I've often found that agents do stupid things when implementing requests. When I only see their changes with limited context in a scroll-by approval blob, it's easy to miss some of that stupidity. In those cases, I'll often only see it when I can review the full change in context in a PR.
This tool adds a local review step, giving the human a final opportunity to be in the loop: your agent commits the changes, pushes to a local 'review' remote, and waits for feedback. The push itself opens a review via post-receive hook, and provides you with a localhost URL to review the changes.
If you request changes via inline comments (like you would for a PR), the agent sees the comments and makes adjustments. It pushes again and the process repeats until you're satisfied.
When you approve the change, your agent can take whatever step is next in your normal workflow - push to open an actual PR, merge locally, etc. These are lightweight instructions (~500 tokens) that should be customized for your workflow.
This tool adds a local review step, giving the human a final opportunity to be in the loop: your agent commits the changes, pushes to a local 'review' remote, and waits for feedback. The push itself opens a review via post-receive hook, and provides you with a localhost URL to review the changes.
If you request changes via inline comments (like you would for a PR), the agent sees the comments and makes adjustments. It pushes again and the process repeats until you're satisfied.
When you approve the change, your agent can take whatever step is next in your normal workflow - push to open an actual PR, merge locally, etc. These are lightweight instructions (~500 tokens) that should be customized for your workflow.