Summary

OpenClaw can retain very old running task rows for days. openclaw tasks maintenance --apply prunes task-flow records, but does not mark these stale running tasks lost/cancelled. In this state the gateway became intermittently non-responsive and required openclaw doctor --fix to recover the control path.

Observed environment

• OpenClaw: 2026.5.2 (8b2a6e5)
• macOS: 26.4.1 arm64
• Node: 25.8.1
• Gateway LaunchAgent: ai.openclaw.gateway

Evidence

openclaw status showed:

• Tasks: 4 active · 0 queued · 4 running · 67 issues
• repeated gateway liveness warnings: event_loop_delay,event_loop_utilization,cpu
• websocket handshake timeouts / connectivity probe timeouts while PID was still listening on port 18789

openclaw tasks maintenance --apply initially reported task-flow pruning, but still left stale running tasks:

Tasks maintenance (applied): tasks 0 reconcile · 0 recovered · 0 cleanup stamp · 0 prune; task-flows 0 reconcile · 65 prune
Tasks health after apply: 0 queued · 3 running · 3 audit errors · 25 audit warnings

The remaining running tasks were days old:

9906af0e-df96-4840-921f-42ced79e82e7  created 2026-04-29 21:53  owner agent:main:web_b015a645b1_47aed28cc700
5e098c13-da22-45b8-91f9-4aa9e88c271e  created 2026-05-01 22:22  owner agent:main:subagent:771e97dc-c137-4af7-9e9b-570e40b8225c
b0fa03c6-460f-4bc6-9d4f-1240f231cc1d  created 2026-05-01 22:22  owner agent:main:subagent:22a5d09c-f5da-474f-a5c3-792f6d48026a

Manual cancellation cleared the task audit errors:

openclaw tasks cancel 9906af0e-df96-4840-921f-42ced79e82e7
openclaw tasks cancel 5e098c13-da22-45b8-91f9-4aa9e88c271e
openclaw tasks cancel b0fa03c6-460f-4bc6-9d4f-1240f231cc1d

Tasks health after apply: 0 queued · 0 running · 0 audit errors · 25 audit warnings

Expected behavior

Task maintenance should detect impossible/stale running tasks, probably by age + missing live run/session/process signal, then mark them lost, timed_out, or cancelled without requiring manual lookup/cancel.

A conservative default would be to treat running tasks older than 24h as stale unless there is clear evidence of active work.

Actual behavior

Maintenance pruned task-flow records but left multi-day running task rows untouched. These showed as audit errors until manually cancelled.

Suggested fix

• Extend openclaw tasks maintenance --apply to reconcile stale running tasks.
• Add an age threshold or config knob, e.g. --max-running-age 24h / default 24h.
• Include a dry-run line that lists stale running tasks that would be marked lost/cancelled.
• Consider having doctor --fix invoke the same stale-running-task reconciliation.

Related liveness symptoms

During the stale-task period the gateway repeatedly hit event-loop liveness warnings and websocket timeouts. A macOS sample during high CPU showed the gateway hot in V8 microtasks/JSON parsing, but the precise JS caller was not available without a Node CPU profile. This issue is specifically about the stale task cleanup gap, not claiming stale tasks are the sole CPU root cause.