[kyle-sexton]

Closes #12659

Summary

The /discovery:research hub carried the practitioner-source test on every load, though a run needs it only when it weighs a person as a recognized author. This PR moves that text word for word into a reference file read at that point. It is the third progressive-disclosure cut under #12243, following the #12387 recipe.

Fix

• New skills/research/reference/practitioner-test.md holds the two moved paragraphs: the proof-of-shipped-work test and the untrusted-profile paragraph. "Reconciling sources at the gate" keeps its heading, first paragraph and scoped-exception quote, and points at the file before a person's post, talk, profile or repository is weighed. "See also" lists the file.
• Hub body (SKILL.md with frontmatter stripped, awk strip piped to wc -c): 45,165 bytes on main 221bfbabf1, 43,646 on this branch. The issue's 44,641 predates fix(discovery): give the research Effort-ceiling sentence headroom in the re-attach slice #12647.
• The two external citations of the moved text, docs/upstream/aiforui-course.md rows L8 and L9, now name the reference file (L9 was fixed in review). No test, contract window, hook or script cites it; the 20,000-byte window pins in contract.test.sh are untouched, since the cut sits after byte 20,000.
• Two new cases read the moved text: rs-practitioner-needs-shipped-work and rs-profile-imperative-is-red-flag. Neither grants Bash, and both end before any web research.

Verification

• /docs-hygiene:compress compare, two independent compare-labeler agents, before against after: both returned VERDICT: PASS. Union of rows: 2 RELOCATED, 2 FALSE POSITIVE, no SEMANTIC LOSS, no AMBIGUITY.
• plugins/discovery/scripts/contract.test.sh: all contract assertions passed. validate-cases.py on plugins/discovery/evals: no finding on the rs-* cases.
• Evals: claude plugin eval --case 'rs-*' --ablation none --runs 2 -j 2 --keep-temp --model claude-opus-5-5, Claude Code 2.1.317, capped scopes, arms launched together. Before is this branch's copy with main's skills/research swapped in. Firing check on main (--runs 1): all 3 cases scored 1.00, VALID. The suite was frozen after it, with one later edit: the spelling check flagged the optional recognised alternative in the not-cited-as-authority grader, so it was dropped. Regrading all 9 observed replies of that case with the old and new pattern gives the same verdict on each.
• The negative control was preregistered on perf(discovery): cut the research skill body through progressive disclosure #12659 before the first paid run.

Round	Arm	Mean	noise-report	run-validity.py --runs 2
B	before	1.00		VALID
B	after	1.00	+0.00 (+0.00 to +0.00), non-inferior at margin 0.05	VALID
C	before	1.00		VALID
C	degraded (body deleted)	0.50 over 2 cases	n too small to call (2 cases comparable)	INVALID: rs-practitioner-needs-shipped-work hit max turns and one own-copy Grep was denied

• After-arm traces show the reference file read in all 4 runs of the new cases; no denied call in either B arm.
• Control decision: rewrite-a-plugin.md step 4 says a case whose degraded runs carry a denial cannot decide the control, so the fall must hold without it. Without rs-practitioner-needs-shipped-work, the mean falls from 1.00 to 0.50, past the 0.15 threshold. The suite detects a loss, so the cut is measurable and ships.

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), perf(source-control): move the commit skill's rationale out of its hub #12704 (commit cut), fix(discovery): give the research Effort-ceiling sentence headroom in the re-attach slice #12647, discovery research SKILL.md: a contract-tested sentence sits 1 byte inside its 20000-byte window #12631.

🤖 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 2026-11-03T17:04:05.549446Z	f64988d	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 7fe6f09b363f9ada63f1179bfc3c496c9177584d: read 1 of 1 changed file(s); 0 finding(s), 0 as inline comments.

Security review (incremental since f64988d): no security issues found

critical-findings: 0

I read the one file changed since the last review, docs/upstream/aiforui-course.md, and its single hunk. The only change is in the L9 row of the lesson table. The "Be selective" citation used to point at plugins/discovery/skills/research/SKILL.md and now points at plugins/discovery/skills/research/reference/practitioner-test.md. That is where this PR moved the practitioner-source rule, and the L8 row on the line above already cites the same file.

• Security surface: none. This is a provenance record in documentation. It doesn't change any standing instruction, gate, workflow, token request or manifest.
• Trusted paths: the file doesn't match any of the TRUSTED PATHS globs.
• Manifests: no plugin.json or marketplace.json changed.
• Prior findings: the last review left only a summary comment with no findings. No prior finding applies to this change.
• Embedded text: none. The change contains no instructions aimed at the review lane.
• Spend: nothing new.

[github-actions]

claude-review: 1 inline finding(s) on f64988da9a7d31a4f7cdf504489ee195949a0102; the summary is in this lane's pull request comment.

claude-lane: claude-review

[github-actions]

claude-review reviewed 7fe6f09b363f9ada63f1179bfc3c496c9177584d: read 1 of 1 changed file(s); 0 finding(s), 0 as inline comments.

Code review (incremental): PR #12876, f64988d → 7fe6f09

I reviewed 1 of 1 changed files: docs/upstream/aiforui-course.md. I found nothing new.

Last round's finding is fixed. Row L9 (line 312) cited plugins/discovery/skills/research/SKILL.md. It now cites plugins/discovery/skills/research/reference/practitioner-test.md, the same file row L8 names. The edit changes only that path. With it, both outside citations of the moved text point at the reference file. That meets the citation part of #12659's criterion, "every citation of it (tests, contract windows, hooks, scripts, fixtures, README) is repointed". The PR body now names rows L8 and L9 as those two citations, which matches the diff.

Not verified: the owner says a sweep of docs/ and plugins/ found no other citations of the moved text. I didn't run that sweep, so this is the owner's claim. The byte-count, /docs-hygiene:compress compare and eval non-inferiority criteria are also still author-claimed. I can't run those tools from this lane.

Spend: the two new eval cases from earlier rounds still use prepaid model credits on each rs-* run. This isn't a billing change.

I found no embedded instructions in the PR context, the settled threads or the prior findings.