openclaw - 💡(How to fix) Fix Subagent sessions not cleaned up after archiveAfterMinutes (orphaned sessions accumulate) [1 participants]

Bug Report: Subagent sessions not cleaned up after archiveAfterMinutes

Issue Summary

Subagent sessions are not being cleaned up after archiveAfterMinutes: 60 (1 hour) as configured. Sessions persist for days, accumulating in agents/*/sessions/ directories. The sessions-archived/ folder remains empty, indicating archival never completed successfully.

Config:

"subagents": {
  "archiveAfterMinutes": 60,
  "maxChildrenPerAgent": 10,
  "maxSpawnDepth": 2,
  "maxConcurrent": 8
}

Reproduction Data

• Environment: Self-hosted OpenClaw on Linux
• First observed: Sessions from 2026-04-17 still present on 2026-04-25 (10+ days old)
• Current session count: 47 .jsonl files in agents/main/sessions/
• Archive folder: sessions-archived/ is empty
• Child session refs: Parent agent shows 15 child session references despite maxChildrenPerAgent: 10

How archiveAfterMinutes Is Supposed to Work

Timer-based sweep in subagent-registry.ts via sweepSubagentRuns(). Deletes from runs.json, calls sessions.delete RPC on sessions.json. .jsonl files are separate.

Confirmed Bug Breakdown

Bug #49000 — Silent RPC failure (root cause for April 17 sessions) sessions.delete is wrapped in try/catch with no retry. If the gateway is unavailable during the sweep window (restart, timeout), the sessions.json entry persists forever with no reconciliation. Archival begins (runs.json deleted) but sessions.json RPC silently fails → sessions survive but are never moved to archive.

Bug #52327 — Three structural gaps

1. In-memory timers lost on gateway restart — if a restart occurs within the 60-min window, those sessions are permanently orphaned
2. mode: "session" thread-bound sessions are entirely excluded from archival — they never receive an archive timestamp
3. No restart-safe catch-up mechanism (unlike cron.sessionRetention which runs on startup)

Bug #9634 — Failed sessions not cleaned Even with cleanup: "delete", failed subagent sessions are not removed from sessions.json.

Bug #58429 — Config workaround is blunt session.maintenance.pruneAfter: "7d" exists as a workaround but has no granularity and is not documented as a mitigation for this specific issue.

Workarounds (documented)

1. Direct sessions.json editing — backup first, then remove orphaned entries
2. Add session.maintenance.pruneAfter: "7d" to config — blunt instrument per #58429
3. Gateway restart — resets in-memory timers but does not fix existing orphans

Expected Behavior

Sessions should be archived (moved to sessions-archived/) after archiveAfterMinutes of inactivity, and orphaned entries in sessions.json should not persist indefinitely.

References

• #49000 (root cause: silent RPC swallow)
• #52327 (3 structural gaps, production incident)
• #9634 (failed sessions not cleaned)
• #58429 (pruneAfter workaround)

Reported via OpenClaw agent with session archival research. Full research notes available on request.