dispatch-prs

name	dispatch-prs
description	Loop B of the autonomous Pegasus dev pipeline — babysit every open PR for a claude-labeled issue assigned to you (fix review feedback, resolve conflicts, re-test, clean up on merge), one pass over all of them per wakeup. Designed as a /loop target, runs independently of dispatch-issues (which owns an issue before a PR exists). Use when invoked by /loop, or standalone for a single dry-run pass.
model	sonnet
effort	low
allowed-tools	Bash(gh issue list:*), Bash(gh api:*), Read, Agent

Run ONE pass over every PR currently in play and stop. Stateless — re-derive everything from GitHub/GitLab every pass, never trust conversation memory. This skill runs inside a long-lived /loop session, so the conversation may already contain your own summaries from earlier passes — treat those as history only. Even if you recall a conclusion from a prior pass ("no actionable threads", "waiting on human review"), re-run every detector command below fresh this pass and act only on its current output — a human can act between wakeups (e.g. leave a new review comment), so a conclusion from 10 minutes ago may already be stale. Single writer: exactly one dispatch-prs loop session may run at a time. Runs independently of dispatch-issues — this skill only ever looks at issues that already have a PR.

Unlike issue development, babysitting several PRs in the same pass is safe: each PR has its own worktree, and fixing one doesn't conflict with fixing another.

Pass

1. Scan: gh issue list --label claude:in-progress --assignee @me --state all --json number,createdAt — open or closed. Closed matters here: a merged PR's Closes #<n> auto-closes the issue, usually before this pass ever runs, so filtering to open issues would miss every merge. --assignee @me scopes the whole pass the same way it scopes dispatch-issues — an issue assigned to someone else never becomes a PR this loop touches. For each, check for a linked PR directly via GraphQL, not via a branch lookup:

   gh api graphql -F owner=<owner> -F repo=<repo> -F issue=<n> -f query='
   query($owner:String!,$repo:String!,$issue:Int!){
     repository(owner:$owner,name:$repo){ issue(number:$issue){
       closedByPullRequestsReferences(first:5){ nodes{
         number state isCrossRepository headRepository{ nameWithOwner } } } } } }'
   

   (Confirmed live: once a PR exists, the issue's linkedBranches/gh issue develop --list <n> goes empty — a branch-first lookup can never find it.) No node → no PR yet, belongs to dispatch-issues, skip. babysit-pr derives the branch, mergeability, etc. straight from gh pr view <pr> itself, so the branch name never needs resolving here at all. isCrossRepository == true → fork PR from an external contributor. Never spawn pr-sitter — its branch lives only on the fork, not origin, so worktree/lint/test/review would either fail unpredictably or, if that ever got "fixed" to properly fetch refs/pull/<pr>/head, run the fork's own code (tests, etc.) with this pipeline's full permissions. Fork PRs are explicitly out of scope; include them in the report as "external PR # on # — needs manual triage" so a human notices, rather than silently ignoring them.
2. For every merged, same-repo PR found → spawn a pr-sitter subagent (fresh context) to run babysit-pr <pr> — its cleanup routine removes the worktree/branch and strips labels.
3. For every open, same-repo PR found → spawn a pr-sitter subagent to run babysit-pr <pr>. Do all of them this pass, not just one.
4. Report one line per PR and what it's waiting on next. Pace the next wakeup: short if a fix just landed and needs a re-check soon; longer (30–60 min) if everything is only waiting on human review.