hermes-agent/gateway/run.py

"""
Gateway runner - entry point for messaging platform integrations.

This module provides:
- start_gateway(): Start all configured platform adapters
- GatewayRunner: Main class managing the gateway lifecycle

Usage:
    # Start the gateway
    python -m gateway.run
    
    # Or from CLI
    python cli.py --gateway
"""

import asyncio
import json
import logging
import os
import re
import shlex
import sys
import signal
import tempfile
import threading
import time
from pathlib import Path
from datetime import datetime
from typing import Dict, Optional, Any, List

# ---------------------------------------------------------------------------
# SSL certificate auto-detection for NixOS and other non-standard systems.
# Must run BEFORE any HTTP library (discord, aiohttp, etc.) is imported.
# ---------------------------------------------------------------------------
def _ensure_ssl_certs() -> None:
    """Set SSL_CERT_FILE if the system doesn't expose CA certs to Python."""
    if "SSL_CERT_FILE" in os.environ:
        return  # user already configured it

    import ssl

    # 1. Python's compiled-in defaults
    paths = ssl.get_default_verify_paths()
    for candidate in (paths.cafile, paths.openssl_cafile):
        if candidate and os.path.exists(candidate):
            os.environ["SSL_CERT_FILE"] = candidate
            return

    # 2. certifi (ships its own Mozilla bundle)
    try:
        import certifi
        os.environ["SSL_CERT_FILE"] = certifi.where()
        return
    except ImportError:
        pass

    # 3. Common distro / macOS locations
    for candidate in (
        "/etc/ssl/certs/ca-certificates.crt",               # Debian/Ubuntu/Gentoo
        "/etc/pki/tls/certs/ca-bundle.crt",                 # RHEL/CentOS 7
        "/etc/pki/ca-trust/extracted/pem/tls-ca-bundle.pem", # RHEL/CentOS 8+
        "/etc/ssl/ca-bundle.pem",                            # SUSE/OpenSUSE
        "/etc/ssl/cert.pem",                                 # Alpine / macOS
        "/etc/pki/tls/cert.pem",                             # Fedora
        "/usr/local/etc/openssl@1.1/cert.pem",               # macOS Homebrew Intel
        "/opt/homebrew/etc/openssl@1.1/cert.pem",            # macOS Homebrew ARM
    ):
        if os.path.exists(candidate):
            os.environ["SSL_CERT_FILE"] = candidate
            return

_ensure_ssl_certs()

# Add parent directory to path
sys.path.insert(0, str(Path(__file__).parent.parent))

# Resolve Hermes home directory (respects HERMES_HOME override)
from hermes_constants import get_hermes_home
from utils import atomic_yaml_write, is_truthy_value
_hermes_home = get_hermes_home()

# Load environment variables from ~/.hermes/.env first.
# User-managed env files should override stale shell exports on restart.
from dotenv import load_dotenv  # backward-compat for tests that monkeypatch this symbol
from hermes_cli.env_loader import load_hermes_dotenv
_env_path = _hermes_home / '.env'
load_hermes_dotenv(hermes_home=_hermes_home, project_env=Path(__file__).resolve().parents[1] / '.env')

# Bridge config.yaml values into the environment so os.getenv() picks them up.
# config.yaml is authoritative for terminal settings — overrides .env.
_config_path = _hermes_home / 'config.yaml'
if _config_path.exists():
    try:
        import yaml as _yaml
        with open(_config_path, encoding="utf-8") as _f:
            _cfg = _yaml.safe_load(_f) or {}
        # Expand ${ENV_VAR} references before bridging to env vars.
        from hermes_cli.config import _expand_env_vars
        _cfg = _expand_env_vars(_cfg)
        # Top-level simple values (fallback only — don't override .env)
        for _key, _val in _cfg.items():
            if isinstance(_val, (str, int, float, bool)) and _key not in os.environ:
                os.environ[_key] = str(_val)
        # Terminal config is nested — bridge to TERMINAL_* env vars.
        # config.yaml overrides .env for these since it's the documented config path.
        _terminal_cfg = _cfg.get("terminal", {})
        if _terminal_cfg and isinstance(_terminal_cfg, dict):
            _terminal_env_map = {
                "backend": "TERMINAL_ENV",
                "cwd": "TERMINAL_CWD",
                "timeout": "TERMINAL_TIMEOUT",
                "lifetime_seconds": "TERMINAL_LIFETIME_SECONDS",
                "docker_image": "TERMINAL_DOCKER_IMAGE",
                "docker_forward_env": "TERMINAL_DOCKER_FORWARD_ENV",
                "singularity_image": "TERMINAL_SINGULARITY_IMAGE",
                "modal_image": "TERMINAL_MODAL_IMAGE",
                "daytona_image": "TERMINAL_DAYTONA_IMAGE",
                "ssh_host": "TERMINAL_SSH_HOST",
                "ssh_user": "TERMINAL_SSH_USER",
                "ssh_port": "TERMINAL_SSH_PORT",
                "ssh_key": "TERMINAL_SSH_KEY",
                "container_cpu": "TERMINAL_CONTAINER_CPU",
                "container_memory": "TERMINAL_CONTAINER_MEMORY",
                "container_disk": "TERMINAL_CONTAINER_DISK",
                "container_persistent": "TERMINAL_CONTAINER_PERSISTENT",
                "docker_volumes": "TERMINAL_DOCKER_VOLUMES",
                "sandbox_dir": "TERMINAL_SANDBOX_DIR",
                "persistent_shell": "TERMINAL_PERSISTENT_SHELL",
            }
            for _cfg_key, _env_var in _terminal_env_map.items():
                if _cfg_key in _terminal_cfg:
                    _val = _terminal_cfg[_cfg_key]
                    if isinstance(_val, list):
                        os.environ[_env_var] = json.dumps(_val)
                    else:
                        os.environ[_env_var] = str(_val)
        # Compression config is read directly from config.yaml by run_agent.py
        # and auxiliary_client.py — no env var bridging needed.
        # Auxiliary model/direct-endpoint overrides (vision, web_extract).
        # Each task has provider/model/base_url/api_key; bridge non-default values to env vars.
        _auxiliary_cfg = _cfg.get("auxiliary", {})
        if _auxiliary_cfg and isinstance(_auxiliary_cfg, dict):
            _aux_task_env = {
                "vision": {
                    "provider": "AUXILIARY_VISION_PROVIDER",
                    "model": "AUXILIARY_VISION_MODEL",
                    "base_url": "AUXILIARY_VISION_BASE_URL",
                    "api_key": "AUXILIARY_VISION_API_KEY",
                },
                "web_extract": {
                    "provider": "AUXILIARY_WEB_EXTRACT_PROVIDER",
                    "model": "AUXILIARY_WEB_EXTRACT_MODEL",
                    "base_url": "AUXILIARY_WEB_EXTRACT_BASE_URL",
                    "api_key": "AUXILIARY_WEB_EXTRACT_API_KEY",
                },
                "approval": {
                    "provider": "AUXILIARY_APPROVAL_PROVIDER",
                    "model": "AUXILIARY_APPROVAL_MODEL",
                    "base_url": "AUXILIARY_APPROVAL_BASE_URL",
                    "api_key": "AUXILIARY_APPROVAL_API_KEY",
                },
            }
            for _task_key, _env_map in _aux_task_env.items():
                _task_cfg = _auxiliary_cfg.get(_task_key, {})
                if not isinstance(_task_cfg, dict):
                    continue
                _prov = str(_task_cfg.get("provider", "")).strip()
                _model = str(_task_cfg.get("model", "")).strip()
                _base_url = str(_task_cfg.get("base_url", "")).strip()
                _api_key = str(_task_cfg.get("api_key", "")).strip()
                if _prov and _prov != "auto":
                    os.environ[_env_map["provider"]] = _prov
                if _model:
                    os.environ[_env_map["model"]] = _model
                if _base_url:
                    os.environ[_env_map["base_url"]] = _base_url
                if _api_key:
                    os.environ[_env_map["api_key"]] = _api_key
        _agent_cfg = _cfg.get("agent", {})
        if _agent_cfg and isinstance(_agent_cfg, dict):
            if "max_turns" in _agent_cfg:
                os.environ["HERMES_MAX_ITERATIONS"] = str(_agent_cfg["max_turns"])
            # Bridge agent.gateway_timeout → HERMES_AGENT_TIMEOUT env var.
            # Env var from .env takes precedence (already in os.environ).
            if "gateway_timeout" in _agent_cfg and "HERMES_AGENT_TIMEOUT" not in os.environ:
                os.environ["HERMES_AGENT_TIMEOUT"] = str(_agent_cfg["gateway_timeout"])
            if "gateway_timeout_warning" in _agent_cfg and "HERMES_AGENT_TIMEOUT_WARNING" not in os.environ:
                os.environ["HERMES_AGENT_TIMEOUT_WARNING"] = str(_agent_cfg["gateway_timeout_warning"])
            if "restart_drain_timeout" in _agent_cfg and "HERMES_RESTART_DRAIN_TIMEOUT" not in os.environ:
                os.environ["HERMES_RESTART_DRAIN_TIMEOUT"] = str(_agent_cfg["restart_drain_timeout"])
        _display_cfg = _cfg.get("display", {})
        if _display_cfg and isinstance(_display_cfg, dict):
            if "busy_input_mode" in _display_cfg and "HERMES_GATEWAY_BUSY_INPUT_MODE" not in os.environ:
                os.environ["HERMES_GATEWAY_BUSY_INPUT_MODE"] = str(_display_cfg["busy_input_mode"])
        # Timezone: bridge config.yaml → HERMES_TIMEZONE env var.
        # HERMES_TIMEZONE from .env takes precedence (already in os.environ).
        _tz_cfg = _cfg.get("timezone", "")
        if _tz_cfg and isinstance(_tz_cfg, str) and "HERMES_TIMEZONE" not in os.environ:
            os.environ["HERMES_TIMEZONE"] = _tz_cfg.strip()
        # Security settings
        _security_cfg = _cfg.get("security", {})
        if isinstance(_security_cfg, dict):
            _redact = _security_cfg.get("redact_secrets")
            if _redact is not None:
                os.environ["HERMES_REDACT_SECRETS"] = str(_redact).lower()
    except Exception:
        pass  # Non-fatal; gateway can still run with .env values

# Apply IPv4 preference if configured (before any HTTP clients are created).
try:
    from hermes_constants import apply_ipv4_preference
    _network_cfg = (_cfg if '_cfg' in dir() else {}).get("network", {})
    if isinstance(_network_cfg, dict) and _network_cfg.get("force_ipv4"):
        apply_ipv4_preference(force=True)
except Exception:
    pass

# Validate config structure early — log warnings so gateway operators see problems
try:
    from hermes_cli.config import print_config_warnings
    print_config_warnings()
except Exception:
    pass

# Gateway runs in quiet mode - suppress debug output and use cwd directly (no temp dirs)
os.environ["HERMES_QUIET"] = "1"

# Enable interactive exec approval for dangerous commands on messaging platforms
os.environ["HERMES_EXEC_ASK"] = "1"

# Set terminal working directory for messaging platforms.
# If the user set an explicit path in config.yaml (not "." or "auto"),
# respect it. Otherwise use MESSAGING_CWD or default to home directory.
_configured_cwd = os.environ.get("TERMINAL_CWD", "")
if not _configured_cwd or _configured_cwd in (".", "auto", "cwd"):
    messaging_cwd = os.getenv("MESSAGING_CWD") or str(Path.home())
    os.environ["TERMINAL_CWD"] = messaging_cwd

from gateway.config import (
    Platform,
    GatewayConfig,
    load_gateway_config,
)
from gateway.session import (
    SessionStore,
    SessionSource,
    SessionContext,
    build_session_context,
    build_session_context_prompt,
    build_session_key,
)
from gateway.delivery import DeliveryRouter
from gateway.platforms.base import (
    BasePlatformAdapter,
    MessageEvent,
    MessageType,
    merge_pending_message_event,
)
from gateway.restart import (
    DEFAULT_GATEWAY_RESTART_DRAIN_TIMEOUT,
    GATEWAY_SERVICE_RESTART_EXIT_CODE,
    parse_restart_drain_timeout,
)


def _normalize_whatsapp_identifier(value: str) -> str:
    """Strip WhatsApp JID/LID syntax down to its stable numeric identifier."""
    return (
        str(value or "")
        .strip()
        .replace("+", "", 1)
        .split(":", 1)[0]
        .split("@", 1)[0]
    )


def _expand_whatsapp_auth_aliases(identifier: str) -> set:
    """Resolve WhatsApp phone/LID aliases using bridge session mapping files."""
    normalized = _normalize_whatsapp_identifier(identifier)
    if not normalized:
        return set()

    session_dir = _hermes_home / "whatsapp" / "session"
    resolved = set()
    queue = [normalized]

    while queue:
        current = queue.pop(0)
        if not current or current in resolved:
            continue

        resolved.add(current)
        for suffix in ("", "_reverse"):
            mapping_path = session_dir / f"lid-mapping-{current}{suffix}.json"
            if not mapping_path.exists():
                continue
            try:
                mapped = _normalize_whatsapp_identifier(
                    json.loads(mapping_path.read_text(encoding="utf-8"))
                )
            except Exception:
                continue
            if mapped and mapped not in resolved:
                queue.append(mapped)

    return resolved

logger = logging.getLogger(__name__)

# Sentinel placed into _running_agents immediately when a session starts
# processing, *before* any await.  Prevents a second message for the same
# session from bypassing the "already running" guard during the async gap
# between the guard check and actual agent creation.
_AGENT_PENDING_SENTINEL = object()


def _resolve_runtime_agent_kwargs() -> dict:
    """Resolve provider credentials for gateway-created AIAgent instances."""
    from hermes_cli.runtime_provider import (
        resolve_runtime_provider,
        format_runtime_provider_error,
    )

    try:
        runtime = resolve_runtime_provider(
            requested=os.getenv("HERMES_INFERENCE_PROVIDER"),
        )
    except Exception as exc:
        raise RuntimeError(format_runtime_provider_error(exc)) from exc

    return {
        "api_key": runtime.get("api_key"),
        "base_url": runtime.get("base_url"),
        "provider": runtime.get("provider"),
        "api_mode": runtime.get("api_mode"),
        "command": runtime.get("command"),
        "args": list(runtime.get("args") or []),
        "credential_pool": runtime.get("credential_pool"),
    }


def _build_media_placeholder(event) -> str:
    """Build a text placeholder for media-only events so they aren't dropped.

    When a photo/document is queued during active processing and later
    dequeued, only .text is extracted.  If the event has no caption,
    the media would be silently lost.  This builds a placeholder that
    the vision enrichment pipeline will replace with a real description.
    """
    parts = []
    media_urls = getattr(event, "media_urls", None) or []
    media_types = getattr(event, "media_types", None) or []
    for i, url in enumerate(media_urls):
        mtype = media_types[i] if i < len(media_types) else ""
        if mtype.startswith("image/") or getattr(event, "message_type", None) == MessageType.PHOTO:
            parts.append(f"[User sent an image: {url}]")
        elif mtype.startswith("audio/"):
            parts.append(f"[User sent audio: {url}]")
        else:
            parts.append(f"[User sent a file: {url}]")
    return "\n".join(parts)


def _dequeue_pending_event(adapter, session_key: str) -> MessageEvent | None:
    """Consume and return the full pending event for a session.

    Queued follow-ups must preserve their media metadata so they can re-enter
    the normal image/STT/document preprocessing path instead of being reduced
    to a placeholder string.
    """
    return adapter.get_pending_message(session_key)


def _check_unavailable_skill(command_name: str) -> str | None:
    """Check if a command matches a known-but-inactive skill.

    Returns a helpful message if the skill exists but is disabled or only
    available as an optional install. Returns None if no match found.
    """
    # Normalize: command uses hyphens, skill names may use hyphens or underscores
    normalized = command_name.lower().replace("_", "-")
    try:
        from tools.skills_tool import _get_disabled_skill_names
        from agent.skill_utils import get_all_skills_dirs
        disabled = _get_disabled_skill_names()

        # Check disabled skills across all dirs (local + external)
        for skills_dir in get_all_skills_dirs():
            if not skills_dir.exists():
                continue
            for skill_md in skills_dir.rglob("SKILL.md"):
                if any(part in ('.git', '.github', '.hub') for part in skill_md.parts):
                    continue
                name = skill_md.parent.name.lower().replace("_", "-")
                if name == normalized and name in disabled:
                    return (
                        f"The **{command_name}** skill is installed but disabled.\n"
                        f"Enable it with: `hermes skills config`"
                    )

        # Check optional skills (shipped with repo but not installed)
        from hermes_constants import get_optional_skills_dir
        repo_root = Path(__file__).resolve().parent.parent
        optional_dir = get_optional_skills_dir(repo_root / "optional-skills")
        if optional_dir.exists():
            for skill_md in optional_dir.rglob("SKILL.md"):
                name = skill_md.parent.name.lower().replace("_", "-")
                if name == normalized:
                    # Build install path: official/<category>/<name>
                    rel = skill_md.parent.relative_to(optional_dir)
                    parts = list(rel.parts)
                    install_path = f"official/{'/'.join(parts)}"
                    return (
                        f"The **{command_name}** skill is available but not installed.\n"
                        f"Install it with: `hermes skills install {install_path}`"
                    )
    except Exception:
        pass
    return None


def _platform_config_key(platform: "Platform") -> str:
    """Map a Platform enum to its config.yaml key (LOCAL→"cli", rest→enum value)."""
    return "cli" if platform == Platform.LOCAL else platform.value


def _load_gateway_config() -> dict:
    """Load and parse ~/.hermes/config.yaml, returning {} on any error."""
    try:
        config_path = _hermes_home / 'config.yaml'
        if config_path.exists():
            import yaml
            with open(config_path, 'r', encoding='utf-8') as f:
                return yaml.safe_load(f) or {}
    except Exception:
        logger.debug("Could not load gateway config from %s", _hermes_home / 'config.yaml')
    return {}


def _resolve_gateway_model(config: dict | None = None) -> str:
    """Read model from config.yaml — single source of truth.

    Without this, temporary AIAgent instances (memory flush, /compress) fall
    back to the hardcoded default which fails when the active provider is
    openai-codex.
    """
    cfg = config if config is not None else _load_gateway_config()
    model_cfg = cfg.get("model", {})
    if isinstance(model_cfg, str):
        return model_cfg
    elif isinstance(model_cfg, dict):
        return model_cfg.get("default") or model_cfg.get("model") or ""
    return ""


def _resolve_hermes_bin() -> Optional[list[str]]:
    """Resolve the Hermes update command as argv parts.

    Tries in order:
    1. ``shutil.which("hermes")`` — standard PATH lookup
    2. ``sys.executable -m hermes_cli.main`` — fallback when Hermes is running
       from a venv/module invocation and the ``hermes`` shim is not on PATH

    Returns argv parts ready for quoting/joining, or ``None`` if neither works.
    """
    import shutil

    hermes_bin = shutil.which("hermes")
    if hermes_bin:
        return [hermes_bin]

    try:
        import importlib.util

        if importlib.util.find_spec("hermes_cli") is not None:
            return [sys.executable, "-m", "hermes_cli.main"]
    except Exception:
        pass

    return None


def _format_gateway_process_notification(evt: dict) -> "str | None":
    """Format a watch pattern event from completion_queue into a [SYSTEM:] message."""
    evt_type = evt.get("type", "completion")
    _sid = evt.get("session_id", "unknown")
    _cmd = evt.get("command", "unknown")

    if evt_type == "watch_disabled":
        return f"[SYSTEM: {evt.get('message', '')}]"

    if evt_type == "watch_match":
        _pat = evt.get("pattern", "?")
        _out = evt.get("output", "")
        _sup = evt.get("suppressed", 0)
        text = (
            f"[SYSTEM: Background process {_sid} matched "
            f"watch pattern \"{_pat}\".\n"
            f"Command: {_cmd}\n"
            f"Matched output:\n{_out}"
        )
        if _sup:
            text += f"\n({_sup} earlier matches were suppressed by rate limit)"
        text += "]"
        return text

    return None


class GatewayRunner:
    """
    Main gateway controller.

    Manages the lifecycle of all platform adapters and routes
    messages to/from the agent.
    """

    # Class-level defaults so partial construction in tests doesn't
    # blow up on attribute access.
    _running_agents_ts: Dict[str, float] = {}
    _busy_input_mode: str = "interrupt"
    _restart_drain_timeout: float = DEFAULT_GATEWAY_RESTART_DRAIN_TIMEOUT
    _exit_code: Optional[int] = None
    _draining: bool = False
    _restart_requested: bool = False
    _restart_task_started: bool = False
    _restart_detached: bool = False
    _restart_via_service: bool = False
    _stop_task: Optional[asyncio.Task] = None
    _session_model_overrides: Dict[str, Dict[str, str]] = {}
    
    def __init__(self, config: Optional[GatewayConfig] = None):
        self.config = config or load_gateway_config()
        self.adapters: Dict[Platform, BasePlatformAdapter] = {}

        # Load ephemeral config from config.yaml / env vars.
        # Both are injected at API-call time only and never persisted.
        self._prefill_messages = self._load_prefill_messages()
        self._ephemeral_system_prompt = self._load_ephemeral_system_prompt()
        self._reasoning_config = self._load_reasoning_config()
        self._service_tier = self._load_service_tier()
        self._show_reasoning = self._load_show_reasoning()
        self._busy_input_mode = self._load_busy_input_mode()
        self._restart_drain_timeout = self._load_restart_drain_timeout()
        self._provider_routing = self._load_provider_routing()
        self._fallback_model = self._load_fallback_model()
        self._smart_model_routing = self._load_smart_model_routing()

        # Wire process registry into session store for reset protection
        from tools.process_registry import process_registry
        self.session_store = SessionStore(
            self.config.sessions_dir, self.config,
            has_active_processes_fn=lambda key: process_registry.has_active_for_session(key),
        )
        self.delivery_router = DeliveryRouter(self.config)
        self._running = False
        self._shutdown_event = asyncio.Event()
        self._exit_cleanly = False
        self._exit_with_failure = False
        self._exit_reason: Optional[str] = None
        self._exit_code: Optional[int] = None
        self._draining = False
        self._restart_requested = False
        self._restart_task_started = False
        self._restart_detached = False
        self._restart_via_service = False
        self._stop_task: Optional[asyncio.Task] = None
        
        # Track running agents per session for interrupt support
        # Key: session_key, Value: AIAgent instance
        self._running_agents: Dict[str, Any] = {}
        self._running_agents_ts: Dict[str, float] = {}  # start timestamp per session
        self._pending_messages: Dict[str, str] = {}  # Queued messages during interrupt

        # Cache AIAgent instances per session to preserve prompt caching.
        # Without this, a new AIAgent is created per message, rebuilding the
        # system prompt (including memory) every turn — breaking prefix cache
        # and costing ~10x more on providers with prompt caching (Anthropic).
        # Key: session_key, Value: (AIAgent, config_signature_str)
        import threading as _threading
        self._agent_cache: Dict[str, tuple] = {}
        self._agent_cache_lock = _threading.Lock()

        # Per-session model overrides from /model command.
        # Key: session_key, Value: dict with model/provider/api_key/base_url/api_mode
        self._session_model_overrides: Dict[str, Dict[str, str]] = {}
        # Track pending exec approvals per session
        # Key: session_key, Value: {"command": str, "pattern_key": str, ...}
        self._pending_approvals: Dict[str, Dict[str, Any]] = {}

        # Track platforms that failed to connect for background reconnection.
        # Key: Platform enum, Value: {"config": platform_config, "attempts": int, "next_retry": float}
        self._failed_platforms: Dict[Platform, Dict[str, Any]] = {}

        # Track pending /update prompt responses per session.
        # Key: session_key, Value: True when a prompt is waiting for user input.
        self._update_prompt_pending: Dict[str, bool] = {}

        # Persistent Honcho managers keyed by gateway session key.
        # This preserves write_frequency="session" semantics across short-lived
        # per-message AIAgent instances.



        # Ensure tirith security scanner is available (downloads if needed)
        try:
            from tools.tirith_security import ensure_installed
            ensure_installed(log_failures=False)
        except Exception:
            pass  # Non-fatal — fail-open at scan time if unavailable
        
        # Initialize session database for session_search tool support
        self._session_db = None
        try:
            from hermes_state import SessionDB
            self._session_db = SessionDB()
        except Exception as e:
            logger.debug("SQLite session store not available: %s", e)
        
        # DM pairing store for code-based user authorization
        from gateway.pairing import PairingStore
        self.pairing_store = PairingStore()
        
        # Event hook system
        from gateway.hooks import HookRegistry
        self.hooks = HookRegistry()

        # Per-chat voice reply mode: "off" | "voice_only" | "all"
        self._voice_mode: Dict[str, str] = self._load_voice_modes()

        # Track background tasks to prevent garbage collection mid-execution
        self._background_tasks: set = set()




    # -- Setup skill availability ----------------------------------------

    def _has_setup_skill(self) -> bool:
        """Check if the hermes-agent-setup skill is installed."""
        try:
            from tools.skill_manager_tool import _find_skill
            return _find_skill("hermes-agent-setup") is not None
        except Exception:
            return False

    # -- Voice mode persistence ------------------------------------------

    _VOICE_MODE_PATH = _hermes_home / "gateway_voice_mode.json"

    def _load_voice_modes(self) -> Dict[str, str]:
        try:
            data = json.loads(self._VOICE_MODE_PATH.read_text())
        except (FileNotFoundError, json.JSONDecodeError, OSError):
            return {}

        if not isinstance(data, dict):
            return {}

        valid_modes = {"off", "voice_only", "all"}
        return {
            str(chat_id): mode
            for chat_id, mode in data.items()
            if mode in valid_modes
        }

    def _save_voice_modes(self) -> None:
        try:
            self._VOICE_MODE_PATH.parent.mkdir(parents=True, exist_ok=True)
            self._VOICE_MODE_PATH.write_text(
                json.dumps(self._voice_mode, indent=2)
            )
        except OSError as e:
            logger.warning("Failed to save voice modes: %s", e)

    def _set_adapter_auto_tts_disabled(self, adapter, chat_id: str, disabled: bool) -> None:
        """Update an adapter's in-memory auto-TTS suppression set if present."""
        disabled_chats = getattr(adapter, "_auto_tts_disabled_chats", None)
        if not isinstance(disabled_chats, set):
            return
        if disabled:
            disabled_chats.add(chat_id)
        else:
            disabled_chats.discard(chat_id)

    def _sync_voice_mode_state_to_adapter(self, adapter) -> None:
        """Restore persisted /voice off state into a live platform adapter."""
        disabled_chats = getattr(adapter, "_auto_tts_disabled_chats", None)
        if not isinstance(disabled_chats, set):
            return
        disabled_chats.clear()
        disabled_chats.update(
            chat_id for chat_id, mode in self._voice_mode.items() if mode == "off"
        )

    # -----------------------------------------------------------------

    def _flush_memories_for_session(
        self,
        old_session_id: str,
        session_key: Optional[str] = None,
    ):
        """Prompt the agent to save memories/skills before context is lost.

        Synchronous worker — meant to be called via run_in_executor from
        an async context so it doesn't block the event loop.
        """
        # Skip cron sessions — they run headless with no meaningful user
        # conversation to extract memories from.
        if old_session_id and old_session_id.startswith("cron_"):
            logger.debug("Skipping memory flush for cron session: %s", old_session_id)
            return

        try:
            history = self.session_store.load_transcript(old_session_id)
            if not history or len(history) < 4:
                return

            from run_agent import AIAgent
            model, runtime_kwargs = self._resolve_session_agent_runtime(
                session_key=session_key,
            )
            if not runtime_kwargs.get("api_key"):
                return

            tmp_agent = AIAgent(
                **runtime_kwargs,
                model=model,
                max_iterations=8,
                quiet_mode=True,
                skip_memory=True,  # Flush agent — no memory provider
                enabled_toolsets=["memory", "skills"],
                session_id=old_session_id,
            )
            # Fully silence the flush agent — quiet_mode only suppresses init
            # messages; tool call output still leaks to the terminal through
            # _safe_print → _print_fn.  Set a no-op to prevent that.
            tmp_agent._print_fn = lambda *a, **kw: None

            # Build conversation history from transcript
            msgs = [
                {"role": m.get("role"), "content": m.get("content")}
                for m in history
                if m.get("role") in ("user", "assistant") and m.get("content")
            ]

            # Read live memory state from disk so the flush agent can see
            # what's already saved and avoid overwriting newer entries.
            _current_memory = ""
            try:
                from tools.memory_tool import get_memory_dir
                _mem_dir = get_memory_dir()
                for fname, label in [
                    ("MEMORY.md", "MEMORY (your personal notes)"),
                    ("USER.md", "USER PROFILE (who the user is)"),
                ]:
                    fpath = _mem_dir / fname
                    if fpath.exists():
                        content = fpath.read_text(encoding="utf-8").strip()
                        if content:
                            _current_memory += f"\n\n## Current {label}:\n{content}"
            except Exception:
                pass  # Non-fatal — flush still works, just without the guard

            # Give the agent a real turn to think about what to save
            flush_prompt = (
                "[System: This session is about to be automatically reset due to "
                "inactivity or a scheduled daily reset. The conversation context "
                "will be cleared after this turn.\n\n"
                "Review the conversation above and:\n"
                "1. Save any important facts, preferences, or decisions to memory "
                "(user profile or your notes) that would be useful in future sessions.\n"
                "2. If you discovered a reusable workflow or solved a non-trivial "
                "problem, consider saving it as a skill.\n"
                "3. If nothing is worth saving, that's fine — just skip.\n\n"
            )

            if _current_memory:
                flush_prompt += (
                    "IMPORTANT — here is the current live state of memory. Other "
                    "sessions, cron jobs, or the user may have updated it since this "
                    "conversation ended. Do NOT overwrite or remove entries unless "
                    "the conversation above reveals something that genuinely "
                    "supersedes them. Only add new information that is not already "
                    "captured below."
                    f"{_current_memory}\n\n"
                )

            flush_prompt += (
                "Do NOT respond to the user. Just use the memory and skill_manage "
                "tools if needed, then stop.]"
            )

            tmp_agent.run_conversation(
                user_message=flush_prompt,
                conversation_history=msgs,
            )
            logger.info("Pre-reset memory flush completed for session %s", old_session_id)
        except Exception as e:
            logger.debug("Pre-reset memory flush failed for session %s: %s", old_session_id, e)

    async def _async_flush_memories(
        self,
        old_session_id: str,
        session_key: Optional[str] = None,
    ):
        """Run the sync memory flush in a thread pool so it won't block the event loop."""
        loop = asyncio.get_event_loop()
        await loop.run_in_executor(
            None,
            self._flush_memories_for_session,
            old_session_id,
            session_key,
        )

    @property
    def should_exit_cleanly(self) -> bool:
        return self._exit_cleanly

    @property
    def should_exit_with_failure(self) -> bool:
        return self._exit_with_failure

    @property
    def exit_reason(self) -> Optional[str]:
        return self._exit_reason

    @property
    def exit_code(self) -> Optional[int]:
        return self._exit_code

    def _session_key_for_source(self, source: SessionSource) -> str:
        """Resolve the current session key for a source, honoring gateway config when available."""
        if hasattr(self, "session_store") and self.session_store is not None:
            try:
                session_key = self.session_store._generate_session_key(source)
                if isinstance(session_key, str) and session_key:
                    return session_key
            except Exception:
                pass
        config = getattr(self, "config", None)
        return build_session_key(
            source,
            group_sessions_per_user=getattr(config, "group_sessions_per_user", True),
            thread_sessions_per_user=getattr(config, "thread_sessions_per_user", False),
        )

    def _resolve_session_agent_runtime(
        self,
        *,
        source: Optional[SessionSource] = None,
        session_key: Optional[str] = None,
        user_config: Optional[dict] = None,
    ) -> tuple[str, dict]:
        """Resolve model/runtime for a session, honoring session-scoped /model overrides.

        If the session override already contains a complete provider bundle
        (provider/api_key/base_url/api_mode), prefer it directly instead of
        resolving fresh global runtime state first.
        """
        resolved_session_key = session_key
        if not resolved_session_key and source is not None:
            try:
                resolved_session_key = self._session_key_for_source(source)
            except Exception:
                resolved_session_key = None

        model = _resolve_gateway_model(user_config)
        override = self._session_model_overrides.get(resolved_session_key) if resolved_session_key else None
        if override:
            override_model = override.get("model", model)
            override_runtime = {
                "provider": override.get("provider"),
                "api_key": override.get("api_key"),
                "base_url": override.get("base_url"),
                "api_mode": override.get("api_mode"),
            }
            if override_runtime.get("api_key"):
                return override_model, override_runtime

        runtime_kwargs = _resolve_runtime_agent_kwargs()
        if override and resolved_session_key:
            model, runtime_kwargs = self._apply_session_model_override(
                resolved_session_key, model, runtime_kwargs
            )
        return model, runtime_kwargs

    def _resolve_turn_agent_config(self, user_message: str, model: str, runtime_kwargs: dict) -> dict:
        from agent.smart_model_routing import resolve_turn_route
        from hermes_cli.models import resolve_fast_mode_overrides

        primary = {
            "model": model,
            "api_key": runtime_kwargs.get("api_key"),
            "base_url": runtime_kwargs.get("base_url"),
            "provider": runtime_kwargs.get("provider"),
            "api_mode": runtime_kwargs.get("api_mode"),
            "command": runtime_kwargs.get("command"),
            "args": list(runtime_kwargs.get("args") or []),
            "credential_pool": runtime_kwargs.get("credential_pool"),
        }
        route = resolve_turn_route(user_message, getattr(self, "_smart_model_routing", {}), primary)

        service_tier = getattr(self, "_service_tier", None)
        if not service_tier:
            route["request_overrides"] = None
            return route

        try:
            overrides = resolve_fast_mode_overrides(route.get("model"))
        except Exception:
            overrides = None
        route["request_overrides"] = overrides
        return route

    async def _handle_adapter_fatal_error(self, adapter: BasePlatformAdapter) -> None:
        """React to an adapter failure after startup.

        If the error is retryable (e.g. network blip, DNS failure), queue the
        platform for background reconnection instead of giving up permanently.
        """
        logger.error(
            "Fatal %s adapter error (%s): %s",
            adapter.platform.value,
            adapter.fatal_error_code or "unknown",
            adapter.fatal_error_message or "unknown error",
        )
        self._update_platform_runtime_status(
            adapter.platform.value,
            platform_state="retrying" if adapter.fatal_error_retryable else "fatal",
            error_code=adapter.fatal_error_code,
            error_message=adapter.fatal_error_message,
        )

        existing = self.adapters.get(adapter.platform)
        if existing is adapter:
            try:
                await adapter.disconnect()
            finally:
                self.adapters.pop(adapter.platform, None)
                self.delivery_router.adapters = self.adapters

        # Queue retryable failures for background reconnection
        if adapter.fatal_error_retryable:
            platform_config = self.config.platforms.get(adapter.platform)
            if platform_config and adapter.platform not in self._failed_platforms:
                self._failed_platforms[adapter.platform] = {
                    "config": platform_config,
                    "attempts": 0,
                    "next_retry": time.monotonic() + 30,
                }
                logger.info(
                    "%s queued for background reconnection",
                    adapter.platform.value,
                )

        if not self.adapters and not self._failed_platforms:
            self._exit_reason = adapter.fatal_error_message or "All messaging adapters disconnected"
            if adapter.fatal_error_retryable:
                self._exit_with_failure = True
                logger.error("No connected messaging platforms remain. Shutting down gateway for service restart.")
            else:
                logger.error("No connected messaging platforms remain. Shutting down gateway cleanly.")
            await self.stop()
        elif not self.adapters and self._failed_platforms:
            # All platforms are down and queued for background reconnection.
            # If the error is retryable, exit with failure so systemd Restart=on-failure
            # can restart the process. Otherwise stay alive and keep retrying in background.
            if adapter.fatal_error_retryable:
                self._exit_reason = adapter.fatal_error_message or "All messaging platforms failed with retryable errors"
                self._exit_with_failure = True
                logger.error(
                    "All messaging platforms failed with retryable errors. "
                    "Shutting down gateway for service restart (systemd will retry)."
                )
                await self.stop()
            else:
                logger.warning(
                    "No connected messaging platforms remain, but %d platform(s) queued for reconnection",
                    len(self._failed_platforms),
                )

    def _request_clean_exit(self, reason: str) -> None:
        self._exit_cleanly = True
        self._exit_reason = reason
        self._shutdown_event.set()

    def _running_agent_count(self) -> int:
        return len(self._running_agents)

    def _status_action_label(self) -> str:
        return "restart" if self._restart_requested else "shutdown"

    def _status_action_gerund(self) -> str:
        return "restarting" if self._restart_requested else "shutting down"

    def _queue_during_drain_enabled(self) -> bool:
        return self._restart_requested and self._busy_input_mode == "queue"

    def _update_runtime_status(self, gateway_state: Optional[str] = None, exit_reason: Optional[str] = None) -> None:
        try:
            from gateway.status import write_runtime_status
            write_runtime_status(
                gateway_state=gateway_state,
                exit_reason=exit_reason,
                restart_requested=self._restart_requested,
                active_agents=self._running_agent_count(),
            )
        except Exception:
            pass

    def _update_platform_runtime_status(
        self,
        platform: str,
        *,
        platform_state: Optional[str] = None,
        error_code: Optional[str] = None,
        error_message: Optional[str] = None,
    ) -> None:
        try:
            from gateway.status import write_runtime_status
            write_runtime_status(
                platform=platform,
                platform_state=platform_state,
                error_code=error_code,
                error_message=error_message,
            )
        except Exception:
            pass
    
    @staticmethod
    def _load_prefill_messages() -> List[Dict[str, Any]]:
        """Load ephemeral prefill messages from config or env var.
        
        Checks HERMES_PREFILL_MESSAGES_FILE env var first, then falls back to
        the prefill_messages_file key in ~/.hermes/config.yaml.
        Relative paths are resolved from ~/.hermes/.
        """
        import json as _json
        file_path = os.getenv("HERMES_PREFILL_MESSAGES_FILE", "")
        if not file_path:
            try:
                import yaml as _y
                cfg_path = _hermes_home / "config.yaml"
                if cfg_path.exists():
                    with open(cfg_path, encoding="utf-8") as _f:
                        cfg = _y.safe_load(_f) or {}
                    file_path = cfg.get("prefill_messages_file", "")
            except Exception:
                pass
        if not file_path:
            return []
        path = Path(file_path).expanduser()
        if not path.is_absolute():
            path = _hermes_home / path
        if not path.exists():
            logger.warning("Prefill messages file not found: %s", path)
            return []
        try:
            with open(path, "r", encoding="utf-8") as f:
                data = _json.load(f)
            if not isinstance(data, list):
                logger.warning("Prefill messages file must contain a JSON array: %s", path)
                return []
            return data
        except Exception as e:
            logger.warning("Failed to load prefill messages from %s: %s", path, e)
            return []

    @staticmethod
    def _load_ephemeral_system_prompt() -> str:
        """Load ephemeral system prompt from config or env var.
        
        Checks HERMES_EPHEMERAL_SYSTEM_PROMPT env var first, then falls back to
        agent.system_prompt in ~/.hermes/config.yaml.
        """
        prompt = os.getenv("HERMES_EPHEMERAL_SYSTEM_PROMPT", "")
        if prompt:
            return prompt
        try:
            import yaml as _y
            cfg_path = _hermes_home / "config.yaml"
            if cfg_path.exists():
                with open(cfg_path, encoding="utf-8") as _f:
                    cfg = _y.safe_load(_f) or {}
                return (cfg.get("agent", {}).get("system_prompt", "") or "").strip()
        except Exception:
            pass
        return ""

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

        Reads agent.reasoning_effort from config.yaml. Valid: "none",
        "minimal", "low", "medium", "high", "xhigh". Returns None to use
        default (medium).
        """
        from hermes_constants import parse_reasoning_effort
        effort = ""
        try:
            import yaml as _y
            cfg_path = _hermes_home / "config.yaml"
            if cfg_path.exists():
                with open(cfg_path, encoding="utf-8") as _f:
                    cfg = _y.safe_load(_f) or {}
                effort = str(cfg.get("agent", {}).get("reasoning_effort", "") or "").strip()
        except Exception:
            pass
        result = parse_reasoning_effort(effort)
        if effort and effort.strip() and result is None:
            logger.warning("Unknown reasoning_effort '%s', using default (medium)", effort)
        return result

    @staticmethod
    def _load_service_tier() -> str | None:
        """Load Priority Processing setting from config.yaml.

        Reads agent.service_tier from config.yaml. Accepted values mirror the CLI:
        "fast"/"priority"/"on" => "priority", while "normal"/"off" disables it.
        Returns None when unset or unsupported.
        """
        raw = ""
        try:
            import yaml as _y
            cfg_path = _hermes_home / "config.yaml"
            if cfg_path.exists():
                with open(cfg_path, encoding="utf-8") as _f:
                    cfg = _y.safe_load(_f) or {}
                raw = str(cfg.get("agent", {}).get("service_tier", "") or "").strip()
        except Exception:
            pass

        value = raw.lower()
        if not value or value in {"normal", "default", "standard", "off", "none"}:
            return None
        if value in {"fast", "priority", "on"}:
            return "priority"
        logger.warning("Unknown service_tier '%s', ignoring", raw)
        return None

    @staticmethod
    def _load_show_reasoning() -> bool:
        """Load show_reasoning toggle from config.yaml display section."""
        try:
            import yaml as _y
            cfg_path = _hermes_home / "config.yaml"
            if cfg_path.exists():
                with open(cfg_path, encoding="utf-8") as _f:
                    cfg = _y.safe_load(_f) or {}
                return bool(cfg.get("display", {}).get("show_reasoning", False))
        except Exception:
            pass
        return False

    @staticmethod
    def _load_busy_input_mode() -> str:
        """Load gateway drain-time busy-input behavior from config/env."""
        mode = os.getenv("HERMES_GATEWAY_BUSY_INPUT_MODE", "").strip().lower()
        if not mode:
            try:
                import yaml as _y
                cfg_path = _hermes_home / "config.yaml"
                if cfg_path.exists():
                    with open(cfg_path, encoding="utf-8") as _f:
                        cfg = _y.safe_load(_f) or {}
                    mode = str(cfg.get("display", {}).get("busy_input_mode", "") or "").strip().lower()
            except Exception:
                pass
        return "queue" if mode == "queue" else "interrupt"

    @staticmethod
    def _load_restart_drain_timeout() -> float:
        """Load graceful gateway restart/stop drain timeout in seconds."""
        raw = os.getenv("HERMES_RESTART_DRAIN_TIMEOUT", "").strip()
        if not raw:
            try:
                import yaml as _y
                cfg_path = _hermes_home / "config.yaml"
                if cfg_path.exists():
                    with open(cfg_path, encoding="utf-8") as _f:
                        cfg = _y.safe_load(_f) or {}
                    raw = str(cfg.get("agent", {}).get("restart_drain_timeout", "") or "").strip()
            except Exception:
                pass
        value = parse_restart_drain_timeout(raw)
        if raw and value == DEFAULT_GATEWAY_RESTART_DRAIN_TIMEOUT:
            try:
                float(raw)
            except (TypeError, ValueError):
                logger.warning(
                    "Invalid restart_drain_timeout '%s', using default %.0fs",
                    raw,
                    DEFAULT_GATEWAY_RESTART_DRAIN_TIMEOUT,
                )
        return value

    @staticmethod
    def _load_background_notifications_mode() -> str:
        """Load background process notification mode from config or env var.

        Modes:
          - ``all``    — push running-output updates *and* the final message (default)
          - ``result`` — only the final completion message (regardless of exit code)
          - ``error``  — only the final message when exit code is non-zero
          - ``off``    — no watcher messages at all
        """
        mode = os.getenv("HERMES_BACKGROUND_NOTIFICATIONS", "")
        if not mode:
            try:
                import yaml as _y
                cfg_path = _hermes_home / "config.yaml"
                if cfg_path.exists():
                    with open(cfg_path, encoding="utf-8") as _f:
                        cfg = _y.safe_load(_f) or {}
                    raw = cfg.get("display", {}).get("background_process_notifications")
                    if raw is False:
                        mode = "off"
                    elif raw not in (None, ""):
                        mode = str(raw)
            except Exception:
                pass
        mode = (mode or "all").strip().lower()
        valid = {"all", "result", "error", "off"}
        if mode not in valid:
            logger.warning(
                "Unknown background_process_notifications '%s', defaulting to 'all'",
                mode,
            )
            return "all"
        return mode

    @staticmethod
    def _load_provider_routing() -> dict:
        """Load OpenRouter provider routing preferences from config.yaml."""
        try:
            import yaml as _y
            cfg_path = _hermes_home / "config.yaml"
            if cfg_path.exists():
                with open(cfg_path, encoding="utf-8") as _f:
                    cfg = _y.safe_load(_f) or {}
                return cfg.get("provider_routing", {}) or {}
        except Exception:
            pass
        return {}

    @staticmethod
    def _load_fallback_model() -> list | dict | None:
        """Load fallback provider chain from config.yaml.

        Returns a list of provider dicts (``fallback_providers``), a single
        dict (legacy ``fallback_model``), or None if not configured.
        AIAgent.__init__ normalizes both formats into a chain.
        """
        try:
            import yaml as _y
            cfg_path = _hermes_home / "config.yaml"
            if cfg_path.exists():
                with open(cfg_path, encoding="utf-8") as _f:
                    cfg = _y.safe_load(_f) or {}
                fb = cfg.get("fallback_providers") or cfg.get("fallback_model") or None
                if fb:
                    return fb
        except Exception:
            pass
        return None

    @staticmethod
    def _load_smart_model_routing() -> dict:
        """Load optional smart cheap-vs-strong model routing config."""
        try:
            import yaml as _y
            cfg_path = _hermes_home / "config.yaml"
            if cfg_path.exists():
                with open(cfg_path, encoding="utf-8") as _f:
                    cfg = _y.safe_load(_f) or {}
                return cfg.get("smart_model_routing", {}) or {}
        except Exception:
            pass
        return {}

    def _snapshot_running_agents(self) -> Dict[str, Any]:
        return {
            session_key: agent
            for session_key, agent in self._running_agents.items()
            if agent is not _AGENT_PENDING_SENTINEL
        }

    def _queue_or_replace_pending_event(self, session_key: str, event: MessageEvent) -> None:
        adapter = self.adapters.get(event.source.platform)
        if not adapter:
            return
        merge_pending_message_event(adapter._pending_messages, session_key, event)

    async def _handle_active_session_busy_message(self, event: MessageEvent, session_key: str) -> bool:
        if not self._draining:
            return False

        adapter = self.adapters.get(event.source.platform)
        if not adapter:
            return True

        thread_meta = {"thread_id": event.source.thread_id} if event.source.thread_id else None
        if self._queue_during_drain_enabled():
            self._queue_or_replace_pending_event(session_key, event)
            message = f"⏳ Gateway {self._status_action_gerund()} — queued for the next turn after it comes back."
        else:
            message = f"⏳ Gateway is {self._status_action_gerund()} and is not accepting another turn right now."

        await adapter._send_with_retry(
            chat_id=event.source.chat_id,
            content=message,
            reply_to=event.message_id,
            metadata=thread_meta,
        )
        return True

    async def _drain_active_agents(self, timeout: float) -> tuple[Dict[str, Any], bool]:
        snapshot = self._snapshot_running_agents()
        last_active_count = self._running_agent_count()
        last_status_at = 0.0

        def _maybe_update_status(force: bool = False) -> None:
            nonlocal last_active_count, last_status_at
            now = asyncio.get_running_loop().time()
            active_count = self._running_agent_count()
            if force or active_count != last_active_count or (now - last_status_at) >= 1.0:
                self._update_runtime_status("draining")
                last_active_count = active_count
                last_status_at = now

        if not self._running_agents:
            _maybe_update_status(force=True)
            return snapshot, False

        _maybe_update_status(force=True)
        if timeout <= 0:
            return snapshot, True

        deadline = asyncio.get_running_loop().time() + timeout
        while self._running_agents and asyncio.get_running_loop().time() < deadline:
            _maybe_update_status()
            await asyncio.sleep(0.1)
        timed_out = bool(self._running_agents)
        _maybe_update_status(force=True)
        return snapshot, timed_out

    def _interrupt_running_agents(self, reason: str) -> None:
        for session_key, agent in list(self._running_agents.items()):
            if agent is _AGENT_PENDING_SENTINEL:
                continue
            try:
                agent.interrupt(reason)
                logger.debug("Interrupted running agent for session %s during shutdown", session_key[:20])
            except Exception as e:
                logger.debug("Failed interrupting agent during shutdown: %s", e)

    def _finalize_shutdown_agents(self, active_agents: Dict[str, Any]) -> None:
        for agent in active_agents.values():
            try:
                from hermes_cli.plugins import invoke_hook as _invoke_hook
                _invoke_hook(
                    "on_session_finalize",
                    session_id=getattr(agent, "session_id", None),
                    platform="gateway",
                )
            except Exception:
                pass
            try:
                if hasattr(agent, "shutdown_memory_provider"):
                    agent.shutdown_memory_provider()
            except Exception:
                pass
            # Close tool resources (terminal sandboxes, browser daemons,
            # background processes, httpx clients) to prevent zombie
            # process accumulation.
            try:
                if hasattr(agent, 'close'):
                    agent.close()
            except Exception:
                pass

    async def _launch_detached_restart_command(self) -> None:
        import shutil
        import subprocess

        hermes_cmd = _resolve_hermes_bin()
        if not hermes_cmd:
            logger.error("Could not locate hermes binary for detached /restart")
            return

        current_pid = os.getpid()
        cmd = " ".join(shlex.quote(part) for part in hermes_cmd)
        shell_cmd = (
            f"while kill -0 {current_pid} 2>/dev/null; do sleep 0.2; done; "
            f"{cmd} gateway restart"
        )
        setsid_bin = shutil.which("setsid")
        if setsid_bin:
            subprocess.Popen(
                [setsid_bin, "bash", "-lc", shell_cmd],
                stdout=subprocess.DEVNULL,
                stderr=subprocess.DEVNULL,
                start_new_session=True,
            )
        else:
            subprocess.Popen(
                ["bash", "-lc", shell_cmd],
                stdout=subprocess.DEVNULL,
                stderr=subprocess.DEVNULL,
                start_new_session=True,
            )

    def request_restart(self, *, detached: bool = False, via_service: bool = False) -> bool:
        if self._restart_task_started:
            return False
        self._restart_requested = True
        self._restart_detached = detached
        self._restart_via_service = via_service
        self._restart_task_started = True

        async def _run_restart() -> None:
            await asyncio.sleep(0.05)
            await self.stop(restart=True, detached_restart=detached, service_restart=via_service)

        task = asyncio.create_task(_run_restart())
        self._background_tasks.add(task)
        task.add_done_callback(self._background_tasks.discard)
        return True

    async def start(self) -> bool:
        """
        Start the gateway and all configured platform adapters.
        
        Returns True if at least one adapter connected successfully.
        """
        logger.info("Starting Hermes Gateway...")
        logger.info("Session storage: %s", self.config.sessions_dir)
        try:
            from hermes_cli.profiles import get_active_profile_name
            _profile = get_active_profile_name()
            if _profile and _profile != "default":
                logger.info("Active profile: %s", _profile)
        except Exception:
            pass
        try:
            from gateway.status import write_runtime_status
            write_runtime_status(gateway_state="starting", exit_reason=None)
        except Exception:
            pass
        
        # Warn if no user allowlists are configured and open access is not opted in
        _any_allowlist = any(
            os.getenv(v)
            for v in ("TELEGRAM_ALLOWED_USERS", "DISCORD_ALLOWED_USERS",
                       "WHATSAPP_ALLOWED_USERS", "SLACK_ALLOWED_USERS",
                       "SIGNAL_ALLOWED_USERS", "SIGNAL_GROUP_ALLOWED_USERS",
                       "EMAIL_ALLOWED_USERS",
                       "SMS_ALLOWED_USERS", "MATTERMOST_ALLOWED_USERS",
                       "MATRIX_ALLOWED_USERS", "DINGTALK_ALLOWED_USERS",
                       "FEISHU_ALLOWED_USERS",
                       "WECOM_ALLOWED_USERS",
                       "WECOM_CALLBACK_ALLOWED_USERS",
                       "WEIXIN_ALLOWED_USERS",
                       "BLUEBUBBLES_ALLOWED_USERS",
                       "GATEWAY_ALLOWED_USERS")
        )
        _allow_all = os.getenv("GATEWAY_ALLOW_ALL_USERS", "").lower() in ("true", "1", "yes") or any(
            os.getenv(v, "").lower() in ("true", "1", "yes")
            for v in ("TELEGRAM_ALLOW_ALL_USERS", "DISCORD_ALLOW_ALL_USERS",
                       "WHATSAPP_ALLOW_ALL_USERS", "SLACK_ALLOW_ALL_USERS",
                       "SIGNAL_ALLOW_ALL_USERS", "EMAIL_ALLOW_ALL_USERS",
                       "SMS_ALLOW_ALL_USERS", "MATTERMOST_ALLOW_ALL_USERS",
                       "MATRIX_ALLOW_ALL_USERS", "DINGTALK_ALLOW_ALL_USERS",
                       "FEISHU_ALLOW_ALL_USERS",
                       "WECOM_ALLOW_ALL_USERS",
                       "WECOM_CALLBACK_ALLOW_ALL_USERS",
                       "WEIXIN_ALLOW_ALL_USERS",
                       "BLUEBUBBLES_ALLOW_ALL_USERS")
        )
        if not _any_allowlist and not _allow_all:
            logger.warning(
                "No user allowlists configured. All unauthorized users will be denied. "
                "Set GATEWAY_ALLOW_ALL_USERS=true in ~/.hermes/.env to allow open access, "
                "or configure platform allowlists (e.g., TELEGRAM_ALLOWED_USERS=your_id)."
            )
        
        # Discover and load event hooks
        self.hooks.discover_and_load()
        
        # Recover background processes from checkpoint (crash recovery)
        try:
            from tools.process_registry import process_registry
            recovered = process_registry.recover_from_checkpoint()
            if recovered:
                logger.info("Recovered %s background process(es) from previous run", recovered)
        except Exception as e:
            logger.warning("Process checkpoint recovery: %s", e)

        # Suspend sessions that were active when the gateway last exited.
        # This prevents stuck sessions from being blindly resumed on restart,
        # which can create an unrecoverable loop (#7536).  Suspended sessions
        # auto-reset on the next incoming message, giving the user a clean start.
        try:
            suspended = self.session_store.suspend_recently_active()
            if suspended:
                logger.info("Suspended %d in-flight session(s) from previous run", suspended)
        except Exception as e:
            logger.warning("Session suspension on startup failed: %s", e)

        connected_count = 0
        enabled_platform_count = 0
        startup_nonretryable_errors: list[str] = []
        startup_retryable_errors: list[str] = []
        
        # Initialize and connect each configured platform
        for platform, platform_config in self.config.platforms.items():
            if not platform_config.enabled:
                continue
            enabled_platform_count += 1
            
            adapter = self._create_adapter(platform, platform_config)
            if not adapter:
                logger.warning("No adapter available for %s", platform.value)
                continue
            
            # Set up message + fatal error handlers
            adapter.set_message_handler(self._handle_message)
            adapter.set_fatal_error_handler(self._handle_adapter_fatal_error)
            adapter.set_session_store(self.session_store)
            adapter.set_busy_session_handler(self._handle_active_session_busy_message)
            
            # Try to connect
            logger.info("Connecting to %s...", platform.value)
            self._update_platform_runtime_status(
                platform.value,
                platform_state="connecting",
                error_code=None,
                error_message=None,
            )
            try:
                success = await adapter.connect()
                if success:
                    self.adapters[platform] = adapter
                    self._sync_voice_mode_state_to_adapter(adapter)
                    connected_count += 1
                    self._update_platform_runtime_status(
                        platform.value,
                        platform_state="connected",
                        error_code=None,
                        error_message=None,
                    )
                    logger.info("✓ %s connected", platform.value)
                else:
                    logger.warning("✗ %s failed to connect", platform.value)
                    if adapter.has_fatal_error:
                        self._update_platform_runtime_status(
                            platform.value,
                            platform_state="retrying" if adapter.fatal_error_retryable else "fatal",
                            error_code=adapter.fatal_error_code,
                            error_message=adapter.fatal_error_message,
                        )
                        target = (
                            startup_retryable_errors
                            if adapter.fatal_error_retryable
                            else startup_nonretryable_errors
                        )
                        target.append(
                            f"{platform.value}: {adapter.fatal_error_message}"
                        )
                        # Queue for reconnection if the error is retryable
                        if adapter.fatal_error_retryable:
                            self._failed_platforms[platform] = {
                                "config": platform_config,
                                "attempts": 1,
                                "next_retry": time.monotonic() + 30,
                            }
                    else:
                        self._update_platform_runtime_status(
                            platform.value,
                            platform_state="retrying",
                            error_code=None,
                            error_message="failed to connect",
                        )
                        startup_retryable_errors.append(
                            f"{platform.value}: failed to connect"
                        )
                        # No fatal error info means likely a transient issue — queue for retry
                        self._failed_platforms[platform] = {
                            "config": platform_config,
                            "attempts": 1,
                            "next_retry": time.monotonic() + 30,
                        }
            except Exception as e:
                logger.error("✗ %s error: %s", platform.value, e)
                self._update_platform_runtime_status(
                    platform.value,
                    platform_state="retrying",
                    error_code=None,
                    error_message=str(e),
                )
                startup_retryable_errors.append(f"{platform.value}: {e}")
                # Unexpected exceptions are typically transient — queue for retry
                self._failed_platforms[platform] = {
                    "config": platform_config,
                    "attempts": 1,
                    "next_retry": time.monotonic() + 30,
                }
        
        if connected_count == 0:
            if startup_nonretryable_errors:
                reason = "; ".join(startup_nonretryable_errors)
                logger.error("Gateway hit a non-retryable startup conflict: %s", reason)
                try:
                    from gateway.status import write_runtime_status
                    write_runtime_status(gateway_state="startup_failed", exit_reason=reason)
                except Exception:
                    pass
                self._request_clean_exit(reason)
                return True
            if enabled_platform_count > 0:
                reason = "; ".join(startup_retryable_errors) or "all configured messaging platforms failed to connect"
                logger.error("Gateway failed to connect any configured messaging platform: %s", reason)
                try:
                    from gateway.status import write_runtime_status
                    write_runtime_status(gateway_state="startup_failed", exit_reason=reason)
                except Exception:
                    pass
                return False
            logger.warning("No messaging platforms enabled.")
            logger.info("Gateway will continue running for cron job execution.")
        
        # Update delivery router with adapters
        self.delivery_router.adapters = self.adapters
        
        self._running = True
        self._update_runtime_status("running")
        
        # Emit gateway:startup hook
        hook_count = len(self.hooks.loaded_hooks)
        if hook_count:
            logger.info("%s hook(s) loaded", hook_count)
        await self.hooks.emit("gateway:startup", {
            "platforms": [p.value for p in self.adapters.keys()],
        })
        
        if connected_count > 0:
            logger.info("Gateway running with %s platform(s)", connected_count)
        
        # Build initial channel directory for send_message name resolution
        try:
            from gateway.channel_directory import build_channel_directory
            directory = build_channel_directory(self.adapters)
            ch_count = sum(len(chs) for chs in directory.get("platforms", {}).values())
            logger.info("Channel directory built: %d target(s)", ch_count)
        except Exception as e:
            logger.warning("Channel directory build failed: %s", e)
        
        # Check if we're restarting after a /update command. If the update is
        # still running, keep watching so we notify once it actually finishes.
        notified = await self._send_update_notification()
        if not notified and any(
            path.exists()
            for path in (
                _hermes_home / ".update_pending.json",
                _hermes_home / ".update_pending.claimed.json",
            )
        ):
            self._schedule_update_notification_watch()

        # Drain any recovered process watchers (from crash recovery checkpoint)
        try:
            from tools.process_registry import process_registry
            while process_registry.pending_watchers:
                watcher = process_registry.pending_watchers.pop(0)
                asyncio.create_task(self._run_process_watcher(watcher))
                logger.info("Resumed watcher for recovered process %s", watcher.get("session_id"))
        except Exception as e:
            logger.error("Recovered watcher setup error: %s", e)

        # Start background session expiry watcher for proactive memory flushing
        asyncio.create_task(self._session_expiry_watcher())

        # Start background reconnection watcher for platforms that failed at startup
        if self._failed_platforms:
            logger.info(
                "Starting reconnection watcher for %d failed platform(s): %s",
                len(self._failed_platforms),
                ", ".join(p.value for p in self._failed_platforms),
            )
        asyncio.create_task(self._platform_reconnect_watcher())

        logger.info("Press Ctrl+C to stop")
        
        return True
    
    async def _session_expiry_watcher(self, interval: int = 300):
        """Background task that proactively flushes memories for expired sessions.
        
        Runs every `interval` seconds (default 5 min).  For each session that
        has expired according to its reset policy, flushes memories in a thread
        pool and marks the session so it won't be flushed again.

        This means memories are already saved by the time the user sends their
        next message, so there's no blocking delay.
        """
        await asyncio.sleep(60)  # initial delay — let the gateway fully start
        _flush_failures: dict[str, int] = {}  # session_id -> consecutive failure count
        _MAX_FLUSH_RETRIES = 3
        while self._running:
            try:
                self.session_store._ensure_loaded()
                # Collect expired sessions first, then log a single summary.
                _expired_entries = []
                for key, entry in list(self.session_store._entries.items()):
                    if entry.memory_flushed:
                        continue
                    if not self.session_store._is_session_expired(entry):
                        continue
                    _expired_entries.append((key, entry))

                if _expired_entries:
                    # Extract platform names from session keys for a compact summary.
                    # Keys look like "agent:main:telegram:dm:12345" — platform is field [2].
                    _platforms: dict[str, int] = {}
                    for _k, _e in _expired_entries:
                        _parts = _k.split(":")
                        _plat = _parts[2] if len(_parts) > 2 else "unknown"
                        _platforms[_plat] = _platforms.get(_plat, 0) + 1
                    _plat_summary = ", ".join(
                        f"{p}:{c}" for p, c in sorted(_platforms.items())
                    )
                    logger.info(
                        "Session expiry: %d sessions to flush (%s)",
                        len(_expired_entries), _plat_summary,
                    )

                for key, entry in _expired_entries:
                    try:
                        await self._async_flush_memories(entry.session_id, key)
                        # Shut down memory provider and close tool resources
                        # on the cached agent.  Idle agents live in
                        # _agent_cache (not _running_agents), so look there.
                        _cached_agent = None
                        _cache_lock = getattr(self, "_agent_cache_lock", None)
                        if _cache_lock is not None:
                            with _cache_lock:
                                _cached = self._agent_cache.get(key)
                                _cached_agent = _cached[0] if isinstance(_cached, tuple) else _cached if _cached else None
                        # Fall back to _running_agents in case the agent is
                        # still mid-turn when the expiry fires.
                        if _cached_agent is None:
                            _cached_agent = self._running_agents.get(key)
                        if _cached_agent and _cached_agent is not _AGENT_PENDING_SENTINEL:
                            try:
                                if hasattr(_cached_agent, 'shutdown_memory_provider'):
                                    _cached_agent.shutdown_memory_provider()
                            except Exception:
                                pass
                            try:
                                if hasattr(_cached_agent, 'close'):
                                    _cached_agent.close()
                            except Exception:
                                pass
                        # Mark as flushed and persist to disk so the flag
                        # survives gateway restarts.
                        with self.session_store._lock:
                            entry.memory_flushed = True
                            self.session_store._save()
                        logger.debug(
                            "Memory flush completed for session %s",
                            entry.session_id,
                        )
                        _flush_failures.pop(entry.session_id, None)
                    except Exception as e:
                        failures = _flush_failures.get(entry.session_id, 0) + 1
                        _flush_failures[entry.session_id] = failures
                        if failures >= _MAX_FLUSH_RETRIES:
                            logger.warning(
                                "Memory flush gave up after %d attempts for %s: %s. "
                                "Marking as flushed to prevent infinite retry loop.",
                                failures, entry.session_id, e,
                            )
                            with self.session_store._lock:
                                entry.memory_flushed = True
                                self.session_store._save()
                            _flush_failures.pop(entry.session_id, None)
                        else:
                            logger.debug(
                                "Memory flush failed (%d/%d) for %s: %s",
                                failures, _MAX_FLUSH_RETRIES, entry.session_id, e,
                            )

                if _expired_entries:
                    _flushed = sum(
                        1 for _, e in _expired_entries if e.memory_flushed
                    )
                    _failed = len(_expired_entries) - _flushed
                    if _failed:
                        logger.info(
                            "Session expiry done: %d flushed, %d pending retry",
                            _flushed, _failed,
                        )
                    else:
                        logger.info(
                            "Session expiry done: %d flushed", _flushed,
                        )
            except Exception as e:
                logger.debug("Session expiry watcher error: %s", e)
            # Sleep in small increments so we can stop quickly
            for _ in range(interval):
                if not self._running:
                    break
                await asyncio.sleep(1)

    async def _platform_reconnect_watcher(self) -> None:
        """Background task that periodically retries connecting failed platforms.

        Uses exponential backoff: 30s → 60s → 120s → 240s → 300s (cap).
        Stops retrying a platform after 20 failed attempts or if the error
        is non-retryable (e.g. bad auth token).
        """
        _MAX_ATTEMPTS = 20
        _BACKOFF_CAP = 300  # 5 minutes max between retries

        await asyncio.sleep(10)  # initial delay — let startup finish
        while self._running:
            if not self._failed_platforms:
                # Nothing to reconnect — sleep and check again
                for _ in range(30):
                    if not self._running:
                        return
                    await asyncio.sleep(1)
                continue

            now = time.monotonic()
            for platform in list(self._failed_platforms.keys()):
                if not self._running:
                    return
                info = self._failed_platforms[platform]
                if now < info["next_retry"]:
                    continue  # not time yet

                if info["attempts"] >= _MAX_ATTEMPTS:
                    logger.warning(
                        "Giving up reconnecting %s after %d attempts",
                        platform.value, info["attempts"],
                    )
                    del self._failed_platforms[platform]
                    continue

                platform_config = info["config"]
                attempt = info["attempts"] + 1
                logger.info(
                    "Reconnecting %s (attempt %d/%d)...",
                    platform.value, attempt, _MAX_ATTEMPTS,
                )

                try:
                    adapter = self._create_adapter(platform, platform_config)
                    if not adapter:
                        logger.warning(
                            "Reconnect %s: adapter creation returned None, removing from retry queue",
                            platform.value,
                        )
                        del self._failed_platforms[platform]
                        continue

                    adapter.set_message_handler(self._handle_message)
                    adapter.set_fatal_error_handler(self._handle_adapter_fatal_error)
                    adapter.set_session_store(self.session_store)
                    adapter.set_busy_session_handler(self._handle_active_session_busy_message)

                    success = await adapter.connect()
                    if success:
                        self.adapters[platform] = adapter
                        self._sync_voice_mode_state_to_adapter(adapter)
                        self.delivery_router.adapters = self.adapters
                        del self._failed_platforms[platform]
                        self._update_platform_runtime_status(
                            platform.value,
                            platform_state="connected",
                            error_code=None,
                            error_message=None,
                        )
                        logger.info("✓ %s reconnected successfully", platform.value)

                        # Rebuild channel directory with the new adapter
                        try:
                            from gateway.channel_directory import build_channel_directory
                            build_channel_directory(self.adapters)
                        except Exception:
                            pass
                    else:
                        # Check if the failure is non-retryable
                        if adapter.has_fatal_error and not adapter.fatal_error_retryable:
                            self._update_platform_runtime_status(
                                platform.value,
                                platform_state="fatal",
                                error_code=adapter.fatal_error_code,
                                error_message=adapter.fatal_error_message,
                            )
                            logger.warning(
                                "Reconnect %s: non-retryable error (%s), removing from retry queue",
                                platform.value, adapter.fatal_error_message,
                            )
                            del self._failed_platforms[platform]
                        else:
                            self._update_platform_runtime_status(
                                platform.value,
                                platform_state="retrying",
                                error_code=adapter.fatal_error_code,
                                error_message=adapter.fatal_error_message or "failed to reconnect",
                            )
                            backoff = min(30 * (2 ** (attempt - 1)), _BACKOFF_CAP)
                            info["attempts"] = attempt
                            info["next_retry"] = time.monotonic() + backoff
                            logger.info(
                                "Reconnect %s failed, next retry in %ds",
                                platform.value, backoff,
                            )
                except Exception as e:
                    self._update_platform_runtime_status(
                        platform.value,
                        platform_state="retrying",
                        error_code=None,
                        error_message=str(e),
                    )
                    backoff = min(30 * (2 ** (attempt - 1)), _BACKOFF_CAP)
                    info["attempts"] = attempt
                    info["next_retry"] = time.monotonic() + backoff
                    logger.warning(
                        "Reconnect %s error: %s, next retry in %ds",
                        platform.value, e, backoff,
                    )

            # Check every 10 seconds for platforms that need reconnection
            for _ in range(10):
                if not self._running:
                    return
                await asyncio.sleep(1)

    async def stop(
        self,
        *,
        restart: bool = False,
        detached_restart: bool = False,
        service_restart: bool = False,
    ) -> None:
        """Stop the gateway and disconnect all adapters."""
        if restart:
            self._restart_requested = True
            self._restart_detached = detached_restart
            self._restart_via_service = service_restart
        if self._stop_task is not None:
            await self._stop_task
            return

        async def _stop_impl() -> None:
            logger.info(
                "Stopping gateway%s...",
                " for restart" if self._restart_requested else "",
            )
            self._running = False
            self._draining = True

            timeout = self._restart_drain_timeout
            active_agents, timed_out = await self._drain_active_agents(timeout)
            if timed_out:
                logger.warning(
                    "Gateway drain timed out after %.1fs with %d active agent(s); interrupting remaining work.",
                    timeout,
                    self._running_agent_count(),
                )
                self._interrupt_running_agents(
                    "Gateway restarting" if self._restart_requested else "Gateway shutting down"
                )
                interrupt_deadline = asyncio.get_running_loop().time() + 5.0
                while self._running_agents and asyncio.get_running_loop().time() < interrupt_deadline:
                    self._update_runtime_status("draining")
                    await asyncio.sleep(0.1)

            if self._restart_requested and self._restart_detached:
                try:
                    await self._launch_detached_restart_command()
                except Exception as e:
                    logger.error("Failed to launch detached gateway restart: %s", e)

            self._finalize_shutdown_agents(active_agents)

            for platform, adapter in list(self.adapters.items()):
                try:
                    await adapter.cancel_background_tasks()
                except Exception as e:
                    logger.debug("✗ %s background-task cancel error: %s", platform.value, e)
                try:
                    await adapter.disconnect()
                    logger.info("✓ %s disconnected", platform.value)
                except Exception as e:
                    logger.error("✗ %s disconnect error: %s", platform.value, e)

            for _task in list(self._background_tasks):
                if _task is self._stop_task:
                    continue
                _task.cancel()
            self._background_tasks.clear()

            self.adapters.clear()
            self._running_agents.clear()
            self._pending_messages.clear()
            self._pending_approvals.clear()
            self._shutdown_event.set()

            # Global cleanup: kill any remaining tool subprocesses not tied
            # to a specific agent (catch-all for zombie prevention).
            try:
                from tools.process_registry import process_registry
                process_registry.kill_all()
            except Exception:
                pass
            try:
                from tools.terminal_tool import cleanup_all_environments
                cleanup_all_environments()
            except Exception:
                pass
            try:
                from tools.browser_tool import cleanup_all_browsers
                cleanup_all_browsers()
            except Exception:
                pass

            from gateway.status import remove_pid_file
            remove_pid_file()

            if self._restart_requested and self._restart_via_service:
                self._exit_code = GATEWAY_SERVICE_RESTART_EXIT_CODE
                self._exit_reason = self._exit_reason or "Gateway restart requested"

            self._draining = False
            self._update_runtime_status("stopped", self._exit_reason)
            logger.info("Gateway stopped")

        self._stop_task = asyncio.create_task(_stop_impl())
        await self._stop_task
    
    async def wait_for_shutdown(self) -> None:
        """Wait for shutdown signal."""
        await self._shutdown_event.wait()
    
    def _create_adapter(
        self, 
        platform: Platform, 
        config: Any
    ) -> Optional[BasePlatformAdapter]:
        """Create the appropriate adapter for a platform."""
        if hasattr(config, "extra") and isinstance(config.extra, dict):
            config.extra.setdefault(
                "group_sessions_per_user",
                self.config.group_sessions_per_user,
            )
            config.extra.setdefault(
                "thread_sessions_per_user",
                getattr(self.config, "thread_sessions_per_user", False),
            )

        if platform == Platform.TELEGRAM:
            from gateway.platforms.telegram import TelegramAdapter, check_telegram_requirements
            if not check_telegram_requirements():
                logger.warning("Telegram: python-telegram-bot not installed")
                return None
            return TelegramAdapter(config)
        
        elif platform == Platform.DISCORD:
            from gateway.platforms.discord import DiscordAdapter, check_discord_requirements
            if not check_discord_requirements():
                logger.warning("Discord: discord.py not installed")
                return None
            return DiscordAdapter(config)
        
        elif platform == Platform.WHATSAPP:
            from gateway.platforms.whatsapp import WhatsAppAdapter, check_whatsapp_requirements
            if not check_whatsapp_requirements():
                logger.warning("WhatsApp: Node.js not installed or bridge not configured")
                return None
            return WhatsAppAdapter(config)
        
        elif platform == Platform.SLACK:
            from gateway.platforms.slack import SlackAdapter, check_slack_requirements
            if not check_slack_requirements():
                logger.warning("Slack: slack-bolt not installed. Run: pip install 'hermes-agent[slack]'")
                return None
            return SlackAdapter(config)

        elif platform == Platform.SIGNAL:
            from gateway.platforms.signal import SignalAdapter, check_signal_requirements
            if not check_signal_requirements():
                logger.warning("Signal: SIGNAL_HTTP_URL or SIGNAL_ACCOUNT not configured")
                return None
            return SignalAdapter(config)

        elif platform == Platform.HOMEASSISTANT:
            from gateway.platforms.homeassistant import HomeAssistantAdapter, check_ha_requirements
            if not check_ha_requirements():
                logger.warning("HomeAssistant: aiohttp not installed or HASS_TOKEN not set")
                return None
            return HomeAssistantAdapter(config)

        elif platform == Platform.EMAIL:
            from gateway.platforms.email import EmailAdapter, check_email_requirements
            if not check_email_requirements():
                logger.warning("Email: EMAIL_ADDRESS, EMAIL_PASSWORD, EMAIL_IMAP_HOST, or EMAIL_SMTP_HOST not set")
                return None
            return EmailAdapter(config)

        elif platform == Platform.SMS:
            from gateway.platforms.sms import SmsAdapter, check_sms_requirements
            if not check_sms_requirements():
                logger.warning("SMS: aiohttp not installed or TWILIO_ACCOUNT_SID/TWILIO_AUTH_TOKEN not set")
                return None
            return SmsAdapter(config)

        elif platform == Platform.DINGTALK:
            from gateway.platforms.dingtalk import DingTalkAdapter, check_dingtalk_requirements
            if not check_dingtalk_requirements():
                logger.warning("DingTalk: dingtalk-stream not installed or DINGTALK_CLIENT_ID/SECRET not set")
                return None
            return DingTalkAdapter(config)

        elif platform == Platform.FEISHU:
            from gateway.platforms.feishu import FeishuAdapter, check_feishu_requirements
            if not check_feishu_requirements():
                logger.warning("Feishu: lark-oapi not installed or FEISHU_APP_ID/SECRET not set")
                return None
            return FeishuAdapter(config)

        elif platform == Platform.WECOM_CALLBACK:
            from gateway.platforms.wecom_callback import (
                WecomCallbackAdapter,
                check_wecom_callback_requirements,
            )
            if not check_wecom_callback_requirements():
                logger.warning("WeComCallback: aiohttp/httpx not installed")
                return None
            return WecomCallbackAdapter(config)

        elif platform == Platform.WECOM:
            from gateway.platforms.wecom import WeComAdapter, check_wecom_requirements
            if not check_wecom_requirements():
                logger.warning("WeCom: aiohttp not installed or WECOM_BOT_ID/SECRET not set")
                return None
            return WeComAdapter(config)

        elif platform == Platform.WEIXIN:
            from gateway.platforms.weixin import WeixinAdapter, check_weixin_requirements
            if not check_weixin_requirements():
                logger.warning("Weixin: aiohttp/cryptography not installed")
                return None
            return WeixinAdapter(config)

        elif platform == Platform.MATTERMOST:
            from gateway.platforms.mattermost import MattermostAdapter, check_mattermost_requirements
            if not check_mattermost_requirements():
                logger.warning("Mattermost: MATTERMOST_TOKEN or MATTERMOST_URL not set, or aiohttp missing")
                return None
            return MattermostAdapter(config)

        elif platform == Platform.MATRIX:
            from gateway.platforms.matrix import MatrixAdapter, check_matrix_requirements
            if not check_matrix_requirements():
                logger.warning("Matrix: mautrix not installed or credentials not set. Run: pip install 'mautrix[encryption]'")
                return None
            return MatrixAdapter(config)

        elif platform == Platform.API_SERVER:
            from gateway.platforms.api_server import APIServerAdapter, check_api_server_requirements
            if not check_api_server_requirements():
                logger.warning("API Server: aiohttp not installed")
                return None
            return APIServerAdapter(config)

        elif platform == Platform.WEBHOOK:
            from gateway.platforms.webhook import WebhookAdapter, check_webhook_requirements
            if not check_webhook_requirements():
                logger.warning("Webhook: aiohttp not installed")
                return None
            adapter = WebhookAdapter(config)
            adapter.gateway_runner = self  # For cross-platform delivery
            return adapter

        elif platform == Platform.BLUEBUBBLES:
            from gateway.platforms.bluebubbles import BlueBubblesAdapter, check_bluebubbles_requirements
            if not check_bluebubbles_requirements():
                logger.warning("BlueBubbles: aiohttp/httpx missing or BLUEBUBBLES_SERVER_URL/BLUEBUBBLES_PASSWORD not configured")
                return None
            return BlueBubblesAdapter(config)

        return None
    
    def _is_user_authorized(self, source: SessionSource) -> bool:
        """
        Check if a user is authorized to use the bot.
        
        Checks in order:
        1. Per-platform allow-all flag (e.g., DISCORD_ALLOW_ALL_USERS=true)
        2. Environment variable allowlists (TELEGRAM_ALLOWED_USERS, etc.)
        3. DM pairing approved list
        4. Global allow-all (GATEWAY_ALLOW_ALL_USERS=true)
        5. Default: deny
        """
        # Home Assistant events are system-generated (state changes), not
        # user-initiated messages.  The HASS_TOKEN already authenticates the
        # connection, so HA events are always authorized.
        # Webhook events are authenticated via HMAC signature validation in
        # the adapter itself — no user allowlist applies.
        if source.platform in (Platform.HOMEASSISTANT, Platform.WEBHOOK):
            return True

        user_id = source.user_id
        if not user_id:
            return False

        platform_env_map = {
            Platform.TELEGRAM: "TELEGRAM_ALLOWED_USERS",
            Platform.DISCORD: "DISCORD_ALLOWED_USERS",
            Platform.WHATSAPP: "WHATSAPP_ALLOWED_USERS",
            Platform.SLACK: "SLACK_ALLOWED_USERS",
            Platform.SIGNAL: "SIGNAL_ALLOWED_USERS",
            Platform.EMAIL: "EMAIL_ALLOWED_USERS",
            Platform.SMS: "SMS_ALLOWED_USERS",
            Platform.MATTERMOST: "MATTERMOST_ALLOWED_USERS",
            Platform.MATRIX: "MATRIX_ALLOWED_USERS",
            Platform.DINGTALK: "DINGTALK_ALLOWED_USERS",
            Platform.FEISHU: "FEISHU_ALLOWED_USERS",
            Platform.WECOM: "WECOM_ALLOWED_USERS",
            Platform.WECOM_CALLBACK: "WECOM_CALLBACK_ALLOWED_USERS",
            Platform.WEIXIN: "WEIXIN_ALLOWED_USERS",
            Platform.BLUEBUBBLES: "BLUEBUBBLES_ALLOWED_USERS",
        }
        platform_allow_all_map = {
            Platform.TELEGRAM: "TELEGRAM_ALLOW_ALL_USERS",
            Platform.DISCORD: "DISCORD_ALLOW_ALL_USERS",
            Platform.WHATSAPP: "WHATSAPP_ALLOW_ALL_USERS",
            Platform.SLACK: "SLACK_ALLOW_ALL_USERS",
            Platform.SIGNAL: "SIGNAL_ALLOW_ALL_USERS",
            Platform.EMAIL: "EMAIL_ALLOW_ALL_USERS",
            Platform.SMS: "SMS_ALLOW_ALL_USERS",
            Platform.MATTERMOST: "MATTERMOST_ALLOW_ALL_USERS",
            Platform.MATRIX: "MATRIX_ALLOW_ALL_USERS",
            Platform.DINGTALK: "DINGTALK_ALLOW_ALL_USERS",
            Platform.FEISHU: "FEISHU_ALLOW_ALL_USERS",
            Platform.WECOM: "WECOM_ALLOW_ALL_USERS",
            Platform.WECOM_CALLBACK: "WECOM_CALLBACK_ALLOW_ALL_USERS",
            Platform.WEIXIN: "WEIXIN_ALLOW_ALL_USERS",
            Platform.BLUEBUBBLES: "BLUEBUBBLES_ALLOW_ALL_USERS",
        }

        # Per-platform allow-all flag (e.g., DISCORD_ALLOW_ALL_USERS=true)
        platform_allow_all_var = platform_allow_all_map.get(source.platform, "")
        if platform_allow_all_var and os.getenv(platform_allow_all_var, "").lower() in ("true", "1", "yes"):
            return True

        # Check pairing store (always checked, regardless of allowlists)
        platform_name = source.platform.value if source.platform else ""
        if self.pairing_store.is_approved(platform_name, user_id):
            return True

        # Check platform-specific and global allowlists
        platform_allowlist = os.getenv(platform_env_map.get(source.platform, ""), "").strip()
        global_allowlist = os.getenv("GATEWAY_ALLOWED_USERS", "").strip()

        if not platform_allowlist and not global_allowlist:
            # No allowlists configured -- check global allow-all flag
            return os.getenv("GATEWAY_ALLOW_ALL_USERS", "").lower() in ("true", "1", "yes")

        # Check if user is in any allowlist
        allowed_ids = set()
        if platform_allowlist:
            allowed_ids.update(uid.strip() for uid in platform_allowlist.split(",") if uid.strip())
        if global_allowlist:
            allowed_ids.update(uid.strip() for uid in global_allowlist.split(",") if uid.strip())

        # "*" in any allowlist means allow everyone (consistent with
        # SIGNAL_GROUP_ALLOWED_USERS precedent)
        if "*" in allowed_ids:
            return True

        check_ids = {user_id}
        if "@" in user_id:
            check_ids.add(user_id.split("@")[0])

        # WhatsApp: resolve phone↔LID aliases from bridge session mapping files
        if source.platform == Platform.WHATSAPP:
            normalized_allowed_ids = set()
            for allowed_id in allowed_ids:
                normalized_allowed_ids.update(_expand_whatsapp_auth_aliases(allowed_id))
            if normalized_allowed_ids:
                allowed_ids = normalized_allowed_ids

            check_ids.update(_expand_whatsapp_auth_aliases(user_id))
            normalized_user_id = _normalize_whatsapp_identifier(user_id)
            if normalized_user_id:
                check_ids.add(normalized_user_id)

        return bool(check_ids & allowed_ids)

    def _get_unauthorized_dm_behavior(self, platform: Optional[Platform]) -> str:
        """Return how unauthorized DMs should be handled for a platform."""
        config = getattr(self, "config", None)
        if config and hasattr(config, "get_unauthorized_dm_behavior"):
            return config.get_unauthorized_dm_behavior(platform)
        return "pair"
    
    async def _handle_message(self, event: MessageEvent) -> Optional[str]:
        """
        Handle an incoming message from any platform.
        
        This is the core message processing pipeline:
        1. Check user authorization
        2. Check for commands (/new, /reset, etc.)
        3. Check for running agent and interrupt if needed
        4. Get or create session
        5. Build context for agent
        6. Run agent conversation
        7. Return response
        """
        source = event.source

        # Internal events (e.g. background-process completion notifications)
        # are system-generated and must skip user authorization.
        if getattr(event, "internal", False):
            pass
        elif source.user_id is None:
            # Messages with no user identity (Telegram service messages,
            # channel forwards, anonymous admin actions) cannot be
            # authorized — drop silently instead of triggering the pairing
            # flow with a None user_id.
            logger.debug("Ignoring message with no user_id from %s", source.platform.value)
            return None
        elif not self._is_user_authorized(source):
            logger.warning("Unauthorized user: %s (%s) on %s", source.user_id, source.user_name, source.platform.value)
            # In DMs: offer pairing code. In groups: silently ignore.
            if source.chat_type == "dm" and self._get_unauthorized_dm_behavior(source.platform) == "pair":
                platform_name = source.platform.value if source.platform else "unknown"
                # Rate-limit ALL pairing responses (code or rejection) to
                # prevent spamming the user with repeated messages when
                # multiple DMs arrive in quick succession.
                if self.pairing_store._is_rate_limited(platform_name, source.user_id):
                    return None
                code = self.pairing_store.generate_code(
                    platform_name, source.user_id, source.user_name or ""
                )
                if code:
                    adapter = self.adapters.get(source.platform)
                    if adapter:
                        await adapter.send(
                            source.chat_id,
                            f"Hi~ I don't recognize you yet!\n\n"
                            f"Here's your pairing code: `{code}`\n\n"
                            f"Ask the bot owner to run:\n"
                            f"`hermes pairing approve {platform_name} {code}`"
                        )
                else:
                    adapter = self.adapters.get(source.platform)
                    if adapter:
                        await adapter.send(
                            source.chat_id,
                            "Too many pairing requests right now~ "
                            "Please try again later!"
                        )
                    # Record rate limit so subsequent messages are silently ignored
                    self.pairing_store._record_rate_limit(platform_name, source.user_id)
            return None
        
        # Intercept messages that are responses to a pending /update prompt.
        # The update process (detached) wrote .update_prompt.json; the watcher
        # forwarded it to the user; now the user's reply goes back via
        # .update_response so the update process can continue.
        _quick_key = self._session_key_for_source(source)
        _update_prompts = getattr(self, "_update_prompt_pending", {})
        if _update_prompts.get(_quick_key):
            raw = (event.text or "").strip()
            # Accept /approve and /deny as shorthand for yes/no
            cmd = event.get_command()
            if cmd in ("approve", "yes"):
                response_text = "y"
            elif cmd in ("deny", "no"):
                response_text = "n"
            else:
                response_text = raw
            if response_text:
                response_path = _hermes_home / ".update_response"
                try:
                    tmp = response_path.with_suffix(".tmp")
                    tmp.write_text(response_text)
                    tmp.replace(response_path)
                except OSError as e:
                    logger.warning("Failed to write update response: %s", e)
                    return f"✗ Failed to send response to update process: {e}"
                _update_prompts.pop(_quick_key, None)
                label = response_text if len(response_text) <= 20 else response_text[:20] + "…"
                return f"✓ Sent `{label}` to the update process."

        # PRIORITY handling when an agent is already running for this session.
        # Default behavior is to interrupt immediately so user text/stop messages
        # are handled with minimal latency.
        #
        # Special case: Telegram/photo bursts often arrive as multiple near-
        # simultaneous updates. Do NOT interrupt for photo-only follow-ups here;
        # let the adapter-level batching/queueing logic absorb them.

        # Staleness eviction: detect leaked locks from hung/crashed handlers.
        # With inactivity-based timeout, active tasks can run for hours, so
        # wall-clock age alone isn't sufficient.  Evict only when the agent
        # has been *idle* beyond the inactivity threshold (or when the agent
        # object has no activity tracker and wall-clock age is extreme).
        _raw_stale_timeout = float(os.getenv("HERMES_AGENT_TIMEOUT", 1800))
        _stale_ts = self._running_agents_ts.get(_quick_key, 0)
        if _quick_key in self._running_agents and _stale_ts:
            _stale_age = time.time() - _stale_ts
            _stale_agent = self._running_agents.get(_quick_key)
            # Never evict the pending sentinel — it was just placed moments
            # ago during the async setup phase before the real agent is
            # created.  Sentinels have no get_activity_summary(), so the
            # idle check below would always evaluate to inf >= timeout and
            # immediately evict them, racing with the setup path.
            _stale_idle = float("inf")  # assume idle if we can't check
            _stale_detail = ""
            if _stale_agent and hasattr(_stale_agent, "get_activity_summary"):
                try:
                    _sa = _stale_agent.get_activity_summary()
                    _stale_idle = _sa.get("seconds_since_activity", float("inf"))
                    _stale_detail = (
                        f" | last_activity={_sa.get('last_activity_desc', 'unknown')} "
                        f"({_stale_idle:.0f}s ago) "
                        f"| iteration={_sa.get('api_call_count', 0)}/{_sa.get('max_iterations', 0)}"
                    )
                except Exception:
                    pass
            # Evict if: agent is idle beyond timeout, OR wall-clock age is
            # extreme (10x timeout or 2h, whichever is larger — catches
            # cases where the agent object was garbage-collected).
            _wall_ttl = max(_raw_stale_timeout * 10, 7200) if _raw_stale_timeout > 0 else float("inf")
            _should_evict = (
                _stale_agent is not _AGENT_PENDING_SENTINEL
                and (
                    (_raw_stale_timeout > 0 and _stale_idle >= _raw_stale_timeout)
                    or _stale_age > _wall_ttl
                )
            )
            if _should_evict:
                logger.warning(
                    "Evicting stale _running_agents entry for %s "
                    "(age: %.0fs, idle: %.0fs, timeout: %.0fs)%s",
                    _quick_key[:30], _stale_age, _stale_idle,
                    _raw_stale_timeout, _stale_detail,
                )
                del self._running_agents[_quick_key]
                self._running_agents_ts.pop(_quick_key, None)

        if _quick_key in self._running_agents:
            if event.get_command() == "status":
                return await self._handle_status_command(event)

            # Resolve the command once for all early-intercept checks below.
            from hermes_cli.commands import resolve_command as _resolve_cmd_inner
            _evt_cmd = event.get_command()
            _cmd_def_inner = _resolve_cmd_inner(_evt_cmd) if _evt_cmd else None

            if _cmd_def_inner and _cmd_def_inner.name == "restart":
                return await self._handle_restart_command(event)

            # /stop must hard-kill the session when an agent is running.
            # A soft interrupt (agent.interrupt()) doesn't help when the agent
            # is truly hung — the executor thread is blocked and never checks
            # _interrupt_requested.  Force-clean _running_agents so the session
            # is unlocked and subsequent messages are processed normally.
            if _cmd_def_inner and _cmd_def_inner.name == "stop":
                running_agent = self._running_agents.get(_quick_key)
                if running_agent and running_agent is not _AGENT_PENDING_SENTINEL:
                    running_agent.interrupt("Stop requested")
                # Force-clean: remove the session lock regardless of agent state
                adapter = self.adapters.get(source.platform)
                if adapter and hasattr(adapter, 'get_pending_message'):
                    adapter.get_pending_message(_quick_key)  # consume and discard
                self._pending_messages.pop(_quick_key, None)
                if _quick_key in self._running_agents:
                    del self._running_agents[_quick_key]
                # Mark session suspended so the next message starts fresh
                # instead of resuming the stuck context (#7536).
                self.session_store.suspend_session(_quick_key)
                logger.info("HARD STOP for session %s — suspended, session lock released", _quick_key[:20])
                return "⚡ Force-stopped. The session is suspended — your next message will start fresh."

            # /reset and /new must bypass the running-agent guard so they
            # actually dispatch as commands instead of being queued as user
            # text (which would be fed back to the agent with the same
            # broken history — #2170).  Interrupt the agent first, then
            # clear the adapter's pending queue so the stale "/reset" text
            # doesn't get re-processed as a user message after the
            # interrupt completes.
            if _cmd_def_inner and _cmd_def_inner.name == "new":
                running_agent = self._running_agents.get(_quick_key)
                if running_agent and running_agent is not _AGENT_PENDING_SENTINEL:
                    running_agent.interrupt("Session reset requested")
                # Clear any pending messages so the old text doesn't replay
                adapter = self.adapters.get(source.platform)
                if adapter and hasattr(adapter, 'get_pending_message'):
                    adapter.get_pending_message(_quick_key)  # consume and discard
                self._pending_messages.pop(_quick_key, None)
                # Clean up the running agent entry so the reset handler
                # doesn't think an agent is still active.
                if _quick_key in self._running_agents:
                    del self._running_agents[_quick_key]
                return await self._handle_reset_command(event)

            # /queue <prompt> — queue without interrupting
            if event.get_command() in ("queue", "q"):
                queued_text = event.get_command_args().strip()
                if not queued_text:
                    return "Usage: /queue <prompt>"
                adapter = self.adapters.get(source.platform)
                if adapter:
                    from gateway.platforms.base import MessageEvent as _ME, MessageType as _MT
                    queued_event = _ME(
                        text=queued_text,
                        message_type=_MT.TEXT,
                        source=event.source,
                        message_id=event.message_id,
                    )
                    adapter._pending_messages[_quick_key] = queued_event
                return "Queued for the next turn."

            # /model must not be used while the agent is running.
            if _cmd_def_inner and _cmd_def_inner.name == "model":
                return "Agent is running — wait or /stop first, then switch models."

            # /approve and /deny must bypass the running-agent interrupt path.
            # The agent thread is blocked on a threading.Event inside
            # tools/approval.py — sending an interrupt won't unblock it.
            # Route directly to the approval handler so the event is signalled.
            if _cmd_def_inner and _cmd_def_inner.name in ("approve", "deny"):
                if _cmd_def_inner.name == "approve":
                    return await self._handle_approve_command(event)
                return await self._handle_deny_command(event)

            # /background must bypass the running-agent guard — it starts a
            # parallel task and must never interrupt the active conversation.
            if _cmd_def_inner and _cmd_def_inner.name == "background":
                return await self._handle_background_command(event)

            if event.message_type == MessageType.PHOTO:
                logger.debug("PRIORITY photo follow-up for session %s — queueing without interrupt", _quick_key[:20])
                adapter = self.adapters.get(source.platform)
                if adapter:
                    merge_pending_message_event(adapter._pending_messages, _quick_key, event)
                return None

            running_agent = self._running_agents.get(_quick_key)
            if running_agent is _AGENT_PENDING_SENTINEL:
                # Agent is being set up but not ready yet.
                if event.get_command() == "stop":
                    # Force-clean the sentinel so the session is unlocked.
                    if _quick_key in self._running_agents:
                        del self._running_agents[_quick_key]
                    logger.info("HARD STOP (pending) for session %s — sentinel cleared", _quick_key[:20])
                    return "⚡ Force-stopped. The agent was still starting — session unlocked."
                # Queue the message so it will be picked up after the
                # agent starts.
                adapter = self.adapters.get(source.platform)
                if adapter:
                    adapter._pending_messages[_quick_key] = event
                return None
            if self._draining:
                if self._queue_during_drain_enabled():
                    self._queue_or_replace_pending_event(_quick_key, event)
                return (
                    f"⏳ Gateway {self._status_action_gerund()} — queued for the next turn after it comes back."
                    if self._queue_during_drain_enabled()
                    else f"⏳ Gateway is {self._status_action_gerund()} and is not accepting another turn right now."
                )
            logger.debug("PRIORITY interrupt for session %s", _quick_key[:20])
            running_agent.interrupt(event.text)
            if _quick_key in self._pending_messages:
                self._pending_messages[_quick_key] += "\n" + event.text
            else:
                self._pending_messages[_quick_key] = event.text
            return None

        # Check for commands
        command = event.get_command()
        
        # Emit command:* hook for any recognized slash command.
        # GATEWAY_KNOWN_COMMANDS is derived from the central COMMAND_REGISTRY
        # in hermes_cli/commands.py — no hardcoded set to maintain here.
        from hermes_cli.commands import GATEWAY_KNOWN_COMMANDS, resolve_command as _resolve_cmd
        if command and command in GATEWAY_KNOWN_COMMANDS:
            await self.hooks.emit(f"command:{command}", {
                "platform": source.platform.value if source.platform else "",
                "user_id": source.user_id,
                "command": command,
                "args": event.get_command_args().strip(),
            })

        # Resolve aliases to canonical name so dispatch only checks canonicals.
        _cmd_def = _resolve_cmd(command) if command else None
        canonical = _cmd_def.name if _cmd_def else command

        if canonical == "new":
            return await self._handle_reset_command(event)
        
        if canonical == "help":
            return await self._handle_help_command(event)

        if canonical == "commands":
            return await self._handle_commands_command(event)
        
        if canonical == "profile":
            return await self._handle_profile_command(event)

        if canonical == "status":
            return await self._handle_status_command(event)

        if canonical == "restart":
            return await self._handle_restart_command(event)
        
        if canonical == "stop":
            return await self._handle_stop_command(event)
        
        if canonical == "reasoning":
            return await self._handle_reasoning_command(event)

        if canonical == "fast":
            return await self._handle_fast_command(event)

        if canonical == "verbose":
            return await self._handle_verbose_command(event)

        if canonical == "yolo":
            return await self._handle_yolo_command(event)

        if canonical == "model":
            return await self._handle_model_command(event)

        if canonical == "provider":
            return await self._handle_provider_command(event)
        
        if canonical == "personality":
            return await self._handle_personality_command(event)

        if canonical == "plan":
            try:
                from agent.skill_commands import build_plan_path, build_skill_invocation_message

                user_instruction = event.get_command_args().strip()
                plan_path = build_plan_path(user_instruction)
                event.text = build_skill_invocation_message(
                    "/plan",
                    user_instruction,
                    task_id=_quick_key,
                    runtime_note=(
                        "Save the markdown plan with write_file to this exact relative path "
                        f"inside the active workspace/backend cwd: {plan_path}"
                    ),
                )
                if not event.text:
                    return "Failed to load the bundled /plan skill."
                canonical = None
            except Exception as e:
                logger.exception("Failed to prepare /plan command")
                return f"Failed to enter plan mode: {e}"
        
        if canonical == "retry":
            return await self._handle_retry_command(event)
        
        if canonical == "undo":
            return await self._handle_undo_command(event)
        
        if canonical == "sethome":
            return await self._handle_set_home_command(event)

        if canonical == "compress":
            return await self._handle_compress_command(event)

        if canonical == "usage":
            return await self._handle_usage_command(event)

        if canonical == "insights":
            return await self._handle_insights_command(event)

        if canonical == "reload-mcp":
            return await self._handle_reload_mcp_command(event)

        if canonical == "approve":
            return await self._handle_approve_command(event)

        if canonical == "deny":
            return await self._handle_deny_command(event)

        if canonical == "update":
            return await self._handle_update_command(event)

        if canonical == "title":
            return await self._handle_title_command(event)

        if canonical == "resume":
            return await self._handle_resume_command(event)

        if canonical == "branch":
            return await self._handle_branch_command(event)

        if canonical == "rollback":
            return await self._handle_rollback_command(event)

        if canonical == "background":
            return await self._handle_background_command(event)

        if canonical == "btw":
            return await self._handle_btw_command(event)

        if canonical == "voice":
            return await self._handle_voice_command(event)

        if self._draining:
            return f"⏳ Gateway is {self._status_action_gerund()} and is not accepting new work right now."

        # User-defined quick commands (bypass agent loop, no LLM call)
        if command:
            if isinstance(self.config, dict):
                quick_commands = self.config.get("quick_commands", {}) or {}
            else:
                quick_commands = getattr(self.config, "quick_commands", {}) or {}
            if not isinstance(quick_commands, dict):
                quick_commands = {}
            if command in quick_commands:
                qcmd = quick_commands[command]
                if qcmd.get("type") == "exec":
                    exec_cmd = qcmd.get("command", "")
                    if exec_cmd:
                        try:
                            proc = await asyncio.create_subprocess_shell(
                                exec_cmd,
                                stdout=asyncio.subprocess.PIPE,
                                stderr=asyncio.subprocess.PIPE,
                            )
                            stdout, stderr = await asyncio.wait_for(proc.communicate(), timeout=30)
                            output = (stdout or stderr).decode().strip()
                            return output if output else "Command returned no output."
                        except asyncio.TimeoutError:
                            return "Quick command timed out (30s)."
                        except Exception as e:
                            return f"Quick command error: {e}"
                    else:
                        return f"Quick command '/{command}' has no command defined."
                elif qcmd.get("type") == "alias":
                    target = qcmd.get("target", "").strip()
                    if target:
                        target = target if target.startswith("/") else f"/{target}"
                        target_command = target.lstrip("/")
                        user_args = event.get_command_args().strip()
                        event.text = f"{target} {user_args}".strip()
                        command = target_command
                        # Fall through to normal command dispatch below
                    else:
                        return f"Quick command '/{command}' has no target defined."
                else:
                    return f"Quick command '/{command}' has unsupported type (supported: 'exec', 'alias')."

        # Plugin-registered slash commands
        if command:
            try:
                from hermes_cli.plugins import get_plugin_command_handler
                # Normalize underscores to hyphens so Telegram's underscored
                # autocomplete form matches plugin commands registered with
                # hyphens. See hermes_cli/commands.py:_build_telegram_menu.
                plugin_handler = get_plugin_command_handler(command.replace("_", "-"))
                if plugin_handler:
                    user_args = event.get_command_args().strip()
                    import asyncio as _aio
                    result = plugin_handler(user_args)
                    if _aio.iscoroutine(result):
                        result = await result
                    return str(result) if result else None
            except Exception as e:
                logger.debug("Plugin command dispatch failed (non-fatal): %s", e)

        # Skill slash commands: /skill-name loads the skill and sends to agent.
        # resolve_skill_command_key() handles the Telegram underscore/hyphen
        # round-trip so /claude_code from Telegram autocomplete still resolves
        # to the claude-code skill.
        if command:
            try:
                from agent.skill_commands import (
                    get_skill_commands,
                    build_skill_invocation_message,
                    resolve_skill_command_key,
                )
                skill_cmds = get_skill_commands()
                cmd_key = resolve_skill_command_key(command)
                if cmd_key is not None:
                    # Check per-platform disabled status before executing.
                    # get_skill_commands() only applies the *global* disabled
                    # list at scan time; per-platform overrides need checking
                    # here because the cache is process-global across platforms.
                    _skill_name = skill_cmds[cmd_key].get("name", "")
                    _plat = source.platform.value if source.platform else None
                    if _plat and _skill_name:
                        from agent.skill_utils import get_disabled_skill_names as _get_plat_disabled
                        if _skill_name in _get_plat_disabled(platform=_plat):
                            return (
                                f"The **{_skill_name}** skill is disabled for {_plat}.\n"
                                f"Enable it with: `hermes skills config`"
                            )
                    user_instruction = event.get_command_args().strip()
                    msg = build_skill_invocation_message(
                        cmd_key, user_instruction, task_id=_quick_key
                    )
                    if msg:
                        event.text = msg
                        # Fall through to normal message processing with skill content
                else:
                    # Not an active skill — check if it's a known-but-disabled or
                    # uninstalled skill and give actionable guidance.
                    _unavail_msg = _check_unavailable_skill(command)
                    if _unavail_msg:
                        return _unavail_msg
                    # Genuinely unrecognized /command: not a built-in, not a
                    # plugin, not a skill, not a known-inactive skill. Warn
                    # the user instead of silently forwarding it to the LLM
                    # as free text (which leads to silent-failure behavior
                    # like the model inventing a delegate_task call).
                    # Normalize to hyphenated form before checking known
                    # built-ins (command may be an alias target set by the
                    # quick-command block above, so _cmd_def can be stale).
                    if command.replace("_", "-") not in GATEWAY_KNOWN_COMMANDS:
                        logger.warning(
                            "Unrecognized slash command /%s from %s — "
                            "replying with unknown-command notice",
                            command,
                            source.platform.value if source.platform else "?",
                        )
                        return (
                            f"Unknown command `/{command}`. "
                            f"Type /commands to see what's available, "
                            f"or resend without the leading slash to send "
                            f"as a regular message."
                        )
            except Exception as e:
                logger.debug("Skill command check failed (non-fatal): %s", e)
        
        # Pending exec approvals are handled by /approve and /deny commands above.
        # No bare text matching — "yes" in normal conversation must not trigger
        # execution of a dangerous command.

        # ── Claim this session before any await ───────────────────────
        # Between here and _run_agent registering the real AIAgent, there
        # are numerous await points (hooks, vision enrichment, STT,
        # session hygiene compression).  Without this sentinel a second
        # message arriving during any of those yields would pass the
        # "already running" guard and spin up a duplicate agent for the
        # same session — corrupting the transcript.
        self._running_agents[_quick_key] = _AGENT_PENDING_SENTINEL
        self._running_agents_ts[_quick_key] = time.time()

        try:
            return await self._handle_message_with_agent(event, source, _quick_key)
        finally:
            # If _run_agent replaced the sentinel with a real agent and
            # then cleaned it up, this is a no-op.  If we exited early
            # (exception, command fallthrough, etc.) the sentinel must
            # not linger or the session would be permanently locked out.
            if self._running_agents.get(_quick_key) is _AGENT_PENDING_SENTINEL:
                del self._running_agents[_quick_key]
            self._running_agents_ts.pop(_quick_key, None)

    async def _prepare_inbound_message_text(
        self,
        *,
        event: MessageEvent,
        source: SessionSource,
        history: List[Dict[str, Any]],
    ) -> Optional[str]:
        """Prepare inbound event text for the agent.

        Keep the normal inbound path and the queued follow-up path on the same
        preprocessing pipeline so sender attribution, image enrichment, STT,
        document notes, reply context, and @ references all behave the same.
        """
        history = history or []
        message_text = event.text or ""

        _is_shared_thread = (
            source.chat_type != "dm"
            and source.thread_id
            and not getattr(self.config, "thread_sessions_per_user", False)
        )
        if _is_shared_thread and source.user_name:
            message_text = f"[{source.user_name}] {message_text}"

        if event.media_urls:
            image_paths = []
            audio_paths = []
            for i, path in enumerate(event.media_urls):
                mtype = event.media_types[i] if i < len(event.media_types) else ""
                if mtype.startswith("image/") or event.message_type == MessageType.PHOTO:
                    image_paths.append(path)
                if mtype.startswith("audio/") or event.message_type in (MessageType.VOICE, MessageType.AUDIO):
                    audio_paths.append(path)

            if image_paths:
                message_text = await self._enrich_message_with_vision(
                    message_text,
                    image_paths,
                )

            if audio_paths:
                message_text = await self._enrich_message_with_transcription(
                    message_text,
                    audio_paths,
                )
                _stt_fail_markers = (
                    "No STT provider",
                    "STT is disabled",
                    "can't listen",
                    "VOICE_TOOLS_OPENAI_KEY",
                )
                if any(marker in message_text for marker in _stt_fail_markers):
                    _stt_adapter = self.adapters.get(source.platform)
                    _stt_meta = {"thread_id": source.thread_id} if source.thread_id else None
                    if _stt_adapter:
                        try:
                            _stt_msg = (
                                "🎤 I received your voice message but can't transcribe it — "
                                "no speech-to-text provider is configured.\n\n"
                                "To enable voice: install faster-whisper "
                                "(`pip install faster-whisper` in the Hermes venv) "
                                "and set `stt.enabled: true` in config.yaml, "
                                "then /restart the gateway."
                            )
                            if self._has_setup_skill():
                                _stt_msg += "\n\nFor full setup instructions, type: `/skill hermes-agent-setup`"
                            await _stt_adapter.send(
                                source.chat_id,
                                _stt_msg,
                                metadata=_stt_meta,
                            )
                        except Exception:
                            pass

        if event.media_urls and event.message_type == MessageType.DOCUMENT:
            import mimetypes as _mimetypes

            _TEXT_EXTENSIONS = {".txt", ".md", ".csv", ".log", ".json", ".xml", ".yaml", ".yml", ".toml", ".ini", ".cfg"}
            for i, path in enumerate(event.media_urls):
                mtype = event.media_types[i] if i < len(event.media_types) else ""
                if mtype in ("", "application/octet-stream"):
                    import os as _os2

                    _ext = _os2.path.splitext(path)[1].lower()
                    if _ext in _TEXT_EXTENSIONS:
                        mtype = "text/plain"
                    else:
                        guessed, _ = _mimetypes.guess_type(path)
                        if guessed:
                            mtype = guessed
                if not mtype.startswith(("application/", "text/")):
                    continue

                import os as _os
                import re as _re

                basename = _os.path.basename(path)
                parts = basename.split("_", 2)
                display_name = parts[2] if len(parts) >= 3 else basename
                display_name = _re.sub(r'[^\w.\- ]', '_', display_name)

                if mtype.startswith("text/"):
                    context_note = (
                        f"[The user sent a text document: '{display_name}'. "
                        f"Its content has been included below. "
                        f"The file is also saved at: {path}]"
                    )
                else:
                    context_note = (
                        f"[The user sent a document: '{display_name}'. "
                        f"The file is saved at: {path}. "
                        f"Ask the user what they'd like you to do with it.]"
                    )
                message_text = f"{context_note}\n\n{message_text}"

        if getattr(event, "reply_to_text", None) and event.reply_to_message_id:
            reply_snippet = event.reply_to_text[:500]
            found_in_history = any(
                reply_snippet[:200] in (msg.get("content") or "")
                for msg in history
                if msg.get("role") in ("assistant", "user", "tool")
            )
            if not found_in_history:
                message_text = f'[Replying to: "{reply_snippet}"]\n\n{message_text}'

        if "@" in message_text:
            try:
                from agent.context_references import preprocess_context_references_async
                from agent.model_metadata import get_model_context_length

                _msg_cwd = os.environ.get("MESSAGING_CWD", os.path.expanduser("~"))
                _msg_ctx_len = get_model_context_length(
                    self._model,
                    base_url=self._base_url or "",
                )
                _ctx_result = await preprocess_context_references_async(
                    message_text,
                    cwd=_msg_cwd,
                    context_length=_msg_ctx_len,
                    allowed_root=_msg_cwd,
                )
                if _ctx_result.blocked:
                    _adapter = self.adapters.get(source.platform)
                    if _adapter:
                        await _adapter.send(
                            source.chat_id,
                            "\n".join(_ctx_result.warnings) or "Context injection refused.",
                        )
                    return None
                if _ctx_result.expanded:
                    message_text = _ctx_result.message
            except Exception as exc:
                logger.debug("@ context reference expansion failed: %s", exc)

        return message_text

    async def _handle_message_with_agent(self, event, source, _quick_key: str):
        """Inner handler that runs under the _running_agents sentinel guard."""
        _msg_start_time = time.time()
        _platform_name = source.platform.value if hasattr(source.platform, "value") else str(source.platform)
        _msg_preview = (event.text or "")[:80].replace("\n", " ")
        logger.info(
            "inbound message: platform=%s user=%s chat=%s msg=%r",
            _platform_name, source.user_name or source.user_id or "unknown",
            source.chat_id or "unknown", _msg_preview,
        )

        # Get or create session
        session_entry = self.session_store.get_or_create_session(source)
        session_key = session_entry.session_key
        
        # Emit session:start for new or auto-reset sessions
        _is_new_session = (
            session_entry.created_at == session_entry.updated_at
            or getattr(session_entry, "was_auto_reset", False)
        )
        if _is_new_session:
            await self.hooks.emit("session:start", {
                "platform": source.platform.value if source.platform else "",
                "user_id": source.user_id,
                "session_id": session_entry.session_id,
                "session_key": session_key,
            })
        
        # Build session context
        context = build_session_context(source, self.config, session_entry)
        
        # Set session context variables for tools (task-local, concurrency-safe)
        _session_env_tokens = self._set_session_env(context)
        
        # Read privacy.redact_pii from config (re-read per message)
        _redact_pii = False
        try:
            import yaml as _pii_yaml
            with open(_config_path, encoding="utf-8") as _pf:
                _pcfg = _pii_yaml.safe_load(_pf) or {}
            _redact_pii = bool((_pcfg.get("privacy") or {}).get("redact_pii", False))
        except Exception:
            pass

        # Build the context prompt to inject
        context_prompt = build_session_context_prompt(context, redact_pii=_redact_pii)
        
        # If the previous session expired and was auto-reset, prepend a notice
        # so the agent knows this is a fresh conversation (not an intentional /reset).
        if getattr(session_entry, 'was_auto_reset', False):
            reset_reason = getattr(session_entry, 'auto_reset_reason', None) or 'idle'
            if reset_reason == "suspended":
                context_note = "[System note: The user's previous session was stopped and suspended. This is a fresh conversation with no prior context.]"
            elif reset_reason == "daily":
                context_note = "[System note: The user's session was automatically reset by the daily schedule. This is a fresh conversation with no prior context.]"
            else:
                context_note = "[System note: The user's previous session expired due to inactivity. This is a fresh conversation with no prior context.]"
            context_prompt = context_note + "\n\n" + context_prompt

            # Send a user-facing notification explaining the reset, unless:
            # - notifications are disabled in config
            # - the platform is excluded (e.g. api_server, webhook)
            # - the expired session had no activity (nothing was cleared)
            try:
                policy = self.session_store.config.get_reset_policy(
                    platform=source.platform,
                    session_type=getattr(source, 'chat_type', 'dm'),
                )
                platform_name = source.platform.value if source.platform else ""
                had_activity = getattr(session_entry, 'reset_had_activity', False)
                # Suspended sessions always notify (they were explicitly stopped
                # or crashed mid-operation) — skip the policy check.
                should_notify = reset_reason == "suspended" or (
                    policy.notify
                    and had_activity
                    and platform_name not in policy.notify_exclude_platforms
                )
                if should_notify:
                    adapter = self.adapters.get(source.platform)
                    if adapter:
                        if reset_reason == "suspended":
                            reason_text = "previous session was stopped or interrupted"
                        elif reset_reason == "daily":
                            reason_text = f"daily schedule at {policy.at_hour}:00"
                        else:
                            hours = policy.idle_minutes // 60
                            mins = policy.idle_minutes % 60
                            duration = f"{hours}h" if not mins else f"{hours}h {mins}m" if hours else f"{mins}m"
                            reason_text = f"inactive for {duration}"
                        notice = (
                            f"◐ Session automatically reset ({reason_text}). "
                            f"Conversation history cleared.\n"
                            f"Use /resume to browse and restore a previous session.\n"
                            f"Adjust reset timing in config.yaml under session_reset."
                        )
                        try:
                            session_info = self._format_session_info()
                            if session_info:
                                notice = f"{notice}\n\n{session_info}"
                        except Exception:
                            pass
                        await adapter.send(
                            source.chat_id, notice,
                            metadata=getattr(event, 'metadata', None),
                        )
            except Exception as e:
                logger.debug("Auto-reset notification failed (non-fatal): %s", e)

            session_entry.was_auto_reset = False
            session_entry.auto_reset_reason = None

        # Auto-load skill(s) for topic/channel bindings (Telegram DM Topics,
        # Discord channel_skill_bindings).  Supports a single name or ordered list.
        # Only inject on NEW sessions — ongoing conversations already have the
        # skill content in their conversation history from the first message.
        _auto = getattr(event, "auto_skill", None)
        if _is_new_session and _auto:
            _skill_names = [_auto] if isinstance(_auto, str) else list(_auto)
            try:
                from agent.skill_commands import _load_skill_payload, _build_skill_message
                _combined_parts: list[str] = []
                _loaded_names: list[str] = []
                for _sname in _skill_names:
                    _loaded = _load_skill_payload(_sname, task_id=_quick_key)
                    if _loaded:
                        _loaded_skill, _skill_dir, _display_name = _loaded
                        _note = (
                            f'[SYSTEM: The "{_display_name}" skill is auto-loaded. '
                            f"Follow its instructions for this session.]"
                        )
                        _part = _build_skill_message(_loaded_skill, _skill_dir, _note)
                        if _part:
                            _combined_parts.append(_part)
                            _loaded_names.append(_sname)
                    else:
                        logger.warning("[Gateway] Auto-skill '%s' not found", _sname)
                if _combined_parts:
                    # Append the user's original text after all skill payloads
                    _combined_parts.append(event.text)
                    event.text = "\n\n".join(_combined_parts)
                    logger.info(
                        "[Gateway] Auto-loaded skill(s) %s for session %s",
                        _loaded_names, session_key,
                    )
            except Exception as e:
                logger.warning("[Gateway] Failed to auto-load skill(s) %s: %s", _skill_names, e)

        # Load conversation history from transcript
        history = self.session_store.load_transcript(session_entry.session_id)
        
        # -----------------------------------------------------------------
        # Session hygiene: auto-compress pathologically large transcripts
        #
        # Long-lived gateway sessions can accumulate enough history that
        # every new message rehydrates an oversized transcript, causing
        # repeated truncation/context failures.  Detect this early and
        # compress proactively — before the agent even starts.  (#628)
        #
        # Token source priority:
        # 1. Actual API-reported prompt_tokens from the last turn
        #    (stored in session_entry.last_prompt_tokens)
        # 2. Rough char-based estimate (str(msg)//4). Overestimates
        #    by 30-50% on code/JSON-heavy sessions, but that just
        #    means hygiene fires a bit early — safe and harmless.
        # -----------------------------------------------------------------
        if history and len(history) >= 4:
            from agent.model_metadata import (
                estimate_messages_tokens_rough,
                get_model_context_length,
            )

            # Read model + compression config from config.yaml.
            # NOTE: hygiene threshold is intentionally HIGHER than the agent's
            # own compressor (0.85 vs 0.50).  Hygiene is a safety net for
            # sessions that grew too large between turns — it fires pre-agent
            # to prevent API failures.  The agent's own compressor handles
            # normal context management during its tool loop with accurate
            # real token counts.  Having hygiene at 0.50 caused premature
            # compression on every turn in long gateway sessions.
            _hyg_model = "anthropic/claude-sonnet-4.6"
            _hyg_threshold_pct = 0.85
            _hyg_compression_enabled = True
            _hyg_config_context_length = None
            _hyg_provider = None
            _hyg_base_url = None
            _hyg_api_key = None
            _hyg_data = {}
            try:
                _hyg_cfg_path = _hermes_home / "config.yaml"
                if _hyg_cfg_path.exists():
                    import yaml as _hyg_yaml
                    with open(_hyg_cfg_path, encoding="utf-8") as _hyg_f:
                        _hyg_data = _hyg_yaml.safe_load(_hyg_f) or {}

                    # Resolve model name (same logic as run_sync)
                    _model_cfg = _hyg_data.get("model", {})
                    if isinstance(_model_cfg, str):
                        _hyg_model = _model_cfg
                    elif isinstance(_model_cfg, dict):
                        _hyg_model = _model_cfg.get("default") or _model_cfg.get("model") or _hyg_model
                        # Read explicit context_length override from model config
                        # (same as run_agent.py lines 995-1005)
                        _raw_ctx = _model_cfg.get("context_length")
                        if _raw_ctx is not None:
                            try:
                                _hyg_config_context_length = int(_raw_ctx)
                            except (TypeError, ValueError):
                                pass
                        # Read provider for accurate context detection
                        _hyg_provider = _model_cfg.get("provider") or None
                        _hyg_base_url = _model_cfg.get("base_url") or None

                    # Read compression settings — only use enabled flag.
                    # The threshold is intentionally separate from the agent's
                    # compression.threshold (hygiene runs higher).
                    _comp_cfg = _hyg_data.get("compression", {})
                    if isinstance(_comp_cfg, dict):
                        _hyg_compression_enabled = str(
                            _comp_cfg.get("enabled", True)
                        ).lower() in ("true", "1", "yes")

                try:
                    _hyg_model, _hyg_runtime = self._resolve_session_agent_runtime(
                        source=source,
                        session_key=session_key,
                        user_config=_hyg_data if isinstance(_hyg_data, dict) else None,
                    )
                    _hyg_provider = _hyg_runtime.get("provider") or _hyg_provider
                    _hyg_base_url = _hyg_runtime.get("base_url") or _hyg_base_url
                    _hyg_api_key = _hyg_runtime.get("api_key") or _hyg_api_key
                except Exception:
                    pass

                # Check custom_providers per-model context_length
                # (same fallback as run_agent.py lines 1171-1189).
                # Must run after runtime resolution so _hyg_base_url is set.
                if _hyg_config_context_length is None and _hyg_base_url:
                    try:
                        _hyg_custom_providers = _hyg_data.get("custom_providers")
                        if isinstance(_hyg_custom_providers, list):
                            for _cp in _hyg_custom_providers:
                                if not isinstance(_cp, dict):
                                    continue
                                _cp_url = (_cp.get("base_url") or "").rstrip("/")
                                if _cp_url and _cp_url == _hyg_base_url.rstrip("/"):
                                    _cp_models = _cp.get("models", {})
                                    if isinstance(_cp_models, dict):
                                        _cp_model_cfg = _cp_models.get(_hyg_model, {})
                                        if isinstance(_cp_model_cfg, dict):
                                            _cp_ctx = _cp_model_cfg.get("context_length")
                                            if _cp_ctx is not None:
                                                _hyg_config_context_length = int(_cp_ctx)
                                    break
                    except (TypeError, ValueError):
                        pass
            except Exception:
                pass

            if _hyg_compression_enabled:
                _hyg_context_length = get_model_context_length(
                    _hyg_model,
                    base_url=_hyg_base_url or "",
                    api_key=_hyg_api_key or "",
                    config_context_length=_hyg_config_context_length,
                    provider=_hyg_provider or "",
                )
                _compress_token_threshold = int(
                    _hyg_context_length * _hyg_threshold_pct
                )
                _warn_token_threshold = int(_hyg_context_length * 0.95)

                _msg_count = len(history)

                # Prefer actual API-reported tokens from the last turn
                # (stored in session entry) over the rough char-based estimate.
                _stored_tokens = session_entry.last_prompt_tokens
                if _stored_tokens > 0:
                    _approx_tokens = _stored_tokens
                    _token_source = "actual"
                else:
                    _approx_tokens = estimate_messages_tokens_rough(history)
                    _token_source = "estimated"
                    # Note: rough estimates overestimate by 30-50% for code/JSON-heavy
                    # sessions, but that just means hygiene fires a bit early — which
                    # is safe and harmless.  The 85% threshold already provides ample
                    # headroom (agent's own compressor runs at 50%).  A previous 1.4x
                    # multiplier tried to compensate by inflating the threshold, but
                    # 85% * 1.4 = 119% of context — which exceeds the model's limit
                    # and prevented hygiene from ever firing for ~200K models (GLM-5).

                # Hard safety valve: force compression if message count is
                # extreme, regardless of token estimates.  This breaks the
                # death spiral where API disconnects prevent token data
                # collection, which prevents compression, which causes more
                # disconnects.  400 messages is well above normal sessions
                # but catches runaway growth before it becomes unrecoverable.
                # (#2153)
                _HARD_MSG_LIMIT = 400
                _needs_compress = (
                    _approx_tokens >= _compress_token_threshold
                    or _msg_count >= _HARD_MSG_LIMIT
                )

                if _needs_compress:
                    logger.info(
                        "Session hygiene: %s messages, ~%s tokens (%s) — auto-compressing "
                        "(threshold: %s%% of %s = %s tokens)",
                        _msg_count, f"{_approx_tokens:,}", _token_source,
                        int(_hyg_threshold_pct * 100),
                        f"{_hyg_context_length:,}",
                        f"{_compress_token_threshold:,}",
                    )

                    _hyg_meta = {"thread_id": source.thread_id} if source.thread_id else None

                    try:
                        from run_agent import AIAgent

                        _hyg_model, _hyg_runtime = self._resolve_session_agent_runtime(
                            source=source,
                            session_key=session_key,
                            user_config=_hyg_data if isinstance(_hyg_data, dict) else None,
                        )
                        if _hyg_runtime.get("api_key"):
                            _hyg_msgs = [
                                {"role": m.get("role"), "content": m.get("content")}
                                for m in history
                                if m.get("role") in ("user", "assistant")
                                and m.get("content")
                            ]

                            if len(_hyg_msgs) >= 4:
                                _hyg_agent = AIAgent(
                                    **_hyg_runtime,
                                    model=_hyg_model,
                                    max_iterations=4,
                                    quiet_mode=True,
                                    enabled_toolsets=["memory"],
                                    session_id=session_entry.session_id,
                                )
                                _hyg_agent._print_fn = lambda *a, **kw: None

                                loop = asyncio.get_event_loop()
                                _compressed, _ = await loop.run_in_executor(
                                    None,
                                    lambda: _hyg_agent._compress_context(
                                        _hyg_msgs, "",
                                        approx_tokens=_approx_tokens,
                                    ),
                                )

                                # _compress_context ends the old session and creates
                                # a new session_id.  Write compressed messages into
                                # the NEW session so the old transcript stays intact
                                # and searchable via session_search.
                                _hyg_new_sid = _hyg_agent.session_id
                                if _hyg_new_sid != session_entry.session_id:
                                    session_entry.session_id = _hyg_new_sid
                                    self.session_store._save()

                                self.session_store.rewrite_transcript(
                                    session_entry.session_id, _compressed
                                )
                                # Reset stored token count — transcript was rewritten
                                session_entry.last_prompt_tokens = 0
                                history = _compressed
                                _new_count = len(_compressed)
                                _new_tokens = estimate_messages_tokens_rough(
                                    _compressed
                                )

                                logger.info(
                                    "Session hygiene: compressed %s → %s msgs, "
                                    "~%s → ~%s tokens",
                                    _msg_count, _new_count,
                                    f"{_approx_tokens:,}", f"{_new_tokens:,}",
                                )

                                if _new_tokens >= _warn_token_threshold:
                                    logger.warning(
                                        "Session hygiene: still ~%s tokens after "
                                        "compression",
                                        f"{_new_tokens:,}",
                                    )

                    except Exception as e:
                        logger.warning(
                            "Session hygiene auto-compress failed: %s", e
                        )

        # First-message onboarding -- only on the very first interaction ever
        if not history and not self.session_store.has_any_sessions():
            context_prompt += (
                "\n\n[System note: This is the user's very first message ever. "
                "Briefly introduce yourself and mention that /help shows available commands. "
                "Keep the introduction concise -- one or two sentences max.]"
            )
        
        # One-time prompt if no home channel is set for this platform
        # Skip for webhooks - they deliver directly to configured targets (github_comment, etc.)
        if not history and source.platform and source.platform != Platform.LOCAL and source.platform != Platform.WEBHOOK:
            platform_name = source.platform.value
            env_key = f"{platform_name.upper()}_HOME_CHANNEL"
            if not os.getenv(env_key):
                adapter = self.adapters.get(source.platform)
                if adapter:
                    await adapter.send(
                        source.chat_id,
                        f"📬 No home channel is set for {platform_name.title()}. "
                        f"A home channel is where Hermes delivers cron job results "
                        f"and cross-platform messages.\n\n"
                        f"Type /sethome to make this chat your home channel, "
                        f"or ignore to skip."
                    )
        
        # -----------------------------------------------------------------
        # Voice channel awareness — inject current voice channel state
        # into context so the agent knows who is in the channel and who
        # is speaking, without needing a separate tool call.
        # -----------------------------------------------------------------
        if source.platform == Platform.DISCORD:
            adapter = self.adapters.get(Platform.DISCORD)
            guild_id = self._get_guild_id(event)
            if guild_id and adapter and hasattr(adapter, "get_voice_channel_context"):
                vc_context = adapter.get_voice_channel_context(guild_id)
                if vc_context:
                    context_prompt += f"\n\n{vc_context}"

        # -----------------------------------------------------------------
        # Auto-analyze images sent by the user
        #
        # If the user attached image(s), we run the vision tool eagerly so
        # the conversation model always receives a text description.  The
        # local file path is also included so the model can re-examine the
        # image later with a more targeted question via vision_analyze.
        #
        # We filter to image paths only (by media_type) so that non-image
        # attachments (documents, audio, etc.) are not sent to the vision
        # tool even when they appear in the same message.
        # -----------------------------------------------------------------
        message_text = await self._prepare_inbound_message_text(
            event=event,
            source=source,
            history=history,
        )
        if message_text is None:
            return

        try:
            # Emit agent:start hook
            hook_ctx = {
                "platform": source.platform.value if source.platform else "",
                "user_id": source.user_id,
                "session_id": session_entry.session_id,
                "message": message_text[:500],
            }
            await self.hooks.emit("agent:start", hook_ctx)

            # Run the agent
            agent_result = await self._run_agent(
                message=message_text,
                context_prompt=context_prompt,
                history=history,
                source=source,
                session_id=session_entry.session_id,
                session_key=session_key,
                event_message_id=event.message_id,
            )

            # Stop persistent typing indicator now that the agent is done
            try:
                _typing_adapter = self.adapters.get(source.platform)
                if _typing_adapter and hasattr(_typing_adapter, "stop_typing"):
                    await _typing_adapter.stop_typing(source.chat_id)
            except Exception:
                pass

            response = agent_result.get("final_response") or ""
            agent_messages = agent_result.get("messages", [])
            _response_time = time.time() - _msg_start_time
            _api_calls = agent_result.get("api_calls", 0)
            _resp_len = len(response)
            logger.info(
                "response ready: platform=%s chat=%s time=%.1fs api_calls=%d response=%d chars",
                _platform_name, source.chat_id or "unknown",
                _response_time, _api_calls, _resp_len,
            )

            # Surface error details when the agent failed silently (final_response=None)
            if not response and agent_result.get("failed"):
                error_detail = agent_result.get("error", "unknown error")
                error_str = str(error_detail).lower()

                # Detect context-overflow failures and give specific guidance.
                # Generic 400 "Error" from Anthropic with large sessions is the
                # most common cause of this (#1630).
                _is_ctx_fail = any(p in error_str for p in (
                    "context", "token", "too large", "too long",
                    "exceed", "payload",
                )) or (
                    "400" in error_str
                    and len(history) > 50
                )

                if _is_ctx_fail:
                    response = (
                        "⚠️ Session too large for the model's context window.\n"
                        "Use /compact to compress the conversation, or "
                        "/reset to start fresh."
                    )
                else:
                    response = (
                        f"The request failed: {str(error_detail)[:300]}\n"
                        "Try again or use /reset to start a fresh session."
                    )

            # If the agent's session_id changed during compression, update
            # session_entry so transcript writes below go to the right session.
            if agent_result.get("session_id") and agent_result["session_id"] != session_entry.session_id:
                session_entry.session_id = agent_result["session_id"]

            # Prepend reasoning/thinking if display is enabled (per-platform)
            try:
                from gateway.display_config import resolve_display_setting as _rds
                _show_reasoning_effective = _rds(
                    _load_gateway_config(),
                    _platform_config_key(source.platform),
                    "show_reasoning",
                    getattr(self, "_show_reasoning", False),
                )
            except Exception:
                _show_reasoning_effective = getattr(self, "_show_reasoning", False)
            if _show_reasoning_effective and response:
                last_reasoning = agent_result.get("last_reasoning")
                if last_reasoning:
                    # Collapse long reasoning to keep messages readable
                    lines = last_reasoning.strip().splitlines()
                    if len(lines) > 15:
                        display_reasoning = "\n".join(lines[:15])
                        display_reasoning += f"\n_... ({len(lines) - 15} more lines)_"
                    else:
                        display_reasoning = last_reasoning.strip()
                    response = f"💭 **Reasoning:**\n```\n{display_reasoning}\n```\n\n{response}"

            # Emit agent:end hook
            await self.hooks.emit("agent:end", {
                **hook_ctx,
                "response": (response or "")[:500],
            })
            
            # Check for pending process watchers (check_interval on background processes)
            try:
                from tools.process_registry import process_registry
                while process_registry.pending_watchers:
                    watcher = process_registry.pending_watchers.pop(0)
                    asyncio.create_task(self._run_process_watcher(watcher))
            except Exception as e:
                logger.error("Process watcher setup error: %s", e)

            # Drain watch pattern notifications that arrived during the agent run.
            # Watch events and completions share the same queue; completions are
            # already handled by the per-process watcher task above, so we only
            # inject watch-type events here.
            try:
                from tools.process_registry import process_registry as _pr
                _watch_events = []
                while not _pr.completion_queue.empty():
                    evt = _pr.completion_queue.get_nowait()
                    evt_type = evt.get("type", "completion")
                    if evt_type in ("watch_match", "watch_disabled"):
                        _watch_events.append(evt)
                    # else: completion events are handled by the watcher task
                for evt in _watch_events:
                    synth_text = _format_gateway_process_notification(evt)
                    if synth_text:
                        try:
                            await self._inject_watch_notification(synth_text, event)
                        except Exception as e2:
                            logger.error("Watch notification injection error: %s", e2)
            except Exception as e:
                logger.debug("Watch queue drain error: %s", e)

            # NOTE: Dangerous command approvals are now handled inline by the
            # blocking gateway approval mechanism in tools/approval.py.  The agent
            # thread blocks until the user responds with /approve or /deny, so by
            # the time we reach here the approval has already been resolved.  The
            # old post-loop pop_pending + approval_hint code was removed in favour
            # of the blocking approach that mirrors CLI's synchronous input().
            
            # Save the full conversation to the transcript, including tool calls.
            # This preserves the complete agent loop (tool_calls, tool results,
            # intermediate reasoning) so sessions can be resumed with full context
            # and transcripts are useful for debugging and training data.
            #
            # IMPORTANT: When the agent failed before producing any response
            # (e.g. context-overflow 400), do NOT persist the user's message.
            # Persisting it would make the session even larger, causing the
            # same failure on the next attempt — an infinite loop. (#1630)
            agent_failed_early = (
                agent_result.get("failed")
                and not agent_result.get("final_response")
            )
            if agent_failed_early:
                logger.info(
                    "Skipping transcript persistence for failed request in "
                    "session %s to prevent session growth loop.",
                    session_entry.session_id,
                )

            ts = datetime.now().isoformat()
            
            # If this is a fresh session (no history), write the full tool
            # definitions as the first entry so the transcript is self-describing
            # -- the same list of dicts sent as tools=[...] in the API request.
            if agent_failed_early:
                pass  # Skip all transcript writes — don't grow a broken session
            elif not history:
                tool_defs = agent_result.get("tools", [])
                self.session_store.append_to_transcript(
                    session_entry.session_id,
                    {
                        "role": "session_meta",
                        "tools": tool_defs or [],
                        "model": _resolve_gateway_model(),
                        "platform": source.platform.value if source.platform else "",
                        "timestamp": ts,
                    }
                )
            
            # Find only the NEW messages from this turn (skip history we loaded).
            # Use the filtered history length (history_offset) that was actually
            # passed to the agent, not len(history) which includes session_meta
            # entries that were stripped before the agent saw them.
            if not agent_failed_early:
                history_len = agent_result.get("history_offset", len(history))
                new_messages = agent_messages[history_len:] if len(agent_messages) > history_len else []
                
                # If no new messages found (edge case), fall back to simple user/assistant
                if not new_messages:
                    self.session_store.append_to_transcript(
                        session_entry.session_id,
                        {"role": "user", "content": message_text, "timestamp": ts}
                    )
                    if response:
                        self.session_store.append_to_transcript(
                            session_entry.session_id,
                            {"role": "assistant", "content": response, "timestamp": ts}
                        )
                else:
                    # The agent already persisted these messages to SQLite via
                    # _flush_messages_to_session_db(), so skip the DB write here
                    # to prevent the duplicate-write bug (#860).  We still write
                    # to JSONL for backward compatibility and as a backup.
                    agent_persisted = self._session_db is not None
                    for msg in new_messages:
                        # Skip system messages (they're rebuilt each run)
                        if msg.get("role") == "system":
                            continue
                        # Add timestamp to each message for debugging
                        entry = {**msg, "timestamp": ts}
                        self.session_store.append_to_transcript(
                            session_entry.session_id, entry,
                            skip_db=agent_persisted,
                        )
            
            # Token counts and model are now persisted by the agent directly.
            # Keep only last_prompt_tokens here for context-window tracking and
            # compression decisions.
            self.session_store.update_session(
                session_entry.session_key,
                last_prompt_tokens=agent_result.get("last_prompt_tokens", 0),
            )

            # Auto voice reply: send TTS audio before the text response
            _already_sent = bool(agent_result.get("already_sent"))
            if self._should_send_voice_reply(event, response, agent_messages, already_sent=_already_sent):
                await self._send_voice_reply(event, response)

            # If streaming already delivered the response, extract and
            # deliver any MEDIA: files before returning None.  Streaming
            # sends raw text chunks that include MEDIA: tags — the normal
            # post-processing in _process_message_background is skipped
            # when already_sent is True, so media files would never be
            # delivered without this.
            #
            # Never skip when the agent failed — the error message is new
            # content the user hasn't seen (streaming only sent earlier
            # partial output before the failure).  Without this guard,
            # users see the agent "stop responding without explanation."
            if agent_result.get("already_sent") and not agent_result.get("failed"):
                if response:
                    _media_adapter = self.adapters.get(source.platform)
                    if _media_adapter:
                        await self._deliver_media_from_response(
                            response, event, _media_adapter,
                        )
                return None

            return response
            
        except Exception as e:
            # Stop typing indicator on error too
            try:
                _err_adapter = self.adapters.get(source.platform)
                if _err_adapter and hasattr(_err_adapter, "stop_typing"):
                    await _err_adapter.stop_typing(source.chat_id)
            except Exception:
                pass
            logger.exception("Agent error in session %s", session_key)
            error_type = type(e).__name__
            error_detail = str(e)[:300] if str(e) else "no details available"
            status_hint = ""
            status_code = getattr(e, "status_code", None)
            _hist_len = len(history) if 'history' in locals() else 0
            if status_code == 401:
                status_hint = " Check your API key or run `claude /login` to refresh OAuth credentials."
            elif status_code == 429:
                # Check if this is a plan usage limit (resets on a schedule) vs a transient rate limit
                _err_body = getattr(e, "response", None)
                _err_json = {}
                try:
                    if _err_body is not None:
                        _err_json = _err_body.json().get("error", {})
                except Exception:
                    pass
                if _err_json.get("type") == "usage_limit_reached":
                    _resets_in = _err_json.get("resets_in_seconds")
                    if _resets_in and _resets_in > 0:
                        import math
                        _hours = math.ceil(_resets_in / 3600)
                        status_hint = f" Your plan's usage limit has been reached. It resets in ~{_hours}h."
                    else:
                        status_hint = " Your plan's usage limit has been reached. Please wait until it resets."
                else:
                    status_hint = " You are being rate-limited. Please wait a moment and try again."
            elif status_code == 529:
                status_hint = " The API is temporarily overloaded. Please try again shortly."
            elif status_code in (400, 500):
                # 400 with a large session is context overflow.
                # 500 with a large session often means the payload is too large
                # for the API to process — treat it the same way.
                if _hist_len > 50:
                    return (
                        "⚠️ Session too large for the model's context window.\n"
                        "Use /compact to compress the conversation, or "
                        "/reset to start fresh."
                    )
                elif status_code == 400:
                    status_hint = " The request was rejected by the API."
            return (
                f"Sorry, I encountered an error ({error_type}).\n"
                f"{error_detail}\n"
                f"{status_hint}"
                "Try again or use /reset to start a fresh session."
            )
        finally:
            # Restore session context variables to their pre-handler state
            self._clear_session_env(_session_env_tokens)
    
    def _format_session_info(self) -> str:
        """Resolve current model config and return a formatted info block.

        Surfaces model, provider, context length, and endpoint so gateway
        users can immediately see if context detection went wrong (e.g.
        local models falling to the 128K default).
        """
        from agent.model_metadata import get_model_context_length, DEFAULT_FALLBACK_CONTEXT

        model = _resolve_gateway_model()
        config_context_length = None
        provider = None
        base_url = None
        api_key = None

        try:
            cfg_path = _hermes_home / "config.yaml"
            if cfg_path.exists():
                import yaml as _info_yaml
                with open(cfg_path, encoding="utf-8") as f:
                    data = _info_yaml.safe_load(f) or {}
                model_cfg = data.get("model", {})
                if isinstance(model_cfg, dict):
                    raw_ctx = model_cfg.get("context_length")
                    if raw_ctx is not None:
                        try:
                            config_context_length = int(raw_ctx)
                        except (TypeError, ValueError):
                            pass
                    provider = model_cfg.get("provider") or None
                    base_url = model_cfg.get("base_url") or None
        except Exception:
            pass

        # Resolve runtime credentials for probing
        try:
            runtime = _resolve_runtime_agent_kwargs()
            provider = provider or runtime.get("provider")
            base_url = base_url or runtime.get("base_url")
            api_key = runtime.get("api_key")
        except Exception:
            pass

        context_length = get_model_context_length(
            model,
            base_url=base_url or "",
            api_key=api_key or "",
            config_context_length=config_context_length,
            provider=provider or "",
        )

        # Format context source hint
        if config_context_length is not None:
            ctx_source = "config"
        elif context_length == DEFAULT_FALLBACK_CONTEXT:
            ctx_source = "default — set model.context_length in config to override"
        else:
            ctx_source = "detected"

        # Format context length for display
        if context_length >= 1_000_000:
            ctx_display = f"{context_length / 1_000_000:.1f}M"
        elif context_length >= 1_000:
            ctx_display = f"{context_length // 1_000}K"
        else:
            ctx_display = str(context_length)

        lines = [
            f"◆ Model: `{model}`",
            f"◆ Provider: {provider or 'openrouter'}",
            f"◆ Context: {ctx_display} tokens ({ctx_source})",
        ]

        # Show endpoint for local/custom setups
        if base_url and ("localhost" in base_url or "127.0.0.1" in base_url or "0.0.0.0" in base_url):
            lines.append(f"◆ Endpoint: {base_url}")

        return "\n".join(lines)

    async def _handle_reset_command(self, event: MessageEvent) -> str:
        """Handle /new or /reset command."""
        source = event.source
        
        # Get existing session key
        session_key = self._session_key_for_source(source)
        
        # Flush memories in the background (fire-and-forget) so the user
        # gets the "Session reset!" response immediately.
        try:
            old_entry = self.session_store._entries.get(session_key)
            if old_entry:
                _flush_task = asyncio.create_task(
                    self._async_flush_memories(old_entry.session_id, session_key)
                )
                self._background_tasks.add(_flush_task)
                _flush_task.add_done_callback(self._background_tasks.discard)
        except Exception as e:
            logger.debug("Gateway memory flush on reset failed: %s", e)
        # Close tool resources on the old agent (terminal sandboxes, browser
        # daemons, background processes) before evicting from cache.
        # Guard with getattr because test fixtures may skip __init__.
        _cache_lock = getattr(self, "_agent_cache_lock", None)
        if _cache_lock is not None:
            with _cache_lock:
                _cached = self._agent_cache.get(session_key)
                _old_agent = _cached[0] if isinstance(_cached, tuple) else _cached if _cached else None
            if _old_agent is not None:
                try:
                    if hasattr(_old_agent, "close"):
                        _old_agent.close()
                except Exception:
                    pass
        self._evict_cached_agent(session_key)

        try:
            from tools.env_passthrough import clear_env_passthrough
            clear_env_passthrough()
        except Exception:
            pass

        try:
            from tools.credential_files import clear_credential_files
            clear_credential_files()
        except Exception:
            pass

        # Reset the session
        new_entry = self.session_store.reset_session(session_key)

        # Clear any session-scoped model override so the next agent picks up
        # the configured default instead of the previously switched model.
        self._session_model_overrides.pop(session_key, None)

        # Fire plugin on_session_finalize hook (session boundary)
        try:
            from hermes_cli.plugins import invoke_hook as _invoke_hook
            _old_sid = old_entry.session_id if old_entry else None
            _invoke_hook("on_session_finalize", session_id=_old_sid,
                         platform=source.platform.value if source.platform else "")
        except Exception:
            pass

        # Emit session:end hook (session is ending)
        await self.hooks.emit("session:end", {
            "platform": source.platform.value if source.platform else "",
            "user_id": source.user_id,
            "session_key": session_key,
        })

        # Emit session:reset hook
        await self.hooks.emit("session:reset", {
            "platform": source.platform.value if source.platform else "",
            "user_id": source.user_id,
            "session_key": session_key,
        })

        # Resolve session config info to surface to the user
        try:
            session_info = self._format_session_info()
        except Exception:
            session_info = ""

        if new_entry:
            header = "✨ Session reset! Starting fresh."
        else:
            # No existing session, just create one
            new_entry = self.session_store.get_or_create_session(source, force_new=True)
            header = "✨ New session started!"

        # Fire plugin on_session_reset hook (new session guaranteed to exist)
        try:
            from hermes_cli.plugins import invoke_hook as _invoke_hook
            _new_sid = new_entry.session_id if new_entry else None
            _invoke_hook("on_session_reset", session_id=_new_sid,
                         platform=source.platform.value if source.platform else "")
        except Exception:
            pass

        # Append a random tip to the reset message
        try:
            from hermes_cli.tips import get_random_tip
            _tip_line = f"\n✦ Tip: {get_random_tip()}"
        except Exception:
            _tip_line = ""

        if session_info:
            return f"{header}\n\n{session_info}{_tip_line}"
        return f"{header}{_tip_line}"
    
    async def _handle_profile_command(self, event: MessageEvent) -> str:
        """Handle /profile — show active profile name and home directory."""
        from hermes_constants import get_hermes_home, display_hermes_home
        from pathlib import Path

        home = get_hermes_home()
        display = display_hermes_home()

        # Detect profile name from HERMES_HOME path
        # Profile paths look like: ~/.hermes/profiles/<name>
        profiles_parent = Path.home() / ".hermes" / "profiles"
        try:
            rel = home.relative_to(profiles_parent)
            profile_name = str(rel).split("/")[0]
        except ValueError:
            profile_name = None

        if profile_name:
            lines = [
                f"👤 **Profile:** `{profile_name}`",
                f"📂 **Home:** `{display}`",
            ]
        else:
            lines = [
                "👤 **Profile:** default",
                f"📂 **Home:** `{display}`",
            ]

        return "\n".join(lines)

    async def _handle_status_command(self, event: MessageEvent) -> str:
        """Handle /status command."""
        source = event.source
        session_entry = self.session_store.get_or_create_session(source)

        connected_platforms = [p.value for p in self.adapters.keys()]

        # Check if there's an active agent
        session_key = session_entry.session_key
        is_running = session_key in self._running_agents

        title = None
        if self._session_db:
            try:
                title = self._session_db.get_session_title(session_entry.session_id)
            except Exception:
                title = None

        lines = [
            "📊 **Hermes Gateway Status**",
            "",
            f"**Session ID:** `{session_entry.session_id}`",
        ]
        if title:
            lines.append(f"**Title:** {title}")
        lines.extend([
            f"**Created:** {session_entry.created_at.strftime('%Y-%m-%d %H:%M')}",
            f"**Last Activity:** {session_entry.updated_at.strftime('%Y-%m-%d %H:%M')}",
            f"**Tokens:** {session_entry.total_tokens:,}",
            f"**Agent Running:** {'Yes ⚡' if is_running else 'No'}",
            "",
            f"**Connected Platforms:** {', '.join(connected_platforms)}",
        ])

        return "\n".join(lines)
    
    async def _handle_stop_command(self, event: MessageEvent) -> str:
        """Handle /stop command - interrupt a running agent.

        When an agent is truly hung (blocked thread that never checks
        _interrupt_requested), the early intercept in _handle_message()
        handles /stop before this method is reached.  This handler fires
        only through normal command dispatch (no running agent) or as a
        fallback.  Force-clean the session lock in all cases for safety.

        When there IS a running/pending agent, the session is also marked
        as *suspended* so the next message starts a fresh session instead
        of resuming the stuck context (#7536).
        """
        source = event.source
        session_entry = self.session_store.get_or_create_session(source)
        session_key = session_entry.session_key

        agent = self._running_agents.get(session_key)
        if agent is _AGENT_PENDING_SENTINEL:
            # Force-clean the sentinel so the session is unlocked.
            if session_key in self._running_agents:
                del self._running_agents[session_key]
            self.session_store.suspend_session(session_key)
            logger.info("HARD STOP (pending) for session %s — suspended, sentinel cleared", session_key[:20])
            return "⚡ Force-stopped. The agent was still starting — your next message will start fresh."
        if agent:
            agent.interrupt("Stop requested")
            # Force-clean the session lock so a truly hung agent doesn't
            # keep it locked forever.
            if session_key in self._running_agents:
                del self._running_agents[session_key]
            self.session_store.suspend_session(session_key)
            return "⚡ Force-stopped. Your next message will start a fresh session."
        else:
            return "No active task to stop."

    async def _handle_restart_command(self, event: MessageEvent) -> str:
        """Handle /restart command - drain active work, then restart the gateway."""
        if self._restart_requested or self._draining:
            count = self._running_agent_count()
            if count:
                return f"⏳ Draining {count} active agent(s) before restart..."
            return "⏳ Gateway restart already in progress..."

        active_agents = self._running_agent_count()
        self.request_restart(detached=True, via_service=False)
        if active_agents:
            return f"⏳ Draining {active_agents} active agent(s) before restart..."
        return "♻ Restarting gateway..."

    async def _handle_help_command(self, event: MessageEvent) -> str:
        """Handle /help command - list available commands."""
        from hermes_cli.commands import gateway_help_lines
        lines = [
            "📖 **Hermes Commands**\n",
            *gateway_help_lines(),
        ]
        try:
            from agent.skill_commands import get_skill_commands
            skill_cmds = get_skill_commands()
            if skill_cmds:
                lines.append(f"\n⚡ **Skill Commands** ({len(skill_cmds)} active):")
                # Show first 10, then point to /commands for the rest
                sorted_cmds = sorted(skill_cmds)
                for cmd in sorted_cmds[:10]:
                    lines.append(f"`{cmd}` — {skill_cmds[cmd]['description']}")
                if len(sorted_cmds) > 10:
                    lines.append(f"\n... and {len(sorted_cmds) - 10} more. Use `/commands` for the full paginated list.")
        except Exception:
            pass
        return "\n".join(lines)

    async def _handle_commands_command(self, event: MessageEvent) -> str:
        """Handle /commands [page] - paginated list of all commands and skills."""
        from hermes_cli.commands import gateway_help_lines

        raw_args = event.get_command_args().strip()
        if raw_args:
            try:
                requested_page = int(raw_args)
            except ValueError:
                return "Usage: `/commands [page]`"
        else:
            requested_page = 1

        # Build combined entry list: built-in commands + skill commands
        entries = list(gateway_help_lines())
        try:
            from agent.skill_commands import get_skill_commands
            skill_cmds = get_skill_commands()
            if skill_cmds:
                entries.append("")
                entries.append("⚡ **Skill Commands**:")
                for cmd in sorted(skill_cmds):
                    desc = skill_cmds[cmd].get("description", "").strip() or "Skill command"
                    entries.append(f"`{cmd}` — {desc}")
        except Exception:
            pass

        if not entries:
            return "No commands available."

        from gateway.config import Platform
        page_size = 15 if event.source.platform == Platform.TELEGRAM else 20
        total_pages = max(1, (len(entries) + page_size - 1) // page_size)
        page = max(1, min(requested_page, total_pages))
        start = (page - 1) * page_size
        page_entries = entries[start:start + page_size]

        lines = [
            f"📚 **Commands** ({len(entries)} total, page {page}/{total_pages})",
            "",
            *page_entries,
        ]
        if total_pages > 1:
            nav_parts = []
            if page > 1:
                nav_parts.append(f"`/commands {page - 1}` ← prev")
            if page < total_pages:
                nav_parts.append(f"next → `/commands {page + 1}`")
            lines.extend(["", " | ".join(nav_parts)])
        if page != requested_page:
            lines.append(f"_(Requested page {requested_page} was out of range, showing page {page}.)_")
        return "\n".join(lines)
    
    async def _handle_model_command(self, event: MessageEvent) -> Optional[str]:
        """Handle /model command — switch model for this session.

        Supports:
          /model                              — interactive picker (Telegram/Discord) or text list
          /model <name>                       — switch for this session only
          /model <name> --global              — switch and persist to config.yaml
          /model <name> --provider <provider> — switch provider + model
          /model --provider <provider>        — switch to provider, auto-detect model
        """
        import yaml
        from hermes_cli.model_switch import (
            switch_model as _switch_model, parse_model_flags,
            list_authenticated_providers,
        )
        from hermes_cli.providers import get_label

        raw_args = event.get_command_args().strip()

        # Parse --provider and --global flags
        model_input, explicit_provider, persist_global = parse_model_flags(raw_args)

        # Read current model/provider from config
        current_model = ""
        current_provider = "openrouter"
        current_base_url = ""
        current_api_key = ""
        user_provs = None
        custom_provs = None
        config_path = _hermes_home / "config.yaml"
        try:
            if config_path.exists():
                with open(config_path, encoding="utf-8") as f:
                    cfg = yaml.safe_load(f) or {}
                model_cfg = cfg.get("model", {})
                if isinstance(model_cfg, dict):
                    current_model = model_cfg.get("default", "")
                    current_provider = model_cfg.get("provider", current_provider)
                    current_base_url = model_cfg.get("base_url", "")
                user_provs = cfg.get("providers")
                custom_provs = cfg.get("custom_providers")
        except Exception:
            pass

        # Check for session override
        source = event.source
        session_key = self._session_key_for_source(source)
        override = self._session_model_overrides.get(session_key, {})
        if override:
            current_model = override.get("model", current_model)
            current_provider = override.get("provider", current_provider)
            current_base_url = override.get("base_url", current_base_url)
            current_api_key = override.get("api_key", current_api_key)

        # No args: show interactive picker (Telegram/Discord) or text list
        if not model_input and not explicit_provider:
            # Try interactive picker if the platform supports it
            adapter = self.adapters.get(source.platform)
            has_picker = (
                adapter is not None
                and getattr(type(adapter), "send_model_picker", None) is not None
            )

            if has_picker:
                try:
                    providers = list_authenticated_providers(
                        current_provider=current_provider,
                        user_providers=user_provs,
                        custom_providers=custom_provs,
                        max_models=50,
                    )
                except Exception:
                    providers = []

                if providers:
                    # Build a callback closure for when the user picks a model.
                    # Captures self + locals needed for the switch logic.
                    _self = self
                    _session_key = session_key
                    _cur_model = current_model
                    _cur_provider = current_provider
                    _cur_base_url = current_base_url
                    _cur_api_key = current_api_key

                    async def _on_model_selected(
                        _chat_id: str, model_id: str, provider_slug: str
                    ) -> str:
                        """Perform the model switch and return confirmation text."""
                        result = _switch_model(
                            raw_input=model_id,
                            current_provider=_cur_provider,
                            current_model=_cur_model,
                            current_base_url=_cur_base_url,
                            current_api_key=_cur_api_key,
                            is_global=False,
                            explicit_provider=provider_slug,
                            user_providers=user_provs,
                            custom_providers=custom_provs,
                        )
                        if not result.success:
                            return f"Error: {result.error_message}"

                        # Update cached agent in-place
                        cached_entry = None
                        _cache_lock = getattr(_self, "_agent_cache_lock", None)
                        _cache = getattr(_self, "_agent_cache", None)
                        if _cache_lock and _cache is not None:
                            with _cache_lock:
                                cached_entry = _cache.get(_session_key)
                        if cached_entry and cached_entry[0] is not None:
                            try:
                                cached_entry[0].switch_model(
                                    new_model=result.new_model,
                                    new_provider=result.target_provider,
                                    api_key=result.api_key,
                                    base_url=result.base_url,
                                    api_mode=result.api_mode,
                                )
                            except Exception as exc:
                                logger.warning("Picker model switch failed for cached agent: %s", exc)

                        # Store model note + session override
                        if not hasattr(_self, "_pending_model_notes"):
                            _self._pending_model_notes = {}
                        _self._pending_model_notes[_session_key] = (
                            f"[Note: model was just switched from {_cur_model} to {result.new_model} "
                            f"via {result.provider_label or result.target_provider}. "
                            f"Adjust your self-identification accordingly.]"
                        )
                        _self._session_model_overrides[_session_key] = {
                            "model": result.new_model,
                            "provider": result.target_provider,
                            "api_key": result.api_key,
                            "base_url": result.base_url,
                            "api_mode": result.api_mode,
                        }

                        # Build confirmation text
                        plabel = result.provider_label or result.target_provider
                        lines = [f"Model switched to `{result.new_model}`"]
                        lines.append(f"Provider: {plabel}")
                        mi = result.model_info
                        if mi:
                            if mi.context_window:
                                lines.append(f"Context: {mi.context_window:,} tokens")
                            if mi.max_output:
                                lines.append(f"Max output: {mi.max_output:,} tokens")
                            if mi.has_cost_data():
                                lines.append(f"Cost: {mi.format_cost()}")
                            lines.append(f"Capabilities: {mi.format_capabilities()}")
                        lines.append("_(session only — use `/model <name> --global` to persist)_")
                        return "\n".join(lines)

                    metadata = {"thread_id": source.thread_id} if source.thread_id else None
                    result = await adapter.send_model_picker(
                        chat_id=source.chat_id,
                        providers=providers,
                        current_model=current_model,
                        current_provider=current_provider,
                        session_key=session_key,
                        on_model_selected=_on_model_selected,
                        metadata=metadata,
                    )
                    if result.success:
                        return None  # Picker sent — adapter handles the response

            # Fallback: text list (for platforms without picker or if picker failed)
            provider_label = get_label(current_provider)
            lines = [f"Current: `{current_model or 'unknown'}` on {provider_label}", ""]

            try:
                providers = list_authenticated_providers(
                    current_provider=current_provider,
                    user_providers=user_provs,
                    custom_providers=custom_provs,
                    max_models=5,
                )
                for p in providers:
                    tag = " (current)" if p["is_current"] else ""
                    lines.append(f"**{p['name']}** `--provider {p['slug']}`{tag}:")
                    if p["models"]:
                        model_strs = ", ".join(f"`{m}`" for m in p["models"])
                        extra = f" (+{p['total_models'] - len(p['models'])} more)" if p["total_models"] > len(p["models"]) else ""
                        lines.append(f"  {model_strs}{extra}")
                    elif p.get("api_url"):
                        lines.append(f"  `{p['api_url']}`")
                    lines.append("")
            except Exception:
                pass

            lines.append("`/model <name>` — switch model")
            lines.append("`/model <name> --provider <slug>` — switch provider")
            lines.append("`/model <name> --global` — persist")
            return "\n".join(lines)

        # Perform the switch
        result = _switch_model(
            raw_input=model_input,
            current_provider=current_provider,
            current_model=current_model,
            current_base_url=current_base_url,
            current_api_key=current_api_key,
            is_global=persist_global,
            explicit_provider=explicit_provider,
            user_providers=user_provs,
            custom_providers=custom_provs,
        )

        if not result.success:
            return f"Error: {result.error_message}"

        # If there's a cached agent, update it in-place
        cached_entry = None
        _cache_lock = getattr(self, "_agent_cache_lock", None)
        _cache = getattr(self, "_agent_cache", None)
        if _cache_lock and _cache is not None:
            with _cache_lock:
                cached_entry = _cache.get(session_key)

        if cached_entry and cached_entry[0] is not None:
            try:
                cached_entry[0].switch_model(
                    new_model=result.new_model,
                    new_provider=result.target_provider,
                    api_key=result.api_key,
                    base_url=result.base_url,
                    api_mode=result.api_mode,
                )
            except Exception as exc:
                logger.warning("In-place model switch failed for cached agent: %s", exc)

        # Store a note to prepend to the next user message so the model
        # knows about the switch (avoids system messages mid-history).
        if not hasattr(self, "_pending_model_notes"):
            self._pending_model_notes = {}
        self._pending_model_notes[session_key] = (
            f"[Note: model was just switched from {current_model} to {result.new_model} "
            f"via {result.provider_label or result.target_provider}. "
            f"Adjust your self-identification accordingly.]"
        )

        # Store session override so next agent creation uses the new model
        self._session_model_overrides[session_key] = {
            "model": result.new_model,
            "provider": result.target_provider,
            "api_key": result.api_key,
            "base_url": result.base_url,
            "api_mode": result.api_mode,
        }

        # Persist to config if --global
        if persist_global:
            try:
                if config_path.exists():
                    with open(config_path, encoding="utf-8") as f:
                        cfg = yaml.safe_load(f) or {}
                else:
                    cfg = {}
                model_cfg = cfg.setdefault("model", {})
                model_cfg["default"] = result.new_model
                model_cfg["provider"] = result.target_provider
                if result.base_url:
                    model_cfg["base_url"] = result.base_url
                from hermes_cli.config import save_config
                save_config(cfg)
            except Exception as e:
                logger.warning("Failed to persist model switch: %s", e)

        # Build confirmation message with full metadata
        provider_label = result.provider_label or result.target_provider
        lines = [f"Model switched to `{result.new_model}`"]
        lines.append(f"Provider: {provider_label}")

        # Rich metadata from models.dev
        mi = result.model_info
        if mi:
            if mi.context_window:
                lines.append(f"Context: {mi.context_window:,} tokens")
            if mi.max_output:
                lines.append(f"Max output: {mi.max_output:,} tokens")
            if mi.has_cost_data():
                lines.append(f"Cost: {mi.format_cost()}")
            lines.append(f"Capabilities: {mi.format_capabilities()}")
        else:
            try:
                from agent.model_metadata import get_model_context_length
                ctx = get_model_context_length(
                    result.new_model,
                    base_url=result.base_url or current_base_url,
                    api_key=result.api_key or current_api_key,
                    provider=result.target_provider,
                )
                lines.append(f"Context: {ctx:,} tokens")
            except Exception:
                pass

        # Cache notice
        cache_enabled = (
            ("openrouter" in (result.base_url or "").lower() and "claude" in result.new_model.lower())
            or result.api_mode == "anthropic_messages"
        )
        if cache_enabled:
            lines.append("Prompt caching: enabled")

        if result.warning_message:
            lines.append(f"Warning: {result.warning_message}")

        if persist_global:
            lines.append("Saved to config.yaml (`--global`)")
        else:
            lines.append("_(session only -- add `--global` to persist)_")

        return "\n".join(lines)

    async def _handle_provider_command(self, event: MessageEvent) -> str:
        """Handle /provider command - show available providers."""
        import yaml
        from hermes_cli.models import (
            list_available_providers,
            normalize_provider,
            _PROVIDER_LABELS,
        )

        # Resolve current provider from config
        current_provider = "openrouter"
        model_cfg = {}
        config_path = _hermes_home / 'config.yaml'
        try:
            if config_path.exists():
                with open(config_path, encoding="utf-8") as f:
                    cfg = yaml.safe_load(f) or {}
                model_cfg = cfg.get("model", {})
                if isinstance(model_cfg, dict):
                    current_provider = model_cfg.get("provider", current_provider)
        except Exception:
            pass

        current_provider = normalize_provider(current_provider)
        if current_provider == "auto":
            try:
                from hermes_cli.auth import resolve_provider as _resolve_provider
                current_provider = _resolve_provider(current_provider)
            except Exception:
                current_provider = "openrouter"

        # Detect custom endpoint from config base_url
        if current_provider == "openrouter":
            _cfg_base = model_cfg.get("base_url", "") if isinstance(model_cfg, dict) else ""
            if _cfg_base and "openrouter.ai" not in _cfg_base:
                current_provider = "custom"

        current_label = _PROVIDER_LABELS.get(current_provider, current_provider)

        lines = [
            f"🔌 **Current provider:** {current_label} (`{current_provider}`)",
            "",
            "**Available providers:**",
        ]

        providers = list_available_providers()
        for p in providers:
            marker = " ← active" if p["id"] == current_provider else ""
            auth = "✅" if p["authenticated"] else "❌"
            aliases = f"  _(also: {', '.join(p['aliases'])})_" if p["aliases"] else ""
            lines.append(f"{auth} `{p['id']}` — {p['label']}{aliases}{marker}")

        lines.append("")
        lines.append("Switch: `/model provider:model-name`")
        lines.append("Setup: `hermes setup`")
        return "\n".join(lines)
    
    async def _handle_personality_command(self, event: MessageEvent) -> str:
        """Handle /personality command - list or set a personality."""
        import yaml

        args = event.get_command_args().strip().lower()
        config_path = _hermes_home / 'config.yaml'

        try:
            if config_path.exists():
                with open(config_path, 'r', encoding="utf-8") as f:
                    config = yaml.safe_load(f) or {}
                personalities = config.get("agent", {}).get("personalities", {})
            else:
                config = {}
                personalities = {}
        except Exception:
            config = {}
            personalities = {}

        if not personalities:
            return "No personalities configured in `~/.hermes/config.yaml`"

        if not args:
            lines = ["🎭 **Available Personalities**\n"]
            lines.append("• `none` — (no personality overlay)")
            for name, prompt in personalities.items():
                if isinstance(prompt, dict):
                    preview = prompt.get("description") or prompt.get("system_prompt", "")[:50]
                else:
                    preview = prompt[:50] + "..." if len(prompt) > 50 else prompt
                lines.append(f"• `{name}` — {preview}")
            lines.append("\nUsage: `/personality <name>`")
            return "\n".join(lines)

        def _resolve_prompt(value):
            if isinstance(value, dict):
                parts = [value.get("system_prompt", "")]
                if value.get("tone"):
                    parts.append(f'Tone: {value["tone"]}')
                if value.get("style"):
                    parts.append(f'Style: {value["style"]}')
                return "\n".join(p for p in parts if p)
            return str(value)

        if args in ("none", "default", "neutral"):
            try:
                if "agent" not in config or not isinstance(config.get("agent"), dict):
                    config["agent"] = {}
                config["agent"]["system_prompt"] = ""
                atomic_yaml_write(config_path, config)
            except Exception as e:
                return f"⚠️ Failed to save personality change: {e}"
            self._ephemeral_system_prompt = ""
            return "🎭 Personality cleared — using base agent behavior.\n_(takes effect on next message)_"
        elif args in personalities:
            new_prompt = _resolve_prompt(personalities[args])

            # Write to config.yaml, same pattern as CLI save_config_value.
            try:
                if "agent" not in config or not isinstance(config.get("agent"), dict):
                    config["agent"] = {}
                config["agent"]["system_prompt"] = new_prompt
                atomic_yaml_write(config_path, config)
            except Exception as e:
                return f"⚠️ Failed to save personality change: {e}"

            # Update in-memory so it takes effect on the very next message.
            self._ephemeral_system_prompt = new_prompt

            return f"🎭 Personality set to **{args}**\n_(takes effect on next message)_"

        available = "`none`, " + ", ".join(f"`{n}`" for n in personalities)
        return f"Unknown personality: `{args}`\n\nAvailable: {available}"
    
    async def _handle_retry_command(self, event: MessageEvent) -> str:
        """Handle /retry command - re-send the last user message."""
        source = event.source
        session_entry = self.session_store.get_or_create_session(source)
        history = self.session_store.load_transcript(session_entry.session_id)
        
        # Find the last user message
        last_user_msg = None
        last_user_idx = None
        for i in range(len(history) - 1, -1, -1):
            if history[i].get("role") == "user":
                last_user_msg = history[i].get("content", "")
                last_user_idx = i
                break
        
        if not last_user_msg:
            return "No previous message to retry."
        
        # Truncate history to before the last user message and persist
        truncated = history[:last_user_idx]
        self.session_store.rewrite_transcript(session_entry.session_id, truncated)
        # Reset stored token count — transcript was truncated
        session_entry.last_prompt_tokens = 0
        
        # Re-send by creating a fake text event with the old message
        retry_event = MessageEvent(
            text=last_user_msg,
            message_type=MessageType.TEXT,
            source=source,
            raw_message=event.raw_message,
        )
        
        # Let the normal message handler process it
        return await self._handle_message(retry_event)
    
    async def _handle_undo_command(self, event: MessageEvent) -> str:
        """Handle /undo command - remove the last user/assistant exchange."""
        source = event.source
        session_entry = self.session_store.get_or_create_session(source)
        history = self.session_store.load_transcript(session_entry.session_id)
        
        # Find the last user message and remove everything from it onward
        last_user_idx = None
        for i in range(len(history) - 1, -1, -1):
            if history[i].get("role") == "user":
                last_user_idx = i
                break
        
        if last_user_idx is None:
            return "Nothing to undo."
        
        removed_msg = history[last_user_idx].get("content", "")
        removed_count = len(history) - last_user_idx
        self.session_store.rewrite_transcript(session_entry.session_id, history[:last_user_idx])
        # Reset stored token count — transcript was truncated
        session_entry.last_prompt_tokens = 0
        
        preview = removed_msg[:40] + "..." if len(removed_msg) > 40 else removed_msg
        return f"↩️ Undid {removed_count} message(s).\nRemoved: \"{preview}\""
    
    async def _handle_set_home_command(self, event: MessageEvent) -> str:
        """Handle /sethome command -- set the current chat as the platform's home channel."""
        source = event.source
        platform_name = source.platform.value if source.platform else "unknown"
        chat_id = source.chat_id
        chat_name = source.chat_name or chat_id
        
        env_key = f"{platform_name.upper()}_HOME_CHANNEL"
        
        # Save to config.yaml
        try:
            import yaml
            config_path = _hermes_home / 'config.yaml'
            user_config = {}
            if config_path.exists():
                with open(config_path, encoding="utf-8") as f:
                    user_config = yaml.safe_load(f) or {}
            user_config[env_key] = chat_id
            atomic_yaml_write(config_path, user_config)
            # Also set in the current environment so it takes effect immediately
            os.environ[env_key] = str(chat_id)
        except Exception as e:
            return f"Failed to save home channel: {e}"
        
        return (
            f"✅ Home channel set to **{chat_name}** (ID: {chat_id}).\n"
            f"Cron jobs and cross-platform messages will be delivered here."
        )
    
    @staticmethod
    def _get_guild_id(event: MessageEvent) -> Optional[int]:
        """Extract Discord guild_id from the raw message object."""
        raw = getattr(event, "raw_message", None)
        if raw is None:
            return None
        # Slash command interaction
        if hasattr(raw, "guild_id") and raw.guild_id:
            return int(raw.guild_id)
        # Regular message
        if hasattr(raw, "guild") and raw.guild:
            return raw.guild.id
        return None

    async def _handle_voice_command(self, event: MessageEvent) -> str:
        """Handle /voice [on|off|tts|channel|leave|status] command."""
        args = event.get_command_args().strip().lower()
        chat_id = event.source.chat_id

        adapter = self.adapters.get(event.source.platform)

        if args in ("on", "enable"):
            self._voice_mode[chat_id] = "voice_only"
            self._save_voice_modes()
            if adapter:
                self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=False)
            return (
                "Voice mode enabled.\n"
                "I'll reply with voice when you send voice messages.\n"
                "Use /voice tts to get voice replies for all messages."
            )
        elif args in ("off", "disable"):
            self._voice_mode[chat_id] = "off"
            self._save_voice_modes()
            if adapter:
                self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=True)
            return "Voice mode disabled. Text-only replies."
        elif args == "tts":
            self._voice_mode[chat_id] = "all"
            self._save_voice_modes()
            if adapter:
                self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=False)
            return (
                "Auto-TTS enabled.\n"
                "All replies will include a voice message."
            )
        elif args in ("channel", "join"):
            return await self._handle_voice_channel_join(event)
        elif args == "leave":
            return await self._handle_voice_channel_leave(event)
        elif args == "status":
            mode = self._voice_mode.get(chat_id, "off")
            labels = {
                "off": "Off (text only)",
                "voice_only": "On (voice reply to voice messages)",
                "all": "TTS (voice reply to all messages)",
            }
            # Append voice channel info if connected
            adapter = self.adapters.get(event.source.platform)
            guild_id = self._get_guild_id(event)
            if guild_id and hasattr(adapter, "get_voice_channel_info"):
                info = adapter.get_voice_channel_info(guild_id)
                if info:
                    lines = [
                        f"Voice mode: {labels.get(mode, mode)}",
                        f"Voice channel: #{info['channel_name']}",
                        f"Participants: {info['member_count']}",
                    ]
                    for m in info["members"]:
                        status = " (speaking)" if m.get("is_speaking") else ""
                        lines.append(f"  - {m['display_name']}{status}")
                    return "\n".join(lines)
            return f"Voice mode: {labels.get(mode, mode)}"
        else:
            # Toggle: off → on, on/all → off
            current = self._voice_mode.get(chat_id, "off")
            if current == "off":
                self._voice_mode[chat_id] = "voice_only"
                self._save_voice_modes()
                if adapter:
                    self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=False)
                return "Voice mode enabled."
            else:
                self._voice_mode[chat_id] = "off"
                self._save_voice_modes()
                if adapter:
                    self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=True)
                return "Voice mode disabled."

    async def _handle_voice_channel_join(self, event: MessageEvent) -> str:
        """Join the user's current Discord voice channel."""
        adapter = self.adapters.get(event.source.platform)
        if not hasattr(adapter, "join_voice_channel"):
            return "Voice channels are not supported on this platform."

        guild_id = self._get_guild_id(event)
        if not guild_id:
            return "This command only works in a Discord server."

        voice_channel = await adapter.get_user_voice_channel(
            guild_id, event.source.user_id
        )
        if not voice_channel:
            return "You need to be in a voice channel first."

        # Wire callbacks BEFORE join so voice input arriving immediately
        # after connection is not lost.
        if hasattr(adapter, "_voice_input_callback"):
            adapter._voice_input_callback = self._handle_voice_channel_input
        if hasattr(adapter, "_on_voice_disconnect"):
            adapter._on_voice_disconnect = self._handle_voice_timeout_cleanup

        try:
            success = await adapter.join_voice_channel(voice_channel)
        except Exception as e:
            logger.warning("Failed to join voice channel: %s", e)
            adapter._voice_input_callback = None
            err_lower = str(e).lower()
            if "pynacl" in err_lower or "nacl" in err_lower or "davey" in err_lower:
                return (
                    "Voice dependencies are missing (PyNaCl / davey). "
                    "Install or reinstall Hermes with the messaging extra, e.g. "
                    "`pip install hermes-agent[messaging]`."
                )
            return f"Failed to join voice channel: {e}"

        if success:
            adapter._voice_text_channels[guild_id] = int(event.source.chat_id)
            self._voice_mode[event.source.chat_id] = "all"
            self._save_voice_modes()
            self._set_adapter_auto_tts_disabled(adapter, event.source.chat_id, disabled=False)
            return (
                f"Joined voice channel **{voice_channel.name}**.\n"
                f"I'll speak my replies and listen to you. Use /voice leave to disconnect."
            )
        # Join failed — clear callback
        adapter._voice_input_callback = None
        return "Failed to join voice channel. Check bot permissions (Connect + Speak)."

    async def _handle_voice_channel_leave(self, event: MessageEvent) -> str:
        """Leave the Discord voice channel."""
        adapter = self.adapters.get(event.source.platform)
        guild_id = self._get_guild_id(event)

        if not guild_id or not hasattr(adapter, "leave_voice_channel"):
            return "Not in a voice channel."

        if not hasattr(adapter, "is_in_voice_channel") or not adapter.is_in_voice_channel(guild_id):
            return "Not in a voice channel."

        try:
            await adapter.leave_voice_channel(guild_id)
        except Exception as e:
            logger.warning("Error leaving voice channel: %s", e)
        # Always clean up state even if leave raised an exception
        self._voice_mode[event.source.chat_id] = "off"
        self._save_voice_modes()
        self._set_adapter_auto_tts_disabled(adapter, event.source.chat_id, disabled=True)
        if hasattr(adapter, "_voice_input_callback"):
            adapter._voice_input_callback = None
        return "Left voice channel."

    def _handle_voice_timeout_cleanup(self, chat_id: str) -> None:
        """Called by the adapter when a voice channel times out.

        Cleans up runner-side voice_mode state that the adapter cannot reach.
        """
        self._voice_mode[chat_id] = "off"
        self._save_voice_modes()
        adapter = self.adapters.get(Platform.DISCORD)
        self._set_adapter_auto_tts_disabled(adapter, chat_id, disabled=True)

    async def _handle_voice_channel_input(
        self, guild_id: int, user_id: int, transcript: str
    ):
        """Handle transcribed voice from a user in a voice channel.

        Creates a synthetic MessageEvent and processes it through the
        adapter's full message pipeline (session, typing, agent, TTS reply).
        """
        adapter = self.adapters.get(Platform.DISCORD)
        if not adapter:
            return

        text_ch_id = adapter._voice_text_channels.get(guild_id)
        if not text_ch_id:
            return

        # Check authorization before processing voice input
        source = SessionSource(
            platform=Platform.DISCORD,
            chat_id=str(text_ch_id),
            user_id=str(user_id),
            user_name=str(user_id),
            chat_type="channel",
        )
        if not self._is_user_authorized(source):
            logger.debug("Unauthorized voice input from user %d, ignoring", user_id)
            return

        # Show transcript in text channel (after auth, with mention sanitization)
        try:
            channel = adapter._client.get_channel(text_ch_id)
            if channel:
                safe_text = transcript[:2000].replace("@everyone", "@\u200beveryone").replace("@here", "@\u200bhere")
                await channel.send(f"**[Voice]** <@{user_id}>: {safe_text}")
        except Exception:
            pass

        # Build a synthetic MessageEvent and feed through the normal pipeline
        # Use SimpleNamespace as raw_message so _get_guild_id() can extract
        # guild_id and _send_voice_reply() plays audio in the voice channel.
        from types import SimpleNamespace
        event = MessageEvent(
            source=source,
            text=transcript,
            message_type=MessageType.VOICE,
            raw_message=SimpleNamespace(guild_id=guild_id, guild=None),
        )

        await adapter.handle_message(event)

    def _should_send_voice_reply(
        self,
        event: MessageEvent,
        response: str,
        agent_messages: list,
        already_sent: bool = False,
    ) -> bool:
        """Decide whether the runner should send a TTS voice reply.

        Returns False when:
        - voice_mode is off for this chat
        - response is empty or an error
        - agent already called text_to_speech tool (dedup)
        - voice input and base adapter auto-TTS already handled it (skip_double)
          UNLESS streaming already consumed the response (already_sent=True),
          in which case the base adapter won't have text for auto-TTS so the
          runner must handle it.
        """
        if not response or response.startswith("Error:"):
            return False

        chat_id = event.source.chat_id
        voice_mode = self._voice_mode.get(chat_id, "off")
        is_voice_input = (event.message_type == MessageType.VOICE)

        should = (
            (voice_mode == "all")
            or (voice_mode == "voice_only" and is_voice_input)
        )
        if not should:
            return False

        # Dedup: agent already called TTS tool
        has_agent_tts = any(
            msg.get("role") == "assistant"
            and any(
                tc.get("function", {}).get("name") == "text_to_speech"
                for tc in (msg.get("tool_calls") or [])
            )
            for msg in agent_messages
        )
        if has_agent_tts:
            return False

        # Dedup: base adapter auto-TTS already handles voice input
        # (play_tts plays in VC when connected, so runner can skip).
        # When streaming already delivered the text (already_sent=True),
        # the base adapter will receive None and can't run auto-TTS,
        # so the runner must take over.
        if is_voice_input and not already_sent:
            return False

        return True

    async def _send_voice_reply(self, event: MessageEvent, text: str) -> None:
        """Generate TTS audio and send as a voice message before the text reply."""
        import uuid as _uuid
        audio_path = None
        actual_path = None
        try:
            from tools.tts_tool import text_to_speech_tool, _strip_markdown_for_tts

            tts_text = _strip_markdown_for_tts(text[:4000])
            if not tts_text:
                return

            # Use .mp3 extension so edge-tts conversion to opus works correctly.
            # The TTS tool may convert to .ogg — use file_path from result.
            audio_path = os.path.join(
                tempfile.gettempdir(), "hermes_voice",
                f"tts_reply_{_uuid.uuid4().hex[:12]}.mp3",
            )
            os.makedirs(os.path.dirname(audio_path), exist_ok=True)

            result_json = await asyncio.to_thread(
                text_to_speech_tool, text=tts_text, output_path=audio_path
            )
            result = json.loads(result_json)

            # Use the actual file path from result (may differ after opus conversion)
            actual_path = result.get("file_path", audio_path)
            if not result.get("success") or not os.path.isfile(actual_path):
                logger.warning("Auto voice reply TTS failed: %s", result.get("error"))
                return

            adapter = self.adapters.get(event.source.platform)

            # If connected to a voice channel, play there instead of sending a file
            guild_id = self._get_guild_id(event)
            if (guild_id
                    and hasattr(adapter, "play_in_voice_channel")
                    and hasattr(adapter, "is_in_voice_channel")
                    and adapter.is_in_voice_channel(guild_id)):
                await adapter.play_in_voice_channel(guild_id, actual_path)
            elif adapter and hasattr(adapter, "send_voice"):
                send_kwargs: Dict[str, Any] = {
                    "chat_id": event.source.chat_id,
                    "audio_path": actual_path,
                    "reply_to": event.message_id,
                }
                if event.source.thread_id:
                    send_kwargs["metadata"] = {"thread_id": event.source.thread_id}
                await adapter.send_voice(**send_kwargs)
        except Exception as e:
            logger.warning("Auto voice reply failed: %s", e, exc_info=True)
        finally:
            for p in {audio_path, actual_path} - {None}:
                try:
                    os.unlink(p)
                except OSError:
                    pass

    async def _deliver_media_from_response(
        self,
        response: str,
        event: MessageEvent,
        adapter,
    ) -> None:
        """Extract MEDIA: tags and local file paths from a response and deliver them.

        Called after streaming has already sent the text to the user, so the
        text itself is already delivered — this only handles file attachments
        that the normal _process_message_background path would have caught.
        """
        from pathlib import Path

        try:
            media_files, _ = adapter.extract_media(response)
            _, cleaned = adapter.extract_images(response)
            local_files, _ = adapter.extract_local_files(cleaned)

            _thread_meta = {"thread_id": event.source.thread_id} if event.source.thread_id else None

            _AUDIO_EXTS = {'.ogg', '.opus', '.mp3', '.wav', '.m4a'}
            _VIDEO_EXTS = {'.mp4', '.mov', '.avi', '.mkv', '.webm', '.3gp'}
            _IMAGE_EXTS = {'.jpg', '.jpeg', '.png', '.webp', '.gif'}

            for media_path, is_voice in media_files:
                try:
                    ext = Path(media_path).suffix.lower()
                    if ext in _AUDIO_EXTS:
                        await adapter.send_voice(
                            chat_id=event.source.chat_id,
                            audio_path=media_path,
                            metadata=_thread_meta,
                        )
                    elif ext in _VIDEO_EXTS:
                        await adapter.send_video(
                            chat_id=event.source.chat_id,
                            video_path=media_path,
                            metadata=_thread_meta,
                        )
                    elif ext in _IMAGE_EXTS:
                        await adapter.send_image_file(
                            chat_id=event.source.chat_id,
                            image_path=media_path,
                            metadata=_thread_meta,
                        )
                    else:
                        await adapter.send_document(
                            chat_id=event.source.chat_id,
                            file_path=media_path,
                            metadata=_thread_meta,
                        )
                except Exception as e:
                    logger.warning("[%s] Post-stream media delivery failed: %s", adapter.name, e)

            for file_path in local_files:
                try:
                    ext = Path(file_path).suffix.lower()
                    if ext in _IMAGE_EXTS:
                        await adapter.send_image_file(
                            chat_id=event.source.chat_id,
                            image_path=file_path,
                            metadata=_thread_meta,
                        )
                    else:
                        await adapter.send_document(
                            chat_id=event.source.chat_id,
                            file_path=file_path,
                            metadata=_thread_meta,
                        )
                except Exception as e:
                    logger.warning("[%s] Post-stream file delivery failed: %s", adapter.name, e)

        except Exception as e:
            logger.warning("Post-stream media extraction failed: %s", e)

    async def _handle_rollback_command(self, event: MessageEvent) -> str:
        """Handle /rollback command — list or restore filesystem checkpoints."""
        from tools.checkpoint_manager import CheckpointManager, format_checkpoint_list

        # Read checkpoint config from config.yaml
        cp_cfg = {}
        try:
            import yaml as _y
            _cfg_path = _hermes_home / "config.yaml"
            if _cfg_path.exists():
                with open(_cfg_path, encoding="utf-8") as _f:
                    _data = _y.safe_load(_f) or {}
                cp_cfg = _data.get("checkpoints", {})
                if isinstance(cp_cfg, bool):
                    cp_cfg = {"enabled": cp_cfg}
        except Exception:
            pass

        if not cp_cfg.get("enabled", False):
            return (
                "Checkpoints are not enabled.\n"
                "Enable in config.yaml:\n```\ncheckpoints:\n  enabled: true\n```"
            )

        mgr = CheckpointManager(
            enabled=True,
            max_snapshots=cp_cfg.get("max_snapshots", 50),
        )

        cwd = os.getenv("MESSAGING_CWD", str(Path.home()))
        arg = event.get_command_args().strip()

        if not arg:
            checkpoints = mgr.list_checkpoints(cwd)
            return format_checkpoint_list(checkpoints, cwd)

        # Restore by number or hash
        checkpoints = mgr.list_checkpoints(cwd)
        if not checkpoints:
            return f"No checkpoints found for {cwd}"

        target_hash = None
        try:
            idx = int(arg) - 1
            if 0 <= idx < len(checkpoints):
                target_hash = checkpoints[idx]["hash"]
            else:
                return f"Invalid checkpoint number. Use 1-{len(checkpoints)}."
        except ValueError:
            target_hash = arg

        result = mgr.restore(cwd, target_hash)
        if result["success"]:
            return (
                f"✅ Restored to checkpoint {result['restored_to']}: {result['reason']}\n"
                f"A pre-rollback snapshot was saved automatically."
            )
        return f"❌ {result['error']}"

    async def _handle_background_command(self, event: MessageEvent) -> str:
        """Handle /background <prompt> — run a prompt in a separate background session.

        Spawns a new AIAgent in a background thread with its own session.
        When it completes, sends the result back to the same chat without
        modifying the active session's conversation history.
        """
        prompt = event.get_command_args().strip()
        if not prompt:
            return (
                "Usage: /background <prompt>\n"
                "Example: /background Summarize the top HN stories today\n\n"
                "Runs the prompt in a separate session. "
                "You can keep chatting — the result will appear here when done."
            )

        source = event.source
        task_id = f"bg_{datetime.now().strftime('%H%M%S')}_{os.urandom(3).hex()}"

        # Fire-and-forget the background task
        _task = asyncio.create_task(
            self._run_background_task(prompt, source, task_id)
        )
        self._background_tasks.add(_task)
        _task.add_done_callback(self._background_tasks.discard)

        preview = prompt[:60] + ("..." if len(prompt) > 60 else "")
        return f'🔄 Background task started: "{preview}"\nTask ID: {task_id}\nYou can keep chatting — results will appear when done.'

    async def _run_background_task(
        self, prompt: str, source: "SessionSource", task_id: str
    ) -> None:
        """Execute a background agent task and deliver the result to the chat."""
        from run_agent import AIAgent

        adapter = self.adapters.get(source.platform)
        if not adapter:
            logger.warning("No adapter for platform %s in background task %s", source.platform, task_id)
            return

        _thread_metadata = {"thread_id": source.thread_id} if source.thread_id else None

        try:
            user_config = _load_gateway_config()
            model, runtime_kwargs = self._resolve_session_agent_runtime(
                source=source,
                user_config=user_config,
            )
            if not runtime_kwargs.get("api_key"):
                await adapter.send(
                    source.chat_id,
                    f"❌ Background task {task_id} failed: no provider credentials configured.",
                    metadata=_thread_metadata,
                )
                return

            platform_key = _platform_config_key(source.platform)

            from hermes_cli.tools_config import _get_platform_tools
            enabled_toolsets = sorted(_get_platform_tools(user_config, platform_key))

            pr = self._provider_routing
            max_iterations = int(os.getenv("HERMES_MAX_ITERATIONS", "90"))
            reasoning_config = self._load_reasoning_config()
            self._reasoning_config = reasoning_config
            self._service_tier = self._load_service_tier()
            turn_route = self._resolve_turn_agent_config(prompt, model, runtime_kwargs)

            def run_sync():
                agent = AIAgent(
                    model=turn_route["model"],
                    **turn_route["runtime"],
                    max_iterations=max_iterations,
                    quiet_mode=True,
                    verbose_logging=False,
                    enabled_toolsets=enabled_toolsets,
                    reasoning_config=reasoning_config,
                    service_tier=self._service_tier,
                    request_overrides=turn_route.get("request_overrides"),
                    providers_allowed=pr.get("only"),
                    providers_ignored=pr.get("ignore"),
                    providers_order=pr.get("order"),
                    provider_sort=pr.get("sort"),
                    provider_require_parameters=pr.get("require_parameters", False),
                    provider_data_collection=pr.get("data_collection"),
                    session_id=task_id,
                    platform=platform_key,
                    user_id=source.user_id,
                    session_db=self._session_db,
                    fallback_model=self._fallback_model,
                )

                return agent.run_conversation(
                    user_message=prompt,
                    task_id=task_id,
                )

            loop = asyncio.get_event_loop()
            result = await loop.run_in_executor(None, run_sync)

            response = result.get("final_response", "") if result else ""
            if not response and result and result.get("error"):
                response = f"Error: {result['error']}"

            # Extract media files from the response
            if response:
                media_files, response = adapter.extract_media(response)
                images, text_content = adapter.extract_images(response)

                preview = prompt[:60] + ("..." if len(prompt) > 60 else "")
                header = f'✅ Background task complete\nPrompt: "{preview}"\n\n'

                if text_content:
                    await adapter.send(
                        chat_id=source.chat_id,
                        content=header + text_content,
                        metadata=_thread_metadata,
                    )
                elif not images and not media_files:
                    await adapter.send(
                        chat_id=source.chat_id,
                        content=header + "(No response generated)",
                        metadata=_thread_metadata,
                    )

                # Send extracted images
                for image_url, alt_text in (images or []):
                    try:
                        await adapter.send_image(
                            chat_id=source.chat_id,
                            image_url=image_url,
                            caption=alt_text,
                        )
                    except Exception:
                        pass

                # Send media files
                for media_path in (media_files or []):
                    try:
                        await adapter.send_document(
                            chat_id=source.chat_id,
                            file_path=media_path,
                        )
                    except Exception:
                        pass
            else:
                preview = prompt[:60] + ("..." if len(prompt) > 60 else "")
                await adapter.send(
                    chat_id=source.chat_id,
                    content=f'✅ Background task complete\nPrompt: "{preview}"\n\n(No response generated)',
                    metadata=_thread_metadata,
                )

        except Exception as e:
            logger.exception("Background task %s failed", task_id)
            try:
                await adapter.send(
                    chat_id=source.chat_id,
                    content=f"❌ Background task {task_id} failed: {e}",
                    metadata=_thread_metadata,
                )
            except Exception:
                pass

    async def _handle_btw_command(self, event: MessageEvent) -> str:
        """Handle /btw <question> — ephemeral side question in the same chat."""
        question = event.get_command_args().strip()
        if not question:
            return (
                "Usage: /btw <question>\n"
                "Example: /btw what module owns session title sanitization?\n\n"
                "Answers using session context. No tools, not persisted."
            )

        source = event.source
        session_key = self._session_key_for_source(source)

        # Guard: one /btw at a time per session
        existing = getattr(self, "_active_btw_tasks", {}).get(session_key)
        if existing and not existing.done():
            return "A /btw is already running for this chat. Wait for it to finish."

        if not hasattr(self, "_active_btw_tasks"):
            self._active_btw_tasks: dict = {}

        import uuid as _uuid
        task_id = f"btw_{datetime.now().strftime('%H%M%S')}_{_uuid.uuid4().hex[:6]}"
        _task = asyncio.create_task(self._run_btw_task(question, source, session_key, task_id))
        self._background_tasks.add(_task)
        self._active_btw_tasks[session_key] = _task

        def _cleanup(task):
            self._background_tasks.discard(task)
            if self._active_btw_tasks.get(session_key) is task:
                self._active_btw_tasks.pop(session_key, None)

        _task.add_done_callback(_cleanup)

        preview = question[:60] + ("..." if len(question) > 60 else "")
        return f'💬 /btw: "{preview}"\nReply will appear here shortly.'

    async def _run_btw_task(
        self, question: str, source, session_key: str, task_id: str,
    ) -> None:
        """Execute an ephemeral /btw side question and deliver the answer."""
        from run_agent import AIAgent

        adapter = self.adapters.get(source.platform)
        if not adapter:
            logger.warning("No adapter for platform %s in /btw task %s", source.platform, task_id)
            return

        _thread_meta = {"thread_id": source.thread_id} if source.thread_id else None

        try:
            user_config = _load_gateway_config()
            model, runtime_kwargs = self._resolve_session_agent_runtime(
                source=source,
                session_key=session_key,
                user_config=user_config,
            )
            if not runtime_kwargs.get("api_key"):
                await adapter.send(
                    source.chat_id,
                    "❌ /btw failed: no provider credentials configured.",
                    metadata=_thread_meta,
                )
                return

            platform_key = _platform_config_key(source.platform)
            reasoning_config = self._load_reasoning_config()
            self._service_tier = self._load_service_tier()
            turn_route = self._resolve_turn_agent_config(question, model, runtime_kwargs)
            pr = self._provider_routing

            # Snapshot history from running agent or stored transcript
            running_agent = self._running_agents.get(session_key)
            if running_agent and running_agent is not _AGENT_PENDING_SENTINEL:
                history_snapshot = list(getattr(running_agent, "_session_messages", []) or [])
            else:
                session_entry = self.session_store.get_or_create_session(source)
                history_snapshot = self.session_store.load_transcript(session_entry.session_id)

            btw_prompt = (
                "[Ephemeral /btw side question. Answer using the conversation "
                "context. No tools available. Be direct and concise.]\n\n"
                + question
            )

            def run_sync():
                agent = AIAgent(
                    model=turn_route["model"],
                    **turn_route["runtime"],
                    max_iterations=8,
                    quiet_mode=True,
                    verbose_logging=False,
                    enabled_toolsets=[],
                    reasoning_config=reasoning_config,
                    service_tier=self._service_tier,
                    request_overrides=turn_route.get("request_overrides"),
                    providers_allowed=pr.get("only"),
                    providers_ignored=pr.get("ignore"),
                    providers_order=pr.get("order"),
                    provider_sort=pr.get("sort"),
                    provider_require_parameters=pr.get("require_parameters", False),
                    provider_data_collection=pr.get("data_collection"),
                    session_id=task_id,
                    platform=platform_key,
                    session_db=None,
                    fallback_model=self._fallback_model,
                    skip_memory=True,
                    skip_context_files=True,
                    persist_session=False,
                )
                return agent.run_conversation(
                    user_message=btw_prompt,
                    conversation_history=history_snapshot,
                    task_id=task_id,
                )

            loop = asyncio.get_event_loop()
            result = await loop.run_in_executor(None, run_sync)

            response = (result.get("final_response") or "") if result else ""
            if not response and result and result.get("error"):
                response = f"Error: {result['error']}"
            if not response:
                response = "(No response generated)"

            media_files, response = adapter.extract_media(response)
            images, text_content = adapter.extract_images(response)
            preview = question[:60] + ("..." if len(question) > 60 else "")
            header = f'💬 /btw: "{preview}"\n\n'

            if text_content:
                await adapter.send(
                    chat_id=source.chat_id,
                    content=header + text_content,
                    metadata=_thread_meta,
                )
            elif not images and not media_files:
                await adapter.send(
                    chat_id=source.chat_id,
                    content=header + "(No response generated)",
                    metadata=_thread_meta,
                )

            for image_url, alt_text in (images or []):
                try:
                    await adapter.send_image(chat_id=source.chat_id, image_url=image_url, caption=alt_text)
                except Exception:
                    pass

            for media_path in (media_files or []):
                try:
                    await adapter.send_file(chat_id=source.chat_id, file_path=media_path)
                except Exception:
                    pass

        except Exception as e:
            logger.exception("/btw task %s failed", task_id)
            try:
                await adapter.send(
                    chat_id=source.chat_id,
                    content=f"❌ /btw failed: {e}",
                    metadata=_thread_meta,
                )
            except Exception:
                pass

    async def _handle_reasoning_command(self, event: MessageEvent) -> str:
        """Handle /reasoning command — manage reasoning effort and display toggle.

        Usage:
            /reasoning              Show current effort level and display state
            /reasoning <level>      Set reasoning effort (none, minimal, low, medium, high, xhigh)
            /reasoning show|on      Show model reasoning in responses
            /reasoning hide|off     Hide model reasoning from responses
        """
        import yaml

        args = event.get_command_args().strip().lower()
        config_path = _hermes_home / "config.yaml"
        self._reasoning_config = self._load_reasoning_config()
        self._show_reasoning = self._load_show_reasoning()

        def _save_config_key(key_path: str, value):
            """Save a dot-separated key to config.yaml."""
            try:
                user_config = {}
                if config_path.exists():
                    with open(config_path, encoding="utf-8") as f:
                        user_config = yaml.safe_load(f) or {}
                keys = key_path.split(".")
                current = user_config
                for k in keys[:-1]:
                    if k not in current or not isinstance(current[k], dict):
                        current[k] = {}
                    current = current[k]
                current[keys[-1]] = value
                atomic_yaml_write(config_path, user_config)
                return True
            except Exception as e:
                logger.error("Failed to save config key %s: %s", key_path, e)
                return False

        if not args:
            # Show current state
            rc = self._reasoning_config
            if rc is None:
                level = "medium (default)"
            elif rc.get("enabled") is False:
                level = "none (disabled)"
            else:
                level = rc.get("effort", "medium")
            display_state = "on ✓" if self._show_reasoning else "off"
            return (
                "🧠 **Reasoning Settings**\n\n"
                f"**Effort:** `{level}`\n"
                f"**Display:** {display_state}\n\n"
                "_Usage:_ `/reasoning <none|minimal|low|medium|high|xhigh|show|hide>`"
            )

        # Display toggle (per-platform)
        platform_key = _platform_config_key(event.source.platform)
        if args in ("show", "on"):
            self._show_reasoning = True
            _save_config_key(f"display.platforms.{platform_key}.show_reasoning", True)
            return (
                "🧠 ✓ Reasoning display: **ON**\n"
                f"Model thinking will be shown before each response on **{platform_key}**."
            )

        if args in ("hide", "off"):
            self._show_reasoning = False
            _save_config_key(f"display.platforms.{platform_key}.show_reasoning", False)
            return f"🧠 ✓ Reasoning display: **OFF** for **{platform_key}**"

        # Effort level change
        effort = args.strip()
        if effort == "none":
            parsed = {"enabled": False}
        elif effort in ("minimal", "low", "medium", "high", "xhigh"):
            parsed = {"enabled": True, "effort": effort}
        else:
            return (
                f"⚠️ Unknown argument: `{effort}`\n\n"
                "**Valid levels:** none, minimal, low, medium, high, xhigh\n"
                "**Display:** show, hide"
            )

        self._reasoning_config = parsed
        if _save_config_key("agent.reasoning_effort", effort):
            return f"🧠 ✓ Reasoning effort set to `{effort}` (saved to config)\n_(takes effect on next message)_"
        else:
            return f"🧠 ✓ Reasoning effort set to `{effort}` (this session only)"

    async def _handle_fast_command(self, event: MessageEvent) -> str:
        """Handle /fast — mirror the CLI Priority Processing toggle in gateway chats."""
        import yaml
        from hermes_cli.models import model_supports_fast_mode

        args = event.get_command_args().strip().lower()
        config_path = _hermes_home / "config.yaml"
        self._service_tier = self._load_service_tier()

        user_config = _load_gateway_config()
        model = _resolve_gateway_model(user_config)
        if not model_supports_fast_mode(model):
            return "⚡ /fast is only available for OpenAI models that support Priority Processing."

        def _save_config_key(key_path: str, value):
            """Save a dot-separated key to config.yaml."""
            try:
                user_config = {}
                if config_path.exists():
                    with open(config_path, encoding="utf-8") as f:
                        user_config = yaml.safe_load(f) or {}
                keys = key_path.split(".")
                current = user_config
                for k in keys[:-1]:
                    if k not in current or not isinstance(current[k], dict):
                        current[k] = {}
                    current = current[k]
                current[keys[-1]] = value
                atomic_yaml_write(config_path, user_config)
                return True
            except Exception as e:
                logger.error("Failed to save config key %s: %s", key_path, e)
                return False

        if not args or args == "status":
            status = "fast" if self._service_tier == "priority" else "normal"
            return (
                "⚡ Priority Processing\n\n"
                f"Current mode: `{status}`\n\n"
                "_Usage:_ `/fast <normal|fast|status>`"
            )

        if args in {"fast", "on"}:
            self._service_tier = "priority"
            saved_value = "fast"
            label = "FAST"
        elif args in {"normal", "off"}:
            self._service_tier = None
            saved_value = "normal"
            label = "NORMAL"
        else:
            return (
                f"⚠️ Unknown argument: `{args}`\n\n"
                "**Valid options:** normal, fast, status"
            )

        if _save_config_key("agent.service_tier", saved_value):
            return f"⚡ ✓ Priority Processing: **{label}** (saved to config)\n_(takes effect on next message)_"
        return f"⚡ ✓ Priority Processing: **{label}** (this session only)"

    async def _handle_yolo_command(self, event: MessageEvent) -> str:
        """Handle /yolo — toggle dangerous command approval bypass for this session only."""
        from tools.approval import (
            disable_session_yolo,
            enable_session_yolo,
            is_session_yolo_enabled,
        )

        session_key = self._session_key_for_source(event.source)
        current = is_session_yolo_enabled(session_key)
        if current:
            disable_session_yolo(session_key)
            return "⚠️ YOLO mode **OFF** for this session — dangerous commands will require approval."
        else:
            enable_session_yolo(session_key)
            return "⚡ YOLO mode **ON** for this session — all commands auto-approved. Use with caution."

    async def _handle_verbose_command(self, event: MessageEvent) -> str:
        """Handle /verbose command — cycle tool progress display mode.

        Gated by ``display.tool_progress_command`` in config.yaml (default off).
        When enabled, cycles the tool progress mode through off → new → all →
        verbose → off for the *current platform*.  The setting is saved to
        ``display.platforms.<platform>.tool_progress`` so each channel can
        have its own verbosity level independently.
        """
        import yaml

        config_path = _hermes_home / "config.yaml"
        platform_key = _platform_config_key(event.source.platform)

        # --- check config gate ------------------------------------------------
        try:
            user_config = {}
            if config_path.exists():
                with open(config_path, encoding="utf-8") as f:
                    user_config = yaml.safe_load(f) or {}
            gate_enabled = user_config.get("display", {}).get("tool_progress_command", False)
        except Exception:
            gate_enabled = False

        if not gate_enabled:
            return (
                "The `/verbose` command is not enabled for messaging platforms.\n\n"
                "Enable it in `config.yaml`:\n```yaml\n"
                "display:\n  tool_progress_command: true\n```"
            )

        # --- cycle mode (per-platform) ----------------------------------------
        cycle = ["off", "new", "all", "verbose"]
        descriptions = {
            "off": "⚙️ Tool progress: **OFF** — no tool activity shown.",
            "new": "⚙️ Tool progress: **NEW** — shown when tool changes (preview length: `display.tool_preview_length`, default 40).",
            "all": "⚙️ Tool progress: **ALL** — every tool call shown (preview length: `display.tool_preview_length`, default 40).",
            "verbose": "⚙️ Tool progress: **VERBOSE** — every tool call with full arguments.",
        }

        # Read current effective mode for this platform via the resolver
        from gateway.display_config import resolve_display_setting
        current = resolve_display_setting(user_config, platform_key, "tool_progress", "all")
        if current not in cycle:
            current = "all"
        idx = (cycle.index(current) + 1) % len(cycle)
        new_mode = cycle[idx]

        # Save to display.platforms.<platform>.tool_progress
        try:
            if "display" not in user_config or not isinstance(user_config.get("display"), dict):
                user_config["display"] = {}
            display = user_config["display"]
            if "platforms" not in display or not isinstance(display.get("platforms"), dict):
                display["platforms"] = {}
            if platform_key not in display["platforms"] or not isinstance(display["platforms"].get(platform_key), dict):
                display["platforms"][platform_key] = {}
            display["platforms"][platform_key]["tool_progress"] = new_mode
            atomic_yaml_write(config_path, user_config)
            return (
                f"{descriptions[new_mode]}\n"
                f"_(saved for **{platform_key}** — takes effect on next message)_"
            )
        except Exception as e:
            logger.warning("Failed to save tool_progress mode: %s", e)
            return f"{descriptions[new_mode]}\n_(could not save to config: {e})_"

    async def _handle_compress_command(self, event: MessageEvent) -> str:
        """Handle /compress command -- manually compress conversation context.

        Accepts an optional focus topic: ``/compress <focus>`` guides the
        summariser to preserve information related to *focus* while being
        more aggressive about discarding everything else.
        """
        source = event.source
        session_entry = self.session_store.get_or_create_session(source)
        history = self.session_store.load_transcript(session_entry.session_id)

        if not history or len(history) < 4:
            return "Not enough conversation to compress (need at least 4 messages)."

        # Extract optional focus topic from command args
        focus_topic = (event.get_command_args() or "").strip() or None

        try:
            from run_agent import AIAgent
            from agent.manual_compression_feedback import summarize_manual_compression
            from agent.model_metadata import estimate_messages_tokens_rough

            session_key = self._session_key_for_source(source)
            model, runtime_kwargs = self._resolve_session_agent_runtime(
                source=source,
                session_key=session_key,
            )
            if not runtime_kwargs.get("api_key"):
                return "No provider configured -- cannot compress."

            msgs = [
                {"role": m.get("role"), "content": m.get("content")}
                for m in history
                if m.get("role") in ("user", "assistant") and m.get("content")
            ]
            original_count = len(msgs)
            approx_tokens = estimate_messages_tokens_rough(msgs)

            tmp_agent = AIAgent(
                **runtime_kwargs,
                model=model,
                max_iterations=4,
                quiet_mode=True,
                enabled_toolsets=["memory"],
                session_id=session_entry.session_id,
            )
            tmp_agent._print_fn = lambda *a, **kw: None

            compressor = tmp_agent.context_compressor
            compress_start = compressor.protect_first_n
            compress_start = compressor._align_boundary_forward(msgs, compress_start)
            compress_end = compressor._find_tail_cut_by_tokens(msgs, compress_start)
            if compress_start >= compress_end:
                return "Nothing to compress yet (the transcript is still all protected context)."

            loop = asyncio.get_event_loop()
            compressed, _ = await loop.run_in_executor(
                None,
                lambda: tmp_agent._compress_context(msgs, "", approx_tokens=approx_tokens, focus_topic=focus_topic)
            )

            # _compress_context already calls end_session() on the old session
            # (preserving its full transcript in SQLite) and creates a new
            # session_id for the continuation.  Write the compressed messages
            # into the NEW session so the original history stays searchable.
            new_session_id = tmp_agent.session_id
            if new_session_id != session_entry.session_id:
                session_entry.session_id = new_session_id
                self.session_store._save()

            self.session_store.rewrite_transcript(new_session_id, compressed)
            # Reset stored token count — transcript changed, old value is stale
            self.session_store.update_session(
                session_entry.session_key, last_prompt_tokens=0
            )
            new_tokens = estimate_messages_tokens_rough(compressed)
            summary = summarize_manual_compression(
                msgs,
                compressed,
                approx_tokens,
                new_tokens,
            )
            lines = [f"🗜️ {summary['headline']}"]
            if focus_topic:
                lines.append(f"Focus: \"{focus_topic}\"")
            lines.append(summary["token_line"])
            if summary["note"]:
                lines.append(summary["note"])
            return "\n".join(lines)
        except Exception as e:
            logger.warning("Manual compress failed: %s", e)
            return f"Compression failed: {e}"

    async def _handle_title_command(self, event: MessageEvent) -> str:
        """Handle /title command — set or show the current session's title."""
        source = event.source
        session_entry = self.session_store.get_or_create_session(source)
        session_id = session_entry.session_id

        if not self._session_db:
            return "Session database not available."

        # Ensure session exists in SQLite DB (it may only exist in session_store
        # if this is the first command in a new session)
        existing_title = self._session_db.get_session_title(session_id)
        if existing_title is None:
            # Session doesn't exist in DB yet — create it
            try:
                self._session_db.create_session(
                    session_id=session_id,
                    source=source.platform.value if source.platform else "unknown",
                    user_id=source.user_id,
                )
            except Exception:
                pass  # Session might already exist, ignore errors

        title_arg = event.get_command_args().strip()
        if title_arg:
            # Sanitize the title before setting
            try:
                sanitized = self._session_db.sanitize_title(title_arg)
            except ValueError as e:
                return f"⚠️ {e}"
            if not sanitized:
                return "⚠️ Title is empty after cleanup. Please use printable characters."
            # Set the title
            try:
                if self._session_db.set_session_title(session_id, sanitized):
                    return f"✏️ Session title set: **{sanitized}**"
                else:
                    return "Session not found in database."
            except ValueError as e:
                return f"⚠️ {e}"
        else:
            # Show the current title and session ID
            title = self._session_db.get_session_title(session_id)
            if title:
                return f"📌 Session: `{session_id}`\nTitle: **{title}**"
            else:
                return f"📌 Session: `{session_id}`\nNo title set. Usage: `/title My Session Name`"

    async def _handle_resume_command(self, event: MessageEvent) -> str:
        """Handle /resume command — switch to a previously-named session."""
        if not self._session_db:
            return "Session database not available."

        source = event.source
        session_key = self._session_key_for_source(source)
        name = event.get_command_args().strip()

        if not name:
            # List recent titled sessions for this user/platform
            try:
                user_source = source.platform.value if source.platform else None
                sessions = self._session_db.list_sessions_rich(
                    source=user_source, limit=10
                )
                titled = [s for s in sessions if s.get("title")]
                if not titled:
                    return (
                        "No named sessions found.\n"
                        "Use `/title My Session` to name your current session, "
                        "then `/resume My Session` to return to it later."
                    )
                lines = ["📋 **Named Sessions**\n"]
                for s in titled[:10]:
                    title = s["title"]
                    preview = s.get("preview", "")[:40]
                    preview_part = f" — _{preview}_" if preview else ""
                    lines.append(f"• **{title}**{preview_part}")
                lines.append("\nUsage: `/resume <session name>`")
                return "\n".join(lines)
            except Exception as e:
                logger.debug("Failed to list titled sessions: %s", e)
                return f"Could not list sessions: {e}"

        # Resolve the name to a session ID
        target_id = self._session_db.resolve_session_by_title(name)
        if not target_id:
            return (
                f"No session found matching '**{name}**'.\n"
                "Use `/resume` with no arguments to see available sessions."
            )

        # Check if already on that session
        current_entry = self.session_store.get_or_create_session(source)
        if current_entry.session_id == target_id:
            return f"📌 Already on session **{name}**."

        # Flush memories for current session before switching
        try:
            _flush_task = asyncio.create_task(
                self._async_flush_memories(current_entry.session_id, session_key)
            )
            self._background_tasks.add(_flush_task)
            _flush_task.add_done_callback(self._background_tasks.discard)
        except Exception as e:
            logger.debug("Memory flush on resume failed: %s", e)

        # Clear any running agent for this session key
        if session_key in self._running_agents:
            del self._running_agents[session_key]

        # Switch the session entry to point at the old session
        new_entry = self.session_store.switch_session(session_key, target_id)
        if not new_entry:
            return "Failed to switch session."

        # Get the title for confirmation
        title = self._session_db.get_session_title(target_id) or name

        # Count messages for context
        history = self.session_store.load_transcript(target_id)
        msg_count = len([m for m in history if m.get("role") == "user"]) if history else 0
        msg_part = f" ({msg_count} message{'s' if msg_count != 1 else ''})" if msg_count else ""

        return f"↻ Resumed session **{title}**{msg_part}. Conversation restored."

    async def _handle_branch_command(self, event: MessageEvent) -> str:
        """Handle /branch [name] — fork the current session into a new independent copy.

        Copies conversation history to a new session so the user can explore
        a different approach without losing the original.
        Inspired by Claude Code's /branch command.
        """
        import uuid as _uuid

        if not self._session_db:
            return "Session database not available."

        source = event.source
        session_key = self._session_key_for_source(source)

        # Load the current session and its transcript
        current_entry = self.session_store.get_or_create_session(source)
        history = self.session_store.load_transcript(current_entry.session_id)
        if not history:
            return "No conversation to branch — send a message first."

        branch_name = event.get_command_args().strip()

        # Generate the new session ID
        from datetime import datetime as _dt
        now = _dt.now()
        timestamp_str = now.strftime("%Y%m%d_%H%M%S")
        short_uuid = _uuid.uuid4().hex[:6]
        new_session_id = f"{timestamp_str}_{short_uuid}"

        # Determine branch title
        if branch_name:
            branch_title = branch_name
        else:
            current_title = self._session_db.get_session_title(current_entry.session_id)
            base = current_title or "branch"
            branch_title = self._session_db.get_next_title_in_lineage(base)

        parent_session_id = current_entry.session_id

        # Create the new session with parent link
        try:
            self._session_db.create_session(
                session_id=new_session_id,
                source=source.platform.value if source.platform else "gateway",
                model=(self.config.get("model", {}) or {}).get("default") if isinstance(self.config, dict) else None,
                parent_session_id=parent_session_id,
            )
        except Exception as e:
            logger.error("Failed to create branch session: %s", e)
            return f"Failed to create branch: {e}"

        # Copy conversation history to the new session
        for msg in history:
            try:
                self._session_db.append_message(
                    session_id=new_session_id,
                    role=msg.get("role", "user"),
                    content=msg.get("content"),
                    tool_name=msg.get("tool_name") or msg.get("name"),
                    tool_calls=msg.get("tool_calls"),
                    tool_call_id=msg.get("tool_call_id"),
                    reasoning=msg.get("reasoning"),
                )
            except Exception:
                pass  # Best-effort copy

        # Set title
        try:
            self._session_db.set_session_title(new_session_id, branch_title)
        except Exception:
            pass

        # Switch the session store entry to the new session
        new_entry = self.session_store.switch_session(session_key, new_session_id)
        if not new_entry:
            return "Branch created but failed to switch to it."

        # Evict any cached agent for this session
        self._evict_cached_agent(session_key)

        msg_count = len([m for m in history if m.get("role") == "user"])
        return (
            f"⑂ Branched to **{branch_title}**"
            f" ({msg_count} message{'s' if msg_count != 1 else ''} copied)\n"
            f"Original: `{parent_session_id}`\n"
            f"Branch: `{new_session_id}`\n"
            f"Use `/resume` to switch back to the original."
        )

    async def _handle_usage_command(self, event: MessageEvent) -> str:
        """Handle /usage command -- show token usage for the current session.

        Checks both _running_agents (mid-turn) and _agent_cache (between turns)
        so that rate limits, cost estimates, and detailed token breakdowns are
        available whenever the user asks, not only while the agent is running.
        """
        source = event.source
        session_key = self._session_key_for_source(source)

        # Try running agent first (mid-turn), then cached agent (between turns)
        agent = self._running_agents.get(session_key)
        if not agent or agent is _AGENT_PENDING_SENTINEL:
            _cache_lock = getattr(self, "_agent_cache_lock", None)
            _cache = getattr(self, "_agent_cache", None)
            if _cache_lock and _cache is not None:
                with _cache_lock:
                    cached = _cache.get(session_key)
                    if cached:
                        agent = cached[0]

        if agent and hasattr(agent, "session_total_tokens") and agent.session_api_calls > 0:
            lines = []

            # Rate limits (when available from provider headers)
            rl_state = agent.get_rate_limit_state()
            if rl_state and rl_state.has_data:
                from agent.rate_limit_tracker import format_rate_limit_compact
                lines.append(f"⏱️ **Rate Limits:** {format_rate_limit_compact(rl_state)}")
                lines.append("")

            # Session token usage — detailed breakdown matching CLI
            input_tokens = getattr(agent, "session_input_tokens", 0) or 0
            output_tokens = getattr(agent, "session_output_tokens", 0) or 0
            cache_read = getattr(agent, "session_cache_read_tokens", 0) or 0
            cache_write = getattr(agent, "session_cache_write_tokens", 0) or 0

            lines.append("📊 **Session Token Usage**")
            lines.append(f"Model: `{agent.model}`")
            lines.append(f"Input tokens: {input_tokens:,}")
            if cache_read:
                lines.append(f"Cache read tokens: {cache_read:,}")
            if cache_write:
                lines.append(f"Cache write tokens: {cache_write:,}")
            lines.append(f"Output tokens: {output_tokens:,}")
            lines.append(f"Total: {agent.session_total_tokens:,}")
            lines.append(f"API calls: {agent.session_api_calls}")

            # Cost estimation
            try:
                from agent.usage_pricing import CanonicalUsage, estimate_usage_cost
                cost_result = estimate_usage_cost(
                    agent.model,
                    CanonicalUsage(
                        input_tokens=input_tokens,
                        output_tokens=output_tokens,
                        cache_read_tokens=cache_read,
                        cache_write_tokens=cache_write,
                    ),
                    provider=getattr(agent, "provider", None),
                    base_url=getattr(agent, "base_url", None),
                )
                if cost_result.amount_usd is not None:
                    prefix = "~" if cost_result.status == "estimated" else ""
                    lines.append(f"Cost: {prefix}${float(cost_result.amount_usd):.4f}")
                elif cost_result.status == "included":
                    lines.append("Cost: included")
            except Exception:
                pass

            # Context window and compressions
            ctx = agent.context_compressor
            if ctx.last_prompt_tokens:
                pct = min(100, ctx.last_prompt_tokens / ctx.context_length * 100) if ctx.context_length else 0
                lines.append(f"Context: {ctx.last_prompt_tokens:,} / {ctx.context_length:,} ({pct:.0f}%)")
            if ctx.compression_count:
                lines.append(f"Compressions: {ctx.compression_count}")

            return "\n".join(lines)

        # No agent at all -- check session history for a rough count
        session_entry = self.session_store.get_or_create_session(source)
        history = self.session_store.load_transcript(session_entry.session_id)
        if history:
            from agent.model_metadata import estimate_messages_tokens_rough
            msgs = [m for m in history if m.get("role") in ("user", "assistant") and m.get("content")]
            approx = estimate_messages_tokens_rough(msgs)
            return (
                f"📊 **Session Info**\n"
                f"Messages: {len(msgs)}\n"
                f"Estimated context: ~{approx:,} tokens\n"
                f"_(Detailed usage available after the first agent response)_"
            )
        return "No usage data available for this session."

    async def _handle_insights_command(self, event: MessageEvent) -> str:
        """Handle /insights command -- show usage insights and analytics."""
        import asyncio as _asyncio

        args = event.get_command_args().strip()
        days = 30
        source = None

        # Parse simple args: /insights 7  or  /insights --days 7
        if args:
            parts = args.split()
            i = 0
            while i < len(parts):
                if parts[i] == "--days" and i + 1 < len(parts):
                    try:
                        days = int(parts[i + 1])
                    except ValueError:
                        return f"Invalid --days value: {parts[i + 1]}"
                    i += 2
                elif parts[i] == "--source" and i + 1 < len(parts):
                    source = parts[i + 1]
                    i += 2
                elif parts[i].isdigit():
                    days = int(parts[i])
                    i += 1
                else:
                    i += 1

        try:
            from hermes_state import SessionDB
            from agent.insights import InsightsEngine

            loop = _asyncio.get_event_loop()

            def _run_insights():
                db = SessionDB()
                engine = InsightsEngine(db)
                report = engine.generate(days=days, source=source)
                result = engine.format_gateway(report)
                db.close()
                return result

            return await loop.run_in_executor(None, _run_insights)
        except Exception as e:
            logger.error("Insights command error: %s", e, exc_info=True)
            return f"Error generating insights: {e}"

    async def _handle_reload_mcp_command(self, event: MessageEvent) -> str:
        """Handle /reload-mcp command -- disconnect and reconnect all MCP servers."""
        loop = asyncio.get_event_loop()
        try:
            from tools.mcp_tool import shutdown_mcp_servers, discover_mcp_tools, _load_mcp_config, _servers, _lock

            # Capture old server names before shutdown
            with _lock:
                old_servers = set(_servers.keys())

            # Read new config before shutting down, so we know what will be added/removed
            # Shutdown existing connections
            await loop.run_in_executor(None, shutdown_mcp_servers)

            # Reconnect by discovering tools (reads config.yaml fresh)
            new_tools = await loop.run_in_executor(None, discover_mcp_tools)

            # Compute what changed
            with _lock:
                connected_servers = set(_servers.keys())

            added = connected_servers - old_servers
            removed = old_servers - connected_servers
            reconnected = connected_servers & old_servers

            lines = ["🔄 **MCP Servers Reloaded**\n"]
            if reconnected:
                lines.append(f"♻️ Reconnected: {', '.join(sorted(reconnected))}")
            if added:
                lines.append(f"➕ Added: {', '.join(sorted(added))}")
            if removed:
                lines.append(f"➖ Removed: {', '.join(sorted(removed))}")
            if not connected_servers:
                lines.append("No MCP servers connected.")
            else:
                lines.append(f"\n🔧 {len(new_tools)} tool(s) available from {len(connected_servers)} server(s)")

            # Inject a message at the END of the session history so the
            # model knows tools changed on its next turn.  Appended after
            # all existing messages to preserve prompt-cache for the prefix.
            change_parts = []
            if added:
                change_parts.append(f"Added servers: {', '.join(sorted(added))}")
            if removed:
                change_parts.append(f"Removed servers: {', '.join(sorted(removed))}")
            if reconnected:
                change_parts.append(f"Reconnected servers: {', '.join(sorted(reconnected))}")
            tool_summary = f"{len(new_tools)} MCP tool(s) now available" if new_tools else "No MCP tools available"
            change_detail = ". ".join(change_parts) + ". " if change_parts else ""
            reload_msg = {
                "role": "user",
                "content": f"[SYSTEM: MCP servers have been reloaded. {change_detail}{tool_summary}. The tool list for this conversation has been updated accordingly.]",
            }
            try:
                session_entry = self.session_store.get_or_create_session(event.source)
                self.session_store.append_to_transcript(
                    session_entry.session_id, reload_msg
                )
            except Exception:
                pass  # Best-effort; don't fail the reload over a transcript write

            return "\n".join(lines)

        except Exception as e:
            logger.warning("MCP reload failed: %s", e)
            return f"❌ MCP reload failed: {e}"

    # ------------------------------------------------------------------
    # /approve & /deny — explicit dangerous-command approval
    # ------------------------------------------------------------------

    _APPROVAL_TIMEOUT_SECONDS = 300  # 5 minutes

    async def _handle_approve_command(self, event: MessageEvent) -> Optional[str]:
        """Handle /approve command — unblock waiting agent thread(s).

        The agent thread(s) are blocked inside tools/approval.py waiting for
        the user to respond.  This handler signals the event so the agent
        resumes and the terminal_tool executes the command inline — the same
        flow as the CLI's synchronous input() approval.

        Supports multiple concurrent approvals (parallel subagents,
        execute_code).  ``/approve`` resolves the oldest pending command;
        ``/approve all`` resolves every pending command at once.

        Usage:
            /approve              — approve oldest pending command once
            /approve all          — approve ALL pending commands at once
            /approve session      — approve oldest + remember for session
            /approve all session  — approve all + remember for session
            /approve always       — approve oldest + remember permanently
            /approve all always   — approve all + remember permanently
        """
        source = event.source
        session_key = self._session_key_for_source(source)

        from tools.approval import (
            resolve_gateway_approval, has_blocking_approval,
        )

        if not has_blocking_approval(session_key):
            if session_key in self._pending_approvals:
                self._pending_approvals.pop(session_key)
                return "⚠️ Approval expired (agent is no longer waiting). Ask the agent to try again."
            return "No pending command to approve."

        # Parse args: support "all", "all session", "all always", "session", "always"
        args = event.get_command_args().strip().lower().split()
        resolve_all = "all" in args
        remaining = [a for a in args if a != "all"]

        if any(a in ("always", "permanent", "permanently") for a in remaining):
            choice = "always"
            scope_msg = " (pattern approved permanently)"
        elif any(a in ("session", "ses") for a in remaining):
            choice = "session"
            scope_msg = " (pattern approved for this session)"
        else:
            choice = "once"
            scope_msg = ""

        count = resolve_gateway_approval(session_key, choice, resolve_all=resolve_all)
        if not count:
            return "No pending command to approve."

        # Resume typing indicator — agent is about to continue processing.
        _adapter = self.adapters.get(source.platform)
        if _adapter:
            _adapter.resume_typing_for_chat(source.chat_id)

        count_msg = f" ({count} commands)" if count > 1 else ""
        logger.info("User approved %d dangerous command(s) via /approve%s", count, scope_msg)
        return f"✅ Command{'s' if count > 1 else ''} approved{scope_msg}{count_msg}. The agent is resuming..."

    async def _handle_deny_command(self, event: MessageEvent) -> str:
        """Handle /deny command — reject pending dangerous command(s).

        Signals blocked agent thread(s) with a 'deny' result so they receive
        a definitive BLOCKED message, same as the CLI deny flow.

        ``/deny`` denies the oldest; ``/deny all`` denies everything.
        """
        source = event.source
        session_key = self._session_key_for_source(source)

        from tools.approval import (
            resolve_gateway_approval, has_blocking_approval,
        )

        if not has_blocking_approval(session_key):
            if session_key in self._pending_approvals:
                self._pending_approvals.pop(session_key)
                return "❌ Command denied (approval was stale)."
            return "No pending command to deny."

        args = event.get_command_args().strip().lower()
        resolve_all = "all" in args

        count = resolve_gateway_approval(session_key, "deny", resolve_all=resolve_all)
        if not count:
            return "No pending command to deny."

        # Resume typing indicator — agent continues (with BLOCKED result).
        _adapter = self.adapters.get(source.platform)
        if _adapter:
            _adapter.resume_typing_for_chat(source.chat_id)

        count_msg = f" ({count} commands)" if count > 1 else ""
        logger.info("User denied %d dangerous command(s) via /deny", count)
        return f"❌ Command{'s' if count > 1 else ''} denied{count_msg}."

    # Platforms where /update is allowed.  ACP, API server, and webhooks are
    # programmatic interfaces that should not trigger system updates.
    _UPDATE_ALLOWED_PLATFORMS = frozenset({
        Platform.TELEGRAM, Platform.DISCORD, Platform.SLACK, Platform.WHATSAPP,
        Platform.SIGNAL, Platform.MATTERMOST, Platform.MATRIX,
        Platform.HOMEASSISTANT, Platform.EMAIL, Platform.SMS, Platform.DINGTALK,
        Platform.FEISHU, Platform.WECOM, Platform.WECOM_CALLBACK, Platform.WEIXIN, Platform.BLUEBUBBLES, Platform.LOCAL,
    })

    async def _handle_update_command(self, event: MessageEvent) -> str:
        """Handle /update command — update Hermes Agent to the latest version.

        Spawns ``hermes update`` in a detached session (via ``setsid``) so it
        survives the gateway restart that ``hermes update`` may trigger. Marker
        files are written so either the current gateway process or the next one
        can notify the user when the update finishes.
        """
        import json
        import shutil
        import subprocess
        from datetime import datetime
        from hermes_cli.config import is_managed, format_managed_message

        # Block non-messaging platforms (API server, webhooks, ACP)
        platform = event.source.platform
        if platform not in self._UPDATE_ALLOWED_PLATFORMS:
            return "✗ /update is only available from messaging platforms. Run `hermes update` from the terminal."

        if is_managed():
            return f"✗ {format_managed_message('update Hermes Agent')}"

        project_root = Path(__file__).parent.parent.resolve()
        git_dir = project_root / '.git'

        if not git_dir.exists():
            return "✗ Not a git repository — cannot update."

        hermes_cmd = _resolve_hermes_bin()
        if not hermes_cmd:
            return (
                "✗ Could not locate the `hermes` command. "
                "Hermes is running, but the update command could not find the "
                "executable on PATH or via the current Python interpreter. "
                "Try running `hermes update` manually in your terminal."
            )

        pending_path = _hermes_home / ".update_pending.json"
        output_path = _hermes_home / ".update_output.txt"
        exit_code_path = _hermes_home / ".update_exit_code"
        session_key = self._session_key_for_source(event.source)
        pending = {
            "platform": event.source.platform.value,
            "chat_id": event.source.chat_id,
            "user_id": event.source.user_id,
            "session_key": session_key,
            "timestamp": datetime.now().isoformat(),
        }
        _tmp_pending = pending_path.with_suffix(".tmp")
        _tmp_pending.write_text(json.dumps(pending))
        _tmp_pending.replace(pending_path)
        exit_code_path.unlink(missing_ok=True)

        # Spawn `hermes update --gateway` detached so it survives gateway restart.
        # --gateway enables file-based IPC for interactive prompts (stash
        # restore, config migration) so the gateway can forward them to the
        # user instead of silently skipping them.
        # Use setsid for portable session detach (works under system services
        # where systemd-run --user fails due to missing D-Bus session).
        # PYTHONUNBUFFERED ensures output is flushed line-by-line so the
        # gateway can stream it to the messenger in near-real-time.
        hermes_cmd_str = " ".join(shlex.quote(part) for part in hermes_cmd)
        update_cmd = (
            f"PYTHONUNBUFFERED=1 {hermes_cmd_str} update --gateway"
            f" > {shlex.quote(str(output_path))} 2>&1; "
            f"status=$?; printf '%s' \"$status\" > {shlex.quote(str(exit_code_path))}"
        )
        try:
            setsid_bin = shutil.which("setsid")
            if setsid_bin:
                # Preferred: setsid creates a new session, fully detached
                subprocess.Popen(
                    [setsid_bin, "bash", "-c", update_cmd],
                    stdout=subprocess.DEVNULL,
                    stderr=subprocess.DEVNULL,
                    start_new_session=True,
                )
            else:
                # Fallback: start_new_session=True calls os.setsid() in child
                subprocess.Popen(
                    ["bash", "-c", update_cmd],
                    stdout=subprocess.DEVNULL,
                    stderr=subprocess.DEVNULL,
                    start_new_session=True,
                )
        except Exception as e:
            pending_path.unlink(missing_ok=True)
            exit_code_path.unlink(missing_ok=True)
            return f"✗ Failed to start update: {e}"

        self._schedule_update_notification_watch()
        return "⚕ Starting Hermes update… I'll stream progress here."

    def _schedule_update_notification_watch(self) -> None:
        """Ensure a background task is watching for update completion."""
        existing_task = getattr(self, "_update_notification_task", None)
        if existing_task and not existing_task.done():
            return

        try:
            self._update_notification_task = asyncio.create_task(
                self._watch_update_progress()
            )
        except RuntimeError:
            logger.debug("Skipping update notification watcher: no running event loop")

    async def _watch_update_progress(
        self,
        poll_interval: float = 2.0,
        stream_interval: float = 4.0,
        timeout: float = 1800.0,
    ) -> None:
        """Watch ``hermes update --gateway``, streaming output + forwarding prompts.

        Polls ``.update_output.txt`` for new content and sends chunks to the
        user periodically.  Detects ``.update_prompt.json`` (written by the
        update process when it needs user input) and forwards the prompt to
        the messenger.  The user's next message is intercepted by
        ``_handle_message`` and written to ``.update_response``.
        """
        import json
        import re as _re

        pending_path = _hermes_home / ".update_pending.json"
        claimed_path = _hermes_home / ".update_pending.claimed.json"
        output_path = _hermes_home / ".update_output.txt"
        exit_code_path = _hermes_home / ".update_exit_code"
        prompt_path = _hermes_home / ".update_prompt.json"

        loop = asyncio.get_running_loop()
        deadline = loop.time() + timeout

        # Resolve the adapter and chat_id for sending messages
        adapter = None
        chat_id = None
        session_key = None
        for path in (claimed_path, pending_path):
            if path.exists():
                try:
                    pending = json.loads(path.read_text())
                    platform_str = pending.get("platform")
                    chat_id = pending.get("chat_id")
                    session_key = pending.get("session_key")
                    if platform_str and chat_id:
                        platform = Platform(platform_str)
                        adapter = self.adapters.get(platform)
                        # Fallback session key if not stored (old pending files)
                        if not session_key:
                            session_key = f"{platform_str}:{chat_id}"
                    break
                except Exception:
                    pass

        if not adapter or not chat_id:
            logger.warning("Update watcher: cannot resolve adapter/chat_id, falling back to completion-only")
            # Fall back to old behavior: wait for exit code and send final notification
            while (pending_path.exists() or claimed_path.exists()) and loop.time() < deadline:
                if exit_code_path.exists():
                    await self._send_update_notification()
                    return
                await asyncio.sleep(poll_interval)
            if (pending_path.exists() or claimed_path.exists()) and not exit_code_path.exists():
                exit_code_path.write_text("124")
                await self._send_update_notification()
            return

        def _strip_ansi(text: str) -> str:
            return _re.sub(r'\x1b\[[0-9;]*[A-Za-z]', '', text)

        bytes_sent = 0
        last_stream_time = loop.time()
        buffer = ""

        async def _flush_buffer() -> None:
            """Send buffered output to the user."""
            nonlocal buffer, last_stream_time
            if not buffer.strip():
                buffer = ""
                return
            # Chunk to fit message limits (Telegram: 4096, others: generous)
            clean = _strip_ansi(buffer).strip()
            buffer = ""
            last_stream_time = loop.time()
            if not clean:
                return
            # Split into chunks if too long
            max_chunk = 3500
            chunks = [clean[i:i + max_chunk] for i in range(0, len(clean), max_chunk)]
            for chunk in chunks:
                try:
                    await adapter.send(chat_id, f"```\n{chunk}\n```")
                except Exception as e:
                    logger.debug("Update stream send failed: %s", e)

        while loop.time() < deadline:
            # Check for completion
            if exit_code_path.exists():
                # Read any remaining output
                if output_path.exists():
                    try:
                        content = output_path.read_text()
                        if len(content) > bytes_sent:
                            buffer += content[bytes_sent:]
                            bytes_sent = len(content)
                    except OSError:
                        pass
                await _flush_buffer()

                # Send final status
                try:
                    exit_code_raw = exit_code_path.read_text().strip() or "1"
                    exit_code = int(exit_code_raw)
                    if exit_code == 0:
                        await adapter.send(chat_id, "✅ Hermes update finished.")
                    else:
                        await adapter.send(chat_id, "❌ Hermes update failed (exit code {}).".format(exit_code))
                    logger.info("Update finished (exit=%s), notified %s", exit_code, session_key)
                except Exception as e:
                    logger.warning("Update final notification failed: %s", e)

                # Cleanup
                for p in (pending_path, claimed_path, output_path,
                          exit_code_path, prompt_path):
                    p.unlink(missing_ok=True)
                (_hermes_home / ".update_response").unlink(missing_ok=True)
                self._update_prompt_pending.pop(session_key, None)
                return

            # Check for new output
            if output_path.exists():
                try:
                    content = output_path.read_text()
                    if len(content) > bytes_sent:
                        buffer += content[bytes_sent:]
                        bytes_sent = len(content)
                except OSError:
                    pass

            # Flush buffer periodically
            if buffer.strip() and (loop.time() - last_stream_time) >= stream_interval:
                await _flush_buffer()

            # Check for prompts
            if prompt_path.exists() and session_key:
                try:
                    prompt_data = json.loads(prompt_path.read_text())
                    prompt_text = prompt_data.get("prompt", "")
                    default = prompt_data.get("default", "")
                    if prompt_text:
                        # Flush any buffered output first so the user sees
                        # context before the prompt
                        await _flush_buffer()
                        # Try platform-native buttons first (Discord, Telegram)
                        sent_buttons = False
                        if getattr(type(adapter), "send_update_prompt", None) is not None:
                            try:
                                await adapter.send_update_prompt(
                                    chat_id=chat_id,
                                    prompt=prompt_text,
                                    default=default,
                                    session_key=session_key,
                                )
                                sent_buttons = True
                            except Exception as btn_err:
                                logger.debug("Button-based update prompt failed: %s", btn_err)
                        if not sent_buttons:
                            default_hint = f" (default: {default})" if default else ""
                            await adapter.send(
                                chat_id,
                                f"⚕ **Update needs your input:**\n\n"
                                f"{prompt_text}{default_hint}\n\n"
                                f"Reply `/approve` (yes) or `/deny` (no), "
                                f"or type your answer directly."
                            )
                        self._update_prompt_pending[session_key] = True
                        logger.info("Forwarded update prompt to %s: %s", session_key, prompt_text[:80])
                except (json.JSONDecodeError, OSError) as e:
                    logger.debug("Failed to read update prompt: %s", e)

            await asyncio.sleep(poll_interval)

        # Timeout
        if not exit_code_path.exists():
            logger.warning("Update watcher timed out after %.0fs", timeout)
            exit_code_path.write_text("124")
            await _flush_buffer()
            try:
                await adapter.send(chat_id, "❌ Hermes update timed out after 30 minutes.")
            except Exception:
                pass
            for p in (pending_path, claimed_path, output_path,
                      exit_code_path, prompt_path):
                p.unlink(missing_ok=True)
            (_hermes_home / ".update_response").unlink(missing_ok=True)
            self._update_prompt_pending.pop(session_key, None)

    async def _send_update_notification(self) -> bool:
        """If an update finished, notify the user.

        Returns False when the update is still running so a caller can retry
        later. Returns True after a definitive send/skip decision.

        This is the legacy notification path used when the streaming watcher
        cannot resolve the adapter (e.g. after a gateway restart where the
        platform hasn't reconnected yet).
        """
        import json
        import re as _re

        pending_path = _hermes_home / ".update_pending.json"
        claimed_path = _hermes_home / ".update_pending.claimed.json"
        output_path = _hermes_home / ".update_output.txt"
        exit_code_path = _hermes_home / ".update_exit_code"

        if not pending_path.exists() and not claimed_path.exists():
            return False

        cleanup = True
        active_pending_path = claimed_path
        try:
            if pending_path.exists():
                try:
                    pending_path.replace(claimed_path)
                except FileNotFoundError:
                    if not claimed_path.exists():
                        return True
            elif not claimed_path.exists():
                return True

            pending = json.loads(claimed_path.read_text())
            platform_str = pending.get("platform")
            chat_id = pending.get("chat_id")

            if not exit_code_path.exists():
                logger.info("Update notification deferred: update still running")
                cleanup = False
                active_pending_path = pending_path
                claimed_path.replace(pending_path)
                return False

            exit_code_raw = exit_code_path.read_text().strip() or "1"
            exit_code = int(exit_code_raw)

            # Read the captured update output
            output = ""
            if output_path.exists():
                output = output_path.read_text()

            # Resolve adapter
            platform = Platform(platform_str)
            adapter = self.adapters.get(platform)

            if adapter and chat_id:
                # Strip ANSI escape codes for clean display
                output = _re.sub(r'\x1b\[[0-9;]*m', '', output).strip()
                if output:
                    if len(output) > 3500:
                        output = "…" + output[-3500:]
                    if exit_code == 0:
                        msg = f"✅ Hermes update finished.\n\n```\n{output}\n```"
                    else:
                        msg = f"❌ Hermes update failed.\n\n```\n{output}\n```"
                else:
                    if exit_code == 0:
                        msg = "✅ Hermes update finished successfully."
                    else:
                        msg = "❌ Hermes update failed. Check the gateway logs or run `hermes update` manually for details."
                await adapter.send(chat_id, msg)
                logger.info(
                    "Sent post-update notification to %s:%s (exit=%s)",
                    platform_str,
                    chat_id,
                    exit_code,
                )
        except Exception as e:
            logger.warning("Post-update notification failed: %s", e)
        finally:
            if cleanup:
                active_pending_path.unlink(missing_ok=True)
                claimed_path.unlink(missing_ok=True)
                output_path.unlink(missing_ok=True)
                exit_code_path.unlink(missing_ok=True)

        return True

    def _set_session_env(self, context: SessionContext) -> list:
        """Set session context variables for the current async task.

        Uses ``contextvars`` instead of ``os.environ`` so that concurrent
        gateway messages cannot overwrite each other's session state.

        Returns a list of reset tokens; pass them to ``_clear_session_env``
        in a ``finally`` block.
        """
        from gateway.session_context import set_session_vars
        return set_session_vars(
            platform=context.source.platform.value,
            chat_id=context.source.chat_id,
            chat_name=context.source.chat_name or "",
            thread_id=str(context.source.thread_id) if context.source.thread_id else "",
            user_id=str(context.source.user_id) if context.source.user_id else "",
            user_name=str(context.source.user_name) if context.source.user_name else "",
            session_key=context.session_key,
        )

    def _clear_session_env(self, tokens: list) -> None:
        """Restore session context variables to their pre-handler values."""
        from gateway.session_context import clear_session_vars
        clear_session_vars(tokens)
    
    async def _enrich_message_with_vision(
        self,
        user_text: str,
        image_paths: List[str],
    ) -> str:
        """
        Auto-analyze user-attached images with the vision tool and prepend
        the descriptions to the message text.

        Each image is analyzed with a general-purpose prompt.  The resulting
        description *and* the local cache path are injected so the model can:
          1. Immediately understand what the user sent (no extra tool call).
          2. Re-examine the image with vision_analyze if it needs more detail.

        Args:
            user_text:   The user's original caption / message text.
            image_paths: List of local file paths to cached images.

        Returns:
            The enriched message string with vision descriptions prepended.
        """
        from tools.vision_tools import vision_analyze_tool
        import json as _json

        analysis_prompt = (
            "Describe everything visible in this image in thorough detail. "
            "Include any text, code, data, objects, people, layout, colors, "
            "and any other notable visual information."
        )

        enriched_parts = []
        for path in image_paths:
            try:
                logger.debug("Auto-analyzing user image: %s", path)
                result_json = await vision_analyze_tool(
                    image_url=path,
                    user_prompt=analysis_prompt,
                )
                result = _json.loads(result_json)
                if result.get("success"):
                    description = result.get("analysis", "")
                    enriched_parts.append(
                        f"[The user sent an image~ Here's what I can see:\n{description}]\n"
                        f"[If you need a closer look, use vision_analyze with "
                        f"image_url: {path} ~]"
                    )
                else:
                    enriched_parts.append(
                        "[The user sent an image but I couldn't quite see it "
                        "this time (>_<) You can try looking at it yourself "
                        f"with vision_analyze using image_url: {path}]"
                    )
            except Exception as e:
                logger.error("Vision auto-analysis error: %s", e)
                enriched_parts.append(
                    f"[The user sent an image but something went wrong when I "
                    f"tried to look at it~ You can try examining it yourself "
                    f"with vision_analyze using image_url: {path}]"
                )

        # Combine: vision descriptions first, then the user's original text
        if enriched_parts:
            prefix = "\n\n".join(enriched_parts)
            if user_text:
                return f"{prefix}\n\n{user_text}"
            return prefix
        return user_text

    async def _enrich_message_with_transcription(
        self,
        user_text: str,
        audio_paths: List[str],
    ) -> str:
        """
        Auto-transcribe user voice/audio messages using the configured STT provider
        and prepend the transcript to the message text.

        Args:
            user_text:   The user's original caption / message text.
            audio_paths: List of local file paths to cached audio files.

        Returns:
            The enriched message string with transcriptions prepended.
        """
        if not getattr(self.config, "stt_enabled", True):
            disabled_note = "[The user sent voice message(s), but transcription is disabled in config."
            if self._has_setup_skill():
                disabled_note += (
                    " You have a skill called hermes-agent-setup that can help "
                    "users configure Hermes features including voice, tools, and more."
                )
            disabled_note += "]"
            if user_text:
                return f"{disabled_note}\n\n{user_text}"
            return disabled_note

        from tools.transcription_tools import transcribe_audio
        import asyncio

        enriched_parts = []
        for path in audio_paths:
            try:
                logger.debug("Transcribing user voice: %s", path)
                result = await asyncio.to_thread(transcribe_audio, path)
                if result["success"]:
                    transcript = result["transcript"]
                    enriched_parts.append(
                        f'[The user sent a voice message~ '
                        f'Here\'s what they said: "{transcript}"]'
                    )
                else:
                    error = result.get("error", "unknown error")
                    if (
                        "No STT provider" in error
                        or error.startswith("Neither VOICE_TOOLS_OPENAI_KEY nor OPENAI_API_KEY is set")
                    ):
                        _no_stt_note = (
                            "[The user sent a voice message but I can't listen "
                            "to it right now — no STT provider is configured. "
                            "A direct message has already been sent to the user "
                            "with setup instructions."
                        )
                        if self._has_setup_skill():
                            _no_stt_note += (
                                " You have a skill called hermes-agent-setup "
                                "that can help users configure Hermes features "
                                "including voice, tools, and more."
                            )
                        _no_stt_note += "]"
                        enriched_parts.append(_no_stt_note)
                    else:
                        enriched_parts.append(
                            "[The user sent a voice message but I had trouble "
                            f"transcribing it~ ({error})]"
                        )
            except Exception as e:
                logger.error("Transcription error: %s", e)
                enriched_parts.append(
                    "[The user sent a voice message but something went wrong "
                    "when I tried to listen to it~ Let them know!]"
                )

        if enriched_parts:
            prefix = "\n\n".join(enriched_parts)
            # Strip the empty-content placeholder from the Discord adapter
            # when we successfully transcribed the audio — it's redundant.
            _placeholder = "(The user sent a message with no text content)"
            if user_text and user_text.strip() == _placeholder:
                return prefix
            if user_text:
                return f"{prefix}\n\n{user_text}"
            return prefix
        return user_text

    async def _inject_watch_notification(self, synth_text: str, original_event) -> None:
        """Inject a watch-pattern notification as a synthetic message event.

        Uses the source from the original user event to route the notification
        back to the correct chat/adapter.
        """
        source = getattr(original_event, "source", None)
        if not source:
            return
        platform_name = source.platform.value if hasattr(source.platform, "value") else str(source.platform)
        adapter = None
        for p, a in self.adapters.items():
            if p.value == platform_name:
                adapter = a
                break
        if not adapter:
            return
        try:
            from gateway.platforms.base import MessageEvent, MessageType
            synth_event = MessageEvent(
                text=synth_text,
                message_type=MessageType.TEXT,
                source=source,
                internal=True,
            )
            logger.info("Watch pattern notification — injecting for %s", platform_name)
            await adapter.handle_message(synth_event)
        except Exception as e:
            logger.error("Watch notification injection error: %s", e)

    async def _run_process_watcher(self, watcher: dict) -> None:
        """
        Periodically check a background process and push updates to the user.

        Runs as an asyncio task. Stays silent when nothing changed.
        Auto-removes when the process exits or is killed.

        Notification mode (from ``display.background_process_notifications``):
          - ``all``    — running-output updates + final message
          - ``result`` — final completion message only
          - ``error``  — final message only when exit code != 0
          - ``off``    — no messages at all
        """
        from tools.process_registry import process_registry

        session_id = watcher["session_id"]
        interval = watcher["check_interval"]
        session_key = watcher.get("session_key", "")
        platform_name = watcher.get("platform", "")
        chat_id = watcher.get("chat_id", "")
        thread_id = watcher.get("thread_id", "")
        user_id = watcher.get("user_id", "")
        user_name = watcher.get("user_name", "")
        agent_notify = watcher.get("notify_on_complete", False)
        notify_mode = self._load_background_notifications_mode()

        logger.debug("Process watcher started: %s (every %ss, notify=%s, agent_notify=%s)",
                      session_id, interval, notify_mode, agent_notify)

        if notify_mode == "off" and not agent_notify:
            # Still wait for the process to exit so we can log it, but don't
            # push any messages to the user.
            while True:
                await asyncio.sleep(interval)
                session = process_registry.get(session_id)
                if session is None or session.exited:
                    break
            logger.debug("Process watcher ended (silent): %s", session_id)
            return

        last_output_len = 0
        while True:
            await asyncio.sleep(interval)

            session = process_registry.get(session_id)
            if session is None:
                break

            current_output_len = len(session.output_buffer)
            has_new_output = current_output_len > last_output_len
            last_output_len = current_output_len

            if session.exited:
                # --- Agent-triggered completion: inject synthetic message ---
                # Skip if the agent already consumed the result via wait/poll/log
                from tools.process_registry import process_registry as _pr_check
                if agent_notify and not _pr_check.is_completion_consumed(session_id):
                    from tools.ansi_strip import strip_ansi
                    _out = strip_ansi(session.output_buffer[-2000:]) if session.output_buffer else ""
                    synth_text = (
                        f"[SYSTEM: Background process {session_id} completed "
                        f"(exit code {session.exit_code}).\n"
                        f"Command: {session.command}\n"
                        f"Output:\n{_out}]"
                    )
                    adapter = None
                    for p, a in self.adapters.items():
                        if p.value == platform_name:
                            adapter = a
                            break
                    if adapter and chat_id:
                        try:
                            from gateway.platforms.base import MessageEvent, MessageType
                            from gateway.session import SessionSource
                            from gateway.config import Platform
                            _platform_enum = Platform(platform_name)
                            _source = SessionSource(
                                platform=_platform_enum,
                                chat_id=chat_id,
                                thread_id=thread_id or None,
                                user_id=user_id or None,
                                user_name=user_name or None,
                            )
                            synth_event = MessageEvent(
                                text=synth_text,
                                message_type=MessageType.TEXT,
                                source=_source,
                                internal=True,
                            )
                            logger.info(
                                "Process %s finished — injecting agent notification for session %s",
                                session_id, session_key,
                            )
                            await adapter.handle_message(synth_event)
                        except Exception as e:
                            logger.error("Agent notify injection error: %s", e)
                    break

                # --- Normal text-only notification ---
                # Decide whether to notify based on mode
                should_notify = (
                    notify_mode in ("all", "result")
                    or (notify_mode == "error" and session.exit_code not in (0, None))
                )
                if should_notify:
                    new_output = session.output_buffer[-1000:] if session.output_buffer else ""
                    message_text = (
                        f"[Background process {session_id} finished with exit code {session.exit_code}~ "
                        f"Here's the final output:\n{new_output}]"
                    )
                    adapter = None
                    for p, a in self.adapters.items():
                        if p.value == platform_name:
                            adapter = a
                            break
                    if adapter and chat_id:
                        try:
                            send_meta = {"thread_id": thread_id} if thread_id else None
                            await adapter.send(chat_id, message_text, metadata=send_meta)
                        except Exception as e:
                            logger.error("Watcher delivery error: %s", e)
                break

            elif has_new_output and notify_mode == "all" and not agent_notify:
                # New output available -- deliver status update (only in "all" mode)
                # Skip periodic updates for agent_notify watchers (they only care about completion)
                new_output = session.output_buffer[-500:] if session.output_buffer else ""
                message_text = (
                    f"[Background process {session_id} is still running~ "
                    f"New output:\n{new_output}]"
                )
                adapter = None
                for p, a in self.adapters.items():
                    if p.value == platform_name:
                        adapter = a
                        break
                if adapter and chat_id:
                    try:
                        send_meta = {"thread_id": thread_id} if thread_id else None
                        await adapter.send(chat_id, message_text, metadata=send_meta)
                    except Exception as e:
                        logger.error("Watcher delivery error: %s", e)

        logger.debug("Process watcher ended: %s", session_id)

    _MAX_INTERRUPT_DEPTH = 3  # Cap recursive interrupt handling (#816)

    @staticmethod
    def _agent_config_signature(
        model: str,
        runtime: dict,
        enabled_toolsets: list,
        ephemeral_prompt: str,
    ) -> str:
        """Compute a stable string key from agent config values.

        When this signature changes between messages, the cached AIAgent is
        discarded and rebuilt.  When it stays the same, the cached agent is
        reused — preserving the frozen system prompt and tool schemas for
        prompt cache hits.
        """
        import hashlib, json as _j

        # Fingerprint the FULL credential string instead of using a short
        # prefix. OAuth/JWT-style tokens frequently share a common prefix
        # (e.g. "eyJhbGci"), which can cause false cache hits across auth
        # switches if only the first few characters are considered.
        _api_key = str(runtime.get("api_key", "") or "")
        _api_key_fingerprint = hashlib.sha256(_api_key.encode()).hexdigest() if _api_key else ""

        blob = _j.dumps(
            [
                model,
                _api_key_fingerprint,
                runtime.get("base_url", ""),
                runtime.get("provider", ""),
                runtime.get("api_mode", ""),
                sorted(enabled_toolsets) if enabled_toolsets else [],
                # reasoning_config excluded — it's set per-message on the
                # cached agent and doesn't affect system prompt or tools.
                ephemeral_prompt or "",
            ],
            sort_keys=True,
            default=str,
        )
        return hashlib.sha256(blob.encode()).hexdigest()[:16]

    def _apply_session_model_override(
        self, session_key: str, model: str, runtime_kwargs: dict
    ) -> tuple:
        """Apply /model session overrides if present, returning (model, runtime_kwargs).

        The gateway /model command stores per-session overrides in
        ``_session_model_overrides``.  These must take precedence over
        config.yaml defaults so the switched model is actually used for
        subsequent messages.  Fields with ``None`` values are skipped so
        partial overrides don't clobber valid config defaults.
        """
        override = self._session_model_overrides.get(session_key)
        if not override:
            return model, runtime_kwargs
        model = override.get("model", model)
        for key in ("provider", "api_key", "base_url", "api_mode"):
            val = override.get(key)
            if val is not None:
                runtime_kwargs[key] = val
        return model, runtime_kwargs

    def _is_intentional_model_switch(self, session_key: str, agent_model: str) -> bool:
        """Return True if *agent_model* matches an active /model session override."""
        override = self._session_model_overrides.get(session_key)
        return override is not None and override.get("model") == agent_model

    def _evict_cached_agent(self, session_key: str) -> None:
        """Remove a cached agent for a session (called on /new, /model, etc)."""
        _lock = getattr(self, "_agent_cache_lock", None)
        if _lock:
            with _lock:
                self._agent_cache.pop(session_key, None)

    async def _run_agent(
        self,
        message: str,
        context_prompt: str,
        history: List[Dict[str, Any]],
        source: SessionSource,
        session_id: str,
        session_key: str = None,
        _interrupt_depth: int = 0,
        event_message_id: Optional[str] = None,
    ) -> Dict[str, Any]:
        """
        Run the agent with the given message and context.
        
        Returns the full result dict from run_conversation, including:
          - "final_response": str (the text to send back)
          - "messages": list (full conversation including tool calls)
          - "api_calls": int
          - "completed": bool
        
        This is run in a thread pool to not block the event loop.
        Supports interruption via new messages.
        """
        from run_agent import AIAgent
        import queue
        
        user_config = _load_gateway_config()
        platform_key = _platform_config_key(source.platform)

        from hermes_cli.tools_config import _get_platform_tools
        enabled_toolsets = sorted(_get_platform_tools(user_config, platform_key))

        display_config = user_config.get("display", {})
        if not isinstance(display_config, dict):
            display_config = {}

        # Per-platform display settings — resolve via display_config module
        # which checks display.platforms.<platform>.<key> first, then
        # display.<key> global, then built-in platform defaults.
        from gateway.display_config import resolve_display_setting

        # Apply tool preview length config (0 = no limit)
        try:
            from agent.display import set_tool_preview_max_len
            _tpl = resolve_display_setting(user_config, platform_key, "tool_preview_length", 0)
            set_tool_preview_max_len(int(_tpl) if _tpl else 0)
        except Exception:
            pass

        # Tool progress mode — resolved per-platform with env var fallback
        _resolved_tp = resolve_display_setting(user_config, platform_key, "tool_progress")
        progress_mode = (
            _resolved_tp
            or os.getenv("HERMES_TOOL_PROGRESS_MODE")
            or "all"
        )
        # Disable tool progress for webhooks - they don't support message editing,
        # so each progress line would be sent as a separate message.
        from gateway.config import Platform
        tool_progress_enabled = progress_mode != "off" and source.platform != Platform.WEBHOOK
        # Natural assistant status messages are intentionally independent from
        # tool progress and token streaming. Users can keep tool_progress quiet
        # in chat platforms while opting into concise mid-turn updates.
        interim_assistant_messages_enabled = (
            source.platform != Platform.WEBHOOK
            and is_truthy_value(
                display_config.get("interim_assistant_messages"),
                default=True,
            )
        )
        
        # Queue for progress messages (thread-safe)
        progress_queue = queue.Queue() if tool_progress_enabled else None
        last_tool = [None]  # Mutable container for tracking in closure
        last_progress_msg = [None]  # Track last message for dedup
        repeat_count = [0]  # How many times the same message repeated
        
        def progress_callback(event_type: str, tool_name: str = None, preview: str = None, args: dict = None, **kwargs):
            """Callback invoked by agent on tool lifecycle events."""
            if not progress_queue:
                return

            # Only act on tool.started events (ignore tool.completed, reasoning.available, etc.)
            if event_type not in ("tool.started",):
                return

            # "new" mode: only report when tool changes
            if progress_mode == "new" and tool_name == last_tool[0]:
                return
            last_tool[0] = tool_name
            
            # Build progress message with primary argument preview
            from agent.display import get_tool_emoji
            emoji = get_tool_emoji(tool_name, default="⚙️")
            
            # Verbose mode: show detailed arguments, respects tool_preview_length
            if progress_mode == "verbose":
                if args:
                    from agent.display import get_tool_preview_max_len
                    _pl = get_tool_preview_max_len()
                    import json as _json
                    args_str = _json.dumps(args, ensure_ascii=False, default=str)
                    _cap = _pl if _pl > 0 else 200
                    if len(args_str) > _cap:
                        args_str = args_str[:_cap - 3] + "..."
                    msg = f"{emoji} {tool_name}({list(args.keys())})\n{args_str}"
                elif preview:
                    msg = f"{emoji} {tool_name}: \"{preview}\""
                else:
                    msg = f"{emoji} {tool_name}..."
                progress_queue.put(msg)
                return
            
            # "all" / "new" modes: short preview, respects tool_preview_length
            # config (defaults to 40 chars when unset to keep gateway messages
            # compact — unlike CLI spinners, these persist as permanent messages).
            if preview:
                from agent.display import get_tool_preview_max_len
                _pl = get_tool_preview_max_len()
                _cap = _pl if _pl > 0 else 40
                if len(preview) > _cap:
                    preview = preview[:_cap - 3] + "..."
                msg = f"{emoji} {tool_name}: \"{preview}\""
            else:
                msg = f"{emoji} {tool_name}..."
            
            # Dedup: collapse consecutive identical progress messages.
            # Common with execute_code where models iterate with the same
            # code (same boilerplate imports → identical previews).
            if msg == last_progress_msg[0]:
                repeat_count[0] += 1
                # Update the last line in progress_lines with a counter
                # via a special "dedup" queue message.
                progress_queue.put(("__dedup__", msg, repeat_count[0]))
                return
            last_progress_msg[0] = msg
            repeat_count[0] = 0
            
            progress_queue.put(msg)
        
        # Background task to send progress messages
        # Accumulates tool lines into a single message that gets edited.
        #
        # Threading metadata is platform-specific:
        # - Slack DM threading needs event_message_id fallback (reply thread)
        # - Telegram uses message_thread_id only for forum topics; passing a
        #   normal DM/group message id as thread_id causes send failures
        # - Other platforms should use explicit source.thread_id only
        if source.platform == Platform.SLACK:
            _progress_thread_id = source.thread_id or event_message_id
        else:
            _progress_thread_id = source.thread_id
        _progress_metadata = {"thread_id": _progress_thread_id} if _progress_thread_id else None

        async def send_progress_messages():
            if not progress_queue:
                return

            adapter = self.adapters.get(source.platform)
            if not adapter:
                return

            # Skip tool progress for platforms that don't support message
            # editing (e.g. iMessage/BlueBubbles) — each progress update
            # would become a separate message bubble, which is noisy.
            from gateway.platforms.base import BasePlatformAdapter as _BaseAdapter
            if type(adapter).edit_message is _BaseAdapter.edit_message:
                while not progress_queue.empty():
                    try:
                        progress_queue.get_nowait()
                    except Exception:
                        break
                return

            progress_lines = []      # Accumulated tool lines
            progress_msg_id = None   # ID of the progress message to edit
            can_edit = True          # False once an edit fails (platform doesn't support it)
            _last_edit_ts = 0.0      # Throttle edits to avoid Telegram flood control
            _PROGRESS_EDIT_INTERVAL = 1.5  # Minimum seconds between edits

            while True:
                try:
                    raw = progress_queue.get_nowait()

                    # Handle dedup messages: update last line with repeat counter
                    if isinstance(raw, tuple) and len(raw) == 3 and raw[0] == "__dedup__":
                        _, base_msg, count = raw
                        if progress_lines:
                            progress_lines[-1] = f"{base_msg} (×{count + 1})"
                        msg = progress_lines[-1] if progress_lines else base_msg
                    else:
                        msg = raw
                        progress_lines.append(msg)

                    # Throttle edits: batch rapid tool updates into fewer
                    # API calls to avoid hitting Telegram flood control.
                    # (grammY auto-retry pattern: proactively rate-limit
                    # instead of reacting to 429s.)
                    _now = time.monotonic()
                    _remaining = _PROGRESS_EDIT_INTERVAL - (_now - _last_edit_ts)
                    if _remaining > 0:
                        # Wait out the throttle interval, then loop back to
                        # drain any additional queued messages before sending
                        # a single batched edit.
                        await asyncio.sleep(_remaining)
                        continue

                    if can_edit and progress_msg_id is not None:
                        # Try to edit the existing progress message
                        full_text = "\n".join(progress_lines)
                        result = await adapter.edit_message(
                            chat_id=source.chat_id,
                            message_id=progress_msg_id,
                            content=full_text,
                        )
                        if not result.success:
                            _err = (getattr(result, "error", "") or "").lower()
                            if "flood" in _err or "retry after" in _err:
                                # Flood control hit — disable further edits,
                                # switch to sending new messages only for
                                # important updates.  Don't block 23s.
                                logger.info(
                                    "[%s] Progress edits disabled due to flood control",
                                    adapter.name,
                                )
                            can_edit = False
                            await adapter.send(chat_id=source.chat_id, content=msg, metadata=_progress_metadata)
                    else:
                        if can_edit:
                            # First tool: send all accumulated text as new message
                            full_text = "\n".join(progress_lines)
                            result = await adapter.send(chat_id=source.chat_id, content=full_text, metadata=_progress_metadata)
                        else:
                            # Editing unsupported: send just this line
                            result = await adapter.send(chat_id=source.chat_id, content=msg, metadata=_progress_metadata)
                        if result.success and result.message_id:
                            progress_msg_id = result.message_id

                    _last_edit_ts = time.monotonic()

                    # Restore typing indicator
                    await asyncio.sleep(0.3)
                    await adapter.send_typing(source.chat_id, metadata=_progress_metadata)

                except queue.Empty:
                    await asyncio.sleep(0.3)
                except asyncio.CancelledError:
                    # Drain remaining queued messages
                    while not progress_queue.empty():
                        try:
                            raw = progress_queue.get_nowait()
                            if isinstance(raw, tuple) and len(raw) == 3 and raw[0] == "__dedup__":
                                _, base_msg, count = raw
                                if progress_lines:
                                    progress_lines[-1] = f"{base_msg} (×{count + 1})"
                            else:
                                progress_lines.append(raw)
                        except Exception:
                            break
                    # Final edit with all remaining tools (only if editing works)
                    if can_edit and progress_lines and progress_msg_id:
                        full_text = "\n".join(progress_lines)
                        try:
                            await adapter.edit_message(
                                chat_id=source.chat_id,
                                message_id=progress_msg_id,
                                content=full_text,
                            )
                        except Exception:
                            pass
                    return
                except Exception as e:
                    logger.error("Progress message error: %s", e)
                    await asyncio.sleep(1)
        
        # We need to share the agent instance for interrupt support
        agent_holder = [None]  # Mutable container for the agent instance
        result_holder = [None]  # Mutable container for the result
        tools_holder = [None]   # Mutable container for the tool definitions
        stream_consumer_holder = [None]  # Mutable container for stream consumer
        
        # Bridge sync step_callback → async hooks.emit for agent:step events
        _loop_for_step = asyncio.get_event_loop()
        _hooks_ref = self.hooks

        def _step_callback_sync(iteration: int, prev_tools: list) -> None:
            try:
                # prev_tools may be list[str] or list[dict] with "name"/"result"
                # keys.  Normalise to keep "tool_names" backward-compatible for
                # user-authored hooks that do ', '.join(tool_names)'.
                _names: list[str] = []
                for _t in (prev_tools or []):
                    if isinstance(_t, dict):
                        _names.append(_t.get("name") or "")
                    else:
                        _names.append(str(_t))
                asyncio.run_coroutine_threadsafe(
                    _hooks_ref.emit("agent:step", {
                        "platform": source.platform.value if source.platform else "",
                        "user_id": source.user_id,
                        "session_id": session_id,
                        "iteration": iteration,
                        "tool_names": _names,
                        "tools": prev_tools,
                    }),
                    _loop_for_step,
                )
            except Exception as _e:
                logger.debug("agent:step hook error: %s", _e)

        # Bridge sync status_callback → async adapter.send for context pressure
        _status_adapter = self.adapters.get(source.platform)
        _status_chat_id = source.chat_id
        _status_thread_metadata = {"thread_id": _progress_thread_id} if _progress_thread_id else None

        def _status_callback_sync(event_type: str, message: str) -> None:
            if not _status_adapter:
                return
            try:
                asyncio.run_coroutine_threadsafe(
                    _status_adapter.send(
                        _status_chat_id,
                        message,
                        metadata=_status_thread_metadata,
                    ),
                    _loop_for_step,
                )
            except Exception as _e:
                logger.debug("status_callback error (%s): %s", event_type, _e)

        def run_sync():
            # The conditional re-assignment of `message` further below
            # (prepending model-switch notes) makes Python treat it as a
            # local variable in the entire function.  `nonlocal` lets us
            # read *and* reassign the outer `_run_agent` parameter without
            # triggering an UnboundLocalError on the earlier read at
            # `_resolve_turn_agent_config(message, …)`.
            nonlocal message

            # session_key is now set via contextvars in _set_session_env()
            # (concurrency-safe). Keep os.environ as fallback for CLI/cron.
            os.environ["HERMES_SESSION_KEY"] = session_key or ""

            # Read from env var or use default (same as CLI)
            max_iterations = int(os.getenv("HERMES_MAX_ITERATIONS", "90"))
            
            # Map platform enum to the platform hint key the agent understands.
            # Platform.LOCAL ("local") maps to "cli"; others pass through as-is.
            platform_key = "cli" if source.platform == Platform.LOCAL else source.platform.value
            
            # Combine platform context with user-configured ephemeral system prompt
            combined_ephemeral = context_prompt or ""
            if self._ephemeral_system_prompt:
                combined_ephemeral = (combined_ephemeral + "\n\n" + self._ephemeral_system_prompt).strip()

            # Re-read .env and config for fresh credentials (gateway is long-lived,
            # keys may change without restart).
            try:
                load_dotenv(_env_path, override=True, encoding="utf-8")
            except UnicodeDecodeError:
                load_dotenv(_env_path, override=True, encoding="latin-1")
            except Exception:
                pass

            try:
                model, runtime_kwargs = self._resolve_session_agent_runtime(
                    source=source,
                    session_key=session_key,
                    user_config=user_config,
                )
            except Exception as exc:
                return {
                    "final_response": f"⚠️ Provider authentication failed: {exc}",
                    "messages": [],
                    "api_calls": 0,
                    "tools": [],
                }

            pr = self._provider_routing
            reasoning_config = self._load_reasoning_config()
            self._reasoning_config = reasoning_config
            self._service_tier = self._load_service_tier()
            # Set up stream consumer for token streaming or interim commentary.
            _stream_consumer = None
            _stream_delta_cb = None
            _scfg = getattr(getattr(self, 'config', None), 'streaming', None)
            if _scfg is None:
                from gateway.config import StreamingConfig
                _scfg = StreamingConfig()

            # Per-platform streaming gate: display.platforms.<plat>.streaming
            # can disable streaming for specific platforms even when the global
            # streaming config is enabled.
            _plat_streaming = resolve_display_setting(
                user_config, platform_key, "streaming"
            )
            # None = no per-platform override → follow global config
            _streaming_enabled = (
                _scfg.enabled and _scfg.transport != "off"
                if _plat_streaming is None
                else bool(_plat_streaming)
            )
            _want_stream_deltas = _streaming_enabled
            _want_interim_messages = interim_assistant_messages_enabled
            _want_interim_consumer = _want_interim_messages
            if _want_stream_deltas or _want_interim_consumer:
                try:
                    from gateway.stream_consumer import GatewayStreamConsumer, StreamConsumerConfig
                    _adapter = self.adapters.get(source.platform)
                    if _adapter:
                        _consumer_cfg = StreamConsumerConfig(
                            edit_interval=_scfg.edit_interval,
                            buffer_threshold=_scfg.buffer_threshold,
                            cursor=_scfg.cursor,
                        )
                        _stream_consumer = GatewayStreamConsumer(
                            adapter=_adapter,
                            chat_id=source.chat_id,
                            config=_consumer_cfg,
                            metadata={"thread_id": _progress_thread_id} if _progress_thread_id else None,
                        )
                        if _want_stream_deltas:
                            _stream_delta_cb = _stream_consumer.on_delta
                        stream_consumer_holder[0] = _stream_consumer
                except Exception as _sc_err:
                    logger.debug("Could not set up stream consumer: %s", _sc_err)

            def _interim_assistant_cb(text: str, *, already_streamed: bool = False) -> None:
                if _stream_consumer is not None:
                    if already_streamed:
                        _stream_consumer.on_segment_break()
                    else:
                        _stream_consumer.on_commentary(text)
                    return
                if already_streamed or not _status_adapter or not str(text or "").strip():
                    return
                try:
                    asyncio.run_coroutine_threadsafe(
                        _status_adapter.send(
                            _status_chat_id,
                            text,
                            metadata=_status_thread_metadata,
                        ),
                        _loop_for_step,
                    )
                except Exception as _e:
                    logger.debug("interim_assistant_callback error: %s", _e)

            turn_route = self._resolve_turn_agent_config(message, model, runtime_kwargs)

            # Check agent cache — reuse the AIAgent from the previous message
            # in this session to preserve the frozen system prompt and tool
            # schemas for prompt cache hits.
            _sig = self._agent_config_signature(
                turn_route["model"],
                turn_route["runtime"],
                enabled_toolsets,
                combined_ephemeral,
            )
            agent = None
            _cache_lock = getattr(self, "_agent_cache_lock", None)
            _cache = getattr(self, "_agent_cache", None)
            if _cache_lock and _cache is not None:
                with _cache_lock:
                    cached = _cache.get(session_key)
                    if cached and cached[1] == _sig:
                        agent = cached[0]
                        logger.debug("Reusing cached agent for session %s", session_key)

            if agent is None:
                # Config changed or first message — create fresh agent
                agent = AIAgent(
                    model=turn_route["model"],
                    **turn_route["runtime"],
                    max_iterations=max_iterations,
                    quiet_mode=True,
                    verbose_logging=False,
                    enabled_toolsets=enabled_toolsets,
                    ephemeral_system_prompt=combined_ephemeral or None,
                    prefill_messages=self._prefill_messages or None,
                    reasoning_config=reasoning_config,
                    service_tier=self._service_tier,
                    request_overrides=turn_route.get("request_overrides"),
                    providers_allowed=pr.get("only"),
                    providers_ignored=pr.get("ignore"),
                    providers_order=pr.get("order"),
                    provider_sort=pr.get("sort"),
                    provider_require_parameters=pr.get("require_parameters", False),
                    provider_data_collection=pr.get("data_collection"),
                    session_id=session_id,
                    platform=platform_key,
                    user_id=source.user_id,
                    session_db=self._session_db,
                    fallback_model=self._fallback_model,
                )
                if _cache_lock and _cache is not None:
                    with _cache_lock:
                        _cache[session_key] = (agent, _sig)
                logger.debug("Created new agent for session %s (sig=%s)", session_key, _sig)

            # Per-message state — callbacks and reasoning config change every
            # turn and must not be baked into the cached agent constructor.
            agent.tool_progress_callback = progress_callback if tool_progress_enabled else None
            agent.step_callback = _step_callback_sync if _hooks_ref.loaded_hooks else None
            agent.stream_delta_callback = _stream_delta_cb
            agent.interim_assistant_callback = _interim_assistant_cb if _want_interim_messages else None
            agent.status_callback = _status_callback_sync
            agent.reasoning_config = reasoning_config
            agent.service_tier = self._service_tier
            agent.request_overrides = turn_route.get("request_overrides")

            # Background review delivery — send "💾 Memory updated" etc. to user
            def _bg_review_send(message: str) -> None:
                if not _status_adapter:
                    return
                try:
                    asyncio.run_coroutine_threadsafe(
                        _status_adapter.send(
                            _status_chat_id,
                            message,
                            metadata=_status_thread_metadata,
                        ),
                        _loop_for_step,
                    )
                except Exception as _e:
                    logger.debug("background_review_callback error: %s", _e)

            agent.background_review_callback = _bg_review_send

            # Store agent reference for interrupt support
            agent_holder[0] = agent
            # Capture the full tool definitions for transcript logging
            tools_holder[0] = agent.tools if hasattr(agent, 'tools') else None
            
            # Convert history to agent format.
            # Two cases:
            #   1. Normal path (from transcript): simple {role, content, timestamp} dicts
            #      - Strip timestamps, keep role+content
            #   2. Interrupt path (from agent result["messages"]): full agent messages
            #      that may include tool_calls, tool_call_id, reasoning, etc.
            #      - These must be passed through intact so the API sees valid
            #        assistant→tool sequences (dropping tool_calls causes 500 errors)
            agent_history = []
            for msg in history:
                role = msg.get("role")
                if not role:
                    continue
                
                # Skip metadata entries (tool definitions, session info)
                # -- these are for transcript logging, not for the LLM
                if role in ("session_meta",):
                    continue
                
                # Skip system messages -- the agent rebuilds its own system prompt
                if role == "system":
                    continue
                
                # Rich agent messages (tool_calls, tool results) must be passed
                # through intact so the API sees valid assistant→tool sequences
                has_tool_calls = "tool_calls" in msg
                has_tool_call_id = "tool_call_id" in msg
                is_tool_message = role == "tool"
                
                if has_tool_calls or has_tool_call_id or is_tool_message:
                    clean_msg = {k: v for k, v in msg.items() if k != "timestamp"}
                    agent_history.append(clean_msg)
                else:
                    # Simple text message - just need role and content
                    content = msg.get("content")
                    if content:
                        # Tag cross-platform mirror messages so the agent knows their origin
                        if msg.get("mirror"):
                            mirror_src = msg.get("mirror_source", "another session")
                            content = f"[Delivered from {mirror_src}] {content}"
                        entry = {"role": role, "content": content}
                        # Preserve reasoning fields on assistant messages so
                        # multi-turn reasoning context survives session reload.
                        # The agent's _build_api_kwargs converts these to the
                        # provider-specific format (reasoning_content, etc.).
                        if role == "assistant":
                            for _rkey in ("reasoning", "reasoning_details",
                                          "codex_reasoning_items"):
                                _rval = msg.get(_rkey)
                                if _rval:
                                    entry[_rkey] = _rval
                        agent_history.append(entry)
            
            # Collect MEDIA paths already in history so we can exclude them
            # from the current turn's extraction. This is compression-safe:
            # even if the message list shrinks, we know which paths are old.
            _history_media_paths: set = set()
            for _hm in agent_history:
                if _hm.get("role") in ("tool", "function"):
                    _hc = _hm.get("content", "")
                    if "MEDIA:" in _hc:
                        for _match in re.finditer(r'MEDIA:(\S+)', _hc):
                            _p = _match.group(1).strip().rstrip('",}')
                            if _p:
                                _history_media_paths.add(_p)
            
            # Register per-session gateway approval callback so dangerous
            # command approval blocks the agent thread (mirrors CLI input()).
            # The callback bridges sync→async to send the approval request
            # to the user immediately.
            from tools.approval import (
                register_gateway_notify,
                reset_current_session_key,
                set_current_session_key,
                unregister_gateway_notify,
            )

            def _approval_notify_sync(approval_data: dict) -> None:
                """Send the approval request to the user from the agent thread.

                If the adapter supports interactive button-based approvals
                (e.g. Discord's ``send_exec_approval``), use that for a richer
                UX.  Otherwise fall back to a plain text message with
                ``/approve`` instructions.
                """
                # Pause the typing indicator while the agent waits for
                # user approval.  Critical for Slack's Assistant API where
                # assistant_threads_setStatus disables the compose box — the
                # user literally cannot type /approve while "is thinking..."
                # is active.  The approval message send auto-clears the Slack
                # status; pausing prevents _keep_typing from re-setting it.
                # Typing resumes in _handle_approve_command/_handle_deny_command.
                _status_adapter.pause_typing_for_chat(_status_chat_id)

                cmd = approval_data.get("command", "")
                desc = approval_data.get("description", "dangerous command")

                # Prefer button-based approval when the adapter supports it.
                # Check the *class* for the method, not the instance — avoids
                # false positives from MagicMock auto-attribute creation in tests.
                if getattr(type(_status_adapter), "send_exec_approval", None) is not None:
                    try:
                        asyncio.run_coroutine_threadsafe(
                            _status_adapter.send_exec_approval(
                                chat_id=_status_chat_id,
                                command=cmd,
                                session_key=_approval_session_key,
                                description=desc,
                                metadata=_status_thread_metadata,
                            ),
                            _loop_for_step,
                        ).result(timeout=15)
                        return
                    except Exception as _e:
                        logger.warning(
                            "Button-based approval failed, falling back to text: %s", _e
                        )

                # Fallback: plain text approval prompt
                cmd_preview = cmd[:200] + "..." if len(cmd) > 200 else cmd
                msg = (
                    f"⚠️ **Dangerous command requires approval:**\n"
                    f"```\n{cmd_preview}\n```\n"
                    f"Reason: {desc}\n\n"
                    f"Reply `/approve` to execute, `/approve session` to approve this pattern "
                    f"for the session, `/approve always` to approve permanently, or `/deny` to cancel."
                )
                try:
                    asyncio.run_coroutine_threadsafe(
                        _status_adapter.send(
                            _status_chat_id,
                            msg,
                            metadata=_status_thread_metadata,
                        ),
                        _loop_for_step,
                    ).result(timeout=15)
                except Exception as _e:
                    logger.error("Failed to send approval request: %s", _e)

            # Prepend pending model switch note so the model knows about the switch
            _pending_notes = getattr(self, '_pending_model_notes', {})
            _msn = _pending_notes.pop(session_key, None) if session_key else None
            if _msn:
                message = _msn + "\n\n" + message

            _approval_session_key = session_key or ""
            _approval_session_token = set_current_session_key(_approval_session_key)
            register_gateway_notify(_approval_session_key, _approval_notify_sync)
            try:
                result = agent.run_conversation(message, conversation_history=agent_history, task_id=session_id)
            finally:
                unregister_gateway_notify(_approval_session_key)
                reset_current_session_key(_approval_session_token)
            result_holder[0] = result

            # Signal the stream consumer that the agent is done
            if _stream_consumer is not None:
                _stream_consumer.finish()
            
            # Return final response, or a message if something went wrong
            final_response = result.get("final_response")

            # Extract actual token counts from the agent instance used for this run
            _last_prompt_toks = 0
            _input_toks = 0
            _output_toks = 0
            _agent = agent_holder[0]
            if _agent and hasattr(_agent, "context_compressor"):
                _last_prompt_toks = getattr(_agent.context_compressor, "last_prompt_tokens", 0)
                _input_toks = getattr(_agent, "session_prompt_tokens", 0)
                _output_toks = getattr(_agent, "session_completion_tokens", 0)
            _resolved_model = getattr(_agent, "model", None) if _agent else None

            if not final_response:
                error_msg = f"⚠️ {result['error']}" if result.get("error") else "(No response generated)"
                return {
                    "final_response": error_msg,
                    "messages": result.get("messages", []),
                    "api_calls": result.get("api_calls", 0),
                    "tools": tools_holder[0] or [],
                    "history_offset": len(agent_history),
                    "last_prompt_tokens": _last_prompt_toks,
                    "input_tokens": _input_toks,
                    "output_tokens": _output_toks,
                    "model": _resolved_model,
                }
            
            # Scan tool results for MEDIA:<path> tags that need to be delivered
            # as native audio/file attachments.  The TTS tool embeds MEDIA: tags
            # in its JSON response, but the model's final text reply usually
            # doesn't include them.  We collect unique tags from tool results and
            # append any that aren't already present in the final response, so the
            # adapter's extract_media() can find and deliver the files exactly once.
            #
            # Uses path-based deduplication against _history_media_paths (collected
            # before run_conversation) instead of index slicing. This is safe even
            # when context compression shrinks the message list. (Fixes #160)
            if "MEDIA:" not in final_response:
                media_tags = []
                has_voice_directive = False
                for msg in result.get("messages", []):
                    if msg.get("role") in ("tool", "function"):
                        content = msg.get("content", "")
                        if "MEDIA:" in content:
                            for match in re.finditer(r'MEDIA:(\S+)', content):
                                path = match.group(1).strip().rstrip('",}')
                                if path and path not in _history_media_paths:
                                    media_tags.append(f"MEDIA:{path}")
                            if "[[audio_as_voice]]" in content:
                                has_voice_directive = True
                
                if media_tags:
                    seen = set()
                    unique_tags = []
                    for tag in media_tags:
                        if tag not in seen:
                            seen.add(tag)
                            unique_tags.append(tag)
                    if has_voice_directive:
                        unique_tags.insert(0, "[[audio_as_voice]]")
                    final_response = final_response + "\n" + "\n".join(unique_tags)
            
            # Sync session_id: the agent may have created a new session during
            # mid-run context compression (_compress_context splits sessions).
            # If so, update the session store entry so the NEXT message loads
            # the compressed transcript, not the stale pre-compression one.
            agent = agent_holder[0]
            _session_was_split = False
            if agent and session_key and hasattr(agent, 'session_id') and agent.session_id != session_id:
                _session_was_split = True
                logger.info(
                    "Session split detected: %s → %s (compression)",
                    session_id, agent.session_id,
                )
                entry = self.session_store._entries.get(session_key)
                if entry:
                    entry.session_id = agent.session_id
                    self.session_store._save()

            effective_session_id = getattr(agent, 'session_id', session_id) if agent else session_id

            # When compression created a new session, the messages list was
            # shortened.  Using the original history offset would produce an
            # empty new_messages slice, causing the gateway to write only a
            # user/assistant pair — losing the compressed summary and tail.
            # Reset to 0 so the gateway writes ALL compressed messages.
            _effective_history_offset = 0 if _session_was_split else len(agent_history)

            # Auto-generate session title after first exchange (non-blocking)
            if final_response and self._session_db:
                try:
                    from agent.title_generator import maybe_auto_title
                    all_msgs = result_holder[0].get("messages", []) if result_holder[0] else []
                    maybe_auto_title(
                        self._session_db,
                        effective_session_id,
                        message,
                        final_response,
                        all_msgs,
                    )
                except Exception:
                    pass

            return {
                "final_response": final_response,
                "last_reasoning": result.get("last_reasoning"),
                "messages": result_holder[0].get("messages", []) if result_holder[0] else [],
                "api_calls": result_holder[0].get("api_calls", 0) if result_holder[0] else 0,
                "tools": tools_holder[0] or [],
                "history_offset": _effective_history_offset,
                "last_prompt_tokens": _last_prompt_toks,
                "input_tokens": _input_toks,
                "output_tokens": _output_toks,
                "model": _resolved_model,
                "session_id": effective_session_id,
                "response_previewed": result.get("response_previewed", False),
            }
        
        # Start progress message sender if enabled
        progress_task = None
        if tool_progress_enabled:
            progress_task = asyncio.create_task(send_progress_messages())

        # Start stream consumer task — polls for consumer creation since it
        # happens inside run_sync (thread pool) after the agent is constructed.
        stream_task = None

        async def _start_stream_consumer():
            """Wait for the stream consumer to be created, then run it."""
            for _ in range(200):  # Up to 10s wait
                if stream_consumer_holder[0] is not None:
                    await stream_consumer_holder[0].run()
                    return
                await asyncio.sleep(0.05)

        stream_task = asyncio.create_task(_start_stream_consumer())
        
        # Track this agent as running for this session (for interrupt support)
        # We do this in a callback after the agent is created
        async def track_agent():
            # Wait for agent to be created
            while agent_holder[0] is None:
                await asyncio.sleep(0.05)
            if session_key:
                self._running_agents[session_key] = agent_holder[0]
                if self._draining:
                    self._update_runtime_status("draining")
        
        tracking_task = asyncio.create_task(track_agent())
        
        # Monitor for interrupts from the adapter (new messages arriving).
        # This is the PRIMARY interrupt path for regular text messages —
        # Level 1 (base.py) catches them before _handle_message() is reached,
        # so the Level 2 running_agent.interrupt() path never fires.
        # The inactivity poll loop below has a BACKUP check in case this
        # task dies (no error handling = silent death = lost interrupts).
        _interrupt_detected = asyncio.Event()  # shared with backup check

        async def monitor_for_interrupt():
            if not session_key:
                return

            while True:
                await asyncio.sleep(0.2)  # Check every 200ms
                try:
                    # Re-resolve adapter each iteration so reconnects don't
                    # leave us holding a stale reference.
                    _adapter = self.adapters.get(source.platform)
                    if not _adapter:
                        continue
                    # Check if adapter has a pending interrupt for this session.
                    # Must use session_key (build_session_key output) — NOT
                    # source.chat_id — because the adapter stores interrupt events
                    # under the full session key.
                    if hasattr(_adapter, 'has_pending_interrupt') and _adapter.has_pending_interrupt(session_key):
                        agent = agent_holder[0]
                        if agent:
                            # Peek at the pending message text WITHOUT consuming it.
                            # The message must remain in _pending_messages so the
                            # post-run dequeue at _dequeue_pending_event() can
                            # retrieve the full MessageEvent (with media metadata).
                            # If we pop here, a race exists: the agent may finish
                            # before checking _interrupt_requested, and the message
                            # is lost — neither the interrupt path nor the dequeue
                            # path finds it.
                            _peek_event = _adapter._pending_messages.get(session_key)
                            pending_text = _peek_event.text if _peek_event else None
                            logger.debug("Interrupt detected from adapter, signaling agent...")
                            agent.interrupt(pending_text)
                            _interrupt_detected.set()
                            break
                except asyncio.CancelledError:
                    raise
                except Exception as _mon_err:
                    logger.debug("monitor_for_interrupt error (will retry): %s", _mon_err)
        
        interrupt_monitor = asyncio.create_task(monitor_for_interrupt())

        # Periodic "still working" notifications for long-running tasks.
        # Fires every 10 minutes so the user knows the agent hasn't died.
        _NOTIFY_INTERVAL = 600  # 10 minutes
        _notify_start = time.time()

        async def _notify_long_running():
            _notify_adapter = self.adapters.get(source.platform)
            if not _notify_adapter:
                return
            while True:
                await asyncio.sleep(_NOTIFY_INTERVAL)
                _elapsed_mins = int((time.time() - _notify_start) // 60)
                # Include agent activity context if available.
                _agent_ref = agent_holder[0]
                _status_detail = ""
                if _agent_ref and hasattr(_agent_ref, "get_activity_summary"):
                    try:
                        _a = _agent_ref.get_activity_summary()
                        _parts = [f"iteration {_a['api_call_count']}/{_a['max_iterations']}"]
                        if _a.get("current_tool"):
                            _parts.append(f"running: {_a['current_tool']}")
                        else:
                            _parts.append(_a.get("last_activity_desc", ""))
                        _status_detail = " — " + ", ".join(_parts)
                    except Exception:
                        pass
                try:
                    await _notify_adapter.send(
                        source.chat_id,
                        f"⏳ Still working... ({_elapsed_mins} min elapsed{_status_detail})",
                        metadata=_status_thread_metadata,
                    )
                except Exception as _ne:
                    logger.debug("Long-running notification error: %s", _ne)

        _notify_task = asyncio.create_task(_notify_long_running())

        try:
            # Run in thread pool to not block.  Use an *inactivity*-based
            # timeout instead of a wall-clock limit: the agent can run for
            # hours if it's actively calling tools / receiving stream tokens,
            # but a hung API call or stuck tool with no activity for the
            # configured duration is caught and killed.  (#4815)
            #
            # Config: agent.gateway_timeout in config.yaml, or
            # HERMES_AGENT_TIMEOUT env var (env var takes precedence).
            # Default 1800s (30 min inactivity).  0 = unlimited.
            _agent_timeout_raw = float(os.getenv("HERMES_AGENT_TIMEOUT", 1800))
            _agent_timeout = _agent_timeout_raw if _agent_timeout_raw > 0 else None
            _agent_warning_raw = float(os.getenv("HERMES_AGENT_TIMEOUT_WARNING", 900))
            _agent_warning = _agent_warning_raw if _agent_warning_raw > 0 else None
            _warning_fired = False
            loop = asyncio.get_event_loop()
            _executor_task = asyncio.ensure_future(
                loop.run_in_executor(None, run_sync)
            )

            _inactivity_timeout = False
            _POLL_INTERVAL = 5.0

            if _agent_timeout is None:
                # Unlimited — still poll periodically for backup interrupt
                # detection in case monitor_for_interrupt() silently died.
                response = None
                while True:
                    done, _ = await asyncio.wait(
                        {_executor_task}, timeout=_POLL_INTERVAL
                    )
                    if done:
                        response = _executor_task.result()
                        break
                    # Backup interrupt check: if the monitor task died or
                    # missed the interrupt, catch it here.
                    if not _interrupt_detected.is_set() and session_key:
                        _backup_adapter = self.adapters.get(source.platform)
                        _backup_agent = agent_holder[0]
                        if (_backup_adapter and _backup_agent
                                and hasattr(_backup_adapter, 'has_pending_interrupt')
                                and _backup_adapter.has_pending_interrupt(session_key)):
                            _bp_event = _backup_adapter._pending_messages.get(session_key)
                            _bp_text = _bp_event.text if _bp_event else None
                            logger.info(
                                "Backup interrupt detected for session %s "
                                "(monitor task state: %s)",
                                session_key[:20],
                                "done" if interrupt_monitor.done() else "running",
                            )
                            _backup_agent.interrupt(_bp_text)
                            _interrupt_detected.set()
            else:
                # Poll loop: check the agent's built-in activity tracker
                # (updated by _touch_activity() on every tool call, API
                # call, and stream delta) every few seconds.
                response = None
                while True:
                    done, _ = await asyncio.wait(
                        {_executor_task}, timeout=_POLL_INTERVAL
                    )
                    if done:
                        response = _executor_task.result()
                        break
                    # Agent still running — check inactivity.
                    _agent_ref = agent_holder[0]
                    _idle_secs = 0.0
                    if _agent_ref and hasattr(_agent_ref, "get_activity_summary"):
                        try:
                            _act = _agent_ref.get_activity_summary()
                            _idle_secs = _act.get("seconds_since_activity", 0.0)
                        except Exception:
                            pass
                    # Staged warning: fire once before escalating to full timeout.
                    if (not _warning_fired and _agent_warning is not None
                            and _idle_secs >= _agent_warning):
                        _warning_fired = True
                        _warn_adapter = self.adapters.get(source.platform)
                        if _warn_adapter:
                            _elapsed_warn = int(_agent_warning // 60) or 1
                            _remaining_mins = int((_agent_timeout - _agent_warning) // 60) or 1
                            try:
                                await _warn_adapter.send(
                                    source.chat_id,
                                    f"⚠️ No activity for {_elapsed_warn} min. "
                                    f"If the agent does not respond soon, it will "
                                    f"be timed out in {_remaining_mins} min. "
                                    f"You can continue waiting or use /reset.",
                                    metadata=_status_thread_metadata,
                                )
                            except Exception as _warn_err:
                                logger.debug("Inactivity warning send error: %s", _warn_err)
                    if _idle_secs >= _agent_timeout:
                        _inactivity_timeout = True
                        break
                    # Backup interrupt check (same as unlimited path).
                    if not _interrupt_detected.is_set() and session_key:
                        _backup_adapter = self.adapters.get(source.platform)
                        _backup_agent = agent_holder[0]
                        if (_backup_adapter and _backup_agent
                                and hasattr(_backup_adapter, 'has_pending_interrupt')
                                and _backup_adapter.has_pending_interrupt(session_key)):
                            _bp_event = _backup_adapter._pending_messages.get(session_key)
                            _bp_text = _bp_event.text if _bp_event else None
                            logger.info(
                                "Backup interrupt detected for session %s "
                                "(monitor task state: %s)",
                                session_key[:20],
                                "done" if interrupt_monitor.done() else "running",
                            )
                            _backup_agent.interrupt(_bp_text)
                            _interrupt_detected.set()

            if _inactivity_timeout:
                # Build a diagnostic summary from the agent's activity tracker.
                _timed_out_agent = agent_holder[0]
                _activity = {}
                if _timed_out_agent and hasattr(_timed_out_agent, "get_activity_summary"):
                    try:
                        _activity = _timed_out_agent.get_activity_summary()
                    except Exception:
                        pass

                _last_desc = _activity.get("last_activity_desc", "unknown")
                _secs_ago = _activity.get("seconds_since_activity", 0)
                _cur_tool = _activity.get("current_tool")
                _iter_n = _activity.get("api_call_count", 0)
                _iter_max = _activity.get("max_iterations", 0)

                logger.error(
                    "Agent idle for %.0fs (timeout %.0fs) in session %s "
                    "| last_activity=%s | iteration=%s/%s | tool=%s",
                    _secs_ago, _agent_timeout, session_key,
                    _last_desc, _iter_n, _iter_max,
                    _cur_tool or "none",
                )

                # Interrupt the agent if it's still running so the thread
                # pool worker is freed.
                if _timed_out_agent and hasattr(_timed_out_agent, "interrupt"):
                    _timed_out_agent.interrupt("Execution timed out (inactivity)")

                _timeout_mins = int(_agent_timeout // 60) or 1

                # Construct a user-facing message with diagnostic context.
                _diag_lines = [
                    f"⏱️ Agent inactive for {_timeout_mins} min — no tool calls "
                    f"or API responses."
                ]
                if _cur_tool:
                    _diag_lines.append(
                        f"The agent appears stuck on tool `{_cur_tool}` "
                        f"({_secs_ago:.0f}s since last activity, "
                        f"iteration {_iter_n}/{_iter_max})."
                    )
                else:
                    _diag_lines.append(
                        f"Last activity: {_last_desc} ({_secs_ago:.0f}s ago, "
                        f"iteration {_iter_n}/{_iter_max}). "
                        "The agent may have been waiting on an API response."
                    )
                _diag_lines.append(
                    "To increase the limit, set agent.gateway_timeout in config.yaml "
                    "(value in seconds, 0 = no limit) and restart the gateway.\n"
                    "Try again, or use /reset to start fresh."
                )

                response = {
                    "final_response": "\n".join(_diag_lines),
                    "messages": result_holder[0].get("messages", []) if result_holder[0] else [],
                    "api_calls": _iter_n,
                    "tools": tools_holder[0] or [],
                    "history_offset": 0,
                    "failed": True,
                }

            # Track fallback model state: if the agent switched to a
            # fallback model during this run, persist it so /model shows
            # the actually-active model instead of the config default.
            # Skip eviction when the run failed — evicting a failed agent
            # forces MCP reinit on the next message for no benefit (the
            # same error will recur).  This was the root cause of #7130:
            # a bad model ID triggered fallback → eviction → recreation →
            # MCP reinit → same 400 → loop, burning 91% CPU for hours.
            _agent = agent_holder[0]
            _result_for_fb = result_holder[0]
            _run_failed = _result_for_fb.get("failed") if _result_for_fb else False
            if _agent is not None and hasattr(_agent, 'model') and not _run_failed:
                _cfg_model = _resolve_gateway_model()
                if _agent.model != _cfg_model and not self._is_intentional_model_switch(session_key, _agent.model):
                    # Fallback activated on a successful run — evict cached
                    # agent so the next message retries the primary model.
                    self._evict_cached_agent(session_key)

            # Check if we were interrupted OR have a queued message (/queue).
            result = result_holder[0]
            adapter = self.adapters.get(source.platform)
            
            # Get pending message from adapter.
            # Use session_key (not source.chat_id) to match adapter's storage keys.
            pending_event = None
            pending = None
            if result and adapter and session_key:
                pending_event = _dequeue_pending_event(adapter, session_key)
                if result.get("interrupted") and not pending_event and result.get("interrupt_message"):
                    pending = result.get("interrupt_message")
                elif pending_event:
                    pending = pending_event.text or _build_media_placeholder(pending_event)
                    logger.debug("Processing queued message after agent completion: '%s...'", pending[:40])

            # Safety net: if the pending text is a slash command (e.g. "/stop",
            # "/new"), discard it — commands should never be passed to the agent
            # as user input.  The primary fix is in base.py (commands bypass the
            # active-session guard), but this catches edge cases where command
            # text leaks through the interrupt_message fallback.
            if pending and pending.strip().startswith("/"):
                _pending_parts = pending.strip().split(None, 1)
                _pending_cmd_word = _pending_parts[0][1:].lower() if _pending_parts else ""
                if _pending_cmd_word:
                    try:
                        from hermes_cli.commands import resolve_command as _rc_pending
                        if _rc_pending(_pending_cmd_word):
                            logger.info(
                                "Discarding command '/%s' from pending queue — "
                                "commands must not be passed as agent input",
                                _pending_cmd_word,
                            )
                            pending_event = None
                            pending = None
                    except Exception:
                        pass

            if self._draining and (pending_event or pending):
                logger.info(
                    "Discarding pending follow-up for session %s during gateway %s",
                    session_key[:20] if session_key else "?",
                    self._status_action_label(),
                )
                pending_event = None
                pending = None

            if pending_event or pending:
                logger.debug("Processing pending message: '%s...'", pending[:40])

                # Clear the adapter's interrupt event so the next _run_agent call
                # doesn't immediately re-trigger the interrupt before the new agent
                # even makes its first API call (this was causing an infinite loop).
                if adapter and hasattr(adapter, '_active_sessions') and session_key and session_key in adapter._active_sessions:
                    adapter._active_sessions[session_key].clear()

                # Cap recursion depth to prevent resource exhaustion when the
                # user sends multiple messages while the agent keeps failing. (#816)
                if _interrupt_depth >= self._MAX_INTERRUPT_DEPTH:
                    logger.warning(
                        "Interrupt recursion depth %d reached for session %s — "
                        "queueing message instead of recursing.",
                        _interrupt_depth, session_key,
                    )
                    adapter = self.adapters.get(source.platform)
                    if adapter and pending_event:
                        merge_pending_message_event(adapter._pending_messages, session_key, pending_event)
                    elif adapter and hasattr(adapter, 'queue_message'):
                        adapter.queue_message(session_key, pending)
                    return result_holder[0] or {"final_response": response, "messages": history}

                was_interrupted = result.get("interrupted")
                if not was_interrupted:
                    # Queued message after normal completion — deliver the first
                    # response before processing the queued follow-up.
                    # Skip if streaming already delivered it.
                    _sc = stream_consumer_holder[0]
                    if _sc and stream_task:
                        try:
                            await asyncio.wait_for(stream_task, timeout=5.0)
                        except (asyncio.TimeoutError, asyncio.CancelledError):
                            stream_task.cancel()
                            try:
                                await stream_task
                            except asyncio.CancelledError:
                                pass
                        except Exception as e:
                            logger.debug("Stream consumer wait before queued message failed: %s", e)
                    _response_previewed = bool(result.get("response_previewed"))
                    _already_streamed = bool(
                        _sc
                        and (
                            getattr(_sc, "final_response_sent", False)
                            or (
                                _response_previewed
                                and getattr(_sc, "already_sent", False)
                            )
                        )
                    )
                    first_response = result.get("final_response", "")
                    if first_response and not _already_streamed:
                        try:
                            await adapter.send(
                                source.chat_id,
                                first_response,
                                metadata=_status_thread_metadata,
                            )
                        except Exception as e:
                            logger.warning("Failed to send first response before queued message: %s", e)
                # else: interrupted — discard the interrupted response ("Operation
                # interrupted." is just noise; the user already knows they sent a
                # new message).

                updated_history = result.get("messages", history)
                next_source = source
                next_message = pending
                next_message_id = None
                if pending_event is not None:
                    next_source = getattr(pending_event, "source", None) or source
                    next_message = await self._prepare_inbound_message_text(
                        event=pending_event,
                        source=next_source,
                        history=updated_history,
                    )
                    if next_message is None:
                        return result
                    next_message_id = getattr(pending_event, "message_id", None)

                return await self._run_agent(
                    message=next_message,
                    context_prompt=context_prompt,
                    history=updated_history,
                    source=next_source,
                    session_id=session_id,
                    session_key=session_key,
                    _interrupt_depth=_interrupt_depth + 1,
                    event_message_id=next_message_id,
                )
        finally:
            # Stop progress sender, interrupt monitor, and notification task
            if progress_task:
                progress_task.cancel()
            interrupt_monitor.cancel()
            _notify_task.cancel()

            # Wait for stream consumer to finish its final edit
            if stream_task:
                try:
                    await asyncio.wait_for(stream_task, timeout=5.0)
                except (asyncio.TimeoutError, asyncio.CancelledError):
                    stream_task.cancel()
                    try:
                        await stream_task
                    except asyncio.CancelledError:
                        pass
            
            # Clean up tracking
            tracking_task.cancel()
            if session_key and session_key in self._running_agents:
                del self._running_agents[session_key]
            if session_key:
                self._running_agents_ts.pop(session_key, None)
            if self._draining:
                self._update_runtime_status("draining")
            
            # Wait for cancelled tasks
            for task in [progress_task, interrupt_monitor, tracking_task, _notify_task]:
                if task:
                    try:
                        await task
                    except asyncio.CancelledError:
                        pass

        # If streaming already delivered the response, mark it so the
        # caller's send() is skipped (avoiding duplicate messages).
        # BUT: never suppress delivery when the agent failed — the error
        # message is new content the user hasn't seen, and it must reach
        # them even if streaming had sent earlier partial output.
        _sc = stream_consumer_holder[0]
        if _sc and isinstance(response, dict) and not response.get("failed"):
            _response_previewed = bool(response.get("response_previewed"))
            if (
                getattr(_sc, "final_response_sent", False)
                or (
                    _response_previewed
                    and getattr(_sc, "already_sent", False)
                )
            ):
                response["already_sent"] = True
        
        return response


def _start_cron_ticker(stop_event: threading.Event, adapters=None, loop=None, interval: int = 60):
    """
    Background thread that ticks the cron scheduler at a regular interval.
    
    Runs inside the gateway process so cronjobs fire automatically without
    needing a separate `hermes cron daemon` or system cron entry.

    When ``adapters`` and ``loop`` are provided, passes them through to the
    cron delivery path so live adapters can be used for E2EE rooms.

    Also refreshes the channel directory every 5 minutes and prunes the
    image/audio/document cache once per hour.
    """
    from cron.scheduler import tick as cron_tick
    from gateway.platforms.base import cleanup_image_cache, cleanup_document_cache

    IMAGE_CACHE_EVERY = 60   # ticks — once per hour at default 60s interval
    CHANNEL_DIR_EVERY = 5    # ticks — every 5 minutes

    logger.info("Cron ticker started (interval=%ds)", interval)
    tick_count = 0
    while not stop_event.is_set():
        try:
            cron_tick(verbose=False, adapters=adapters, loop=loop)
        except Exception as e:
            logger.debug("Cron tick error: %s", e)

        tick_count += 1

        if tick_count % CHANNEL_DIR_EVERY == 0 and adapters:
            try:
                from gateway.channel_directory import build_channel_directory
                build_channel_directory(adapters)
            except Exception as e:
                logger.debug("Channel directory refresh error: %s", e)

        if tick_count % IMAGE_CACHE_EVERY == 0:
            try:
                removed = cleanup_image_cache(max_age_hours=24)
                if removed:
                    logger.info("Image cache cleanup: removed %d stale file(s)", removed)
            except Exception as e:
                logger.debug("Image cache cleanup error: %s", e)
            try:
                removed = cleanup_document_cache(max_age_hours=24)
                if removed:
                    logger.info("Document cache cleanup: removed %d stale file(s)", removed)
            except Exception as e:
                logger.debug("Document cache cleanup error: %s", e)

        stop_event.wait(timeout=interval)
    logger.info("Cron ticker stopped")


async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool = False, verbosity: Optional[int] = 0) -> bool:
    """
    Start the gateway and run until interrupted.
    
    This is the main entry point for running the gateway.
    Returns True if the gateway ran successfully, False if it failed to start.
    A False return causes a non-zero exit code so systemd can auto-restart.
    
    Args:
        config: Optional gateway configuration override.
        replace: If True, kill any existing gateway instance before starting.
                 Useful for systemd services to avoid restart-loop deadlocks
                 when the previous process hasn't fully exited yet.
    """
    # ── Duplicate-instance guard ──────────────────────────────────────
    # Prevent two gateways from running under the same HERMES_HOME.
    # The PID file is scoped to HERMES_HOME, so future multi-profile
    # setups (each profile using a distinct HERMES_HOME) will naturally
    # allow concurrent instances without tripping this guard.
    import time as _time
    from gateway.status import get_running_pid, remove_pid_file, terminate_pid
    existing_pid = get_running_pid()
    if existing_pid is not None and existing_pid != os.getpid():
        if replace:
            logger.info(
                "Replacing existing gateway instance (PID %d) with --replace.",
                existing_pid,
            )
            try:
                terminate_pid(existing_pid, force=False)
            except ProcessLookupError:
                pass  # Already gone
            except (PermissionError, OSError):
                logger.error(
                    "Permission denied killing PID %d. Cannot replace.",
                    existing_pid,
                )
                return False
            # Wait up to 10 seconds for the old process to exit
            for _ in range(20):
                try:
                    os.kill(existing_pid, 0)
                    _time.sleep(0.5)
                except (ProcessLookupError, PermissionError):
                    break  # Process is gone
            else:
                # Still alive after 10s — force kill
                logger.warning(
                    "Old gateway (PID %d) did not exit after SIGTERM, sending SIGKILL.",
                    existing_pid,
                )
                try:
                    terminate_pid(existing_pid, force=True)
                    _time.sleep(0.5)
                except (ProcessLookupError, PermissionError, OSError):
                    pass
            remove_pid_file()
            # Also release all scoped locks left by the old process.
            # Stopped (Ctrl+Z) processes don't release locks on exit,
            # leaving stale lock files that block the new gateway from starting.
            try:
                from gateway.status import release_all_scoped_locks
                _released = release_all_scoped_locks()
                if _released:
                    logger.info("Released %d stale scoped lock(s) from old gateway.", _released)
            except Exception:
                pass
        else:
            hermes_home = str(get_hermes_home())
            logger.error(
                "Another gateway instance is already running (PID %d, HERMES_HOME=%s). "
                "Use 'hermes gateway restart' to replace it, or 'hermes gateway stop' first.",
                existing_pid, hermes_home,
            )
            print(
                f"\n❌ Gateway already running (PID {existing_pid}).\n"
                f"   Use 'hermes gateway restart' to replace it,\n"
                f"   or 'hermes gateway stop' to kill it first.\n"
                f"   Or use 'hermes gateway run --replace' to auto-replace.\n"
            )
            return False

    # Sync bundled skills on gateway start (fast -- skips unchanged)
    try:
        from tools.skills_sync import sync_skills
        sync_skills(quiet=True)
    except Exception:
        pass

    # Centralized logging — agent.log (INFO+), errors.log (WARNING+),
    # and gateway.log (INFO+, gateway-component records only).
    # Idempotent, so repeated calls from AIAgent.__init__ won't duplicate.
    from hermes_logging import setup_logging
    setup_logging(hermes_home=_hermes_home, mode="gateway")

    # Optional stderr handler — level driven by -v/-q flags on the CLI.
    # verbosity=None (-q/--quiet): no stderr output
    # verbosity=0    (default):    WARNING and above
    # verbosity=1    (-v):         INFO and above
    # verbosity=2+   (-vv/-vvv):   DEBUG
    if verbosity is not None:
        from agent.redact import RedactingFormatter

        _stderr_level = {0: logging.WARNING, 1: logging.INFO}.get(verbosity, logging.DEBUG)
        _stderr_handler = logging.StreamHandler()
        _stderr_handler.setLevel(_stderr_level)
        _stderr_handler.setFormatter(RedactingFormatter('%(levelname)s %(name)s: %(message)s'))
        logging.getLogger().addHandler(_stderr_handler)
        # Lower root logger level if needed so DEBUG records can reach the handler
        if _stderr_level < logging.getLogger().level:
            logging.getLogger().setLevel(_stderr_level)

    runner = GatewayRunner(config)
    
    # Set up signal handlers
    def shutdown_signal_handler():
        asyncio.create_task(runner.stop())

    def restart_signal_handler():
        runner.request_restart(detached=False, via_service=True)
    
    loop = asyncio.get_event_loop()
    for sig in (signal.SIGINT, signal.SIGTERM):
        try:
            loop.add_signal_handler(sig, shutdown_signal_handler)
        except NotImplementedError:
            pass
    if hasattr(signal, "SIGUSR1"):
        try:
            loop.add_signal_handler(signal.SIGUSR1, restart_signal_handler)
        except NotImplementedError:
            pass
    
    # Start the gateway
    success = await runner.start()
    if not success:
        return False
    if runner.should_exit_cleanly:
        if runner.exit_reason:
            logger.error("Gateway exiting cleanly: %s", runner.exit_reason)
        return True
    
    # Write PID file so CLI can detect gateway is running
    import atexit
    from gateway.status import write_pid_file, remove_pid_file
    write_pid_file()
    atexit.register(remove_pid_file)
    
    # Start background cron ticker so scheduled jobs fire automatically.
    # Pass the event loop so cron delivery can use live adapters (E2EE support).
    cron_stop = threading.Event()
    cron_thread = threading.Thread(
        target=_start_cron_ticker,
        args=(cron_stop,),
        kwargs={"adapters": runner.adapters, "loop": asyncio.get_running_loop()},
        daemon=True,
        name="cron-ticker",
    )
    cron_thread.start()
    
    # Wait for shutdown
    await runner.wait_for_shutdown()

    if runner.should_exit_with_failure:
        if runner.exit_reason:
            logger.error("Gateway exiting with failure: %s", runner.exit_reason)
        return False
    
    # Stop cron ticker cleanly
    cron_stop.set()
    cron_thread.join(timeout=5)

    # Close MCP server connections
    try:
        from tools.mcp_tool import shutdown_mcp_servers
        shutdown_mcp_servers()
    except Exception:
        pass

    if runner.exit_code is not None:
        raise SystemExit(runner.exit_code)

    return True


def main():
    """CLI entry point for the gateway."""
    import argparse
    
    parser = argparse.ArgumentParser(description="Hermes Gateway - Multi-platform messaging")
    parser.add_argument("--config", "-c", help="Path to gateway config file")
    parser.add_argument("--verbose", "-v", action="store_true", help="Verbose output")
    
    args = parser.parse_args()
    
    config = None
    if args.config:
        import json
        with open(args.config, encoding="utf-8") as f:
            data = json.load(f)
            config = GatewayConfig.from_dict(data)
    
    # Run the gateway - exit with code 1 if no platforms connected,
    # so systemd Restart=on-failure will retry on transient errors (e.g. DNS)
    success = asyncio.run(start_gateway(config))
    if not success:
        sys.exit(1)


if __name__ == "__main__":
    main()