review-cycle_skill

Overview

This skill is a unified, multi-dimensional code review orchestrator that runs session-based (git changes) or module-based (path patterns) reviews and optionally executes an automated fix pipeline. It coordinates parallel analysis across seven review dimensions, performs iterative deep dives on critical findings, and can export and execute batched fix tasks. The tool is an orchestrator only — it manages phases, agents, progress state, and artifact output for reliable, repeatable review workflows.

How this skill works

The orchestrator detects mode (session, module, or fix) and runs a phased pipeline: discovery, parallel review (7 agents), aggregation, conditional iterative deep-dive, and completion. If --fix is provided or user confirms, it continues into fix discovery, parallel planning, export-to-task JSON, staged execution, and final aggregation. Each phase is loaded on demand, subagents are spawned and closed explicitly, and outputs (JSON, reports, tasks, plan.json) are passed between phases.

When to use it

• Run code-wide automated reviews across many files or a specific git session.
• Detect and triage cross-cutting problems that appear in multiple review dimensions.
• Automate end-to-end fix planning and export FIX-*.json tasks for work orchestration.
• Iteratively investigate critical issues with repeated deep-dive cycles.
• Integrate review-run triggers into CI/CD via workflow:review-cycle events.

Best practices

• Start with default all-7 dimensions; narrow with --dimensions only when needed.
• Use session mode for incremental reviews and module mode for scoped sweeps by path pattern.
• Preserve review artifacts and review-state.json for auditability and to resume fix sessions.
• Set sensible --max-iterations (default 3) to bound deep-dive loops and avoid runaway cycles.
• Use --export-tasks to produce .task/FIX-*.json and plan.json for downstream task orchestration.

Example use cases

• Run review-cycle src/auth/** to perform a targeted security and architecture sweep of auth code.
• Trigger review-cycle with an empty input in CI to auto-detect current session changes and run a review.
• After a review, run review-cycle --fix <reviewDir> to generate FIX tasks, export plan.json, and execute staged fixes.
• Resume an interrupted fix run with review-cycle --fix --resume to continue planning or execution without re-running discovery.
• Limit analysis to specific concerns: review-cycle src/** --dimensions=security,performance to focus agent effort.