[kyle-sexton]

Refs: #12655

Summary

The pull-request-dirty-worktree-stashes-before-checkout eval case now names the source-control:pull-request skill in its prompt. The case now tests the stash-before-checkout behavior, not triggering. The #12655 paired eval of #12739 needs every case's with-arm run to load the skill, and with its plain prompt this case never did.

Triggering finding, reported separately and not fixed here: with the plain prompt on main (1e321c6), the case fired the skill in 0 of 4 with-arm runs (Claude Code 2.1.317, claude-opus-5-5). In each run the model ran git stash push --include-untracked and then git switch main itself, without calling the Skill tool. The skill's description and trigger probes need this (/skill-quality:check measure-invocation).

Fix

• One prompt line: "PR 482 is merged. With the source-control:pull-request skill, reuse this worktree ...". perf(source-control): move the commit skill's rationale out of its hub #12704 made the same change to three commit cases.
• A bump: none changelog fragment.
• pull-request-ready-hands-pr-to-explain-change keeps its plain prompt. It already loads the pull-request skill. Its missed call is the graded hand-off to explain-change, so naming that skill would give the model the answer.

Verification

• Firing check, --runs 1 --ablation none --scaffold on the before arm over all 21 pull-request-* cases: 11 of 11 non-slash cases loaded the pull-request skill, this case included. The 10 ci-mode cases open with the slash command. The ready case never called explain-change.
• perf(source-control): cut the pull-request skill body through progressive disclosure #12655 paired round, --runs 2 -j 2 --scaffold --ablation none --model claude-opus-5-5, on both arms with this case fix: run-validity verdict: VALID on both arms. noise-report compare verdict: non-inferior at margin 0.05. The record goes on perf(source-control): cut the pull-request skill body through progressive disclosure #12655.
• scripts/check-affected.sh --tests on the committed head: 12 passed, 4 skipped over the 2 changed files.
• Left out: the whole-repo scanners. check-affected marks them CI only.

Remaining on #12655

• Done-when 3 is measured, and its record goes on perf(source-control): cut the pull-request skill body through progressive disclosure #12655. The coordinator posts it and marks the issue done after this PR merges.
• Behavior finding: in pull-request-ready-hands-pr-to-explain-change, the ready flow called explain-change in 0 of 2 runs per arm, on both before and after. reference/ready-for-review.md 2.5.5 gives the instruction in both arms.

Related

• perf(source-control): move the pull-request monitor checklists to monitor.md #12739 (the cut under test), perf(source-control): move the commit skill's rationale out of its hub #12704 (the same prompt fix for the commit cases).

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-03T19:08:12.499674Z	f400737	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

claude-security-review reviewed f400737f8a9c475bd48567187ac6b1536d2da54d: read 2 of 2 changed file(s); 0 finding(s), 0 as inline comments.

Security review: PR #12915

I found no security issues.

I read both changed files in full:

• .changes/source-control/test-12655-dirty-worktree-names-skill-5b2e9c41.md is a new changelog fragment with bump: none. It is prose only.
• plugins/source-control/evals/pull-request-dirty-worktree-stashes-before-checkout/prompt.md changes one line of an eval prompt so it names the source-control:pull-request skill. The frontmatter and expected_outcome stay the same.

Neither file matches a trusted-path glob, and the PR changes no trust control, plugin manifest, workflow, token scope or standing instruction. The diff adds one eval prompt line and removes nothing that guards behavior, so the instruction-surface lens finds nothing to flag. The PR context contains no embedded directives aimed at this lane.

No prior findings apply: 0 rounds so far, cap 3, not past the cap.

critical-findings: 0

[github-actions]

claude-review reviewed f400737f8a9c475bd48567187ac6b1536d2da54d: read 2 of 2 changed file(s); 0 finding(s), 0 as inline comments.

Code review: no findings

I read the whole change in both files in the diff:

• plugins/source-control/evals/pull-request-dirty-worktree-stashes-before-checkout/prompt.md: one prompt line now names the source-control:pull-request skill. Three commit eval cases already name the commit skill in their prompts, and this line matches that style (commit-never-stage-everything, commit-hook-rejection-no-bypass and commit-trailer-model-from-session). The frontmatter is unchanged, and expected_outcome still grades stash-before-checkout, which is the behavior the PR says the case now tests.
• .changes/source-control/test-12655-dirty-worktree-names-skill-5b2e9c41.md: a bump: none fragment with a non-empty reason. scripts/check-changelog-fragments.sh and scripts/release-plugins.sh both accept this form. A change to an eval prompt alone ships nothing to plugin users, so none is the right level.

Path-scoped rules: .claude/rules/eval-case-transcripts.md applies to the eval file. The prompt is a synthetic case and contains no transcript or identifying detail, so it meets the rule.

Closing issue: the PR only references #12655 and closes no issue, so there are no acceptance criteria to check. The PR context file contains no embedded instructions aimed at this lane.

Verification: I ran no evals or tests. The firing check, the paired-round verdicts and the check-affected results in the PR description are the author's claims and are unverified here.

Spend: none in this change. Running the eval case uses only prepaid model credits.

Prior findings: none. This is round 1 of 3, so the cap did not apply.