Introduction
Open-CR-Agent (ocra) is an open-source, multi-agent code review system that favors precision over noise.
Open-CR-Agent (ocra) reviews code changes with a pipeline of deterministic stages around a small number of LLM agents. Code handles what must not go wrong (which files to review, how to group them, where a comment lands); models handle judgment.
Why ocra
- Precision first. Every reviewer states what it must not flag, and findings that fail validation never reach you. A quiet review you trust beats a noisy one you skim.
- Comments land on the right line. Models quote the code they mean; ocra resolves the quote to exact lines itself. Models never make up line numbers.
- Plugins all the way down. VCS platforms, agent runtimes, reviewers, rule packs, tools and listeners are all plugins.
- Any model, with failback. Configure a chain of models per tier; overloads move to the next model, short rate limits are waited out, and a model out of quota is dropped for the run.
- Transparent cost. Every run reports input, output, reasoning and cached tokens and the cost of each attempt.
- Measured, not guessed. Changes are evaluated on the AACR-Bench benchmark.
Status
ocra is in early development.
| Milestone | Scope | Status |
|---|---|---|
| M1 | Local CLI, selection, bundling, anchoring, correctness reviewer, OpenCode runtime, plugins, evaluation | Done |
| M2 | Specialized reviewers, review matrix, verification and judging, risk tiers | Done |
| M3 | GitHub Action with inline comments and incremental re-review | Done; not yet exercised on a live repository |
| M4 | Circuit breakers, remote configuration, long-term review memory, --ultra mode | Done (--ultra without planning and impact analysis) |