pr-autopilot: opt-in stages from open PR to green merge — never silent by default
May 6 FelipeOFF Claude Code skill: Reviewer + Author subagents triage every human/bot comment, attribute CI before patching, stop on business rules; --auto is explicit, bare run only opens the PR.
FelipeOFF/pr-autopilot (created 6 May 2026; still shipping features through Sep) is a Claude Code skill that turns the long PR ritual into one orchestrated command — create → review → triage every comment → respond → re-review → resolve conflicts & fix CI → wait for CI → merge — with the important default that every stage is off until you ask. Bare pr-autopilot opens the PR and stops. --review, --resolve, --merge, or --auto turn the factory on deliberately.
The split is Reviewer vs Author. The Reviewer runs parallel code and test tracks (deep maintainability audit plus test-value assessment that will recommend deleting tautologies), posts structured findings, and treats “works” as insufficient for approval. The Author, under --resolve, pulls every comment already on the PR — humans and bots (Copilot, CodeRabbit, Sonar) — classifies critique vs noise vs already-handled, can refute a wrong blocker with evidence, resolves merge conflicts by merging base in (no force-push), and fixes failing CI only after attributing the failure: is the file in your diff, does the same job fail on the base branch or sibling PRs, is it a missing secret? Your breakage gets patched; someone else’s never gets papered over with a skipped test. Two hard stops stay human: business-rule changes route through an explicit confirm skill before the Author ships a decision you did not make, and “this isn’t my PR’s fault” CI triage posts only with permission (once), staying quiet under --auto with a drafted note instead. Verification gates (lint, typecheck, tests) must stay green before any push; state under .pr-autopilot/<PR>/ makes runs resumable. Optional --jev Specgate and stacked --cascade forests are further opt-ins — --auto does not silently enable them.
Why it matters for code review
Continuous verified momentum is not the same as a skill that merges by surprise. pr-autopilot’s receipt is admission control on autonomy: the same lifecycle that drains overnight comment and CI debt only runs as far as the flags allow, and CI attribution plus business-rule stops are the review surfaces that keep volume from becoming silent policy change. FOMO that the PR is still red at 9am is real — so is review debt from an agent that “fixed” a flaky base-branch job by weakening the suite. Prefer an opt-in loop that resumes from disk over another peak session that re-discovers the same threads from scratch.