auto-pr

Invoke the skill with "/auto-pr <PR number>" on a task bound to the repository. It runs as a self-paced goal: each iteration reads the gate verdict, classifies or fixes findings, schedules its own continuation, and squash-merges once zero blocking findings remain on reviewed code.

Each review thread is classified as BLOCKING, NIT, or WRONG. Only blocking findings — real defects like data loss, races, or broken contracts — get fixed; nits and wrong findings get a marked reply and are resolved without a commit.

A reviewer handed a diff always produces new findings, so waiting for zero findings never terminates above roughly one finding per push. This skill exits on zero unresolved blocking findings instead, and enforces a prose freeze, one push per iteration, and a round budget.

Yes. The gate detects bot logins with substring matching, treats bot CHANGES_REQUESTED as classifiable findings rather than human blocks, and handles shared-login setups where the reviewing bot and replying account are the same identity.

It escalates on merge conflicts, a human reviewer's CHANGES_REQUESTED, an unreviewable PR that edits the review workflow, or a spent round budget with blocking findings or a still-red check. Escalation stops the goal and reports which condition fired.

It requires git, the gh CLI, and python3, with network access to github.com and api.github.com. jq is deliberately not used; JSON parsing goes through gh's built-in --jq flag or python3.