juror

Multi-model AI pull request review CLI with cost receipts

TLDR

Review the current branch against a base

$ juror review --base [main]

Review a pull request and print findings

$ juror review --pr [1234] --repo [owner/name]

Review a PR and post the sticky comment

$ juror review --pr [1234] --repo [owner/name] --post

Use a jury preset

$ juror review --preset [balanced] --base [main]

Run a shadow benchmark from a corpus file

$ juror benchmark --file [benchmarks/case.json]

SYNOPSIS

juror review [options]juror benchmark [--file path]

juror (npm package juror-ai) is a multi-model code review tool that runs several frontier models in parallel through their native agent harnesses (Claude Code, Codex, OpenCode, Grok Build, Kimi Code, and generic OpenAI-compatible endpoints), then merges and deduplicates findings into one report with a merge-confidence score and a per-model cost receipt.It is designed for GitHub pull requests and local branch diffs. Typical install is npm i -g juror-ai (binary names juror and juror-ai). The same code path powers the juror-ai/juror GitHub Action, which posts a sticky summary comment and batched inline review comments on PRs.Findings are anchored to the diff, blocked by file/line window, collapsed by exact and similarity/referee stages, and optionally filtered for higher precision via publish_mode: consensus. Provider keys are isolated per harness; models never receive GITHUB_TOKEN.

PARAMETERS

review

Run a multi-model review of a local branch or a GitHub pull request

--base REF

Base ref to diff against for a local working-branch review

--pr NUMBER

Pull request number to review

--repo OWNER/NAME

GitHub repository for --pr reviews

--post

Post (or update) the sticky review comment on the pull request

--preset, --mode NAME

Jury preset: fast (default), balanced, high, or ultra

--models A,B,...

Narrow the selected preset (or custom jury) to these model ids for one run

--config PATH

Path to a Juror config file (default: .juror.yml at the repo root)

--cost-target USD

Per-PR planning cost target in USD

--repo-dir PATH

Local repository directory (used by CI and advanced invocations)

--json PATH

Write structured review output as JSON to PATH

--dry-run

Do not treat the run as a live publish (used with Action dry runs)

benchmark

Evaluate reviewer quality against an adjudicated corpus file

--file PATH

Benchmark corpus JSON file

CONFIGURATION

Repo-root .juror.yml (optional). Keys include preset, models, consensus, review (publishmode, severityfloor, paths_ignore), budget, and output. An explicit models: list replaces the preset entirely.Provider keys in environment or a local .env (loaded automatically; untracked files are not exposed to model checkouts).

CAVEATS

Requires Node.js 20+ and at least one LLM provider API key (JUROR_OPENAI_API_KEY, JUROR_ANTHROPIC_API_KEY, JUROR_XAI_API_KEY, JUROR_FIREWORKS_API_KEY, or unprefixed vendor fallbacks). Actual review quality and cost depend on installed harness CLIs and available keys. Cost for some harnesses is estimated from tokens × list price rather than provider-reported USD. Fork PRs in GitHub Actions are skipped by design when secrets are withheld. Not an autofix bot or linter.

RESOURCES

Source code · Documentation