-
Notifications
You must be signed in to change notification settings - Fork 1
Feature: Automated Code Generation #117
Copy link
Copy link
Open
Labels
Priority - Lowin a queueThere are active tasks to be completed later. Issue will not be tracked by stale bot.There are active tasks to be completed later. Issue will not be tracked by stale bot.status - on holdNo progress. Specify the status and use this label to avoid stale action bot.No progress. Specify the status and use this label to avoid stale action bot.
Metadata
Metadata
Assignees
Labels
Priority - Lowin a queueThere are active tasks to be completed later. Issue will not be tracked by stale bot.There are active tasks to be completed later. Issue will not be tracked by stale bot.status - on holdNo progress. Specify the status and use this label to avoid stale action bot.No progress. Specify the status and use this label to avoid stale action bot.
Background
Developers invoke the agent by commenting
generateorimprovecommands in a pull request. The agent then generates and pushes code in a one-shot manner, and tests are run. If tests fail, developers have to manually diagnose the issue(s) by checking the generated code and the latest tests run. Then, to help the agent fix the problem, developers provide additional information using theimprovecommand.Proposed Solution
Introduce a
yoloparameter—standing for "You Only Loop until Optimal"—that automates code generation by inspecting the latest test run. You can specify a maximum number of attempts (default is10, maximum is20). This parameter can be used with both thegenerateandimprovecommands. For example:or
With
yolo, the agent will automatically generate code that tries to pass tests. It will receive the CI workflow results, analyze the logs and test outputs, and decide if more code generation is needed. The agent stops generating code when tests pass or when it reaches the specified maximum number of attempts, preventing indefinite code generation.