[ovhirup]

Closes #30.

Problem

The Context layer says to fork research, but not how that research should retrieve. A forked scan runs one grep or search, takes what comes back, and hands it on as if it were complete. The plan is then built on thin context, and nothing records which source went unchecked.

Policy change (policy-rev 8 → 9)

One Context-layer bullet:

A research fork states its question and done condition. On thin results, rephrase or switch source (code, docs, history, issues, web), at most twice more, then return Sources:, NOT FOUND:, UNCHECKED:. UNCHECKED: stays open.

• Word budget: the block is now 734 of its 735-word budget. To fit, the fork bullet drops "exploration", which "research" already covers. I didn't raise the budget. Since this leaves one word of headroom, the next policy change will have to cut something before it can add anything.
• Advisor: advisor.md now lists NOT FOUND: and UNCHECKED: items under Open questions and never builds a step on them. That part lives in the agent file because only the advisor plans.
• Retry cap: "at most twice more" matches the existing loop cap, as Agentic retrieval loop for forked research: rewrite, route, check, retry #30 asked.

How this meets #30's acceptance criteria

• The policy block and both ports state the rule, and every revision is bumped. Codex rev 7 adds the rule to the Context layer and the open-questions line to the advisor template. Grok rev 6 adds the rule, and its PLAN phase treats unconfirmed items as open questions. The CHANGELOG 3.6.0 entry tells users to re-run the installer and accept the agent refresh.
• The research output includes Sources:, NOT FOUND: and UNCHECKED: lines.
• The rule stays short, with no taxonomy prose: one bullet of 36 words.
• docs/enforcement.md lists the rule as instructed, not mechanical, for all three: Claude Code, Codex and Grok.
• CI and the install checks pass with the new revision (see Verification).
• A behavior scenario is added: evals/evals.json case 10, where the answer is only found after a rephrase or a second source.

Also fixed

• docs/evals.md: it said "Seven scenarios" and stopped at case 7, though cases 8 and 9 shipped in 3.3.0 and 3.4.0. Adding case 10 changed the count anyway, so it now says ten and lists cases 8–10.
• README layout comment: the stale eval count there is corrected too.

Verification

• python3 tools/validate.py: all 9 categories OK. The budget step from validate.yml reports policy block: 734 of 735 words OK.
• tools/: 63 tests pass. measured/: 12 tests pass. None skipped.
• No policy-rev 8, grok-rev 5, codex-rev 6 or 3.5.0 references remain outside the CHANGELOG history.
• Upgrade from a real rev-8 install (the block and advisor.md from main):
  • Before: check_install.py fails on the policy revision and on agent advisor — body differs from assets/advisor.md. diagnose.py reports both problems.
  • After replacing the block and the advisor: INSTALL: PASS.
  • This exercises the stale-agent detection from Upgrades leave stale scaffolded agents in place, and nothing detects it #31 end to end.

Limits

• Whether agents follow the rule in practice isn't measured. That needs the runtime evals parked in Execute installation fixtures and agent-behavior eval scenarios #12.
• Publishing 3.6.0 (Actions → release) is a manual step after merge.

---

Generated by 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 2026-11-02T08:40:30.746153Z	8b17660	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.