[tmchow]

Summary

A subagent brief now has to carry what the child cannot recover on its own: the user's stated limits on its work, a stop rule that says what to hold and what may continue, and a report of anything refused, skipped, or left unverified. Separately, ce-setup now counts distinct pull requests, not repeated bot comments, when it decides a finding keeps getting declined.

After #1974, the dispatching skills said a spawned worker "does not continue the parent conversation", which states a fact but asks for nothing. One recorded Codex ce-code-review brief said exactly that and still dropped a constraint the user had given.

What changed

• Authoring standard. portable-agent-skill-authoring.md states the three brief conditions once, with a checklist line. Decisions a child exists to judge on its own, such as settled design choices a reviewer should re-derive, stay out of the brief.
• ce-code-review. The Stage 2 intent summary, which every reviewer and the cross-model peer receive, now carries the user's limits on the review in their own words.
• ce-work. The worker brief now carries user limits, which win over the plan when the two conflict. It also says what to do when the harness refuses an action or a step can't run here: hold that step, don't reach the same result another way, keep finished edits, and finish the independent parts. The worker reports what did not go as briefed, and a held step keeps the unit open on both the serial and the parallel path.
• ce-setup. At the evidence step, one condition governs both recurrence judgments: the "declined across several pull requests" gate and the refresh signal "findings under a written rule that the team keeps declining". An incident is a pull request, or a change when no pull request carries it. A partial history is named in the preview and summary.

ce-standards-bar is unchanged. It admits a rule on a single fixed finding, so it doesn't count anything.

Validation

Live A/B evals on Claude and Codex, one trial per cell. Pre is main at d307bdb and post is this branch. The fixtures and rows are added to tests/skill-eval-cell/catalog.ts.

• ce-work/live-worker-brief-limits (live parallel wave, a user limit that conflicts with the plan, and a docs generator that refuses to run): all 4 cells passed the artifact grade. tests/ledger.test.js was untouched, docs/api.md was not hand-edited, and the docs step came back as a blocker in every cell. In Claude's worker prompts, the user limit and a blocked-step rule were already present before the change (4/4). The "did not go as briefed" report item appeared only after it (0/4 to 4/4). On Codex the brief can't be graded: spawn messages are encrypted, and fork_turns: "all" replays the parent's context into each child, so the outcomes are the only evidence there.
• ce-code-review/live-user-limit-reaches-reviewers (user asks to leave a staged file out): Claude reviewer prompts carried the limit in both arms (6/6), and no arm on either host produced a reviewer finding on the excluded file. The recorded Codex miss did not reproduce, so this shows no regression and no measured gain.
• ce-setup/review-recurrence-counts-incidents and ce-setup/refresh-recurrence-counts-incidents: all 8 cells passed in both arms. Four re-posts in one PR counted as one incident, and the partial export was stated. On current models the condition is insurance.
• A fresh-reader pass found ambiguities in the first draft, all fixed before the post arms ran:
  • the serial path could still commit a unit over a held step
  • "a check fails" could be read as covering expected red tests
  • the brief didn't say whether the plan or a user limit wins
  • "changes" could be read as pushes
• bun run test --changed-since=45ccf35e4 (4971 pass, 0 fail), typecheck, release:validate, and plugin:validate all pass.

grading-live-delegation-evals-from-host-session-logs.md now records where the eval cells keep Codex rollouts and why forked Codex children can't show what the brief said.

Follow-up: ce-doc-review still uses the bare #1974 sentence. Its reviewers return residual risk through a schema, and it has no observed gap.

Security Disclosure

No security-relevant changes.

Agent Disclosure

• Model: Claude Code · claude-opus-5-5

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-09T19:09:52.514510Z	63c33fe	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.