The problem
Most AI code reviewers only read the diff, so they guess. That produces confident comments about bugs that don’t exist, and misses ones that do.
What it does
PR URL → fetch diff + metadata
→ git clone at the head commit
→ verify agent (LLM + sandboxed run_command tool) → structured findings
→ reconcile findings to diff lines
→ post inline comments + summary back to the provider- Verifies by execution: the agent runs imports, tests and edge cases before claiming a bug.
- Works with any provider: GitHub, GitLab and Bitbucket adapters.
- Combines deterministic and LLM checks: static analyzers (ruff, pyflakes, codespell…) run alongside the agent.
- Per-language playbooks: review prompts are picked by file type and framework, and are editable.
- Optional Docker sandbox for running untrusted code.