[mldangelo-oai]

Summary

Deep Scan reducers are told to use JavaScript code mode to read paged findings. In a fresh Codex 0.164.0 worker, code mode is off and the two reducer MCP tools are available directly. A reducer that follows the old instructions can finish its turn without writing a result, then receive the same instructions on retry.

Fixes #1865.

Changes

Explain how to follow each cursor, read the fragments in order, and submit the aggregate directly when code mode is unavailable. Apply the same guidance when a reducer resumes. Workers with code mode can continue using the existing pager to keep large documents inside a cell.

Testing

• Reproduced the mismatch on main at 6696705319. The new regression failed against the original prompt and passed against both the initial and resumed prompts after the fix.
• Ran one real Codex 0.164.0 turn with code mode off against the production artifact server. The reducer called both tools directly and saved two synthetic findings with their identities, source IDs, and original bodies intact.
• Exercised the packaged plugin with a scripted local model endpoint. The real worker read five pages with a 480-byte limit and submitted once. A separate interrupted run resumed the same thread at its next unread cursor and committed the reduction.
• The coordinator, template and large-input paging tests passed, including the stdio response limit and native code-mode IPC limit. The plugin build, MCP typecheck, Prettier check on the changed TypeScript, and all five portable source checks passed.

Risk and rollout

This change affects the instructions sent to reducers. The direct path keeps input fragments in the model conversation, so the model's context window still bounds what it can process this way. The live-model smoke covered a small batch.

Public disclosure review

• No customer, partner, prospect, or user identities, data, or identifying details are included.
• No credentials, personal data, private source, scan findings, or nonpublic links or tickets are included.
• I reviewed the branch name, title, description, commits, changes, comments, logs, screenshots, attachments, and links for public disclosure.

[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 2026-10-24T02:24:56.168952Z	1d415c1	PR opened
🔒 Security Review	✅ Completed 2026-10-24T02:26:21.876752Z	1d415c1	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.

[alandelong-oai]

Thanks for fixing this. No blocking findings in this owner-level merge-safety pass. Both initial and resumed instructions now match the directly available reducer tools while retaining the code-mode path.

Verified a local scripted direct-tool run with five 480-byte pages, one successful submission, and intact source bodies and identities, plus focused coordinator, template and paging checks, typecheck, plugin build, and portable checks. The full coordinator suite is locally blocked by sandbox filesystem ownership; its seven isolated relevant cases pass. No live-model run was performed. Some Windows CI checks were still pending at review time.