Summary

Follow-up for the lock-scope field report in #88168 after landing the focused synthetic-placeholder race fix in #88588.

A user reported that the synthetic tool-result flush path can coincide with a per-session .jsonl.lock file staying write-held by the live gateway process. The visible symptom is subsequent turns failing with session file locked (timeout 60000ms) until the lock expires or the gateway is killed.

Source report: #88168 (comment)

Scope

• Version reported: openclaw 2026.5.27
• Channel: webchat
• Model API family: Anthropic Messages-compatible
• Trigger reported: adjacent long-running exec/process polling turns where synthetic missing-result insertion happened while real tool results were racing
• Impact: session/channel can remain blocked even after the placeholder race itself is fixed

Desired proof

• Reproduce or disprove that synthetic tool-result writes can acquire a session JSONL lock without releasing it.
• Confirm whether fix(agents): avoid synthetic tool results during parallel races (#88168) #88588 eliminates the trigger or only reduces the probability.
• If the lock leak remains possible, route the synthetic write path through the same acquire/release discipline as normal appends, or remove the stale synthetic write path entirely where no longer needed.

Notes

#88588 fixes the central transcript placeholder race by avoiding synthetic insertion at the new assistant tool-call boundary and letting transcript repair move late real results. This issue tracks the separate lock-lifetime claim so #88168 can close for the placeholder behavior without losing the lock report.