acceptance-runner

Files (1)

• SKILL.md

SKILL.md (first lines)

# Acceptance Runner

Stage 6 (verification) owner — the gate that turns "the builder says it's
done" into "it is provably done". Method references: Bab 12 (Writer/Reviewer
pattern), Bab 17 #4 (TRUST-THEN-VERIFY GAP), plus the community-verified
Iron Law from obra/superpowers `verification-before-completion`.

## The Iron Law

```
NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE
```

If you did not run the verification command **in this session**, you cannot
claim it passes. Reporting a verdict without evidence is lying, not verifying.

## Input

- `planning/sprints/sprint-XXX/ACCEPTANCE.MD` (the contract)
- `BLUEPRINT.MD` (for scope) + builder report (claims to check, not trust)
- Access to run tests/build/lint in the project

## Workflow

### Step 1: Load Fresh

Confirm you are NOT the session that wrote the code. If you are, hand this
skill to a fresh context (new session / subagent). A reviewer that inherits
the writer's context reviews its own memories, not the code.

### Step 2: Map Every Criterion to a Command

For each ACCEPTANCE.MD item, identify the command that proves it and the
expected observable output. Criteria that cannot be mapped → record as
`UNTESTED` and report them back to `/sprint-planner` as gaps.

### Step 3: The Gate Function (per criterion)

```
1. IDENTIFY  — what command proves this claim?
2. RUN       — execute the FULL command (fresh, complete)
3. READ      — full output, exit code, count failures
4. VERIFY    — does the output actually confirm the claim?
5. ONLY THEN — record the verdict WITH the evidence
```

### Step 4: Record the Evidence

Write `planning/sprints/sprint-XXX/VERIFICATION-REPORT.MD`:

```markdown
# VERIFICATION-REPORT — Sprint XXX

| # | Kriteria | Perintah | Hasil | Bukti |
|---|----------|----------|-------|-------|
| 1 | Parse Excel valid | npm test -- parser | PASS | "12 passed, 0 failed" |
| 2 | Excel kosong → error rapi | npm test -- parser-empty | FAIL | expected graceful, got throw |
| 3 | Dashboard tampil | (screenshot) | PASS | dashboard-2026-09-11.png |

## GAPS (correctness only)
- Item 2: edge case kosong tidak tertangani...

## Verdict: REWORK (1 FAIL, 0 UNTESTED)
```

### Step 5: Verdict

- **APPROVED** — all criteria PASS with evidence → next:
  `/security-dependency-auditor`
- **REWORK** — any FAIL or load-bearing UNTESTED → return to the builder with
  the gap list.

### Step 6: Bounded Fix Loop

- Rounds 1–3: builder (same context) fixes; re-verify only the failed items.
- Round 4–5 (if still failing): hand rework to a **fresh** builder with a
  stronger model.
- After 5 failed rounds: STOP. Escalate to the human with the full gap list.
  A plan whose every path forward is a guess is not a plan.

## Rules

MUST:
- Run every verification command yourself, this session
- Report GAPS only (correctness) — never style preferences
- Include raw evidence: real command output, exit codes, screenshots
- Judge ONLY what ACCEPTANCE.MD defines (no scope creep)
- Stay read-only on src/ and tests — you verify, you do not fix

MUST NOT:
- Accept "should pass", "probably works", or a builder's success report as
  evidence
- Count a previous run as verification
- Skip UNTESTED items silently — they are gaps

## Evidence Table (what each claim requires)

| Klaim | Butuh | TIDAK cukup |
|-------|-------|-------------|
| Tests pass | output: 0 failures | run sebelumnya |
| Build sukses | exit code 0 | linter lolos |
| Bug fixed | test gejala asli: pass | kode berubah |
| Requirements met | checklist per baris | "tests passing" |
| Agent selesai | git diff shows changes | laporan "success" |

## False Positives

| You are tempted to | Do instead |
|--------------------|------------|
| "Output looks fine, probably passes" | Read the exit code and failure count |
| Suggest a nicer architecture while reviewing | Stop — report gaps only |
| Fix the failing line yourself | Stop — verdict REWORK, builder fixes |
| Mark a flaky test PASS because it passed once | Run again; flaky = FAIL + gap |

## Next Step

→ APPROVED: `/security-dependency-auditor`. REWORK: builder loop (max 5
rounds, then human). After the final sprint APPROVED:
`/deploy-production-gate`, then `/state-keeper`.

From susantosanto/config-opencode, MIT.

More from susantosanto/config-opencode