hongming
5196de0836
molecule-core's ephemeral-CP e2e failed with "Scheduler never fired
'e2e-fire-local' within 360s". The daemon this plugin ships (generated from
SDK 0.5.5) delivers each due self-turn synchronously inside the tick loop and
writes its heartbeat only AFTER that loop. A self-turn is a blocking A2A
message/send with a 600s default timeout, so when the agent was settling the
tick never returned:
target health: {"armed": 1, "errors": {}, "last_tick": "21:16:57"} <- probed 21:23:19
sibling health: {"armed": 0, "errors": {}, "last_tick": "21:23:15"} <- still ticking
last_tick froze 2s after arming and never advanced for 6m22s while sibling
daemons kept ticking, and health kept reporting a healthy-looking armed:1 /
errors:{} the whole time. The process was alive-but-hung, so the runtime
supervisor never restarted it: a tenant's schedules stop firing until the
container is replaced.
SDK 0.5.6 fixes this at the SSOT (scaffold commit ab21b58, released via
molecule-ai-sdk PRs #173/#187): delivery is DECOUPLED from ticking. The tick
only enqueues due fires and always advances last_tick; a serial background
worker delivers, with a watchdog that cancels and RE-QUEUES a wedged delivery
so at-least-once holds (the earlier bounded-timeout idea was rejected upstream
because a short deadline cannot tell a hung agent from a healthy long turn,
and advancing on timeout silently drops the fire).
- scheduler.py regenerated from the 0.5.6 scaffold (the only gated file that
changed; trigger_schedule.py and channel_sdk.py are byte-identical).
- ci.yml SDK_VERSION 0.5.5 -> 0.5.6 so the drift gate regenerates from the
version that carries the fix.
- plugin.yaml 0.2.0 -> 0.2.1.
- tests/test_scaffold.py synced from the 0.5.6 scaffold. The instance copy
still exercised `scheduler.run_once`, which no longer exists (it is now
SchedulerDaemon.tick), so it would not even import against the new daemon.
The synced suite carries the regression guards:
test_hung_delivery_does_not_freeze_the_tick_loop,
test_watchdog_requeues_a_wedged_delivery,
test_healthy_long_turn_completes_and_advances_without_error,
test_poke_whose_delivery_fails_is_requeued_not_dropped.
Verified locally against the published 0.5.6 wheel, running this repo's own CI
steps: manifest valid, DRIFT-GATE OK (byte-identical), 16/16 tests pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
41 lines
1.8 KiB
YAML
41 lines
1.8 KiB
YAML
name: ci
|
|
|
|
on: [push, pull_request]
|
|
|
|
permissions:
|
|
contents: read
|
|
|
|
jobs:
|
|
validate:
|
|
runs-on: ubuntu-latest
|
|
steps:
|
|
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
|
|
with:
|
|
persist-credentials: false
|
|
- uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
|
|
with:
|
|
python-version: "3.11"
|
|
# No `cache: pip` — this plugin ships no requirements.txt/pyproject.toml
|
|
# (it vendors channel_sdk and installs the authoring SDK ad-hoc below),
|
|
# and setup-python's pip cache hard-fails without a deps file to hash.
|
|
- name: Install authoring SDK (scaffold SSOT)
|
|
# This plugin is a materialized instance of the SDK trigger scaffold, so
|
|
# `validate` and the drift gate need the SDK. Install it from its git
|
|
# RELEASED-wheel pin: bumped to sdk 0.5.5, which adds the RFC-#8
|
|
# schedule_name outcome correlation to the trigger scaffold (drift gate
|
|
# byte-identical vs the 0.5.5 wheel). Bump SDK_VERSION (and regenerate
|
|
# this plugin) whenever the SDK trigger scaffold changes.
|
|
env:
|
|
SDK_VERSION: 0.5.6
|
|
run: |
|
|
# --extra-index-url (not --index-url): the Gitea registry hosts only
|
|
# molecule packages; deps (pyyaml) must still resolve from PyPI.
|
|
python -m pip install "molecule-ai-sdk==${SDK_VERSION}" \
|
|
--extra-index-url https://git.moleculesai.app/api/packages/molecule-ai/pypi/simple/
|
|
- name: Manifest validation
|
|
run: python -m molecule_plugin validate plugin .
|
|
- name: SSOT drift gate (daemon logic == SDK scaffold)
|
|
run: python scripts/check_scaffold_drift.py
|
|
- name: Daemon unit tests
|
|
run: python -m unittest discover -s tests -v
|