Three OpenClaw Fixes Stop Maintenance Windows From Destroying Live Sessions

Three repairs merged to main on August 21, 2026 share a signature: a routine, well-intended maintenance action — compacting a long session, repairing a stuck lane, honoring /restart — damaged the very conversations it was supposed to protect.

Compaction dropped genuine user turns

PR #127563 closes several ways a real conversation turn could silently vanish when a long session compacted. The reported triggers are uncomfortably ordinary: repeating an instruction after an assistant reply, attaching different images, videos, or documents that shared a caption, or referencing case-sensitive paths. Out-of-order timestamps could additionally discard historical turns.

The failure class matters because compaction is invisible when it works. An operator who repeats an instruction and sees it answered assumes continuity; if that turn was dropped from the preserved transcript, later behavior degrades without any error to investigate.

The merged fix preserves real user turns through compaction regardless of those trigger patterns.

Interpretation: treat any pre-fix long-running session as potentially missing user input near its compaction boundaries. Upgrading restores correct behavior going forward; it cannot recover what earlier compactions discarded.

Operator actions

1. For long-lived agents, spot-check whether instructions given twice, media with reused captions, or mixed-case paths produced the behavior you expected on affected builds.
2. After upgrading, re-run a representative long session with harmless markers before and after a compaction event and confirm the markers survive.

The watchdog killed live work during legitimate flushes

PR #127510 fixes a pair of losses around session-lane repair. A user waiting on a reply could lose the active turn or become permanently stuck when another message arrived during a legitimately long compaction or memory flush — the diagnostic watchdog treated the slow maintenance as a hang and repaired over it. Separately, recovery could silently forget queued user work after releasing a stale session lane.

The merged fix keeps queued replies alive through watchdog repair and stops the watchdog from firing on legitimate maintenance windows.

Operator action: if users of an affected deployment report messages that “never got answered” or lanes that stuck until restart, treat those reports as plausible casualties of this bug rather than user error, and re-examine spool state after upgrading rather than assuming the work vanished legitimately.

/restart looped instead of restarting once

PR #127179 fixes the sharpest of the three: sending /restart in a chat channel produced an unbounded restart loop — the gateway restarted roughly every thirty seconds and did not settle — instead of exactly one restart.

The PR notes the diagnosis and production evidence were reviewed by a human before filing; it carried a needs-proof label at filing time and subsequently merged, which places the final verification with upstream maintainers. OpenClaw Academy reviewed the official merged record.

Interpretation: an operator who used /restart to apply a config change on an affected build may have experienced a restart storm instead — and blamed flapping infrastructure rather than the command. Any historical “gateway kept bouncing after I told it to restart” incident deserves a second look.

Operator actions

1. Prefer gateway-native restart mechanisms over chat-triggered ones until your build verifiably contains this fix.
2. After upgrading, exercise /restart once in a disposable channel and confirm exactly one restart cycle completes.

The common lesson

All three bugs converted protective machinery — compaction, watchdog repair, restart handling — into a destruction or looping trigger at one edge case. None requires new configuration; all three change what you should assume about incidents on pre-fix builds. When maintenance and recovery code is involved, “it worked until someone did X during Y” is the pattern to hunt, and upgrades answer it going forward but not backward.