Phase 1 — Detect Context

Purpose: Validate implementation against the active plan. Decides: AUTO-RECAP, FIX PLAN, or MANUAL REVIEW.

Phase 1 — Detect Context

Check system-reminder for plan file path (same pattern as /recap).

If no active plan found: → Output: ⚠ No plan found. Run /verify after an implementation session. → Stop.

Read the plan file. Extract:

• Livrables — expected deliverables (files created/modified, MCP memories, etc.)
• Fichiers modifiés — explicit file list if present
• Vérification — any test commands or checks defined in the plan

Phase 2 — Build Test Matrix

For each deliverable, generate and run automated tests:

Livrable type	Test method	Pass condition
Fichier créé	Read	File exists + non-empty
Fichier modifié	Read	Contains expected content/sections
Contenu spécifique	Grep	Pattern found (≥1 match)
Absence de pattern	Grep	0 matches
MCP Memory créé/modifié	search_by_tag or retrieve_memory	Content matches expected
MCP Memory supprimé	search_by_tag	0 results
Frontmatter YAML	Read lines 1-3	--- + required fields present
Skill/command runtime	(skip)	Mark ⏳ — not statically testable

Run all tests. Record each result: ✓ pass / ✗ fail / ⏳ pending.

Phase 3 — Score & Decide

Tests: ✓ {pass} / ✗ {fail} / ⏳ {pending}

fail = 0 AND pending ≤ 2  →  ► AUTO-RECAP
fail > 0                   →  ► FIX PLAN
pending > 2                →  ► MANUAL REVIEW

Phase 4a — AUTO-RECAP

Display compact test table:

╔══════════════════════════════════════════════╗
║  VERIFY  ✓ {n} passed  ⏳ {n} pending        ║
╠══════════════════════════════════════════════╣
║  ✓  {livrable}                               ║
║  ⏳  {livrable} — runtime, not tested        ║
╚══════════════════════════════════════════════╝

Output the summary table and inform the user all checks passed — ready to wrap up the session.

Phase 4b — FIX PLAN

Display full test table with failure details:

╔══════════════════════════════════════════════╗
║  VERIFY  ✓ {n}  ✗ {n}  ⏳ {n}               ║
╠══════════════════════════════════════════════╣
║  ✓  {livrable}                               ║
║  ✗  {livrable} — {reason: missing/wrong/...} ║
║  ⏳  {livrable} — runtime, not tested        ║
╚══════════════════════════════════════════════╝

Enter plan mode if not already active (call EnterPlanMode). Then write a Fix Plan section into the plan file:

── Fix Plan ──────────────────────────────────
1. [action] — [reason: what failed and why]
2. [action] — [reason: what failed and why]

── Verification ──────────────────────────────
Re-run after fix:
- [test command / Read path / Grep pattern]

Call ExitPlanMode for user review.

Phase 4c — MANUAL REVIEW

Call AskUserQuestion with:

• "Proceed to /recap (trust pending tests)"
• "Test manually first (I'll run the commands)"
• "Generate fix plan anyway"

Act on user choice:

• Option 1 → run Phase 4a (skip pending check)
• Option 2 → output the pending test commands for user to run manually; stop
• Option 3 → run Phase 4b

Critical

• Run ALL tests before deciding (don't stop at first failure)
• Output test table BEFORE calling Skill or ExitPlanMode
• Phase 4b: write Fix Plan BEFORE calling ExitPlanMode
• Execute silently: no intermediate commentary, just actions and final output
• /verify never modifies the codebase — read-only except for writing Fix Plan to plan file

Typical workflow

[impl done]  →  /verify  →  AUTO-RECAP  →  wrap up session
                          →  FIX PLAN   →  [fix]  →  /verify (retry)