polyreview

Deterministic code-review skill: runs Semgrep, Radon, Bandit and friends across a diff and emits a review bundle of raw per-tool output (review-bundle.v1.json) — one capture per analyzer for an agent to interpret (ADR-0020).

Status

Alpha. No API stability guarantees before 1.0. Expect breaking changes to CLI flags, config schema, and the bundle contract. The canonical version lives in pyproject.toml; check the installed version with polyreview --version once the flag lands or python -c "import code_review; print(code_review.__version__)".

Renamed from claude-code-review. Early development used the working name claude-code-review; the PyPI distribution is polyreview (it was renamed before its first release, so there is no claude-code-review package to migrate from). The rename drops the vendor prefix because the tool is agent-agnostic — its Agent Skill bundle is read by GitHub Copilot, Cursor, Codex, and other agents, not just Claude — and polyreview reads as a multi-language, multi-tool reviewer rather than an Anthropic-only one. The Python import name stays code_review.

Install

pip install polyreview
pipx install polyreview
uv tool install polyreview

The PyPI distribution and console-script binary are both polyreview; the Python import name stays code_review (PEP-423 allows the distribution name to differ from the import name).

0.1.0 is the first GA release to PyPI. It was staged as 0.1.0rc1 on TestPyPI first; see sdlc/docs/runbooks/release.md for the release process.

Analyzer prerequisites

pip install polyreview ships the Python analyzers — Semgrep, Bandit, Radon, Vulture, pydeps, cohesion — ready to run. The other analyzers depend on external tooling that pip can't bundle:

• JavaScript/TypeScript (ESLint, knip, jscpd, dependency-cruiser, jscomplexity) need a vendored node_modules.
• Secret & dependency scanning (gitleaks, Trivy) are standalone binaries that must be on your PATH.

Run polyreview run --capabilities and read analyzers[] to see which are active: each reports status: available or unavailable with the reason. To provision the full set from a source checkout, run ./scripts/setup.sh (Node tooling + offline caches) and install gitleaks/Trivy via your package manager. Analyzers that aren't available are skipped silently — so a finding-free run on a stack you expected coverage for may just mean the analyzer wasn't installed.

Use as an Agent Skill

polyreview is also an Agent Skill bundle: agents (Claude Code, Codex, GitHub Copilot, Gemini CLI, …) discover it from their user-level skills directory. After installing the package, place the bundle where your agents look — polyreview install is agent-independent, idempotent, and creates missing directories:

polyreview install                 # neutral ~/.agents/skills/ + every agent home present
polyreview install --agent claude  # one target: agents | claude | copilot | gemini
polyreview install --all           # every known agent location
polyreview install --force         # refresh an already-installed bundle in place

Install places the skill (SKILL.md, the config example, the vendored Semgrep rules) for discovery; it does not fetch the analyzer caches (node_modules, Trivy DB) — run ./scripts/setup.sh for the full toolchain. Remove it with the same target scoping:

polyreview uninstall               # mirrors install's default + --agent/--all scoping

Uninstall is marker-guarded: it removes only a directory that is verifiably the polyreview bundle, and never touches a sibling skill, an agent's own files (e.g. Claude's agents/reviewer.md), or the skills directory itself.

Quick start

polyreview run --review security --depth quick --diff HEAD~1..HEAD --output review.json

Writes a review bundle to review.json (review-bundle.v1.json schema): the request echo plus one entry per analyzer in the security/quick set, each carrying that tool's raw stdout/stderr verbatim and an ADR-0019 status (ok / error / timeout / unavailable). The bundle is for an agent to interpret — polyreview does not parse, rank, or merge findings.

What it does

• Deterministic analyzer layer — Semgrep, Bandit, Radon, and other rule-based scanners.
• Emits a raw review bundle (one capture per tool) so the consuming agent reads each tool's native output directly, rather than a lossy normalised summary (ADR-0020).
• Runs under /sandbox — analyzers are isolated from the host filesystem and network.

What it doesn't do

• LLM-based code review — that's the sibling intent-review project.
• Cross-tool aggregation, dedup, or severity ranking — it captures each tool raw; the agent interprets.
• CI orchestration — invoke it from your existing pipeline; it doesn't replace one.

Full reference

Complete review-set taxonomy, CLI resolution rules, and configuration knobs: .claude/skills/code-review/SKILL.md.

Development

git clone https://github.com/jiludvik2/agentic-skills
cd agentic-skills/code-review
./scripts/setup.sh
uv run pytest

License

MIT.