mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-06 05:17:15 +00:00
Teknium
0e89fb1f2f
SessionStore publishes a route transition in two halves: the new entry under _lock, then the state.db row (insert for /new and a fresh route, reopen for /resume) after releasing it. Anything that resolved the route in between saw the named conversation's row missing or still ended and took the #54878 stale-route self-heal: drop the entry and recover the newest open row, which is the conversation the user just left. A Desktop/Webapp turn finishing on a chat's previous conversation does exactly that (its post-turn bookkeeping calls get_or_create_session), so after /new the chat's route named A again and its next message could land there. The store now remembers the ids it published whose row is not live yet, and _is_session_ended_in_db does not call those ended. Every reader of that check (the routing self-heal, the reaped-run eviction, the cached-agent liveness check) gets the fix.
308 lines
15 KiB
Python
308 lines
15 KiB
Python
"""SessionStore explicit suspension, crash-recovery markers, pruning and shared clock/id helpers."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import logging
|
|
import os
|
|
import time
|
|
import uuid
|
|
from datetime import datetime, timedelta, timezone, UTC
|
|
from typing import TYPE_CHECKING, Optional
|
|
|
|
from hermes_state_ids import new_session_id
|
|
|
|
if TYPE_CHECKING:
|
|
from gateway.session import SessionEntry, SessionSource
|
|
|
|
# Log-record parity with the origin module.
|
|
logger = logging.getLogger("gateway.session")
|
|
|
|
|
|
def _now() -> datetime:
|
|
"""Return the current local time."""
|
|
return datetime.now()
|
|
|
|
|
|
def _new_session_id(now: datetime) -> str:
|
|
return new_session_id(now, hex_len=8)
|
|
|
|
|
|
def _iso(dt: Optional[datetime]) -> Optional[str]:
|
|
return dt.isoformat() if dt else None
|
|
|
|
|
|
def _parse_iso(value) -> Optional[datetime]:
|
|
"""``datetime.fromisoformat`` that returns None for empty/malformed input."""
|
|
if not value:
|
|
return None
|
|
try:
|
|
return datetime.fromisoformat(value)
|
|
except (TypeError, ValueError):
|
|
return None
|
|
|
|
|
|
# Auto-continue freshness window (1 hour) after the ``resume_pending`` mark; ``gateway/run.py``
|
|
# bridges config.yaml ``agent.gateway_auto_continue_freshness`` into the env var at startup.
|
|
_AUTO_CONTINUE_FRESHNESS_SECS_DEFAULT = 60 * 60
|
|
|
|
|
|
def auto_continue_freshness_window() -> float:
|
|
"""Resume-scheduler freshness window; stale automation never discards the transcript."""
|
|
raw = os.environ.get("HERMES_AUTO_CONTINUE_FRESHNESS")
|
|
try:
|
|
return float(raw) if raw else float(_AUTO_CONTINUE_FRESHNESS_SECS_DEFAULT)
|
|
except (TypeError, ValueError):
|
|
return float(_AUTO_CONTINUE_FRESHNESS_SECS_DEFAULT)
|
|
|
|
|
|
class SessionLifecycleMixin:
|
|
"""SessionStore explicit boundaries and crash-recovery markers."""
|
|
|
|
def _is_session_ended_in_db(self, session_id: str, session_key: Optional[str] = None) -> bool:
|
|
"""True iff state.db says the session is gone: ended (non-null end_reason) or hard-deleted
|
|
(no row in a readable owning DB). No DB or a DB error -> False (same failure mode as
|
|
``_prune_stale_sessions_locked``). Lets routing self-heal a session finalized or deleted
|
|
while the gateway stays alive. Store resolved from the owning profile.
|
|
|
|
Used by ``get_or_create_session`` to self-heal at routing time: ``_prune_stale_sessions_locked``
|
|
only runs at startup, so a session ended in the DB while the gateway stays alive (any path that
|
|
finalizes the row without clearing sessions.json) would otherwise be reused as a live routing key
|
|
and silently swallow every subsequent message until the next restart (#54878 — the live-gateway
|
|
variant of #52804/FM9). A hard delete is the same shape one step further: the row is GONE, not
|
|
merely ended, and reusing the route makes run_agent's INSERT OR IGNORE resurrect the deleted
|
|
session with its old id (#42422) — so a missing row is treated exactly like an ended one. DB
|
|
errors are non-fatal — never block routing on a failed lookup.
|
|
The store is resolved from the row's owning profile rather than the ambient scope: an unscoped
|
|
background writer keeps its own copy of the same session, and comparing against that copy reports a
|
|
live session as ended (#66887).
|
|
|
|
Pass *session_key* when the owning key is known but the id may have left the routing index:
|
|
after the self-heal re-homes the key, the id has no owner and would resolve to the launch
|
|
store, which never holds a routed profile's rows (#118862).
|
|
"""
|
|
with self._lock:
|
|
if session_id in self._rows_pending:
|
|
# A route transition published this id and has not inserted/reopened its row yet
|
|
# (a /new racing a turn that resolves the route as it finishes): healing it here
|
|
# would hand the route back to the conversation the user just left.
|
|
return False
|
|
db = self._db_for_key(session_key) if session_key else self._db_for_session_id(session_id)
|
|
if not db or not session_id:
|
|
return False
|
|
try:
|
|
row = db.get_session(session_id)
|
|
except Exception:
|
|
return False
|
|
return row is None or row.get("end_reason") is not None
|
|
|
|
def _local_receipt_target(self, entry: Optional[SessionEntry]) -> Optional[str]:
|
|
"""The physical target the authority's receipt names for *entry*'s local route, else None
|
|
(not a local route, no receipt, or the receipt belongs to another route).
|
|
|
|
Such a route moves only through the authority (compression and reset advance the receipt
|
|
and the route in one transaction), and its first admitted turn reopens a stamped row
|
|
(``reopen_local_session``). A row stamped ended (a viewer's ``ws_orphan_reap``) is
|
|
therefore never a stale route: dropping it minted a replacement id the authority refuses
|
|
(``admission_conflict``) and stranded the conversation (#140642)."""
|
|
origin = getattr(entry, "origin", None)
|
|
if origin is None or getattr(origin, "platform", None) is None or origin.platform.value != "local":
|
|
return None
|
|
db = self._db_for_key(entry.session_key)
|
|
if db is None or getattr(db, "_read_ctx", None) is None:
|
|
return None
|
|
import sqlite3
|
|
|
|
from hermes_state_local import local_receipt
|
|
from hermes_state_runtime import RuntimeStoreError
|
|
try:
|
|
receipt = local_receipt(db, str(origin.chat_id))
|
|
if receipt.get("route") != entry.session_key:
|
|
return None
|
|
return str(receipt["entry"]["session_id"]) or None
|
|
except (RuntimeStoreError, sqlite3.Error, KeyError, TypeError, AttributeError):
|
|
return None
|
|
|
|
def _api_bound(self, entry: Optional[SessionEntry]) -> bool:
|
|
"""*entry*'s API route is the one its binding receipt names (``bind_api_session``). Like a
|
|
local receipt route it moves only through the authority, and its next admitted turn reopens a
|
|
stamped row (``reopen_api_session``): replacing it forked the caller's conversation onto a
|
|
minted id every later run on it refused (``admission_conflict``)."""
|
|
origin = getattr(entry, "origin", None)
|
|
if origin is None or getattr(origin, "platform", None) is None or origin.platform.value != "api_server":
|
|
return False
|
|
db = self._db_for_key(entry.session_key)
|
|
if db is None or getattr(db, "_read_ctx", None) is None:
|
|
return False
|
|
import json
|
|
import sqlite3
|
|
|
|
from hermes_state_local import API_BINDING_PREFIX
|
|
try:
|
|
with db._read_ctx() as conn:
|
|
saved = conn.execute("SELECT value FROM state_meta WHERE key=?",
|
|
(API_BINDING_PREFIX + str(origin.chat_id),)).fetchone()
|
|
return saved is not None and json.loads(saved[0])["entry"]["session_key"] == entry.session_key
|
|
except (sqlite3.Error, ValueError, KeyError, TypeError):
|
|
return False
|
|
|
|
def _route_reset_reason(self, entry: SessionEntry) -> Optional[str]:
|
|
"""Only explicit suspension replaces a routed conversation; time never does."""
|
|
return "suspended" if entry.suspended else None
|
|
|
|
def _update_entry(self, session_key: str, mutate) -> bool:
|
|
"""Apply ``mutate(entry)`` under ``_lock`` and full-save; False when the entry is missing
|
|
or *mutate* returned False (nothing to persist)."""
|
|
with self._lock:
|
|
entry = self._entry_locked(session_key)
|
|
if entry is None or mutate(entry) is False:
|
|
return False
|
|
self._save()
|
|
return True
|
|
|
|
def _update_all_entries_locked(self, mutate) -> int:
|
|
"""Apply ``mutate(entry) -> bool`` to every entry under ``_lock``; save once if any
|
|
returned True. Returns the count that did."""
|
|
with self._lock:
|
|
self._ensure_loaded_locked()
|
|
changed = sum(1 for entry in self._entries.values() if mutate(entry))
|
|
if changed:
|
|
self._save()
|
|
return changed
|
|
|
|
def suspend_session(self, session_key: str) -> bool:
|
|
"""Mark a session suspended so it auto-resets on next access (/stop). True if it existed.
|
|
|
|
Used by ``/stop`` to prevent stuck sessions from being resumed after a gateway restart (#7536).
|
|
"""
|
|
return self._update_entry(session_key, lambda e: setattr(e, "suspended", True))
|
|
|
|
def _set_turn_marker_locked(self, session_key: str, entry: SessionEntry, token, started_at) -> None:
|
|
"""Persist the active-turn pair BEFORE publishing it in memory, so a failed write can
|
|
neither leak an unowned token nor drop a live one. Lock held."""
|
|
candidate = entry.to_dict()
|
|
candidate["active_turn_token"] = token
|
|
candidate["active_turn_started_at"] = _iso(started_at)
|
|
touched = _now() if started_at is not None else None
|
|
if touched is not None:
|
|
# Keeps the legacy 120s startup heuristic working for an older binary during a rolling
|
|
# downgrade/upgrade window.
|
|
candidate["updated_at"] = touched.isoformat()
|
|
self._save_entry(session_key, entry_data=candidate, lock_held=True)
|
|
entry.active_turn_token = token
|
|
entry.active_turn_started_at = started_at
|
|
if touched is not None:
|
|
entry.updated_at = touched
|
|
|
|
def mark_turn_active(self, session_key: str) -> Optional[str]:
|
|
"""Persist exact ownership of the running agent turn; returns the opaque token for
|
|
:meth:`clear_turn_active`. Re-marking replaces the previous token so a stale asynchronous
|
|
unwind cannot clear a newer turn."""
|
|
token = uuid.uuid4().hex
|
|
with self._lock:
|
|
entry = self._entry_locked(session_key)
|
|
if entry is None:
|
|
return None
|
|
# Aware UTC, unlike the local wall clock elsewhere: the next process compares it with
|
|
# epoch transcript timestamps and may run in another zone (DST, container vs unit TZ).
|
|
self._set_turn_marker_locked(session_key, entry, token, datetime.now(UTC))
|
|
return token
|
|
|
|
def clear_turn_active(self, session_key: str, token: str) -> bool:
|
|
"""Compare-and-swap clear an active-turn marker; ``False`` when the entry disappeared or a
|
|
newer turn owns it."""
|
|
with self._lock:
|
|
entry = self._entry_locked(session_key)
|
|
if entry is None or entry.active_turn_token != token:
|
|
return False
|
|
self._set_turn_marker_locked(session_key, entry, None, None)
|
|
return True
|
|
|
|
def recover_interrupted_turns(self, max_age_seconds: int = 60 * 60) -> int:
|
|
"""Promote crash-left turn markers into ``resume_pending`` (unclean startup only).
|
|
Old/invalid markers are cleared without resuming; suspended sessions are never re-armed.
|
|
Returns the number of newly promoted sessions."""
|
|
now, epoch_now = _now(), time.time()
|
|
promoted = 0
|
|
|
|
def _promote(entry: SessionEntry) -> bool:
|
|
nonlocal promoted
|
|
if not entry.active_turn_token:
|
|
return False
|
|
started_at = entry.active_turn_started_at
|
|
# Epoch arithmetic: a pre-upgrade naive marker reads as local time, an aware one exactly.
|
|
marker_is_stale = started_at is None or (
|
|
max_age_seconds > 0 and epoch_now - started_at.timestamp() > max_age_seconds
|
|
)
|
|
if not marker_is_stale and not entry.suspended:
|
|
if entry.resume_pending:
|
|
# A drain-timeout marker is more specific; keep it.
|
|
if entry.last_resume_marked_at is None:
|
|
entry.last_resume_marked_at = now
|
|
else:
|
|
entry.resume_pending = True
|
|
entry.resume_reason = "restart_interrupted"
|
|
entry.last_resume_marked_at = now # freshness starts at discovery
|
|
promoted += 1
|
|
entry.active_turn_token = None
|
|
entry.active_turn_started_at = None
|
|
return True
|
|
|
|
self._update_all_entries_locked(_promote)
|
|
return promoted
|
|
|
|
def discard_active_turn_markers(self) -> int:
|
|
"""Clear orphan turn markers after a verified clean shutdown."""
|
|
def _discard(entry: SessionEntry) -> bool:
|
|
if not entry.active_turn_token and entry.active_turn_started_at is None:
|
|
return False
|
|
entry.active_turn_token = None
|
|
entry.active_turn_started_at = None
|
|
return True
|
|
return self._update_all_entries_locked(_discard)
|
|
|
|
def mark_resume_pending(self, session_key: str, reason: str = "restart_timeout") -> bool:
|
|
"""Mark a session resumable after a restart interruption (keeps the session_id/transcript,
|
|
unlike ``suspend_session``). True if marked."""
|
|
def _apply(entry: SessionEntry):
|
|
if entry.suspended: # never override an explicit ``suspended`` (hard forced-wipe)
|
|
return False
|
|
entry.resume_pending = True
|
|
entry.resume_reason = reason
|
|
entry.last_resume_marked_at = _now()
|
|
return self._update_entry(session_key, _apply)
|
|
|
|
def clear_resume_pending(self, session_key: str) -> bool:
|
|
"""Clear the resume-pending flag after a successful resumed turn; True if cleared."""
|
|
def _apply(entry: SessionEntry):
|
|
if not entry.resume_pending:
|
|
return False
|
|
entry.resume_pending = False
|
|
entry.resume_reason = None
|
|
entry.last_resume_marked_at = None
|
|
return self._update_entry(session_key, _apply)
|
|
|
|
def prune_old_entries(self, max_age_days: int) -> int:
|
|
"""Drop routing entries idle (by ``updated_at``) for more than max_age_days; suspended
|
|
entries and entries with active background processes are kept. Only the key -> session_id
|
|
mapping is dropped (the transcript stays). ``max_age_days <= 0`` disables. Returns count."""
|
|
if max_age_days is None or max_age_days <= 0:
|
|
return 0
|
|
cutoff = _now() - timedelta(days=max_age_days)
|
|
with self._lock:
|
|
self._ensure_loaded_locked()
|
|
removed_keys = [
|
|
key for key, entry in list(self._entries.items())
|
|
if not entry.suspended
|
|
# The callback is keyed by session_key, NOT session_id.
|
|
and not self._has_active_processes_safe(entry.session_key, context="prune")
|
|
and entry.updated_at < cutoff
|
|
]
|
|
for key in removed_keys:
|
|
self._entries.pop(key, None)
|
|
if removed_keys:
|
|
self._save()
|
|
if removed_keys:
|
|
logger.info("SessionStore pruned %d entries older than %d days",
|
|
len(removed_keys), max_age_days)
|
|
return len(removed_keys)
|