Repository navigation
feat(domain-driven-design): concise curate-language description, measured no-loss - #12114
Conversation
…ured no-loss
The description drops paraphrase ("Actively", "as domain understanding
changes", "what-it-IS", "already-known") and keeps all four quoted
trigger phrases, both unquoted triggers and the three exclusions. Body,
spoke, evals and probes are unchanged.
Trigger passes on the frozen probes and a blind held-out set, before and
after launched together: every pass VALID, every score unchanged on Opus
5.5 and Sonnet 5.5.
Refs: #6604
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ain-driven-design
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. |
|
claude-security-review reviewed Security review: PR #12114, no security issues foundI read both changed files in full in
Instruction-surface check: The new description drops some words ("Actively", "a consuming project's", "as domain understanding changes", "what-it-IS", "already-"). It keeps all four quoted triggers, both unquoted triggers and all three "Not for" exclusions. The removed words only shape when the skill triggers. None of them forbids an action or guards a protected class, so nothing that used to be blocked is now reachable unchecked. I could not read the instruction exception register from this lane (no shell or web access), so I judged these removals by what they do rather than against its list of protected classes. Other checks:
Spend: this change costs nothing to merge. The PR describes past eval runs on prepaid usage (about 57 USD at list price); that is not a hold. critical-findings: 0 |
|
claude-review reviewed Code review: no findings. I read both changed files in full: the changelog fragment and the What changed. Only the Checks against the change
Closing issue #6604, done-when items
No test can fail on these criteria because they are eval and process checks, not code behavior, so I report no test-gap finding. Spend. The PR reports about 57 USD of prepaid eval API credit, at list price. That is prepaid usage, not a billing change, so it is not a hold. Embedded text. The PR context contains no instructions aimed at this lane. |
Closes #6604
Summary
Concise description for
domain-driven-design's one skill,curate-language, on the issue'sdescription-only path, run as a measured before/after pass under the
rewrite protocol.
The new description triggers the same as the old one on every probe, on both target models. The
body, spoke, eval cases and probes are unchanged.
Fix
plugins/domain-driven-design/skills/curate-language/SKILL.md:descriptiononly, 587 to 506characters. It drops "Actively", "a consuming project's", "as domain understanding changes",
"what-it-IS" and "already-" (in "already-known"), and the colon after "Use when". It keeps all
four quoted trigger phrases verbatim, both unquoted triggers (domain modeling resolves
vocabulary; planning resolves domain language worth preserving), every branch, and the three
exclusions.
Decision table:
Verification
The pre-registration was written before any paid run, and each later rule went into its Run log
before the run it governs.
SKILL.md line 2. A fresh rewriter, barred from probes, evals and the held-out set, wrote the after
text. It had merged the two unquoted triggers so that "worth preserving" also qualified
domain modeling; an earlier round's labeler on this issue flagged that merge, so the original
split was restored before labeling.
docs-hygiene:compare-labeleragents, REASONS none. Both returnedVERDICT: PASS, 0 SEMANTIC LOSS, 0 AMBIGUITY. Each cut was labeled RELOCATED (the body andcontext/glossary-contract.mdstill say it) or FALSE POSITIVE.probes/curate-language.json(18 queries, 9 positive, 9 negative, fromtest(domain-driven-design): add outcome eval cases and frozen trigger probes for curate-language #11828), plus a held-out set of 16 queries (8/8) written blind by a fresh agent from a neutral
purpose statement. The held-out set is not committed.
/skill-quality:check measure-invocationemit-plugin-eval, thenclaude plugin eval --ablation none --disallow-tools Read Glob Grep --runs 3 -j 2 --keep-temp.Claude Code 2.1.316. In each pair, before and after were launched together.
run-validity.pycame from main.
It is partial, INVALID and not read. A should-trigger Opus run works the request after the skill
fires, so 24 positive runs alone exceed 5 USD. The ceiling was raised to 10 USD, a deviation
recorded before the rerun, and the pair was re-run in full.
Hits over queries x 3 runs (should-trigger fired / should-not stayed quiet), before / after:
measure-invocation compareEvery pass had 0 permission denials. The one run-validity WARN on each is that no judge model is
recorded, and the suite has no judge grader.
What this cannot show:
rival skills from other plugins win.
skill by slash command, and the body did not change.
Other checks:
/skill-quality:check plugins/domain-driven-design/skills: 1 passed, 0 failed, on main and onthis head. The same two warnings both times: the numbered procedure, and the description length
above Pocock's measured max (now 506, was 587). Run on the plugin root,
check-skill.shreports "No skills found" on both main and this head, so the
skills/path is the one thatchecks the skill.
check-affectedand the guardrails reference and content checks passed. gitleaks overthe branch: no leaks.
Spend: 57.36 USD list-price estimate, CLI-reported, across 10 result files including the two
partial passes. Eval wall time: 18:31-19:38 UTC, 2026-11-01.
Done-when on #6604: the diff under
skills/is thedescriptionline only. The probe file is onmain from #11828. The scores are above. After equals before. The skill check shows no new FAIL.
The changelog fragment is included.
Related
cl-existing-conventionprompt wording ("never the insuredorganization") and its
claimant-rejects-insured-orggrader, noted on test(domain-driven-design): add outcome eval cases and frozen trigger probes for curate-language #11828. The frozen outcomesuite was not run or edited here.
🤖 Generated with Claude Code