Repository navigation
revert: restore four lane-3 skill cuts that failed negative controls - #12247
Conversation
…cription The shorter description shipped in #12114 on an eval that cannot tell it from "Helper for repository maintenance tasks.": on Sonnet 5.5 both fired 51/51 should-trigger probes, and the one should-not case drop did not reproduce in a confirmation run. Under the negative-control rule the cut was unmeasurable, so the earlier text returns until a suite can measure it. Refs: #6604 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t removed A copy of the setup body with its five procedure sections emptied scored 0.942 against main's 1.000 on Sonnet 5.5 (3 cases merged, no case drop), a 0.058 drop. Under the negative-control rule the body cut was unmeasurable, so its two removed Gotchas return. The description change stays: the skill has disable-model-invocation, so its description never reaches routing. Refs: #6599 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With its check and Output sections emptied, the sync body still scored 1.000 against main's 1.000 on sy-check-report-only (Sonnet 5.5, 5 runs each, both VALID, no grader miss). Under the negative-control rule the cut was unmeasurable, so the earlier body returns until a suite can measure it. Refs: #6605 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The manage body's negative control failed: the suite scored the shipped body 0.898 and a copy with six procedure sections emptied 0.889, with no case drop. The manage description and setup body keep #12121's text; their controls passed. Refs: #6609 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c5f96a4fc7
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".
|
claude-security-review reviewed Security review: PR #12247, incremental (c5f96a4 → 14f34bd)No security issues found. I read the one hunk in What changed: one paragraph in the body of
This is the fix for the settled owner thread on this file (resolution head 14f34bd, lines changed since resolution: no). I am not raising that point again. Instruction-surface lens: the change keeps every rule. Nothing is removed or narrowed:
Because the rules are unchanged, I did not need to consult the instruction-exception register. Checks:
critical-findings: 0 |
|
claude-review reviewed Code review (incremental since c5f96a4): PR #12247, no findingsScope. This review covers one change: commit What changed. In the "Consumer vs maintainer role split" paragraph, the old sentence explained what
Checks
Scope and inputs
|
The restored role-split sentence presented the flag's upstream behavior as the reason. State the decision (a person runs the check) and point at the docs section with an as-of date and recheck trigger. Refs: #6605 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Refs: #6604
Refs: #6599
Refs: #6605
Refs: #6609
Summary
Restores four skill-text cuts that lane 3 merged as "measured no-loss". Retroactive negative controls (the owner's rule in
rewrite-a-plugin.md, #12222) showed that the suite behind each cut scores a deliberately broken copy about the same as the shipped text. So none of these cuts was measured, and under the rule none stays. Each issue stays open with its control result posted.Fix
One commit per part, each with a patch changelog fragment:
domain-driven-designcurate-language: the description returns to its text before feat(domain-driven-design): concise curate-language description, measured no-loss #12114.attributionsetup: the two Gotchas that feat(attribution): concise setup skill rewrite, measured no-loss (round 2) #12159 removed return. The description stays, because the skill setsdisable-model-invocation, so its description never reaches routing and a control cannot measure it.dometrainsync: the body returns to its text before feat(dometrain): concise sync body (round 2, measured no-loss) #12162.kindle-dedrmmanage: the body returns to its text before feat(kindle-dedrm): concise manage and setup skills, measured no-loss #12121. Themanagedescription and thesetupbody keep feat(kindle-dedrm): concise manage and setup skills, measured no-loss #12121's text, because their controls passed.Verification
Every control ran on Sonnet 5.5 with judge Opus. Each broken copy was launched together with a fresh pass on main's text. The rule's PASS bar is a noise-report case drop, or a mean drop of at least 0.15.
checkand Output emptiedcheckandapplyemptiedgit log origin/main -- <file>.git diff <cut>^ HEADon each restored file is empty, except for the two descriptions that stay (attributionsetup, kindle-dedrmmanage).managebody keeps the no-batch-cleanup and Calibre-Library-never-offered rules from fix(kindle-dedrm): call sync-finalize.sh without its nonexistent argument, keep cleanup per-item and the library kept, and add manage and setup eval cases #11967.Related
🤖 Generated with Claude Code