PatchGarden 🌱: I Built an AI Maintenance Engineer That Doesn't Trust Its Own Patches

Description and use

PatchGarden automates investigation, planning, patch generation, verification, and review while keeping developer control and producing auditable repair records.

Problem

Open-source maintainers struggle with failing tests, neglected issues, and maintenance work that never ends, while AI-generated patches may introduce regressions and lack trustworthiness.

Outcome

PatchGarden reduces manual maintenance effort by automating investigation and patch generation while ensuring verification and developer control, producing auditable repair records.

Workflow

• Investigate the failure and gather diagnostic evidence
• Plan a structured repair with hypothesis and verification criteria
• Patch by generating a minimal, reviewable candidate change in an isolated workspace
• Verify by running relevant tests and comparing results to baseline
• Review the diff and evidence for human approval or rejection

Evidence

PatchGarden is designed to help developers investigate repository failures, generate minimal repair proposals, verify changes, and preserve an auditable history of the process.

Source

The architecture separates reasoning from execution: 1. Open-weight AI for investigation 2. A controlled execution layer 3. Independent verification

Source

PatchGarden relies on Temporal durable workflow orchestration paired with persistent SQLite domain state and immutable JSONL audit logs

Source