harlan-zw PR Triage
Turn the owned PR backlog into a ranked merge queue. Never merge.
harlan-github-agent merges pull requests labelled harlan-agent-auto-merge on its own. This skill never does. See auto merge.
Worktree isolation
Before repair edits, follow the worktree isolation contract. It provides the atomic live-agent claim used below.
An existing worktree alone does not prove another agent is active.
wt is the only worktree tool. Never run git worktree add, and never use a harness worktree option such as EnterWorktree or isolation: "worktree". Those write to .claude/worktrees/, which is banned. wt places every worktree at <parent>/<repo>.<branch-slug>.
Discovery stays read only and may use the primary checkout. Every Repair uses a task-owned worktree. Run wt list --format=json. Reuse the task's worktree with wt switch <branch>, or create one with wt switch --create <branch> --base <base>. Read its absolute path from the JSON, then pass that path as workdir to every later command.
Load contracts
Read these completely before discovery:
../adversarial-review/SKILL.md, the complete workflow for one PR.
The adversarial-review skill loads pr, humanize-writing, and unit-tests when required. Do not duplicate those rules here.
Discover
Run scripts/discover-prs.sh from this skill directory.
It returns every human-authored open PR in the allowed base repositories. Exclude GitHub Apps and bot accounts unless the user requests --include-bots.
Skip automated authors before dispatch, review attempts, or comments. Do not post a skip status.
For an author outside configured writable_pr_authors, wait for exact-Revision Review Approval in harlan-github-agent. Do not create a status comment before Approval. Require separate fix Approval after open findings and before edits.
If discovery reaches GitHub's 1,000 result cap, stop before mutations and report incomplete discovery.
Process the backlog
Process one PR at a time. Parallelize read-only GitHub queries when useful.
For each PR, run the complete adversarial-review workflow. A PR is unfinished until its marked bot comment is confirmed on GitHub.
Treat one trusted marked comment for the current head commit as the active review. Never dispatch a second agent for that head commit.
For a personal repository mapped under ~/sites, use take-ownership by convention unless repository policy or the user disables it. Other PRs stop after adversarial review.
Continue to the next PR after a confirmed READY, WAITING, or BLOCKED status. Preserve exact failure evidence.
Finish repairs across the backlog, then revisit WAITING pull requests once. Never sign off a stale SHA.
When required CI also fails on the current base of an owned repository, dispatch one focused baseline repair worker. Open its repair pull request, link it from the original PR status, and revisit the original after merge.
Return the merge queue
List READY pull requests first, sorted by confidence. Then list WAITING and
BLOCKED pull requests without confidence scores.
Confidence | PR | Outcome | Checked | Next action
96/100 high | owner/repo#42 | READY | Base, review, tests, CI | Human merge decision
— | owner/repo#17 | BLOCKED | Review complete; CI failed | Fix Linux job
Include skipped PRs and the exact safety reason. Never merge unless the user separately asks.