Orchestration MVP

Scope

• Cross‑cutting ACS planner with CEO budgets/risk; no streaming in v0 (batch output).
• Candidate providers: L2 (primary), L4 hints (optional), L1 (expansion, not fallback).

Inputs

• render_request.v0 (CEO→ACS): { version, intent, budgets { tokens_max, time_ms }, risk, privacy_mode, request_id }

Outputs

• render_context_reply.v0 (ACS→CEO): { fragments[], kv_policy, metrics }
• Streaming: disabled in v0 (HCS reserved for future)

Success Criteria

• Quality first, then token budget adherence (0% overshoot), then latency (comfort target)
• Latency: comfort targets only (no hard fail); document observed p95
• Coverage: will be derived from initial data collection (first telemetry pass)

Non‑Goals

• Learned planner (use heuristics v0)
• Full KG traversal (L1 is vector‑first expansion)

Risks

• Provider timeouts; mitigate with per‑provider deadlines and retries before any expansion.
• Fragment inflation; control via LOD + KV policy and strict token budget.

Milestones

• M1: Provider API + simple planner (2–3 days)
• M2: Telemetry (no streaming), quality notes + CLI smoke test (1–2 days)
• M3: Integrate expansion‑to‑L1 path with explicit warnings and retry policy (1 day)