[kyle-sexton]

Closes #12666

Summary

The /source-control:commit hub loaded explanatory prose on every invocation that a commit rarely needs. This PR moves that prose word for word into reference files read on demand, and keeps every command, table and gate in the hub. It is the second progressive-disclosure cut under #12243, following the #12387 recipe.

Fix

• New reference/config-probes.md, reference/changelog-fragments.md and reference/design-rationale.md hold the moved explanations. The three native-surface descriptions move into the existing reference/native-commit.md.
• The hub's reference index has a row for each file, so the agent still knows where each topic lives.
• Hub body (SKILL.md with frontmatter stripped): 28,967 bytes on main, 25,098 bytes on this branch.
• Four commit-* eval cases never fired the skill on main, so no round could be VALID. Each prompt now names the skill: commit-hook-rejection-no-bypass, commit-never-stage-everything, commit-no-push-no-pr and commit-trailer-model-from-session. Only the prompts changed; graders and case.yaml are untouched.

Verification

• awk frontmatter strip piped to wc -c on both copies printed 28967 and 25098.

• /docs-hygiene:compress compare (fresh compare-labeler, before against after) returned VERDICT: PASS with no SEMANTIC LOSS and no AMBIGUITY rows.

• A repo-wide grep found no external citation of the moved text to repoint.

• markdownlint passed on every changed file; the pre-push check-affected hook passed.

• Paired commit-* eval. Command: claude plugin eval --case 'commit-*' --runs 2 --scaffold --ablation none --keep-temp --model claude-opus-5-5, Claude Code 2.1.317, under a capped scope.

  • Before is this branch's copy with main's skills/commit swapped in. The two arms differ only in the commit skill.
  • Result: both arms VALID under run-validity.py --runs 2 --scaffolded, with 12 of 12 traces resolving per arm. The only warning is that the judge model was not recorded.

  Round	Before	After	Change, 95% interval	noise-report at margin 0.05	run-validity.py
  3	1.00	1.00	+0.00 (+0.00 to +0.00)	non-inferior	VALID both arms

  • Rounds 1 and 2 ran before the case fixes. They were INVALID on the skill-fired check, so they do not count.

Related

• perf: cut the bodies of the most-invoked skills through progressive disclosure - part 1 of 2 #12243 (parent), perf(source-control): move the worktree nesting measurement out of the skill hub #12387 (cut 1 recipe).

🤖 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 2026-11-03T05:56:49.928495Z	5d80bd5	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.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5d80bd5dcb

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-security-review reviewed 2ea70e7cdb622e2971e100a0e3d8129615dd4441: read 1 of 1 changed file(s); 0 finding(s), 0 as inline comments.

Security review: PR #12704, incremental since 6e8daa73

No security issues found.

I read the one hunk in the one file this push changed (.claude-lane/incremental.diff): plugins/source-control/evals/commit-trailer-model-from-session/prompt.md. The file is not on a trusted path. It is not a trust control, a merge hold or a plugin manifest.

• The only change is to the user prompt of an eval case. It now ends "Commit the staged fix with the source-control:commit skill." instead of "Commit the staged fix.", so the skill fires in that case. This matches the other eval cases the previous push fixed and the PR's "Remaining" note.
• The case's adversarial setup is unchanged: the prompt still carries the mismatched Co-Authored-By: Claude Fable 5.1 trailer. expected_outcome still says the trailer must name the actual model, and allowed_tools ([Read, Glob, Grep, Skill, Bash]) is the same.
• This is an eval fixture, not a standing instruction. The change removes no guardrail and grants no tool, so the instruction-surface lens finds nothing.

The PR context file contains no text aimed at the review. The one settled thread, on design-rationale.md, is outside this push's files and stays settled. The only prior post is the last round's clean summary, so there is nothing to repeat. The round cap was not reached (0 of 3 rounds).

critical-findings: 0

[github-actions]

claude-review reviewed 2ea70e7cdb622e2971e100a0e3d8129615dd4441: read 1 of 1 changed file(s); 0 finding(s), 0 as inline comments.

Code review (incremental): PR #12704, changes since 6e8daa7

No new findings. I read the whole of the one file changed since the last review.

• plugins/source-control/evals/commit-trailer-model-from-session/prompt.md

What changed. The prompt now ends "Commit the staged fix with the source-control:commit skill." The last push did the same for commit-hook-rejection-no-bypass and commit-never-stage-everything. The point is to make the skill fire, so the paired eval round isn't INVALID on the skill-fired check.

Does the case still test what it says? Yes:

• The prompt still sets the trap. It carries the mismatched Co-Authored-By: Claude Fable 5.1 instruction and states the real model (Opus 5.5).
• The expected_outcome is unchanged and still matches the prompt.
• Naming the skill changes what the case tests: whether the skill's trailer rule overrides the stale string, rather than whether the model picks the skill. That fits a before/after comparison of the hub content.

Carried-over observation (not a finding). commit-no-push-no-pr still doesn't name the skill. The previous review raised this, and it still holds: on the before arm it could still miss the skill-fired check. Done-when 3 of #12666 needs a round that is VALID on both arms, so this case may need the same edit.

Other checks

• Path-scoped rule (.claude/rules/eval-case-transcripts.md): the prompt is a short synthetic line with no transcript text or identifying details, so the rule is met.
• Plugin manifests: none changed in this increment.
• Closing issue: none. The PR only refs perf(source-control): cut the commit skill body through progressive disclosure #12666, so there were no acceptance criteria to check.
• Settled threads / prior findings: the one settled thread (design-rationale.md, no lines changed since it was resolved) is outside this increment. The only prior post is the last round's summary, which had no findings, so nothing is repeated.
• Embedded instructions: none found in the PR context, the settled-threads file or the prior-findings file.
• Round cap: 0 rounds with findings so far against a cap of 3, so the cap did not apply.
• Verification: I had no shell, so I ran nothing. Whether the skill now fires in this case stays unverified until an eval round runs.

Spend (not a finding): only the normal lane runs and the planned paired eval rounds, which use prepaid API credit.

[kyle-sexton]

Reply to the claude-review finding on plugins/source-control/skills/commit/SKILL.md:195:

#	Finding	Classification	Basis
1	Step 7's release-record obligation left the hub	VALID, fixed	verified: the hub's step 7 named only exit 2 and the reference file. The rule that a plugin the list does not name still needs the hand version bump and CHANGELOG entry was only in reference/changelog-fragments.md

Fixed in 9a365b4. Step 7 now states that a fragment replaces the hand release record, and that for a plugin the list does not name the script writes nothing and the hand record still applies.