Multi-agent branch code review plugin for Claude Code. Reviews conventions, architecture, design craft, bugs, performance, security, release readiness, and git history in parallel using specialized subagents.
/code-review:full [focus area or specific concerns]
/code-review:post
/code-review:full— runs the multi-phase review and writes a report./code-review:post— posts the report from the current session to the GitLab MR as inline diff comments + one summary comment. Requires a/code-review:fullrun earlier in the same session. GitLab-only.
The review runs in 4 phases:
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
An Explore subagent builds understanding of the touched code areas — callers, callees, data flow, downstream effects, and previous state.
Eight 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-code-design |
Opinionated design-craft improvement hints: functional-core purity, rich domain models, value objects, right-sized abstraction, intent-revealing clarity | orange |
review-bugs |
Logic errors, edge cases, error handling, race conditions, lost updates, resource leaks | red |
review-performance |
Data-access efficiency (N+1, eager/lazy loading, preload-before-logic), query cost, transaction scope, caching, memory | purple |
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).
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.
- 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
claude plugin marketplace add fprochazka/claude-code-plugins --scope user
claude plugin install code-review@fprochazka-claude-code-plugins --scope user