plan-review

name	plan-review
description	Use when a plan needs validation through a planner-critic loop. Triggers on "플랜 리뷰", "plan review", or when explicitly requested.

Plan Review — Planner ↔ Critic Validation Loop

Overview

Iterative refinement of a plan through structured debate between planner and critic agents. Runs until the critic approves or max iterations reached.

Process

Phase 1: Critic Review

Delegate to critic agent (complex category):

Review the plan against:

• All TODOs have concrete acceptance criteria
• File references are correct
• No unvalidated assumptions
• Scope is clear
• QA scenarios are agent-executable
• Dependencies identified
• No YAGNI violations

Output structured findings (CRITICAL / IMPORTANT / MINOR), then judge:

• APPROVE if zero CRITICAL findings remain — IMPORTANT and MINOR findings may still exist but do not block approval. Note them as recommendations in the approval.
• REQUEST_CHANGES if any CRITICAL finding exists.

Phase 1.5: Codex Adversarial Review (Optional)

Requires codex-plugin-cc installed. Check hud.json → codex field. If false, skip this phase entirely and proceed to Phase 2.

If Phase 1 result is REQUEST_CHANGES, skip Phase 1.5 entirely — Codex adversarial review only runs after critic APPROVE.

After Phase 1 critic outputs APPROVE, dispatch codex:codex-rescue agent to get a second opinion from a different model. This combats self-review bias by having GPT challenge the plan that Claude created.

How to dispatch:

Agent(subagent_type="codex:codex-rescue", prompt="""
You are reviewing a plan document as an adversarial reviewer.
Your job is to challenge the plan, find pitfalls, and identify hidden assumptions.

Focus on:
- Practical pitfalls the reviewer may have missed
- Alternative architectures worth considering
- Hidden assumptions in the plan
- Real-world failure modes

Plan content:
{plan_content}

Return your findings as structured JSON:
{ "verdict": "ALLOW" | "BLOCK", "findings": [...] }
""")

Handling Codex results:

• Codex returns ALLOW (no issues) → proceed to Phase 2 normally
• Codex returns BLOCK with findings → add findings as IMPORTANT-level items in a separate "Codex Adversarial Review" section within Phase 2 User Feedback. The user decides whether to accept or dismiss each finding.
• Codex unavailable/timeout/error → skip silently, proceed to Phase 2 with critic findings only

Key constraint: Codex findings NEVER override the critic's APPROVE judgment. They are additional perspective only. The user has final say on whether to act on Codex feedback.

Phase 2: User Feedback

If the critic has findings, present them to the user. Ask if they agree with the critique.

Phase 3: Planner Revision

If changes are needed, delegate to planner agent (complex category):

• Address each CRITICAL finding (mandatory)
• Address IMPORTANT findings (recommended)
• MINOR findings are optional
• Steelman the critic's concerns before dismissing any

Phase 4: Re-Review

Loop back to Phase 1 with the revised plan.

Max iterations: 5. If CRITICAL findings persist after 5 rounds, present the best version to the user with remaining CRITICAL concerns noted and let the user decide whether to proceed or continue refining.

Phase 5: Finalize

When critic outputs APPROVE:

1. Present final plan to user
2. The planner agent saves the revised plan directly to vault using vault-save.sh (same pattern as plan skill). Return the vault path to the orchestrator.
3. Ready for execution

When to Use

• Complex plans with multiple components
• Plans that touch critical systems (auth, payments, data migration)
• When the user explicitly requests review ("plan review", "플랜 리뷰")
• NOT needed for simple, single-feature plans