PR Watch

PR Watch is an experiment in using agent orchestrators to watch pull requests and help authors refine their work. I built the same two workflows on two runtimes: Gas City and CAO.

  flowchart LR
    D[Discover eligible PRs] --> G{Safety gate}
    G -->|review request| R[Fresh reviewer session]
    G -->|owned PR| B[Long-lived babysitter session]
    R --> V[Validated review or local draft]
    B --> F[Fix CI and review feedback]

Rather than using an agent to orchestrate, the workflows programmatically search pull requests and react to them by setting up tooling, bare clones, and additional context. The first workflow finds pull requests that need my review and starts a clean review for the current head revision. The second watches my open pull requests with agent instructions to keep CI green, handle review feedback, and request approval from an optional reviewer.

CAO has been the better fit. Its project-local server, profiles, sessions, work directories, and dashboard make the long-running agent model easier to reason about and inspect. The setup is also a straightforward interface for a task runner instead of a collection of custom process glue.

The automation still has hard limits. Pull-request titles, diffs, comments, logs, and linked content are untrusted. Reviewers only publish validated reviews. Babysitters can make small verified fixes and push to the PR branch, but cannot merge, force-push, dismiss reviews, or approve for someone else.

In CAO, you have workflows that inherit an environment with agents running in OpenCode, in a terminal multiplexer like Herdr, between python scripts, in a web dashboard. It sounds like spaghetti but it actually comes together very nicely.