# The Safe Fix Pipeline — Safe Autonomous Code Fixing

Source: https://www.skillbyai.com/en/safe-autonomous-code-fixing/i-pipeline

> Reproduce, localise, patch, verify, review.

## Stages with gates

A safe pipeline has explicit stages: (1) **reproduce** the bug with a failing test; (2) **localise** the cause using traces, logs or bisect; (3) **generate** one or more minimal candidate patches in a sandbox; (4) **verify** each candidate in isolation: the new test passes, the full suite passes, nothing protected is touched, no secrets, small diff, static checks pass; (5) open a **pull request** with evidence for **human review**; (6) **monitor** after merge and revert quickly if needed. Each gate can stop the run.

## The pipeline with its gates

Any failed gate stops the run.

```text
issue / failing CI
 -> reproduce: failing test committed to a branch        [gate: test fails before fix]
 -> localise: trace, logs, git bisect                   [gate: suspect files identified]
 -> patch: candidates in a sandbox                      [gate: diff size, allowed files]
 -> verify: new test + full suite + lint/types          [gate: all green, no tampering, no secrets]
 -> PR with evidence -> human review                    [gate: approval]
 -> merge -> monitor -> revert if needed
```

## Make the failing test the contract

If you cannot write a test that fails before and passes after, the problem is not ready for automated fixing.

**Quiz:** What should come first in a safe fix pipeline?

- [ ] Rewriting the module
- [ ] Merging a patch
- [ ] Deleting old tests
- [x] Reproducing the bug with a failing test

*Answer:* Reproducing the bug with a failing test. Reproduce before fixing.