midas-harness: Skill for Claude Code

close-sprint (Phase 8 — Per-sprint Audit & Adjust)

Guard + state: <paths.engine>/templates/skill-state-ritual.md (+ AGENTS.md § Safety / Path resolution).

The closing half of the 7 ⇄ 8 loop. After a sprint's code lands, the orchestrator audits the living code against the rules frozen in Phase 5 and the scope in the business case, resolves drift, freezes the verdict, and decides what's next. The producer never grades its own homework — this is an independent orchestrate-tier audit.

vs /start-sprint: start-sprint ran the pre-sprint drift audit and set the sprint active. close-sprint is the formal Phase-8 gate — only after tasks, tests, and UI /midas-verify journeys (web and/or mobile sections in verify-NN.md, when applicable) are green. Never open a sprint with close-sprint; never ship a gate without it.

Precondition. A sprint must be active in paths.state → sprints[] with its work landed (tasks done, tests run). If no active sprint, stop and report.

Shared audit fragments: <paths.engine>/templates/audit-checklists.md (gate semantics, evidence rule, tally format, hygiene hook). Record shape: <paths.engine>/templates/audit-record.md.

Procedure

Full procedure: <paths.engine>/pipeline/8-audit-adjust.md.

Step outline (Steps 0–6):

• Step 0 — Hygiene pass (path-pass, not Skill-tool): Read and follow <paths.engine>/skills/midas-hygiene/SKILL.md (product scope — path-passes sweep product + lean on fat diffs) when brownfield requires a sweep record or the diff is large (greenfield: recommended; document any skip). Do not Skill-tool / auto-slash; power-users may still type /midas-hygiene or internals /midas-sweep / /midas-lean-review.
• Step 0.5 — Diff gate receipts (when production paths changed): require {paths.cache}/gates/<run>/test.json and quality.json with isPassingReceipt semantics (pass or skipped + reason) or record an explicit skip with reason in {runs}/sprints/NN-progress.md or audit notes. Engine-only / docs-only diffs: skip OK. When receipts are missing or stale, path-pass <paths.engine>/skills/midas-diff-gates/SKILL.md and run its receipt commands in this same close run — still not Skill-tool invoke. Does not change verify-NN.md semantics — UI/API proof stays /midas-verify (primary; human-typed when needed).
• Step 1 — Read state + frozen rules: paths.state (honor track: lite — {product}/market.md optional; {product}/business-plan.md still required), active sprint, all effective rules, architecture + idea docs, design system.
• Step 2 — Conformance audit: every rule, pass/fail with on-disk evidence; confirm Context7 doc coverage; trigger every matching playbook. Risk-selected lenses: follow <paths.engine>/pipeline/8-audit-adjust.md § Risk-selected lenses (law: <paths.engine>/rules/verification.md).
• Step 3 — Scope audit: delivered scope vs plan and business-case success metrics; no scope creep, no silent drops.
• Step 4 — Resolve drift: fix now or consciously amend the rule (+ ADR); re-render adapters if amended; nothing left silent.
• Step 5 — Freeze the audit: write {runs}/audits/audit-NN.md with MIDAS_AUDIT_RESULT tally; unresolved=0 required before closing.
• Step 6 — Plan adjustment + update state: set sprint status: done; set the matching {product}/roadmap.md Status cell to done; reconcile sprints[]; select next sprint or declare MVP complete.
• Optional — lifecycle journal: node <paths.scripts>/lifecycle-journal.mjs close_sprint --detail "sprint-NN" (fail-open metrics under {paths.cache}/metrics/lifecycle.jsonl).
• Optional — quality log: node <paths.scripts>/quality-log.mjs audit pass --detail "sprint-NN" (metadata-only JSONL; never secrets).
• Optional — carryover refresh: node <paths.scripts>/carryover-refresh.mjs after sprint done (snapshot → idle). Resume ladder: <paths.engine>/templates/session-resume-precedence.md.
• Optional — capture proposals: after {runs}/sprints/NN-progress.md § Learned is filled, the human may run node <paths.scripts>/capture-candidates.mjs --progress {runs}/sprints/NN-progress.md — propose only; never auto-write rules/playbooks.