[kyle-sexton]

Closes #12631

Summary

Replaces #12647, which pr-unblock could not finish: the open review finding needed an edit to plugins/discovery/scripts/contract.test.sh, outside that lane's scope. This branch carries #12647's commits and adds the missing test.

Fix

• From fix(discovery): give the research Effort-ceiling sentence headroom in the re-attach slice #12647: the sentence "The Effort row is the ceiling over discipline 8" moves into the Disciplines intro of plugins/discovery/skills/research/SKILL.md, so it now ends at byte 15,332 instead of near byte 20,000.
• New: contract.test.sh asserts that this whole sentence ends at or before byte 18,000 (20,000 minus the 2,000 bytes of headroom discovery research SKILL.md: a contract-tested sentence sits 1 byte inside its 20000-byte window #12631 asks for), measured with head -c the same way assert_in_slice measures. The check covers that sentence only. The "Effort, source breadth" heading starts at byte 19,592, so a blanket 18,000 limit on every pinned phrase would fail today.

Verification

• bash plugins/discovery/scripts/contract.test.sh: all assertions pass, including research effort ceiling sentence keeps 2,000 bytes of headroom.
• Negative check: 2,700 filler bytes inserted before the sentence moved its end to byte 18,033. The new assertion failed, while research effort ceiling sentence is inside the re-attach slice (the 20,000-byte check) still passed. The file was then restored from HEAD.

Related

• fix(discovery): give the research Effort-ceiling sentence headroom in the re-attach slice #12647

🤖 Generated with Claude Code

https://claude.ai/code/session_01Vcupom3BKF6QuC5U3sRJPj

[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 2026-11-03T06:06:09.118040Z	56752f9	PR opened

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 56752f9e79d36aff7aef011e4905167e74cc8563: read 3 of 3 changed file(s); 0 finding(s), 0 as inline comments.

Code review: no findings

I read all 3 changed files in full: the changelog fragment, contract.test.sh and research/SKILL.md.

Closing issue #12631. The done-when criterion is "the sentence has at least 2000 bytes of headroom … and contract.test.sh passes".

• Headroom is met and tested. The sentence "The Effort row is the ceiling over discipline 8" now sits in the Disciplines intro, at SKILL.md#L103. The new check at contract.test.sh#L389-L394 matches the whole sentence inside head -c 18000, so the test fails if the headroom shrinks below 2,000 bytes. It measures the same way assert_in_slice does.
• The test passing is the author's claim only. I had no shell, so I did not run the test. The byte offsets (15,332 for the sentence, 19,592 for the heading) are also the author's figures.

Other pinned assertions still hold, from reading the head file:

• The fan-out paragraph moved under "The deep tier" (SKILL.md#L167). It keeps its leading indent, and "Fanning out over N topics … fresh verifier for the verifier-owned rows" is still on one line. So assert_present at contract.test.sh#L563 still matches.
• The 2-space indent makes it a continuation paragraph of the "Two or more separable topics" bullet, which is the right place for it.
• The ### Effort, source breadth heading and verification: skipped (cost) are both still present.

Note, not a finding. The fan-out paragraph's details (the --sub-slice restatement exception and the cost skip) now sit past the 20,000-byte slice that compaction re-attaches. Acceptance gate step 2 keeps a short pointer (SKILL.md#L42), so after compaction the gate still says to grade each sub-slice and send the synthesis to a fresh verifier. Losing those details fails safe: a compacted run would always send the synthesis to a verifier and never take the cost skip.

Changelog. The fragment's bump: patch fits a doc move plus a test. Its "about byte 15,200" differs slightly from the PR body's 15,332, which doesn't matter.

No plugin manifest changed. There was no settled-threads file, and the prior-findings file lists no earlier findings (0 rounds, cap 3). The PR context had no embedded instructions.

[github-actions]

claude-security-review reviewed 56752f9e79d36aff7aef011e4905167e74cc8563: read 3 of 3 changed file(s); 0 finding(s), 0 as inline comments.

Security review: PR #12745 (head 56752f9)

I found no security issues in this pull request. I read every hunk of all 3 changed files, so this is a complete review. None of the files matches a trusted-path glob.

• plugins/discovery/skills/research/SKILL.md: the text is moved, nothing is removed.
  • The sentence "The Effort row is the ceiling over discipline 8" now sits in the Disciplines intro.
  • The paragraph on grading each N-topic fan-out run against its own sub-slice now sits under "The deep tier". The rule is unchanged: each sub-slice is graded before synthesis, the synthesis goes to a fresh verifier, and verification is skipped only on the restatement check. The only addition is an "(acceptance gate step 2)" cross-reference.
  • Acceptance gate step 2 still keeps a short pointer to that paragraph.
  • I read this under the instruction-surface lens. No guardrail is deleted, narrowed or softened, so no protected class is touched. I did not open the instruction exception register file.
• plugins/discovery/scripts/contract.test.sh: one new assertion runs head -c 18000 on a file inside the repo and does a fixed-string glob match. It takes no outside input and runs no injected code.
• .changes/discovery/...md: a changelog fragment only.

There are no manifest, workflow, token or trust-control changes. Nothing in the pull request context tries to steer the review. No prior findings exist; the round cap (3) did not apply.

critical-findings: 0