babysit-pr

Bot comments are checked against the source before they lead to code changes. The agent separates real repository failures from flaky infrastructure and keeps fixes within the PR's intended behavior.

Install the skill with the command above, then name the PR and ask the agent to watch it. For example: "babysit PR 42". The agent selects the skill for requests to monitor, watch, or babysit a PR.

Use it after opening a PR, while CI and automated reviewers run, or when failures and bot findings need attention. Full-send and full-stop use it for this part of their work.

The skill uses T3 Code's PR watcher there, or its bundled pr-wait script elsewhere. One watcher owns each PR; the agent does not create a polling loop. SKILL.md has the host-specific waiting and recovery instructions.

The agent handles existing checks and comments before waiting for new events. It acts only on checks and comments newer than the latest push. When nothing needs action, it stays quiet.

Real findings get fixes. False positives get a written dismissal. Extra tests, unneeded edge-case guards, and small speedups outside the goal become follow-ups with reasons. CodeRabbit findings get the same scrutiny, but the agent never waits for CodeRabbit or holds a merge for it.

Before reporting readiness, the agent checks the base branch, conflicts, and whether the PR is behind. If base commits after the last green run touch the PR's files, it rebases, pushes, and waits for green again.

• Real CI failures and findings resolved, with false positives answered.
• A ready PR or a concrete blocker.
• Declined changes recorded as follow-ups.

Merging requires your explicit request. Immediately before a merge, the agent fetches and rechecks the PR. If another PR makes this one obsolete, it stops and asks before closing it unless closure was already authorized.