sprint-superviselisted

# sprint-supervise — Re-entrant Sprint Supervisor Loop Runs inside a `kind=sprint` worker. The sprint frame (roles/frames/sprint.md) defines the loop structure; this skill covers the operational detail: how to decompose, how to write the handoff, and which dimensions to assign to children. ## The 5-Step Loop **Step 1 — Load prior handoffs.** Read the `[handoff]` turns in this thread (kind=event, source_module=arc-sprint rows on the sprint's thread_id, oldest first). On cycle 1 there are none. These are what previous cycles tried and shipped. **Step 2 — Check evidence.** Re-read the sprint's own body (Requirements + Success criteria). Inspect child ledger rows (via `bun ledger.ts list --parent <id>`) for terminal states. Read issue_events for evidence. Do not trust memory — read the ledger. **Step 3 — Complete if done.** If all success criteria are met: ask bookie to `update <id> --state merged --evidence "<summary of what shipped>"`. Then tear down. **Step 4 — Decompose if not done.** Plan the next increment. Decompose into N capability/cell children via `ledger decompose --child '{...}'`. Per-child dimension rules: | Dimension | Rule | |---|---| | **pool** | Blocking HITL presentation → `interactive`. Production task → `ops`. Fix/implement basic functionality → `build`. Improve/optimize → `explore`. | | **agent** | Code work → `developer`. Human gate → `chat`. Ops/infra → `admin`. | | **tier** | Child inherits sprint's tier by default. Escalate only if new risk class