AI code improvement tool
AI Steward
An improvement loop that proposes and verifies code changes while leaving the final commit decision with the operator.
A tool and a reference implementation
AI Steward connects the practical problem of improving a codebase with the governance question of accountable delegation. The operator writes a destination, the loop scans and proposes an improvement, and the result is verified and staged for review. The documented workflow ends with the human choosing to commit or discard, rather than treating model confidence as authorization to merge.
Separate model work from deterministic checks
The pipeline distinguishes pre-flight checks, scanning, implementation, verification and recording. Not every operation needs an LLM call. Repository state, syntax and tests can be checked through ordinary tooling, while model calls handle the interpretation and proposal work. This makes the improvement cycle easier to inspect than one undifferentiated request to “make the code better.”
Two kinds of record
The semantic audit trail explains the agent’s decision; the harness-captured session records model-layer interactions outside that narration. They are complementary, not interchangeable. A staged diff and test results supply further evidence of repository behavior. Linking those surfaces is important because a persuasive account of a change can exist even when a tool failed or the resulting code does not do what was intended.
Scope of the public evidence
The repository documents its versioned status, prerequisites and continuing work. This overview does not turn historical test counts, model prices or a successful self-targeting run into current guarantees for every repository. Test selection, protected paths, runtime isolation and the quality of the operator’s destination still matter. The project is a concrete implementation to examine and run under appropriate controls, not proof that autonomous software maintenance is solved.