Coding-agent skills for pull-request work and orchestration: an adversarial multi-pass PR reviewer, PR description and title writing, answering review comments, CI failure triage and PR nursing, a DAG orchestrator that delegates to cheaper subagents, and file-channel coordination between agents. Supports GitHub and GitLab.
95%
Does it follow best practices?
Impact
98%
1.88xAverage score across 10 eval scenarios
Low
Low-risk findings worth noting
Fallback for any agent without a dedicated file here.
With a shell, gh and glab are the canonical tools and the skill runs as written. With only a host MCP server, use it for Step 2 metadata (head and base SHA must come from one response) and for linked issues; Step 4's clone still needs a shell. With neither, the user must check out the PR head and paste the metadata; the skill then verifies HEAD == headRefOid and runs the "current branch" path. The skill never manufactures data.
| Tier | Requirement | Used for |
|---|---|---|
| Hard | the strongest model available for a sub-task | assessor; escalated reviewer |
| Medium | a mid-cost model | reviewers A, B, C |
| Easy | the cheapest | summary |
With one model only, record it for every pass; the lenses, the assessor's steel-man and the state file still apply.
If the runtime can start isolated sub-tasks with their own prompt (and ideally their own model), each reviewer prompt from references/passes.md is one sub-task, dispatched together, with no reviewer's output in another's context. The assessor and summary are two more sub-tasks, in sequence. If it cannot, run single mode and record mode: single (fallback: no isolated sub-agent primitive) in findings.md so the report says so. Do not emulate independence by role-playing three reviewers in one context: that is one reviewer with three hats, and the report would overstate it.
Most runtimes have no loop primitive (Claude Code is the exception). In increasing order of automation:
/pr-reviewer <num> after each round until converged: true or --rounds=N./pr-reviewer <num> --rounds=3 repeatedly until findings.md records converged: true for the current revision."for i in $(seq 1 3); do
<agent-invocation> "/pr-reviewer <num> --rounds=3"
awk '/^## Revision/{c=0} /^- converged: true$/{c=1} END{exit !c}' "<base-dir>/findings.md" && break
doneThe awk resets at every ## Revision header, so only the latest revision's convergence stops the loop. Always pass --rounds=N; an invocation that finds the cap already reached renders the report without running a pass.
Step 11 of the skill, after the report, with explicit confirmation, through whatever the runtime offers. Without a posting path, print the report and let the user post it.