"""Hot-serve for ``gateway.multiplex_profiles``: keep the served-profile set in step with ``profiles/`` while the multiplexer runs, instead of snapshotting it once at boot. Three things were start-time snapshots: the secondary adapter set (``_start_secondary_profile_adapters``), the ``served_profiles`` record in ``gateway_state.json`` (``_record_served_profiles``) and the cron ticker's ``profile_homes`` list. Everything else (``/p//`` prefixes, profile-route eligibility, handoff/kanban watchers, shared ingress) already reads ``profiles_to_serve()`` / ``_profile_adapters`` live, so reconciling those three is enough for a profile created after boot to be served. ``reconcile_served_profiles`` runs on the loop under one lock, triggered by the ``rescan-profiles`` control verb (``hermes_cli/profiles.py`` create/delete fire it through the control socket) and by the supervised ``_profile_reconcile_watcher`` every ``_PROFILE_RESCAN_INTERVAL_SECS`` as the safety net. A served profile whose ``config.yaml``/``.env`` changed since its adapters were last built is re-scanned too: creators make the profile first and add the bot token afterwards, and without this an adapter-less profile would stay adapter-less forever. The same watcher tick (on every gateway, multiplexed or not) re-syncs a home's shell hooks when its ``shell-hooks-allowlist.json`` (or ``config.yaml``) changed: a non-TTY gateway skips unconsented hooks, so consent recorded afterwards must register them without a restart (#132993). That re-runs hook registration only, never an adapter rebuild. """ from __future__ import annotations import asyncio import logging from functools import partial from pathlib import Path from typing import Any, Dict, Optional from gateway.run_shutdown import _log_suppressed from hermes_cli.config_backend import ConfigBackendUnavailable from utils import file_cache_version logger = logging.getLogger(__name__) _PROFILE_RESCAN_INTERVAL_SECS = 30.0 _PROFILE_SIGNATURE_FILES = (".env",) # config.yaml: the config backend's version, below # Bound for one own-gateway liveness probe (``live_gateway_pid_for_home``): the control-socket # read it can end in has no timeout of its own — on Windows the named pipe stalls there until # its peer answers — so the await needs one. A healthy identify answers in milliseconds. _OWN_GATEWAY_PROBE_TIMEOUT_SECS = 5.0 def _stat_signature(path: Path) -> Optional[tuple]: try: return file_cache_version(path, always_digest=True) # stored, compared on a later tick except OSError: return None def profile_serve_signature(home: Path) -> tuple: """Change detector for a served profile's credentials/config: the config backend's signature of its config layer, then ``file_cache_version`` of ``.env``. Each is hashed every time (``always_digest``) where it is a file: it is compared long after it was taken, and a racy digest going empty would read as a change and reconnect the profile's adapters. A remote layer changes with no local file change, so its signature is the backend's version. A config backend outage raises ``ConfigBackendUnavailable``; the reconcile cycle contains it and retries.""" from hermes_cli.config_backend import config_version_or_none, supports_file_tooling cfg = Path(home) / "config.yaml" sig = [_stat_signature(cfg) if supports_file_tooling() else config_version_or_none(cfg)] for name in _PROFILE_SIGNATURE_FILES: try: sig.append(file_cache_version(Path(home) / name, always_digest=True)) except OSError: sig.append(None) return tuple(sig) def _added_profile_blocker(launch: Path) -> Optional[str]: """Why a single-profile boot must not start multiplexing now that a second profile exists: the ``gateway.auto_multiplex_migration: false`` opt-out on the default profile, else boot's preflight.""" from hermes_cli.gateway_migrate_guards import auto_migration_opted_out from hermes_cli.gateway_multiplex_mode import AUTO_MIGRATION_OPTED_OUT_REASON, implicit_multiplex_blocker if auto_migration_opted_out(launch): return AUTO_MIGRATION_OPTED_OUT_REASON return implicit_multiplex_blocker() class GatewayProfileReconcileMixin: """Runtime reconciliation of the multiplexed served-profile set (hot add / unroute / credential-add).""" _served_profile_homes: Optional[dict[str, Path]] = None _served_profile_signatures: Optional[dict[str, tuple]] = None _profile_reconcile_lock: Optional[asyncio.Lock] = None _profile_own_gateway_warned: Optional[set[str]] = None _profile_probe_timeout_warned: Optional[set[str]] = None _hook_consent_signatures: Optional[dict[str, Optional[tuple]]] = None _single_profile_blocker_logged: Optional[str] = None # ── state helpers ───────────────────────────────────────────────────────────────────────────── def _reconcile_lock(self) -> asyncio.Lock: if self._profile_reconcile_lock is None: self._profile_reconcile_lock = asyncio.Lock() return self._profile_reconcile_lock def served_profile_names(self) -> list: """Profiles this multiplexer currently serves (active first), from the live bookkeeping.""" homes = self._served_profile_homes or {} active = getattr(self, "_primary_profile_name", None) or "default" return ([active] if active in homes or not homes else []) + sorted(n for n in homes if n != active) def _note_served_profiles(self, profile_homes) -> None: """Called by ``_record_served_profiles``: remember the served set and each home's signature.""" homes = {str(name): Path(home) for name, home in profile_homes} self._served_profile_homes = homes sigs = self._served_profile_signatures if isinstance(self._served_profile_signatures, dict) else {} self._served_profile_signatures = {name: sigs.get(name) or profile_serve_signature(home) for name, home in homes.items()} # ── watcher ─────────────────────────────────────────────────────────────────────────────────── async def _profile_reconcile_watcher(self, interval: float = _PROFILE_RESCAN_INTERVAL_SECS) -> None: """Supervised safety net, every ``interval`` seconds: re-sync shell hooks whose consent changed (every gateway), then rescan ``profiles/`` when multiplexing (creators signal the control socket for an immediate rescan); a gateway standalone only because the host had one profile starts multiplexing here once a second one exists.""" while self._running: await asyncio.sleep(interval) if not self._running: return try: await self._sync_hooks_on_consent_change() # A standalone gateway never runs a profile reconcile from the watcher, unless it is the # single-profile boot that starts multiplexing in place for a newly created profile. if self._multiplex_on() or await self._multiplex_for_added_profile(): await self.reconcile_served_profiles(reason="watcher") except asyncio.CancelledError: raise except (Exception, ConfigBackendUnavailable): # a config outage is a SystemExit by design logger.warning("Served-profile reconcile failed; retrying next cycle", exc_info=True) async def _sync_hooks_on_consent_change(self) -> None: """Re-sync the shell hooks of each home (launch + served) whose consent allowlist or config changed since the last tick, under that home's scope: hooks consented since start register, revoked ones stop firing. A home seen for the first time syncs once (consent may have landed after startup registered). Hook registration only: adapters and sessions are never touched.""" from hermes_constants import get_process_hermes_home homes = {Path(h) for h in (get_process_hermes_home(), *(self._served_profile_homes or {}).values())} self._hook_consent_signatures = await asyncio.to_thread( self._resync_changed_hook_consent, homes, self._hook_consent_signatures or {}) def _resync_changed_hook_consent(self, homes, seen: dict) -> dict: """Blocking half of ``_sync_hooks_on_consent_change``; returns the signatures it applied. The key covers ``config.yaml`` too: a sync that read an empty/unreadable config applied nothing, so its recovery must re-run it. A failed sync is left unrecorded and retried next tick.""" from agent.shell_hooks import ALLOWLIST_FILENAME, sync_from_config from gateway.run import _load_gateway_config, _profile_runtime_scope from gateway.run_turn import GatewayTurnMixin current = {} for home in (h.resolve() for h in homes): sig = tuple(_stat_signature(home / name) for name in (ALLOWLIST_FILENAME, "config.yaml")) if seen.get(str(home), ()) == sig: current[str(home)] = sig continue load = partial(_load_gateway_config, home / "config.yaml") scope = (_profile_runtime_scope(home) if self._multiplex_on() else GatewayTurnMixin._standalone_launch_scope()) try: with scope: sync_from_config(load) except Exception: # health: allow BLE001 -- one home's failure never blocks the others logger.warning("shell-hook re-sync failed for %s; retrying next cycle", home, exc_info=True) continue current[str(home)] = sig return current # ── reconcile ───────────────────────────────────────────────────────────────────────────────── async def reconcile_served_profiles(self, *, reason: str = "request") -> dict[str, Any]: """Diff ``profiles/`` (what exists now) against the served set (the process reservation): reserve + build the runtime and start adapters for new profiles, tear down, unroute and unreserve deleted ones, (re)build adapters for served profiles whose config/.env changed. Other profiles' adapters are never touched. A new profile whose home another gateway owns or whose store is unusable is parked (logged, not served) and left for an explicit rescan. Returns ``{"added", "removed", "rescanned", "parked", "served_profiles"}``.""" from gateway.run import MultiplexConfigError from hermes_cli.profiles import profiles_to_serve result: dict[str, Any] = {"added": [], "removed": [], "rescanned": [], "parked": [], "reason": reason} if not self._multiplex_on() and not await self._multiplex_for_added_profile(): return {**result, "multiplex": False, "served_profiles": self.served_profile_names()} if not self._running or self._served_profile_homes is None: # Startup enumerates profiles/ itself; a rescan before it finishes has nothing to diff against. return {**result, "pending": True, "served_profiles": self.served_profile_names()} async with self._reconcile_lock(): active = getattr(self, "_primary_profile_name", None) or "default" live = {str(name): Path(home) for name, home in profiles_to_serve(multiplex=True)} listed = set(live) # a probe skip below is not a deletion: it never clears a parked verdict known = dict(self._served_profile_homes or {}) from gateway.status import live_gateway_pid_for_home blocked = set() timed_out = set() warned = self._profile_own_gateway_warned or set() timeout_warned = self._profile_probe_timeout_warned or set() for name in list(live): if name == active or name in known: continue # The probe can end in a control-socket read that has no timeout of its own # (a Windows named pipe stalls there until its peer answers). Inline on the # loop thread it parked shutdown_watchdog liveness probes and the multiplexer # was hard-killed with exit 75 (#132547). Probe off the loop, bounded: a # stalled probe only wedges one bounded housekeeping worker for this cycle # (wait_for cancels the still-queued await), never the event loop. try: pid = await asyncio.wait_for( self._run_housekeeping_in_executor(live_gateway_pid_for_home, live[name]), timeout=_OWN_GATEWAY_PROBE_TIMEOUT_SECS) except TimeoutError: # Unprovable is not "own gateway running": skip it this cycle without # warning about (or remembering) a gateway that may not exist. Warn once # per stall; a peer that stays wedged repeats at DEBUG every cycle. timed_out.add(name) log = logger.debug if name in timeout_warned else logger.warning log("[MULTIPLEX] Own-gateway probe for profile '%s' timed out; " "not serving it this cycle", name) del live[name] continue if pid is not None: blocked.add(name) if name not in warned: logger.warning("[MULTIPLEX] Profile '%s' still runs its own gateway; " "stop it before the host can serve this profile", name) del live[name] self._profile_own_gateway_warned = blocked self._profile_probe_timeout_warned = timed_out sigs = self._served_profile_signatures or {} from gateway.run_runtime import unpark_profile for name in [n for n in self._parked_profile_names() if n not in listed]: unpark_profile(self, name) parked = self._parked_profile_names() # The watcher never re-parks the same profile every cycle; a creator's explicit signal does. retry_parked = reason == "control-socket" added = [n for n in live if n not in known and n != active and (retry_parked or n not in parked)] removed = [n for n in known if n not in live and n != active] changed = [n for n in live if n in known and n != active and n not in added and profile_serve_signature(live[n]) != sigs.get(n)] current = {n: h for n, h in live.items() if n in known or n == active} return await self._apply_profile_changes( current, added, removed, changed, reason=reason, live=live, result=result) async def _apply_profile_changes(self, current, added, removed, changed, *, reason, live=None, result=None): """Apply a selected diff under the reconcile lock, shared by the watcher and the serve/unserve control verbs. New profiles first get a runtime (reservation + store); one whose home another gateway owns or whose store is unusable is parked, not served.""" from gateway.run import MultiplexConfigError from gateway.run_runtime import park_profile, unpark_profile from hermes_cli.profiles import profiles_to_serve active = getattr(self, "_primary_profile_name", None) or "default" result = dict(result or {"added": [], "removed": [], "rescanned": [], "parked": [], "reason": reason}) result.setdefault("parked", []) live = live if live is not None else dict(current) known = dict(self._served_profile_homes or {}) sigs = self._served_profile_signatures or {} if not (added or removed or changed): return {**result, "served_profiles": self.served_profile_names()} for name in removed: await self._unserve_profile(name, known[name]) result["removed"].append(name) for name in list(added): reason_ = await self._serve_profile_runtime(name, live[name]) if reason_ is not None: added.remove(name) park_profile(self, name, reason_) result["parked"].append(name) continue unpark_profile(self, name) current[name] = live[name] claimed = self._live_resource_claims(active) transient_failed = set() config_parked = set() for name in added + changed: # Only acknowledge the configuration observed before connecting; # a setup save during an awaited handshake needs another scan. scan_signature = profile_serve_signature(current[name]) if name in changed: self._drop_stale_platform_verdicts(name) try: connected = await self._start_one_profile_adapters(name, current[name], claimed) except MultiplexConfigError as exc: # Runtime config refusal is a real park, not an adapter-less served profile: # withdraw the authority/tickets, release the reservation and publish the # terminal parked verdict. A control-socket rescan can explicitly retry it. logger.error("[MULTIPLEX] Profile '%s' not served: %s", name, exc) home = current.pop(name) await self._unserve_profile(name, home) park_profile(self, name, str(exc)) result["parked"].append(name) config_parked.add(name) continue except Exception: logger.error("[MULTIPLEX] Failed to start adapters for profile '%s'", name, exc_info=True) connected = 0 # A transient failure is not the deliberate park above: leave the signature # unacknowledged so the next reconcile retries the connect. transient_failed.add(name) else: sigs[name] = scan_signature if name in added: logger.info("[MULTIPLEX] Now serving profile '%s' (%s adapter(s) connected; %s)", name, connected, reason) result["added"].append(name) else: logger.info("[MULTIPLEX] Re-scanned profile '%s' after config/.env change (%s adapter(s) connected)", name, connected) result["rescanned"].append(name) if config_parked: added = [name for name in added if name not in config_parked] self._served_profile_signatures = sigs # A profile deleted while an adapter above was still connecting must not be recorded back # (the deleter's signal timed out against this lock and rmtree already ran). live_now = {str(name) for name, _home in profiles_to_serve(multiplex=True)} for name in [n for n in current if n not in live_now and n != active]: await self._unserve_profile(name, current.pop(name)) result["removed"].append(name) added = [n for n in added if n != name] self._record_served_profiles(active, list(current.items())) # ``_note_served_profiles`` fills a missing signature with the current one; that refill # would park a transiently-failed profile exactly like the config-error case above. for name in transient_failed: if isinstance(self._served_profile_signatures, dict): self._served_profile_signatures.pop(name, None) # A cached config with no live adapters is owed a home-channel notice nothing can # deliver, and the planned-restart marker then never clears. configs = getattr(self, "_profile_configs", None) if isinstance(configs, dict): configs.pop(name, None) if added: from gateway.run_runtime import recover_served_native_rows await recover_served_native_rows(self) await self._after_profiles_added([(n, current[n]) for n in added]) result["served_profiles"] = self.served_profile_names() return result def _drop_stale_platform_verdicts(self, name: str) -> None: """Before a changed profile's rescan, drop its ``:`` status entries for platforms with no live or queued adapter: the rescan re-publishes every verdict that still holds, so a fatal/disabled entry about a credential the user just removed can't outlive it.""" from gateway.run import _write_runtime_status_quiet live = set((getattr(self, "_profile_adapters", None) or {}).get(name) or {}) live |= set((getattr(self, "_profile_failed_platforms", None) or {}).get(name) or {}) noted = getattr(self, "_unserved_secondary_platforms", None) if isinstance(noted, set): noted -= {entry for entry in noted if entry[0] == name} _write_runtime_status_quiet(drop_profile_platforms=name, keep_platforms={f"{name}:{getattr(p, 'value', p)}" for p in live}) def _parked_profile_names(self) -> list: """Profiles that exist but could not be served (unusable store, home owned elsewhere); boot's ``initialize_gateway_runtime`` parks into the same published ``name -> reason`` map.""" from gateway.run_runtime import parked_profile_map return list(parked_profile_map(self)) async def _multiplex_for_added_profile(self) -> bool: """Start multiplexing in place when this gateway booted standalone ONLY because the host had one profile and a second now exists (#141623). Boot settles the mode once, so without this a profile created afterwards (Desktop's New Agent, ``hermes profile create``) was never served: every ticket, session and delete for it was refused until a manual restart. No restart and no drain: the launch env is frozen and unscoped secret reads fail closed (``activate_multi_profile_hosting``, the hosted-room precedent), the launch home becomes the first entry of the reservation, and the caller's ordinary hot-serve reconcile serves the new profile. Turns already running keep their bound scope; the launch profile's later work binds its own scope through ``_standalone_launch_scope`` and the re-wired primary handlers. Any other standalone verdict, or a preflight blocker now, leaves the gateway as it is; nothing is latched, so the next rescan or watcher tick retries once the blocker clears. The switch is an automatic migration, so ``gateway.auto_multiplex_migration: false`` holds it back the same way (re-read on every rescan: setting it back to true converges without a restart). True when armed.""" from hermes_cli.gateway_multiplex_mode import ( AUTO_MIGRATION_OPTED_OUT_REASON, SINGLE_PROFILE_REASON, MultiplexDecision, log_multiplex_decision, record_multiplex_decision, recorded_standalone_reason) from hermes_cli.profiles import profiles_to_serve from hermes_constants import get_default_hermes_root, get_process_hermes_home # Serialized with reconcile: a second rescan (watcher + creator) resuming mid-activation # reset the reservation the first one's reconcile had already grown. async with self._reconcile_lock(): if self._multiplex_on(): return True # a concurrent rescan activated while this one waited if not self._running or getattr(self, "_draining", False): return False # shutdown began: never arm fail-closed secrets under drain-time work if recorded_standalone_reason() not in (SINGLE_PROFILE_REASON, AUTO_MIGRATION_OPTED_OUT_REASON): return False launch = get_process_hermes_home().resolve() if launch != get_default_hermes_root().resolve(): return False # only the default home runs the multiplexer (``_launch_home_may_multiplex``) if len(profiles_to_serve(multiplex=True, include_parked=True)) < 2: return False blocker = await self._run_housekeeping_in_executor(_added_profile_blocker, launch) if blocker is not None: if blocker != self._single_profile_blocker_logged: logger.warning("[MULTIPLEX] A new profile exists but this gateway stays standalone: %s", blocker) self._single_profile_blocker_logged = blocker if blocker == AUTO_MIGRATION_OPTED_OUT_REASON: # Published so `gateway ensure` refuses the new profile naming the opt-out (Desktop shows # that detail) instead of "only one profile exists". record_multiplex_decision(MultiplexDecision(False, "guard", blocker)) return False if not self._running or getattr(self, "_draining", False): return False # shutdown began during the preflight await from hermes_constants import hermes_home_key from tools.mcp_tool_scope import rekey_unscoped_connections from tui_gateway.launch_profile_policy import activate_multi_profile_hosting activate_multi_profile_hosting() rekey_unscoped_connections(hermes_home_key(launch)) self.config.multiplex_profiles = True self.config._runtime_profile_homes = (("default", launch),) registry = getattr(self, "session_authorities", None) if registry is not None: registry.multiplexed = True for adapter in list(self.adapters.values()): adapter.set_message_handler(self._primary_message_handler()) adapter.set_busy_session_handler(self._primary_busy_session_handler()) adapter.set_platform_event_handler(self._primary_platform_event_handler()) self._record_served_profiles("default", [("default", launch)]) log_multiplex_decision(MultiplexDecision(True, "default", "a second profile was created")) logger.info("[MULTIPLEX] A second profile exists; serving every profile from this process (no restart)") return True async def _serve_profile_runtime(self, name: str, home: Path) -> Optional[str]: """Grow the reservation by *home* and build its session authority (boot's per-secondary steps). Returns the park reason — reservation released — when another gateway owns the home (a stray per-profile daemon) or its store cannot be opened; None when served. An adapters-only runner (no authority registry) grows the reservation alone.""" from gateway.run_runtime import release_profile_home, reserve_profile_home_in_worker, serve_profile_runtime from gateway.runtime_ownership import OwnershipConflict try: await reserve_profile_home_in_worker(self, name, home) except (OwnershipConflict, OSError) as exc: logger.error("[MULTIPLEX] Profile '%s' not served: %s", name, exc) return f"home owned by another gateway: {exc}" if getattr(self, "session_authorities", None) is None: return None try: await serve_profile_runtime(self, name, home) except (Exception, ConfigBackendUnavailable) as exc: # unservable now: parked, retried by rescan if self.session_authorities.for_home(home) is not None: # Retirement of the failed attempt missed its deadline: its writer may still use # this home, so the reservation and authority stay; the profile is parked (no # adapters) and the batch goes on for every other profile. logger.error("[MULTIPLEX] Profile '%s' not served (%s); a turn outlived its Stop, " "ownership retained", name, exc, exc_info=True) return f"stopping after a failed serve; ownership retained: {exc}" logger.error("[MULTIPLEX] Profile '%s' not served: its session store is unusable (%s): %s", name, home, exc, exc_info=True) release_profile_home(self, home) return f"session store unusable: {exc}" return None def _live_resource_claims(self, active: str) -> dict[tuple, str]: """Startup's ``claimed`` map rebuilt from what is live now: primary claims plus every connected secondary's credential/listener, so a hot-added profile reusing a token is parked, never a second poller.""" claimed = self._primary_resource_claims(active) for profile_name, adapters in (getattr(self, "_profile_adapters", None) or {}).items(): for platform, adapter in list(adapters.items()): for claim in self._adapter_claims(platform, adapter): if claim is not None: claimed[claim] = profile_name return claimed async def _after_profiles_added(self, profile_homes) -> None: """Per-profile startup side effects for hot-added profiles: log routing + scoped MCP discovery.""" from gateway.run import _enable_multiplex_log_routing, _profile_runtime_scope from contextvars import copy_context with _log_suppressed(logging.DEBUG, "log routing refresh failed", exc_info=True): _enable_multiplex_log_routing(self.config) from tools.mcp_oauth import suppress_interactive_oauth loop = asyncio.get_running_loop() for profile_name, profile_home in profile_homes: try: from tools.mcp_tool_discovery import discover_mcp_tools with _profile_runtime_scope(Path(profile_home)), suppress_interactive_oauth(): await loop.run_in_executor(None, copy_context().run, discover_mcp_tools) except Exception: logger.warning("MCP tool discovery failed for profile '%s'", profile_name, exc_info=True) async def _unserve_profile(self, name: str, home: Path) -> None: """Stop and unroute one profile: cancel its reconnects, tear down its adapters, drop its bookkeeping and release this process's handles into its home so the deleter's rmtree succeeds. The releasing process is not always the deleter (#130244): under multiplexing THIS gateway process routes the profile's logs and owns its scoped MCP servers, so both must be released here too or the deleter's rmtree still fails on ``logs/.__agent.lock`` / ``mcp-stderr.log``. The whole teardown runs inside the removed profile's own runtime scope: adapter disconnect hooks, the agent-cache eviction (provider/memory shutdown) and the state/memory handle release all read config and credentials at call time, and this coroutine runs on the reconcile task with no profile bound — unscoped they resolved against the LAUNCH home, so a teardown hook needing this profile's credential failed closed (or, worse, borrowed the launch profile's). Secrets are not re-hydrated: teardown must not block the loop on a source fetch. """ from gateway.run import _profile_runtime_scope, _write_runtime_status_quiet from gateway.run_runtime import release_profile_home, unserve_profile_runtime # Retire before teardown; a False verdict (writer outlived its Stop) is honoured below. retired = getattr(self, 'session_authorities', None) is None or await unserve_profile_runtime(self, home) pending = (getattr(self, "_profile_failed_platforms", None) or {}).pop(name, None) or {} tasks = [t for t in pending.values() if isinstance(t, asyncio.Task) and not t.done()] # Its adapters stay in the map (live owners of their bots) while the reconnects cancel; the retiring # mark keeps anything finishing meanwhile (a fatal teardown, a startup handoff) from queuing another. retiring = getattr(self, "_retiring_profiles", None) if retiring is None: retiring = self._retiring_profiles = set() retiring.add(name) try: for task in tasks: task.cancel() if tasks: await asyncio.wait(tasks, timeout=self._adapter_disconnect_timeout_secs()) finally: retiring.discard(name) with _profile_runtime_scope(Path(home), hydrate_secrets=False): adapters = (getattr(self, "_profile_adapters", None) or {}).pop(name, None) or {} for platform, adapter in adapters.items(): # no await since the pop: each holds its bot until torn down self._hold_secondary_claims(name, platform, adapter, self._adapter_claims(platform, adapter)) try: for platform, adapter in list(adapters.items()): await self._bounded_adapter_teardown(adapter, platform, profile=name) self._release_secondary_claims(name, platform, adapter) # its bot is free once it is down finally: # Every hold of the profile goes, a fatal teardown's still disconnecting included (the profile # is gone, so its handler queues nothing): none may block a re-serve. held = getattr(self, "_secondary_held_claims", None) or {} for key in [key for key in held if key[0] == name]: del held[key] # Its ``:`` runtime entries describe a profile that no longer exists. _write_runtime_status_quiet(drop_profile_platforms=name) for attr in ("pairing_stores", "_busy_text_modes_by_profile", "_busy_input_modes_by_profile", "_busy_text_timing_by_profile", "_human_delay_by_profile", "_profile_configs", "_profile_bot_platforms"): store = getattr(self, attr, None) if isinstance(store, dict): store.pop(name, None) if isinstance(self._served_profile_homes, dict): self._served_profile_homes.pop(name, None) if isinstance(self._served_profile_signatures, dict): self._served_profile_signatures.pop(name, None) from gateway.session import _session_key_namespace prefix = _session_key_namespace(name) + ":" cache = getattr(self, "_agent_cache", None) for key in [k for k in list(cache or {}) if str(k).startswith(prefix)]: with _log_suppressed(logging.DEBUG, "agent eviction failed for %s", key, exc_info=True): self._evict_cached_agent(key) if not retired: # A writer outlived its Stop and may still write this home: the profile leaves the # served set (ingress is gone above) but keeps its reservation and store handles until # this process exits. Raising here aborted the reconcile for every other profile. logger.error("[MULTIPLEX] Profile '%s' unrouted, but a turn outlived its Stop; ownership retained", name) return # The reservation (gateway.lock) is released LAST: a deleter waiting on it may rmtree the # home as soon as it goes, so every handle this process holds there must close first. try: with _log_suppressed(logging.DEBUG, "profile handle release failed", exc_info=True): from hermes_state_registry import close_all_under close_all_under(home) with _log_suppressed(logging.DEBUG, "memory-store release failed", exc_info=True): from plugins.memory import import_provider_module # catalog plugin since it left core import_provider_module("holographic", "store").MemoryStore.release_all_under(home) # The gateway process is a second holder the CLI-side delete cannot reach: its routed # log files (logs/.__agent.lock, logs/.__errors.lock) and its scoped MCP servers' # mcp-stderr.log handle live HERE, so unserve must release them in this process too # (#130244) — the same two calls ``delete_profile`` makes for the deleting process. # Both calls BLOCK (the log release stops/joins the QueueListener, which can be inside # a ConcurrentRotatingFileHandler emit waiting on another process's rotation lock; the # MCP shutdown waits on future.result for the MCP loop), so they run off the loop in a # worker thread — otherwise every other served profile's events stall behind them. # ``asyncio.to_thread`` copies the context, so the surrounding # ``_profile_runtime_scope`` home/secret override stays visible in the worker. with _log_suppressed(logging.DEBUG, "scoped MCP shutdown failed", exc_info=True): from hermes_constants import hermes_home_key from tools.mcp_tool_lifecycle import shutdown_mcp_servers await asyncio.to_thread(shutdown_mcp_servers, scope=hermes_home_key(home)) with _log_suppressed(logging.DEBUG, "profile log handler release failed", exc_info=True): from hermes_logging import release_profile_log_handlers await asyncio.to_thread(release_profile_log_handlers, home) finally: release_profile_home(self, home) logger.info("[MULTIPLEX] Profile '%s' unserved — %d adapter(s) stopped and unrouted", name, len(adapters)) def _profile_lifecycle_verb(runner, *, serve: bool): """Build on the owning loop; socket handlers themselves run on executor threads.""" loop = asyncio.get_running_loop() async def apply(name): from hermes_cli.profiles import profiles_to_serve, profile_is_parked if not runner._multiplex_on() or not runner._running or runner._served_profile_homes is None: return {"error": "host multiplexer is not ready"} async with runner._reconcile_lock(): active = getattr(runner, "_primary_profile_name", None) or "default" if not isinstance(name, str) or not name or name == active: return {"error": "a non-launch profile name is required"} known = dict(runner._served_profile_homes) if not serve: if name not in known: return {"error": f"profile '{name}' is not served"} await runner._unserve_profile(name, known.pop(name)) runner._record_served_profiles(active, list(known.items())) return {"unserved": name, "served_profiles": runner.served_profile_names()} installed = dict(profiles_to_serve(True, include_parked=True)) if name not in installed: return {"error": f"unknown profile '{name}'"} if name != "default" and profile_is_parked(installed[name]): return {"error": f"profile '{name}' is parked (gateway.parked)"} if name in known: return {"error": f"profile '{name}' is already served"} known[name] = installed[name] result = await runner._apply_profile_changes(known, [name], [], [], reason="control-socket") if name not in result["served_profiles"]: return {"error": f"profile '{name}' was removed or parked during startup"} return {"served": name, "served_profiles": result["served_profiles"]} def handler(params): future = asyncio.run_coroutine_threadsafe(apply(params.get("name")), loop) try: return future.result(timeout=5.0) except TimeoutError: # Keep running: the client must not interpret an incomplete teardown as stopped. return {"pending": True, "served_profiles": runner.served_profile_names()} except Exception as exc: logger.warning("Profile lifecycle request failed", exc_info=True) return {"error": f"{type(exc).__name__}: {exc}"} return handler def unserve_profile_verb(runner): return _profile_lifecycle_verb(runner, serve=False) def serve_profile_verb(runner): return _profile_lifecycle_verb(runner, serve=True) def _mcp_config_reconciler(runner=None): """Housekeeping chore keeping live MCP servers in step with ``mcp_servers`` on disk, every tick after the first (startup discovery owns that one). Reconciling on DRIFT rather than only on a config EDIT is what brings back a server whose FIRST connect failed (#112445): it never reached ``_servers``, so the parked self-probe — a property of a task that connected once — cannot revive it, and its config never changes. The reconcile is a cached config read plus set compares when nothing moved; a server dropped from config is torn down (a parked one otherwise self-probes every ``_PARKED_RETRY_INTERVAL`` for the life of the process) and a missing one is reconnected only once its per-server connect cooldown (30s→600s backoff) has lapsed, so a chronically failing server is retried on that schedule, not every tick. Interactive OAuth is suppressed — this runs on a housekeeping thread nobody is watching.""" primed: set = set() def _reconcile_current(label: str) -> None: from tools.mcp_oauth import suppress_interactive_oauth from tools.mcp_tool_discovery import reconcile_mcp_servers_with_config if label not in primed: primed.add(label) return # first tick: startup discovery already reflects this config (or is still running) with suppress_interactive_oauth(): result = reconcile_mcp_servers_with_config() if result["removed"] or result["added"]: logger.info("MCP servers reconciled with config (%s): removed=%s added=%s", label, result["removed"], result["added"]) return lambda: _for_each_served_profile(runner, _reconcile_current) def _for_each_served_profile(runner, body) -> None: """Run ``body(profile_label)`` once per served profile, inside that profile's runtime scope. Housekeeping runs on a bare thread with no turn on the stack, so nothing binds a profile for it: ``get_hermes_home()`` and ``get_secret()`` see the LAUNCH profile's values, and under ``gateway.multiplex_profiles`` a fail-closed credential read logs ``no profile secret scope on a multiplexed call`` on every tick (the skills-sync pulls resolved Nous credentials this way, four WARNINGs per hourly tick per chore). A single-profile gateway runs ``body`` once, unscoped: there the process env IS the profile's own value — unless a hosted room already flipped the process-wide guard (#112878), in which case the launch profile's OWN scope is bound, as ``run_turn.py::_standalone_launch_scope`` does for turns.""" from gateway.run import _multiplex_profile_homes, _profile_runtime_scope config = getattr(runner, "config", None) if not getattr(config, "multiplex_profiles", False): from gateway.run_turn import GatewayTurnMixin with GatewayTurnMixin._standalone_launch_scope(): body("default") return for profile_name, profile_home in _multiplex_profile_homes(config): # One boundary per profile: callers (``_housekeeping_chore``) catch only at the tick level, # so one profile's unreadable store or broken .env abandoned every profile after it, on # every tick. The launch store failing is reachable (``_init_session_db`` tolerates it and # keeps running), and serve defers every served profile's sweep to this loop. try: with _profile_runtime_scope(Path(profile_home)): body(str(profile_name)) except Exception as exc: logger.debug("Housekeeping for profile %s skipped: %s", profile_name, exc) def profile_scoped_chore(runner, chore): """Wrap a zero-arg housekeeping chore that reads the profile's home, config or credentials so it runs once per served profile under that profile's scope (see ``_for_each_served_profile``).""" return lambda: _for_each_served_profile(runner, lambda _label: chore()) def migrate_profile_identity_verb(runner): """Build the ``migrate-profile-identity`` control-verb handler for ``hermes profile rename`` (#111926). The live multiplexer owns the routing index in memory and writes it back periodically, so a CLI-side rewrite of ``agent::*`` would be clobbered on the next save; the CLI therefore asks this process to rekey both durable stores AND ``SessionStore._entries``. Runs on the control-socket executor thread; ``rekey_profile_routing`` takes the store lock.""" def _handler(params: dict) -> dict: old, new = str(params.get("old") or "").strip(), str(params.get("new") or "").strip() if not old or not new or old == new: return {"ok": False, "error": "old/new required and must differ"} store = getattr(runner, "session_store", None) if store is None: return {"ok": False, "error": "live gateway has no session store"} acquired = [] try: from hermes_state_registry import acquire, release_or_close db_counts: dict[str, dict[str, int]] = {} routing_db = getattr(store, "_routing_db", None) if routing_db is not None and hasattr(routing_db, "rekey_profile_state"): db_counts["routing"] = routing_db.rekey_profile_state(old, new) routing_home = getattr(store, "_routing_home", None) profile_path = Path(routing_home) / "profiles" / new / "state.db" if routing_home else None if profile_path is not None and profile_path.exists(): profile_db = acquire(profile_path) acquired.append(profile_db) db_counts["profile"] = profile_db.rekey_profile_state(old, new) rekeyed = store.rekey_profile_routing(old, new) return {"ok": True, "rekeyed": rekeyed, "db": db_counts} except Exception as exc: logger.warning("Profile identity migration failed for %r->%r: %s", old, new, exc) return {"ok": False, "error": f"{type(exc).__name__}: {exc}"} finally: for db in acquired: try: release_or_close(db) except Exception: logger.debug("Failed to release renamed profile state DB", exc_info=True) return _handler def purge_profile_identity_verb(runner): """Build the ``purge-profile-identity`` control-verb handler for ``hermes profile delete`` (#111926, delete side). The live multiplexer owns the routing index in memory and writes it back periodically, so a CLI-side DELETE of ``agent::*`` rows would be undone by its next save; the CLI therefore asks this process to drop the durable rows AND ``SessionStore._entries``. Deliberately NOT part of ``_unserve_profile()``: that path also unserves names that are still alive elsewhere in the identity story — a rename's old name leaves the served set exactly like a delete does (its directory is gone either way) — and purging there would race the rekey it is supposed to leave intact. Only the delete path invokes this verb. Runs on the control-socket executor thread; ``purge_profile_routing`` takes the store lock.""" def _handler(params: dict) -> dict: name = str(params.get("name") or "").strip() if not name: return {"ok": False, "error": "name required"} store = getattr(runner, "session_store", None) if store is None: return {"ok": False, "error": "live gateway has no session store"} try: db_counts: dict[str, dict[str, int]] = {} routing_db = getattr(store, "_routing_db", None) if routing_db is not None and hasattr(routing_db, "purge_profile_state"): db_counts["routing"] = routing_db.purge_profile_state(name) dropped = store.purge_profile_routing(name) return {"ok": True, "dropped": dropped, "db": db_counts} except Exception as exc: logger.warning("Profile identity purge failed for %r: %s", name, exc) return {"ok": False, "error": f"{type(exc).__name__}: {exc}"} return _handler