-
released this
2026-08-04 22:36:00 +00:00 | 0 commits to main since this releasev0.5 Live Assist — On-the-Job Voice Companion — Milestone Release
Summary
- Milestone: v0.5 (Live Assist — on-the-job voice companion)
- Type: Feature milestone — final patch v0.1.13 IS the milestone release
- Phases: P0 (v0.1.10) → P1 (v0.1.11) → P2 (v0.1.12) → P3 (v0.1.13)
What's New
v0.5 adds the companion surface to Praxis: a hands-free voice assistant a learner invokes while actually working on the job, context-aware of their current scenario/skill path, coaching in real time without doing the job for them. The key distinction from the practice surface (v0.1-v0.4): in v0.5, the learner is on a real call with a real customer and the AI is in their ear.
Phases
Phase 0 — Pre-Execution (v0.1.10)
SPECIFY → CLARIFY → RESEARCH → IDEATE → PLAN → GRILL. 16 active REQs activated (3 ASSIST + 4 NFR + 9 IDEATE). 760-line research, 1075-line plan, 627-line grill (Proceed-with-conditions, 2 MUSTs, 1 escalation).
Phase 1 — Assist Core + Guardrail (v0.1.11)
8 slices, 4 waves, 24 tasks. The safety-critical, on-voice-path surface:
- server/assist/ — full assist module (session, context, lifecycle, routes, pipeline, webrtc, consent, pii_policy, mode_conflict, guardrail_processor)
- server/guardrails/live_assist.py — 3-layer LiveAssistGuardrail (prompt + regex filter + audit log)
- client/src/AssistControl.tsx — tap-to-talk client control
- 92 new tests (409 total). G-049 + G-067 MUSTs resolved.
Phase 2 — Integration + Tech-Debt + NFR (v0.1.12)
4 slices, 2 waves, 9 tasks:
- NFR measurement (p95 latency ≤650ms pilot, guardrail FP<5% / FN measured + trended nightly)
- Cohort aggregation assist metrics (5 new metrics, no schema change)
- Assist cost tracking + C-3 budget check
- 8 v0.4 P1+ tech-debt findings addressed (argon2id offload, cookie-secret, credential enum, f-string SQL, cache persistence, zoneinfo, audit log, 429 mock)
- 60 new tests (469 total)
Phase 3 — Final Review + Ship (v0.1.13)
Multi-persona review (APPROVE_WITH_NOTES, 1 P0 fixed — guardrail processor must buffer LLM text before TTS for REQ-ASSIST-03 safety). Audit (NEEDS_ATTENTION — 12 stale-status fields advanced, no critical issues). 8 P1+ flagged for v0.6.
Requirements Covered (16/16)
- REQ-ASSIST-01: Hands-free voice companion (tap-to-talk in v0.5; wake-word deferred to v0.6 per D-071)
- REQ-ASSIST-02: Context-aware (path week + scenario tag → ≤150-token prompt)
- REQ-ASSIST-03: Guardrails (3-layer: prompt + regex filter + audit log; FN 13.3% ≤ 20% pilot threshold)
- REQ-NFR-ASSIST-01: Latency <600ms target, ≤650ms pilot (D-072)
- REQ-NFR-ASSIST-02: Tap-to-talk on existing React-Web client
- REQ-NFR-ASSIST-03: Guardrail enforcement (FP 0%, FN 13.3%, nightly trending)
- REQ-NFR-ASSIST-04: Shift-bounded sessions, session_type=assist, no mastery update (D-063)
- REQ-IDEATE-01..09: Guardrail tuning corpus, in-loop processor test, mode-conflict, measurable NFRs, PII policy, tech-debt wave, cost tracking, WebRTC reconnect, incremental audit log
v0.6 Backlog (4 ideation-derived requirements)
- REQ-IDEATE-10: LLM-as-judge guardrail evaluation (nightly)
- REQ-IDEATE-11: Assist-weaning metric (mastery signal)
- REQ-IDEATE-12: Offline assist degraded mode
- REQ-IDEATE-13: Voice-only context declaration
Open Escalation
ESCALATION-01 (PIPEDA consent-law review): The ambient mic captures the real customer (a third party). Canada PIPEDA + provincial consent law review needed before assist surface go-live. The consent disclosure (D-070) is implemented regardless. Human legal review required.
Tests
469 passed, 45 skipped, 0 failed
Downloads