Build a self-hosted AI meeting delegate with two modes:

• Ghost Mode: listen + transcribe + summarize
• Agent Mode: respond in-meeting with policy guardrails

What

• Finalize architecture, scope, and mode split
• Define out-of-scope for v0.1

DoD

• requirements.md updated and approved
• This 10-step plan documented

What

• Chrome extension captures tab audio
• Streams chunks to local backend via WebSocket
• Backend stores chunks and exposes basic stats

DoD

• Start capture from extension popup
• backend receives chunks (/ws/audio) and writes .webm
• GET /api/sessions shows received bytes/chunks

What

• Convert incoming audio chunks to transcribable frames
• Run faster-whisper incremental transcription

DoD

• transcript lines appear in backend logs/API in near-real-time
• configurable model (tiny/base/small)

What

• Side panel for live transcript view
• Speaker segments + timestamps

DoD

• captions update in UI while meeting audio is streaming

What

• Post-meeting summary generation
• Decisions / Action items / Follow-ups format

DoD

• summary generated from transcript via API
• deterministic output schema

What

• Pre-meeting context form:
  • meeting topic
  • user role
  • talking points
  • constraints

DoD

• summary/responses incorporate briefing context

What

• Build response policy layer:
  • when to respond
  • safe fallback lines
  • no unauthorized commitments

DoD

• policy tests pass on sample scenarios
• generated responses include confidence/fallback when uncertain

What

• TTS + local virtual audio routing path
• optional (off by default)

DoD

• one-click “speak response” works on local setup guide

What

• one-command backend run
• extension load instructions
• sample config and troubleshooting

DoD

• fresh machine setup in under 10 minutes

What

• produce polished demo video
• README narrative + architecture diagram + examples

DoD

• public repo ready
• demo covers Ghost mode end-to-end

• ✅ Do only Step 1 and Step 2
• ❌ Do not implement STT/summary yet