docs(pr-flow): make an exhaustive Copilot review loop the step after opening a PR

[cliffhall]

Closes #2464

What changed

• AGENTS.md, Issue-driven Work Style: new rule. After opening a PR, run a Copilot review loop to exhaustion without being asked, weigh each finding against the issue the PR closes, and decline scope expansion. The same scope rule is added to Responding to Code Reviews.
• pr-flow step 7 is now the loop:
  • 7a requests a round (the mutation is unchanged).
  • 7b waits for a posted review or for Copilot's pending request to disappear without one, plus a 25-minute cap. The old wait only watched for a review, so it hung forever when Copilot's session ended without posting.
  • 7c is a stop/continue table. The loop stops on the first clean round, with no confirming round "just to be sure"; on a round holding only out-of-scope findings; or on two consecutive rounds that end without a review. A round counts as clean only when the inline comments, the body headline and the Suppressed comments block are all empty.
• pr-flow step 6 now links the PR to its issue with the addCloseIssueReferences GraphQL mutation. Closing keywords do not link a PR whose base is v2/main, so a v2 card showed nothing in the board's Linked pull requests field. Step 6 also ends with "go straight to step 7". The AGENTS.md note on Closes #N points to the recipe.
• The pr-flow description leads with "create a PR for an issue", and three new eval cases pin that prompt shape and the review-loop request.

Verification

• npm run local:gate: EXIT=0.
• verify:skills: OK, listing share 3,775/4,000.
• skills:eval -- pr-flow, RUNS=4: all 10 cases at 100% under both AGENT=claude and AGENT=copilot, including create a PR for #2463.
• Full Claude skills:eval: 68/74. All six misses are security-advisory and reproduce at 0% on unchanged v2/main, so they are unrelated to this change. Filed as security-advisory skill never auto-loads under Claude: all six skills:eval positives score 0% (Copilot: 100%) #2465, which is Claude-only; Copilot passes those cases.

Verified live on this PR

• 7b's pending-request check: while Copilot's review was pending, reviewRequests returned {"__typename":"Bot","login":"copilot-pull-request-reviewer"}, which the query matches. Round 1 resolved as ROUND=posted.
• Linking: with the manual link removed first, addCloseIssueReferences alone put docs(pr-flow): make an exhaustive Copilot review loop the step after opening a PR #2466 in pr-flow: make an exhaustive Copilot review loop the default step after opening a PR #2464's closingIssuesReferences and in the card's Linked pull requests field on board Add tab and approval flow for server -> client sampling #28.

🤖 Generated with Claude Code

[Copilot]

Copilot review overview

🟡 Changes recommended

A timed-out pending request cannot be retried as a distinct round using the documented mutation.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 1

Open (1)

• Retrying timed-out review never starts a new request · New

What changed in this PR

Adds a mandatory post-creation Copilot review loop to the repository’s PR workflow.

Changes:

• Documents review-loop and scope-control rules.
• Adds request, wait, retry, and exit procedures.
• Expands skill-trigger evaluation cases.

File	Description
AGENTS.md	Establishes review-loop rules.
.claude/​skills/​pr-flow/​SKILL.md	Documents the operational workflow.
.claude/​skills/​pr-flow/​evals/​evals.json	Adds trigger cases.

[cliffhall]

Copilot round 1: 1 finding, fixed.

• Retrying a timed-out review never starts a new request (pr-flow 7c). Fixed in 3be44dc. timed-out now stops and reports the round as still pending; only ended-without-review is retried.

Also in this push (2623167), at the maintainer's request: pr-flow step 6 now links the PR to its issue with the addCloseIssueReferences GraphQL mutation. Closing keywords do not link a PR whose base is v2/main, so a v2 card showed no linked PR. Verified on this PR: with the manual link removed first, the mutation alone put #2466 in #2464's Linked pull requests field on board #28. The matching AGENTS.md note is updated too.

local:gate: EXIT=0. Requesting round 2.

[Copilot]

Copilot review overview

🟢 Approval recommended

The documented workflow is coherent, complete, and consistent with issue #2464.

Review effort: Balanced
Findings: None

Resolved since last review (1)

• Retrying timed-out review never starts a new request

[cliffhall]

Copilot round 2: clean. Review loop closed.

No inline comments, no finding in the headline ("coherent, complete, and consistent with issue #2464"), and no Suppressed comments block. Round 1's finding is listed as resolved. Under the new 7c a clean round is the stopping point, so no further round is requested. The loop ran as documented: two rounds, one fix, and both waits resolved as ROUND=posted.