[kyle-sexton]

Closes #8328

Summary

/work-items:decompose's window-bar row keeps the one-fresh-context-window bar and adds that a vertical slice can merge to the mainline at least daily. The cadence is practice guidance, the merge floor, not an estimate: the row states no hours unit, no time estimate and no day cap.

Fix

• plugins/work-items/skills/decompose/SKILL.md: the window-bar row adds the daily-integration clause, forbids a time estimate or day cap beside the existing token-budget ban, and points at DORA, trunk-based development and Fowler, Continuous Integration with an as-of date and recheck trigger.
• New eval case plugins/work-items/evals/decompose-slice-checks-daily-integration: an LLM judge on the window-bar verdict, a regex that the reply states daily integration, and a regex that it gives no hours estimate, with labeled samples for both regexes.
• The first paid run showed every with-arm reply wording the floor as "merge it the same day" or "within a day", which the judge passed and which meets the floor, while the regex demanded the literal "at least daily". The regex now also accepts those two forms (commit 78a0c17), with samples for each and a negated same-day fail.
• Changelog fragment .changes/work-items/feat-8328-decompose-daily-integration-6b1e4d9a.md.

Acceptance criteria

• grep -c "at least daily" on the decompose SKILL.md returns 1; grep -ciE "[0-9]+ hours" returns 0.
• Eval case (a case.yaml dir under the work-items plugin's evals/): a slice sized to the window bar is also checked for daily integration, with no hours estimate.
• The wording is confirmed against the Problem First effort's time-units rule (comment on #8328).
• After the commit, bash scripts/check-changed-skills.sh origin/main exits 0 (PASS, 0 errors, 4 pre-existing size and style warnings).
• A changelog fragment for the work-items plugin.

Verification

• scripts/check-affected.sh --tests: 7 passed, 2 skipped (shellcheck, ruff: no files of that kind changed).

• validate-cases.py plugins/work-items/evals: exit 0 (two WARN lines on the unrelated decompose-no-container-carries-brief case).

• Paid eval, claude plugin eval on Claude Code 2.1.303, model claude-opus-5-5, judge opus, 3 runs per arm, case decompose-slice-checks-daily-integration, plugin copied outside the repository:

  Plugin	With	Without	Delta	Validity
  This branch (grader fix applied)	1.00	0.33	+0.67	VALID
  main at 211a341 (the branch base) with the same case	0.33	0.33	0.00	VALID
  This branch, before the grader fix	0.67	0.33	+0.33	(superseded)

  The judge voted unanimously in all 6 runs of each pass. On main the with-arm applies the window bar but never states daily integration, so the judge and the daily regex fail; on the branch all three graders pass in every with-arm run. Noise report: n too small to call (one case; 3 or more needed for an interval), so this measures the case, not a suite-wide gain. List-price cost about 0.63 USD for the branch pass.

• Left out: judge calibration for applies-window-bar (it has no labeled samples, so it is reported uncalibrated); the validity gate's skill fired warning (the prompt loads the skill by its slash name, which the with-arm replies quote, and the case has no tool_used grader); main was measured at the branch base, not the main checkout, which was 81 commits behind.

Related

• Parent spec container: feat(planning): carry universal, overridable design defaults through planning output #8326.
• Agreed wording: Problem First effort's time-units rule, comment on #8328.

🤖 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 2026-10-19T03:30:29.550960Z	57f73d2	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 35d4dba60ecabd4299262cf69203020a249ab903: 0 finding(s), 0 as inline comments.

No security issues found in the incremental change since 57f73d2.

• Eval grader regex (mentions-daily-integration.md): the pattern now names the integrate/integrating/integration word forms. The (\w+\s+){0,5}? repetition is bounded, so there is no catastrophic-backtracking risk, and it only grades sample text.
• Eval sample JSON files: the only change is a trailing newline.
• decompose/SKILL.md: the only change is anchor fragments added to two citation URLs. No standing instruction was deleted, narrowed or softened, so the instruction-surface lens finds nothing. The instruction exception register was not consulted because the diff removes no rule.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 57f73d2360

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-review reviewed 35d4dba60ecabd4299262cf69203020a249ab903: 0 finding(s), 0 as inline comments.

No findings on the incremental change since 57f73d2.

I read .claude-lane/incremental.diff, the PR context and the labeled samples for the grader.

• Grader regex (graders/mentions-daily-integration.md): the stem integrat\w* became integrate[sd]?|integrating|integration. I traced the labeled samples against it by hand and the pass/fail labels all hold. The negation lookbehinds (cannot, not, n't, never) still block the "cannot integrate at least daily" and "cannot merge it the same day" fail samples. The narrower alternation only stops matching words like "integrator" and "integral", which a daily-cadence statement would not use. I did not run the regex: the command that would have run it was denied approval. The author's paid eval is the only run evidence, and I did not verify it.
• Sample JSON files: the only change is a trailing newline at the end of each file. This is harmless.
• decompose/SKILL.md window-bar row: the DORA and Fowler links now carry fragment anchors, and the "as of 2026-10-18" stamp and recheck trigger are kept. Both anchors look plausible, but I did not fetch either page to confirm the fragments resolve. The row still states no hours estimate and no day cap, and "at least daily" appears once, which meets the issue's acceptance criteria.