Implementation Plan
Audience: an AI coding agent plus one human operator.
Companions: Contributing for Agents · Contracts · Progress Tracker.
Source: docs/IMPLEMENTATION_PLAN.md.
Work milestones in order. Do not start M(n+1) until M(n) exit criteria are checked. Each milestone has Goal, Why here, Tasks, Deliverables, Tests, UAT, Exit criteria, Risks, Cursor prompt.
Effort is ideal focused hours, not calendar time (~6–10 ideal hours/week solo evenings).
Architecture note (2026-08-03): Speech compute prefers AO-advertised sidecars via Reach SpeechClient. Prefer tracker + runbook where they diverge from older milestone text.
| # | Name | Effort | Ships |
|---|---|---|---|
| M0 | Ground truth | 6h | Verified facts about real services, no product code |
| M1 | Skeleton & config | 8h (+3h DEV_LOOP) | Three processes start, log, exit cleanly |
| M2 | Vision client | 10h | CPAI client + poll loop, mocked & tested |
| M3 | Attention state machine | 12h (+4h console) | Pure logic, exhaustively tested |
| M4 | Audio pipeline | 14h | Wake + VAD + capture on real hardware |
| M5 | AO Reach session | 12h | Identity-mapped sessions, text turn E2E |
| M6 | Voice round trip | 10h | Speak → transcript → AO → spoken answer |
| M7 | Avatar & kiosk | 14h | SVG/TalkingHead driven by bridge |
| M8 | First contact | 10h | Full demo under 15s |
| M9 | Hardening & soak | 16h | Reconnects, degradation, 24h soak |
Total ≈ 112 ideal hours.
Per-milestone detail: Milestones.
Walk into the room, be greeted by name within two seconds, ask a question in plain speech without touching anything, and get a spoken answer in under fifteen seconds — reliably, for a week, with the public internet unplugged.