hermes-agent/gateway/run_inbound.py

"""Inbound message pipeline (_handle_message, text/media preparation, durable-turn markers, plugin injection) for GatewayRunner.

Split out of ``gateway/run.py``; bound onto ``GatewayRunner`` via the MRO.
``gateway.run`` internals are imported lazily inside method bodies (import cycle),
so ``patch("gateway.run.X")`` keeps intercepting them at call time.
"""

from __future__ import annotations

import logging
from typing import TYPE_CHECKING
import asyncio
import concurrent.futures
import dataclasses
import json
import os
import re
import time
from contextlib import suppress
from gateway.config import Platform
from gateway.platforms.base import EphemeralReply
from gateway.platforms.event import MessageEvent, MessageType
from gateway.run_common import _UNSET
from gateway.session import (
    SessionSource, is_shared_multi_user_session, neutralize_untrusted_inline_text
)
from gateway.turn_lease import TurnLeaseTimeoutError
from typing import Any, Dict, List, Optional, Tuple

if TYPE_CHECKING:  # string annotations only; never imported at runtime (cycle)
    from gateway.run import GatewayRunner  # noqa: F401
    from gateway.run_turn_runner import TurnRunner  # noqa: F401

# Log-record parity with the origin module.
logger = logging.getLogger("gateway.run")


class GatewayInboundMixin:
    """Inbound message pipeline (_handle_message, text/media preparation, durable-turn markers, plugin injection) for GatewayRunner."""

    def _hm_pre_gateway_dispatch_hook(
        self, event: "MessageEvent", source: SessionSource
    ) -> Optional["MessageEvent"]:
        """Run the ``pre_gateway_dispatch`` plugin hook; None = drop, else the (maybe rewritten) event.
        Results: ``{"action": "skip"}`` → drop; ``{"action": "rewrite", "text"}`` → replace ``event.text``;
        ``allow``/None → normal dispatch. Runs BEFORE auth so plugins can handle unauthorized senders."""
        try:
            from hermes_cli.lifecycle import invoke_hook as _invoke_hook
            _hook_results = _invoke_hook(
                "pre_gateway_dispatch", event=event, gateway=self,
                # getattr: bare-runner tests build GatewayRunner via object.__new__ without __init__.
                session_store=getattr(self, "session_store", None),
            )
        except Exception as _hook_exc:
            logger.warning("pre_gateway_dispatch invocation failed: %s", _hook_exc)
            _hook_results = []

        for _result in _hook_results:
            if not isinstance(_result, dict):
                continue
            _action = _result.get("action")
            if _action == "skip":
                logger.info(
                    "pre_gateway_dispatch skip: reason=%s platform=%s chat=%s",
                    _result.get("reason"), source.platform.value if source.platform else "unknown",
                    source.chat_id or "unknown",
                )
                return None
            if _action == "rewrite":
                _new_text = _result.get("text")
                if isinstance(_new_text, str):
                    event = dataclasses.replace(event, text=_new_text)
                break
            if _action == "allow":
                break
        return event

    async def _hm_offer_pairing_code(self, source: SessionSource) -> None:
        """DM an unauthorized sender a pairing code (rate-limited; groups never reach here)."""
        platform_name = source.platform.value if source.platform else "unknown"
        pairing_store = self._pairing_store_for(source)
        if pairing_store is None:
            logger.error("Cannot offer pairing code on %s: no pairing store", platform_name)
            return
        # Rate-limit ALL pairing responses (code or rejection) so a burst of DMs doesn't spam.
        if pairing_store._is_rate_limited(platform_name, source.user_id):
            return
        code = pairing_store.generate_code(platform_name, source.user_id, source.user_name or "")
        adapter = self._adapter_for_source(source)
        if code:
            store_profile = getattr(pairing_store, "profile", None)
            profile_arg = (
                f"-p {store_profile} "
                if isinstance(store_profile, str) and store_profile and store_profile != "default"
                else ""
            )
            reply = (
                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 {profile_arg}pairing approve "
                f"{platform_name} {code}`"
            )
        else:
            reply = "Too many pairing requests right now~ Please try again later!"
        if adapter:
            await adapter.send(source.chat_id, reply)
        if not code:
            # Record rate limit so subsequent messages are silently ignored
            pairing_store._record_rate_limit(platform_name, source.user_id)

    async def _hm_admit_event(
        self, event: "MessageEvent"
    ) -> Optional[Tuple["MessageEvent", SessionSource, bool]]:
        """Ingress gates for ``_handle_message``; None when dropped, else ``(event, source, is_internal)``
        (the ``pre_gateway_dispatch`` hook may have rewritten ``event``)."""
        from gateway.run import _is_slack_ignored_channel
        source = event.source
        # getattr(self, ...) throughout: bare test runners build GatewayRunner via object.__new__.
        _config = getattr(self, "config", None)

        # 🔴 Cross-session leak guard: this per-message task was create_task()'d with a copy of the
        # spawning context, which may carry ANOTHER message's HERMES_SESSION_* ContextVars; until
        # _set_session_env binds ours a subprocess would read the foreign identity. Reset to _UNSET.
        try:
            from gateway.session_context import reset_session_vars
            reset_session_vars()
        except Exception:
            logger.debug("reset_session_vars failed at handler entry", exc_info=True)

        # Most adapters resolve profile routes in build_source(); internal/voice paths construct
        # SessionSource directly, so resolve those here as the shared fail-closed ingress gate.
        # Strict boolean marker: require the literal True so duck-typed test/internal sources with
        # dynamic attributes are not mistaken for a rejection.
        if (
            getattr(_config, "multiplex_profiles", False)
            and not getattr(source, "profile", None)
            and getattr(source, "profile_route_rejected", False) is not True
        ):
            from gateway.profile_routing import ProfileRouteRejected

            try:
                source.profile = self._profile_name_for_source(source)
            except ProfileRouteRejected:
                source.profile_route_rejected = True
        if getattr(source, "profile_route_rejected", False) is True:
            logger.warning(
                "Dropping inbound message because its explicit profile route "
                "targets an unserved profile"
            )
            return None

        is_internal = bool(getattr(event, "internal", False))  # e.g. background-process notifications

        # Ignored-channel guard runs FIRST — before startup-restore queueing, plugin hooks, auth,
        # and session setup — so an ignored channel can never reach pairing/auth/session state.
        _chat_id = getattr(source, "chat_id", None)
        if (
            # See #51899.
            not is_internal
            and getattr(source, "platform", None) == Platform.SLACK
            and _is_slack_ignored_channel(_config, _chat_id)
        ):
            logger.info("Dropping Slack message from configured ignored channel %s", _chat_id)
            return None

        if (
            getattr(self, "_startup_restore_in_progress", False)
            and not is_internal
            and not getattr(event, "_hermes_startup_restore_replay", False)
        ):
            self._queue_startup_restore_event(event)
            return None

        if is_internal:
            return event, source, True

        # scale-to-zero: only real user-originated inbound stamps the last-inbound clock;
        # counting internal/system events would keep a genuinely idle gateway awake.
        self._scale_to_zero_note_real_inbound()
        event = self._hm_pre_gateway_dispatch_hook(event, source)
        if event is None:
            return None
        source = event.source

        if not self._is_user_authorized_for_source(source):
            if source.user_id is None:
                # No user identity (Telegram service messages, channel forwards, anonymous admin
                # posts, sender_chat): can't be paired but may be authorized via a chat allowlist.
                logger.debug("Ignoring message with no user_id from %s", source.platform.value)
                return None
            logger.warning("Unauthorized user: %s (%s) on %s", source.user_id, source.user_name, source.platform.value)
            # DMs get a pairing code, groups are ignored. A bot cannot pair, and answering one mid-cooldown is outbound traffic.
            if (
                source.chat_type == "dm"
                and not getattr(source, "is_bot", False)
                and self._get_unauthorized_dm_behavior(source.platform, profile=source.profile) == "pair"
            ):
                await self._hm_offer_pairing_code(source)
            return None
        # The busy path charged this event on arrival; a drained follow-up must not pay twice.
        if not getattr(event, "_bot_loop_admitted", False) and not self._admit_bot_message_for_source(source):
            return None
        return event, source, False

    def _hm_estop_turn_allowed(self, event: "MessageEvent", source: SessionSource) -> bool:
        """Whether a turn may bypass the global emergency stop: pause blocks NEW agent turns, never
        running work or control traffic — recognized slash commands (incl. /pause off, the in-band
        resume) and replies owned by in-flight work (pending update prompt, running session,
        pending slash-confirm, dangerous-command approval) all pass through."""
        with suppress(Exception):
            _estop_cmd = event.get_command()
            if _estop_cmd:
                from hermes_cli.commands import resolve_command as _resolve_estop_cmd
                if _resolve_estop_cmd(_estop_cmd) is not None:
                    return True
        with suppress(Exception):
            _estop_key = self._session_key_for_source(source)
            _estop_state = self._peek_session_state(_estop_key)
            if _estop_state is not None and _estop_state.persistent.update_prompt_pending:
                return True
            # A running session covers steering plus pending clarify / tool approvals it holds.
            if self._is_session_running(_estop_key):
                return True
            from tools import slash_confirm as _estop_confirm_mod
            if _estop_confirm_mod.get_pending(_estop_key):
                return True
            from tools.approval import has_blocking_approval as _estop_has_approval
            if _estop_has_approval(_estop_key):
                return True
        return False

    def _hm_estop_gate(
        self, event: "MessageEvent", source: SessionSource, is_internal: bool
    ) -> Optional[str]:
        """Global emergency-stop (`hermes pause`) notice when this turn must be blocked, else None.
        Placed after auth so unauthorized senders can't probe pause state."""
        if is_internal:
            return None
        try:
            from agent.estop import paused_reply as _estop_paused_reply
        except ImportError:
            return None
        _paused_notice = _estop_paused_reply()
        if _paused_notice is None or self._hm_estop_turn_allowed(event, source):
            return None
        logger.info(
            "Gateway turn paused by global emergency stop (platform=%s chat=%s)",
            getattr(getattr(source, "platform", None), "value", "unknown"),
            getattr(source, "chat_id", None) or "unknown",
        )
        return _paused_notice

    @staticmethod
    def _hm_write_update_response(response_text: str) -> Optional[str]:
        """Atomically hand *response_text* to the detached update process; returns the OSError str."""
        from gateway.run import _hermes_home
        response_path = _hermes_home / ".update_response"
        try:
            tmp = response_path.with_suffix(".tmp")
            tmp.write_text(response_text, encoding="utf-8")
            tmp.replace(response_path)
            (_hermes_home / ".update_prompt.json").unlink(missing_ok=True)
        except OSError as e:
            return str(e)
        return None

    def _hm_update_prompt_reply(self, event: "MessageEvent", _quick_key: str) -> Optional[str]:
        """Consume a reply to a pending ``/update`` prompt (routed to the detached update process via
        ``.update_response``); None when nothing was consumed. Recognized slash commands must bypass
        this or /new, /help etc. get silently consumed as update answers."""
        _up_state = self._peek_session_state(_quick_key)
        if _up_state is None or not _up_state.persistent.update_prompt_pending:
            return None
        # Accept /approve and /deny as shorthand for yes/no
        cmd = event.get_command()
        _recognized_cmd = None
        if cmd in {"approve", "yes"}:
            response_text = "y"
        elif cmd in {"deny", "no"}:
            response_text = "n"
        else:
            if cmd:
                with suppress(Exception):
                    from hermes_cli.commands import resolve_command as _resolve_update_cmd
                    _cmd_def = _resolve_update_cmd(cmd)
                    _recognized_cmd = _cmd_def.name if _cmd_def else None
            response_text = "" if _recognized_cmd else (event.text or "").strip()
        if response_text:
            err = self._hm_write_update_response(response_text)
            if err is not None:
                logger.warning("Failed to write update response: %s", err)
                return f"✗ Failed to send response to update process: {err}"
            _up_state.persistent.update_prompt_pending = False
            label = response_text if len(response_text) <= 20 else response_text[:20] + "…"
            return f"✓ Sent `{label}` to the update process."
        # Recognized slash command during a pending update prompt: write a blank response so the
        # detached update's ``_gateway_prompt`` returns the prompt's default (typically a safe
        # "n" / skip) and exits instead of blocking on stdin until the watcher timeout.
        if _recognized_cmd:
            err = self._hm_write_update_response("")
            if err is None:
                logger.info(
                    "Recognized /%s during pending update prompt for %s; "
                    "cancelled prompt with default and dispatching command",
                    _recognized_cmd, _quick_key,
                )
            else:
                logger.warning("Failed to write cancel response for pending update prompt: %s", err)
            _up_state.persistent.update_prompt_pending = False
        return None

    async def _hm_clarify_reply(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str
    ) -> Optional[str]:
        """Intercept a reply to a pending clarify prompt; None when the message falls through.
        Free text answers open-ended/"Other" prompts; "2" answers a multi-choice one. Resolved/retained
        replies return "" so adapters don't double-post — the agent produces the next user-facing message."""
        try:
            from tools import clarify_gateway as _clarify_mod
            _pending_clarify = _clarify_mod.get_pending_for_session(_quick_key, include_choice_prompts=True)
        except Exception:
            return None
        if _pending_clarify is None:
            return None
        _clarify_has_audio = bool(self._pending_event_audio_paths(event))
        _raw_clarify_reply = await self._prepare_clarify_reply_text(event)

        def _retain(why: str) -> str:
            logger.info(
                "Gateway retained pending clarify after %s (session=%s, id=%s)",
                why, _quick_key, _pending_clarify.clarify_id,
            )
            return ""

        if _clarify_has_audio and not _raw_clarify_reply:
            return _retain("voice transcription produced no usable text")
        # Slash commands: the user wanted a command, not to answer the clarify. Leave it pending so
        # they can retry; on timeout the agent unblocks with an empty response.
        if not _raw_clarify_reply or _raw_clarify_reply.startswith("/"):
            return None
        _text_outcome = _clarify_mod.attempt_text_response_for_session(_quick_key, _raw_clarify_reply)
        if _text_outcome == _clarify_mod.TEXT_RESOLVED:
            logger.info(
                "Gateway intercepted clarify text response (session=%s, id=%s)",
                _quick_key, _pending_clarify.clarify_id,
            )
            # The clarify callback pauses the platform typing/status indicator while waiting so
            # Slack users can type; the active agent resumes now, so re-enable its indicator.
            _clarify_adapter = self._adapter_for_source(source)
            if _clarify_adapter:
                try:
                    _clarify_adapter.resume_typing_for_chat(source.chat_id)
                except Exception:
                    logger.debug("Failed to resume typing after clarify response", exc_info=True)
            return ""
        if _text_outcome == _clarify_mod.TEXT_REJECTED_SELECTION:
            # Selection-shaped but invalid (out-of-range number, bad comma-list): keep the clarify
            # armed for retry — don't cancel, don't treat as an unrelated follow-up.
            return _retain("invalid selection attempt")
        if _text_outcome == _clarify_mod.TEXT_REJECTED_PROSE:
            # Native-choice prompts reject unmatched prose so it continues through normal busy
            # routing. Release this clarify first: redirect() degrades to steer() while tools
            # execute, and that steer cannot drain until the clarify tool returns.
            _clarify_mod.resolve_gateway_clarify(_pending_clarify.clarify_id, "")
        return None

    # Reply → choice for a pending slash-confirm prompt; the command spelling wins over the
    # bang/slash-stripped free-text spelling.
    _SLASH_CONFIRM_CMD_CHOICES = {
        "approve": "once", "yes": "once", "ok": "once", "confirm": "once",
        "always": "always", "remember": "always",
        "cancel": "cancel", "no": "cancel", "deny": "cancel", "nevermind": "cancel",
    }
    _SLASH_CONFIRM_TEXT_CHOICES = {
        "approve": "once", "approve once": "once", "once": "once",
        "always": "always", "always approve": "always",
        "cancel": "cancel", "nevermind": "cancel", "no": "cancel",
    }

    async def _hm_slash_confirm_reply(self, event: "MessageEvent", _quick_key: str) -> Optional[str]:
        """Resolve a reply (/approve, /always, /cancel + aliases) to a pending slash-confirm prompt;
        None when it falls through — a stale pending confirm does NOT block other commands. A pending
        dangerous-command approval takes precedence: /approve there unblocks the waiting tool thread."""
        from tools import slash_confirm as _slash_confirm_mod
        _pending_confirm = _slash_confirm_mod.get_pending(_quick_key)
        if not _pending_confirm:
            return None
        with suppress(Exception):
            from tools.approval import has_blocking_approval
            if has_blocking_approval(_quick_key):
                return None
        # Accept bang-prefixed replies (`!always`, `!cancel`) verbatim: Slack/Matrix show the `!`
        # prefix (typed `/` is blocked in Slack threads) and adapters only rewrite
        # `!<known-command>` — confirm keywords aren't commands, so the `!` survives to here.
        _norm_reply = (event.text or "").strip().lstrip("!/").lower()
        _confirm_choice = (
            self._SLASH_CONFIRM_CMD_CHOICES.get(event.get_command())
            or self._SLASH_CONFIRM_TEXT_CHOICES.get(_norm_reply)
        )
        if _confirm_choice is not None:
            _resolved = await _slash_confirm_mod.resolve(
                _quick_key, _pending_confirm.get("confirm_id"), _confirm_choice,
            )
            return _resolved or ""
        # Stale pending + unrelated command: the user moved on, so drop the pending state rather
        # than let the confirm block normal usage indefinitely.
        _slash_confirm_mod.clear_if_stale(_quick_key)
        return None

    def _hm_evict_idle_stale_agent(self, _quick_key: str) -> None:
        """Evict a leaked lock from a hung/crashed handler: only when the agent has been *idle* past
        the threshold (active tasks can run for hours), or has no activity tracker and an extreme
        wall-clock age. The pending sentinel is never evicted (no get_activity_summary() → idle
        reads inf and would race the async setup path)."""
        from gateway.run import _AGENT_PENDING_SENTINEL, _float_env
        _raw_stale_timeout = _float_env("HERMES_AGENT_TIMEOUT", 1800)
        _quick_state = self._peek_session_state(_quick_key)
        _stale_ts = _quick_state.turn.started_ts if _quick_state else 0
        if _quick_state is None or _quick_state.turn.agent is None or not _stale_ts:
            return
        _stale_age = time.time() - _stale_ts
        _stale_agent = _quick_state.turn.agent
        _stale_idle = float("inf")  # assume idle if we can't check
        _stale_detail = ""
        _activity_summary_valid = False
        if _stale_agent and hasattr(_stale_agent, "get_activity_summary"):
            with suppress(Exception):
                _sa = _stale_agent.get_activity_summary()
                from gateway.session_stall import resolve_session_idle_seconds_from_activity

                _sa_d = _sa if isinstance(_sa, dict) else {}
                _resolved_idle = resolve_session_idle_seconds_from_activity(
                    _sa if isinstance(_sa, dict) else None, now=time.time(),
                )
                if _resolved_idle is not None:
                    _stale_idle = _resolved_idle
                    _activity_summary_valid = True
                _stale_detail = (
                    f" | last_activity={_sa_d.get('last_activity_desc', 'unknown')} "
                    f"({_stale_idle:.0f}s ago) "
                    f"| iteration={_sa_d.get('api_call_count', 0)}/{_sa_d.get('max_iterations', 0)}"
                )
        # A valid activity clock is authoritative: total age alone never makes an actively
        # progressing turn stale. The emergency wall TTL is only a fallback when the agent cannot
        # report usable activity.
        _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 (
            (_activity_summary_valid and _raw_stale_timeout > 0 and _stale_idle >= _raw_stale_timeout)
            or (not _activity_summary_valid and _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, _stale_age, _stale_idle, _raw_stale_timeout, _stale_detail,
            )
            self._hm_evict_running_agent(_quick_key, "stale_running_agent_eviction")

    def _hm_evict_reaped_agent(self, _quick_key: str) -> None:
        """Evict the in-memory turn slot of a session whose durable row was ended while the gateway
        lived (``ws_orphan_reap`` / ``agent_close``): otherwise the fast-path queues every next
        message into the dead runtime. The cold path re-attaches via ``get_or_create_session``."""
        try:
            # #99106: durable-reaped guard. This is the live-gateway variant of #54878 and the #632
            # detached/ 405 suppressions in production. Evict the stale slot so the next message falls
            # through to the cold path and re-attaches or creates a fresh session; /status then correctly
            # shows 代理运行中: 否 before the heal and a live turn after.
            _reap_store = getattr(self, "session_store", None)
            # Public, lock-held accessors: peek_session_id returns a non-str on stubbed stores in
            # bare test runners — the isinstance() / ``is True`` gates keep this inert unless a
            # real SessionStore answers.
            _reap_peek = getattr(_reap_store, "peek_session_id", None)
            _is_ended = getattr(_reap_store, "_is_session_ended_in_db", None)
            _reap_sid = _reap_peek(_quick_key) if callable(_reap_peek) else None
            if isinstance(_reap_sid, str) and _reap_sid and callable(_is_ended) and _is_ended(_reap_sid) is True:
                logger.warning(
                    "Evicting stale _running_agents entry for %s — "
                    "durable session %s is ended (reaped) in state.db; "
                    "healing routing on next message (#99106)", _quick_key, _reap_sid,
                )
                self._hm_evict_running_agent(_quick_key, "reaped_session_eviction")
        except Exception:
            logger.debug("reaped-session staleness check failed", exc_info=True)

    def _hm_evict_running_agent(self, _quick_key: str, reason: str) -> None:
        self._invalidate_session_run_generation(_quick_key, reason=reason)
        self._release_running_agent_state(_quick_key)

    def _hm_merge_pending_for_source(
        self, source: SessionSource, _quick_key: str, event: "MessageEvent", *, merge_text: bool = False
    ) -> None:
        """Merge *event* into the source adapter's pending slot (no-op without an adapter)."""
        from gateway.platforms.base import merge_pending_message_event
        adapter = self._adapter_for_source(source)
        if adapter:
            merge_pending_message_event(adapter._pending_messages, _quick_key, event, merge_text=merge_text)

    async def _hm_busy_slash_or_photo(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str
    ) -> Tuple[bool, Optional[str]]:
        """Slash-command / photo-burst handling on the busy fast-path → ``(handled, result)``. Each
        command's mid-run behavior is declared on its CommandDef (busy_policy / busy_handler)."""
        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:
            # /status and /context are intentionally pre-gate so users always see session state.
            if _cmd_def_inner.name == "status":
                return True, await self._handle_status_command(event)
            if _cmd_def_inner.name == "context":
                return True, await self._handle_context_command(event)
            # Slash access control mirrors the cold-path gate so non-admins can't bypass gating
            # just because an agent is busy. /help and /whoami are the always-allowed floor.
            _denied = self._check_slash_access(source, _cmd_def_inner.name)
            if _denied is not None:
                return True, _denied
            # Any recognized slash command dispatches per its declared busy_policy (dispatch /
            # interrupt_then_dispatch / reject). Unrecognized commands and plain text fall through.
            return True, await self._dispatch_busy_slash_command(event, _cmd_def_inner, _quick_key, source)

        # Telegram photo bursts arrive as near-simultaneous updates — never interrupt for a
        # photo-only follow-up; adapter-level batching absorbs them.
        if event.message_type == MessageType.PHOTO:
            logger.debug("PRIORITY photo follow-up for session %s — queueing without interrupt", _quick_key)
            self._hm_merge_pending_for_source(source, _quick_key, event)
            return True, None
        return False, None

    def _hm_busy_telegram_grace_queue(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str, effective_busy_input_mode: str
    ) -> bool:
        """Queue a Telegram text follow-up that lands within the post-start grace window."""
        _grace = float(os.getenv("HERMES_TELEGRAM_FOLLOWUP_GRACE_SECONDS", "3.0"))
        _grace_state = self._peek_session_state(_quick_key)
        _started_at = _grace_state.turn.started_ts if _grace_state else 0
        if not (
            source.platform == Platform.TELEGRAM and event.message_type == MessageType.TEXT
            and _grace > 0 and _started_at and (time.time() - _started_at) <= _grace
        ):
            return False
        logger.debug(
            "Telegram follow-up arrived %.2fs after run start for %s — queueing without interrupt",
            time.time() - _started_at, _quick_key,
        )
        if effective_busy_input_mode != "queue":
            self._hm_merge_pending_for_source(source, _quick_key, event, merge_text=True)
        else:
            adapter = self._adapter_for_source(source)
            if adapter:
                self._enqueue_fifo(_quick_key, event, adapter)
        return True

    @staticmethod
    def _hm_text_only(event: "MessageEvent") -> bool:
        return event.message_type == MessageType.TEXT and not event.media_urls and not event.media_types

    def _hm_busy_steer(self, event: "MessageEvent", running_agent: Any, _quick_key: str) -> None:
        """Steer mode: inject text mid-run via ``agent.steer()``, else fall back to queue semantics."""
        steer_text = (event.text or "").strip()
        steered = False
        if self._hm_text_only(event) and steer_text and hasattr(running_agent, "steer"):
            try:
                steered = bool(running_agent.steer(self._steer_text_with_origin(steer_text, event)))
            except Exception as exc:
                logger.warning("PRIORITY steer failed for session %s: %s", _quick_key, exc)
        if steered:
            logger.debug("PRIORITY steer for session %s", _quick_key)
            return
        logger.debug("PRIORITY steer-fallback-to-queue for session %s", _quick_key)
        self._queue_or_replace_pending_event(_quick_key, event)

    async def _hm_busy_interrupt(
        self, event: "MessageEvent", source: SessionSource, running_agent: Any, _quick_key: str
    ) -> None:
        """Interrupt path: redirect text-only corrections when supported, else ``agent.interrupt()``."""
        from gateway.run import _build_media_placeholder
        # Text-only corrections redirect the live turn (preserving displayed context) when the
        # runtime supports it; media/voice and older runtimes use the interrupt path below.
        _can_redirect = getattr(running_agent, "_supports_active_turn_redirect", False) is True
        if self._hm_text_only(event) and _can_redirect and hasattr(running_agent, "redirect"):
            try:
                if running_agent.redirect(
                    self._steer_text_with_origin((event.text or "").strip(), event)
                ):
                    logger.debug("PRIORITY redirect for session %s", _quick_key)
                    return
            except Exception as exc:
                logger.warning("PRIORITY redirect failed for session %s: %s", _quick_key, exc)
        logger.debug("PRIORITY interrupt for session %s", _quick_key)
        _interrupt_text = event.text
        if self._pending_event_audio_paths(event):
            _interrupt_text, _ = await self._transcribe_and_echo_pending_voice(
                event, self._adapter_for_source(source), source, event.text or "",
                log_context="Voice-priority-interrupt",
            )
        elif not _interrupt_text and getattr(event, "media_urls", None):
            _interrupt_text = _build_media_placeholder(event)
        # Delivered via adapter._pending_messages (read by _run_agent); never also buffered on self
        # — that copy was never consumed and grew unbounded.
        running_agent.interrupt(_interrupt_text)

    async def _hm_handle_running_session_message(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str
    ) -> Optional[str]:
        """Fast-path while this session's agent is running: interrupt by default (minimal latency);
        busy_input_mode queue/steer, subagent and compression protection demote to queue."""
        from gateway.run import _AGENT_PENDING_SENTINEL
        _handled, _result = await self._hm_busy_slash_or_photo(event, source, _quick_key)
        if _handled:
            return _result

        effective_busy_input_mode = self._effective_busy_input_mode(source)
        if self._hm_busy_telegram_grace_queue(event, source, _quick_key, effective_busy_input_mode):
            return None

        _ra_state = self._peek_session_state(_quick_key)
        running_agent = _ra_state.turn.agent if _ra_state else None
        if running_agent is _AGENT_PENDING_SENTINEL:  # agent still being set up
            if event.get_command() == "stop":  # force-clean the sentinel so the session is unlocked
                self._release_running_agent_state(_quick_key)
                logger.info("HARD STOP (pending) for session %s — sentinel cleared", _quick_key)
                return EphemeralReply("⚡ Force-stopped. The agent was still starting — session unlocked.")
            self._hm_merge_pending_for_source(source, _quick_key, event, merge_text=True)  # picked up after start
            return None
        if self._draining:
            queue_during_drain = self._queue_during_drain_enabled(effective_busy_input_mode)
            if queue_during_drain:
                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 queue_during_drain
                else f"⏳ Gateway is {self._status_action_gerund()} and is not accepting another turn right now."
            )
        if effective_busy_input_mode == "queue":
            logger.debug("PRIORITY queue follow-up for session %s", _quick_key)
            self._queue_or_replace_pending_event(_quick_key, event)
            return None
        if effective_busy_input_mode == "steer":
            self._hm_busy_steer(event, running_agent, _quick_key)
            return None
        # Subagent protection: an interrupt cascades through ``_active_children`` and aborts
        # in-flight delegate_task work (/stop reached its handler above — still an escape hatch).
        # Compression protection: an interrupt would start a new turn on the pre-rotation parent
        # while compression rotates the id away, forking orphaned siblings.
        if self._agent_has_active_subagents(running_agent):
            _demote = "because the running agent has active subagents (#30170)"
        elif await self._session_has_compression_in_flight(_quick_key):
            _demote = "because context compression is in flight (#56391)"
        else:
            await self._hm_busy_interrupt(event, source, running_agent, _quick_key)
            return None
        logger.info("PRIORITY interrupt demoted to queue for session %s %s", _quick_key, _demote)
        self._queue_or_replace_pending_event(_quick_key, event)
        return None

    def _hm_quick_commands(self) -> dict:
        """User-defined ``quick_commands`` mapping from config (empty dict when unset/malformed)."""
        cfg = self.config
        qc = (cfg.get("quick_commands") if isinstance(cfg, dict) else getattr(cfg, "quick_commands", None)) or {}
        return qc if isinstance(qc, dict) else {}

    @staticmethod
    def _hm_expand_alias_quick_command(event: "MessageEvent", qcmd: dict) -> Optional[str]:
        """Rewrite ``event.text`` to an alias quick command's target; returns the new command name."""
        target = (qcmd.get("target") or "").strip()
        if not target:
            return None
        target = target if target.startswith("/") else f"/{target}"
        event.text = f"{target} {event.get_command_args().strip()}".strip()
        target_command = target.lstrip("/")
        return target_command.split()[0] if target_command else target_command

    async def _hm_command_hooks(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str, command: str, canonical: str
    ) -> Tuple[bool, Optional[str], Optional[str]]:
        """Fire ``pre_command`` (observer) and ``command:<canonical>`` (interceptor) hooks →
        ``(handled, result, new_command)`` (``new_command`` set when a handler rewrote the command).
        The running-agent path deliberately does NOT fire these — a slow or hostile plugin must not
        interfere with the operator's escape hatches for a live agent."""
        raw_args = event.get_command_args().strip()
        platform = source.platform.value if source.platform else ""
        try:
            from hermes_cli.plugins import fire_pre_command_hook
            fire_pre_command_hook(
                surface="gateway", command=str(canonical), alias_used=str(command),
                args_raw=raw_args, session_key=_quick_key, platform=platform,
            )
        except Exception as _pre_cmd_err:
            logger.debug("pre_command hook dispatch failed (non-fatal): %s", _pre_cmd_err)

        # Handlers may return ``{"decision": "deny" | "handled" | "rewrite", ...}`` to intercept
        # dispatch; handlers returning nothing behave as plain observers.
        hook_ctx = {
            "platform": platform, "user_id": source.user_id, "command": canonical,
            "raw_command": command, "args": raw_args, "raw_args": raw_args,
        }
        try:
            hook_results = await self.hooks.emit_collect(f"command:{canonical}", hook_ctx)
        except Exception as _hook_err:
            logger.debug("command:%s hook dispatch failed (non-fatal): %s", canonical, _hook_err)
            hook_results = []

        for hook_result in hook_results:
            if not isinstance(hook_result, dict):
                continue
            decision = str(hook_result.get("decision", "")).strip().lower()
            message = hook_result.get("message")
            message = message if isinstance(message, str) and message else None
            if decision == "deny":
                return True, message or f"Command `/{command}` was blocked by a hook.", None
            if decision == "handled":
                return True, message, None
            if decision == "rewrite":
                new_command = str(hook_result.get("command_name", "")).strip().lstrip("/")
                if new_command:
                    event.text = f"/{new_command} {str(hook_result.get('raw_args', '')).strip()}".strip()
                    return False, None, event.get_command()
        return False, None, None

    async def _hm_resolve_command(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str
    ) -> Tuple[bool, Optional[str], Optional[str], Optional[str]]:
        """Resolve the slash command (aliases, access gate, hooks) → ``(handled, result, command,
        canonical)``; when ``handled`` the caller returns ``result`` as-is (may be None)."""
        from hermes_cli.commands import is_gateway_known_command, resolve_command as _resolve_cmd

        def _canon(cmd):
            # Aliases resolve to the canonical name so dispatch and hook names don't depend on them.
            _def = _resolve_cmd(cmd) if cmd else None
            return _def, (_def.name if _def else cmd)

        command = event.get_command()
        _cmd_def, canonical = _canon(command)

        # Expand alias quick commands before built-in dispatch so targets like /model openai/gpt-5.5
        # --provider openrouter reach the /model handler. Built-ins keep precedence: aliases only
        # need early handling when the typed command is not already known.
        if command and _cmd_def is None:
            qcmd = self._hm_quick_commands().get(command)
            if qcmd is not None and qcmd.get("type") == "alias":
                new_command = self._hm_expand_alias_quick_command(event, qcmd)
                if new_command is not None:
                    command = new_command
                    _cmd_def, canonical = _canon(command)

        if not (command and canonical and is_gateway_known_command(canonical)):
            return False, None, command, canonical

        # Per-platform slash access control: only active when the operator set ``allow_admin_from``
        # for the source's scope; then non-admins get ``user_allowed_commands`` plus the
        # /help, /whoami floor. Plain chat is never gated.
        _denied = self._check_slash_access(source, canonical)
        if _denied is not None:
            return True, _denied, command, canonical

        _handled, _result, new_command = await self._hm_command_hooks(
            event, source, _quick_key, command, canonical
        )
        if _handled:
            return True, _result, command, canonical
        if new_command is not None:
            command = new_command
            _cmd_def, canonical = _canon(command)
        return False, None, command, canonical

    async def _hm_confirm_destructive(self, event, command: str, detail: str, handler) -> Tuple[bool, Optional[str]]:
        async def _execute():
            return await handler(event)
        return True, await self._maybe_confirm_destructive_slash(
            event=event, command=command, title=f"/{command}", detail=detail, execute=_execute,
        )

    async def _hm_cmd_new(self, event, source, _quick_key):
        if await asyncio.to_thread(self._is_telegram_topic_root_lobby, source):
            return True, self._telegram_topic_root_new_message()
        return await self._hm_confirm_destructive(
            event, "new", "This starts a fresh session and discards the current conversation history.",
            self._handle_reset_command,
        )

    async def _hm_cmd_start(self, event, source, _quick_key):
        if event.get_command_args().strip().startswith("wisdom_"):
            return True, await self._continue_wisdom_start(event, source)
        logger.info("Ignoring /start platform ping for session %s", _quick_key)
        return True, ""

    async def _hm_cmd_egress(self, event, source, _quick_key):
        from hermes_cli.proxy_cli import format_status_text
        return True, format_status_text()

    async def _hm_rewrite_turn_to_prompt(self, event, source, name: str, ack: str, build) -> Tuple[bool, Optional[str]]:
        """Ack, then rewrite the turn to ``build()`` and fall through to the agent (keeps role
        alternation; works on any backend). A failing builder replies with a retry hint."""
        await self._send_command_ack(source, ack, name)
        try:
            event.text = build()
        except Exception:
            return True, f"Could not start /{name} — please try again."
        return False, None

    # /learn and /plan: ack, rewrite the turn to a builder prompt, fall through to the agent.
    async def _hm_cmd_learn(self, event, source, _quick_key):
        from agent.learn_prompt import build_learn_prompt

        req = event.get_command_args().strip()
        _ack = f"Learning a skill from {'what you described' if req else 'this conversation'}…"
        return await self._hm_rewrite_turn_to_prompt(event, source, "learn", _ack, lambda: build_learn_prompt(req))

    async def _hm_cmd_plan(self, event, source, _quick_key):
        from agent.plan_prompt import build_plan_prompt

        task = event.get_command_args().strip()
        _ack = f"Planning: {task[:80]}{'…' if len(task) > 80 else ''}" if task else "Planning from this conversation's context…"
        return await self._hm_rewrite_turn_to_prompt(event, source, "plan", _ack, lambda: build_plan_prompt(task))

    async def _hm_cmd_init(self, event, source, _quick_key):
        # /init builds the prompt first: the ack wording depends on whether AGENTS.md exists.
        from hermes_cli.init_command import build_init_prompt_for_cwd

        try:
            _init_prompt = build_init_prompt_for_cwd(extra=event.get_command_args().strip())
        except Exception:
            return True, "Could not start /init — please try again."
        _ack = (
            "Updating AGENTS.md from a project scan…"
            if "UPDATE the existing AGENTS.md" in _init_prompt
            else "Generating AGENTS.md from a project scan…"
        )
        await self._send_command_ack(source, _ack, "init")
        event.text = _init_prompt
        return False, None

    async def _hm_cmd_blueprint(self, event, source, _quick_key):
        _blueprint_result = await self._handle_blueprint_command(event)
        _text = getattr(_blueprint_result, "text", "") or ""
        _blueprint_seed = getattr(_blueprint_result, "agent_seed", None)
        if not _blueprint_seed:
            return True, _text or None
        # Blueprint matched — rewrite the turn to the seed and fall through so the agent collects
        # each slot value conversationally, then calls the cronjob tool (the /steer pattern).
        if _text:
            await self._send_command_ack(source, _text, "blueprint")
        try:
            event.text = _blueprint_seed
        except Exception:
            return True, _text or None
        return False, None

    async def _hm_cmd_undo(self, event, source, _quick_key):
        _undo_n = 1
        _undo_raw = event.get_command_args().strip()
        if _undo_raw:
            with suppress(ValueError, IndexError):
                _undo_n = max(1, int(_undo_raw.split()[0]))
        _undo_detail = (
            "This removes the last user/assistant exchange from history."
            if _undo_n == 1
            else f"This removes the last {_undo_n} user turns from history."
        )
        return await self._hm_confirm_destructive(event, "undo", _undo_detail, self._handle_undo_command)

    # /queue and /steer on the idle path: no agent is running, so strip the prefix and send the
    # payload as a regular user turn; an empty payload surfaces the usage hint.
    async def _hm_cmd_queue(self, event, source, _quick_key):
        return self._hm_send_payload_as_turn(event, "Usage: /queue <prompt>")

    async def _hm_cmd_steer(self, event, source, _quick_key):
        return self._hm_send_payload_as_turn(
            event, "Usage: /steer <prompt>  (no agent is running; sending as a normal message)"
        )

    @staticmethod
    def _hm_send_payload_as_turn(event, usage: str) -> Tuple[bool, Optional[str]]:
        payload = event.get_command_args().strip()
        if not payload:
            return True, usage
        with suppress(Exception):
            event.text = payload
        return False, None

    async def _hm_cmd_moa(self, event, source, _quick_key):
        # /moa is one-shot sugar only: run a single prompt through the default MoA preset, then
        # restore the prior model. To *switch* to a MoA preset for the session, pick it from the
        # model picker (MoA presets surface as a virtual "Mixture of Agents" provider).
        from hermes_cli.moa_config import moa_usage, normalize_moa_config
        from hermes_cli.config import load_config

        moa_payload = event.get_command_args().strip()
        if not moa_payload:
            return True, moa_usage()
        try:
            cfg = load_config()
            moa_cfg = normalize_moa_config(cfg.get("moa") if isinstance(cfg, dict) else {})
        except Exception:
            moa_cfg = normalize_moa_config({})
        try:
            event.text = moa_payload
            _moa_state = self._session_state(_quick_key)
            event._moa_restore_override = _moa_state.conversation.model_override
            _moa_state.conversation.model_override = {
                "provider": "moa", "model": moa_cfg["default_preset"], "base_url": "moa://local",
                "api_key": "moa-virtual-provider", "api_mode": "chat_completions",
            }
            self._evict_cached_agent(_quick_key)
            event._moa_disable_after_turn = True
        except Exception:
            return True, "Failed to prepare MoA turn."
        return False, None

    # Idle-path built-ins with bespoke flow (confirmations, prompt rewrites, one-shot MoA), each
    # handled by ``_hm_cmd_<name>`` → ``(handled, result)``; ``(False, None)`` falls through to the agent.
    _HM_CANONICAL_COMMANDS = frozenset({
        "new", "start", "egress", "learn", "plan", "init", "blueprint", "undo", "queue", "steer", "moa",
    })

    async def _hm_dispatch_canonical_command(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str,
        canonical: Optional[str],
    ) -> Tuple[bool, Optional[str]]:
        """Dispatch built-in idle-path commands → ``(handled, result)``; prompt-rewriting commands
        mutate ``event.text`` and return ``(False, None)`` to fall through to the agent."""
        plain_handler = (
            self._gateway_plain_command_handlers().get(canonical)
            or self._gateway_idle_command_handlers().get(canonical)
        )
        if plain_handler is not None:
            return True, await plain_handler(event)
        if canonical in self._HM_CANONICAL_COMMANDS:
            return await getattr(self, f"_hm_cmd_{canonical}")(event, source, _quick_key)
        return False, None

    async def _hm_run_exec_quick_command(self, command: str, exec_cmd: str) -> str:
        """Run a ``type: exec`` quick command in the gateway process (30 s cap, sanitized env — the
        gateway process has every API key in os.environ; output is redacted too)."""
        try:
            from tools.environments.local import build_subprocess_env
            proc = await asyncio.create_subprocess_shell(
                exec_cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE,
                env=build_subprocess_env(),
            )
            stdout, stderr = await asyncio.wait_for(proc.communicate(), timeout=30)
            output = (stdout or stderr).decode().strip()
            if output:
                from agent.redact import redact_sensitive_text
                output = redact_sensitive_text(output)
            return output or "Command returned no output."
        except asyncio.TimeoutError:
            return "Quick command timed out (30s)."
        except Exception as e:
            return f"Quick command error: {e}"

    async def _hm_dispatch_quick_and_plugin_commands(
        self, event: "MessageEvent", source: SessionSource, command: Optional[str]
    ) -> Tuple[bool, Optional[str], Optional[str]]:
        """Drain gate, user-defined quick commands (exec/alias) and plugin slash commands →
        ``(handled, result, command)``; an alias quick command rewrites ``command``."""
        if self._draining:
            return True, f"⏳ Gateway is {self._status_action_gerund()} and is not accepting new work right now.", command

        # User-defined quick commands (bypass agent loop, no LLM call)
        qcmd = self._hm_quick_commands().get(command) if command else None
        if qcmd is not None:
            # Quick commands are slash capabilities too — and type:exec ones run a shell command in
            # the gateway process. They are never in the registry, so the early gate never fires for
            # them; apply the same admin/user policy to the raw typed name here.
            # The early gate above only fires for registry-known commands, so quick commands (never in the
            # registry) would otherwise reach this dispatch sink unchecked. (#44727)
            _denied = self._check_slash_access(source, command)
            if _denied is not None:
                return True, _denied, command
            qtype = qcmd.get("type")
            if qtype == "exec":
                exec_cmd = qcmd.get("command", "")
                if not exec_cmd:
                    return True, f"Quick command '/{command}' has no command defined.", command
                return True, await self._hm_run_exec_quick_command(command, exec_cmd), command
            if qtype != "alias":
                return True, f"Quick command '/{command}' has unsupported type (supported: 'exec', 'alias').", command
            new_command = self._hm_expand_alias_quick_command(event, qcmd)
            if new_command is None:
                return True, f"Quick command '/{command}' has no target defined.", command
            command = new_command  # Fall through to normal command dispatch below

        # Plugin-registered slash commands. Underscores normalize to hyphens so Telegram's
        # underscored autocomplete form matches plugin commands registered with hyphens.
        if command:
            try:
                from hermes_cli.plugins import get_plugin_command_handler
                plugin_handler = get_plugin_command_handler(command.replace("_", "-"))
                if plugin_handler:
                    result = plugin_handler(event.get_command_args().strip())
                    if asyncio.iscoroutine(result):
                        result = await result
                    return True, str(result) if result else None, command
            except Exception as e:
                logger.warning("Plugin command dispatch failed: %s", e)
        return False, None, command

    def _hm_bundle_slash_rewrite(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str, command: str
    ) -> bool:
        """Rewrite ``/<bundle>`` to the bundle invocation message; True when handled.
        Skill bundles take precedence over individual skill commands (mirrors CLI dispatch)."""
        try:
            from agent.skill_bundles import (
                build_bundle_invocation_message, resolve_bundle_command_key
            )
            bundle_key = resolve_bundle_command_key(command)
            if bundle_key is None:
                return False
            # Pass the platform explicitly: bundle skill loading bypasses get_skill_commands()'
            # scan-time disabled filter, and one gateway process serves several platforms, so
            # env-var platform resolution can't be trusted here.
            # Mirrors the stacked-skill gate (#58888).
            bundle_result = build_bundle_invocation_message(
                bundle_key, event.get_command_args().strip(), task_id=_quick_key,
                platform=source.platform.value if source.platform else None,
            )
            if not bundle_result:
                return False
            event.text, _loaded, missing = bundle_result
            if missing:
                logger.info("Bundle %s skipped missing skills: %s", bundle_key, ", ".join(missing))
            return True  # Fall through to normal message processing with bundle content
        except Exception as exc:
            logger.warning("Bundle dispatch failed: %s", exc)
            return False

    @staticmethod
    def _hm_unknown_slash_reply(command: str, source: SessionSource) -> Optional[str]:
        """Reply for a /command that is not built-in/plugin/skill; None when it is known."""
        from gateway.run import _check_unavailable_skill
        from hermes_cli.commands import GATEWAY_KNOWN_COMMANDS
        # Known-but-disabled or uninstalled skill → actionable guidance.
        _unavail_msg = _check_unavailable_skill(command)
        if _unavail_msg:
            return _unavail_msg
        # Genuinely unrecognized: warn instead of forwarding to the LLM as free text (it invents
        # tool calls). Normalize to hyphenated form first: the quick-command block may have set an
        # alias target, so the resolved def can be stale.
        if command.replace("_", "-") in GATEWAY_KNOWN_COMMANDS:
            return None
        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."
        )

    def _hm_skill_slash_rewrite(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str, command: Optional[str]
    ) -> Optional[str]:
        """Rewrite ``/<bundle>`` / ``/<skill>`` invocations into the skill prompt on ``event.text``;
        returns a reply string when the command is disabled/unknown/failed, else None.
        resolve_skill_command_key() handles the Telegram underscore/hyphen round-trip (/claude_code)."""
        if not command or self._hm_bundle_slash_rewrite(event, source, _quick_key, command):
            return None
        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 None:
                return self._hm_unknown_slash_reply(command, source)
            _plat = source.platform.value if source.platform else None
            user_instruction = event.get_command_args().strip()
            # Stacked slash-skill invocations: `/skill-a /skill-b do XYZ` loads every leading skill
            # (up to 5), not just the first. Mirrors CLI.
            try:
                from agent.skill_commands import (
                    build_stacked_skill_invocation_message as _build_stacked,
                    split_stacked_skill_commands,
                )
                extra_keys, stacked_instruction = split_stacked_skill_commands(user_instruction)
            except Exception:
                _build_stacked = None
                extra_keys, stacked_instruction = [], user_instruction
            _skill_name = skill_cmds[cmd_key].get("name", "")
            if _plat and (_skill_name or extra_keys):
                # Per-platform disabled check: get_skill_commands() only applies the *global*
                # disabled list at scan time (process-global cache across platforms), and
                # split_stacked_skill_commands() only checks each extra token is a KNOWN skill.
                from agent.skill_utils import get_disabled_skill_names as _get_plat_disabled
                _plat_disabled = _get_plat_disabled(platform=_plat)
                if _skill_name and _skill_name in _plat_disabled:
                    return (
                        f"The **{_skill_name}** skill is disabled for {_plat}.\n"
                        f"Enable it with: `hermes skills config`"
                    )
                _disabled_extra = [
                    skill_cmds.get(k, {}).get("name", "")
                    for k in extra_keys
                    if skill_cmds.get(k, {}).get("name", "") in _plat_disabled
                ]
                if _disabled_extra:
                    return (
                        f"The **{', '.join(_disabled_extra)}** skill(s) in this "
                        f"stacked invocation are disabled for {_plat}.\n"
                        f"Enable them with: `hermes skills config`"
                    )
            if extra_keys and _build_stacked is not None:
                stacked_result = _build_stacked(
                    [cmd_key, *extra_keys], stacked_instruction, task_id=_quick_key,
                )
                if not stacked_result:
                    return f"Failed to load stacked skills for /{command}."
                event.text, _loaded, _missing = stacked_result
            else:
                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
        except Exception as e:
            logger.debug("Skill command check failed (non-fatal): %s", e)
        return None

    async def _hm_pending_reply_intercepts(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str
    ) -> Optional[str]:
        """Replies owned by in-flight work: pending /update prompt, clarify, slash-confirm.
        Only events that may control the gateway (``allow_gateway_control``) can answer them."""
        if not event.allow_gateway_control:
            return None
        _reply = self._hm_update_prompt_reply(event, _quick_key)
        if _reply is None:
            _reply = await self._hm_clarify_reply(event, source, _quick_key)
        if _reply is None:
            _reply = await self._hm_slash_confirm_reply(event, _quick_key)
        return _reply

    async def _hm_dispatch_idle_commands(
        self, event: "MessageEvent", source: SessionSource, _quick_key: str
    ) -> Tuple[bool, Optional[str]]:
        """Idle path: resolve + dispatch slash commands; rewriting commands fall through to the agent."""
        _handled, _result, command, canonical = await self._hm_resolve_command(event, source, _quick_key)
        if not _handled:
            _handled, _result = await self._hm_dispatch_canonical_command(event, source, _quick_key, canonical)
        if not _handled:
            _handled, _result, command = await self._hm_dispatch_quick_and_plugin_commands(event, source, command)
        if not _handled:
            _result = self._hm_skill_slash_rewrite(event, source, _quick_key, command)
            _handled = _result is not None
        return _handled, _result

    def _hm_rescue_orphaned_fifo(
        self, event: "MessageEvent", source: SessionSource, is_internal: bool, _quick_key: str
    ) -> Tuple["MessageEvent", SessionSource, bool]:
        """FIFO orphan rescue: a session that went idle with a populated overflow (post-turn drain
        never promoted, e.g. a compression-demoted follow-up) silently orphaned those events. The
        oldest orphan runs as THIS turn and the incoming event is parked behind the chain. Skipped
        for control commands and internal events."""
        try:
            # ── FIFO orphan rescue (#99882) ──────────────────────────────── If this session went idle with
            # a populated overflow (queued during a busy window whose post-turn drain never promoted — e.g.
            # a compression-demoted follow-up after the compression window ended through an exit that
            # skipped the promotion site), those events were silently orphaned. We are starting the next
            # turn for this session NOW: re-stage the orphans in FIFO order and enqueue the incoming event
            # behind them, so arrival order (#28503) holds: oldest orphan runs as this turn, the rest drain
            # in order, the new message last.
            _orphan_adapter = self._adapter_for_source(source)
            if _orphan_adapter is None or getattr(event, "internal", False) or event.get_command():
                return event, source, is_internal
            _rescued = self._rescue_orphaned_overflow(_quick_key, _orphan_adapter)
            if _rescued is None:
                return event, source, is_internal
            # Into the slot when the chain was a single orphan (post-turn drain picks it up),
            # otherwise into overflow behind the already-staged next orphan.
            self._enqueue_fifo(_quick_key, event, _orphan_adapter)
            # Same session key by construction; carry the orphan's own source so reply anchors /
            # thread metadata point at the message actually being answered.
            _rescued_source = getattr(_rescued, "source", None)
            source = _rescued_source if _rescued_source is not None else source
            return _rescued, source, bool(getattr(_rescued, "internal", False))
        except Exception:
            logger.debug("FIFO orphan rescue pre-claim failed for %s", _quick_key, exc_info=True)
            return event, source, is_internal

    async def _handle_message(self, event: MessageEvent) -> Optional[str]:
        """Handle an incoming message from any platform: auth → command check → running-agent
        interrupt → get/create session → build context → run agent → return response."""
        from gateway.run import _AGENT_PENDING_SENTINEL
        _admitted = await self._hm_admit_event(event)
        if _admitted is None:
            return None
        event, source, is_internal = _admitted
        # TERMINAL-DECLINE LATCH TEARDOWN. Deliberately placed AFTER admission,
        # not on the adapter's raw inbound: profile routing, the ignored-channel
        # guard, plugin hooks and user authorization all reject events above,
        # and a rejected event must not be able to clear a refusal belonging to
        # an active turn. This is also the single entry point every lane shares
        # — Discord interaction passthrough builds its own MessageEvent and
        # calls handle_message directly, so a teardown on the relay's inbound
        # handler left those turns muted.

        _paused_notice = self._hm_estop_gate(event, source, is_internal)
        if _paused_notice is not None:
            return _paused_notice

        _quick_key = self._session_key_for_source(source)
        _reply = await self._hm_pending_reply_intercepts(event, source, _quick_key)
        if _reply is not None:
            return _reply

        # Evict a leaked/reaped ``_running_agents`` slot before the busy-session fast-path.
        self._hm_evict_idle_stale_agent(_quick_key)
        if self._is_session_running(_quick_key):
            self._hm_evict_reaped_agent(_quick_key)
        if self._is_session_running(_quick_key):
            return await self._hm_handle_running_session_message(event, source, _quick_key)

        _handled, _result = await self._hm_dispatch_idle_commands(event, source, _quick_key)
        if _handled:
            return _result

        # Pending exec approvals go through /approve and /deny only — no bare-text matching, or a
        # conversational "yes" would execute a dangerous command.
        if not is_internal:
            if await asyncio.to_thread(self._is_telegram_topic_root_lobby, source):
                # Debounced so a user who forgets about topic mode doesn't get ten reminders.
                if self._should_send_telegram_lobby_reminder(source):
                    return self._telegram_topic_root_lobby_message()
                return None
            # External-drain new-turn gate: when NAS engaged an external drain (.drain_request.json,
            # seen by _drain_control_watcher), refuse to START new turns so the in-flight set can
            # only fall to zero. Reversible.
            if self._external_drain_active:
                logger.info("Refusing new turn for session %s — external drain active.", _quick_key)
                return (
                    "⏳ This agent is draining for a maintenance action and isn't "
                    "accepting new turns right now. It'll be back in a moment — "
                    "please resend shortly."
                )

        # Claim this session before any await: many awaits sit between here and _run_agent
        # registering the real AIAgent; without this sentinel a second message during any of them
        # passes the "already running" guard and spins up a duplicate agent for the same session.
        _active_session_lease, _limit_message = self._claim_active_session_slot(_quick_key, source)
        if _limit_message is not None:
            logger.info("Rejecting new active session %s: max_concurrent_sessions reached", _quick_key)
            return _limit_message

        event, source, is_internal = self._hm_rescue_orphaned_fifo(event, source, is_internal, _quick_key)

        _claim_state = self._session_state(_quick_key)
        if _active_session_lease is not None:
            _claim_state.turn.lease = _active_session_lease
        _claim_state.turn.agent = _AGENT_PENDING_SENTINEL
        _claim_state.turn.started_ts = time.time()
        self._persist_active_agents()
        _run_generation = self._begin_session_run_generation(_quick_key)

        try:
            try:
                _agent_result = await self._handle_message_with_agent(event, source, _quick_key, _run_generation)
            except TurnLeaseTimeoutError as exc:
                # A rejected message, not a completed turn: return before the /goal judge so it
                # cannot consume the resend notice and enqueue a synthetic continuation loop.
                logger.error(
                    "Rejecting turn for routing key %s on session %s after "
                    "turn-lease timeout; transcript load was not started and "
                    "the user must resend",
                    _quick_key, exc.session_id,
                )
                return (
                    "⏳ Another turn is still running on this session. To "
                    "protect the transcript, this message was not processed. "
                    "Wait for the active turn to finish, then resend it."
                )
            try:
                await self._run_post_turn_hooks(
                    agent_result=_agent_result, source=source, is_internal=is_internal, event=event,
                )
            except Exception as _goal_exc:
                logger.debug("post-turn hook failed: %s", _goal_exc)
            return _agent_result
        finally:
            # MoA one-shot restore must run on EVERY exit path (success, exception, interrupt):
            # the restore data lives on the per-turn event and would leak permanently otherwise.
            self._restore_moa_one_shot(event, _quick_key)
            self._restore_pending_one_turn_model_override(_quick_key)
            # SIGKILL/OOM skips finally, leaving the durable marker for the next unclean startup's
            # recovery pass.
            await self._clear_durable_active_turn(event)
            # Unconditional, idempotent release without a run_generation guard: evicts the zombie
            # left when session_reset bumps the generation mid-flight (gen-N's guarded release in
            # _run_agent returns False; a sentinel-only check would lock forever).
            self._release_running_agent_state(_quick_key)
            # Turn lease is keyed by (routing key, run generation) so this unwind can only free
            # the lease its own turn acquired, never a newer turn's.
            # Unconditional release covers every exit path. _release_running_agent_state is idempotent
            # (pop-on-absent is harmless) and, called without a run_generation guard, always clears the slot
            # regardless of which generation it holds. This evicts the zombie left when session_reset bumps
            # the generation (N -> N+1) mid-flight: gen-N's guarded release inside _run_agent returns False,
            # and the old sentinel-only check here missed the leftover real agent — locking the session out
            # forever (#28686).
            self._release_turn_lease(_quick_key, _run_generation)

    def _restore_moa_one_shot(self, event: "MessageEvent", quick_key: str) -> None:
        """Revert a ``/moa <prompt>`` one-shot model override after its turn (called from the
        message-handling ``finally``). ``_moa_restore_override`` holds the prior per-session
        override (``None`` = clear the MoA override outright)."""
        if not getattr(event, "_moa_disable_after_turn", False):
            return
        with suppress(Exception):
            self._session_state(quick_key).conversation.model_override = getattr(event, "_moa_restore_override", None)
            self._evict_cached_agent(quick_key)

    def _restore_pending_one_turn_model_override(self, session_key: str) -> None:
        """Restore a per-session model override after ``/model --once`` runs."""
        if not session_key:
            return
        try:
            _otr_state = self._peek_session_state(session_key)
            snapshot = _otr_state.conversation.one_turn_restore if _otr_state else None
            if _otr_state is not None:
                _otr_state.conversation.one_turn_restore = None
            if snapshot:
                self._restore_session_model_override(session_key, snapshot)
        except Exception:
            logger.debug("Failed to restore one-turn model override", exc_info=True)

    def _prefix_inbound_sender_context(self, event: MessageEvent, source: SessionSource, message_text: str) -> str:
        """Attribute the sender in shared multi-user sessions and prepend history-backfill channel context."""
        _is_shared_multi_user = is_shared_multi_user_session(
            source, group_sessions_per_user=getattr(self.config, "group_sessions_per_user", True),
            thread_sessions_per_user=getattr(self.config, "thread_sessions_per_user", False),
        )
        if _is_shared_multi_user and source.user_name:
            # Display names are attacker-influenceable: neutralize newlines/control chars or a
            # hostile name masquerades as a fake markdown section (mirrors build_session_context_prompt).
            _safe_user_name = neutralize_untrusted_inline_text(source.user_name)
            # Slack: expose the CURRENT speaker's verifiable `<@U...>` id so "mention me again" has a
            # trusted target (display names are ambiguous). user_id comes from the envelope, not user-editable.
            # See #17916.
            if source.platform == Platform.SLACK and source.user_id:
                _safe_user_name = f"{_safe_user_name} | Slack user <@{source.user_id}>"
            message_text = f"[{_safe_user_name}] {message_text}"
        # After the sender-prefix so the prefix applies only to the trigger message, not the backfill.
        if getattr(event, "channel_context", None):
            message_text = f"{event.channel_context}\n\n[New message]\n{message_text}"
        return message_text

    @staticmethod
    def _classify_inbound_media(
        event: MessageEvent, pending_stt_prepared: bool
    ) -> Tuple[list, list, list, list]:
        """Split ``event.media_urls`` into (image, STT-voice, audio-file, video) paths. Per-attachment
        MIME wins over the message-level type (a document sent alongside an image must not be routed
        as an image). MessageType.AUDIO / mixed DOCUMENT audio is a file attachment, never STT."""
        from gateway.run import _event_media_is_audio, _event_media_is_image, _event_media_is_stt_input
        image_paths, audio_paths, audio_file_paths, video_paths = [], [], [], []
        for i, path in enumerate(event.media_urls or []):
            mtype = event.media_types[i] if i < len(event.media_types) else ""
            if _event_media_is_image(event, i):
                image_paths.append(path)
            if _event_media_is_audio(event, i):
                if event.message_type in {MessageType.AUDIO, MessageType.DOCUMENT}:
                    audio_file_paths.append(path)
                elif not pending_stt_prepared and _event_media_is_stt_input(event, i):
                    audio_paths.append(path)
            if mtype.startswith("video/") or (not mtype and event.message_type == MessageType.VIDEO):
                video_paths.append(path)
        return image_paths, audio_paths, audio_file_paths, video_paths

    async def _enrich_inbound_images(
        self, source: SessionSource, session_key: str, message_text: str, image_paths: list[str]
    ) -> str:
        """Route images natively (attach pixels at run_conversation) or pre-analyze them into text."""
        # See agent/image_routing.py. Offloaded to a thread: the decision does blocking network I/O
        # (models.dev fetch on cache miss, Ollama /api/show probe) that would stall the event loop.
        _img_mode = await asyncio.to_thread(
            self._decide_image_input_mode, source=source, session_key=session_key,
        )
        if _img_mode == "native":
            self._session_state(session_key).persistent.native_image_paths = list(image_paths)
            logger.info(
                "Image routing: native (model supports vision). %d image(s) will be attached inline.",
                len(image_paths),
            )
            return message_text
        logger.info(
            "Image routing: text (mode=%s). Pre-analyzing %d image(s) via vision_analyze.",
            _img_mode, len(image_paths),
        )
        # Vision enrichment runs before AIAgent.run_conversation(), so bind this session's resolved
        # runtime explicitly rather than consulting process-global compatibility mirrors.
        vision_runtime = None
        try:
            turn_model, runtime_kwargs = self._resolve_session_agent_runtime(
                source=source, session_key=session_key,
            )
            vision_runtime = {**(runtime_kwargs or {}), "model": turn_model}
        except Exception:
            logger.debug("vision enrichment: session runtime resolution failed", exc_info=True)

        from agent.auxiliary_client import scoped_runtime_main

        with scoped_runtime_main(vision_runtime):
            return await self._enrich_message_with_vision(message_text, image_paths)

    async def _echo_stt_transcripts(
        self, adapter, source: SessionSource, transcripts: List[str], *, metadata=None, log_context: str = "Transcript"
    ) -> None:
        """Send each transcript back as ``🎙️ "…"`` (best-effort; failures are logged, never raised)."""
        for tx in transcripts:
            try:
                await adapter.send(source.chat_id, f'🎙️ "{tx}"', metadata=metadata)
            except Exception as echo_exc:
                logger.debug("%s echo failed (non-fatal): %s", log_context, echo_exc)

    async def _enrich_inbound_voice(
        self, event: MessageEvent, source: SessionSource, message_text: str, audio_paths: list[str]
    ) -> str:
        message_text, _successful_transcripts = await self._enrich_message_with_transcription(
            message_text, audio_paths,
        )
        # Echo each successful transcript back immediately when configured so users can verify STT
        # quality in real time. On transcription failure do NOT send a hardcoded notice: that
        # bypassed the LLM and produced two replies; enrichment leaves one neutral marker instead.
        if _successful_transcripts and self._should_echo_stt_transcripts():
            _echo_adapter = self._adapter_for_source(source)
            if _echo_adapter:
                _echo_meta = self._thread_metadata_for_source(source, self._reply_anchor_for_event(event))
                await self._echo_stt_transcripts(_echo_adapter, source, _successful_transcripts, metadata=_echo_meta)
        return message_text

    @staticmethod
    def _inbound_attachment_display_name(path: str) -> Tuple[str, str]:
        """``(display_name, agent_visible_path)``: cache filename is ``<id>_<id>_<original>``; the
        path is translated to the in-container mount under a Docker backend."""
        from tools.credential_files import to_agent_visible_cache_path
        basename = os.path.basename(path)
        parts = basename.split("_", 2)
        return re.sub(r'[^\w.\- ]', '_', parts[2] if len(parts) >= 3 else basename), to_agent_visible_cache_path(path)

    @classmethod
    def _prepend_inbound_media_file_notes(cls, message_text: str, audio_file_paths: list[str], video_paths: list[str]) -> str:
        """Prepend a path-pointing note per audio-file / video attachment (content is not inlined)."""
        for kind, noun, verb, tool, paths in (
            ("an audio file attachment", "audio", "transcribe or process", "a transcription or media tool", audio_file_paths),
            ("a video attachment", "video", "inspect or process", "a video analysis or media tool", video_paths),
        ):
            for _path in paths:
                _display, _agent_path = cls._inbound_attachment_display_name(_path)
                message_text = (
                    f"[The user sent {kind}: '{_display}'. "
                    f"It is saved at: {_agent_path}. "
                    f"Its content is not inlined here. If the user's request involves "
                    f"what the {noun} contains, {verb} it yourself — for "
                    f"example by passing the path to {tool} — "
                    f"instead of asking the user to describe it. Only ask what to do "
                    f"with it if their intent is genuinely unclear.]"
                    f"\n\n{message_text}"
                )
        return message_text

    @classmethod
    def _prepend_inbound_document_notes(cls, event: MessageEvent, message_text: str) -> str:
        """Prepend a context note per non-media attachment (anything not routed as image/audio/video)."""
        from gateway.run import (
            _build_document_context_note, _event_media_is_audio, _event_media_is_image,
            _event_media_is_video,
        )
        if not event.media_urls:
            return message_text
        import mimetypes as _mimetypes

        _TEXT_EXTENSIONS = {".txt", ".md", ".csv", ".log", ".json", ".xml", ".yaml", ".yml", ".toml", ".ini", ".cfg"}
        inline_flags = getattr(event, "media_text_inlined", None) or []
        for i, path in enumerate(event.media_urls):
            # A document mixed into a PHOTO/VOICE message (message-level type != DOCUMENT) still
            # reaches the agent; only genuine non-media files get a note.
            if any(f(event, i) for f in (_event_media_is_image, _event_media_is_audio, _event_media_is_video)):
                continue
            mtype = event.media_types[i] if i < len(event.media_types) else ""
            if mtype in {"", "application/octet-stream"}:
                _is_text = os.path.splitext(path)[1].lower() in _TEXT_EXTENSIONS
                mtype = "text/plain" if _is_text else (_mimetypes.guess_type(path)[0] or "application/octet-stream")
            # Every accepted file gets a note — a non-text/non-application MIME (font/*, model/*)
            # must still tell the agent the file exists.
            display_name, agent_path = cls._inbound_attachment_display_name(path)
            inline_flag = inline_flags[i] if i < len(inline_flags) else None
            context_note = _build_document_context_note(
                display_name, agent_path, mtype, content_inlined=inline_flag is not False,
            )
            message_text = f"{context_note}\n\n{message_text}"
        return message_text

    @staticmethod
    def _prepend_inbound_reply_context(event: MessageEvent, source: SessionSource, message_text: str) -> str:
        """Prepend the Discord triggering-message id and the reply-to pointer."""
        # Discord: the triggering message id goes on the per-turn user message, never the cached
        # system prompt — it changes every turn and would bust the agent-cache signature.
        if (
            source is not None
            and getattr(source, "platform", None) == Platform.DISCORD
            and getattr(event, "message_id", None)
        ):
            from gateway.session import _discord_tools_loaded as _disc_tools_loaded
            if _disc_tools_loaded():
                message_text = (
                    f"[Triggering message id: `{event.message_id}` — use as "
                    f"`message_id` for reply/react/pin via the discord tools.]\n\n"
                    f"{message_text}"
                )

        if getattr(event, "reply_to_text", None) and event.reply_to_message_id:
            # Always inject the reply-to pointer even when the quoted text is already in history:
            # it's disambiguation (*which* prior message), not deduplication.
            # Adapters resolve the original message (or the user's native partial quote).
            # A preview here silently loses later list items and code; keep that context intact.
            reply_text = event.reply_to_text
            _who = " your previous message" if getattr(event, "reply_to_is_own_message", False) else ""
            message_text = f'[Replying to{_who}: "{reply_text}"]\n\n{message_text}'
        return message_text

    async def _inbound_model_context_length(self, source: SessionSource, session_key: str) -> int:
        """Context length of the model this turn runs on. A global ``model.context_length`` pin
        belongs to the configured model, not a /model or channel override; custom-provider limits win."""
        from gateway.run import _load_gateway_config
        from agent.model_metadata import get_model_context_length_async

        _msg_config_ctx = None
        _msg_cfg = None
        _msg_model_cfg = {}
        _msg_custom_providers = []
        with suppress(Exception):
            _msg_cfg = _load_gateway_config()
            _msg_model_cfg = _msg_cfg.get("model", {})
            if isinstance(_msg_model_cfg, dict):
                _msg_raw_ctx = _msg_model_cfg.get("context_length")
                if _msg_raw_ctx is not None:
                    _msg_config_ctx = int(_msg_raw_ctx)
            try:
                from hermes_cli.config import get_compatible_custom_providers

                _msg_custom_providers = get_compatible_custom_providers(_msg_cfg)
            except Exception:
                _msg_custom_providers = _msg_cfg.get("custom_providers") or []
        # GatewayRunner has no self._model/self._base_url; resolve the session's actual runtime.
        _msg_model, _msg_runtime = self._resolve_session_agent_runtime(
            source=source, session_key=session_key, user_config=_msg_cfg,
        )
        _msg_base_url = _msg_runtime.get("base_url") or ""
        if isinstance(_msg_model_cfg, dict):
            _msg_configured_model = _msg_model_cfg.get("default") or _msg_model_cfg.get("model")
        else:
            _msg_configured_model = _msg_model_cfg  # (no dict → no pin was read; ctx is already None)
        if _msg_model != _msg_configured_model:
            _msg_config_ctx = None
        if _msg_config_ctx is not None:
            try:
                from hermes_cli.route_identity import should_clear_context_pin_async

                if await should_clear_context_pin_async(
                    None, None,  # model match already checked above
                    _msg_model_cfg.get("base_url"), _msg_base_url,
                    _msg_model_cfg.get("provider"), _msg_runtime.get("provider"),
                ):
                    _msg_config_ctx = None
            except Exception:
                _msg_config_ctx = None
        if _msg_custom_providers and _msg_base_url:
            with suppress(Exception):
                from hermes_cli.config import get_custom_provider_context_length

                _msg_config_ctx = get_custom_provider_context_length(
                    model=_msg_model, base_url=_msg_base_url, custom_providers=_msg_custom_providers,
                ) or _msg_config_ctx
        return await get_model_context_length_async(
            _msg_model, base_url=_msg_base_url, api_key=_msg_runtime.get("api_key") or "",
            config_context_length=_msg_config_ctx, provider=_msg_runtime.get("provider") or "",
            custom_providers=_msg_custom_providers,
        )

    async def _expand_inbound_context_references(
        self, source: SessionSource, session_key: str, message_text: str
    ) -> Optional[str]:
        """Expand ``@`` context references; returns None when the injection was refused (user notified)."""
        try:
            from agent.context_references import preprocess_context_references_async

            try:
                from tools.terminal_scope import terminal_env as _ts_env
            except ImportError:
                _ts_env = os.environ.get
            _msg_cwd = _ts_env("TERMINAL_CWD", os.path.expanduser("~"))
            _msg_ctx_len = await self._inbound_model_context_length(source, session_key)
            _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._adapter_for_source(source)
                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.warning("@ context reference expansion failed: %s", exc)
            logger.debug("@ context reference expansion failure detail", exc_info=True)
        return message_text

    async def _prepare_inbound_message_text(
        self, *, event: MessageEvent, source: SessionSource, history: List[Dict[str, Any]],
        session_key: Optional[str] = None,
    ) -> Optional[str]:
        """Prepare inbound event text for the agent. Shared by the normal inbound and queued
        follow-up paths so attribution, image enrichment, STT, document notes, reply context and
        @ references behave the same. Side effect: buffers per-session native image paths when the
        model supports native vision; the caller consumes that buffer at ``run_conversation``."""
        _pending_stt_prepared = hasattr(event, "_gateway_pending_stt_text")
        message_text = (event._gateway_pending_stt_text if _pending_stt_prepared else event.text) or ""
        # Prefer the caller's resolved session key so this write key matches the consume key at the
        # run_conversation site; derive it here only for tests and legacy standalone callers.
        session_key = session_key or self._session_key_for_source(source)
        # Reset only this session's per-call buffer; other sessions may be concurrently preparing.
        self._consume_pending_native_image_paths(session_key)

        message_text = self._prefix_inbound_sender_context(event, source, message_text)
        image_paths, audio_paths, audio_file_paths, video_paths = self._classify_inbound_media(event, _pending_stt_prepared)
        if image_paths:
            message_text = await self._enrich_inbound_images(source, session_key, message_text, image_paths)
        if audio_paths:
            message_text = await self._enrich_inbound_voice(event, source, message_text, audio_paths)
        message_text = self._prepend_inbound_media_file_notes(message_text, audio_file_paths, video_paths)
        message_text = self._prepend_inbound_document_notes(event, message_text)
        if "@" in message_text:
            message_text = await self._expand_inbound_context_references(source, session_key, message_text)
            if message_text is None:
                return None
        # After expansion: the quoted reply is someone else's text and stays literal — an
        # ``@file:`` inside it must never read a local file on the replier's behalf.
        return self._prepend_inbound_reply_context(event, source, message_text)

    async def _prepare_profile_scoped_inbound_message_text(
        self, *, event: MessageEvent, source: SessionSource, history: List[Dict[str, Any]],
        session_key: Optional[str] = None,
    ) -> Optional[str]:
        """Run inbound preprocessing under the routed profile when multiplexed."""
        from gateway.run import _async_profile_runtime_scope
        kwargs = dict(event=event, source=source, history=history, session_key=session_key)
        if getattr(getattr(self, "config", None), "multiplex_profiles", False):
            async with _async_profile_runtime_scope(self._resolve_profile_home_for_source(source)):
                return await self._prepare_inbound_message_text(**kwargs)
        return await self._prepare_inbound_message_text(**kwargs)

    async def _prepare_clarify_reply_text(self, event) -> str:
        """Return raw text or successful voice transcripts for a clarify reply."""
        if not self._pending_event_audio_paths(event):
            return (event.text or "").strip()
        _, successful_transcripts = await self._transcribe_pending_audio_event_once(event, "")
        return "\n\n".join(t.strip() for t in successful_transcripts if t.strip())

    def _consume_pending_native_image_paths(self, session_key: str) -> List[str]:
        state = self._peek_session_state(session_key)
        paths = list(state.persistent.native_image_paths or []) if state is not None else []
        if paths:
            state.persistent.native_image_paths = []
        return paths

    async def _mark_durable_active_turn(self, event: "MessageEvent", session_key: str) -> bool:
        """Persist the exact resolved routing key for this running turn."""
        try:
            token = await self.async_session_store.mark_turn_active(session_key)
        except Exception as exc:
            logger.warning("Could not persist active-turn marker for %s: %s", session_key, exc)
            return False
        if not token:
            return False
        # Private event attributes are process-local ownership state: keep the token out of public
        # metadata, transcripts, and platform payloads.
        event._gateway_active_turn_session_key = session_key
        event._gateway_active_turn_token = token
        return True

    async def _clear_durable_active_turn(self, event: "MessageEvent") -> bool:
        """Best-effort CAS clear of the marker owned by *event* (3 attempts; never blocks agent/lease
        release — a stale marker is bounded by the agent timeout and clean-start discard)."""
        session_key = getattr(event, "_gateway_active_turn_session_key", None)
        token = getattr(event, "_gateway_active_turn_token", None)
        try:
            if not session_key or not token:
                return False
            last_error: Optional[Exception] = None
            for attempt in range(1, 4):
                try:
                    return bool(await self.async_session_store.clear_turn_active(session_key, token))
                except Exception as exc:
                    last_error = exc
                    if attempt < 3:
                        logger.debug(
                            "Retrying active-turn marker cleanup for %s (%d/3): %s",
                            session_key, attempt, exc,
                        )
            logger.warning(
                "Could not clear active-turn marker for %s after 3 attempts: %s", session_key, last_error,
            )
            return False
        finally:
            for attr in ("_gateway_active_turn_session_key", "_gateway_active_turn_token"):
                with suppress(AttributeError):
                    delattr(event, attr)

    def _install_plugin_message_injector(self) -> None:
        """Publish this live gateway's plugin message scheduler."""
        from hermes_cli.plugins import get_plugin_manager

        get_plugin_manager().set_gateway_message_injector(
            self, self._schedule_plugin_message_injection
        )

    def _clear_plugin_message_injector(self) -> None:
        """Remove this runner's scheduler without clobbering a newer owner."""
        from hermes_cli.plugins import get_plugin_manager

        get_plugin_manager().clear_gateway_message_injector(self)

    def _schedule_plugin_message_injection(
        self, *, session_key: str, content: str, plugin_id: str
    ) -> bool:
        """Schedule a plugin-triggered turn on the live gateway loop (thread-safe)."""
        from gateway.run import safe_schedule_threadsafe
        loop = getattr(self, "_gateway_loop", None)
        if not getattr(self, "_running", False) or loop is None or loop.is_closed():
            return False

        coro = self._dispatch_plugin_message_injection(
            session_key=session_key, content=content, plugin_id=plugin_id,
        )
        try:
            current_loop = asyncio.get_running_loop()
        except RuntimeError:
            current_loop = None

        if current_loop is loop:
            try:
                future = loop.create_task(coro)
            except Exception:
                coro.close()
                logger.warning("Plugin message injection scheduling failed", exc_info=True)
                return False
            self._background_tasks.add(future)
            future.add_done_callback(self._background_tasks.discard)
        else:
            future = safe_schedule_threadsafe(
                coro, loop, logger=logger, log_message="Plugin message injection scheduling failed",
                log_level=logging.WARNING,
            )
            if future is None:
                return False

        def _log_result(completed) -> None:
            try:
                if completed.result():
                    return
                what, exc = "was not routed", None
            except (asyncio.CancelledError, concurrent.futures.CancelledError):
                return
            except Exception as err:
                what, exc = "failed", err
            logger.warning(
                "Plugin message injection %s: plugin=%s session=%s", what, plugin_id, session_key, exc_info=exc,
            )

        future.add_done_callback(_log_result)
        return True

    async def _dispatch_plugin_message_injection(
        self, *, session_key: str, content: str, plugin_id: str
    ) -> bool:
        """Route a plugin-triggered turn through the session's live adapter."""
        def _accepting() -> bool:
            return getattr(self, "_running", False) and not getattr(self, "_draining", False)

        if not _accepting():
            return False
        entry = await self.async_session_store.lookup_by_session_key(session_key)
        if entry is None or entry.origin is None or not _accepting():
            return False

        source = dataclasses.replace(entry.origin)
        try:
            authorized = self._is_user_authorized(source, allow_adapter_delegation=False)
        except Exception:
            logger.warning(
                "Plugin message injection authorization check failed: plugin=%s session=%s",
                plugin_id, session_key, exc_info=True,
            )
            return False
        if not authorized:
            logger.warning(
                "Plugin message injection denied by current gateway authorization: "
                "plugin=%s session=%s", plugin_id, session_key,
            )
            return False

        adapter = self._adapter_for_source(source)
        if adapter is None:
            return False

        await adapter.handle_message(MessageEvent(
            text=content, message_type=MessageType.TEXT, source=source, internal=True,
            allow_gateway_control=False,
            metadata={
                "hermes_plugin_id": plugin_id, "hermes_plugin_injection": True,
                "gateway_session_key": session_key, "gateway_session_id": entry.session_id,
                "gateway_session_strict": True,
            },
        ))
        logger.info(
            "Plugin message injection dispatched: plugin=%s session=%s session_id=%s",
            plugin_id, session_key, entry.session_id,
        )
        return True

    def _decide_image_input_mode(
        self, *, source: Optional[SessionSource] = None, session_key: Optional[str] = None,
        user_config: Optional[dict] = None, provider: Optional[str] = None,
        model: Optional[str] = None,
    ) -> str:
        """Resolve image-input routing (``"native"`` / ``"text"``) for the effective model this turn
        (see agent/image_routing.py). Sessions can carry /model overrides and this runs before AIAgent
        sets the auxiliary_client runtime globals, so resolve the per-session runtime bundle the
        upcoming turn will use, not just the persisted default."""
        try:
            from agent.image_routing import decide_image_input_mode
            from agent.auxiliary_client import _read_main_model, _read_main_provider
            from hermes_cli.config import load_config

            cfg = user_config if isinstance(user_config, dict) else load_config()
            resolved_provider = (provider or "").strip()
            resolved_model = (model or "").strip()
            resolved_requested_provider = ""

            if (not resolved_provider or not resolved_model) and (source is not None or session_key):
                try:
                    turn_model, runtime_kwargs = self._resolve_session_agent_runtime(
                        source=source, session_key=session_key, user_config=cfg,
                    )
                    rk = runtime_kwargs if isinstance(runtime_kwargs, dict) else {}
                    if not resolved_model and isinstance(turn_model, str):
                        resolved_model = turn_model.strip()
                    if not resolved_provider and isinstance(rk.get("provider"), str):
                        resolved_provider = rk["provider"].strip()
                    if isinstance(rk.get("requested_provider"), str):
                        resolved_requested_provider = rk["requested_provider"].strip()
                except Exception as exc:
                    logger.debug(
                        "image_routing: session runtime resolution failed, falling back to config — %s",
                        exc,
                    )

            return decide_image_input_mode(
                resolved_provider or _read_main_provider(), resolved_model or _read_main_model(),
                cfg, requested_provider=resolved_requested_provider,
            )
        except Exception as exc:
            logger.debug("image_routing: decision failed, falling back to text — %s", exc)
            return "text"

    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.
        Description *and* local cache path are injected so the model understands the image without
        a tool call and can re-examine it with vision_analyze."""
        from tools.vision_tools import vision_analyze_tool
        from agent.memory_manager import sanitize_context

        analysis_prompt = (
            "Concisely describe this image in 2-4 sentences "
            "(~200 Chinese characters or ~150 English words). "
            "Cover the main subject, key visible text/data/code, and overall context. "
            "If it is a chart, diagram, or scientific figure, include the important "
            "labels, legend, and key values. Skip decorative details."
        )
        enriched_parts = []
        for path in image_paths:
            try:
                logger.debug("Auto-analyzing user image: %s", path)
                result = json.loads(await vision_analyze_tool(image_url=path, user_prompt=analysis_prompt))
                if result.get("success"):
                    description = sanitize_context(result.get("analysis", ""))
                    note = (
                        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:
                    note = (
                        "[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)
                note = (
                    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}]"
                )
            enriched_parts.append(note)
        if not enriched_parts:
            return user_text
        prefix = "\n\n".join(enriched_parts)
        return f"{prefix}\n\n{user_text}" if user_text else prefix

    _EMPTY_TEXT_PLACEHOLDER = "(The user sent a message with no text content)"

    @classmethod
    def _prepend_media_prefix(cls, prefix: str, user_text: str) -> str:
        """``prefix`` + the user's text; the Discord empty-content placeholder is dropped as redundant."""
        if user_text and user_text.strip() != cls._EMPTY_TEXT_PLACEHOLDER:
            return f"{prefix}\n\n{user_text}"
        return prefix

    @staticmethod
    def _untranscribed_audio_note(path: str) -> str:
        """One minimal neutral marker for every STT failure. Never mention "no STT provider" or setup
        steps — persisted in history they make the model keep volunteering STT-setup advice."""
        from tools.credential_files import to_agent_visible_cache_path
        agent_path = to_agent_visible_cache_path(os.path.abspath(path))
        return f"[voice message could not be transcribed automatically; the audio is available at: {agent_path}]"

    async def _transcribe_one_clip(self, path: str, transcribe_audio, transcribe_audio_local_fallback) -> Tuple[Optional[str], str]:
        """``(transcript_or_None, note)`` for one clip via configured STT with local fallback."""
        result = await asyncio.to_thread(transcribe_audio, path, None, "gateway")
        if not result.get("success"):
            fallback = await asyncio.to_thread(transcribe_audio_local_fallback, path)
            if fallback.get("success"):
                logger.info("Configured STT failed for %s; recovered with local STT", path)
                result = fallback
        if not result["success"]:
            logger.info("Voice transcription failed for %s: %s", path, result.get("error", "unknown error"))
            return None, self._untranscribed_audio_note(path)
        transcript = result["transcript"]
        # STT may return success=True with an empty/whitespace transcript (silence, cut-off);
        # empty quotes make the agent reply to nothing and can loop, so emit a sentinel note.
        # See #41603.
        if not (transcript or "").strip():
            return None, (
                "[The user sent a voice message but it came through "
                "empty or inaudible — speech-to-text returned no "
                "words. Do not guess at the content; ask the user "
                "to resend or type it out.]"
            )
        # Plain quoted line: a "The user sent a voice message..." wrapper read as a meta-instruction
        # and made the LLM comment on voice mode instead.
        return transcript, f'"{transcript}"'

    async def _enrich_message_with_transcription(
        self, user_text: str, audio_paths: List[str]
    ) -> tuple[str, List[str]]:
        """Transcribe voice clips with the configured STT provider and prepend the transcripts →
        ``(enriched_text, successful_transcripts)``; the transcripts (input order; empty if every clip
        failed or STT is disabled) let callers echo them back before the agent loop."""
        from gateway.run import _probe_audio_duration
        audio_paths = list(dict.fromkeys(audio_paths))
        if not getattr(self.config, "stt_enabled", True):
            notes = []
            for path in audio_paths:
                abs_path = os.path.abspath(path)
                duration_str = await _probe_audio_duration(abs_path)
                suffix = f" (duration: {duration_str})" if duration_str else ""
                notes.append(f"[The user sent a voice message: {abs_path}{suffix}]")
            return (self._prepend_media_prefix("\n\n".join(notes), user_text) if notes else user_text), []

        try:
            from tools.transcription_tools import (
                transcribe_audio, transcribe_audio_local_fallback
            )
        except ModuleNotFoundError as e:
            logger.error("Transcription module unavailable: %s", e)
            return self._prepend_media_prefix("[voice message could not be transcribed]", user_text), []

        enriched_parts = []
        successful_transcripts: List[str] = []
        for path in audio_paths:
            try:
                logger.debug("Transcribing user voice: %s", path)
                transcript, note = await self._transcribe_one_clip(
                    path, transcribe_audio, transcribe_audio_local_fallback,
                )
                if transcript is not None:
                    successful_transcripts.append(transcript)
                enriched_parts.append(note)
            except Exception as e:
                logger.error("Transcription error: %s", e)
                enriched_parts.append(self._untranscribed_audio_note(path))

        if enriched_parts:
            user_text = self._prepend_media_prefix("\n\n".join(enriched_parts), user_text)
        return user_text, successful_transcripts

    def _pending_event_audio_paths(self, event) -> List[str]:
        """Return STT-eligible paths from a pending voice message."""
        from gateway.run import _event_media_is_stt_input
        return [
            path for i, path in enumerate(getattr(event, "media_urls", None) or [])
            if _event_media_is_stt_input(event, i)
        ]

    async def _transcribe_pending_audio_event_once(
        self, event, user_text: Optional[str] = None
    ) -> tuple[str | None, List[str]]:
        """Transcribe a pending audio event once and cache the result on the event: the interrupt
        monitor and the pending-drain path both need it — one STT call and one echo per message."""
        if hasattr(event, "_gateway_pending_stt_text"):
            return event._gateway_pending_stt_text, list(getattr(event, "_gateway_pending_stt_transcripts", []) or [])
        audio_paths = self._pending_event_audio_paths(event)
        if not audio_paths:
            return user_text if user_text is not None else (getattr(event, "text", None) or None), []
        text = user_text if user_text is not None else (getattr(event, "text", "") or "")
        enriched_text, successful_transcripts = await self._enrich_message_with_transcription(text, audio_paths)
        event._gateway_pending_stt_text = enriched_text
        event._gateway_pending_stt_transcripts = list(successful_transcripts)
        return enriched_text, successful_transcripts

    async def _echo_pending_stt_transcripts_once(
        self, event, adapter, source, transcripts: List[str], *, metadata=None,
        log_context: str = "Transcript",
    ) -> None:
        """Echo pending-event STT transcripts to the chat at most once. Tracked as a COUNT (not a
        set — identical transcripts are distinct deliveries): ``merge_pending_message_event`` can
        append a second voice note and invalidate the cache; the re-run returns earlier transcripts
        as a prefix, so only the unsent tail is echoed."""
        if not transcripts or not self._should_echo_stt_transcripts() or adapter is None:
            return
        already_echoed = int(getattr(event, "_gateway_pending_stt_echoed", 0) or 0)
        event._gateway_pending_stt_echoed = max(already_echoed, len(transcripts))
        await self._echo_stt_transcripts(
            adapter, source, transcripts[already_echoed:], metadata=metadata, log_context=log_context,
        )

    async def _transcribe_and_echo_pending_voice(
        self, event, adapter, source, text: str, *, log_context: str, metadata=_UNSET
    ) -> tuple[str, List[str]]:
        """Transcribe a pending voice event and echo transcripts once → ``(enriched_text,
        transcripts)`` for ``agent.interrupt()`` or the pending-drain flow; ``(text, [])`` when there
        is no STT-eligible media (caller owns the ``_build_media_placeholder`` fallback)."""
        if not self._pending_event_audio_paths(event):
            return text, []
        try:
            enriched_text, transcripts = await self._transcribe_pending_audio_event_once(event, text)
            if metadata is _UNSET:
                metadata = self._thread_metadata_for_source(source, self._reply_anchor_for_event(event))
            await self._echo_pending_stt_transcripts_once(
                event, adapter, source, transcripts, metadata=metadata, log_context=log_context
            )
            return enriched_text or text, transcripts
        except Exception as trans_exc:
            logger.warning("%s transcription failed: %s", log_context, trans_exc)
            return text, []