Symptom
The gateway SIGTERM-restarts on any openclaw.json change, even for keys that don't require a runtime restart. In-flight embedded agent runs are killed mid-turn, producing silent failures downstream (compounds with the ACP silent-failure handler issue).
Evidence
From ~/.openclaw/logs/gateway.log on 2026-04-12:
- 10:36 — SIGTERM restart triggered by
tools.media.audio.maxBytes change
- 10:48 — SIGTERM restart triggered by
agents.list change
- 12:27 — SIGTERM restart triggered by
plugins.entries.google.config.webSearch change
- 09:33:11 — force-kill event with the following log line:
[reload] restart timeout after 300060ms with 3 operation(s), 1 embedded run(s), 3 task run(s) still active; restarting anyway
That single force-kill terminated 3 active task runs, 1 embedded run, and 3 operations mid-flight.
Impact
- In-flight embedded agent turns are killed without notice — produces
ACP_TURN_FAILED downstream (see companion issue)
- User-visible effect: agent goes silent, user has no idea why
- Config reload churn is a normal operational pattern (tuning audio limits, toggling web search, adding agents) — these should not be restart-worthy
- The 300s wait-for-drain is honored in name only: if drain doesn't complete, the gateway restarts anyway and nukes active work
Trigger keys observed (non-critical — should not force restart)
tools.media.audio.maxBytesplugins.entries.google.config.webSearchagents.list (adding/editing an agent config)
Recommendation
Two complementary fixes, ideally both:
(a) Defer restart until in-flight runs complete — replace the "restart anyway" fallback with either:
- Indefinite wait with a warning log every 30s, OR
- Queue the restart and apply it at the next idle window
(b) Exempt non-critical keys from the restart-trigger list — maintain an explicit allowlist of keys that require restart (e.g. listeners.*, providers.*, top-level server.*). Everything else should hot-reload.
The current "any config change = SIGTERM" behavior is too aggressive for a 24/7 control-plane process.
Context
Surfaced during debug session 2026-04-12 investigating why Ranveer (eng-hygiene agent) went silent on coding tasks. Full artifact at .planning/debug/resolved/ranveer-silent-on-coding-tasks.md in the user's openclaw config repo.
Symptom
The gateway SIGTERM-restarts on any
openclaw.jsonchange, even for keys that don't require a runtime restart. In-flight embedded agent runs are killed mid-turn, producing silent failures downstream (compounds with the ACP silent-failure handler issue).Evidence
From
~/.openclaw/logs/gateway.logon 2026-04-12:tools.media.audio.maxByteschangeagents.listchangeplugins.entries.google.config.webSearchchangeThat single force-kill terminated 3 active task runs, 1 embedded run, and 3 operations mid-flight.
Impact
ACP_TURN_FAILEDdownstream (see companion issue)Trigger keys observed (non-critical — should not force restart)
tools.media.audio.maxBytesplugins.entries.google.config.webSearchagents.list(adding/editing an agent config)Recommendation
Two complementary fixes, ideally both:
(a) Defer restart until in-flight runs complete — replace the "restart anyway" fallback with either:
(b) Exempt non-critical keys from the restart-trigger list — maintain an explicit allowlist of keys that require restart (e.g.
listeners.*,providers.*, top-levelserver.*). Everything else should hot-reload.The current "any config change = SIGTERM" behavior is too aggressive for a 24/7 control-plane process.
Context
Surfaced during debug session 2026-04-12 investigating why Ranveer (eng-hygiene agent) went silent on coding tasks. Full artifact at
.planning/debug/resolved/ranveer-silent-on-coding-tasks.mdin the user's openclaw config repo.