---
gsd_state_version: 1.0
milestone: v1.6
milestone_name: Voice Pipeline + Minimal Message Bridge
status: executing
stopped_at: Completed 37-03-PLAN.md — Voice Output Components (ChatVoicePlayer, ChatVoiceBadge, VoiceModeToggle)
last_updated: "2026-04-04T02:52:59.506Z"
last_activity: 2026-04-04
progress:
  total_phases: 4
  completed_phases: 2
  total_plans: 7
  completed_plans: 7
  percent: 0
---

# Project State

## Project Reference

See: .planning/PROJECT.md (updated 2026-04-03)

**Core value:** A fresh onboard asks for ONE thing (root directory), auto-creates PM + Engineer agents, and drops you in the dashboard.
**Current focus:** Phase 37 — web-chat-voice-ui

## Current Position

Phase: 38
Plan: Not started
Status: Ready to execute
Last activity: 2026-04-04

Progress: [░░░░░░░░░░] 0%

## Performance Metrics

**Velocity:**

- Total plans completed: 0 (v1.6)
- Average duration: -
- Total execution time: 0 hours

## Accumulated Context

### Decisions

Decisions are logged in PROJECT.md Key Decisions table.
Key constraints for v1.6:

- voicePipelineService is the keystone — Phase 37 and Phase 38 both depend on it; build Phase 36 first
- Telegram bridge uses long polling (grammY `bot.start()`) — no public HTTPS required on Mac Mini
- Audio transcoding via ffmpeg-static ^5.2.0 — NOT archived fluent-ffmpeg (archived May 2025)
- Voice mode flag must survive every pipeline layer: client → Express → message persistence → agent codec
- COOP/COEP headers required for @ricky0123/vad-react SharedArrayBuffer (add to Express static middleware)
- Phase 37 and Phase 38 are independent once Phase 36 ships; sequential ordering for single-developer delivery
- Telegram bridge must stay under 500 lines (TGRAM-06 is a hard constraint)
- [Phase 36]: Export nexusSettingsSchema for direct testing, use nexusSettingsSchema.parse({}) for consistent defaults in catch blocks
- [Phase 36]: Used manual execFileAsync wrapper instead of promisify(execFileCb) to avoid util.promisify.custom symbol incompatibility with vitest mocks
- [Phase 36]: Voice routes are dedicated voice.ts module (not added to chat-files.ts) for clean separation — voice pipeline is its own subsystem
- [Phase 36]: voiceMode typed as text|voice_input|full_voice union in stream endpoint, persisted as voice_full/voice_input messageType for downstream rendering
- [Phase 37]: Cherry-picked Phase 36 commits to bring voice pipeline, nexus-settings, and voiceMode wiring to phase-37 branch
- [Phase 37]: COOP/COEP headers placed as first Express middleware — applies to all responses including API, static, and Vite dev
- [Phase 37]: VAD ONNX assets served from ui/public/ same-origin to avoid COEP blocking CDN-served binary files
- [Phase 37]: useVadRecorder requests separate MediaStream ref for VoiceWaveform AnalyserNode — useMicVAD manages its own stream internally
- [Phase 37]: AudioContext not closed on cleanup in VoiceWaveform — reused across recording cycles to avoid repeated autoplay unlock prompts
- [Phase 37]: useVoiceMode hook created in plan 37-03 to unblock VoiceModeToggle during parallel execution
- [Phase 37]: Auto-play preference stored in localStorage (nexus:voice:autoplay), not nexus-settings — avoids server round-trip for fast UX

### Pending Todos

None yet.

### Blockers/Concerns

- [v1.5 carryover] smart-whisper Apple Silicon acceleration unverified on Mac Mini M4 — fall back to `tiny.en` if `base.en` acceleration not confirmed
- [v1.6] grammY session management approach not yet chosen: lightweight `Map<chatId, sessionId>` vs. grammY conversation plugin — decide at Phase 38 planning
- [v1.6] Dual output prompt reliability on 7B models is ~90% — Approach B fallback (post-process markdown strip) must be implemented as safety net, not optional

## Session Continuity

Last session: 2026-04-04T02:38:29.163Z
Stopped at: Completed 37-03-PLAN.md — Voice Output Components (ChatVoicePlayer, ChatVoiceBadge, VoiceModeToggle)
Resume file: None