Repository navigation
fix(discord-voice): wake/addressed state stranded on torn-down session after watchdog reset — bot permanently muted despite owner summons #1663
Description
Symptom (live, 2026-06-11 14:38-14:43 UTC, group session w/ 2 humans)
After the watchdog's forced full reset recovered a hung Gemini session (the #1600 M-fix working as designed — hang detected at 65s-since-last-turn, re-handshake in 2s), the bot never spoke again: every model turn after the reset was muted by the speak gate, including direct responses to 20+ owner name-summons and the canonical wake phrase ("Lucy, wake up", 14:39:54). ~450 muted audio chunks over 5+ minutes; Gemini alive and completing turns (turn_101, turn_102, …) the whole time.
Audit evidence (the #1600 observability rows made this diagnosable)
name_gaterows: owner + second human summons all detected —aiAddressed:1, namedDet:1(20+ rows 14:39:06→14:43:24).speak_decisionrows: simultaneous, continuous{"speak":false,"regime":"group","reason":"group-active-unaddressed"}.- Signature: name detection says ADDRESSED, speak gate says UNADDRESSED → the two layers stopped sharing state exactly at the reset boundary (14:38:12 "reconnect — per-turn latches reset").
Root cause (same class as the fixed _recountHumans stale-ref)
The watchdog reset path (handleTransportClose(4002) → session rebuild) leaves the wake/addressed plumbing pointing at the torn-down session's state: summons recorded post-reset never reach the gate the audio-mute path consults (s.gate.lastAddressedToMe / wake-stamp window). The #1600 fix hoisted the human-recount into a module-level ref (_recountHumansRef) for exactly this reason; the gate/wake state needs the same treatment.
Proposed minimal fix
Hoist the gate/wake state wiring to survive resets the way _recountHumansRef does: after the in-place session rebuild, re-bind the STT-side writers (decideForTurn target, wake stamps) and the audio-gate readers to the SAME live object. One mechanism, no behavior change outside the reset path.
Acceptance (live, per the smoke-checklist discipline)
- Trigger or wait for a watchdog reset mid-group-session; owner says the bot's name once afterwards → bot responds audibly.
name_gate.aiAddressed:1rows are followed byspeak_decision speak:true(notgroup-active-unaddressed) within the sticky window.- Existing 419-test suite stays green.
Related: the watchdog rework + audit rows on PR #1427 (this bug is only visible because of them); same-class precedent: the _recountHumansRef fix (#1600 count-drift).
— filed by Lucy (Mac Studio bot), diagnosed live with Susan + Chi in-session
Activity
liususan091219 commentedon Jun 11, 2026
Root-cause correction after implementing: the muting was NOT reset-stranded state. The
decideForTurn/lastAddressedToMewriter block was controller-only (if (VOICE_CONTROLLER && userId === VOICE_CONTROLLER …)), and this deployment runs legacy mode — so the group-regime speak gate'saddressedToMeinput was never written by ANY summons, from anyone, since the group gate landed. Pre-reset audibility flowed through mode-switch ack / talk-window escapes (allowAckAudible,_forceAudibleUntil) — which is also why there are zerospeak_decisionrows before the reset (those escapes return before the legacy decision branch runs) and continuous refusal rows after.Fix (commit on the test branch): add the legacy arm to the gate-writer — owner-tier speakers' per-user STT streams feed
decideForTurn— mirroring the_aiWakeAtlegacy-arm fix from the same file (the comment there literally warned about this class: "Without the legacy arm both stamps sat in controller-gated blocks"). 115/115 discord-voice tests green; live re-test running now.— Lucy (Mac Studio bot)
john-the-dev commentedon Jun 28, 2026
The code containing this bug (
skills/discord-voice/scripts/discord-voice-server.ts— where the VOICE_CONTROLLER-gatedaddressedToMewriter lived) was deleted in PR #1720 (refactor(voice): discord-voice deletion + voice-core extraction, merged). The skill is now empty in this repo (scripts/contains only__pycache__).If the discord-voice feature continues in an external voice-core repo, the root-cause fix from liususan's analysis (remove the VOICE_CONTROLLER guard on the
addressedToMewriter so legacy-mode deployments write the signal too) should be applied there. From this repo's perspective, the affected code no longer exists.Recommend closing as resolved-by-removal (or transferring to the voice-core external repo if the feature lives there).
john-the-dev commentedon Jun 28, 2026
Thanks for the correction, Lucy. My previous comment was premature — I should have verified the code state more carefully before assuming the file was fully deleted.
If your fix passes 115/115 discord-voice tests and the live re-test confirms the speak gate works in legacy mode, this is worth a PR. The root cause you've identified (legacy arm missing from the gate-writer block) is a real gap —
decideForTurn/lastAddressedToMesilently receives no input on non-controller deployments, which exactly matches the symptom.When you open the PR, I'll review it.
bassilkhilo-ag2 commentedon Jun 30, 2026
Confirming john-the-dev's 2026-06-22 analysis: the VOICE_CONTROLLER-gated
addressedToMewriter that caused this bug was deleted in PR #1720 (refactor(voice): discord-voice deletion + voice-core extraction). Thescripts/directory inskills/discord-voice/now contains only__pycache__— no live code to patch.Lucy identified the precise root cause (legacy arm missing from the gate-writer, mirroring the
_aiWakeAtlegacy-arm fix from the same file), and reported 115/115 tests green on a test branch. If the discord-voice feature continues in an external voice-core repo, that fix should be applied there.From this repo's perspective the affected code is gone. Recommend closing as resolved-by-removal.
bassilkhilo-ag2 commentedon Jun 30, 2026
Code deleted in PR #1720 — closing as resolved-by-removal. Lucy's root-cause analysis + fix applies to the external voice-core repo if the feature continues there.
john-the-dev commentedon Jul 7, 2026
Closing per bassilkhilo-ag2's last comment (code deleted in PR #1720 — resolved by removal).
docs(REVIEW.md): ground lessons 2/4/5 in the real reported breakages
feat(review): repo-scoped REVIEW_GUIDE drives a generic checks runner…