Shepherd a factory pull request

Answer the feedback on one existing factory pull request: Tessl Code Review findings, comments from people, and failed checks. Push fixes to the same branch. Never open a new branch or a new pull request.

Inputs

• The triggering events at .cloud-launch/trigger-events.json. Several events can be folded into one run; read them all and require that they name one pull request on a factory/ branch.
• An environment with githubAccess: write, for pushes, replies and resolving review threads.

Procedure

1. Count attempts. Every top-level comment this skill posts ends with <!-- perihelion-factory:shepherd -->. If the pull request already has three, add factory:needs-human, comment that the budget is spent with a summary of what is still open, and stop. Tessl has no run budget of its own, so this count is the only thing that stops a loop between review and fix.
2. Gather the current state. Fetch the pull request, its unresolved review threads, its conversation and its check runs. Work from that state, not only from the events; earlier feedback may still be open.
3. Read the standards. Read the AGENTS.md or CLAUDE.md chain and the team standards, as the review lenses do. A comment that cites a standard is judged against the standard's text.
4. Resolve the feedback. Follow this plugin's resolve-feedback skill for each unresolved item: verify it against the code, fix what is right, and decline what is wrong with a reason. For a failed check, read its log, reproduce it locally, and fix the cause the change introduced.
5. Verify and push. Run the build, the tests and tessl change verify where configured, commit, and push to the existing branch.
6. Reply. Reply in each thread you acted on, saying what changed or why nothing did, and resolve the threads you fixed. Post one top-level summary ending with the marker from step 1.

Stop and hand back

Add factory:needs-human and say why, without pushing, when feedback items contradict each other, a fix would go beyond what the issue asked for, or a failed check is infrastructure rather than code. Approving and merging always stay with a person.