Review Cycle

Unified multi-dimensional code review orchestrator with dual-mode (session/module) file discovery, 7-dimension parallel analysis, iterative deep-dive on critical findings, and optional automated fix pipeline with intelligent batching and parallel planning.

Architecture Overview

┌──────────────────────────────────────────────────────────────────────┐
│  Review Cycle Orchestrator (SKILL.md)                                │
│  → Pure coordinator: mode detection, phase dispatch, state tracking  │
└───────────────────────────────┬──────────────────────────────────────┘
                                │
  ┌─────────────────────────────┼─────────────────────────────────┐
  │           Review Pipeline (Phase 1-5)                          │
  │                                                                │
  │  ┌─────────┐  ┌─────────┐  ┌─────────┐  ┌─────────┐  ┌─────────┐
  │  │ Phase 1 │→ │ Phase 2 │→ │ Phase 3 │→ │ Phase 4 │→ │ Phase 5 │
  │  │Discovery│  │Parallel │  │Aggregate│  │Deep-Dive│  │Complete │
  │  │  Init   │  │ Review  │  │         │  │(cond.)  │  │         │
  │  └─────────┘  └─────────┘  └─────────┘  └─────────┘  └─────────┘
  │   session|      7 agents     severity     N agents     finalize
  │   module        ×cli-explore  calc        ×cli-explore  state
  │                                  ↕ loop
  └────────────────────────────────────────────────────────────────┘
                                │
                          (optional --fix)
                                │
  ┌─────────────────────────────┼─────────────────────────────────┐
  │           Fix Pipeline (Phase 6-9)                             │
  │                                                                │
  │  ┌─────────┐  ┌─────────┐  ┌─────────┐  ┌─────────┐
  │  │ Phase 6 │→ │ Phase 7 │→ │ Phase 8 │→ │ Phase 9 │
  │  │Discovery│  │Parallel │  │Execution│  │Complete │
  │  │Batching │  │Planning │  │Orchestr.│  │         │
  │  └─────────┘  └─────────┘  └─────────┘  └─────────┘
  │   grouping     N agents     M agents     aggregate
  │   + batch      ×cli-plan    ×cli-exec    + summary
  └────────────────────────────────────────────────────────────────┘

Key Design Principles

1. Dual-Mode Review: Session-based (git changes) and module-based (path patterns) share the same review pipeline (Phase 2-5), differing only in file discovery (Phase 1)
2. Pure Orchestrator: Execute phases in sequence, parse outputs, pass context between them
3. Progressive Phase Loading: Phase docs are read on-demand when that phase executes, not all at once
4. Auto-Continue: All phases run autonomously without user intervention between phases
5. Task Attachment Model: Sub-tasks attached/collapsed dynamically in TaskCreate/TaskUpdate
6. Optional Fix Pipeline: Phase 6-9 triggered only by explicit --fix flag or user confirmation after Phase 5
7. Content Preservation: All agent prompts, code, schemas preserved verbatim from source commands

Usage

# Review Pipeline (Phase 1-5)
Skill(skill="review-cycle", args="<path-pattern>")                                    # Module mode
Skill(skill="review-cycle", args="[session-id]")                                      # Session mode
Skill(skill="review-cycle", args="[session-id|path-pattern] [FLAGS]")                 # With flags

# Fix Pipeline (Phase 6-9)
Skill(skill="review-cycle", args="--fix <review-dir|export-file>")                    # Fix mode
Skill(skill="review-cycle", args="--fix <review-dir> [FLAGS]")                        # Fix with flags

# Flags
--dimensions=dim1,dim2,...    Custom dimensions (default: all 7)
--max-iterations=N           Max deep-dive iterations (default: 3)
--fix                        Enter fix pipeline after review or standalone
--resume                     Resume interrupted fix session
--batch-size=N               Findings per planning batch (default: 5, fix mode only)

# Examples
Skill(skill="review-cycle", args="src/auth/**")                                       # Module: review auth
Skill(skill="review-cycle", args="src/auth/**,src/payment/**")                        # Module: multiple paths
Skill(skill="review-cycle", args="src/auth/** --dimensions=security,architecture")    # Module: custom dims
Skill(skill="review-cycle", args="WFS-payment-integration")                           # Session: specific
Skill(skill="review-cycle", args="")                                                  # Session: auto-detect
Skill(skill="review-cycle", args="--fix .workflow/active/WFS-123/.review/")           # Fix: from review dir
Skill(skill="review-cycle", args="--fix --resume")                                    # Fix: resume session

Mode Detection

// Input parsing logic (orchestrator responsibility)
function detectMode(args) {
  if (args.includes('--fix')) return 'fix';
  if (args.match(/\*|\.ts|\.js|\.py|src\/|lib\//)) return 'module';  // glob/path patterns
  if (args.match(/^WFS-/) || args.trim() === '') return 'session';   // session ID or empty
  return 'session';  // default
}

Execution Flow

Input Parsing:
   └─ Detect mode (session|module|fix) → route to appropriate phase entry

Review Pipeline (session or module mode):

Phase 1: Discovery & Initialization
   └─ Ref: phases/01-discovery-initialization.md
      ├─ Session mode: session discovery → git changed files → resolve
      ├─ Module mode: path patterns → glob expand → resolve
      └─ Common: create session, output dirs, review-state.json, review-progress.json

Phase 2: Parallel Review Coordination
   └─ Ref: phases/02-parallel-review.md
      ├─ Launch 7 cli-explore-agent instances (Deep Scan mode)
      ├─ Each produces dimensions/{dimension}.json + reports/{dimension}-analysis.md
      └─ CLI fallback: Gemini → Qwen → Codex

Phase 3: Aggregation
   └─ Ref: phases/03-aggregation.md
      ├─ Load dimension JSONs, calculate severity distribution
      ├─ Identify cross-cutting concerns (files in 3+ dimensions)
      └─ Decision: critical > 0 OR high > 5 OR critical files → Phase 4
                   Else → Phase 5

Phase 4: Iterative Deep-Dive (conditional)
   └─ Ref: phases/04-iterative-deep-dive.md
      ├─ Select critical findings (max 5 per iteration)
      ├─ Launch deep-dive agents for root cause analysis
      ├─ Re-assess severity → loop back to Phase 3 aggregation
      └─ Exit when: no critical findings OR max iterations reached

Phase 5: Review Completion
   └─ Ref: phases/05-review-completion.md
      ├─ Finalize review-state.json + review-progress.json
      ├─ Prompt user: "Run automated fixes? [Y/n]"
      └─ If yes → Continue to Phase 6

Fix Pipeline (--fix mode or after Phase 5):

Phase 6: Fix Discovery & Batching
   └─ Ref: phases/06-fix-discovery-batching.md
      ├─ Validate export file, create fix session
      └─ Intelligent grouping by file+dimension similarity → batches

Phase 7: Fix Parallel Planning
   └─ Ref: phases/07-fix-parallel-planning.md
      ├─ Launch N cli-planning-agent instances (≤10 parallel)
      ├─ Each outputs partial-plan-{batch-id}.json
      └─ Orchestrator aggregates → fix-plan.json

Phase 8: Fix Execution
   └─ Ref: phases/08-fix-execution.md
      ├─ Stage-based execution per aggregated timeline
      ├─ Each group: analyze → fix → test → commit/rollback
      └─ 100% test pass rate required

Phase 9: Fix Completion
   └─ Ref: phases/09-fix-completion.md
      ├─ Aggregate results → fix-summary.md
      └─ Optional: complete workflow session if all fixes successful

Complete: Review reports + optional fix results

Phase Reference Documents (read on-demand when phase executes):

Core Rules

1. Start Immediately: First action is TaskCreate initialization, second action is Phase 1 execution
2. Mode Detection First: Parse input to determine session/module/fix mode before Phase 1
3. Parse Every Output: Extract required data from each phase for next phase
4. Auto-Continue: Check TaskList status to execute next pending phase automatically
5. Progressive Phase Loading: Read phase docs ONLY when that phase is about to execute
6. DO NOT STOP: Continuous multi-phase workflow until all applicable phases complete
7. Conditional Phase 4: Only execute if aggregation triggers iteration (critical > 0 OR high > 5 OR critical files)
8. Fix Pipeline Optional: Phase 6-9 only execute with explicit –fix flag or user confirmation

Data Flow

User Input (path-pattern | session-id | --fix export-file)
    ↓
[Mode Detection: session | module | fix]
    ↓
Phase 1: Discovery & Initialization
    ↓ Output: sessionId, reviewId, resolvedFiles, reviewMode, outputDir
    ↓         review-state.json, review-progress.json
Phase 2: Parallel Review Coordination
    ↓ Output: dimensions/*.json, reports/*-analysis.md
Phase 3: Aggregation
    ↓ Output: severityDistribution, criticalFiles, deepDiveFindings
    ↓ Decision: iterate? → Phase 4 : Phase 5
Phase 4: Iterative Deep-Dive (conditional, loops with Phase 3)
    ↓ Output: iterations/*.json, reports/deep-dive-*.md
    ↓ Loop: re-aggregate → check criteria → iterate or exit
Phase 5: Review Completion
    ↓ Output: final review-state.json, review-progress.json
    ↓ Decision: fix? → Phase 6 : END
Phase 6: Fix Discovery & Batching
    ↓ Output: finding batches (in-memory)
Phase 7: Fix Parallel Planning
    ↓ Output: partial-plan-*.json → fix-plan.json (aggregated)
Phase 8: Fix Execution
    ↓ Output: fix-progress-*.json, git commits
Phase 9: Fix Completion
    ↓ Output: fix-summary.md, fix-history.json

TaskCreate/TaskUpdate Pattern

Review Pipeline Initialization:

TaskCreate({ subject: "Phase 1: Discovery & Initialization", activeForm: "Initializing review" });
TaskCreate({ subject: "Phase 2: Parallel Reviews (7 dimensions)", activeForm: "Reviewing" });
TaskCreate({ subject: "Phase 3: Aggregation", activeForm: "Aggregating findings" });
TaskCreate({ subject: "Phase 4: Deep-dive (conditional)", activeForm: "Deep-diving" });
TaskCreate({ subject: "Phase 5: Review Completion", activeForm: "Completing review" });

During Phase 2 (sub-tasks for each dimension):

// Attach dimension sub-tasks
TaskCreate({ subject: "  → Security review", activeForm: "Analyzing security" });
TaskCreate({ subject: "  → Architecture review", activeForm: "Analyzing architecture" });
TaskCreate({ subject: "  → Quality review", activeForm: "Analyzing quality" });
// ... other dimensions
// Collapse: Mark all dimension tasks completed when Phase 2 finishes

Fix Pipeline (added after Phase 5 if triggered):

TaskCreate({ subject: "Phase 6: Fix Discovery & Batching", activeForm: "Batching findings" });
TaskCreate({ subject: "Phase 7: Parallel Planning", activeForm: "Planning fixes" });
TaskCreate({ subject: "Phase 8: Execution", activeForm: "Executing fixes" });
TaskCreate({ subject: "Phase 9: Fix Completion", activeForm: "Completing fixes" });

Error Handling

Review Pipeline Errors

Fix Pipeline Errors

CLI Fallback Chain

Gemini → Qwen → Codex → degraded mode

Fallback Triggers: HTTP 429/5xx, connection timeout, invalid JSON output, low confidence < 0.4, analysis too brief (< 100 words)

Output File Structure

.workflow/active/WFS-{session-id}/.review/
├── review-state.json                    # Orchestrator state machine
├── review-progress.json                 # Real-time progress
├── dimensions/                          # Per-dimension results (Phase 2)
│   ├── security.json
│   ├── architecture.json
│   ├── quality.json
│   ├── action-items.json
│   ├── performance.json
│   ├── maintainability.json
│   └── best-practices.json
├── iterations/                          # Deep-dive results (Phase 4)
│   ├── iteration-1-finding-{uuid}.json
│   └── iteration-2-finding-{uuid}.json
├── reports/                             # Human-readable reports
│   ├── security-analysis.md
│   ├── security-cli-output.txt
│   ├── deep-dive-1-{uuid}.md
│   └── ...
└── fixes/{fix-session-id}/             # Fix results (Phase 6-9)
    ├── partial-plan-*.json
    ├── fix-plan.json
    ├── fix-progress-*.json
    ├── fix-summary.md
    ├── active-fix-session.json
    └── fix-history.json

Related Commands

View Progress

ccw view

Workflow Pipeline

# Step 1: Review (this skill)
Skill(skill="review-cycle", args="src/auth/**")

# Step 2: Fix (continue or standalone)
Skill(skill="review-cycle", args="--fix .workflow/active/WFS-{session-id}/.review/")