Process
Phase 1 — Find Plan
- If path given, use it. Otherwise find latest plan file using
saml-latest-planfile. - Read the plan file. If not found, tell user to run
saml-planfirst.
Phase 2 — Gather Context
- Read codebase context:
AGENTS.md,.github/copilot-instructions.md, or similar. - Understand project conventions relevant to the plan.
Phase 3 — Review-Fix Loop
Launch a general-purpose sub-agent (fresh context window) to review the plan. Use saml-model skill to pick the model.
The reviewer must NOT inherit any prior conversation history — pass all context explicitly in the prompt:
- The full plan file contents
- Codebase context gathered in Phase 2
- Project conventions and constraints
You are a senior engineer. You HATE this plan and want to criticize it. Review this plan for simplicity, completeness, correctness, and actionability. Return PASS or FAIL with specific issues. Write review to
PLANFILE.review.md.
- If PASS → done.
- If FAIL → fix the plan file based on feedback, re-run.
- Max 3 iterations.
Done
Tell user: "✅ Plan reviewed: <path>"
If iteration limit hit, show remaining issues.