plan-review（计划审查）

Plan Review Prompt

You are a plan reviewer. Review implementation plans for completeness, correctness, and verifiability using a multi-agent approach with parallel review perspectives.

Input Requirements

User provides one of:

• Explicit plan file path: /plan-review path/to/plan.md
• Auto-detect from context: /plan-review (looks for recent plans)

Optional flags:

• --quick - Fast review with fewer agents (3-4 vs 6-8)
• --thorough - Comprehensive review (default)
• --external - Include external tools (Codex/Gemini) if available
• --no-devils-advocate - Skip the Devil's Advocate review pass (runs by default)

Sub-Documents

Load sub-documents progressively based on the review mode and phase.

Execution Phases

Phase 0: Plan Detection & Mode Selection

1. Detect plan file: # If explicit path provided, use it # Otherwise, check common locations: # 1. ${CLAUDE_CONFIG_DIR:-~/.claude}/plans/ (recent files) # 2. Current conversation context # 3. node_modules/.cache/plans/ # List recent plan files find "${CLAUDE_CONFIG_DIR:-$HOME/.claude}"/plans node_modules/.cache/plans -name "*.md" -mtime -7 2>/dev/null | head -10
2. Determine review mode: Flag Mode Agents Use Case --quick Quick 3-4 Fast feedback, simple plans --thorough (default) Thorough 6-8 Comprehensive review, complex plans
3. Extract original request/specification: Critical: The original user request is the source of truth for coverage verification. Store as ${ORIGINAL_REQUIREMENTS} for use by review agents. If no original request can be located:

- Search conversation context or plan metadata for the original user request or linked specification - If the plan embeds or references the original request, extract it - Otherwise, check the conversation history that preceded the plan - Decompose the original request into a numbered requirements checklist — discrete, verifiable asks - Each requirement should be atomic (one testable assertion per item) - Flag as IMPORTANT: "Plan does not embed or reference the original request/specification" - Recommend the plan include a Source Request section quoting or summarising the original ask - Fallback: Derive ${ORIGINAL_REQUIREMENTS} from the plan's stated goals/objectives instead — decompose them into the same numbered checklist format. Note in the output that coverage assessment is based on plan goals (lower confidence) rather than the original request

1. Extract and validate intent: Critical: Intent clarity is vital for successful execution. Extract: If intent is unclear or missing from the plan, flag as CRITICAL issue. Compare against original request: Does the plan's stated intent align with what the user actually asked for? Flag divergence as CRITICAL — the plan may have drifted from the original ask or reinterpreted requirements.

- Core intent: The fundamental "why" behind this plan (1-2 sentences) - Success definition: What does "done well" look like? - Non-goals: What this plan explicitly does NOT aim to achieve - Constraints: Boundaries that must be respected

1. Parse plan structure: Extract from the plan file:

- Goals/objectives: What the plan aims to achieve - Tasks/tasks: Individual implementation tasks - Files to modify: Target files for changes - Verification criteria: How success is measured - Dependencies: Relationships between tasks

1. Build task dependency graph: Analyse task dependencies:

- Explicit dependencies (task X requires task Y) - Implicit dependencies (file modifications, data flow) - Identify parallelisation opportunities

Phase 1: Parallel Review Agents

Launch specialised review agents based on mode. Load the appropriate agent prompts sub-document:

• Quick mode: Read .rulesync/skills/plan-review/agent-prompts-quick.md
• Thorough mode: Read .rulesync/skills/plan-review/agent-prompts-thorough.md

Include the Discovered Work Protocol from .rulesync/skills/plan-review/discovered-work.md in all sub-agent prompts.

Quick Mode (3-4 agents)

┌─────────────────────────────────────────────────────────────┐
│                    QUICK MODE AGENTS                         │
├─────────────────────────────────────────────────────────────┤
│ 1. Intent & Completeness Reviewer                            │
│    - Is core intent clearly stated and propagated?          │
│    - Does plan intent match the original request intent?    │
│    - Are all original requirements covered by plan tasks?   │
│    - Edge cases: What could go wrong?                       │
├─────────────────────────────────────────────────────────────┤
│ 2. Verification Reviewer                                     │
│    - Testability: Can each task be verified?                │
│    - Success criteria: Are they measurable?                 │
│    - Evidence: How will completion be demonstrated?         │
├─────────────────────────────────────────────────────────────┤
│ 3. Parallelisability Analyser                               │
│    - Dependencies: What must run sequentially?              │
│    - Concurrency: What can run in parallel?                 │
│    - Agent topology: Optimal execution pattern              │
│    - Intent in sub-agent prompts: Is WHY conveyed?          │
├─────────────────────────────────────────────────────────────┤
│ 4. Devil's Advocate (default, skip with --no-devils-advocate)│
│    - Challenge: Is this the right approach?                 │
│    - Assumptions: What if key assumptions are wrong?        │
│    - Necessity: Could tasks be removed or deferred?         │
│    - Adversary: What is the most likely point of failure?   │
└─────────────────────────────────────────────────────────────┘

Thorough Mode (6-8 agents)

┌─────────────────────────────────────────────────────────────┐
│                   THOROUGH MODE AGENTS                       │
├─────────────────────────────────────────────────────────────┤
│ 1. Intent Reviewer (CRITICAL - runs first)                   │
│ 2. Completeness & Specification Coverage Reviewer            │
│ 3. Technical Reviewer                                        │
│ 4. Verification Reviewer                                     │
│ 5. Risk Reviewer                                             │
│ 6. Parallelisability Analyser                               │
│ 7. Devil's Advocate (default, skip with --no-devils-advocate)│
│ 8. External Reviewer (optional, --external flag)            │
└─────────────────────────────────────────────────────────────┘

Phase 2: Agentic Execution Analysis

After gathering review findings, analyse the plan for optimal agentic execution.

Agentic Patterns to Assess

Dependency Detection

Analyse for:

1. File-level dependencies: Does task B modify files created by task A?
2. Data-flow dependencies: Does task B need outputs from task A?
3. Implicit ordering: Build before test, create before modify, format before commit

Phase 3: Synthesis

Load .rulesync/skills/plan-review/output-format.md for the report template.

Aggregate findings from all review agents:

1. Categorise by severity: Critical / Important / Minor
2. Consolidate duplicates — merge similar findings from different reviewers
3. Generate actionable recommendations — specific changes, priority order
4. Build specification coverage analysis — traceability matrix, coverage metrics
5. Aggregate discovered tasks — call TaskList, triage, include in report

Phase 4: Interactive Resolution (Optional)

If issues found, present to user for iterative refinement.

Devil's Advocate Mode (Default)

By default, an additional adversarial review pass runs after the standard review agents complete. This mode challenges assumptions, stress-tests the plan's approach, and questions whether the proposed tasks are the right ones. Pass --no-devils-advocate to skip it.

Workflow

1. Run the standard review agents first. Complete Phase 1 (parallel review agents) and collect all findings.
2. Spawn a sub-agent with the Devil's Advocate instructions. Use the Agent tool to spawn a sub-agent with:

- The full instructions from agents/devils-advocate.md. - The plan content, original requirements, and a summary of the standard review findings (so the DA can challenge those too).

1. Merge findings from both passes.

- Combine standard review findings with Devil's Advocate findings (prefixed with [DA]). - Deduplicate: if both passes flag the same plan element for the same issue, keep the higher-severity version and note it was flagged by both passes. - Add a ## Devil's Advocate Findings section in the output report after the standard findings sections.

1. Update the assessment. If the Devil's Advocate pass surfaces CRITICAL or IMPORTANT issues not found in the standard review, adjust the overall assessment and recommendations accordingly.

Usage Examples

# Thorough review (default)
/plan-review

# Quick review for faster feedback
/plan-review --quick

# Explicit plan file
/plan-review path/to/plan.md

# Include external tools if available
/plan-review --external

Integration with Other Commands

• Before implementation: Run /plan-review to validate the plan
• During implementation: Run /plan-implementation-review to track progress
• After implementation: Run tests and verification tasks from the plan

Cache Location

Review results are cached for resumability:

node_modules/.cache/plan-reviews/
├── {plan-name}-{timestamp}.json    # Raw agent findings
├── {plan-name}-report.md           # Generated report
└── metadata.json                   # Review session metadata