Why it mattersA coding agent that has to write and freeze the test before it writes the fix cannot pass by rewriting the check, and the maintainer sees the evidence before approving anything.
Indada released RepoPilot 1.3.0 on GitHub this week, a self-hosted agent framework that turns a GitHub Issue or a pull-request comment into a test, then asks the OpenAI Codex SDK to write the fix and prove it against that test. The repository was created on 18 September and reached 152 stars and 7 forks within two days.
The pipeline, in the order it runs
RepoPilot pins the commit at both ends of the change and loads what the project's rules file says is allowed. The Codex SDK reads the repository, writes a test plan against the reported failure, and runs the plan against the base commit and the proposed head commit independently. The controller reports what came back from both. Only if the test failed on the base and there is a genuine defect does Codex propose a repair. The repair is checked against the frozen test and the policy. If it passes, the controller opens a draft pull request on an isolated repair branch, and the maintainer reads the evidence and decides whether to merge.
Indada's own line for this is worth quoting: "verification-driven software iteration, powered by the OpenAI Codex SDK. Self-hosted agents turn goals, GitHub Issues and PR feedback into tested, reviewable changes, with bounded execution and human control over merging and deployment."
The bounds it will not cross
The controller enforces five hard limits on any run: a task timeout, a maxCalls budget on API calls, maxAttempts on repair proposals, maxTaskExecutions on retries within a phase, and a maxTokens budget the run tallies as it goes. Retry delays are bounded and the process tree is cleaned up when a task ends, whether the run succeeded, failed or was killed. The agent container reads the OPENAI_API_KEY you supply; the GitHub credentials stay in the controller, so the model never sees them.
The pipeline supports JavaScript and TypeScript through the project's own test runner, Python through pytest, Go through its native test tools, and Java through any framework that emits JUnit XML. Nothing about the framework is Codex-specific beyond the SDK call itself.
Why freezing the test first is the point
An agent that can rewrite the check while it is fixing the code has no meaningful check. RepoPilot's answer is that the failing test is written before the repair is attempted and cannot be edited during the repair, so the pull request either passes the same test the base commit failed or it is rejected without leaving the controller. The maintainer opening the draft PR sees the base failure, the frozen test, the repair diff and the head result on one screen, and can merge or throw the branch away. Deployment is deliberately not automated.
For a team already running Codex on internal tasks, the shift is that the agent stops being asked "please fix this issue" and starts being asked "please write the test that would demonstrate this issue, and then write the fix." The evidence arrives before the diff.
Source
Source: GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.