Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

README.md

code-review

Multi-agent branch code review plugin for Claude Code. Reviews conventions, architecture, bugs, security, release readiness, and git history in parallel using specialized subagents.

Usage

/code-review:full [focus area or specific concerns]

How it works

The review runs in 4 phases:

Phase 1 — Load Context (main thread)

Loads everything into the main context window unsummarized:

  • Branch diff and commit list
  • MR/PR description, comments, and pipeline status
  • Ticket description and acceptance criteria

Phase 2 — Explore Surrounding Code (single subagent)

An Explore subagent builds understanding of the touched code areas — callers, callees, data flow, downstream effects, and previous state.

Phase 3 — Parallel Review (6 subagents)

Six specialized review agents run in parallel, each with its own checklist and scope. They receive the branch range and fetch their own git data.

Agent Scope Color
review-conventions Documented conventions, naming, test structure, annotations cyan
review-architecture Module placement, layers, coupling, abstractions, API design, dependency direction blue
review-bugs Logic errors, edge cases, error handling, race conditions, resource leaks red
review-security Injection, auth, secrets, input validation, XSS, OWASP yellow
review-release Deployment risks: migrations, messaging infra, config changes, API contracts, rollback safety magenta
review-git-history Commit atomicity, refactoring separation, fixup detection, message format green

Each agent returns structured findings with confidence ratings (0-100).

Phase 4 — Validate & Report (main thread)

The main agent validates every finding against its full unsummarized context. Findings that can't be verified are dropped. The final report is written to .claude/review-report/<topic>.md in the project directory.

Design decisions

  • Context stays unsummarized — Phase 1 loads MR/ticket/diff directly into main context so validation in Phase 4 has full fidelity
  • Agents fetch their own git data — the orchestrator passes only the branch range (<base>...HEAD), avoiding context-passing errors
  • Each agent knows its boundaries — explicit "Out of Scope" sections prevent duplicate findings across agents
  • No automatic filtering by confidence — the main agent verifies findings manually rather than relying on a numeric threshold
  • Platform-agnostic — works with any code hosting platform and issue tracker

Installation

claude plugin marketplace add fprochazka/claude-code-plugins --scope user
claude plugin install code-review@fprochazka-claude-code-plugins --scope user