{"content":"IDBots dev journal — fix/online-chat-message-order @ 99335867c (2/2)\n\nfix: cancel un-broadcast wait notices and retract the promise on silent turns. Two send-path defects in the private-chat skill wait-notice lifecycle, both reproduced on session c3171c2c:\n\n(1) Stale-notice race — the notice pipeline (LLM ack generation + chain broadcast) runs concurrently with the skill turn and can outlive a fast turn, so the ack broadcast AFTER the turn's real reply. The bridge now hands the daemon a SkillExecutionStartControls cancel flag, flipped the moment the turn settles (complete) or terminates (error/stop; watchdog timeout deliberately does NOT cancel since the turn keeps running in the runner). The daemon checks the flag after generation and before createSimpleMsgPin — the point of no return. The finish path still joins the notice promise so a broadcast already in flight keeps its notice-before-reply chain order.\n\n(2) Dangling promise on silence — when a notice WAS broadcast and the turn settles silent ([NO_REPLY] sentinel / echo-loop guard), the notice became the bot's last word and the peer waited forever (the 02:10 incident). The silent settlement now sends exactly one wake-compatible retraction in the notice's own language (privateChatWaitNoticeRetraction metadata), before the wake arm so the re-drive context carries the closure. The retraction is excluded from echo-loop counting and from extractFinalAssistantReply so a later turn never delivers it as a reply; the notice prompt also stops promising timeframes like 'in a few minutes'.\n\nTests: bridge cancellation contract (5 cases, fake runner), retraction integration via the wake harness (2 cases: exactly one retraction across silent wake re-drives; none when a reply closes the loop), and the previously-orphaned privateChatNoReplyWake suite registered into test:release-regressions-extra. Gates: compile:electron + root tsc clean; full chain 157+68+22 = 247 tests, 0 failed.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}