You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Scheduler observability: APScheduler logs are now visible in Chibi's loguru sinks. APScheduler 3.x emits everything through stdlib logging on the apscheduler.* logger family, and Chibi never bridged it — so every INFO/DEBUG record (Scheduler started, Added job ... to job store, Next wakeup is due at ..., No jobs; waiting until a job is added) was discarded outright, and even the surviving WARNING+ records (e.g. Run time of job ... was missed by 2:00:00) leaked to stderr as bare, unformatted lines with no timestamp, no level and no scheduler marker, making "did not fire" and "fired but delivery failed" indistinguishable in the logs. A dedicated stdlib→loguru bridge (chibi.config.logging) now installs a single intercepting handler on the apscheduler logger (idempotent, propagation disabled — no duplicate or double-formatted output, no logging.lastResort leak), scoped narrowly to the apscheduler.* family so all other loggers and every runner sink behave exactly as before. Bridged records render in the standard datetime | LEVEL | message shape with the stdlib logger name embedded in the message ([apscheduler.scheduler] ...), so a plain grep -i scheduler over the logs finally matches.
New SCHEDULER_DEBUG_LOGS setting (default false): the bridge floor is flag-gated — with the flag off the floor is INFO (start, scheduling, missed-run warnings; no per-wakeup chatter), with SCHEDULER_DEBUG_LOGS=true it drops to DEBUG so Next wakeup is due at ... and Added job ... become visible for diagnosing wakeup-arming issues. Toggling the env var only requires a process restart, not a rebuild. Documented in the generated configuration reference and the settings docstring.
Explicit Chibi-level scheduler logging, independent of the stdlib bridge: every successful scheduling call (schedule_interval_job / schedule_cron_job / schedule_once_job) now logs an INFO line carrying the job id, the owner (derived from the job-id namespace — system or the agent's user id), a trigger summary and the next run time, and a failed add_job logs a WARNING with the job id and error before re-raising. Additionally, a Chibi-level WARNING is logged whenever APScheduler dispatches a missed-run event (EVENT_JOB_MISSED), carrying the job id and the missed-by duration — the exact signature of the macOS-sleep digest failures, previously invisible.
Docs: self-sufficient trigger_text contract for self-type scheduled jobs. The schedule_task tool now states explicitly (in the trigger_text parameter description, the action-type descriptions and the SCHEDULER_HINT_PROMPT system-prompt hint) that for self jobs the trigger text must be exhaustive and self-sufficient — goal, requirements, constraints and expected output — because at fire time the model may have no memory of the scheduling conversation (thread cleared, context evicted, or a different chat). Includes good vs. bad examples. Docs-only: no schema or runtime changes.
Fixed
Scheduler wakeup guard against silent job starvation: a transient job store failure (e.g. a Redis hiccup) inside APScheduler's wakeup() left the wakeup timer permanently disarmed — the scheduler stayed "running" but silently stopped executing jobs until an unrelated external wakeup() (a new job or a chat message) happened to re-arm it. This mirrored an upstream APScheduler bug (agronholm/apscheduler#1132, merged to the 3.x branch on 2026-08-31 but not yet released). Chibi now ships GuardedAsyncIOScheduler, a local guard that mirrors the upstream fix verbatim: an exception escaping _process_jobs() is logged through the standard apscheduler.scheduler logger and the wakeup timer is re-armed with jobstore_retry_interval. Both scheduler flavors (Telegram ChibiScheduler and stdio/IDE StdioScheduler) use it. The guard is marked TODO(apscheduler-3.11.4): remove guard and is a pure-deletion revert once APScheduler 3.11.4 ships. Regression tests simulate a jobstore.update_job() failure mid-wakeup and verify the scheduler keeps firing jobs.
Local semantic memory no longer breaks after OS temp cleanup: the LOCAL fastembed embedding model was cached in the system temp directory ($TMPDIR/fastembed_cache), which the OS wipes periodically — a wiped or interrupted download then made the ONNX model unloadable and killed Chroma archival and semantic search for the whole deployment. The model is now cached in a persistent per-user directory (~/.cache/chibi/fastembed), so a temp-directory cleanup can no longer destroy it; the location is overridable with the FASTEMBED_CACHE_PATH environment variable. Embedding vectors, model and dimension are unchanged, so existing collections need no re-embedding.
Changed
APScheduler upgraded 3.11.2 → 3.11.3 across all dependency artifacts: poetry.lock, and the exported requirements.txt (the production Docker image pin) and requirements-dev.txt (pyproject.toml already allowed ^3.11); regenerating the requirements via the deps task (poetry export) keeps everything consistent. 3.11.3 fixes #1103 — sub-minute interval jobs stalling across a DST spring-forward gap with ZoneInfo timezones, relevant ahead of the 2026-10-26 CEST→CET transition (our tzlocal yields ZoneInfo). No API changes; PTB 22.x compatibility range (>=3.10.4,<3.12.0) unaffected.