claude-pr-deep-review
v1.0.1
Published
Deep PR review skill for Claude Code. Combines code-review, security-review, simplify, test runs, and spec cross-check in a single pass.
Maintainers
Readme
claude-pr-deep-review
End-to-end PR/branch review skill for Claude Code. Combines code-review, security-review, simplify, real test runs, and spec cross-check in a single pass with evidence-backed reporting.
What It Does
10-phase review process:
- Setup (worktree or checkout)
- Inspect diff (stat, log, full diff, read changed files)
- Codebase pattern comparison (adjacent siblings, convention divergence)
- Code review (5-angle: line-by-line, removed-behavior, cross-file, language-pitfall, wrapper correctness)
- Security review (AuthN/AuthZ bypass, SQL injection, info leakage)
- Simplify (dead code, test divergence, trailing whitespace)
- Run tests (unit, integration, typecheck, lint)
- Test coverage gap analysis (happy path, 401, 403, cross-tenant, edge cases)
- Spec cross-check (requirement table with evidence)
- Final report (structured markdown with evidence for every claim)
Installation
npx (one command)
npx claude-pr-deep-reviewcurl
curl -fsSL https://raw.githubusercontent.com/danolez1/claude-pr-deep-review/main/install.sh | bashClone
git clone https://github.com/danolez1/claude-pr-deep-review.git
cd claude-pr-deep-review && ./install.shUsage
In any Claude Code session:
/pr-deep-review feature/my-branchArguments (free-form text):
- branch (required): PR branch name or PR number
- base (optional): base branch (default: develop, then main)
- task details (optional): acceptance criteria from ticket
- reference doc (optional): absolute path to spec file for cross-check
Iron Rules
- No completion claim without evidence (command + exit code + counts)
- Pre-existing failures must be proven (re-run on base)
- Regen before test (if generated files exist)
- Adjacency proves convention (quote sibling file:line)
- Never trust agent summaries (read actual output)
License
MIT
