| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302 |
- """Wake an existing agent session from a background completion event. Push-capable adapters
- (``supports_async_delivery``) get a synthetic ``MessageEvent(internal=True)`` via handle_message;
- stateless adapters (API server) would run that under a ``build_session_key()`` key that never
- matches the raw ``X-Hermes-Session-Id`` real turns use (invisible parallel session), so we self-POST
- ``/v1/chat/completions`` with the raw id header to resume the REAL session. Exception:
- async-delegation completions: the CLIENT owns the next turn, so they are never self-POSTed as a
- new ``role=user`` prompt (could cross a pending human-confirmation gate); instead
- ``persist_delegation_delivery`` writes a durable DELIVERY row (``display_kind=
- "async_delegation_complete"``, read by TUI/desktop pollers). Failures RAISE (after bounded retries
- on transient errors) so callers can rewind cursors / retry instead of silently losing the event."""
- from __future__ import annotations
- import asyncio
- import logging
- from typing import Any, Optional
- logger = logging.getLogger(__name__)
- # A wake self-post runs the whole agent turn synchronously (stream=false); generous ceiling so long
- # tool-using turns aren't killed mid-flight.
- WAKE_TURN_TIMEOUT_SECONDS = 600.0
- # Backoff between retries on transient failures. The API server has no per-session lock (concurrent
- # turns are last-writer-wins) but DOES enforce a global max_concurrent_runs cap via HTTP 429, which
- # is worth waiting out.
- _RETRY_DELAYS_SECONDS = (2.0, 5.0, 10.0)
- def adapter_supports_push(adapter: Any) -> bool:
- """Whether this adapter can push a message to the user after a turn ends. Reads
- ``supports_async_delivery`` off the adapter class rather than the request-scoped contextvar
- (background watchers run outside any bound session context). Adapters that don't declare
- the flag are push-capable."""
- return bool(getattr(adapter, "supports_async_delivery", True))
- # Store refusal codes (``RuntimeStoreError.reason``) that refuse a notice for good by themselves: the
- # route has no session (``not_found``) or the payload is malformed (``invalid_params``). A bare
- # ``admission_conflict`` proves nothing: an unowned completion, or one pinned to a live session the
- # route does not run, may still find its owner (d6635ffe03b), so it is permanent only when the
- # authority's state proves the pin gone (``refusal_proven_permanent``). Every other refusal
- # (``runtime_draining`` during a restart's after-turn wait or a parked profile, ``queue_full``,
- # ``session_busy``, ``stale_epoch``, ...) clears on its own: back off, never give the notice up.
- PERMANENT_REFUSAL_REASONS = frozenset({'not_found', 'invalid_params'})
- class WakeNotAccepted(RuntimeError):
- """No adapter admission: retry without treating a healthy chat as dead. ``reason`` is the
- store's refusal code when the session authority refused it, ``runtime_draining`` when the
- adapter path refused it during a drain (None: the adapter did not accept);
- ``permanent`` is set once the refusal is proven final (``refusal_proven_permanent``)."""
- def __init__(self, message: str = "internal wake not accepted by adapter", *, reason: Optional[str] = None):
- super().__init__(message)
- self.reason = reason
- self.permanent = False
- def session_owned_by_profile(config: Any, profile: Optional[str], session_id: Any) -> bool:
- """True when a stateless (``api_server``) destination's raw session id is canonically owned by
- served *profile*'s own session store.
- A shared-listener mirror platform has no chat/thread/guild anchor a ``profile_routes`` entry
- could match, so the session store itself is the ownership proof: the row must exist in that
- profile's ``state.db`` under its own home and carry that profile's stamp (a NULL legacy stamp
- belongs to the store's own profile — the same rule the dashboard's session routes apply). An
- unserved profile, a missing row, a row stamped for another profile, or an unreadable store all
- fail closed. Shared by the Kanban notifier and the background-process wake path.
- """
- import contextlib
- from pathlib import Path
- if not session_id or not profile:
- return False
- profile = str(profile)
- try:
- from gateway.run import _multiplex_profile_homes
- home = dict(_multiplex_profile_homes(config)).get(profile)
- if home is None:
- return False
- from hermes_state import SessionDB
- db = SessionDB(Path(home) / "state.db", read_only=True)
- except Exception as exc:
- logger.debug("wake: session ownership check unavailable for %s/%s: %s", profile, session_id, exc)
- return False
- try:
- row = db.get_session(str(session_id))
- except Exception as exc:
- logger.debug("wake: session ownership lookup failed for %s/%s: %s", profile, session_id, exc)
- return False
- finally:
- with contextlib.suppress(Exception):
- db.close()
- return bool(row) and (row.get("profile_name") or profile) == profile
- async def admit_internal_event(adapter: Any, event: Any) -> None:
- """Require a concrete adapter admission, not merely a handler returning None.
- The public handler return stays unchanged. This receipt means scheduled/queued,
- not model execution, authorization of a later turn, or successful outbound delivery.
- """
- event._gateway_accepted = False
- # A draining gateway answers new work with a drain notice and discards queued follow-ups,
- # so a wake admitted now would never run. Refuse it so the caller keeps it for a retry.
- if _runner_draining(adapter):
- raise WakeNotAccepted("gateway is draining; retry after restart", reason="runtime_draining")
- await adapter.handle_message(event)
- if event._gateway_accepted is not True:
- raise WakeNotAccepted()
- # handle_message can await before it claims or queues the event (Telegram topic recovery, the
- # busy handler's profile scope). A drain seen here dooms the stored event: the drain gate refuses
- # a claimed turn and the pending discard drops a queued follow-up. Refuse it so the producer
- # retries it after the restart.
- if _runner_draining(adapter):
- raise WakeNotAccepted("gateway began draining during admission; retry after restart",
- reason="runtime_draining")
- # No drain began before or during admission, so the producer takes this receipt as final and
- # will not retry: a drain that begins later must run the wake rather than drop it (#138123).
- event._admitted_before_drain = True
- def _runner_draining(adapter: Any) -> bool:
- return getattr(getattr(adapter, "gateway_runner", None), "_draining", False) is True
- async def deliver_wake(adapter: Any, *, text: str, session_id: str = "", source: Any = None,
- notification_category: str = "result", profile: Optional[str] = None) -> None:
- """Deliver a wake turn to the session behind ``adapter``. ``session_id`` is the RAW session id
- (``X-Hermes-Session-Id`` / state.db key) — required for non-push adapters. ``source`` is the
- ``SessionSource`` for the synthetic event — required for push-capable adapters. ``profile``
- names the served profile that canonically owns a non-push destination; a non-default value is
- delivered in-process under the caller's profile scope (see ``_self_post_chat_completion``).
- Raises on failure so the caller can rewind/retry."""
- if adapter_supports_push(adapter):
- if source is None:
- raise ValueError("deliver_wake: push-capable adapter requires a SessionSource")
- from gateway.platforms.event import MessageEvent, MessageType
- runner = getattr(adapter, "gateway_runner", None)
- if runner is not None:
- from gateway.session_automation import admit_route_wake
- from hermes_state_runtime import RuntimeStoreError
- category = {"notification_category": notification_category} if notification_category != "result" else {}
- try:
- if await admit_route_wake(runner, adapter, source, text, "kanban-wake", category):
- return # committed to the session FIFO: never interrupts the admitted turn
- except RuntimeStoreError as exc:
- raise WakeNotAccepted(str(exc), reason=exc.reason) from exc
- synth_event = MessageEvent(text=text, message_type=MessageType.TEXT, source=source, internal=True,
- metadata={"notification_category": notification_category})
- await admit_internal_event(adapter, synth_event)
- return
- if not session_id:
- raise ValueError("deliver_wake: non-push adapter (supports_async_delivery=False) "
- "requires the raw session id to self-post the wake turn")
- extra: dict = {}
- if profile:
- extra["profile"] = profile
- if notification_category == "diagnostic":
- extra["notification_category"] = notification_category
- await _self_post_chat_completion(adapter, text=text, session_id=session_id, **extra)
- def _delegation_display_metadata(evt: dict) -> dict:
- """Display-only metadata for a persisted delegation delivery row. Mirrors
- ``tui_gateway.server._async_delegation_display_metadata`` (same ``display_kind`` consumer
- contract) without importing the TUI stack."""
- raw_results = evt.get("results")
- results = [r for r in raw_results if isinstance(r, dict)] if isinstance(raw_results, list) else []
- task_count = len(results) or 1
- completed_count = sum(1 for r in results if r.get("status") in {"completed", "success"})
- failed_count = sum(1 for r in results if r.get("status") in {"failed", "error"})
- metadata = {"delegation_id": str(evt.get("delegation_id") or ""), "task_count": task_count,
- "completed_count": completed_count or task_count - failed_count,
- "failed_count": failed_count}
- if evt.get("task_failure_notice"):
- metadata["delivery_notice"] = f"task_failure:{results[0].get('task_index', '') if results else ''}"
- metadata["notification_category"] = "diagnostic"
- from gateway.warning_notifications import warning_notifications_enabled
- if not warning_notifications_enabled("api_server"):
- metadata["presentation_suppressed"] = True
- duration = evt.get("total_duration_seconds") or evt.get("duration_seconds")
- if isinstance(duration, (int, float)):
- metadata["duration_seconds"] = duration
- return metadata
- async def persist_delegation_delivery(adapter: Any, *, text: str, session_id: str, evt: Optional[dict] = None) -> None:
- """Persist an async-delegation completion as a durable DELIVERY row (see module docstring)
- WITHOUT running any agent turn. Raises on failure so the caller can release the durable claim
- and retry.
- 85957: on stateless api_server sessions the client owns the turn after ``event.complete`` — a completion
- must never become a new ``role=user`` prompt via the self-post (that starts an unauthorized agent turn
- and can cross a pending human-confirmation gate). Instead, append the completion to the session
- transcript as a timeline bookkeeping row (``display_kind="async_delegation_complete"`` + display
- metadata — the exact shape the TUI/desktop delivery path persists), WITHOUT running any agent turn.
- Clients polling ``GET /api/sessions/{id}/messages`` see it immediately; the pre-request repair belt
- folds it into the next real client turn as context. See #85957.
- """
- if not session_id:
- raise ValueError("persist_delegation_delivery: raw session id required to persist "
- "the completion on the api_server session transcript")
- ensure = getattr(adapter, "_ensure_session_db", None)
- db: Any = await asyncio.to_thread(ensure) if callable(ensure) else None
- if db is None:
- raise RuntimeError("persist_delegation_delivery: api_server SessionDB unavailable — "
- f"cannot persist completion for session {session_id}")
- # #98619: the parent run may have compressed/rotated between dispatch and this detached
- # completion — the captured origin id is then a closed parent and the append below is
- # rejected with CompressionSessionClosedError forever (the watcher retries the same stale
- # id). Adopt the live continuation tip first, the same canonical resolution
- # /api/sessions/{id}/messages reads use, so the delivery row lands where the next run and
- # the messages endpoint both resolve. Fails open to the original id.
- resolver = getattr(db, "resolve_resume_session_id", None)
- if callable(resolver):
- try:
- resolved = await asyncio.to_thread(resolver, session_id)
- if resolved:
- session_id = str(resolved)
- except Exception:
- logger.debug("delegation delivery continuation resolve failed for %s", session_id, exc_info=True)
- await asyncio.to_thread(
- db.append_delegation_delivery, session_id, text, _delegation_display_metadata(evt or {}),
- )
- logger.info(
- "async delegation completion persisted as delivery row for api_server session %s (no wake turn)", session_id
- )
- async def _self_post_chat_completion(adapter: Any, *, text: str, session_id: str,
- notification_category: str = "result",
- profile: Optional[str] = None) -> None:
- """POST the wake text to the in-pod API server as a normal session turn, using the adapter's
- own bind host/port/key. Session continuation via ``X-Hermes-Session-Id`` is 403-gated on
- ``API_SERVER_KEY``, so a missing key is a hard error rather than a wake in a fresh session
- nobody watches.
- ``profile`` (a served secondary under ``gateway.multiplex_profiles``) takes the IN-PROCESS
- route instead of the HTTP one: the shared listener's ``/p/<profile>/`` mirror authenticates
- with that profile's own ``API_SERVER_KEY`` — which a route-only profile legitimately does not
- have — and an unprefixed self-post would resume the session in the DEFAULT profile's store.
- The caller already holds the owner profile's runtime scope, so the turn lands in that profile's
- own session. A non-default profile whose adapter cannot run in-process fails closed.
- """
- if profile and str(profile) != "default":
- in_process: Any = getattr(adapter, "run_internal_session_turn", None)
- if not callable(in_process):
- raise RuntimeError(
- f"wake self-post for served profile {profile!r} requires in-process session "
- "delivery; refusing to self-post as the default profile")
- await in_process(session_id=session_id, text=text, profile=str(profile),
- notification_category=notification_category)
- return
- import aiohttp
- host = str(getattr(adapter, "_host", "") or "127.0.0.1")
- if host in ("0.0.0.0", "::", "*"):
- host = "127.0.0.1" # wildcard bind — connect over loopback
- port = int(getattr(adapter, "_port", 0) or 8642)
- api_key = str(getattr(adapter, "_api_key", "") or "")
- if not api_key:
- raise RuntimeError("wake self-post requires API_SERVER_KEY: session continuation via "
- "X-Hermes-Session-Id is rejected (403) on an unauthenticated API "
- "server, so the wake cannot reach the target session")
- if ":" in host and not host.startswith("["):
- host = f"[{host}]" # bare IPv6 literal
- url = f"http://{host}:{port}/v1/chat/completions"
- headers = {"Authorization": f"Bearer {api_key}", "X-Hermes-Session-Id": session_id}
- payload = {"model": str(getattr(adapter, "_model_name", "") or "hermes-agent"),
- "messages": [{"role": "user", "content": text}], "stream": False}
- if notification_category == "diagnostic":
- payload["hermes_notification_category"] = "diagnostic"
- last_err: Optional[BaseException] = None
- attempts = 1 + len(_RETRY_DELAYS_SECONDS)
- for attempt in range(attempts):
- if attempt:
- await asyncio.sleep(_RETRY_DELAYS_SECONDS[attempt - 1])
- try:
- timeout = aiohttp.ClientTimeout(total=WAKE_TURN_TIMEOUT_SECONDS)
- async with aiohttp.ClientSession(timeout=timeout) as http:
- async with http.post(url, json=payload, headers=headers) as resp:
- if resp.status == 429: # concurrency cap — transient; back off and retry
- last_err = RuntimeError(
- f"wake self-post got HTTP 429 (concurrency cap) for session {session_id}"
- )
- logger.warning("%s; attempt %d/%d", last_err, attempt + 1, attempts)
- continue
- if resp.status >= 400: # non-transient (auth/validation): fail immediately
- body = (await resp.text())[:300]
- raise RuntimeError(
- f"wake self-post failed for session {session_id}: HTTP {resp.status}: {body}"
- )
- await resp.read()
- logger.info("wake self-post delivered for session %s (attempt %d)", session_id, attempt + 1)
- return
- except (TimeoutError, aiohttp.ClientError, OSError) as exc:
- last_err = exc
- logger.warning("wake self-post transient failure for session %s (attempt %d/%d): %s",
- session_id, attempt + 1, attempts, exc)
- raise RuntimeError(
- f"wake self-post gave up for session {session_id} after {attempts} attempts: {last_err}"
- ) from last_err
|