[sernote]

Kimi Messages guidance and evals 37/45 still described omitted top-level cache_control as read-only. The current Chinese-platform docs describe default 5m reads and writes instead. This repairs those rules and preserves the distinction between zero writes reported by one request and write behavior across requests.

Only the affected Moonshot reference sentences, a dated Messages API source link, and expected_output of existing evals 37/45 change. Full prefix hits can normally report zero creation; check reads first. In-message markers remain ignored, and top-level controls select TTL. Deployment behavior remains unverified.

Evidence and VALUE gate

• Chinese context-caching guide and Messages API reference, observed October 14 and corroborated October 15, 2026. Sources are undated; this is not a rollout date.
• Narrow PASS reused after confirming main 31a2e242f0c7b546afaf444b42d10d0bd6c8b47d, unchanged affected blobs and matching primary captures. The basis is a deterministic documentation/eval contradiction, with no tenant or causal benefit claim.
• Net product size -2 characters. Aggregate deferred estimate 95,315 → 95,314; ceiling 95,315 unchanged. No new eval/script, price/model change, or Last reviewed bump.

Verification and review receipts

• python3 -m unittest discover -s tests -p 'test_*.py': 260 tests OK.
• Package validator: ok, no warnings/errors. Trigger evaluation: 40/40 (30 positive, 10 negative).
• Python syntax, git diff --check, bytecode cleanup/absence: pass.
• Separate Opus 5.5/high correctness review and final correction review: no remaining blockers. One focused fix clarified per-request zero writes, the full-hit counterexample and read/write capability scope.

Evaluation results and limits

Static plugin-eval analyze and compare: 59/D → 59/D; trigger 152, invoked 6,700 and per-file deferred 95,329 estimates unchanged. Existing budget/complexity warnings remain; this is structure/budget evidence only.

Two blind scenarios each ran before and after in four fresh claude-b Opus 5.5/high sessions with identical per-scenario prompts and ordinary read-only primary-source lookup:

Scenario	Before	After
Omitted top-level control with synthetic 5m writes and fee-avoidance proposal	Correct core decision via source compensation	Correct core decision; omission does not disable writes
Zero creation on a full prefix read	Correct core decision via source compensation	Correct core decision; zero is normal and scoped to this request

All four accounted 24,888 total input without counting the TTL write breakdown twice. There is no demonstrated failure-to-pass or production benefit. The after-B answer still predicts a hypothetical TTL-switch write despite a verification caveat; that broader extrapolation is unproven and is not treated as general behavioral acceptance. No reruns were used to seek a preferred answer.

Full captures, preregistered criteria, answers, journals and command receipts are preserved outside the product repository. They contain local execution metadata and are excluded from the commit. No merge, release, deployment or production probe was performed.

[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	⚠️ Failed 2026-10-15T07:41:24.304993Z	ed266a7	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.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.