Vérification de spécification

---

name: verify-spec description: Check implemented work against executable spec.yml before completion or deploy by validating acceptance criteria coverage and red/green test evidence user-invocable: true argument-hint: "docs/specs/<slug>/spec.yml [--phase red|green|complete]"

Instructions

You are performing a conformance check for spec-driven work. This is not a replacement for tests; it confirms that tests and implementation evidence match the executable spec.yml.

Step 1: Read Inputs

Read:

1. The provided docs/specs/<slug>/spec.yml
2. The implementation_plan and tasks sections inside that spec
3. docs/contributing/guides/spec-driven-development.md
4. Current session status and related test tracking:

   python3 scripts/sessions.py status --json
   python3 scripts/sessions.py check-tests --session <SESSION_ID>
   

If there is no active session, ask for the session ID or perform a document-only review.

Step 2: Build The Coverage Table

First run:

python3 scripts/spec_validate.py docs/specs/<slug>/spec.yml
python3 scripts/spec_verify.py docs/specs/<slug>/spec.yml --phase complete

For every scenario and acceptance criterion, record:

• Implemented: yes/no/partial/unknown
• Coverage: covered/ambiguous/uncovered/blocked/waived based on coverage_status, linked tests, and verification_ids
• Evidence: test file, command, manual verification, code reference, or gap
• Risk: none/low/medium/high

Step 3: Validate Required Evidence

Pass only when:

• Every completed acceptance criterion has green evidence.
• Every required acceptance criterion is covered, waived, or blocked by an accepted user/external dependency; ambiguous and uncovered criteria fail.
• Every required assumption whose required_before is implementation is confirmed, corrected and applied, waived, or externally blocked.
• required assumptions with unchecked, checking, or contradicted status fail the implementation gate unless the user accepts a waiver or external blocker.
• Every required red phase has evidence before implementation. Red evidence may be failed as expected, passed unexpectedly with a reason, missing-test with a reason, skipped with a reason, or not applicable.
• Every required green/final check has passing, user-confirmed, waived, or accepted-blocker evidence.
• Privacy/security criteria have concrete code or test evidence.
• Changed source files have related tests or an explicit skip reason.
• Open questions are resolved or listed as accepted residual risk.

Failed required checks must not be treated as a summary-only issue. The report must identify the affected acceptance criteria and confirm that follow-up tasks exist or that the user explicitly accepted a waiver/blocker. failed required checks keep the spec active until the linked follow-up work is done or the user accepts a waiver/blocker.

Playwright green evidence is only valid after the implementation has been deployed to dev, Vercel is Ready, and the spec has run against app.dev.openmates.org.

Step 4: Output Report

Use this format:

## Spec Verification

Spec: docs/specs/<slug>/spec.yml
Status: pass | fail | partial

| ID | Status | Evidence | Risk |
| --- | --- | --- | --- |
| S-1 | pass | `frontend/...spec.ts` | none |
| AC-1 | pass | `python3 scripts/tests.py run ...` | none |

Gaps:
- <gap or none>

Deploy note:
- Spec: docs/specs/<slug>/spec.yml
- Tests: <commands/results>
- Residual risk: <none or concise note>

Step 5: Stop On Failure

If status is fail, do not deploy. Fix the gap, update the spec if product intent changed, or ask the user to accept a documented risk.

Rules

• Do not mark criteria complete based on intent alone.
• Do not weaken or remove acceptance criteria to make verification pass.
• Do not mark ambiguous coverage as complete; normalize vague criteria into concrete checks first.
• Prefer Playwright/pytest/vitest via repo-approved test commands over manual browser checks.
• Manual verification is allowed only when automation is impractical and the reason is documented.
• Do not mark a full spec verified while required green evidence is missing.