[kyle-sexton]

Closes #12491
Siblings: #13113

Summary

Records what the description controls showed on the three code-metrics audit leaves, in invocation-probes.md "Description controls", and keeps the certification rule as it is. With the name kept and a generic description, should-trigger probes fired 81 of 81; with the skill's folder renamed too, 0 of 81, on all-VALID passes (result). So the name keeps these controls firing on Opus, and these three descriptions stay as they are.

This closes #12491 under its own done-when, amended by the owner to three leaves plus the router: "A description that fails keeps its main text", and each did.

Fix

• One table and one paragraph in plugins/skill-quality/reference/invocation-probes.md, plus a bump: none fragment.
• The rule does not change. Basis: a renaming control removes the skill from view, so it falls all or nothing and could never fail a description cut, and naming:name-it-better shows the description control can still drop on Sonnet 5.5. Accepting it would make the control always pass.

Verification

• I read the first result and the name-control result on the issue; the table copies their counts. The runs themselves are a worker's, and I did not re-score them.
• The rest of refactor(code-metrics): give the six audit-* skills distinct, trimmed descriptions #12491's done-when is met on main: distinct first clauses and /skill-quality:check on the before text, probes in plugins/code-metrics/probes, and three leaves plus the router.
• Left out: a Sonnet run of these three controls and the map skills under the renaming control, since neither would change this decision.

Related

• test(architecture): give four map-* skills probes a description cut can fail, then trim them #13050, closed not planned: the same result on four architecture:map-* skills.
• PR 13113: the probes.

🤖 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 2026-11-04T14:06:13.609583Z	da1ce77	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-review reviewed da1ce77d29235840cc4f6351bb00adcb91def152: read 2 of 2 changed file(s); 0 finding(s), 0 as inline comments.

Code review: no findings. This was a complete review: I read every hunk of both changed files.

The PR changes only documentation:

• plugins/skill-quality/reference/invocation-probes.md ("Description controls"): it adds a table and a paragraph recording the name-and-description control on the three code-metrics audit leaves.
• .changes/skill-quality/docs-12491-description-control-record-a7c41e93.md: a bump: none changelog fragment.

What I checked

• Table counts: the rows (27/27 / 27/27 / 0/27 for coverage, duplication and suppressions) match the two result comments on refactor(code-metrics): give the six audit-* skills distinct, trimmed descriptions #12491 in the PR context. Those runs are the author's. I did not re-run or re-score them.
• Run conditions: the prose matches the issue's result comments. The runs used the same version and model as the architecture table above (Claude Code 2.1.318, Opus 5.5), with 9 should-trigger probes at 3 runs each and all passes VALID.
• Certification rule: it is unchanged, as the PR says. The new paragraph matches the issue's own limits on the renaming control: it shows the probe set can fail, but it says nothing about how much the description contributes.
• Release fragment: bump: none fits a reference-doc-only change.
• Path rules: .claude/rules/cost-claims.md matches plugins/*/reference/**. The added text makes no cost claims.

Closing issue #12491 (done-when)

• Probes: plugins/code-metrics/probes/ holds the three probe files.
• Skill layout: plugins/code-metrics/skills/ has three audit-* leaves plus the audit router, which matches the amended count of 3.
• Descriptions kept: the "A description that fails keeps its main text" criterion holds, because no description changes in this PR.
• Not verified by me: the claims that the first clauses are distinct, that /skill-quality:check passes, that the meaning diff passes and that scripts/validate-plugins.sh passes. Those are the author's claims, based on work already on main.

This PR adds no testable behavior.

Embedded text and spend: none of the input files contains embedded instructions to this lane. The PR spends no new API credit or plan usage; the runs it records were done before it.

[github-actions]

claude-security-review reviewed da1ce77d29235840cc4f6351bb00adcb91def152: read 2 of 2 changed file(s); 0 finding(s), 0 as inline comments.

Security review: PR #13251 (head da1ce77d)

No security issues found.

Scope. Both changed files were read in full:

• .changes/skill-quality/docs-12491-description-control-record-a7c41e93.md is a new bump: none changelog fragment.
• plugins/skill-quality/reference/invocation-probes.md adds one table and one paragraph to "Description controls". They record the eval results for the name-and-description control on three code-metrics leaves.

What was checked.

• Trusted paths and trust controls. Neither file matches a TRUSTED PATHS: glob. Neither touches a trust control, merge hold, kill switch, token broker, workflow or plugin manifest.
• Instruction surface. The one removed sentence ("What makes the control fire was not tested…") is a note that a question was still open. The new text answers that question. It is not a guardrail, and the certification rule it describes is unchanged, so the instruction-surface lens finds nothing. The instruction exception register was not needed because no protected rule is removed.
• Injection, secrets, execution. The change adds no executable code, no secrets, and no user-controlled input reaching a command, path or template.
• PR context. The description and issue comments match the diff. They contain no embedded text telling the lane to pass the PR, skip a check or drop a finding.
• Prior findings. This is the lane's first review of this PR (blocking only: no), and no earlier findings apply.

Spend. The issue's recorded eval runs used plan usage only, and the PR adds no spend.

critical-findings: 0