This article describes a method for automating the maintenance of software forks using AI coding agents, applying it to Cohere's fork of vLLM. The approach compresses the time required to absorb upstream releases from weeks to days by replacing manual intervention with an automated feedback loop.
- The workflow decomposes maintenance into three agent-automatable components: disturbance injection (rebase and conflict resolution), measurement collection (running tests and benchmarks), and a controller (iterative fixing).
- The system uses git rebase to replay custom commits on top of new upstream tags, leveraging rerere for known resolutions.
- An agent monitors test results and iteratively applies fixes until all checks pass, with humans only reviewing the final outcome.
- This method reduces developer attention to intermittent tasks, allowing the fork to stay in sync with upstream improvements automatically.
By automating the feedback loop of syncing, measuring, and fixing, this approach allows teams to absorb upstream features and bug fixes with minimal human intervention.