[kyle-sexton]

Closes #7129

Summary

Kept-trace reruns of the two write-css rule cases that scored 0/5 in every arm found two grader defects. In css-rule-logical, no-physical failed 5/5 runs whose CSS was fully logical: every reply ended with a right-to-left note naming the physical alternative in inline code (`text-align: left`), and the pattern matched it. The issue's guess (top: 0 on a positioned element) was not the cause. In css-rule-ease-in, all 5 runs routed to user-interface:write-animation, so the skill-fired grader, which named only write-css, failed every run.

Fix

• no-physical no longer counts a physical property inside inline code, whether a bare declaration or a quoted whole rule; each alternative skips a match with a closing backtick later on the line. Physical declarations in CSS still fail it, including in a fenced block. Must-pass samples cover both prose forms.
• The ease-in skill-fired grader accepts write-css or write-animation. Both carry the no-ease-in rule, and an animation request routes to write-animation by design (its Step 7 hands CSS to write-css).
• Implements owner decision Q105 (2026-10-24, recorded on test(user-interface): css-rule-logical and css-rule-ease-in fail 0/5 in every arm #7129: test(user-interface): css-rule-logical and css-rule-ease-in fail 0/5 in every arm #7129 (comment)): an explicit request for an exit that accelerates away wins over the ease-in rule. write-css (SKILL.md, reference/motion.md) and the animation catalog say to use ease-in for that exit and warn that its slow start can read as lag. write-css also says the review script and audit-animation still report that ease-in, and that the finding is expected. The default is unchanged.
• The css-rule-ease-in eval follows: no-ease-in and its samples are removed, and fast-start-curve becomes exit-request-honored, which passes an ease-out curve, or an honored request with the lag warning. The judge has four labeled samples.

Verification

• Claude Code 2.1.310, claude plugin eval <copy>/plugins/user-interface --case <case> --ablation none --keep-temp --judge-model opus --threshold 0.8, 5 runs, run-validity verdict VALID:
  • css-rule-logical: before, no-physical failed 5/5 (score 0.67); after, 5/5 (score 1.00).
  • css-rule-ease-in on the final skill text: 5/5 (score 1.00), every grader passing. A run on the earlier wording ("say it starts slow") scored 4/5. The miss called the slow start a feature and never named the lag, so the skill text now asks for the lag warning.
• exit-request-honored judge calibration (calibrate-judge.py, opus, 3 runs per sample): 12/12 agreement over 4 samples. Three independent labelers agreed with every sample label.
• Offline: validate-cases.py plugins/user-interface/evals exits 0, checking each deterministic grader against its samples; the quoted-rule sample failed the old pattern first. validate-plugins.sh --only user-interface, check-changelog-fragments.sh --check and check-affected.sh --tests pass.

Related

• test(user-interface): run the write-css three-arm rule eval and settle which rules stay in the body #7054: the three-arm write-css eval that surfaced these two cases.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X1zWtmVHsCdqJk7AEeNE2b

[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 2026-10-25T07:05:13.402182Z	740dd8a	Draft marked ready

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 3173c75d82864fd1a94ff0da252ecf814cbf2632: 0 finding(s), 0 as inline comments.

No security issues in the incremental change (8 files: eval prompt, grader, and samples; write-css and animation reference prose; a changelog fragment).

• No manifests, workflows, hooks, scripts, trusted paths, or userConfig keys are touched, so no manifest <field> lines apply.
• Instruction-surface lens: the ease-in wording changes from "say it starts slow" to "warn that its slow start can read as lag", and adds a note that the review script and audit-animation still report that finding. This deletes or softens no control. The ease-in rule is a UI-quality convention, not a protected class. The explicit-request exception was already in the earlier reviewed head. I did not read the instruction exception register, so I judged this by consequence.
• Grader regex: the text-align alternative of no-physical now ignores a match followed by a backtick on the same line, as the other alternatives already did. Real physical declarations in fenced CSS still fail. This is an eval-quality matter with no exploit path.
• PR context contains no embedded instructions aimed at the review.
• Spend: eval reruns draw only on prepaid API credit, so there is no hold.

critical-findings: 0

[github-actions]

claude-review: 2 inline finding(s) on 740dd8aa6957d45dc7481934130c2c91c9773b70; the summary is in this lane's pull request comment.

claude-lane: claude-review

[github-actions]

claude-review reviewed 3173c75d82864fd1a94ff0da252ecf814cbf2632: 0 finding(s), 0 as inline comments.

Incremental review of the 8 files changed since 740dd8a: no findings.

• Wording change. The reply must now "warn that its slow start can read as lag" instead of "name that it starts slow". The change fragment, prompt.md, catalog.md, write-css/SKILL.md and reference/motion.md all use the new wording.
• Ease-in judge. graders/exit-request-honored.md was not touched. It still accepts "starts slow or reads as lag", which is slightly looser than the new expected_outcome but does not contradict it.
• Lint note. The new SKILL.md text says the review script's ease-in finding on an accelerating exit is expected and to keep the curve. This matches the audit-animation note.
• no-physical regex. The text-align alternative now uses (?![^\n]*) instead of (?!)`, which matches the other two alternatives.
  • I traced the new sample `.rtl { text-align: left; }`. The lookahead sees the closing backtick on the same line, so it does not match, and the pass sample holds.
  • The existing fail samples (.bubble { text-align: left; }, the margin and offset ones) contain no backtick, so they still match and still fail.
  • A physical declaration in a fenced block on its own line still fails. A one-line reply that puts CSS and a later inline-code span on the same line would slip through. That is an unlikely shape and not worth blocking.
• Not run. I read the files and traced the regex by hand. I did not run validate-cases.py or any live eval. The PR's verification claims are the author's and unverified here. The PR itself says the opus judge for exit-request-honored is uncalibrated and was not re-run live.
• Manifests. The diff changes no plugin.json or marketplace.json, so there are no manifest lines. The fragment's bump: patch fits an eval and docs-wording change.
• Spend. The change is eval and doc content only. It adds no billing or paid-plan change, so there is no hold.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 740dd8aa69

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".