$ the-wire · showcase
Gateway PR comparisons go local, auto-reply compaction stops eating turns
By RepoJournal · Filed · About OpenClaw · Composed from the cited sources · methodology
The day's work is mostly repair and prevention: openclaw's gateway stops re-fetching objects during PR branch comparisons, manual /compact stops cancelling active turns, and crabhelm closes three runtime holes that could silently drop messages or credentials.
fix(auto-reply): settle active run before compacting instead of unconditional abort openclaw/openclaw
Manual /compact used to cancel an active turn and silently discard the user's pending answer; an active turn now gets up to 60 seconds to settle first, and if cancellation does go through the status reply asks the user to resend the interrupted request. A rejected cancellation no longer produces that resend warning.
perf(gateway): keep PR branch comparisons local openclaw/openclaw
Session PR comparisons in partial clones could repeatedly hydrate the same missing historical PR head and then start fresh 120-second waits; comparisons now stay local, use an ancestry query instead of counting every ahead commit, and drop the per-diff Git capability probe. A timed-out read leaves the branch visible without statistics or a GitHub Create PR link and logs recovery guidance, while...
fix(runtime): refresh an old token and stop on rejection openclaw/crabhelm
The runtime bridge retried rejected connection tickets indefinitely, could claim work before refreshing an old credential, and could exit while an agent process group was still running. Now a refresh request is remembered once disconnected, token files at least five minutes old are refreshed before new claims, HTTP 401/403 stops reconnects, outstanding process groups are awaited, and resistant ...
fix(aws): runtime messages sent as the socket opens are no longer dropped openclaw/crabhelm
An AWS runtime could send a job claim or heartbeat immediately after WebSocket upgrade and have it dropped, because the socket stayed paused until attachment finished. The authenticated socket now resumes with listeners already installed, pre-attachment buffering is bounded to eight frames or 256 KiB, and buffered frames are replayed once in order; a closed or overflowing replacement is rejecte...
fix(memory): no visible reason when a subagent loses memory access after its parent resets openclaw/openclaw
A subagent that permanently lost memory access after its parent session was reset logged nothing above debug, leaving operators and agents unable to tell why memory tools stopped working or how to fix it. Operators now get one warning per attempt naming the denied session, the stale ancestor, and the repair, with transient verification failures logging at info.
