Completion audit

/completion-audit — Claimed vs Real

"Marked Done" is a claim, not a fact. This audit treats every completion claim as false until an executed check fails to break it, then prices the difference between claimed and real as a remediation plan. Right-sized is the target: happy-path-only work is incomplete; gold-plating that misses the actual need is also incomplete.

Report format, ruling grammar (CONFIRMED/REFUTED/UNVERIFIED), severity tiers, JSON trailer, degradation rules, and routing table: see _shared/audit-report-contract.md. Task state via the task-manager adapter ({task_prefix}-###); decisions via the storage adapter.

Invocation

/completion-audit                     # audit this session's claimed work
/completion-audit PR #42              # audit a PR's claims
/completion-audit --task {task_prefix}-123   # single-task scope = validate-completion's job

--task <id> narrows scope to one claim and runs only steps 2–4; that scoped mode IS /validate-completion (which is a thin alias for it).

Procedure

1. Collect the claims. Session summary, PR description, task statuses via the adapter. One falsifiable sentence per claim — vague claims get rewritten before they get audited.
2. Cross-check every planning source. Plan doc, tracker, spec files, original prompt/brief, roadmap, decision log. A claim contradicted by any source is a finding; contradictions BETWEEN sources are findings too.
3. Rule on each claim via /refute's procedure. Construct the break most likely to succeed and run it. Delegate narrow checks where they belong (per-claim execution → {agents.validate_completion}; spec conformance → {agents.spec_audit}; rule breaches → {agents.rules_audit}; over-engineering → {agents.pragmatism_audit}) and reconcile their trailers — a delegate fail can't sit under an overall pass.
4. Catch the classic false-completes: works only under ideal conditions; code that exists but is never wired/reachable end-to-end; missing error handling that makes it unusable; over-abstraction posing as done work; fragility posing as MVP scope; missing basics excused as "design choice".
5. Price the gap. For each REFUTED/UNVERIFIED claim: severity tier, file:line, evidence, and a remediation item with a testable definition of done, dependency order, and honest effort framing.
6. Report. What actually works (plainly), the gap list (most severe first), the remediation plan, process recommendations to prevent recurrence, the JSON trailer, and the verdict line logged via /verdict:

AUDIT: completion-audit | <scope> | <pass|fail|inconclusive> | <blocking>/<total> findings

Rules

• Never average: one S0 finding means the session is not done, regardless of how many claims CONFIRMED.
• "Couldn't run it" is UNVERIFIED, reported as such — not a pass.
• The remediation plan must be executable by someone with no session context.

Judgment weave

• Each claim ruling runs /refute (this skill is /refute applied at scale).
• Gaps found repeatedly across sessions → author a hook or checklist line per judgment/MANUAL.md §4; a lesson not banked evaporates.
• Verdict line → /verdict ledger; scoring later via /calibrate.
• Borderline pass/fail calls → /escalate, don't round in your own favor.

Composes with

• /validate-completion — the --task scope of this skill, as its own entry point.
• /spec-audit, /rules-audit, /pragmatism-audit — delegated narrow checks (routing table in the shared contract).
• /epic close, /milestone close, /ccb — invoke this before accepting Done.
• claim-evidence hook — mechanically demands the evidence this audit produces.