Problem
After opening a PR, the Copilot review cycle is run only when the prompt spells it out ("request a review, wait for it, respond, request again until there are no more comments, resist scope creep"). Nothing in AGENTS.md or the pr-flow skill makes it the default next action, so a bare "create a PR for #N" stops at the open PR.
The skill's wait loop also only watches for a posted review. When Copilot's session ends without posting (it fails, or occasionally has nothing to say), the loop waits forever.
Two habits make the cycle longer than it needs to be:
- agents request a confirming round after a clean one "just to be sure", which spends Copilot tokens re-reviewing unchanged code;
- agents accept findings that expand the PR beyond its issue, and every such fix is fresh surface for the next round.
Proposed change
AGENTS.md (Issue-driven Work Style): a rule that opening a PR is followed, unprompted, by a Copilot review loop run to exhaustion, with findings weighed against the issue and scope expansion declined. The same scope rule is added to "Responding to Code Reviews".pr-flow step 7 becomes the loop: 7a request, 7b wait for a posted review or for the pending request to disappear without one (plus a 25-minute cap), 7c a stop/continue table. The loop stops on the first clean round (clean on inline comments, the body headline and the Suppressed comments block), on a round holding only out-of-scope findings, or on two consecutive rounds that end without a review.- The
pr-flow description leads with "create a PR for an issue", and eval cases pin that exact prompt shape.
Verification
skills:eval for pr-flow, 4 runs each: every case at 100% under both AGENT=claude and AGENT=copilot, including create a PR for #2463. Full Claude eval: 68/74. The six misses are all security-advisory, and they reproduce at 0% on unchanged v2/main, so they come from somewhere else.
Problem
After opening a PR, the Copilot review cycle is run only when the prompt spells it out ("request a review, wait for it, respond, request again until there are no more comments, resist scope creep"). Nothing in
AGENTS.mdor thepr-flowskill makes it the default next action, so a bare "create a PR for #N" stops at the open PR.The skill's wait loop also only watches for a posted review. When Copilot's session ends without posting (it fails, or occasionally has nothing to say), the loop waits forever.
Two habits make the cycle longer than it needs to be:
Proposed change
AGENTS.md(Issue-driven Work Style): a rule that opening a PR is followed, unprompted, by a Copilot review loop run to exhaustion, with findings weighed against the issue and scope expansion declined. The same scope rule is added to "Responding to Code Reviews".pr-flowstep 7 becomes the loop: 7a request, 7b wait for a posted review or for the pending request to disappear without one (plus a 25-minute cap), 7c a stop/continue table. The loop stops on the first clean round (clean on inline comments, the body headline and theSuppressed commentsblock), on a round holding only out-of-scope findings, or on two consecutive rounds that end without a review.pr-flowdescription leads with "create a PR for an issue", and eval cases pin that exact prompt shape.Verification
skills:evalforpr-flow, 4 runs each: every case at 100% under bothAGENT=claudeandAGENT=copilot, includingcreate a PR for #2463. Full Claude eval: 68/74. The six misses are allsecurity-advisory, and they reproduce at 0% on unchangedv2/main, so they come from somewhere else.