacy format)"}
-            new_name = value
-            db_sess = _session_query(db).first()
-            if not db_sess:
-                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
-            db_sess.name = new_name
-            db.commit()
-            _session_manager.update_session_name(target_sid, new_name)
-            return {"action": "rename", "session_id": target_sid, "name": new_name,
-                    "results": f"Session renamed to '{new_name}'"}
-
-        elif action == "archive":
-            db_sess = _session_query(db).first()
-            if not db_sess:
-                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
-            db_sess.archived = True
-            db.commit()
-            return {"action": "archive", "session_id": target_sid,
-                    "results": f"Session '{db_sess.name}' archived"}
-
-        elif action == "unarchive":
-            db_sess = _session_query(db).first()
-            if not db_sess:
-                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
-            db_sess.archived = False
-            db.commit()
-            return {"action": "unarchive", "session_id": target_sid,
-                    "results": f"Session '{db_sess.name}' unarchived"}
-
-        elif action == "delete":
-            if target_sid == session_id:
-                return {"error": "Cannot delete the current session while chatting in it. Delete other sessions first."}
-            db_sess = _session_query(db).first()
-            if not db_sess:
-                return {"error": f"Session '{target_sid}' not found. Refusing to delete an unknown chat id; use the exact id from list_sessions."}
-            if db_sess and db_sess.is_important:
-                return {"error": f"Session '{db_sess.name}' is starred/favorited. Unstar it first before deleting."}
-            try:
-                ok = _session_manager.delete_session(target_sid)
-                if not ok:
-                    return {"error": f"Session '{target_sid}' was not deleted because it no longer exists."}
-                return {"action": "delete", "session_id": target_sid,
-                        "results": f"Session '{db_sess.name or target_sid}' deleted"}
-            except Exception as e:
-                return {"error": f"Failed to delete session: {e}"}
-
-        elif action in ("important", "unimportant"):
-            is_important = action == "important"
-            db_sess = _session_query(db).first()
-            if not db_sess:
-                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
-            # Prevent AI from unstarring sessions — only the user can do that manually
-            if not is_important and db_sess.is_important:
-                return {"error": f"Session '{db_sess.name}' is starred by the user. Only the user can unstar sessions manually."}
-            db_sess.is_important = is_important
-            db.commit()
-            status = "marked as important" if is_important else "unmarked as important"
-            return {"action": action, "session_id": target_sid,
-                    "results": f"Session '{db_sess.name}' {status}"}
-
-        elif action == "truncate":
-            db_sess = _session_query(db).first()
-            if not db_sess:
-                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
-            keep_count = 10
-            if value:
-                try:
-                    keep_count = int(value)
-                except ValueError:
-                    pass
-            success = _session_manager.truncate_messages(target_sid, keep_count)
-            if success:
-                return {"action": "truncate", "session_id": target_sid,
-                        "results": f"Session truncated to last {keep_count} messages"}
-            return {"error": f"Failed to truncate session '{target_sid}'"}
-
-        elif action == "fork":
-            db_sess = _session_query(db).first()
-            if not db_sess:
-                return {"error": f"Session '{target_sid}' not found. Use list_sessions and pass the exact id it returned."}
-            keep_count = 0  # 0 = all messages
-            if value:
-                try:
-                    keep_count = int(value)
-                except ValueError:
-                    pass
-
-            source = _session_manager.get_session(target_sid)
-            if not source:
-                return {"error": f"Session '{target_sid}' not found"}
-
-            new_sid = str(uuid.uuid4())[:8]
-            _session_manager.create_session(
-                session_id=new_sid,
-                name=f"Fork: {source.name}",
-                endpoint_url=source.endpoint_url,
-                model=source.model,
-                rag=False,
-                owner=owner,
-            )
-            # Copy messages
-            history = source.get_context_messages()
-            if keep_count > 0:
-                history = history[:keep_count]
-            from core.models import ChatMessage as InMemoryMsg
-            new_sess = _session_manager.get_session(new_sid)
-            for msg in history:
-                new_sess.add_message(InMemoryMsg(msg["role"], msg["content"]))
-            try:
-                from src.event_bus import fire_event
-                fire_event("session_created", owner)
-            except Exception:
-                logger.debug("session_created event dispatch failed", exc_info=True)
-
-            return {"action": "fork", "session_id": new_sid,
-                    "source_session": target_sid, "messages_copied": len(history),
-                    "results": f"Forked session '{source.name}' -> new session {new_sid} ({len(history)} messages)"}
-
-        else:
-            return {"error": f"Unknown action '{action}'. Use: list, switch, rename, archive, unarchive, delete, important, unimportant, truncate, fork"}
-    except Exception as e:
-        logger.error(f"manage_session failed: {e}")
-        return {"error": str(e)}
-    finally:
-        db.close()
-
-
 # ---------------------------------------------------------------------------
 # Memory management tool
 # ---------------------------------------------------------------------------
@@ -1104,83 +453,6 @@ async def do_manage_memory(content: str, session_id: Optional[str] = None, owner
         return {"error": f"Unknown action '{action}'. Use: list, add, edit, delete, search"}
 
 
-# ---------------------------------------------------------------------------
-# List models tool
-# ---------------------------------------------------------------------------
-
-async def do_list_models(content: str, session_id: Optional[str] = None, owner: Optional[str] = None) -> Dict:
-    """List all available models across configured endpoints.
-
-    Content = optional filter keyword.
-    """
-    import httpx
-    from src.database import SessionLocal, ModelEndpoint
-    from src.llm_core import _detect_provider, ANTHROPIC_MODELS
-    from src.auth_helpers import owner_filter
-
-    keyword = content.strip().lower() if content.strip() else None
-
-    db = SessionLocal()
-    try:
-        query = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True)
-        if owner:
-            query = owner_filter(query, ModelEndpoint, owner)
-        endpoints = query.all()
-        if not endpoints:
-            return {"results": "No enabled model endpoints configured."}
-
-        result_lines = []
-        total_models = 0
-
-        for ep in endpoints:
-            try:
-                base, api_key = resolve_endpoint_runtime(ep, owner=owner)
-            except Exception:
-                continue
-            provider = _detect_provider(base)
-            headers = build_headers(api_key, base)
-
-            model_ids = []
-            if provider == "anthropic":
-                model_ids = list(ANTHROPIC_MODELS)
-            else:
-                try:
-                    models_url = build_models_url(base)
-                    if models_url:
-                        r = httpx.get(models_url, headers=headers, timeout=5)
-                        r.raise_for_status()
-                        data = r.json()
-                        model_ids = [m.get("id") for m in (data.get("data") or []) if m.get("id")]
-                        if not model_ids:
-                            model_ids = [
-                                m.get("name") or m.get("model")
-                                for m in (data.get("models") or [])
-                                if m.get("name") or m.get("model")
-                            ]
-                    else:
-                        model_ids = json.loads(ep.cached_models or "[]")
-                except Exception:
-                    model_ids = ["(endpoint offline)"]
-
-            if keyword:
-                model_ids = [m for m in model_ids if keyword in m.lower() or keyword in (ep.name or "").lower()]
-
-            if model_ids:
-                result_lines.append(f"\n**{ep.name or base}** ({provider}):")
-                for mid in model_ids:
-                    result_lines.append(f"  - `{mid}`")
-                    total_models += 1
-
-        if not result_lines:
-            return {"results": "No models found" + (f" matching '{keyword}'" if keyword else "") + "."}
-
-        header = f"Available models ({total_models} total):"
-        return {"results": header + "\n".join(result_lines)}
-    except Exception as e:
-        logger.error(f"list_models failed: {e}")
-        return {"error": str(e)}
-    finally:
-        db.close()
 
 
 # ---------------------------------------------------------------------------
@@ -1831,55 +1103,20 @@ async def dispatch_ai_tool(
 ) -> Tuple[str, Dict]:
     """Dispatch an AI interaction tool. Returns (description, result_dict)."""
 
-    if tool == "chat_with_model":
-        model_spec = content.split("\n")[0].strip()[:60]
-        desc = f"chat_with_model: {model_spec}"
-        result = await do_chat_with_model(content, session_id, owner=owner)
-
-    elif tool == "create_session":
-        name = content.split("\n")[0].strip()[:60]
-        desc = f"create_session: {name}"
-        result = await do_create_session(content, session_id, owner=owner)
-
-    elif tool == "list_sessions":
-        keyword = content.strip()[:40]
-        desc = f"list_sessions{': ' + keyword if keyword else ''}"
-        result = await do_list_sessions(content, session_id, owner=owner)
-
-    elif tool == "send_to_session":
-        sid = content.split("\n")[0].strip()[:20]
-        desc = f"send_to_session: {sid}"
-        result = await do_send_to_session(content, session_id, owner=owner)
-
-    elif tool == "pipeline":
+    if tool == "pipeline":
         desc = "pipeline: running steps"
         result = await do_pipeline(content, session_id, owner=owner)
 
-    elif tool == "manage_session":
-        action = content.split("\n")[0].strip()[:40]
-        desc = f"manage_session: {action}"
-        result = await do_manage_session(content, session_id, owner=owner)
-
     elif tool == "manage_memory":
         action = content.split("\n")[0].strip()[:40]
         desc = f"manage_memory: {action}"
         result = await do_manage_memory(content, session_id, owner=owner)
 
-    elif tool == "list_models":
-        keyword = content.strip()[:40]
-        desc = f"list_models{': ' + keyword if keyword else ''}"
-        result = await do_list_models(content, session_id, owner=owner)
-
     elif tool == "ui_control":
         action = content.split("\n")[0].strip()[:60]
         desc = f"ui_control: {action}"
         result = await do_ui_control(content, session_id, owner=owner)
 
-    elif tool == "ask_teacher":
-        problem = content.split("\n", 1)[-1].strip()[:60]
-        desc = f"ask_teacher: {problem}"
-        result = await do_ask_teacher(content, session_id, owner=owner)
-
     else:
         desc = f"unknown ai tool: {tool}"
         result = {"error": f"Unknown AI interaction tool: {tool}"}

src/bg_jobs.py

@@ -263,10 +263,32 @@ def list_for_session(session_id: str) -> List[Dict[str, Any]]:
     return [r for r in refresh().values() if r.get("session_id") == session_id]
 
 
+def kill(job_id: str) -> Optional[Dict[str, Any]]:
+    """Terminate a running job's process tree and mark it killed. Returns the
+    updated record, or None if the id is unknown. Idempotent: a job that already
+    finished is returned unchanged. Sets followed_up so the monitor does not also
+    fire an auto-continue for a job the agent deliberately stopped."""
+    jobs = _load()
+    rec = jobs.get(job_id)
+    if rec is None:
+        return None
+    if rec.get("status") == "running":
+        _kill(rec.get("pid"))
+        rec["status"] = "failed"
+        rec["exit_code"] = -1
+        rec["ended_at"] = time.time()
+        rec["killed"] = True
+        rec["followed_up"] = True
+        _save(jobs)
+    return rec
+
+
 def result_text(rec: Dict[str, Any]) -> str:
     """Human/agent-readable summary of a finished job, for the follow-up."""
     out = _read_output(rec)
-    if rec.get("timed_out"):
+    if rec.get("killed"):
+        head = "Background job was killed."
+    elif rec.get("timed_out"):
         head = f"Background job timed out after {rec.get('max_runtime_s')}s."
     elif rec.get("died"):
         head = "Background job process died unexpectedly (no exit code)."

src/builtin_actions.py

@@ -76,8 +76,7 @@ async def action_consolidate_memory(owner: str, **kwargs) -> Tuple[str, bool]:
         import json
         import re
         from src.constants import DATA_DIR
-        from src.endpoint_resolver import resolve_endpoint
-        from src.llm_core import llm_call_async
+        from src.llm_core import llm_call_async_with_fallback
         from src.memory import MemoryManager
 
         manager = MemoryManager(DATA_DIR)
@@ -116,10 +115,9 @@ async def action_consolidate_memory(owner: str, **kwargs) -> Tuple[str, bool]:
             if len(group_memories) < 2:
                 return False
 
-            url, model, headers = resolve_endpoint("utility", owner=group_owner or None)
-            if not url or not model:
-                url, model, headers = resolve_endpoint("default", owner=group_owner or None)
-            if not url or not model:
+            from src.task_endpoint import resolve_task_candidates
+            candidates = resolve_task_candidates(owner=group_owner or None)
+            if not candidates:
                 return False
 
             try:
@@ -147,13 +145,11 @@ async def action_consolidate_memory(owner: str, **kwargs) -> Tuple[str, bool]:
                     "\"drop\":[{\"id\":\"existing id\",\"reason\":\"short reason\"}]}\n\n"
                     f"MEMORIES:\n{json.dumps(items, ensure_ascii=False)}"
                 )
-                raw = await llm_call_async(
-                    url=url,
-                    model=model,
+                raw = await llm_call_async_with_fallback(
+                    candidates,
                     messages=[{"role": "user", "content": prompt}],
                     temperature=0.0,
                     max_tokens=4096,
-                    headers=headers,
                     timeout=120,
                 )
                 from src.text_helpers import strip_think
@@ -604,8 +600,7 @@ async def action_classify_events(owner: str, **kwargs) -> Tuple[str, bool]:
     try:
         from datetime import timedelta
         from core.database import SessionLocal, CalendarEvent
-        from src.endpoint_resolver import resolve_endpoint
-        from src.llm_core import llm_call_async
+        from src.llm_core import llm_call_async_with_fallback
         import re as _re, json as _json
 
         db = SessionLocal()
@@ -620,10 +615,9 @@ async def action_classify_events(owner: str, **kwargs) -> Tuple[str, bool]:
             if not events:
                 return "No upcoming events to classify", True
 
-            llm_url, llm_model, llm_headers = resolve_endpoint("utility", owner=owner)
-            if not llm_url:
-                llm_url, llm_model, llm_headers = resolve_endpoint("default", owner=owner)
-            llm_available = bool(llm_url and llm_model)
+            from src.task_endpoint import resolve_task_candidates
+            llm_candidates = resolve_task_candidates(owner=owner)
+            llm_available = bool(llm_candidates)
 
             # Pull user memories so the LLM has personal context (relationships,
             # job, hobbies). Helps it know e.g. "<name> is your spouse" so their
@@ -699,11 +693,11 @@ async def action_classify_events(owner: str, **kwargs) -> Tuple[str, bool]:
                     f"EVENTS: {_json.dumps(items)}"
                 )
                 try:
-                    raw = await llm_call_async(
-                        url=llm_url, model=llm_model,
+                    raw = await llm_call_async_with_fallback(
+                        llm_candidates,
                         messages=[{"role": "user", "content": prompt}],
                         temperature=0.1, max_tokens=16384,
-                        headers=llm_headers, timeout=180,
+                        timeout=180,
                     )
                     from src.text_helpers import strip_think as _st
                     raw = _st(raw or "", prose=False, prompt_echo=False)
@@ -810,8 +804,7 @@ async def action_learn_sender_signatures(owner: str, **kwargs) -> Tuple[str, boo
         import asyncio as _aio
         from datetime import datetime as _dt, timedelta as _td
         from routes.email_helpers import _email_cache_owner_clause, _imap_connect, SCHEDULED_DB
-        from src.endpoint_resolver import resolve_endpoint
-        from src.llm_core import llm_call_async
+        from src.llm_core import llm_call_async_with_fallback
 
         # 1. Pull recent UIDs + From headers cheaply (header-only fetch).
         def _pull_headers():
@@ -891,11 +884,11 @@ async def action_learn_sender_signatures(owner: str, **kwargs) -> Tuple[str, boo
         if not eligible:
             return "All sender sigs already cached (or no eligible senders)", True
 
-        url, model, headers = resolve_endpoint("utility", owner=owner)
-        if not url or not model:
-            url, model, headers = resolve_endpoint("default", owner=owner)
-        if not url or not model:
+        from src.task_endpoint import resolve_task_candidates
+        candidates = resolve_task_candidates(owner=owner)
+        if not candidates:
             return "No LLM endpoint available", False
+        model = candidates[0][1]
 
         analyzed = 0
         no_sig = 0
@@ -949,11 +942,11 @@ async def action_learn_sender_signatures(owner: str, **kwargs) -> Tuple[str, boo
             )
 
             try:
-                raw = await llm_call_async(
-                    url=url, model=model,
+                raw = await llm_call_async_with_fallback(
+                    candidates,
                     messages=[{"role": "user", "content": prompt}],
                     temperature=0.0, max_tokens=600,
-                    headers=headers, timeout=60,
+                    timeout=60,
                 )
                 from src.text_helpers import strip_think as _st
                 sig = _st(raw or "", prose=False, prompt_echo=False).strip()
@@ -1137,7 +1130,6 @@ async def action_test_skills(owner: str, **kwargs) -> Tuple[str, bool]:
         from services.memory.skills import SkillsManager
         from src.constants import DATA_DIR
         from routes.skills_routes import _run_skill_test_once, _skill_test_task
-        from src.endpoint_resolver import resolve_endpoint
 
         # #3 SCOPE GUARD: refuse to run on a None/empty owner — otherwise
         # `sm.load(owner=None)` returns every user's skills and we'd cross-
@@ -1152,27 +1144,40 @@ async def action_test_skills(owner: str, **kwargs) -> Tuple[str, bool]:
         if not names:
             raise TaskNoop("no skills to test")
 
-        url, model, headers = resolve_endpoint("default", owner=owner)
-        if not url or not model:
+        from src.task_endpoint import resolve_task_candidates
+        candidates = resolve_task_candidates(owner=owner)
+        if not candidates:
             return "No Default/Utility model configured — set one in Settings.", False
 
         # #2 NO SILENT MODEL SWAP: if the configured model isn't served by the
         # endpoint, try a basename match — but fail loudly instead of grabbing
         # `avail[0]` which could be an embedding-only model and produce 36
         # garbage transcripts → 36 'unknown' verdicts with no hint why.
+        url, model, headers = candidates[0]
         try:
             from src.llm_core import list_model_ids
-            avail = list_model_ids(url, headers=headers)
-            if avail and model not in avail:
-                import os as _os
-                base = _os.path.basename((model or "").rstrip("/"))
-                m = next((a for a in avail if _os.path.basename(a.rstrip("/")) == base), None)
-                if m:
-                    model = m
-                else:
-                    return (f"Default model '{model}' not served by endpoint {url}. "
-                            f"Available: {', '.join(avail[:8])}{'…' if len(avail) > 8 else ''}. "
-                            "Set a valid Default model in Settings."), False
+            import os as _os
+
+            selected = None
+            mismatch_notes = []
+            for cand_url, cand_model, cand_headers in candidates:
+                avail = list_model_ids(cand_url, headers=cand_headers)
+                if not avail or cand_model in avail:
+                    selected = (cand_url, cand_model, cand_headers)
+                    break
+                base = _os.path.basename((cand_model or "").rstrip("/"))
+                matched = next((a for a in avail if _os.path.basename(a.rstrip("/")) == base), None)
+                if matched:
+                    selected = (cand_url, matched, cand_headers)
+                    break
+                mismatch_notes.append(
+                    f"{cand_model} not served by {cand_url}; available: "
+                    f"{', '.join(avail[:8])}{'...' if len(avail) > 8 else ''}"
+                )
+            if selected:
+                url, model, headers = selected
+            elif mismatch_notes:
+                return "No configured task fallback model is served. " + " | ".join(mismatch_notes[:3]), False
         except Exception as _e:
             logger.warning(f"test_skills model resolve check failed (continuing): {_e}")
 
@@ -1483,7 +1488,6 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
         from pathlib import Path as _P
         from core.database import SessionLocal as _SL, EmailAccount as _EA
         from routes.email_helpers import _imap_connect, _decode_header
-        from src.endpoint_resolver import resolve_endpoint, resolve_utility_fallback_candidates
         from src.llm_core import llm_call_async_with_fallback
 
         # Per-owner state file so multi-user runs don't clobber each other's
@@ -1505,12 +1509,10 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
 
         # ── 1. Resolve LLM candidates (utility primary + utility fallbacks; fall
         # through to default chat as a last resort).
-        url, model, headers = resolve_endpoint("utility", owner=owner)
-        if not url or not model:
-            url, model, headers = resolve_endpoint("default", owner=owner)
-        if not url or not model:
+        from src.task_endpoint import resolve_task_candidates
+        candidates = resolve_task_candidates(owner=owner)
+        if not candidates:
             return "No LLM endpoint available", False
-        candidates = [(url, model, headers)] + resolve_utility_fallback_candidates(owner=owner)
 
         # ── 2. Enumerate enabled accounts. Match this task's owner AND fall
         # back to the legacy "unowned account whose imap_user / from_address

src/builtin_mcp.py

@@ -14,6 +14,7 @@ import subprocess
 import sys
 
 from core.platform_compat import IS_WINDOWS, which_tool
+from src.runtime_paths import get_app_root
 
 logger = logging.getLogger(__name__)
 
@@ -81,7 +82,7 @@ _BUILTIN_NPX_SERVERS = {
         "name": "Built-in: Browser",
         "command": "npx",
         "args": ["-y", "@playwright/mcp@latest", "--headless", "--caps", "vision"],
-    },
+    }
 }
 
 # Global flag to disable MCP if there are compatibility issues
@@ -94,7 +95,7 @@ async def register_builtin_servers(mcp_manager):
         logger.info("Built-in MCP servers disabled via ODYSSEUS_DISABLE_MCP")
         return
 
-    base_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+    base_dir = get_app_root()
     python = sys.executable
 
     async def _connect_python_server(server_id: str, script_path: str, name: str):

src/config.py

@@ -5,6 +5,7 @@ from pydantic_settings import BaseSettings, SettingsConfigDict
 from pydantic import Field, field_validator
 
 from src.constants import DATA_DIR as _DATA_DIR_CONST
+from src.runtime_paths import get_app_root
 
 # Cross-platform OS flag, exposed here so callers can `from src.config import
 # IS_WINDOWS`. Defined locally (a trivial `os.name == "nt"`) rather than imported
@@ -19,7 +20,7 @@ IS_WINDOWS = os.name == "nt"
 class DataConfig(BaseSettings):
     """Configuration for data storage and file handling."""
     # Base directory
-    base_dir: Path = Field(default=Path(__file__).parent.parent, description="Base directory for the application")
+    base_dir: Path = Field(default=Path(get_app_root()), description="Base directory for the application")
     
     # Data paths
     data_dir: Path = Field(default=Path(_DATA_DIR_CONST), description="Main data directory")
@@ -138,7 +139,7 @@ class AppConfig(BaseSettings):
         if isinstance(v, dict) and "base_dir" in v:
             base_dir = v["base_dir"]
         else:
-            base_dir = Path(__file__).parent.parent
+            base_dir = Path(get_app_root())
         
         # Convert string paths to Path objects relative to base_dir
         data_dir = Path(_DATA_DIR_CONST)

src/constants.py

@@ -2,12 +2,14 @@
 """Application-wide constants and configuration values."""
 import os
 
-APP_VERSION = "1.0.0"
+from src.runtime_paths import get_app_root, get_default_data_dir
+
+APP_VERSION = "1.0.1"
 
 # Base paths
-BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) + "/"
+BASE_DIR = os.path.join(get_app_root(), "")
 STATIC_DIR = os.path.join(BASE_DIR, "static")
-DATA_DIR = os.getenv("ODYSSEUS_DATA_DIR", os.path.join(BASE_DIR, "data"))
+DATA_DIR = os.getenv("ODYSSEUS_DATA_DIR", get_default_data_dir())
 
 # Data file paths
 # Single source of truth: every persisted file/dir lives under DATA_DIR, which
@@ -55,7 +57,13 @@ MEMORY_VECTORS_DIR = os.path.join(DATA_DIR, "memory_vectors")
 
 # Paths with an intentional dedicated env override, defaulting under DATA_DIR.
 MAIL_ATTACHMENTS_DIR = os.getenv("ODYSSEUS_MAIL_ATTACHMENTS_DIR", os.path.join(DATA_DIR, "mail-attachments"))
-FASTEMBED_CACHE_DIR = os.getenv("FASTEMBED_CACHE_PATH", os.path.join(DATA_DIR, "fastembed_cache"))
+# `or` (not os.getenv's default arg) so a PRESENT-but-EMPTY value falls back to
+# the default. docker-compose.yml injects `FASTEMBED_CACHE_PATH=${FASTEMBED_CACHE_PATH:-}`,
+# which sets the var to "" when the host hasn't defined it. os.getenv(name, default)
+# only returns the default when the var is ABSENT, so the empty string would win →
+# os.makedirs("") raises [Errno 2] No such file or directory: '' → FastEmbed fails to
+# init and all vector features (RAG, semantic memory, tool index) silently degrade.
+FASTEMBED_CACHE_DIR = os.getenv("FASTEMBED_CACHE_PATH") or os.path.join(DATA_DIR, "fastembed_cache")
 
 # Agent tool output limits (single source of truth — imported by tool_execution.py,
 # tool_implementations.py, agent_tools.py, and any other module that needs them)
@@ -63,11 +71,26 @@ MAX_OUTPUT_CHARS = 10_000       # cap for bash/python/web_search/web_fetch outpu
 MAX_READ_CHARS = 20_000         # cap for read_file / document preview
 MAX_DIFF_LINES = 400            # cap for edit_file unified-diff display
 
+# web_fetch response-size policy (#3812). MAX_OUTPUT_CHARS above only trims
+# what the agent SEES; these caps bound what the server downloads, parses,
+# and writes to the content cache. The soft cap is the default download
+# budget; the agent can raise it per call (full/max_bytes) but never past
+# the hard cap, so a model can't decide to pull a multi-GB file.
+WEB_FETCH_SOFT_MAX_BYTES = 2_000_000    # default download budget (2 MB)
+WEB_FETCH_HARD_MAX_BYTES = 20_000_000   # absolute ceiling, even with override (20 MB)
+
 # API Configuration
 MAX_CONTEXT_MESSAGES = 90
 REQUEST_TIMEOUT = 20
 OPENAI_COMPAT_PATH = "/v1/chat/completions"
 
+# Outbound UA for web_fetch / web_search scraping; common desktop UA so pages serve normal HTML.
+WEB_FETCH_USER_AGENT = os.environ.get(
+    "WEB_FETCH_USER_AGENT",
+    "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
+    "(KHTML, like Gecko) Chrome/148.0.0.0 Safari/537.36",
+)
+
 # Environment variables with defaults
 DEFAULT_HOST = os.getenv("LLM_HOST", "localhost")
 LLM_HOSTS = [h.strip() for h in os.getenv("LLM_HOSTS", "").split(",") if h.strip()]
@@ -79,6 +102,9 @@ SEARXNG_INSTANCE = os.getenv("SEARXNG_INSTANCE", "http://localhost:8080")
 CLEANUP_ENABLED = os.getenv("CLEANUP_ENABLED", "True").lower() == "true"
 CLEANUP_INTERVAL_HOURS = int(os.getenv("CLEANUP_INTERVAL_HOURS", "24"))
 
+# Auth policy
+PASSWORD_MIN_LENGTH = 8
+
 # Default parameters
 DEFAULT_TEMPERATURE = 1.0
 DEFAULT_MAX_TOKENS = 0

src/embeddings.py

@@ -31,6 +31,8 @@ import numpy as np
 import httpx
 from typing import List, Optional
 
+from src.runtime_paths import get_app_root
+
 logger = logging.getLogger(__name__)
 
 _DEFAULT_MODEL = "all-minilm:l6-v2"

src/endpoint_resolver.py

@@ -161,6 +161,32 @@ def normalize_base(url: str) -> str:
     return url
 
 
+def _validated_endpoint_base(url: str) -> str:
+    """Return a base URL that is safe for endpoint path appends."""
+    base = (url or "").strip().rstrip("/")
+    if "?" in base or "#" in base:
+        raise ValueError("Endpoint base URL must not include query or fragment")
+    return urlunparse(urlparse(base)._replace(query="", fragment="")).rstrip("/")
+
+
+def _prepare_endpoint_base(base: str) -> str:
+    base = _validated_endpoint_base(normalize_base(base))
+    return _validated_endpoint_base(normalize_base(resolve_url(base)))
+
+
+def _append_endpoint_path(base: str, suffix: str) -> str:
+    parsed = urlparse(base)
+    current = (parsed.path or "").rstrip("/")
+    extra = "/" + suffix.lstrip("/")
+    path = f"{current}{extra}" if current else extra
+    return urlunparse(parsed._replace(path=path, query="", fragment=""))
+
+
+def _pathless_host(base: str, host: str) -> bool:
+    parsed = urlparse(base)
+    return (parsed.hostname or "").lower() == host and not (parsed.path or "").strip("/")
+
+
 def _anthropic_api_root(base: str) -> str:
     """Return Anthropic's API root, preserving /v1 for OpenAI-compatible APIs elsewhere."""
     base = (base or "").strip().rstrip("/")
@@ -171,15 +197,17 @@ def _anthropic_api_root(base: str) -> str:
 
 def build_chat_url(base: str) -> str:
     """Return the correct chat endpoint URL for a given base."""
-    base = resolve_url(base)
+    base = _prepare_endpoint_base(base)
     provider = _detect_provider(base)
     if provider == "anthropic":
-        return _anthropic_api_root(base) + "/v1/messages"
+        return _append_endpoint_path(_anthropic_api_root(base), "/v1/messages")
     if provider == "ollama":
-        return _ollama_api_root(base) + "/chat"
+        return _append_endpoint_path(_ollama_api_root(base), "/chat")
     if provider == "chatgpt-subscription":
-        return base.rstrip("/") + "/responses"
-    return base + "/chat/completions"
+        return _append_endpoint_path(base, "/responses")
+    if _pathless_host(base, "api.openai.com"):
+        base = _append_endpoint_path(base, "/v1")
+    return _append_endpoint_path(base, "/chat/completions")
 
 
 def build_models_url(base: str) -> Optional[str]:
@@ -193,12 +221,12 @@ def build_models_url(base: str) -> Optional[str]:
     untouched (so custom prefixes like ``/openai`` or ``/api/openai/v1`` keep
     their semantics).
     """
-    base = normalize_base(resolve_url(base))
+    base = _prepare_endpoint_base(base)
     provider = _detect_provider(base)
     if provider == "anthropic":
-        return _anthropic_api_root(base) + "/v1/models"
+        return _append_endpoint_path(_anthropic_api_root(base), "/v1/models")
     if provider == "ollama":
-        return _ollama_api_root(base) + "/tags"
+        return _append_endpoint_path(_ollama_api_root(base), "/tags")
     if provider == "chatgpt-subscription":
         return None
     # Generic OpenAI-compatible fallback: local model servers with no explicit
@@ -208,10 +236,10 @@ def build_models_url(base: str) -> Optional[str]:
     parsed = urlparse(base)
     host = (parsed.hostname or "").lower()
     is_local = host in {"localhost", "127.0.0.1", "::1", "host.docker.internal"}
-    uses_v1_models_by_default = is_local or host in {"api.deepseek.com"}
+    uses_v1_models_by_default = is_local or host in {"api.deepseek.com", "api.openai.com"}
     if not parsed.path and uses_v1_models_by_default:
-        base = base + "/v1"
-    return base + "/models"
+        base = _append_endpoint_path(base, "/v1")
+    return _append_endpoint_path(base, "/models")
 
 
 def build_headers(api_key: Optional[str], base: str) -> Dict[str, str]:
@@ -396,6 +424,9 @@ def resolve_utility_fallback_candidates(owner: Optional[str] = None) -> list:
         settings = load_settings()
         utility_ep = (get_user_setting("utility_endpoint_id", owner or "", settings.get("utility_endpoint_id", "")) or "").strip()
         if not utility_ep:
+            utility_chain = get_user_setting("utility_model_fallbacks", owner or "", settings.get("utility_model_fallbacks") or []) or []
+            if utility_chain:
+                return _resolve_fallback_candidates("utility_model_fallbacks", owner=owner)
             return _resolve_fallback_candidates("default_model_fallbacks", owner=owner)
     except Exception:
         pass

src/integrations.py

@@ -4,6 +4,7 @@ import uuid
 import logging
 import re
 from typing import Dict, List, Optional, Any
+from urllib.parse import urljoin, urlparse, urlunparse
 
 import httpx
 from fastapi import HTTPException
@@ -202,6 +203,22 @@ def mask_integration_secret(integration: Dict[str, Any]) -> Dict[str, Any]:
     return safe
 
 
+def _normalize_integration_base_url(base_url: Any) -> str:
+    if not isinstance(base_url, str) or not base_url.strip():
+        raise ValueError("Integration base URL is required")
+    cleaned = base_url.strip().rstrip("/")
+    if "?" in cleaned or "#" in cleaned:
+        raise ValueError("Integration base URL must not include query or fragment")
+    parsed = urlparse(cleaned)
+    if parsed.scheme.lower() not in ("http", "https") or not parsed.hostname:
+        raise ValueError("Integration base URL must be an HTTP(S) URL")
+    return urlunparse(parsed._replace(scheme=parsed.scheme.lower(), query="", fragment="")).rstrip("/")
+
+
+def _join_integration_url(base_url: str, path: str) -> str:
+    return urljoin(base_url.rstrip("/") + "/", path.lstrip("/"))
+
+
 def load_integrations() -> List[Dict[str, Any]]:
     """Load all integrations from disk with secrets decrypted for runtime use."""
     if not os.path.exists(DATA_FILE):
@@ -261,8 +278,10 @@ def add_integration(data: Dict[str, Any]) -> Dict[str, Any]:
 
     if not isinstance(integration.get("name"), str) or not integration["name"].strip():
         raise HTTPException(400, "Integration name is required")
-    if not isinstance(integration.get("base_url"), str) or not integration["base_url"].strip():
-        raise HTTPException(400, "Integration base URL is required")
+    try:
+        integration["base_url"] = _normalize_integration_base_url(integration.get("base_url"))
+    except ValueError as exc:
+        raise HTTPException(400, str(exc)) from exc
 
     integrations = load_integrations()
     integrations.append(integration)
@@ -272,10 +291,14 @@ def add_integration(data: Dict[str, Any]) -> Dict[str, Any]:
 
 def update_integration(integration_id: str, data: Dict[str, Any]) -> Optional[Dict[str, Any]]:
     """Update fields on an existing integration. Returns updated integration or None."""
+    data = dict(data)
     if "name" in data and (not isinstance(data["name"], str) or not data["name"].strip()):
         raise HTTPException(400, "Integration name is required")
-    if "base_url" in data and (not isinstance(data["base_url"], str) or not data["base_url"].strip()):
-        raise HTTPException(400, "Integration base URL is required")
+    if "base_url" in data:
+        try:
+            data["base_url"] = _normalize_integration_base_url(data["base_url"])
+        except ValueError as exc:
+            raise HTTPException(400, str(exc)) from exc
 
     integrations = load_integrations()
     for item in integrations:
@@ -341,9 +364,10 @@ async def execute_api_call(
     if not integration.get("enabled", True):
         return {"error": f"Integration '{integration.get('name')}' is disabled", "exit_code": 1}
 
-    base_url = integration.get("base_url", "").rstrip("/")
-    if not base_url:
-        return {"error": "Integration has no base_url configured", "exit_code": 1}
+    try:
+        base_url = _normalize_integration_base_url(integration.get("base_url", ""))
+    except ValueError as exc:
+        return {"error": str(exc), "exit_code": 1}
 
     # Strip common API path suffixes users might accidentally include
     # (e.g. "http://host/v1/" → "http://host"). The integration's preset
@@ -366,7 +390,10 @@ async def execute_api_call(
     if re.search(r"^https?://", path) or "://" in path:
         return {"error": "Path must not contain a protocol scheme", "exit_code": 1}
 
-    url = base_url + path
+    if "#" in path:
+        return {"error": "Path must not contain a fragment", "exit_code": 1}
+
+    url = _join_integration_url(base_url, path)
     method = method.upper()
 
     # Build headers

src/llm_core.py

@@ -87,7 +87,7 @@ _host_health_lock = threading.Lock()
 _model_activity: Dict[str, float] = {}
 
 _HARMONY_MARKER_RE = re.compile(
-    r"<\|channel\|>(analysis|final)"
+    r"<\|channel\|>(analysis|commentary|final)"
     r"|<\|start\|>(?:assistant|system|user|tool)?"
     r"|<\|message\|>"
     r"|<\|end\|>"
@@ -96,6 +96,7 @@ _HARMONY_MARKER_RE = re.compile(
 )
 _HARMONY_MARKERS = (
     "<|channel|>analysis",
+    "<|channel|>commentary",
     "<|channel|>final",
     "<|start|>assistant",
     "<|start|>system",
@@ -145,7 +146,10 @@ class _HarmonyStreamRouter:
             out.append((text, False))
             return
         if self._in_message:
-            out.append((text, self._channel == "analysis"))
+            # analysis + commentary (tool-call preambles / function-arg bodies)
+            # are internal, not user-facing — route them to thinking so they
+            # don't leak into the visible answer; only `final` is visible.
+            out.append((text, self._channel in ("analysis", "commentary")))
 
     def _handle_marker(self, match: re.Match[str]) -> None:
         marker = match.group(0)
@@ -283,7 +287,8 @@ def _is_ollama_native_url(url: str) -> bool:
     """Return True for native Ollama API URLs, including Ollama Cloud."""
     try:
         parsed = urlparse(url or "")
-    except Exception:
+    except Exception as e:
+        logger.warning("Failed to parse URL for Ollama detection", exc_info=e)
         return False
     host = parsed.hostname or ""
     path = (parsed.path or "").rstrip("/")
@@ -902,7 +907,10 @@ def _anthropic_rejects_temperature(model: str) -> bool:
     return (int(match.group(1)), int(match.group(2))) >= (4, 7)
 
 # Models that support structured thinking — may output </think> without opening tag
-_THINKING_MODEL_PATTERNS = ("qwen3", "qwq", "deepseek-r1", "deepseek-reasoner", "minimax", "m2-reap", "gemma")
+_THINKING_MODEL_PATTERNS = (
+    "qwen3", "qwq", "deepseek-r1", "deepseek-reasoner", "minimax",
+    "m2-reap", "gemma", "stepfun", "step-3", "step3",
+)
 
 def _supports_thinking(model: str) -> bool:
     """Check if model supports structured thinking output."""
@@ -1345,8 +1353,8 @@ def list_model_ids(
                 r = httpx.get(root + "/api/tags", timeout=timeout)
                 r.raise_for_status()
                 return [m.get("name") or m.get("model") for m in (r.json().get("models") or []) if m.get("name") or m.get("model")]
-        except Exception:
-            pass
+        except Exception as e:
+            logger.warning("Failed to fetch model list from configured endpoint", exc_info=e)
         return []
 
 def normalize_model_id(
@@ -2130,6 +2138,8 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
                                             yield _stream_delta_event(reasoning, thinking=True)
                                         content = delta.get("content") or ""
                                         if content:
+                                            content = re.sub(r"<mm:think(\s+[^>]*)?>", r"<think\1>", content, flags=re.IGNORECASE)
+                                            content = re.sub(r"</mm:think>", "</think>", content, flags=re.IGNORECASE)
                                             stripped = content.lstrip()
                                             # gpt-oss harmony format (<|channel|>analysis/final): route via the harmony
                                             # stream router. Sticky once the first marker appears — distinct from the

src/mcp_manager.py

@@ -11,6 +11,8 @@ import os
 import re
 from typing import Any, Dict, List, Optional, Set, Tuple
 
+from src.runtime_paths import get_app_root
+
 logger = logging.getLogger(__name__)
 
 def _format_mcp_connection_error(name: str, command: str = "", args: Optional[List[str]] = None, error: Exception = None) -> str:
@@ -508,7 +510,7 @@ class McpManager:
             return False
 
         script_rel, name = _BUILTIN_SERVERS[server_id]
-        base_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+        base_dir = get_app_root()
         script_path = os.path.join(base_dir, script_rel)
 
         # Clean up old connection

src/model_context.py

@@ -17,10 +17,11 @@ import httpx
 logger = logging.getLogger(__name__)
 
 _LOCAL_HOSTS = {"localhost", "127.0.0.1", "0.0.0.0", "::1", "host.docker.internal"}
-_PRIVATE_PREFIXES = ("10.", "172.16.", "172.17.", "172.18.", "172.19.",
-                     "172.20.", "172.21.", "172.22.", "172.23.", "172.24.",
-                     "172.25.", "172.26.", "172.27.", "172.28.", "172.29.",
-                     "172.30.", "172.31.", "192.168.")
+_PRIVATE_NETWORKS = (
+    ipaddress.ip_network("10.0.0.0/8"),
+    ipaddress.ip_network("172.16.0.0/12"),
+    ipaddress.ip_network("192.168.0.0/16"),
+)
 
 # Tailscale uses the CGNAT range 100.64.0.0/10, NOT all of 100.0.0.0/8.
 # A bare "100." prefix would classify public addresses (e.g. AWS ranges
@@ -36,6 +37,14 @@ def _in_tailscale_range(host: str) -> bool:
         return False
 
 
+def _is_private_ip_literal(host: str) -> bool:
+    try:
+        ip = ipaddress.ip_address(host)
+    except ValueError:
+        return False
+    return any(ip in network for network in _PRIVATE_NETWORKS)
+
+
 def _normalize_base_for_compare(url: str) -> str:
     url = (url or "").strip().rstrip("/")
     for suffix in ("/chat/completions", "/models", "/completions", "/v1/messages"):
@@ -87,7 +96,7 @@ def is_local_endpoint(url: str) -> bool:
         return True
     try:
         host = urlparse(url).hostname or ""
-        return host in _LOCAL_HOSTS or host.startswith(_PRIVATE_PREFIXES) or _in_tailscale_range(host)
+        return host in _LOCAL_HOSTS or _is_private_ip_literal(host) or _in_tailscale_range(host)
     except Exception:
         return False
 

src/personal_docs.py

@@ -322,6 +322,47 @@ class PersonalDocsManager:
         else:
             logger.info(f"Directory not in index: {directory}")
 
+    def rename_directory(self, old_directory: str, new_directory: str, *, path_map: Dict[str, str] = None):
+        """Rewrite tracked directory and excluded-file paths after an owner rename."""
+        old_directory = os.path.abspath(old_directory)
+        new_directory = os.path.abspath(new_directory)
+        path_map = {os.path.abspath(k): os.path.abspath(v) for k, v in (path_map or {}).items()}
+
+        def rewrite(path: str) -> str:
+            abs_path = os.path.abspath(path)
+            mapped = path_map.get(abs_path)
+            if mapped:
+                return mapped
+            if abs_path == old_directory:
+                return new_directory
+            if abs_path.startswith(old_directory + os.sep):
+                return new_directory + abs_path[len(old_directory):]
+            return abs_path
+
+        changed_dirs = False
+        rewritten_dirs = []
+        for directory in self.indexed_directories:
+            rewritten = rewrite(directory)
+            changed_dirs = changed_dirs or rewritten != os.path.abspath(directory)
+            if rewritten not in rewritten_dirs:
+                rewritten_dirs.append(rewritten)
+        if changed_dirs:
+            self.indexed_directories = rewritten_dirs
+            self.save_directories()
+
+        changed_excluded = False
+        rewritten_excluded = set()
+        for path in self.excluded_files:
+            rewritten = rewrite(path)
+            changed_excluded = changed_excluded or rewritten != os.path.abspath(path)
+            rewritten_excluded.add(rewritten)
+        if changed_excluded:
+            self.excluded_files = rewritten_excluded
+            self._save_excluded()
+
+        if changed_dirs or changed_excluded:
+            self.refresh_index()
+
     def get_indexed_directories(self):
         """Get the list of all indexed directories."""
         return self.indexed_directories.copy()

src/rag_singleton.py

@@ -7,6 +7,7 @@ import time
 from pathlib import Path
 
 from src.constants import RAG_DIR
+from src.runtime_paths import get_app_root
 
 logger = logging.getLogger(__name__)
 

src/rag_vector.py

@@ -50,6 +50,23 @@ def _generate_doc_id(text: str, owner: str = "") -> str:
     return f"doc_{hashlib.sha256(key.encode('utf-8')).hexdigest()[:16]}"
 
 
+def _rewrite_owner_path(value: str, path_map: Dict[str, str], path_prefixes: List[tuple]) -> str:
+    if not isinstance(value, str) or not value:
+        return value
+    abs_value = os.path.abspath(value)
+    mapped = path_map.get(abs_value)
+    if mapped:
+        return mapped
+    for old_prefix, new_prefix in path_prefixes:
+        old_abs = os.path.abspath(old_prefix)
+        new_abs = os.path.abspath(new_prefix)
+        if abs_value == old_abs:
+            return new_abs
+        if abs_value.startswith(old_abs + os.sep):
+            return new_abs + abs_value[len(old_abs):]
+    return value
+
+
 class VectorRAG:
     """RAG system using ChromaDB vector storage with hybrid search."""
 
@@ -250,6 +267,75 @@ class VectorRAG:
             "failed_count": len(docs) - len(valid),
         }
 
+    def rename_owner(
+        self,
+        old_owner: str,
+        new_owner: str,
+        *,
+        path_map: Optional[Dict[str, str]] = None,
+        path_prefixes: Optional[List[tuple]] = None,
+    ) -> Dict[str, Any]:
+        """Rewrite existing RAG metadata after an auth username rename."""
+        if not self.healthy:
+            return {"success": False, "updated_count": 0, "message": "Collection not initialized"}
+
+        old_owner = (old_owner or "").strip().lower()
+        new_owner = (new_owner or "").strip().lower()
+        if not old_owner or not new_owner or old_owner == new_owner:
+            return {"success": True, "updated_count": 0, "message": "No owner rename needed"}
+
+        path_map = {os.path.abspath(k): os.path.abspath(v) for k, v in (path_map or {}).items()}
+        path_prefixes = path_prefixes or []
+        updated_ids = set()
+        failed_count = 0
+
+        for lane_name, collection in self._collections_for_delete():
+            try:
+                results = collection.get(
+                    where={"owner": old_owner},
+                    include=["metadatas"],
+                )
+            except Exception as e:
+                logger.warning("rename_owner metadata scan failed in %s lane: %s", lane_name, e)
+                failed_count += 1
+                continue
+
+            ids = results.get("ids") or []
+            metadatas = results.get("metadatas") or []
+            if not ids:
+                continue
+
+            new_metas = []
+            selected_ids = []
+            for doc_id, meta in zip(ids, metadatas):
+                if not isinstance(meta, dict):
+                    continue
+                next_meta = dict(meta)
+                if str(next_meta.get("owner", "")).strip().lower() == old_owner:
+                    next_meta["owner"] = new_owner
+                for key in ("source", "directory"):
+                    next_meta[key] = _rewrite_owner_path(next_meta.get(key), path_map, path_prefixes)
+                selected_ids.append(doc_id)
+                new_metas.append(next_meta)
+
+            if not selected_ids:
+                continue
+
+            try:
+                collection.update(ids=selected_ids, metadatas=new_metas)
+                updated_ids.update(selected_ids)
+            except Exception as e:
+                logger.warning("rename_owner metadata update failed in %s lane: %s", lane_name, e)
+                failed_count += len(selected_ids)
+
+        success = failed_count == 0
+        return {
+            "success": success,
+            "updated_count": len(updated_ids),
+            "failed_count": failed_count,
+            "message": f"Updated {len(updated_ids)} RAG chunk(s)",
+        }
+
     # ------------------------------------------------------------------
     # Search — hybrid: vector similarity + keyword overlap
     # ------------------------------------------------------------------

src/runtime_paths.py

@@ -0,0 +1,30 @@
+"""Helpers for resolving runtime paths in source and frozen builds."""
+
+import os
+import sys
+
+
+def get_app_root() -> str:
+    """Return the app root directory.
+
+    In normal source runs, this is the repository root. In a frozen Windows
+    build, it is the bundle content root (PyInstaller's internal directory)
+    so bundled runtime folders like `static/`, `scripts/`, and `data/` stay
+    together with the executable payload.
+    """
+    if getattr(sys, "frozen", False):
+        return getattr(sys, "_MEIPASS", os.path.dirname(os.path.abspath(sys.executable)))
+    return os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+
+def get_default_data_dir() -> str:
+    """Return the default path to the data directory.
+
+    In normal runs, this is a 'data' subdirectory under the app root.
+    In frozen builds, it is a persistent user directory (~/.odysseus/data)
+    to prevent SQLite databases and other persistent files from being
+    written to the ephemeral, temporary extraction bundle directory.
+    """
+    if getattr(sys, "frozen", False):
+        return os.path.join(os.path.expanduser("~"), ".odysseus", "data")
+    return os.path.join(get_app_root(), "data")

src/task_endpoint.py

@@ -1,6 +1,11 @@
-"""Shared resolver for background-task AI endpoint (auto-naming, memory, sorting)."""
+"""Shared resolver for background-task AI endpoints."""
 
-from src.endpoint_resolver import resolve_endpoint
+from src.endpoint_resolver import (
+    resolve_chat_fallback_candidates,
+    resolve_endpoint,
+    resolve_utility_fallback_candidates,
+)
+from src.llm_core import llm_call_async_with_fallback
 
 
 def resolve_task_endpoint(fallback_url=None, fallback_model=None, fallback_headers=None, owner=None):
@@ -11,3 +16,60 @@ def resolve_task_endpoint(fallback_url=None, fallback_model=None, fallback_heade
     endpoint cannot be resolved.
     """
     return resolve_endpoint("task", fallback_url, fallback_model, fallback_headers, owner=owner)
+
+
+def resolve_task_candidates(
+    fallback_url=None,
+    fallback_model=None,
+    fallback_headers=None,
+    owner=None,
+):
+    """Return ordered background-task LLM candidates.
+
+    Order:
+    1. configured Background Tasks endpoint/model, or caller fallback
+    2. Utility endpoint/model
+    3. Default endpoint/model
+    4. Utility fallback chain
+    5. Default fallback chain
+    """
+    candidates = []
+
+    def _append(url, model, headers):
+        if not url or not model:
+            return
+        key = (url, model)
+        if any((u, m) == key for u, m, _ in candidates):
+            return
+        candidates.append((url, model, headers or {}))
+
+    _append(*resolve_task_endpoint(fallback_url, fallback_model, fallback_headers, owner=owner))
+    _append(*resolve_endpoint("utility", owner=owner))
+    _append(*resolve_endpoint("default", owner=owner))
+    for url, model, headers in resolve_utility_fallback_candidates(owner=owner):
+        _append(url, model, headers)
+    for url, model, headers in resolve_chat_fallback_candidates(owner=owner):
+        _append(url, model, headers)
+
+    return candidates
+
+
+async def task_llm_call_async(
+    messages,
+    *,
+    fallback_url=None,
+    fallback_model=None,
+    fallback_headers=None,
+    owner=None,
+    **kwargs,
+):
+    """Call the shared background-task LLM candidate chain."""
+    candidates = resolve_task_candidates(
+        fallback_url=fallback_url,
+        fallback_model=fallback_model,
+        fallback_headers=fallback_headers,
+        owner=owner,
+    )
+    if not candidates:
+        raise RuntimeError("No LLM endpoint available for background task")
+    return await llm_call_async_with_fallback(candidates, messages=messages, **kwargs)

src/task_scheduler.py

@@ -9,6 +9,8 @@ import uuid
 from datetime import datetime, timedelta, timezone
 from typing import Any, Awaitable, Callable, Dict, Tuple
 
+from core.auth import RESERVED_USERNAMES
+
 logger = logging.getLogger(__name__)
 
 
@@ -17,6 +19,34 @@ def _utcnow() -> datetime:
     return datetime.now(timezone.utc).replace(tzinfo=None)
 
 
+# Shell/file tools a scheduled task's agent should be offered by default,
+# mirroring the chat agent (where these are on unless a privilege or global
+# setting turns them off). The RAG tool selector + ASSISTANT_ALWAYS_AVAILABLE
+# never include bash/python, so on a host with an empty/degraded tool-embedding
+# index a task could not run shell or Python even for an admin owner. Offering
+# them here is safe: stream_agent_loop's blocked_tools_for_owner() still strips
+# this whole group for non-admin multi-user owners, and only admits it for
+# admins and single-user (AUTH_ENABLED=false) deployments.
+TASK_DEFAULT_SHELL_TOOLS = frozenset({
+    "bash", "python", "read_file", "write_file", "edit_file",
+    "grep", "glob", "ls", "get_workspace",
+})
+
+
+def compose_task_relevant_tools(rag_tools, assistant_always, disabled_tools):
+    """Compose the relevant-tools set offered to a scheduled task's agent.
+
+    Unions the RAG-retrieved tools, the assistant's always-available set, and
+    the default shell/file group, then removes anything the task's crew
+    explicitly disabled via its `enabled_tools` allowlist. Per-owner admin
+    gating is applied later by stream_agent_loop (blocked_tools_for_owner).
+    """
+    tools = set(rag_tools) | set(assistant_always) | set(TASK_DEFAULT_SHELL_TOOLS)
+    if disabled_tools:
+        tools -= set(disabled_tools)
+    return tools
+
+
 # ── Shared TTL cache (singleflight) ────────────────────────────────────────
 # Multiple scheduled tasks firing in the same minute often need the same
 # external data (Miniflux unreads, MCP tool snapshots, etc.). This cache
@@ -236,6 +266,29 @@ def _digest_windows(now):
     ]
 
 
+def _checkin_calendar_events(db, owner, start, end):
+    """Calendar events in [start, end] for ONE owner, for the check-in digest.
+
+    Ownership lives on CalendarCal.owner; events inherit it via calendar_id.
+    The digest query had no owner scope, so it pulled EVERY user's events into
+    one user's check-in (a cross-tenant leak of summaries/locations). Scope it
+    by joining CalendarCal, mirroring routes/calendar_routes.list_events.
+    """
+    from core.database import CalendarEvent as _CE, CalendarCal as _CC
+    return (
+        db.query(_CE)
+        .join(_CC, _CE.calendar_id == _CC.id)
+        .filter(
+            _CC.owner == owner,
+            _CE.dtstart >= start,
+            _CE.dtstart <= end,
+            _CE.status != "cancelled",
+        )
+        .order_by(_CE.dtstart)
+        .all()
+    )
+
+
 class TaskScheduler:
     def __init__(self, session_manager):
         self._session_manager = session_manager
@@ -833,6 +886,14 @@ class TaskScheduler:
                     owner=task.owner,
                     body=run.result if output == "notification" else None,
                 )
+            elif run.status == "error":
+                self.add_notification(
+                    task.name,
+                    "error",
+                    task_id,
+                    owner=task.owner,
+                    body=run.error or run.result,
+                )
 
             # Log result to the assistant chat so all task activity is visible.
             # Skip skipped/error rows — user shouldn't see "skipped: …" noise
@@ -1127,11 +1188,7 @@ class TaskScheduler:
                     # Strip timezone for naive DB comparison
                     _s = start.replace(tzinfo=None) if start.tzinfo else start
                     _e = end.replace(tzinfo=None) if end.tzinfo else end
-                    evs = _db.query(_CE).filter(
-                        _CE.dtstart >= _s,
-                        _CE.dtstart <= _e,
-                        _CE.status != "cancelled",
-                    ).order_by(_CE.dtstart).all()
+                    evs = _checkin_calendar_events(_db, task.owner, _s, _e)
                     if not evs:
                         continue
                     # Group by importance for richer output
@@ -1370,17 +1427,30 @@ class TaskScheduler:
             time_str = _utcnow().strftime("%A, %B %d %Y, %H:%M UTC")
         system_prompt = f"Current time: {time_str}\n\n{system_prompt}"
 
-        # Compute tool filter from CrewMember.enabled_tools if set
-        disabled_tools = None
+        # Compute the disabled-tools set: the crew's enabled_tools allowlist
+        # (inverted) plus the operator's global disabled_tools setting. The
+        # global list must be merged here — chat does the same merge before
+        # entering the agent loop (routes/chat_routes.py) — otherwise an admin
+        # or AUTH_ENABLED=false scheduled task would still see and call shell/
+        # file tools after the operator disabled them globally, because the
+        # prompt/schema/execution gates only enforce what is passed in.
+        disabled_tools: set[str] = set()
         if crew and crew.enabled_tools:
             try:
                 enabled = json.loads(crew.enabled_tools)
                 if isinstance(enabled, list) and enabled:
                     from src.tool_index import BUILTIN_TOOL_DESCRIPTIONS
                     all_tools = set(BUILTIN_TOOL_DESCRIPTIONS.keys())
-                    disabled_tools = all_tools - set(enabled)
+                    disabled_tools |= all_tools - set(enabled)
             except Exception:
                 pass
+        try:
+            from src.settings import get_setting
+            _global_disabled = get_setting("disabled_tools", [])
+            if isinstance(_global_disabled, list):
+                disabled_tools.update(_global_disabled)
+        except Exception:
+            pass
 
         # RAG-select relevant tools for this prompt + always-available assistant tools.
         # Without this, all 40+ tools get sent and models hit their tool limit.
@@ -1390,10 +1460,10 @@ class TaskScheduler:
             tool_idx = get_tool_index()
             if tool_idx:
                 rag_tools = tool_idx.get_tools_for_query(task.prompt or "", k=8)
-                relevant_tools = (rag_tools | ASSISTANT_ALWAYS_AVAILABLE)
-                if disabled_tools:
-                    relevant_tools -= disabled_tools
-                logger.info(f"[assistant] RAG selected {len(rag_tools)} tools + {len(ASSISTANT_ALWAYS_AVAILABLE)} always-available = {len(relevant_tools)} total for '{task.name}'")
+                relevant_tools = compose_task_relevant_tools(
+                    rag_tools, ASSISTANT_ALWAYS_AVAILABLE, disabled_tools
+                )
+                logger.info(f"[assistant] RAG selected {len(rag_tools)} tools + {len(ASSISTANT_ALWAYS_AVAILABLE)} always-available + shell/file defaults = {len(relevant_tools)} total for '{task.name}'")
         except Exception as e:
             logger.warning(f"[assistant] RAG tool selection failed, using all: {e}")
 
@@ -1401,17 +1471,23 @@ class TaskScheduler:
         try:
             result = await self._run_agent_loop(
                 endpoint_url, model, task, session_id,
-                system_prompt=system_prompt, disabled_tools=disabled_tools,
+                system_prompt=system_prompt, disabled_tools=disabled_tools or None,
                 relevant_tools=relevant_tools,
             )
         except Exception as e:
             logger.warning(f"Agent loop failed for task '{task.name}', falling back to simple call: {e}")
-            from src.llm_core import llm_call_async
+            from src.task_endpoint import task_llm_call_async
             messages = [
                 {"role": "system", "content": system_prompt},
                 {"role": "user", "content": task.prompt},
             ]
-            result = await llm_call_async(url=endpoint_url, model=model, messages=messages, timeout=120)
+            result = await task_llm_call_async(
+                messages,
+                fallback_url=endpoint_url,
+                fallback_model=model,
+                owner=task.owner,
+                timeout=120,
+            )
 
         # Strip the model's chain-of-thought before saving/delivering. Task
         # output is LLM-only, so prose=True (which also removes untagged
@@ -1636,13 +1712,17 @@ class TaskScheduler:
         # Honor per-task max_steps (defense against runaway agent loops).
         # Falls back to 20 if not set — the historical default.
         _task_max_rounds = task.max_steps if task.max_steps and task.max_steps > 0 else 20
-        # Tasks are background workloads — they share the Utility model's
-        # fallback chain (Settings → Utility Model → Fallbacks). A downed
-        # primary endpoint won't silently yield `(no output)` — same recipe
-        # chat uses but with the utility list (`utility_model_fallbacks`).
+        # Tasks are background workloads: use the shared task fallback chain
+        # behind the primary endpoint so a downed primary won't silently yield
+        # `(no output)`.
         try:
-            from src.endpoint_resolver import resolve_utility_fallback_candidates
-            _task_fallbacks = resolve_utility_fallback_candidates(owner=task.owner or None)
+            from src.task_endpoint import resolve_task_candidates
+            _task_fallbacks = resolve_task_candidates(
+                fallback_url=endpoint_url,
+                fallback_model=model,
+                fallback_headers=headers,
+                owner=task.owner or None,
+            )[1:]
         except Exception:
             _task_fallbacks = []
         async for event_str in stream_agent_loop(
@@ -1679,21 +1759,22 @@ class TaskScheduler:
         # asking it to summarize what it did. Guarantees output.
         if not full_text.strip():
             try:
-                from src.llm_core import llm_call_async_with_fallback
-                from src.endpoint_resolver import resolve_utility_fallback_candidates
+                from src.task_endpoint import task_llm_call_async
                 grace_context = "You ran out of steps. "
                 if tool_results:
                     grace_context += "Here's what your tools returned:\n" + "\n".join(tool_results[-5:])
                 else:
                     grace_context += "No tool results were captured."
                 grace_context += "\n\nSummarize what you accomplished and what's still pending. Be concise."
-                _grace_candidates = [(endpoint_url, model, headers)] + resolve_utility_fallback_candidates(owner=task.owner or None)
-                full_text = await llm_call_async_with_fallback(
-                    _grace_candidates,
+                full_text = await task_llm_call_async(
                     messages=[
                         {"role": "system", "content": system_content},
                         {"role": "user", "content": grace_context},
                     ],
+                    fallback_url=endpoint_url,
+                    fallback_model=model,
+                    fallback_headers=headers,
+                    owner=task.owner or None,
                     timeout=30,
                 )
                 full_text = (full_text or "").strip()
@@ -2202,7 +2283,7 @@ class TaskScheduler:
         # check-ins seeded, which then double-fire alongside the human user's
         # check-ins. This was the root cause of the duplicate 'Morning check-in'
         # rows we had to manually clean up.
-        if not owner or owner in {"internal-tool", "api", "demo", "system"}:
+        if not owner or owner in RESERVED_USERNAMES:
             logger.info(f"ensure_assistant_defaults: skip synthetic owner {owner!r}")
             return
         from core.database import SessionLocal, CrewMember, ScheduledTask

src/tool_execution.py

@@ -323,6 +323,24 @@ _MCP_TOOL_MAP = {
     "web_fetch":      ("web_fetch",  "web_fetch"),
     "generate_image": ("image_gen",  "generate_image"),
 }
+_EMAIL_MCP_OWNER_ARG = "_odysseus_owner"
+
+
+def _parse_qualified_mcp_args(tool: str, content: str) -> tuple[Dict, Optional[str]]:
+    raw = (content or "").strip()
+    if not raw:
+        return {}, None
+    try:
+        parsed = json.loads(raw)
+    except (json.JSONDecodeError, TypeError):
+        if tool.startswith("mcp__email__"):
+            return {}, "Email MCP tool arguments must be a JSON object."
+        return {}, None
+    if not isinstance(parsed, dict):
+        if tool.startswith("mcp__email__"):
+            return {}, "Email MCP tool arguments must be a JSON object."
+        return {}, None
+    return parsed, None
 
 
 def _parse_generate_image(content: str) -> Dict:
@@ -453,6 +471,8 @@ async def _direct_fallback(
     tool: str,
     content: str,
     progress_cb: Optional[Callable[[Dict], Awaitable[None]]] = None,
+    session_id: Optional[str] = None,
+    owner: Optional[str] = None,
 ) -> Optional[Dict]:
     _subproc_env = {
         **os.environ,
@@ -466,6 +486,8 @@ async def _direct_fallback(
         ctx = {
             "progress_cb": progress_cb,
             "subproc_env": _subproc_env,
+            "session_id": session_id,
+            "owner": owner,
         }
 
         from src.agent_tools import TOOL_HANDLERS
@@ -713,10 +735,13 @@ async def _execute_tool_block_impl(
             desc = f"bash (background): {short}"
             result = {
                 "output": (
-                    f"Started background job `{rec['id']}`. It is running detached — "
+                    f"Started background job `{rec['id']}`. It is running detached; "
                     f"do NOT wait for it or poll it. You will be automatically re-invoked "
                     f"with its full output when it finishes. Continue with other work, or "
-                    f"end your turn now and resume when the result arrives."
+                    f"end your turn now and resume when the result arrives. If the user "
+                    f"later asks to check progress or stop it, call the manage_bg_jobs "
+                    f"tool yourself (output or kill); do not tell them to run a tool "
+                    f"command, and do not surface raw tool syntax in your reply."
                 ),
                 "exit_code": 0,
                 "bg_job_id": rec["id"],
@@ -737,6 +762,11 @@ async def _execute_tool_block_impl(
         desc = f"{tool}: {first_line}"
         result = await _direct_fallback(tool, content, progress_cb=progress_cb) \
             or {"error": f"{tool}: execution failed", "exit_code": 1}
+    elif tool == "manage_bg_jobs":
+        # Inspect/kill detached `bash` jobs; needs session_id to scope to chat.
+        desc = f"manage_bg_jobs: {content.split(chr(10))[0][:80]}"
+        result = await _direct_fallback(tool, content, session_id=session_id, owner=owner) \
+            or {"error": "manage_bg_jobs: execution failed", "exit_code": 1}
     elif tool in ("create_document", "update_document", "edit_document",
                   "suggest_document", "manage_documents"):
         desc = f"{tool}: {content.split(chr(10))[0][:80]}"
@@ -748,10 +778,24 @@ async def _execute_tool_block_impl(
         query = content.split("\n")[0].strip()
         desc = f"search_chats: {query[:80]}"
         result = await do_search_chats(query, owner=owner)
-    elif tool in ("chat_with_model", "create_session", "list_sessions",
-                  "send_to_session", "pipeline",
-                  "manage_session", "manage_memory", "list_models",
-                  "ui_control", "ask_teacher"):
+    elif tool in ("chat_with_model", "ask_teacher", "list_models"):
+        # Migrated to the agent_tools registry (#3629): dispatched through
+        # TOOL_HANDLERS with the owner/session ctx these tools need, instead
+        # of the legacy dispatch_ai_tool elif. The impls live in
+        # src/agent_tools/model_interaction_tools.py.
+        first_line = content.split(chr(10))[0].strip()[:60]
+        desc = f"{tool}: {first_line}" if first_line else tool
+        result = await _document_tool_dispatch(tool, content, session_id, owner) \
+            or {"error": f"{tool}: execution failed", "exit_code": 1}
+    elif tool in ("create_session", "list_sessions", "send_to_session", "manage_session"):
+        # Migrated to the agent_tools registry (#3629): dispatched through
+        # TOOL_HANDLERS with the owner/session ctx these tools need. The impls
+        # live in src/agent_tools/session_tools.py.
+        first_line = content.split(chr(10))[0].strip()[:60]
+        desc = f"{tool}: {first_line}" if first_line else tool
+        result = await _document_tool_dispatch(tool, content, session_id, owner) \
+            or {"error": f"{tool}: execution failed", "exit_code": 1}
+    elif tool in ("pipeline", "manage_memory", "ui_control"):
         from src.ai_interaction import dispatch_ai_tool
         desc, result = await dispatch_ai_tool(tool, content, session_id, owner=owner)
     elif tool == "manage_tasks":
@@ -858,12 +902,15 @@ async def _execute_tool_block_impl(
         # MCP tool dispatch
         mcp = get_mcp_manager()
         if mcp:
-            try:
-                args = json.loads(content) if content.strip().startswith("{") else {}
-            except (json.JSONDecodeError, TypeError):
-                args = {}
             desc = f"mcp: {tool}"
-            result = await mcp.call_tool(tool, args)
+            args, parse_error = _parse_qualified_mcp_args(tool, content)
+            if parse_error:
+                result = {"error": parse_error, "exit_code": 1}
+            else:
+                if tool.startswith("mcp__email__") and owner:
+                    args = dict(args)
+                    args[_EMAIL_MCP_OWNER_ARG] = owner
+                result = await mcp.call_tool(tool, args)
         else:
             desc = f"mcp: {tool}"
             result = {"error": "MCP manager not available", "exit_code": 1}

src/tool_implementations.py

@@ -12,12 +12,24 @@ import os
 import re
 from typing import Any, Dict, List, Optional
 
+from fastapi import HTTPException
 from src.constants import MAX_READ_CHARS, DEEP_RESEARCH_DIR, VAULT_FILE
 from src.tool_utils import get_mcp_manager
 from core.constants import internal_api_base
+from routes._validators import validate_remote_host, validate_ssh_port
 
 logger = logging.getLogger(__name__)
 
+
+def _string_arg(value: Any) -> str:
+    return "" if value is None else str(value).strip()
+
+
+def _validate_cookbook_ssh_target(remote_host: Any, ssh_port: Any = "") -> tuple[str, str]:
+    remote = validate_remote_host(_string_arg(remote_host) or None) or ""
+    sport = validate_ssh_port(_string_arg(ssh_port) or None) or ""
+    return remote, sport
+
 # ---------------------------------------------------------------------------
 # Active email state
 # ---------------------------------------------------------------------------
@@ -645,6 +657,137 @@ async def do_manage_endpoints(content: str, owner: Optional[str] = None) -> Dict
 # MCP server management tool
 # ---------------------------------------------------------------------------
 
+# Parallel to routes/cookbook_helpers._validate_serve_cmd but deliberately the
+# opposite policy: that gate guards an admin-only serve command and allows
+# interpreters (python3/etc) because model-serving needs them, whereas this is
+# the model/prompt-injection-reachable manage_mcp path, so interpreters and
+# runners are denied here.
+#
+# Commands that can execute arbitrary code regardless of their arguments. These
+# are NEVER accepted on the manage_mcp agent path, even if an operator lists one
+# in ODYSSEUS_MCP_ALLOWED_COMMANDS -- a stdio server that genuinely needs an
+# interpreter or package runner must be registered via the trusted admin route.
+_MCP_DENIED_COMMANDS = frozenset({
+    "sh", "bash", "zsh", "fish", "dash", "ksh", "csh", "tcsh", "ash", "busybox",
+    "cmd", "command.com", "powershell", "pwsh",
+    "python", "pypy", "node", "nodejs", "deno", "bun", "ruby", "jruby",
+    "perl", "raku", "php", "lua", "luajit", "tclsh", "wish", "expect", "rscript",
+    "groovy", "scala", "elixir", "erl", "iex", "java", "javac", "jshell", "jbang",
+    "kotlin", "kotlinc", "dotnet", "mono", "swift", "osascript", "tsx", "ts-node",
+    "npx", "bunx", "uvx", "pipx", "npm", "pnpm", "yarn", "pip", "uv",
+    "gem", "cargo", "go", "bundle", "poetry", "conda", "mamba", "brew",
+    "apt", "apt-get", "yum", "dnf", "pacman", "apk",
+    "env", "xargs", "nohup", "setsid", "nice", "ionice", "time", "timeout",
+    "watch", "stdbuf", "unbuffer", "script", "ssh", "scp", "sshpass", "sudo",
+    "doas", "su", "make", "cmake", "docker", "podman", "kubectl", "find",
+    "awk", "gawk", "sed", "vi", "vim", "nvim", "emacs", "ed", "tee", "eval",
+})
+
+# Argv flags that make even an allowlisted binary execute inline code. Matched
+# by prefix so glued forms (-cimport os, --eval=...) are caught, not just the
+# exact-token form.
+_MCP_CODE_EXEC_SHORT_FLAGS = ("-c", "-e", "-m")
+_MCP_CODE_EXEC_LONG_FLAGS = ("--eval", "--exec", "--print", "--module", "--command", "--require")
+
+_MCP_URL_SCHEMES = ("http://", "https://", "ftp://", "ftps://", "file://", "data:", "jar:", "blob:")
+
+# Shell metacharacters refused in command/args. Args are passed as an argv list
+# (no shell), but refusing these keeps the surface narrow and obvious.
+_MCP_SHELL_METACHARS = set(";|&$`><\n\r")
+
+# Env vars that let a child process load attacker-supplied code before main().
+_MCP_DANGEROUS_ENV = frozenset({
+    "LD_PRELOAD", "LD_LIBRARY_PATH", "LD_AUDIT", "DYLD_INSERT_LIBRARIES",
+    "DYLD_LIBRARY_PATH", "DYLD_FRAMEWORK_PATH", "PYTHONPATH", "PYTHONSTARTUP",
+    "PYTHONHOME", "PYTHONEXECUTABLE", "NODE_OPTIONS", "NODE_PATH", "BASH_ENV",
+    "ENV", "SHELLOPTS", "PERL5LIB", "PERL5OPT", "RUBYOPT", "RUBYLIB", "GEM_PATH",
+    "R_PROFILE", "R_HOME", "PATH", "IFS", "PROMPT_COMMAND",
+})
+
+
+def _mcp_allowed_commands() -> set:
+    """Operator-configured allowlist of safe MCP launcher basenames for the agent
+    path. Empty by default; set ODYSSEUS_MCP_ALLOWED_COMMANDS (comma-separated)
+    to opt specific trusted binaries in. Denied commands are rejected even if
+    listed here."""
+    raw = os.environ.get("ODYSSEUS_MCP_ALLOWED_COMMANDS", "")
+    return {c.strip().lower() for c in raw.split(",") if c.strip()}
+
+
+def _validate_mcp_command(command, args, env) -> Optional[str]:
+    """Validate a model-supplied stdio MCP registration. Returns an error string
+    if it must be rejected, else None.
+
+    Closes the RCE where manage_mcp 'add' passed prompt-injection-controlled
+    command/args/env straight to a subprocess spawn (issue #438): a payload
+    smuggled into a skill description, memory entry, fetched page, or email body
+    could register a stdio server running arbitrary code as the app UID.
+    """
+    if not isinstance(command, str) or not command.strip():
+        return "command must be a non-empty string"
+    command = command.strip()
+    if "/" in command or "\\" in command:
+        return "command must be a bare executable name, not a path"
+    if any(ch in _MCP_SHELL_METACHARS for ch in command):
+        return "command contains shell metacharacters"
+    base = command.lower()
+    if base.endswith(".exe") or base.endswith(".cmd") or base.endswith(".bat"):
+        base = base.rsplit(".", 1)[0]
+    # Canonicalize a trailing version suffix so versioned aliases collapse to the
+    # family name (python3.11 -> python, node18 -> node, pip3 -> pip); both the
+    # raw basename and the canonical form are denied, so an operator cannot
+    # accidentally allowlist a runtime alias back into the path.
+    canon = re.sub(r"[-_.]?\d+(?:\.\d+)*$", "", base)
+    if base in _MCP_DENIED_COMMANDS or canon in _MCP_DENIED_COMMANDS:
+        return (
+            f"command '{command}' is not allowed on the agent MCP path: "
+            "interpreters, runtimes, package runners, and shells can execute "
+            "arbitrary code. Register such a server via the admin route instead."
+        )
+    if base not in _mcp_allowed_commands():
+        return (
+            f"command '{command}' is not in the MCP allowlist. Add it to "
+            "ODYSSEUS_MCP_ALLOWED_COMMANDS if you trust it, or register the "
+            "server via the admin route."
+        )
+
+    if args is not None:
+        if isinstance(args, str):
+            try:
+                args = json.loads(args)
+            except Exception:
+                return "args must be a JSON list"
+        if not isinstance(args, list):
+            return "args must be a list"
+        for a in args:
+            if not isinstance(a, str):
+                return "args must all be strings"
+            s = a.strip()
+            low = s.lower()
+            if any(s == f or s.startswith(f) for f in _MCP_CODE_EXEC_SHORT_FLAGS):
+                return f"arg '{a}' is a code-execution flag and is not allowed"
+            if any(low == f or low.startswith(f + "=") for f in _MCP_CODE_EXEC_LONG_FLAGS):
+                return f"arg '{a}' is a code-execution flag and is not allowed"
+            if any(low.startswith(u) for u in _MCP_URL_SCHEMES):
+                return f"arg '{a}' is a remote URL and is not allowed"
+            if any(ch in _MCP_SHELL_METACHARS for ch in a):
+                return f"arg '{a}' contains shell metacharacters"
+
+    if env:
+        if isinstance(env, str):
+            try:
+                env = json.loads(env)
+            except Exception:
+                return "env must be a JSON object"
+        if not isinstance(env, dict):
+            return "env must be an object"
+        for k in env:
+            if str(k).strip().upper() in _MCP_DANGEROUS_ENV:
+                return f"env var '{k}' can inject code into the child process and is not allowed"
+
+    return None
+
+
 async def do_manage_mcp(content: str, owner: Optional[str] = None) -> Dict:
     """Manage MCP servers: list, add, delete, enable, disable, reconnect."""
     try:
@@ -684,6 +827,12 @@ async def do_manage_mcp(content: str, owner: Optional[str] = None) -> Dict:
         env = args.get("env", {})
         if not name or not command:
             return {"error": "name and command are required", "exit_code": 1}
+        # Validate BEFORE any DB write or spawn: a rejected registration must
+        # leave no enabled row (which would otherwise auto-reconnect on restart)
+        # and must not attempt a connection.
+        _mcp_err = _validate_mcp_command(command, cmd_args, env)
+        if _mcp_err:
+            return {"error": f"manage_mcp: refused unsafe server registration: {_mcp_err}", "exit_code": 1}
         sid = str(_uuid.uuid4())[:8]
         db = SessionLocal()
         try:
@@ -1119,8 +1268,8 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
             _ALIASES = {
                 "shell": ["bash"],
                 "terminal": ["bash"],
-                "search": ["web_search"],
-                "web": ["web_search"],
+                "search": ["web_search", "web_fetch"],
+                "web": ["web_search", "web_fetch"],
                 "browser": ["builtin_browser"],
                 "documents": ["create_document", "edit_document", "update_document", "suggest_document"],
                 "doc": ["create_document", "edit_document", "update_document", "suggest_document"],
@@ -1132,7 +1281,7 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
                 "notes": ["manage_notes"],
                 "calendar": ["manage_calendar"],
                 "email": ["mcp__email__list_emails", "mcp__email__read_email", "mcp__email__send_email"],
-                "research": ["web_search"],  # research is a per-request flag, not a tool — closest analog
+                "research": ["web_search", "web_fetch"],  # research is a per-request flag, not a tool — closest analog
             }
 
             if action == "list_tools":
@@ -2714,13 +2863,25 @@ async def do_serve_model(content: str, owner: Optional[str] = None) -> Dict:
                 endpoint_added=endpoint_added, endpoint_id=endpoint_id or "",
             )
             note = "" if registered else " (state-write failed — task may not show in UI)"
+            where = host or "local"
+            log_path = f"/tmp/odysseus-tmux/{sid}.log"
             return {
-                "output": f"Serving {repo_id} (session: {sid}){note}",
+                "output": (
+                    f"Serving {repo_id} on {where} (session: {sid}){note}\n"
+                    f"Next required check: call list_served_models. If this task is not ready, "
+                    f"call tail_serve_output with session_id={sid} and tail=400 before answering. "
+                    f"Do not tell the user to check logs; you have the log tool."
+                ),
                 "session_id": sid,
                 "task_type": "serve",
                 "phase": "running",
                 "host": host,
                 "endpoint_id": endpoint_id,
+                "log_path": log_path,
+                "next_tools": [
+                    {"name": "list_served_models", "arguments": {}},
+                    {"name": "tail_serve_output", "arguments": {"session_id": sid, "tail": 400}},
+                ],
                 "exit_code": 0,
             }
         # FastAPI HTTPException puts the message under `detail`, not `error`.
@@ -2888,6 +3049,10 @@ async def _cookbook_kill_session(session_id: str, *, remote_host: str = "",
             break
 
     if remote:
+        try:
+            remote, sport = _validate_cookbook_ssh_target(remote, sport)
+        except HTTPException as e:
+            return {"error": str(getattr(e, "detail", e)), "exit_code": 1}
         _pf = f"-p {shlex.quote(str(sport))} " if sport and str(sport) != "22" else ""
         cmd = (
             f"ssh -o ConnectTimeout=5 -o StrictHostKeyChecking=no "
@@ -2976,8 +3141,8 @@ async def do_tail_serve_output(content: str, owner: Optional[str] = None) -> Dic
         tail = 400
     tail = max(20, min(tail, 4000))
     headers = _internal_headers()
-    remote = (args.get("remote_host") or args.get("host") or "").strip()
-    sport = (args.get("ssh_port") or "").strip()
+    remote = _string_arg(args.get("remote_host") or args.get("host"))
+    sport = _string_arg(args.get("ssh_port"))
     # Resolve host from cookbook state if caller didn't pass one — same
     # lookup _cookbook_kill_session uses.
     if not remote:
@@ -2995,6 +3160,12 @@ async def do_tail_serve_output(content: str, owner: Optional[str] = None) -> Dic
                     if not sport:
                         sport = t.get("sshPort") or ""
                     break
+    if remote:
+        try:
+            remote, sport = _validate_cookbook_ssh_target(remote, sport)
+        except HTTPException as e:
+            return {"error": str(getattr(e, "detail", e)), "exit_code": 1}
+
     # Prefer the persisted /tmp/odysseus-tmux/SESSION.log file over the
     # live tmux pane. The pane is what the user would see scrolling on
     # their screen — including the post-crash neofetch banner and the
@@ -3057,8 +3228,17 @@ async def do_tail_serve_output(content: str, owner: Optional[str] = None) -> Dic
         MAX_CHARS = 8000
         if len(output_text) > MAX_CHARS:
             output_text = "…(earlier output truncated)…\n" + output_text[-MAX_CHARS:]
+        if not output_text:
+            output_text = (
+                f"No log output captured yet for {session_id} on {host_label}. "
+                "This usually means the tmux wrapper has started but the model process "
+                "has not printed anything yet. Do not stop here: call list_served_models "
+                "again to check whether it is still loading, ready, or crashed; if it is "
+                "still not ready, call tail_serve_output again with a larger tail after "
+                "the next status check."
+            )
         return {
-            "output": output_text or "(empty pane)",
+            "output": output_text,
             "session_id": session_id,
             "host": host_label,
             "tail_lines": tail,
@@ -3172,7 +3352,7 @@ async def do_adopt_served_model(content: str, owner: Optional[str] = None) -> Di
     except ValueError:
         return {"error": "Invalid JSON arguments", "exit_code": 1}
 
-    host = (args.get("host") or args.get("remote_host") or "").strip()
+    host = _string_arg(args.get("host") or args.get("remote_host"))
     sess = (args.get("tmux_session") or args.get("session_id") or "").strip()
     model = (args.get("model") or args.get("repo_id") or "").strip()
     port = args.get("port") or 8000
@@ -3183,6 +3363,12 @@ async def do_adopt_served_model(content: str, owner: Optional[str] = None) -> Di
         return {"error": "tmux_session and model are required", "exit_code": 1}
 
     # Verify tmux session exists on the target host
+    if host:
+        try:
+            host, _ = _validate_cookbook_ssh_target(host)
+        except HTTPException as e:
+            return {"error": str(getattr(e, "detail", e)), "exit_code": 1}
+
     headers = _internal_headers()
     if host:
         check = f"ssh -o ConnectTimeout=5 -o StrictHostKeyChecking=no {shlex.quote(host)} 'tmux has-session -t {shlex.quote(sess)} 2>&1'"
@@ -3797,7 +3983,7 @@ async def do_resolve_contact(content: str, owner: Optional[str] = None) -> Dict:
     if not name:
         return {"error": "name is required", "exit_code": 1}
 
-    contacts = {}  # email -> {name, source}
+    contacts = {}  # email_or_phone -> {name, source, phone?}
 
     # 1. CardDAV (Radicale) — structured contacts. Call in-process: a
     # server-side httpx GET to /api/contacts/search carries no session
@@ -3812,10 +3998,18 @@ async def do_resolve_contact(content: str, owner: Optional[str] = None) -> Dict:
             match = q in hay_name or any(q in (e or "").lower() for e in c.get("emails", []))
             if not match:
                 continue
+            has_email = False
             for email in (c.get("emails") or []):
                 email = (email or "").strip().lower()
                 if email and "@" in email:
                     contacts[email] = {"name": c.get("name") or email, "source": "contacts"}
+                    has_email = True
+            # Fall back to phone numbers when the contact has no email address
+            if not has_email:
+                for phone in (c.get("phones") or []):
+                    phone = (phone or "").strip()
+                    if phone:
+                        contacts[phone] = {"name": c.get("name") or phone, "source": "contacts", "phone": phone}
     except Exception:
         pass
 
@@ -3835,8 +4029,11 @@ async def do_resolve_contact(content: str, owner: Optional[str] = None) -> Dict:
         return {"output": f"No contacts found matching '{name}'.", "exit_code": 0}
 
     lines = [f"Contacts matching '{name}':"]
-    for email, info in contacts.items():
-        lines.append(f"- {info['name']} <{email}> ({info['source']})")
+    for key, info in contacts.items():
+        if info.get("phone"):
+            lines.append(f"- {info['name']} — phone: {info['phone']} ({info['source']})")
+        else:
+            lines.append(f"- {info['name']} <{key}> ({info['source']})")
     return {"output": "\n".join(lines), "exit_code": 0}
 
 

src/tool_index.py

@@ -94,6 +94,7 @@ BUILTIN_TOOL_DESCRIPTIONS: Dict[str, str] = {
     "manage_endpoints": "Endpoint management: list, add, delete, enable, or disable model API endpoints.",
     "manage_mcp": "MCP server management: list, add, delete, reconnect servers, or list available tools.",
     "manage_webhooks": "Webhook management: list, add, delete, enable, or disable webhooks.",
+    "api_call": "Call a configured API integration by name (Home Assistant, Miniflux, Gitea, Linkding, Jellyfin, RSS reader, git forge, bookmark manager, smart home, or any other registered service). Make a GET/POST/PUT/PATCH/DELETE request to the integration's endpoint path, with an optional JSON body. Use whenever the user asks to query or control one of their connected integrations/services.",
     "manage_tokens": "API token management: list, create, or delete API access tokens.",
     "manage_documents": "List, read, delete, or tidy documents in the editor panel. action='list' returns clickable rows (most-recent first) so the user can open any doc by clicking. action='read' (aka view/open/get) with document_id returns the content; supports offset=<N> + limit=<N> to page through large docs (response includes next_offset when more remains, so you can keep calling with offset=next_offset). action='delete' with document_id removes a doc (only way to delete). Use this for ANY 'show/read/list/open my documents/docs/files/notes' request — never shell or curl.",
     "manage_research": "List, read/open, or delete saved DEEP RESEARCH results from the Library. action='list' returns clickable [query](#research-<id>) rows (most-recent first). action='read' (aka open/view/get) with id returns the report + sources. action='delete' with id removes it. Use this for ANY 'open/read/find/delete my research / that report / the research on X' request. NOTE: this is for EXISTING research; to START new research use trigger_research.",
@@ -134,6 +135,7 @@ BUILTIN_TOOL_DESCRIPTIONS: Dict[str, str] = {
     "app_api": "Generic loopback to allowed Odysseus internal endpoints. Use this when the user wants something the UI can do but there's no named tool for it. Covers calendar, gallery, library/documents, memory, notes, tasks, settings, research, compare, cookbook GPUs/state — allowed UI buttons hit /api/* endpoints and you can hit them too. Sensitive auth/user/admin/shell paths and host-control Cookbook mutation routes are blocked; do NOT use app_api for shell commands, package installs, engine rebuilds, or PID signalling. Use named command tooling for shell commands. action='endpoints' with filter=<keyword> lists available endpoints. action='call' takes method+path+body. Hits same routes the UI uses — auth flows free. NOTE: themes are NOT an API endpoint — use the ui_control tool (create_theme / set_theme), not app_api. SESSIONS/CHATS: do NOT use app_api for these — GET /api/sessions returns EMPTY for tool calls (it's owner-filtered and tool calls authenticate as a different identity). EMAIL ACCOUNTS: do NOT use /api/email/accounts via app_api; use list_email_accounts, list_emails, and read_email instead. To list/rename/archive/delete/fork chats use the list_sessions and manage_session tools instead.",
     "edit_image": "Edit an image in the gallery: upscale (increase resolution), remove background (rembg), inpaint (fill selected area), or harmonize (blend edits). Specify image ID and action.",
     "trigger_research": "Start a deep research job on any topic — appears in the Deep Research sidebar, streams progress, produces a detailed report. Use for 'research X', 'look into Y', 'do deep research on Z', 'investigate'. NOT a scheduled task — it runs now and surfaces in the sidebar.",
+    "manage_bg_jobs": "Inspect and control detached background `bash` jobs (the ones started with a `#!bg` marker). action='list' shows this chat's jobs (id/status/age/command); action='output' returns a job's captured output so far (check on a long-running job, or re-read a finished one); action='kill' stops a runaway job by id. Use for 'is the background job done', 'check on that job', 'show the build output', 'kill the background job', 'stop the bg task'. output/kill need a job_id from list.",
 }
 
 
@@ -348,6 +350,12 @@ class ToolIndex:
             {"list_email_accounts", "list_emails", "read_email", "send_email", "reply_to_email", "bulk_email", "delete_email", "archive_email", "mark_email_read", "resolve_contact", "ui_control"},
         frozenset({"calendar", "event", "meeting", "schedule", "appointment"}):
             {"manage_calendar"},
+        # Detached background `bash` jobs (#!bg): check on / read output / kill.
+        frozenset({"background job", "background jobs", "bg job", "bg jobs",
+                   "background task", "is the job done", "check the job",
+                   "check on that job", "job output", "kill the job",
+                   "kill the background", "stop the background", "running job"}):
+            {"manage_bg_jobs"},
         frozenset({"note", "todo", "reminder", "remind", "checklist", "remember to"}):
             {"manage_notes"},
         # Chat/session management. "rename" alone maps to documents below, so a
@@ -414,6 +422,14 @@ class ToolIndex:
                    "my settings", "change setting", "change a setting", "set setting",
                    "preference", "preferences", "configure"}):
             {"manage_settings", "ui_control"},
+        # API-integration intent → the api_call tool. Mirrors the agent-loop
+        # "integrations" domain so api_call still surfaces on the retrieval and
+        # keyword-fallback paths (not just the deterministic domain seed) when a
+        # user names a connected service.
+        frozenset({"api_call", "api call", "integration", "integrations",
+                   "home assistant", "homeassistant", "miniflux", "gitea",
+                   "linkding", "jellyfin"}):
+            {"api_call"},
         # Managing EXISTING research in the Library — open/read/find/delete.
         frozenset({"my research", "the research", "research on", "open research",
                    "read research", "find research", "delete research",

src/tool_parsing.py

@@ -39,6 +39,10 @@ _XML_TOOL_CALL_RE = re.compile(
     r"<(?:[\w]+:)?(?:tool_call|function_call)>\s*([\s\S]*?)</(?:[\w]+:)?(?:tool_call|function_call)>",
     re.IGNORECASE,
 )
+_XML_OPEN_TOOL_CALL_RE = re.compile(
+    r"<(?:[\w]+:)?(?:tool_call|function_call)>\s*([\s\S]*)\Z",
+    re.IGNORECASE,
+)
 _XML_INVOKE_RE = re.compile(
     r'<invoke\s+name=["\'](\w+)["\']>\s*([\s\S]*?)</invoke>',
     re.IGNORECASE,
@@ -47,6 +51,21 @@ _XML_PARAM_RE = re.compile(
     r'<parameter\s+name=["\'](\w+)["\']>([\s\S]*?)</parameter>',
     re.IGNORECASE,
 )
+_XML_DIRECT_TOOL_RE = re.compile(
+    r"<\s*([A-Za-z_][\w-]*)\s*>([\s\S]*?)</\s*\1\s*>",
+    re.IGNORECASE,
+)
+
+# Pattern 3b: StepFun Step-3.x native tool-call tokens. The tokenizer defines:
+#   <｜tool▁calls▁begin｜> ... <｜tool▁calls▁end｜>
+#   <｜tool▁call▁begin｜>tool_name<｜tool▁sep｜>{...}<｜tool▁call▁end｜>
+# These can leak as text through llama.cpp/Ollama-style endpoints when the
+# engine does not return structured OpenAI tool_calls.
+_STEPFUN_CALL_BEGIN = "<｜tool▁call▁begin｜>"
+_STEPFUN_CALL_SEP = "<｜tool▁sep｜>"
+_STEPFUN_CALL_END = "<｜tool▁call▁end｜>"
+_STEPFUN_CALLS_BEGIN = "<｜tool▁calls▁begin｜>"
+_STEPFUN_CALLS_END = "<｜tool▁calls▁end｜>"
 
 # Pattern 4: <tool_code> blocks (MiniMax-M2.5 style)
 # {tool => 'tool_name', args => '<param>value</param>'}
@@ -175,6 +194,9 @@ _TOOL_NAME_MAP = {
     "notes": "manage_notes",
     "todo": "manage_notes",
     "todos": "manage_notes",
+    "manage_bg_jobs": "manage_bg_jobs",
+    "bg_jobs": "manage_bg_jobs",
+    "background_jobs": "manage_bg_jobs",
 }
 
 _MISFENCED_WEB_TOOL_NAMES = {
@@ -443,6 +465,138 @@ def _parse_xml_invoke(inv_match) -> Optional[ToolBlock]:
     return function_call_to_tool_block(tool_name, json.dumps(params))
 
 
+def _parse_xml_direct_tool(tool_match) -> Optional[ToolBlock]:
+    """Parse direct XML tool tags inside <tool_call>.
+
+    Some local models emit:
+      <tool_call><web_search>query</web_search></tool_call>
+    instead of the invoke/parameter shape:
+      <tool_call><invoke name="web_search"><parameter name="query">query</parameter></invoke></tool_call>
+    Keep this as an adapter to the canonical function-call converter so aliases
+    and per-tool argument formatting stay in one place.
+    """
+    tool_name = tool_match.group(1).lower().replace("-", "_")
+    if tool_name in {"invoke", "parameter", "tool_call", "function_call"}:
+        return None
+    mapped = _TOOL_NAME_MAP.get(tool_name) or (tool_name if tool_name in TOOL_TAGS else None)
+    if not mapped:
+        return None
+    body = tool_match.group(2).strip()
+    if not body:
+        return None
+    try:
+        params = json.loads(body)
+        if not isinstance(params, dict):
+            params = {}
+    except json.JSONDecodeError:
+        if mapped == "web_search":
+            params = {"query": body}
+        elif mapped == "web_fetch":
+            params = {"url": body}
+        elif mapped == "bash":
+            params = {"command": body}
+        elif mapped == "python":
+            params = {"code": body}
+        elif mapped in ("read_file", "write_file"):
+            params = {"path": body}
+        else:
+            params = {"content": body}
+    from src.tool_schemas import function_call_to_tool_block
+    return function_call_to_tool_block(mapped, json.dumps(params))
+
+
+def _iter_stepfun_tool_calls(text: str):
+    """Yield StepFun native tool-call token bodies without regex backtracking."""
+    pos = 0
+    while True:
+        start = text.find(_STEPFUN_CALL_BEGIN, pos)
+        if start < 0:
+            return
+        name_start = start + len(_STEPFUN_CALL_BEGIN)
+        sep = text.find(_STEPFUN_CALL_SEP, name_start)
+        if sep < 0:
+            return
+        end = text.find(_STEPFUN_CALL_END, sep + len(_STEPFUN_CALL_SEP))
+        if end < 0:
+            return
+        raw_name = text[name_start:sep].strip()
+        body = text[sep + len(_STEPFUN_CALL_SEP):end].strip()
+        if raw_name and len(raw_name) <= 128:
+            yield raw_name, body
+        pos = end + len(_STEPFUN_CALL_END)
+
+
+def _strip_stepfun_tool_markup(text: str) -> str:
+    """Remove StepFun tool-call token blocks and wrappers using literal scans."""
+    out = []
+    pos = 0
+    while True:
+        start = text.find(_STEPFUN_CALL_BEGIN, pos)
+        if start < 0:
+            out.append(text[pos:])
+            break
+        end = text.find(_STEPFUN_CALL_END, start + len(_STEPFUN_CALL_BEGIN))
+        if end < 0:
+            out.append(text[pos:])
+            break
+        out.append(text[pos:start])
+        pos = end + len(_STEPFUN_CALL_END)
+    cleaned = "".join(out)
+    return cleaned.replace(_STEPFUN_CALLS_BEGIN, "").replace(_STEPFUN_CALLS_END, "")
+
+
+def _strip_bare_invoke_markup(text: str) -> str:
+    """Remove bare <invoke ...>...</invoke> blocks without regex backtracking."""
+    out = []
+    pos = 0
+    while True:
+        start = text.lower().find("<invoke", pos)
+        if start < 0:
+            out.append(text[pos:])
+            break
+        tag_end = text.find(">", start)
+        if tag_end < 0:
+            out.append(text[pos:])
+            break
+        close = text.lower().find("</invoke>", tag_end + 1)
+        if close < 0:
+            out.append(text[pos:])
+            break
+        out.append(text[pos:start])
+        pos = close + len("</invoke>")
+    return "".join(out)
+
+
+def _parse_stepfun_tool_call(tool_name: str, body: str) -> Optional[ToolBlock]:
+    """Parse StepFun native tool-call tokens into an Odysseus ToolBlock."""
+    tool_name = tool_name.lower().replace("-", "_").replace(".", "_")
+    mapped = _TOOL_NAME_MAP.get(tool_name) or (tool_name if tool_name in TOOL_TAGS else None)
+    if not mapped:
+        return None
+    body = (body or "").strip()
+    if not body:
+        return None
+    try:
+        params = json.loads(body)
+        if not isinstance(params, dict):
+            params = {}
+    except json.JSONDecodeError:
+        if mapped == "web_search":
+            params = {"query": body}
+        elif mapped == "web_fetch":
+            params = {"url": body}
+        elif mapped == "bash":
+            params = {"command": body}
+        elif mapped == "python":
+            params = {"code": body}
+        elif mapped in ("read_file", "write_file"):
+            params = {"path": body}
+        else:
+            params = {"content": body}
+    from src.tool_schemas import function_call_to_tool_block
+    return function_call_to_tool_block(mapped, json.dumps(params))
+
+
 def _parse_tool_code_block(raw: str) -> Optional[ToolBlock]:
     """Parse a <tool_code>{tool => 'name', args => '...'}</tool_code> block (MiniMax style)."""
     # Extract tool name
@@ -508,8 +662,9 @@ def parse_tool_blocks(text: str, skip_fenced: bool = False) -> List[ToolBlock]:
     2. [TOOL_CALL] ... [/TOOL_CALL] blocks (some models)
     3. XML-style <tool_call>/<invoke> blocks
     4. <tool_code> blocks (MiniMax-M2.5 style)
-    5. DeepSeek DSML markup (normalized to <invoke> first)
-    6. Non-native local model fallback: prose mentioning web_search followed by
+    5. StepFun Step-3 native <｜tool▁call▁begin｜> tokens
+    6. DeepSeek DSML markup (normalized to <invoke> first)
+    7. Non-native local model fallback: prose mentioning web_search followed by
        bare JSON args, e.g. {"query":"...", "time_filter":"week"}
 
     `skip_fenced`: when True, Pattern 1 (fenced ```bash/```python/```json code
@@ -564,12 +719,38 @@ def parse_tool_blocks(text: str, skip_fenced: bool = False) -> List[ToolBlock]:
 
     # Pattern 3: XML-style <tool_call>/<invoke> blocks
     if not blocks:
+        for tool_name, body in _iter_stepfun_tool_calls(text):
+            block = _parse_stepfun_tool_call(tool_name, body)
+            if block:
+                blocks.append(block)
+        if blocks:
+            return blocks
         # Try wrapped: <tool_call><invoke ...>...</invoke></tool_call>
         for m in _XML_TOOL_CALL_RE.finditer(text):
             for inv in _XML_INVOKE_RE.finditer(m.group(1)):
                 block = _parse_xml_invoke(inv)
                 if block:
                     blocks.append(block)
+            if not blocks:
+                for direct in _XML_DIRECT_TOOL_RE.finditer(m.group(1)):
+                    block = _parse_xml_direct_tool(direct)
+                    if block:
+                        blocks.append(block)
+        # Some local models stream an opening <tool_call> wrapper and a
+        # complete inner tool tag, but forget the closing </tool_call>.
+        if not blocks:
+            for m in _XML_OPEN_TOOL_CALL_RE.finditer(text):
+                body = m.group(1)
+                for inv in _XML_INVOKE_RE.finditer(body):
+                    block = _parse_xml_invoke(inv)
+                    if block:
+                        blocks.append(block)
+                if blocks:
+                    break
+                for direct in _XML_DIRECT_TOOL_RE.finditer(body):
+                    block = _parse_xml_direct_tool(direct)
+                    if block:
+                        blocks.append(block)
         # Try bare <invoke> without wrapper
         if not blocks:
             for inv in _XML_INVOKE_RE.finditer(text):
@@ -611,7 +792,9 @@ def strip_tool_blocks(text: str, skip_fenced: bool = False) -> str:
     text = _normalize_dsml(text)
     cleaned = text if skip_fenced else _TOOL_BLOCK_RE.sub('', text)
     cleaned = _TOOL_CALL_RE.sub('', cleaned)
+    cleaned = _strip_stepfun_tool_markup(cleaned)
     cleaned = _XML_TOOL_CALL_RE.sub('', cleaned)
+    cleaned = _XML_OPEN_TOOL_CALL_RE.sub('', cleaned)
     cleaned = _TOOL_CODE_RE.sub('', cleaned)
     if not skip_fenced:
         raw_web_json = _parse_raw_web_json_lookup(cleaned)
@@ -619,6 +802,6 @@ def strip_tool_blocks(text: str, skip_fenced: bool = False) -> str:
             _, (start, end) = raw_web_json
             cleaned = cleaned[:start] + cleaned[end:]
     # Strip bare <invoke> blocks not wrapped in <tool_call>
-    cleaned = re.sub(r'<invoke\s+name=["\'].*?</invoke>', '', cleaned, flags=re.DOTALL | re.IGNORECASE)
+    cleaned = _strip_bare_invoke_markup(cleaned)
     cleaned = re.sub(r'\n{3,}', '\n\n', cleaned)
     return cleaned.strip()

src/tool_schemas.py

@@ -68,11 +68,12 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "web_fetch",
-            "description": "Fetch and read the text content of a specific URL the user names (e.g. 'check example.com', 'what's on this page <url>'). Use when you already have a concrete URL/domain. NOT for open-ended searches (use web_search) or 'research X' jobs (use trigger_research).",
+            "description": "Fetch and read the text content of a specific URL the user names (e.g. 'check example.com', 'what's on this page <url>'). Use when you already have a concrete URL/domain. NOT for open-ended searches (use web_search) or 'research X' jobs (use trigger_research). Downloads are size-budgeted; a '[partial content: ...]' notice in the result means the body was cut short and you can re-call with full=true for the rest.",
             "parameters": {
                 "type": "object",
                 "properties": {
-                    "url": {"type": "string", "description": "The URL or domain to fetch (http/https; a bare domain like example.com is fine)"}
+                    "url": {"type": "string", "description": "The URL or domain to fetch (http/https; a bare domain like example.com is fine)"},
+                    "full": {"type": "boolean", "description": "Raise the download budget to the hard cap for large pages/files. Use only after a result reported partial content."}
                 },
                 "required": ["url"]
             }
@@ -1008,7 +1009,7 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "resolve_contact",
-            "description": "Look up a contact's email address by name. Searches CardDAV address book and sent email history. Use when the user says 'message [name]' or 'email [name]' without an email address.",
+            "description": "Look up a contact by name. Searches CardDAV address book and sent email history. Returns email addresses (when available) or phone numbers. Use when the user says 'message [name]', 'email [name]', or asks for someone's contact details.",
             "parameters": {
                 "type": "object",
                 "properties": {
@@ -1187,6 +1188,21 @@ FUNCTION_TOOL_SCHEMAS = [
             }
         }
     },
+    {
+        "type": "function",
+        "function": {
+            "name": "manage_bg_jobs",
+            "description": "Inspect and control detached background `bash` jobs (started with the `#!bg` marker). action='list' shows this chat's jobs with id/status/age/command; action='output' returns a job's captured output so far (use for a still-running job, or to re-read a finished one); action='kill' terminates a runaway job's process tree instead of waiting out its max-runtime. output and kill need job_id from list.",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "action": {"type": "string", "enum": ["list", "output", "kill"], "description": "list | output | kill (default: list)"},
+                    "job_id": {"type": "string", "description": "Background job id (required for output/kill; from action='list')"},
+                },
+                "required": ["action"]
+            }
+        }
+    },
 ]
 
 
@@ -1205,23 +1221,26 @@ def function_call_to_tool_block(name: str, arguments: str) -> Optional[ToolBlock
         logger.error(f"Failed to parse function call arguments for {name}: {arguments}")
         return None
 
+    tool_type = _TOOL_NAME_MAP.get(name, name)
+    _BUILTIN_EMAIL_TOOLS = {"list_email_accounts", "send_email", "list_emails", "read_email", "reply_to_email",
+                            "archive_email", "delete_email", "mark_email_read", "bulk_email", "download_attachment"}
+
     # Some models emit valid JSON that isn't an object (e.g. a bare array
-    # ["ls -la"], string, or number) as the function arguments. Every branch
-    # below assumes a dict and calls args.get(...), so a non-dict would raise
-    # AttributeError and abort the whole agent stream. Coerce to {} instead.
+    # ["ls -la"], string, or number) as function arguments. Most local tools keep
+    # the legacy empty-object coercion for stream robustness, but email MCP tools
+    # must fail closed so a malformed call cannot read the default mailbox.
     if not isinstance(args, dict):
+        if tool_type.startswith("mcp__email__") or name in _BUILTIN_EMAIL_TOOLS:
+            logger.warning(f"Non-object email function call arguments for {name}: {args!r}; rejecting")
+            return None
         logger.warning(f"Non-object function call arguments for {name}: {args!r}; treating as empty")
         args = {}
 
-    tool_type = _TOOL_NAME_MAP.get(name, name)
-
     # Allow MCP tools through (namespaced as mcp__serverid__toolname)
     if tool_type.startswith("mcp__"):
         content = json.dumps(args) if args else "{}"
         return ToolBlock(tool_type, content)
     # Email tools are implemented as MCP — route them to email
-    _BUILTIN_EMAIL_TOOLS = {"list_email_accounts", "send_email", "list_emails", "read_email", "reply_to_email",
-                            "archive_email", "delete_email", "mark_email_read", "bulk_email", "download_attachment"}
     if name in _BUILTIN_EMAIL_TOOLS:
         return ToolBlock(f"mcp__email__{name}", json.dumps(args) if args else "{}")
     if tool_type not in TOOL_TAGS:

src/tool_security.py

@@ -14,6 +14,7 @@ logger = logging.getLogger(__name__)
 NON_ADMIN_BLOCKED_TOOLS = {
     "bash",
     "python",
+    "manage_bg_jobs",
     "read_file",
     "write_file",
     "edit_file",
@@ -114,6 +115,8 @@ _PLAN_MODE_KNOWN_MUTATORS = {
     # Shell is never read-only-safe; block it explicitly so it stays out of plan
     # mode even if the schema list fails to load.
     "bash", "python",
+    # Controls shell processes (kill); plan mode can't run bash anyway.
+    "manage_bg_jobs",
 }