"""Process startup phases; diagnostics remain in the gateway facade.""" from __future__ import annotations import logging import threading import time from pathlib import Path from typing import Any, Optional from gateway.config import GatewayConfig def _start_deferred_state_db_integrity_check(evidence: dict) -> None: """Run the store scan an unclean-exit report deferred, once the runner is live. On the event loop it held every local listener open but unanswered for the whole ``quick_check`` (minutes on a multi-GB ``state.db``), so local clients hung instead of being refused. A daemon thread nothing waits on; it only appends the forensic exit-diag record. """ from agent.memory_provider import spawn_context_thread from gateway.lifecycle_ledger import STATE_DB_INTEGRITY_DEADLINE_S, run_deferred_integrity_check # Under the caller's profile scope: the scan resolves this gateway's state.db through it. spawn_context_thread( run_deferred_integrity_check, args=(evidence,), kwargs={"deadline_s": STATE_DB_INTEGRITY_DEADLINE_S}, name="state-db-integrity-check").start() # Log-record parity with the origin module. logger = logging.getLogger("gateway.run") async def _start_gateway_replace_existing_instance(existing_pid: int, replace: bool) -> bool: """Handle a live gateway PID under this HERMES_HOME: replace it (``--replace``) or refuse. Returns False when startup must abort (refused, permission denied, target still alive).""" from gateway.run import (_clear_takeover_marker_quiet, _replace_target_belongs_to_other_profile, _wait_for_pid_exit, get_hermes_home, suppress) from gateway.status import get_process_start_time, remove_pid_file, terminate_pid if not replace: hermes_home = str(get_hermes_home()) logger.error( "Another gateway instance is already running (PID %d, HERMES_HOME=%s) and did not " "publish a host record this process could attach to.", existing_pid, hermes_home) print( f"\n❌ A gateway already owns this host (PID {existing_pid}).\n" f" One gateway per host serves every profile, so there is nothing to start here.\n" f" Attach is impossible: PID {existing_pid} published no usable host record\n" f" (an older build, or an unwritable lock directory).\n" f" Take the host over: hermes gateway run --replace\n" f" Or stop it first: hermes gateway stop\n") return False # Never signal a process not provably ours (a poisoned PID record → cross-profile restart loop). if _replace_target_belongs_to_other_profile(existing_pid): from gateway.status import _get_process_hermes_home, write_runtime_status logger.error( "Refusing --replace: PID %d cannot be proven to belong " "to this profile's gateway (HERMES_HOME %s). Remove the " "stale PID record or stop the owning profile explicitly.", existing_pid, _get_process_hermes_home()) # The stamp must be ON DISK before this process exits: every CLI exit path funnels # through _exit_after_graceful_shutdown(), whose os._exit() never flushes the queued # runtime-status writer. A fire-and-forget stamp can die in the queue, the supervisor # retries into the same stale record, and the loop this refusal exists to break keeps # going (#135392). Bounded wait: a stalled disk delays the exit by at most 5 s, it # never hangs the start. try: write_runtime_status( gateway_state="startup_failed", exit_reason="gateway_replace_ownership_unproven", wait_timeout=5.0, ) except Exception: # Persistence must never abort the start; the traceback (DEBUG) keeps the # failure visible for the next bug report. logger.debug("Could not stamp runtime status after the ownership refusal", exc_info=True) return False existing_start_time = get_process_start_time(existing_pid) logger.info("Replacing existing gateway instance (PID %d) with --replace.", existing_pid) # Takeover marker: target exits 0 on our SIGTERM (exit 1 → systemd Restart=on-failure flap loop). try: from gateway.status import write_takeover_marker write_takeover_marker(existing_pid) except Exception as e: logger.debug("Could not write takeover marker: %s", e) # Snapshot children BEFORE signalling: reparented orphans are invisible yet hold scoped token locks. try: from gateway.status import _snapshot_gateway_children _old_gateway_children = _snapshot_gateway_children(existing_pid) except Exception: _old_gateway_children = [] try: terminate_pid(existing_pid, force=False) except ProcessLookupError: pass # Already gone except (PermissionError, OSError): logger.error("Permission denied killing PID %d. Cannot replace.", existing_pid) _clear_takeover_marker_quiet() return False # Up to 10s for SIGTERM, then SIGKILL. if not await _wait_for_pid_exit(existing_pid, 20, 0.5): logger.warning("Old gateway (PID %d) did not exit after SIGTERM, sending SIGKILL.", existing_pid) old_gateway_exited = False try: terminate_pid(existing_pid, force=True, expected_start_time=existing_start_time) except ProcessLookupError: old_gateway_exited = True except (PermissionError, OSError): pass # Confirm SIGKILL took (D-state/zombie) before clearing PID/locks, or two gateways share a token. if not old_gateway_exited and not await _wait_for_pid_exit(existing_pid, 20, 0.25): logger.error( "Old gateway (PID %d) still appears alive after SIGKILL; " "aborting replacement to avoid a duplicate gateway.", existing_pid) _clear_takeover_marker_quiet() return False # Reap orphaned children (POSIX; mirrors Windows taskkill /T) so they stop holding scoped token locks. try: from gateway.status import reap_gateway_children reap_gateway_children(_old_gateway_children, parent_pid=existing_pid) except Exception: logger.debug("Child reap for replaced gateway PID %d failed", existing_pid, exc_info=True) remove_pid_file() # remove_pid_file() is a no-op when the PID doesn't match; force-unlink covers a crashed old process. with suppress(Exception): (get_hermes_home() / "gateway.pid").unlink(missing_ok=True) # The old process may not have consumed the marker (SIGKILL'd before its handler read it). _clear_takeover_marker_quiet() # Stopped (Ctrl+Z) processes don't release scoped locks on exit; stale lock files block the new gateway. try: from gateway.status_scoped_locks import release_all_scoped_locks _released = release_all_scoped_locks(owner_pid=existing_pid, owner_start_time=existing_start_time) if _released: logger.info("Released %d stale scoped lock(s) from old gateway.", _released) except Exception: logger.debug("Stale scoped lock release for replaced gateway PID %d failed", existing_pid, exc_info=True) return True def _start_gateway_configure_logging(verbosity: Optional[int]) -> None: """Sync bundled skills, set up file logging + startup security audit, and the -v/-q stderr handler.""" from gateway.run import (_best_effort, _gateway_stderr_formatter, _hermes_home, logging) def _sync_skills() -> None: from tools.skills_sync import sync_skills sync_skills(quiet=True) _best_effort(_sync_skills) # Centralized logging (agent.log INFO+, errors.log WARNING+, gateway.log gateway-only); idempotent. from hermes_logging import setup_logging, _safe_stderr setup_logging(hermes_home=_hermes_home, mode="gateway") def _security_audit() -> None: # Warn-on-load, never blocks: surfaces root / weak-SSH / unauthenticated-listener exposure. from hermes_cli.security_audit_startup import log_startup_security_warnings def _raw_cfg(): from hermes_cli.config import read_raw_config return read_raw_config() log_startup_security_warnings(hermes_home=_hermes_home, config=_best_effort(_raw_cfg)) _best_effort(_security_audit, "Startup security audit failed (non-fatal): %s") # Optional stderr handler from -v/-q: None (quiet) = none; 0 = WARNING; 1 = INFO; 2+ = DEBUG. if verbosity is not None: _stderr_level = {0: logging.WARNING, 1: logging.INFO}.get(verbosity, logging.DEBUG) _stderr_handler = logging.StreamHandler(_safe_stderr()) _stderr_handler.setLevel(_stderr_level) _stderr_handler.setFormatter(_gateway_stderr_formatter()) root = logging.getLogger() root.addHandler(_stderr_handler) if _stderr_level < root.level: # so DEBUG records can reach the handler root.setLevel(_stderr_level) def _start_gateway_make_shutdown_signal_handler(runner, _signal_initiated_shutdown: list): """Build the SIGINT/SIGTERM handler; ``_signal_initiated_shutdown[0]`` records an unplanned signal.""" from gateway.run import (_best_effort, _hermes_home, asyncio, logger, signal) planned_stop_seen = [False] def shutdown_signal_handler(received_signal=None): # Planned --replace takeover (sibling marked this PID): exit 0 so systemd won't revive us. def _takeover() -> bool: from gateway.status import consume_takeover_marker_for_self return consume_takeover_marker_for_self() # Planned stop: CLI marks first, else its SIGTERM looks like an external kill. SIGINT = Ctrl+C. def _planned_stop() -> bool: from gateway.status import consume_planned_stop_marker_for_self return consume_planned_stop_marker_for_self() # Fast (<10ms) sync snapshot: stdlib + /proc, no subprocesses (`ps aux` here once blocked ~3s). def _snapshot(): from gateway.shutdown_forensics import snapshot_shutdown_context return snapshot_shutdown_context(received_signal) planned_takeover = bool(_best_effort(_takeover, "Takeover marker check failed: %s")) planned_stop = received_signal == signal.SIGINT or ( not planned_takeover and bool(_best_effort(_planned_stop, "Planned stop marker check failed: %s"))) # `hermes gateway stop` writes the marker, THEN signals: the planned-stop watcher can consume # the marker in between, and the CLI's own SIGTERM must not then read as an external kill. if planned_stop: planned_stop_seen[0] = True elif planned_stop_seen[0] and not planned_takeover: planned_stop = True _shutdown_ctx = _best_effort(_snapshot, "snapshot_shutdown_context failed: %s") sig_name = _shutdown_ctx["signal"] if _shutdown_ctx else None if planned_takeover: logger.info("Received %s as a planned --replace takeover — exiting cleanly", sig_name or "SIGTERM") elif planned_stop: logger.info("Received %s as a planned gateway stop — exiting cleanly", sig_name or "SIGTERM/SIGINT") else: # Mirrored onto the runner so _stop_impl suppresses the gateway_state=stopped persist for # unexpected signals; operator stops take the `planned_stop` branch and leave it False (DO persist). _signal_initiated_shutdown[0] = runner._signal_initiated_shutdown = True logger.info("Received %s — initiating shutdown", sig_name or "SIGTERM/SIGINT") if _shutdown_ctx is not None: def _log_context() -> None: # The most useful line for "gateway keeps dying" tickets. from gateway.shutdown_forensics import format_context_for_log logger.warning("Shutdown context: %s", format_context_for_log(_shutdown_ctx)) def _diagnostic() -> None: # Heavyweight (ps auxf, pstree, dmesg), detached so it finishes even if our cgroup is torn # down; bounded by an internal timeout, never blocks. from gateway.shutdown_forensics import spawn_async_diagnostic spawn_async_diagnostic( _hermes_home / "logs" / "gateway-shutdown-diag.log", _shutdown_ctx["signal"], timeout_seconds=5.0) _best_effort(_log_context, "format_context_for_log failed: %s") _best_effort(_diagnostic, "spawn_async_diagnostic failed: %s") if not planned_takeover: # Supervisor/operator SIGNAL stop (bootout, kickstart -k, systemd, s6, bare kill) — the # only kind launchd times with ExitTimeOut. In-band SIGUSR1 restarts never pass through # here, and a sibling-driven --replace takeover is not launchd-timed either, so both # keep the configured drain. _stop_impl uses this to cap the drain to the live budget. runner._stop_requested_by_signal = True asyncio.create_task(runner.stop()) return shutdown_signal_handler def _start_gateway_claim_pid_file(force: bool = False) -> bool: """Claim the runtime lock + PID file (O_EXCL winner is the authoritative gateway). False = lost.""" from gateway.run import (_claim_host_gateway_role, logger, os) import atexit from gateway.status import ( acquire_gateway_runtime_lock, get_running_pid, release_gateway_runtime_lock, remove_pid_file, write_pid_file) _current_pid = get_running_pid() if _current_pid is not None and _current_pid != os.getpid(): logger.error("Another gateway instance (PID %d) started during our startup. " "Exiting to avoid double-running.", _current_pid) return False if not acquire_gateway_runtime_lock(): logger.error("Gateway runtime lock is already held by another instance. Exiting.") return False try: write_pid_file() except FileExistsError: release_gateway_runtime_lock() logger.error("PID file race lost to another gateway instance. Exiting.") return False atexit.register(remove_pid_file) atexit.register(release_gateway_runtime_lock) from hermes_cli.gateway_multiplex_mode import publish_pending_multiplex_decision publish_pending_multiplex_decision() _claim_host_gateway_role(force=force) return True async def _start_gateway_start_control_socket(runner): """Start the gateway control socket (identify/status/pause-for-update); None when unavailable.""" from gateway.run import (asyncio, logger, os, threading) import atexit import concurrent.futures _control_server = None try: # Started immediately after the PID-file claim: winning that O_EXCL race is the moment this process # becomes the authoritative gateway for its HERMES_HOME, so from here on "does a socket answer?" is # a truthful liveness/identity query for updater and fleet consumers. Returns None on failure and # start_gateway aborts startup: the listener mints every local client's attach ticket, so a # gateway without it can serve no local session. See #92091. from gateway.control_socket import GatewayControlServer, build_identify_payload from gateway.run_profile_reconcile import ( migrate_profile_identity_verb, purge_profile_identity_verb, unserve_profile_verb, serve_profile_verb, ) from gateway.run_plugin_rewire import reload_plugins_verb from gateway.run_active_work_verb import active_work_verb from gateway.relay.link import relay_link_verb from gateway.drain_control import RESTART_WAIT_REASON descriptor = runner.session_runtime_descriptor def _identify_runtime(): payload = build_identify_payload() payload.update({key: descriptor[key] for key in ( "instance_id", "runtime_protocol", "authority_epoch", "state", "api_origin", "served_profiles", "parked_profiles", "capabilities") if key in descriptor}) if getattr(runner, '_draining', False): payload.update(state='draining') # The restart's after-turn wait still owns its turns: clients reach it to answer a # parked approval/clarify (#139041), so the endpoint stays dialable until stop(). if descriptor.get('drain_reason') != RESTART_WAIT_REASON: payload['capabilities'] = [] if payload.get('state') == 'draining' and descriptor.get('drain_reason'): payload['drain_reason'] = descriptor['drain_reason'] payload["supervisor"] = {"manual": "none", "desktop": "none"}.get( payload.get("supervisor"), payload.get("supervisor", "none")) return payload # pause-for-update: the updater asks us to drain + exit (freeing venv handles) vs. a tree-kill # (same path as SIGUSR1). Handler runs on the socket executor thread, so marshal onto the loop. # pause-for-update (#92091 step 2): the updater asks this gateway to drain in-flight turns and exit # cleanly — releasing every venv file handle — instead of being tree-killed mid-turn. Same drain # path as SIGUSR1/service restarts (request_restart(via_service=True)); the updater (or the service # manager) relaunches after the code swap. _main_loop = asyncio.get_running_loop() def _pause_for_update_handler() -> dict: try: from hermes_cli.gateway import _get_restart_drain_timeout _drain = float(_get_restart_drain_timeout()) except Exception: _drain = 30.0 accepted_box: list[bool] = [] _done = threading.Event() def _request() -> None: try: accepted_box.append(runner.request_restart(detached=False, via_service=True)) finally: _done.set() _main_loop.call_soon_threadsafe(_request) _done.wait(timeout=5.0) accepted = bool(accepted_box and accepted_box[0]) return { "pausing": accepted, "already_stopping": not accepted, "pid": os.getpid(), "drain_timeout": _drain} def _rescan_profiles_handler() -> dict: """``hermes profile create/delete`` asks the multiplexer to reconcile ``profiles/`` now (the watcher also rescans periodically). Runs on the socket executor: marshal onto the loop and wait briefly so the caller learns whether the profile is served. A gateway standalone only because the host had one profile starts multiplexing in place on this rescan.""" future = asyncio.run_coroutine_threadsafe( runner.reconcile_served_profiles(reason="control-socket"), _main_loop) try: # Bounded: a token-less create reconciles in milliseconds; a credential-add whose adapter # connect outlasts this keeps running and the caller sees ``pending`` (not an error). result = future.result(timeout=5.0) except concurrent.futures.TimeoutError: result = {"pending": True, "served_profiles": runner.served_profile_names()} return {"multiplex": bool(getattr(runner.config, "multiplex_profiles", False)), **result} _control_server = GatewayControlServer( verb_handlers={"pause-for-update": _pause_for_update_handler, "identify": _identify_runtime, "rescan-profiles": _rescan_profiles_handler, "migrate-profile-identity": migrate_profile_identity_verb(runner), "unserve-profile": unserve_profile_verb(runner), "serve-profile": serve_profile_verb(runner), "purge-profile-identity": purge_profile_identity_verb(runner), # `hermes gateway stop/restart` refuse while this names live work. "active-work": active_work_verb(runner, _main_loop), # A plugin installed/enabled by another process loads now and re-wires the # live adapters' handlers (#87770); tools/prompt still wait for the next session. "reload-plugins": reload_plugins_verb(runner, _main_loop), # The per-gateway relay secret exists only in this process: the dashboard # mints Team Gateway link codes through it. "relay-link": relay_link_verb}) _control_server.ticket_store = runner.session_ticket_store # The Webapp's verbs answer from the socket's first moment ("starting" until the listener # settles): a client polling a restarting gateway must never read "unknown verb" as a failure. from gateway.webapp_listener import register_control_verbs from hermes_constants import get_hermes_home register_control_verbs(runner, get_hermes_home(), server=_control_server) if not await _control_server.start(): _control_server = None else: atexit.register(_control_server.cleanup_files) # Hosted-room transports install onto this listener; bind it here so every launcher # that starts the control socket (not only start_gateway) exposes it. runner.session_control_server = _control_server except Exception as _cs_exc: logger.error("Control socket startup failed; the gateway cannot start without it: %s", _cs_exc) _control_server = None return _control_server def _start_gateway_start_cron_and_housekeeping(runner): """Start the cron scheduler thread + gateway housekeeping thread; returns ``(cron_stop, cron_provider, cron_thread, housekeeping_thread)``.""" from gateway.run import (Dict, Platform, _cron_profile_gate, _cron_tick_profile_homes, _start_gateway_housekeeping, asyncio, logger, threading) # The event loop is passed so cron delivery can use live adapters (E2EE support). from cron.scheduler_provider import ( InProcessCronScheduler, resolve_cron_scheduler, scheduler_for_profile_mode) cron_stop = threading.Event() # ONE gateway process per host multiplexes every profile, so its cron ticker owns EVERY # profile's store — `gateway.multiplex_profiles` gates adapters, not cron. Gating the tick set # on that flag left every non-launch profile's jobs in a store no ticker visited: they # silently never fired. try: cron_profile_homes = _cron_tick_profile_homes(runner.config) except Exception as exc: logger.warning("Could not resolve profile homes for cron: %s", exc) cron_profile_homes = [] # External providers own one unscoped remote registry, so they can only serve a single home. cron_provider = scheduler_for_profile_mode( resolve_cron_scheduler(), multiplex_profiles=len(cron_profile_homes) > 1) cron_start_kwargs: Dict[str, Any] = {"adapters": runner.adapters, "loop": asyncio.get_running_loop()} if isinstance(cron_provider, InProcessCronScheduler) and cron_profile_homes: # Live enumerator: the ticker re-reads profiles/ every cycle so a profile created while # the gateway runs gets its jobs fired without a restart (hot-serve). cron_start_kwargs["profile_homes"] = lambda: _cron_tick_profile_homes(runner.config) # Stand down, per tick, for a profile whose OWN gateway process ticks it. cron_start_kwargs["profile_gate"] = _cron_profile_gate # Per-profile adapters so each profile's cron output goes via its own bot, not the # default's. Absent (no multiplexed adapters), delivery for a secondary profile falls # back to the primary's routed adapters or fails closed — the job still FIRES. cron_start_kwargs["profile_adapters"] = getattr(runner, "_profile_adapters", None) # A secondary bot that dropped is still that profile's own: never borrow meanwhile. cron_start_kwargs["profile_bot_platforms"] = getattr(runner, "_profile_bot_platforms", None) # runner.adapters belongs to the LAUNCH profile (``default``, or the ``--profile`` # name); naming it keeps the ticker from routing a secondary's cron through that bot # and lets a named multiplexer's own jobs reuse its live adapters. cron_start_kwargs["default_profile"] = runner._primary_profile_name logger.info( "Cron scheduler will tick %d profile(s): %s", len(cron_profile_homes), [p[0] if isinstance(p, tuple) else p for p in cron_profile_homes]) # Only the in-process ticker polls local due jobs, so only it gets the external-drain dispatch gate. if isinstance(cron_provider, InProcessCronScheduler): cron_start_kwargs["can_dispatch"] = lambda: not ( runner._draining or runner._external_drain_active) # Supervised: a ticker that dies without a stop request is respawned by housekeeping (#111010). from cron.scheduler_thread import SupervisedTickerThread cron_thread = SupervisedTickerThread( cron_provider.start, args=(cron_stop,), kwargs=cron_start_kwargs, stop_event=cron_stop) from gateway.runtime_ownership import process_ownership process_ownership.start_writer(cron_thread) # External providers fire over loopback HTTP to THIS process's api_server; if it never came up (usually # API_SERVER_KEY missing) every fire fails while manual runs work — misread as a job bug. Say it ONCE. if not isinstance(cron_provider, InProcessCronScheduler): try: _has_api_server = Platform.API_SERVER in (runner.adapters or {}) except Exception: _has_api_server = True # never let the tell break startup if not _has_api_server: logger.warning( "Cron provider '%s' is active but the api_server adapter is " "NOT running in this gateway — scheduled fires arrive over " "loopback HTTP and will all fail (jobs only run when " "triggered manually). Most common cause: API_SERVER_KEY is " "missing from this gateway process's environment. Restart " "the gateway through its supervisor (`hermes gateway " "restart`) so the profile env loads.", getattr(cron_provider, "name", "external")) # Gateway-only housekeeping runs independently of the cron provider; shares cron_stop for shutdown. housekeeping_thread = threading.Thread( target=_start_gateway_housekeeping, args=(cron_stop,), kwargs={"adapters": runner.adapters, "loop": asyncio.get_running_loop(), "cron_provider": cron_provider, "runner": runner, "cron_thread": cron_thread}, daemon=True, name="gateway-housekeeping") process_ownership.start_writer(housekeeping_thread) return cron_stop, cron_provider, cron_thread, housekeeping_thread async def _start_gateway_shutdown_tail( runner, _control_server, cron_stop: threading.Event, cron_provider, cron_thread: Any, housekeeping_thread: threading.Thread, _planned_stop_watcher_stop: threading.Event, _planned_stop_watcher_thread: threading.Thread, _signal_initiated_shutdown: list) -> bool: """Post-``wait_for_shutdown`` teardown; returns the process exit verdict (True = exit 0).""" from gateway.run import (_CRON_SHUTDOWN_DRAIN_TIMEOUT, _HOUSEKEEPING_SHUTDOWN_DRAIN_TIMEOUT, _await_thread_exit, _best_effort, _resolve_gateway_exit_verdict, _shutdown_mcp_servers_nonblocking, _stop_cron_provider) # Control socket first: once shutdown begins we are no longer a truthful "serving here" answer and a # successor must be able to bind. Early-exit paths rely on the atexit cleanup_files hook instead. if _control_server is not None: try: await _control_server.stop() except Exception: logger.debug("Control socket stop failed (non-fatal)", exc_info=True) def _stop_keepalive() -> None: from hermes_cli.nous_auth_keepalive import stop_nous_auth_keepalive stop_nous_auth_keepalive() _best_effort(_stop_keepalive) # Never join(): an in-flight cron delivery is a coroutine on THIS loop; a sync join would drop it. # Stop cron scheduler + housekeeping cleanly. These MUST be awaited cooperatively, not join()ed. A cron # delivery in flight when the gateway restarts is a coroutine scheduled onto THIS event loop # (safe_schedule_threadsafe); the ticker thread is blocked on its future.result(). A synchronous # cron_thread.join() would block the loop, so that delivery could never run — it timed out and the # message was silently dropped (#58818). Awaiting keeps the loop alive so the in-flight delivery # finishes before we tear down. cron_stop.set() _stop_cron_provider(cron_provider) if not await _await_thread_exit(cron_thread, timeout=_CRON_SHUTDOWN_DRAIN_TIMEOUT): logger.warning("Cron ticker did not exit within %.0fs of shutdown — an in-flight " "delivery may have been dropped.", _CRON_SHUTDOWN_DRAIN_TIMEOUT) await _await_thread_exit(housekeeping_thread, timeout=_HOUSEKEEPING_SHUTDOWN_DRAIN_TIMEOUT) # Stop the planned-stop watcher (daemon=True so this is belt-and-suspenders). _planned_stop_watcher_stop.set() _planned_stop_watcher_thread.join(timeout=2) # Never suppressed: a raise here is a real teardown failure (it once hid a changed signature, # leaving every MCP connection and the shared loop up while the gateway reported a clean exit). try: await _shutdown_mcp_servers_nonblocking(config=getattr(runner, "config", None)) except Exception: logger.warning("MCP shutdown failed; connections may be left open", exc_info=True) # The failure verdict comes AFTER the cooperative teardown: returning early here leaked the # cron ticker + housekeeping threads (and open MCP connections) for embedded callers (#12175). return _resolve_gateway_exit_verdict(runner, _signal_initiated_shutdown[0]) def _launch_home_may_multiplex(config=None) -> bool: """A named profile serving ``default`` as a secondary splits default's local sessions between the two stores (routing rows follow the launch home, receipts follow the authority), so the multiplexer is the default profile's job. An EXPLICIT ``gateway.multiplex_profiles: true`` (config.yaml or ``GATEWAY_MULTIPLEX_PROFILES``) on a named launch home exits EX_CONFIG: a config verdict, never a retry. The implicit default (unset key, or the retired ``false`` that ``resolve_multiplex_mode`` settles like unset) is not the operator's choice, so it is downgraded the way every other implicit-multiplex blocker is: this process comes up standalone, says why, and *config* is flipped in place so the boot below reserves and serves only the launch profile.""" from hermes_constants import profile_name_for_home from gateway.run import get_hermes_home, logger name = profile_name_for_home(get_hermes_home()) if name in (None, "default"): return True from hermes_cli.gateway_multiplex_mode import explicit_multiplex_flag if config is not None and not explicit_multiplex_flag(get_hermes_home()): from hermes_cli.gateway_multiplex_mode import MultiplexDecision, log_multiplex_decision decision = MultiplexDecision( False, "guard", f"profile {name!r} launched the gateway, and only the default profile runs the multiplexer " "(start it with `hermes gateway run` from the default profile to serve every profile)") config.multiplex_profiles = False log_multiplex_decision(decision) return True from gateway.restart import GATEWAY_FATAL_CONFIG_EXIT_CODE from gateway.run import _write_runtime_status_quiet logger.error( "gateway.multiplex_profiles is set on profile %r, but only the default profile can run the " "multiplexer. Unset the flag in this profile's config.yaml, or start the gateway from the " "default profile (`hermes gateway run`), which serves this profile too.", name) _write_runtime_status_quiet(gateway_state="startup_failed", exit_reason="multiplex_requires_default_profile") raise SystemExit(GATEWAY_FATAL_CONFIG_EXIT_CODE) def _start_gateway_resolve_terminal_cwd() -> None: """Messaging-only defaults belong to startup, not incidental imports by the TUI.""" import os from gateway.cwd_placeholder import CWD_PLACEHOLDERS, resolve_placeholder_terminal_cwd configured_cwd = os.environ.get("TERMINAL_CWD", "") if not configured_cwd or configured_cwd in CWD_PLACEHOLDERS: resolved_cwd = resolve_placeholder_terminal_cwd( configured_cwd=configured_cwd, terminal_backend=os.environ.get("TERMINAL_ENV", ""), messaging_cwd=os.getenv("MESSAGING_CWD"), docker_mount_cwd_to_workspace=os.getenv( "TERMINAL_DOCKER_MOUNT_CWD_TO_WORKSPACE", "false").lower() in {"true", "1", "yes"}, home_fallback=str(Path.home())) if resolved_cwd is None: os.environ.pop("TERMINAL_CWD", None) else: os.environ["TERMINAL_CWD"] = resolved_cwd # A browse that catches a lagging state.db up holds the profile's gateway.lock only for the drift # re-check and the additive DDL, cut off at hermes_state_reconcile.BROWSE_RECONCILE_BUDGET_S by a # sqlite progress handler. Its integrity scan (which scales with file size) runs BEFORE the lock, so # a large store never extends this window. A gateway booting in it waits it out, with margin for lock # publication and the drift probes, instead of exiting as if another gateway owned the profile. def _browse_reconcile_reserve_wait_s() -> float: from hermes_state_reconcile import BROWSE_RECONCILE_BUDGET_S return BROWSE_RECONCILE_BUDGET_S + 2.0 def _reserve_waiting_out_browse_reconcile(ownership, homes, *, wait_s: float | None = None): """Boot's all-or-nothing reserve of every served home, waiting out transient holders (a browse reconcile, an unpublished record, a client's lock probe) for up to the browse window.""" from gateway.runtime_ownership import reserve_waiting_out_transient_holders if wait_s is None: wait_s = _browse_reconcile_reserve_wait_s() return reserve_waiting_out_transient_holders(ownership, homes, wait_s=wait_s) def _start_gateway_reserve_profiles(profile_homes) -> bool: """Reserve the launch home plus every served profile home; False when startup must abort.""" from gateway.run import get_hermes_home from gateway.runtime_ownership import process_ownership, OwnershipConflict homes = [get_hermes_home(), *(home for _, home in profile_homes)] try: _reserve_waiting_out_browse_reconcile(process_ownership, homes) except OwnershipConflict as exc: logger.error("Cannot reserve gateway profiles: %s", exc) # A named profile that the LIVE default multiplexer already serves is a configuration # verdict (gateway.multiplex_profiles), not a transient fault: exit EX_CONFIG so systemd's # RestartPreventExitStatus=78 parks the unit instead of restart-looping (#51228, #97120). # Any other same-user contender keeps the ordinary "already running" exit 1. from hermes_cli.gateway import named_profile_served_by_running_multiplexer if named_profile_served_by_running_multiplexer(): from gateway.restart import GATEWAY_FATAL_CONFIG_EXIT_CODE from gateway.run import _write_runtime_status_quiet logger.error( "The default gateway is running as a profile multiplexer and already serves this profile. " "Manage it with `hermes gateway restart` from the default profile instead of starting a " "second gateway for the profile.") _write_runtime_status_quiet(gateway_state="startup_failed", exit_reason=str(exc)) raise SystemExit(GATEWAY_FATAL_CONFIG_EXIT_CODE) return False except OSError as exc: logger.error("Cannot reserve gateway profiles: %s", exc) return False return True def _start_gateway_install_signal_handlers(loop, shutdown_signal_handler, restart_signal_handler) -> None: """SIGINT/SIGTERM drain, SIGUSR1 restart; main thread only.""" import signal from contextlib import suppress if threading.current_thread() is threading.main_thread(): # add_signal_handler raises NotImplementedError on Windows; SIGUSR1 is POSIX-only. handlers = [(sig, shutdown_signal_handler, (sig,)) for sig in (signal.SIGINT, signal.SIGTERM)] if hasattr(signal, "SIGUSR1"): handlers.append((signal.SIGUSR1, restart_signal_handler, ())) # windows-footgun: ok — hasattr-guarded for sig, handler, args in handlers: with suppress(NotImplementedError): loop.add_signal_handler(sig, handler, *args) # windows-footgun: ok — suppress(NotImplementedError) else: logger.info("Skipping signal handlers (not running in main thread).") async def _start_gateway_run_runner(runner, _signal_initiated_shutdown: list) -> Optional[bool]: """MCP discovery, then ``runner.start()``. Returns the exit verdict when the process must end before running mode (start failed, clean exit, aborted startup); None to continue into cron.""" from gateway.run import (_discover_gateway_mcp_tools, _resolve_gateway_exit_verdict, _shutdown_gateway_health_export, _shutdown_mcp_servers_nonblocking) # discover_mcp_tools() blocks up to 120s; on the loop thread it would freeze platform heartbeats. try: # MCP tool discovery — run in an executor so the asyncio event loop stays responsive even when a # configured MCP server is slow or unreachable. discover_mcp_tools() uses a blocking 120s wait # internally; calling it from the loop thread would freeze platform heartbeats (Discord shard, # Telegram polling) until it returned. See #16856. await _discover_gateway_mcp_tools(runner.config) except Exception: logger.debug("MCP tool discovery failed (non-fatal)", exc_info=True) try: success = await runner.start() except BaseException: _shutdown_gateway_health_export(runner) raise if not success: _shutdown_gateway_health_export(runner) return False if runner.should_exit_cleanly: _shutdown_gateway_health_export(runner) if runner.exit_reason: logger.error("Gateway exiting cleanly: %s", runner.exit_reason) # Explicit exit codes (GATEWAY_FATAL_CONFIG_EXIT_CODE) must propagate so s6 finish maps 78 → 125. if runner.exit_code is not None: raise SystemExit(runner.exit_code) return True if not runner._running: # Startup aborted by restart/shutdown before running mode; preserve that path without starting cron. try: await runner.wait_for_shutdown() try: await _shutdown_mcp_servers_nonblocking(config=getattr(runner, "config", None)) except Exception: logger.warning("MCP shutdown failed; connections may be left open", exc_info=True) return _resolve_gateway_exit_verdict(runner, _signal_initiated_shutdown[0]) finally: _shutdown_gateway_health_export(runner) return None def _start_gateway_early_environment_probe() -> None: """Background-start the host toolchain probe the warm-up and first prompt read (idempotent; honours ``agent.environment_probe: false``). Never raises: the warm-up still runs it lazily.""" try: from hermes_cli.config import load_config_readonly agent_cfg = load_config_readonly().get("agent") if isinstance(agent_cfg, dict) and not agent_cfg.get("environment_probe", True): return from tools.env_probe import warm_environment_probe_async warm_environment_probe_async() except Exception: logger.debug("early environment probe did not start", exc_info=True) async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool = False, verbosity: Optional[int] = 0, force: bool = False, idle_exit: bool = False) -> bool: """Start the gateway and run until interrupted; False if it failed to start (non-zero exit so systemd can auto-restart). ``replace`` kills any existing instance first (avoids restart-loop deadlocks); ``force`` starts without consulting the host owner at all; ``idle_exit`` arms the unmanaged idle exit (``gateway.run_idle_exit``) for a client-started daemon.""" from gateway.run import ( GatewayRunner, _best_effort, _enable_multiplex_log_routing, _ensure_windows_gateway_venv_imports, _gateway_loop_exception_handler, _host_attach_or_none, _log_standalone_profiles_at_boot, _multiplex_profile_homes, _refresh_host_gateway_record, _run_planned_stop_watcher, _start_gateway_make_restart_signal_handler, asyncio, load_gateway_config_for_runner, os, threading, ) # Set here (not at import) so incidental gateway.run imports from CLI code don't poison it. os.environ["HERMES_EXEC_ASK"] = "1" from hermes_cli.update_proxy_env import restore_user_proxy_env restore_user_proxy_env() # a gateway an update restarted keeps the user's proxy env, not the update's _start_gateway_resolve_terminal_cwd() from hermes_cli.resource_limits import apply_nofile_soft_limit apply_nofile_soft_limit() # Snapshot the revision while sys.modules matches disk so a later `git pull` is detected safely. from gateway.code_skew import record_boot_fingerprint record_boot_fingerprint() # Config verdicts come before the duplicate-instance guard: `--replace` must not stop a healthy # gateway for a launch that cannot start. resolved_config = config if config is not None else load_gateway_config_for_runner() # The first turn's prompt needs the host toolchain probe (five subprocess calls); start it now # so it runs beside the boot imports instead of serially inside the pre-READY warm-up. _start_gateway_early_environment_probe() profile_homes = (_multiplex_profile_homes(resolved_config) if getattr(resolved_config, 'multiplex_profiles', False) else []) if profile_homes and not _launch_home_may_multiplex(resolved_config): return False if profile_homes and not getattr(resolved_config, 'multiplex_profiles', False): profile_homes = [] # implicit default downgraded: this named launch serves only itself # Multiplex-only: the ONE host gateway decides first. Attach to it, make it serve this profile, # replace it (--replace) or refuse — before anything below binds a port or claims a PID file. _host_decision = await _host_attach_or_none(replace, force) if _host_decision is not None: return _host_decision # Duplicate-instance guard scoped to HERMES_HOME (the host record is absent or unusable here). from gateway.status import get_running_pid existing_pid = get_running_pid() if (existing_pid is not None and existing_pid != os.getpid() and not await _start_gateway_replace_existing_instance(existing_pid, replace)): return False from gateway.status import remove_pid_file, release_gateway_runtime_lock if not _start_gateway_reserve_profiles(profile_homes): return False # Freeze discovery: later profile additions must restart and reserve before opening stores. if profile_homes: resolved_config._runtime_profile_homes = tuple(profile_homes) # PID file BEFORE adapters: of two concurrent `run --replace`, only the O_EXCL winner opens sockets. # Only --force skips the host-lock refusal. Every generated unit carries --replace, so reading it as # --force there disabled the one arbiter of the two-units-at-once race; a replace that took the # owner over already freed the lock with that process. if not _start_gateway_claim_pid_file(force=force): release_gateway_runtime_lock() return False _control_server = None runner = None _planned_stop_watcher_stop = None try: _start_gateway_configure_logging(verbosity) from gateway.run_startup import recover_left_core_at_gateway_start await asyncio.to_thread(recover_left_core_at_gateway_start) # before the runner loads platform config from gateway.run_runtime import initialize_gateway_runtime, pin_launch_store, unpin_launch_store _launch_store_pin = await asyncio.to_thread(pin_launch_store) try: runner = GatewayRunner(resolved_config) finally: unpin_launch_store(_launch_store_pin) await initialize_gateway_runtime(runner) # Multiplex: swap the launch-home file handlers for per-profile routers so each profile's records # land in its own logs/. Must run after the runner resolved (possibly None) config and setup_logging. # See #82936. _enable_multiplex_log_routing(runner.config) # ``--replace`` is explicit startup authority, not a durable reconnect policy: GatewayRunner scopes # it to cold adapter connects and clears it before the background reconnect watcher starts. runner._platform_lock_takeover_on_start = bool(replace) # Unexpected signals exit non-zero so service managers revive us; planned stops write a marker first. _signal_initiated_shutdown = [False] shutdown_signal_handler = _start_gateway_make_shutdown_signal_handler( runner, _signal_initiated_shutdown) restart_signal_handler = _start_gateway_make_restart_signal_handler(runner) loop = asyncio.get_running_loop() # Swallow transient network errors from background tasks; one unhandled httpx error would kill us. # Issues #31066 / #31110: an unhandled ``telegram.error.TimedOut`` (or peer NetworkError / httpx # connection error) in any awaited coroutine would propagate to the loop and kill the gateway process, # taking down every profile attached to the same runner. systemd then restarts the service after ~5s but # the active conversation turn is lost. The fix is intentionally narrow: only well-known transient # network errors are swallowed (and logged with full traceback so the originating call site is still # discoverable). Anything else is forwarded to the default handler so real bugs still surface. loop.set_exception_handler(_gateway_loop_exception_handler) _start_gateway_install_signal_handlers(loop, shutdown_signal_handler, restart_signal_handler) # Windows has no add_signal_handler, so `hermes gateway stop`'s SIGTERM would never drain; poll the # planned-stop marker (written BEFORE the kill) instead. Runs everywhere so masked-SIGTERM drains. # Windows fallback: asyncio.add_signal_handler raises NotImplementedError on Windows, so `hermes gateway # stop`'s SIGTERM (which Python maps to TerminateProcess on Windows) never invokes # shutdown_signal_handler. That means the drain loop never runs, mark_resume_pending never fires, and # sessions are silently lost across restarts (issue #33778). The fix is a marker-polling thread: `hermes # gateway stop` writes the planned-stop marker BEFORE killing, and this thread notices it and drives the # same shutdown path the signal handler would have. Runs on every platform (cheap, defensive) so # non-signal-bearing environments (Windows native, sandboxed CI runners that mask SIGTERM) still get a # clean drain. _planned_stop_watcher_stop = threading.Event() _planned_stop_watcher_thread = threading.Thread( target=_run_planned_stop_watcher, args=(_planned_stop_watcher_stop, runner, loop, shutdown_signal_handler), daemon=True, name="planned-stop-watcher") _planned_stop_watcher_thread.start() # Right after the PID claim (which makes us authoritative). Fatal: local clients attach only # with tickets minted on this listener, so startup aborts without it. _control_server = await _start_gateway_start_control_socket(runner) if _control_server is None: raise RuntimeError("gateway session bootstrap control listener unavailable") # Now the attach channel answers: republish the host record with the settled served set. _refresh_host_gateway_record(runner) _log_standalone_profiles_at_boot(runner) from gateway.run_runtime import start_gateway_runtime_api await start_gateway_runtime_api(runner) def _lifecycle_record_startup() -> Optional[dict]: # Report if the previous life died uncleanly (SIGKILL / OOM / VM death), then claim the # sentinel for this life. After the PID-file claim so a --replace loser can't clobber it. # The store scan that report used to run here is deferred: see below. from gateway.lifecycle_ledger import record_startup return record_startup() def _start_keepalive() -> None: from hermes_cli.nous_auth_keepalive import start_nous_auth_keepalive start_nous_auth_keepalive() _unclean_exit = _best_effort(_lifecycle_record_startup, "Lifecycle ledger startup record failed: %s") _best_effort(_start_keepalive, "Nous auth keepalive did not start: %s") _ensure_windows_gateway_venv_imports() _early_verdict = await _start_gateway_run_runner(runner, _signal_initiated_shutdown) if _early_verdict is not None: return _early_verdict if _unclean_exit is not None: _start_deferred_state_db_integrity_check(_unclean_exit) cron_stop, cron_provider, cron_thread, housekeeping_thread = ( _start_gateway_start_cron_and_housekeeping(runner)) # From here every exit (a READY publication that raises included) runs the shutdown tail, # which stops the cron ticker and housekeeping before the adapters they deliver through close. from gateway.run_runtime import publish_gateway_runtime_ready, wait_gateway_runtime try: # READY only once adapters, cron and housekeeping run; missing systemd state just disables watchdog. publish_gateway_runtime_ready(runner) runner._start_systemd_watchdog() if idle_exit: from gateway.run_idle_exit import arm_unmanaged_idle_exit arm_unmanaged_idle_exit(runner) else: # A supervised/explicit gateway serves Desktop and the dashboard SPA, which call the # dashboard routes right away: mount them now, after READY. A client's own idle-exit # gateway (chat -q, TUI) only uses the native WebSocket; it mounts on first request. from gateway.run_api import warm_gateway_dashboard warm_gateway_dashboard(runner) await wait_gateway_runtime(runner) finally: try: await runner.stop() finally: verdict = await _start_gateway_shutdown_tail( runner, _control_server, cron_stop, cron_provider, cron_thread, housekeeping_thread, _planned_stop_watcher_stop, _planned_stop_watcher_thread, _signal_initiated_shutdown) return verdict finally: try: if runner is not None: from gateway.run_runtime import drain_gateway_runtime await drain_gateway_runtime(runner) # Startup may have opened adapters before an exception or an early exit verdict. The # same stop path owns their writers and DB handles on every exit, but when no stop ran # yet this is cleanup, not a shutdown: keep startup_failed and the crash markers. if getattr(runner, "_stop_task", None) is None and not getattr(runner, "_running", False): runner._startup_cleanup_only = True await runner.stop() finally: if _planned_stop_watcher_stop is not None: _planned_stop_watcher_stop.set() if _control_server is not None: await _control_server.stop() remove_pid_file() release_gateway_runtime_lock()