[kyle-sexton]

Closes #11423

Summary

The last three items on #11423. /user-interface:design now auto-invokes on a question about a single visual value, such as a nested element's corner radius or a layer's shadow, even when the question gives every number. Before this change the model answered those questions itself and loaded no skill, which is why shape-and-depth-nested-radius fired no skill on base or head. The probe sets are re-measured on current main with the four probes reworded after #11765. /user-interface:write-css gets one added trigger, chosen by a pass rule written before any result was read.

Fix

• plugins/user-interface/skills/design/SKILL.md: "to decide how something should look" now reads "to decide how something should look, down to one corner radius or shadow, even with every number given".
• plugins/user-interface/skills/write-css/SKILL.md: adds the quoted trigger 'fix this layout in the stylesheet'.
• plugins/user-interface/probes/design.json: adds three shape probes (pos-train-shape-01, pos-train-shape-02, pos-validation-shape-01); one of them gives every number.
• .changes/user-interface/fix-11423-shape-and-css-triggers-7d3a91c2.md.

Verification

Round 4 chose these two descriptions (design A3 and write-css B2) under a pass rule written before any result was read, but run-validity.py judged every round-4 probe side INVALID, so its numbers count for nothing. Round 5 re-measured the same two descriptions on runs the gate judges VALID, under a decision rule written before any round-5 run started.

Round 5 setup. claude plugin eval, 3 runs per case, opus judge, --ablation none, Claude Code 2.1.316, both sides concurrently. Base is plugins/user-interface at main c3a4498. Head is base with only the two description: lines replaced by this PR's. Probe cases: main's design.json and write-css.json, the three shape probes above, and a held-out set of 6 design and 6 write-css probes (written by a fresh agent from neutral purpose statements, not committed), emitted by measure-invocation.sh emit-plugin-eval --runs 3 and run with --disallow-tools Read Glob Grep Agent.

Gate. run-validity.py --runs 3: base probes VALID, head probes VALID (297 of 297 traces read, no errored runs, no permission denials).

group	base	head	change
design positives, main's	60/78 (0.77)	67/78 (0.86)	+9.0 points
shape positives (visible + held-out)	11/21 (0.52)	21/21 (1.00)	+47.6 points
design positives, all	71/99 (0.72)	88/99 (0.89)	+17.2 points
design negatives fired	9/60	9/60	0
write-css positives, main's	59/66 (0.89)	60/66 (0.91)	+1.5 points
write-css positives, all	71/78 (0.91)	72/78 (0.92)	+1.3 points
write-css negatives fired	0/60	0/60	0
CSS-only design near-misses fired	0/9	0/9	0

noise-report.py <head> --baseline <base> --margin 0.05: paired change +0.06 over 99 cases (95% interval +0.02 to +0.10), non-inferior.

shape-and-depth-nested-radius (the case the design change targets; 3 runs per side, no errored runs, no denials):

side	design fired	answer grader (border-subtracted) passed
base	0/3	3/3
head	3/3	3/3

The gate marks the named-four result files INVALID on both sides for reasons outside this measurement: its skill-fired check counts base not firing on this case, which is the behavior the change fixes, as an invalid run; base also had one denied plugin-root Glob in states-projects-list, and head one unfired accessibility-mobile-nav-drawer run. The owner decided to judge this case on its runs with no error or denial, counting unfired runs as data.

Outcome against the round-5 rule.

• Design (A3): all five conditions pass. Nested-radius fires in 3/3 head runs with no answer drop; main's design positives +9.0 points; design negatives +0; write-css main positives +1.5 points; shape positives rise.
• write-css (B2): all three conditions pass. write-css positives +1.3 points; design positives +17.2 points (with design conditions 1 and 5); write-css negatives +0.

measure-invocation.sh validate passes on the four probe files. check-skill.sh passes design and write-css.

Related

• fix(user-interface): trigger design on build-a-component requests #11765 (design's build-a-component triggers, the first part of fix(user-interface): design and write-css rarely fire on build-a-component prompts #11423)
• feat: land the rest of the UI course migration: interface design guidance, build-generator and eval cases #11169 (the course migration that added the first probe sets)
• fix(evals): a deep TMPDIR stops the eval sandbox from starting, and the run still scores #11740 (long TMPDIR paths break claude plugin eval)

🤖 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 2026-11-01T07:28:20.337961Z	6798385	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 67983857dfc952f32b8f45ea5672d88981bfec0e: read 4 of 4 changed file(s); 0 finding(s), 0 as inline comments.

Security review: PR #11895

No security issues found.

I read every hunk in all 4 changed files:

• .changes/user-interface/fix-11423-shape-and-css-triggers-7d3a91c2.md: a new changelog fragment containing only prose.
• plugins/user-interface/probes/design.json: an extended ranking_note string and three new should-trigger eval probes. These are test data for invocation measurement and are never executed.
• plugins/user-interface/skills/design/SKILL.md: the frontmatter description gains "down to one corner radius or shadow, even with every number given". No other frontmatter key changes. user-invocable and disable-model-invocation stay the same, and no tool grant changes.
• plugins/user-interface/skills/write-css/SKILL.md: the frontmatter description gains one quoted trigger, 'fix this layout in the stylesheet'. No other key changes.

What I checked:

• Trusted paths: none of the changed files match a trusted-path glob.
• Plugin manifests: no .claude-plugin/plugin.json or marketplace.json changes, so there are no manifest lines to report.
• Instruction surface: the two description edits widen when each skill auto-triggers. They delete, narrow or soften no guardrail or standing rule, so the instruction-surface lens does not apply.
• Trust controls and credentials: the diff touches no trust control, token scope, kill switch or merge hold. It touches no workflow, script or credential handling.
• PR context file: it contains no embedded directives aimed at this lane.

Spend: none beyond normal trigger behaviour. The changed descriptions may make the design skill auto-load slightly more often, which only uses subscription usage.

critical-findings: 0

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 67983857df

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".

[github-actions]

claude-review reviewed 67983857dfc952f32b8f45ea5672d88981bfec0e: read 4 of 4 changed file(s); 0 finding(s), 0 as inline comments.

Code review: PR #11895, no findings

I read every hunk of all 4 changed files:

• the changelog fragment
• probes/design.json
• skills/design/SKILL.md
• skills/write-css/SKILL.md

None of them matches a trusted-path glob.

What changed. The PR changes two skill descriptions. In design, "to decide how something should look" becomes "…down to one corner radius or shadow, even with every number given". write-css gains the trigger 'fix this layout in the stylesheet'. The PR also adds three shape probes to probes/design.json and a bump: patch changelog fragment.

Checks

• Probes. The new ids (pos-train-shape-01, pos-train-shape-02, pos-validation-shape-01) are unique in the file. Their split and expect_trigger values match the file's convention. pos-train-shape-01 gives every number (12px corners, 8px inset), which the updated ranking_note says one shape probe does.
• Description length. I estimate the design description at roughly 970 characters, still under the common 1024-character skill-description cap. The author says check-skill.sh passes. I could not run it, so that is unverified.
• Path-scoped rules. .claude/rules/skill-bodies-state-current-rules.md and .claude/rules/cost-claims.md cover the SKILL.md files. Neither is broken: the new text states current triggers, and it makes no cost claims.
• Plugin manifests. The PR changes no plugin.json or marketplace.json. The fragment's patch bump suits a trigger-wording fix.

Closing issue #11423. The reopen comment lists three open items:

1. shape-and-depth-nested-radius should route. The design description now covers questions that give every number, and a committed probe (pos-train-shape-01) would show a regression. The author reports the case fired design 0/3 on base and 3/3 on head. I did not run this, so it is unverified.
2. Re-measure the probe sets. This is measurement only, with no code surface. The PR body reports results; I did not run them.
3. Decide write-css's description. The PR rewrites it, and the existing probes/write-css.json positives are where a drop past the 5-point bar would show. The author reports 57/66 on both base and head for main's positives. I did not run this either.

Embedded text. The PR context contains no instructions aimed at the review lane.

Spend. None. The change only alters model-invocation wording and probe data.

[kyle-sexton]

Arm request (Emil lane, for the hub): #11895 is ready to arm once ci-status re-runs green on head 6798385. do-not-merge is removed, the Codex P1 thread is answered with round-5 VALID probe runs and resolved (0 open threads), both review lanes posted for this head, and the body cites round 5. Owner decision (2026-11-01, in session): the nested-radius condition is judged on runs with no error or denial, counting unfired runs as data, because the gate's skill-fired check rejects base not firing on that case, which is the behavior this change fixes.