OpenClaw merged PR #97164, "fix: prevent daily rollover splitting active transcripts", a P1 session-state fix for active agent runs that cross a stale-session reset boundary.
The bug affected a narrow but painful timing window. If an inbound message arrived just after a configured daily or idle reset boundary while the previous turn was still running, stale-session rollover could archive the active transcript before reply-turn admission had a chance to wait, queue, or steer correctly.
The active writer could then recreate the old transcript path, while later output landed elsewhere. The result was a split transcript and potentially orphaned output from a run that was still in progress.
What Changed
The new logic defers implicit daily or idle rollover when the reply run registry shows the same session key and session id already has a non-queued active operation.
That scope is intentionally tight. Normal stale-session cleanup still works when no run owns the session. Queued pre-dispatch reservations still allow ordinary rollover, and if rollover creates a new session, the queued reservation is rebound to the new session id before the run starts.
Explicit /new and /reset behavior is unchanged.
Why It Matters
OpenClaw transcripts are not just logs. They are the durable record that helps agents resume context, operators audit behavior, and channel replies stay explainable after a long-running turn.
When a transcript splits mid-run, the user-facing symptom can be confusing: output appears incomplete, later tool activity seems detached from the initiating prompt, or follow-up context lands in the wrong place. For long-running automated tasks, that can make a successful run look lost.
PR #97164 makes reset-boundary behavior respect active ownership. If a run is still writing to a session, rollover waits instead of archiving the ground underneath it.
Active Runs Versus Queued Work
The distinction between running work and queued pre-dispatch work is important. The patch does not freeze all rollover behavior just because a future run has been reserved.
Instead, it checks whether the same session has a non-queued active operation. That gives OpenClaw the behavior operators actually want:
- Do not split the transcript of a running turn.
- Do keep ordinary stale-session cleanup available when the current inbound turn has not started yet.
- Do rebind queued work to the new session id when rollover legitimately creates one.
That keeps the stale-session feature useful without letting it interrupt live output.
Validation
The PR includes focused tests for session handling, media-only reply runs, reply-turn admission, Gateway session reset hooks, and Gateway reset cleanup.
It also includes current-commit Crabbox proof using a stale daily-reset session with a running replyRunRegistry operation for the same session id. The observed result was isNewSession:false, resetTriggered:false, no archive files, and all transcript markers staying in one file.
An autoreview pass initially found issues around queued pre-dispatch reservations, which were fixed before merge. The final autoreview run exited clean.
For OpenClaw users, this is a reliability polish with real operational weight. Long-running turns can cross daily or idle boundaries without having their transcript split out from under them.