devflow-verification-gate

Skill: devflow-verification-gate

Iron-Law port of Superpowers verification-before-completion. The review-time guard that makes self-approval possible: prove the work, don't claim it.

When to use

In review state, in this exact order:

1. devflow-receiving-review — triage code-review findings (Critical / Important / Minor)
2. devflow-plan-reconciliation (DF-310) — reconcile plan ↔ actual implementation (done/partial/missing/extra/moved)
3. devflow-verification-gate (this skill) — prove every AC with a real evidence command
4. devflow-adr-compliance — ADR-glob path-check
5. devflow-knowledge-completer — close any wiki gaps
6. flow_update review → done with all 4 discipline-tokens

Without this exact order under agent_with_discipline, the gate blocks with discipline_incomplete.

Process

1 · Gather acceptance criteria

Pull the flow's acceptance-criteria array. Each entry needs an evidence command — the kind of command a reviewer would run to confirm it themselves, not an internal assertion.

AC kind	Typical evidence command
New API endpoint	curl -X POST … | jq showing the 201 response
New backend behavior	npm run test:api -- <suite>
New frontend	npx playwright test <spec> or screenshot path
Migration	sqlite3 data/devflow.db ".schema <table>"
Doc-page	curl /api/projects/X/docs/<id> | jq .data.content
Glob/regex helper	vitest run <unit-test>

If a criterion does not lend itself to a single command, split it into smaller criteria first. Do not write a "summary" command that doesn't actually verify anything.

2 · Run each command and capture output

Execute, capture full stdout + stderr, exit code, duration. Truncate to ~30 lines per command for the testing-instructions write-out (full log goes into the discipline-token evidence).

3 · Format testingInstructions

Write a markdown section per criterion:

### AC-N: <criterion text>

**Command:** `npm run test:api -- adr-compliance.test.ts`
**Exit code:** 0
**Output (truncated):**
\`\`\`
... 8 passed in 141ms ...
\`\`\`