Self-review & checks
When an agent finishes, a second agent reviews the diff for bugs and security issues, and your project's tests run automatically. You get a plain traffic-light verdict — and on anything but green, the issues go straight back to the agent to fix.
Automatic self-review#
With Auto self-review & checks enabled (the default, in Settings → General), every successful run that produced changes triggers a second Claude Code agent. It reads the uncommitted diff and writes a concise, severity-grouped review — and is explicitly told not to make any edits, so it only comments.
Traffic-light verdicts#
The review ends in one of three color-striped verdicts, so you can tell at a glance whether the change is safe:
| Verdict | Meaning |
|---|---|
| 🟢 All good | No correctness or security blockers — review when you're ready. |
| 🟡 Issues | Minor problems worth fixing before merge. |
| 🔴 Problems | Real bugs, failing tests, or security risks. |
The auto-fix loop#
On an amber or red verdict, NeuralCode automatically sends the issues back to the main agent to fix, then re-reviews — capped at two loopsso it can't spin forever. A green verdict stops the loop immediately. If it's still not green after two passes, the run hands back to you with the outstanding findings.
Automatic checks#
Alongside the review, NeuralCode detects and runs your test command:
| Project type | Command run |
|---|---|
Node (with a real test script) | npm test |
Rust (Cargo.toml) | cargo test |
Python (pyproject.toml / pytest.ini) | pytest -q |
Output is captured and surfaced under a Test output toggle, and the pass/fail result feeds the verdict — a failing suite pushes an otherwise-green review to amber or red. Checks have a generous timeout so long suites still complete.