fix(discord-voice): wake/addressed state stranded on torn-down session after watchdog reset — bot permanently muted despite owner summons

[liususan091219]

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_gate rows: owner + second human summons all detected — aiAddressed:1, namedDet:1 (20+ rows 14:39:06→14:43:24).
• speak_decision rows: 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)

1. Trigger or wait for a watchdog reset mid-group-session; owner says the bot's name once afterwards → bot responds audibly.
2. name_gate.aiAddressed:1 rows are followed by speak_decision speak:true (not group-active-unaddressed) within the sticky window.
3. 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

[liususan091219]

Root-cause correction after implementing: the muting was NOT reset-stranded state. The decideForTurn/lastAddressedToMe writer block was controller-only (if (VOICE_CONTROLLER && userId === VOICE_CONTROLLER …)), and this deployment runs legacy mode — so the group-regime speak gate's addressedToMe input 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 zero speak_decision rows 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 _aiWakeAt legacy-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]

The code containing this bug (skills/discord-voice/scripts/discord-voice-server.ts — where the VOICE_CONTROLLER-gated addressedToMe writer 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 addressedToMe writer 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]

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/lastAddressedToMe silently receives no input on non-controller deployments, which exactly matches the symptom.

When you open the PR, I'll review it.

[bassilkhilo-ag2]

Confirming john-the-dev's 2026-06-22 analysis: the VOICE_CONTROLLER-gated addressedToMe writer that caused this bug was deleted in PR #1720 (refactor(voice): discord-voice deletion + voice-core extraction). The scripts/ directory in skills/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 _aiWakeAt legacy-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]

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.