openclaw - 💡(How to fix) Fix Gateway heap OOM: cron runtime contexts retain sessions.json and skillsSnapshot.prompt [1 pull requests]

Observed on OpenClaw 2026.5.7 during gateway uptime on 2026-05-21.

Problem

• The gateway crashed with V8 heap exhaustion, not a kernel OOM.
• Heap snapshots showed ~2.0 GiB retained by ~113k copies of the skill prompt string.
• The retainer chain runs through skillsSnapshot.prompt -> skillsSnapshot -> session entry -> session store -> cronSession.store -> cron run context.
• The on-disk sessions.json footprint was much smaller (~4.7 MiB of prompt text in the main store), so this is runtime retention, not just durable store size.

What existing maintenance already does

• OpenClaw session maintenance and the archive job help with old session files and sessions.json pruning.
• That is useful for disk pressure, but it does not release the live cron/subagent execution context that keeps the store graph reachable.

Recommended fix

• Add a shallow completion finalizer for isolated cron/subagent runs.
• Release cronSession.store after the final durable write.
• Clear the run context from agent-events after completion.
• Keep the finalizer shallow and O(1), with no deep traversal and no hot-path disk rewrite.
• Keep backward compatibility for existing sessions.json entries.

Why this matters

• The failure recurred twice in one day, so disk cleanup cadence alone is not enough.
• The right fix is runtime lifecycle cleanup, which should reduce GC pressure and avoid latency spikes.

Validation performed locally

• Focused tests passed in the isolated worktree for the cleanup path and existing model-switch regression coverage.