The agent that wrote the code is a bad reviewer of it. Better to launch a dedicated agent with minimal context and a different harness.
In a single repo: easy.
If multiple repos are involved: still easy 👇
Starting a session off of a problem someone in the org already hit.
search_sessions turns up 13 past sessions and the closest one gets pulled in as context. Its PR was closed unmerged, so what the agent inherits is why that approach was rejected.
- you make a change to a design system component in its repo
- you ask Polygraph to open a PR on the target repo with the change
- Polygraph packages the changes into a tarball & pushes a PR with it
- the target repo runs all CI tests against it
👇
