nm-pensive-unified-review（nm 沉思统一审查）

Night Market Skill — ported from claude-night-market/pensive. For the full experience with agents, hooks, and commands, install the Claude Code plugin.

Table of Contents

• Quick Start
• When to Use
• Review Skill Selection Matrix
• Workflow
• 1. Analyze Repository Context
• 2. Select Review Skills
• 3. Execute Reviews
• 4. Integrate Findings
• Review Modes
• Auto-Detect (default))
• Focused Mode
• Full Review Mode
• Quality Gates
• Deliverables
• Executive Summary
• Domain-Specific Reports
• Integrated Action Plan
• Modular Architecture
• Exit Criteria

Unified Review Orchestration

Intelligently selects and executes appropriate review skills based on codebase analysis and context.

Quick Start

# Auto-detect and run appropriate reviews
/full-review

# Focus on specific areas
/full-review api          # API surface review
/full-review architecture # Architecture review
/full-review bugs         # Bug hunting
/full-review tests        # Test suite review
/full-review all          # Run all applicable skills

Verification: Run pytest -v to verify tests pass.

When To Use

• Starting a full code review
• Reviewing changes across multiple domains
• Need intelligent selection of review skills
• Want integrated reporting from multiple review types
• Before merging major feature branches

When NOT To Use

• Specific review type known

- use bug-review

• Test-review
• Architecture-only focus - use

architecture-review

• Specific review type known

- use bug-review

Review Skill Selection Matrix

Workflow

1. Analyze Repository Context

• Detect primary languages from extensions and manifests
• Analyze git status and diffs for change scope
• Identify project structure (monorepo, microservices, library)
• Detect build systems, testing frameworks, documentation

2. Select Review Skills

# Detection logic
if has_rust_files():
    schedule_skill("rust-review")
if has_api_changes():
    schedule_skill("api-review")
if has_test_files():
    schedule_skill("test-review")
if has_makefiles():
    schedule_skill("makefile-review")
if has_math_code():
    schedule_skill("math-review")
if has_architecture_changes():
    schedule_skill("architecture-review")
# Default
schedule_skill("bug-review")

Verification: Run pytest -v to verify tests pass.

3. Execute Reviews

Dispatch selected skills concurrently via the Agent tool. Use this mapping to resolve skill names to agent types:

Rules:

• Never use skill names as agent types (e.g., pensive:math-review is NOT an agent)
• When pensive:code-reviewer covers multiple domains, dispatch once with combined scope
• For skills without dedicated agents, use general-purpose and instruct it to invoke the Skill tool
• Maintain consistent evidence logging across all agents
• Track progress via TodoWrite

4. Integrate Findings

• Consolidate findings across domains
• Identify cross-domain patterns
• Prioritize by impact and effort
• Generate unified action plan

Deferred capture for backlog findings: Findings that are triaged to the backlog (out-of-scope for the current review or deferred by the team) should be preserved so they are not lost between review cycles. For each finding assigned to the backlog, run:

python3 scripts/deferred_capture.py \
  --title "<finding title>" \
  --source review \
  --context "Review dimension: <dimension>. <finding description>"

The <dimension> value should match the review skill that surfaced the finding (e.g. bug-review, api-review, architecture-review). This runs automatically after the action plan is finalised, without prompting the user.

Review Modes

Auto-Detect (default)

Automatically selects skills based on codebase analysis.

Focused Mode

Run specific review domains:

• /full-review api → api-review only
• /full-review architecture → architecture-review only
• /full-review bugs → bug-review only
• /full-review tests → test-review only

Full Review Mode

Run all applicable review skills:

• /full-review all → Execute all detected skills

Quality Gates

Each review must:

1. Establish proper context
2. Execute all selected skills successfully
3. Document findings with evidence
4. Prioritize recommendations by impact
5. Create action plan with owners

Deliverables

Executive Summary

• Overall codebase health assessment
• Critical issues requiring immediate attention
• Review frequency recommendations

Domain-Specific Reports

• API surface analysis and consistency
• Architecture alignment with ADRs
• Test coverage gaps and improvements
• Bug analysis and security findings
• Performance and maintainability recommendations

Integrated Action Plan

• Prioritized remediation tasks
• Cross-domain dependencies
• Assigned owners and target dates
• Follow-up review schedule

Modular Architecture

All review skills use a hub-and-spoke architecture with progressive loading:

• pensive:shared: Common workflow, output templates, quality checklists
• Each skill has modules/: Domain-specific details loaded on demand
• Cross-plugin deps: imbue:proof-of-work, imbue:diff-analysis/modules/risk-assessment-framework

This reduces token usage by 50-70% for focused reviews while maintaining full capabilities.

Exit Criteria

• All selected review skills executed
• Findings consolidated and prioritized
• Action plan created with ownership
• Evidence logged per structured output format

Supporting Modules

• Review workflow core - standard 5-step workflow pattern for all pensive reviews
• Output format templates - finding entry, severity, action item templates
• Quality checklist patterns - pre-review, analysis, evidence, deliverable checklists

Troubleshooting

Common Issues

If the auto-detection fails to identify the correct review skills, explicitly specify the mode (e.g., /full-review rust instead of just /full-review). If integration fails, check that TodoWrite logs are accessible and that evidence files were correctly written by the individual skills.