[sernote]

The Bedrock reference categorically restricted GPT-6 Astra caching to Responses and applied Responses-shaped usage/accounting claims to it. AWS's guide and Astra card support explicit caching on Runtime Chat Completions and InvokeModel, while Converse supports implicit caching only. A valid Chat Completions configuration could therefore receive a false unsupported-surface finding and an unnecessary migration recommendation.

Remove Astra from the GPT-5.6-only paragraph and link it, per its own card, to the existing surface, placement, minimum and usage-ambiguity guidance. Explicitly exclude Sol's Routes bullet because Astra routes and tiers differ. Only original bedrock.md lines 20–21 change: 8,391 → 8,319 characters (−72); every other line is byte-identical. GPT-5.6 support wording is preserved apart from unsharing Astra.

Evidence and VALUE gate:

• AWS prompt-caching guide: Astra row lists input_text for Responses and text parts for Chat Completions/InvokeModel; no explicit Converse caching.
• AWS GPT-6 Astra card: Runtime implicit/explicit surfaces are distinguished; Mantle supports both modes on Responses and Chat Completions.
• Baseline reference: applicable categorical contradiction. VALUE PASS is limited to deterministic contradiction repair; the existing Freshness Gate can already produce correct source-grounded answers.
• Original undated AWS change observed within 2026-10-17T06:57:59Z .. 2026-10-19T07:08:44Z; unchanged sources reverified 2026-10-20T08:56:53Z. No publication date is inferred from HTTP rebuild stamps. Accepted normalized SHA-256: guide b2729b952c356010bddcae804b50485872a142a844b53bd7ade0faf2b8cc82f8; Astra card 1fb4dca684852c76d971fcc779b57a6042b40ee62b73cba571c04c2a743c95db. Direct fresh HTML captures matched the accepted captures. Fresh base 4784358cdcd2df9900107a8e0bc03abe367c6a7f; relevant reference, SKILL and rules blobs match the accepted gate.

Fresh local verification output:

python3 -m unittest discover -s tests -p 'test_*.py'
Ran 260 tests in 40.526s — OK
python3 audit-prompt-caching/scripts/validate_skill_package.py audit-prompt-caching
status: ok; errors: []; warnings: []
python3 audit-prompt-caching/scripts/run_trigger_eval.py audit-prompt-caching
status: ok; cases: 40 (30 positive, 10 negative); errors: []
Python syntax compile: all 12 bundled script/test files OK
git diff --check: clean
Bytecode cleanup: complete; residual __pycache__ / *.pyc check empty

Static evaluation (plugin-eval analyze before/after, then compare): trigger estimate 152 unchanged; invocation estimate 6,700 unchanged; deferred estimate 95,329 → 95,311 (−18). Score remains 59/D, with the same pre-existing budget/structure/code warnings and no new failures. No budget ceiling changed. These are static estimates.

Targeted inspection results, with criteria fixed before runs:

Scenario	Source-grounded before	Source-grounded after
Astra Runtime Chat Completions, explicit, text breakpoint, 30m, ≥1,024 prefix, low reads and writes present	Supported surface; no Responses migration; inspect prefix/route and raw usage	Same correct answer; categorical reference contradiction removed
Astra Runtime Converse with native cachePoint	Explicit unsupported; retain implicit-prefix audit	Same correct answer; reference now agrees

Both were informed before/after inspections, not blind experiments. No behavioral uplift, cache hit, per-route usage schema/accounting, ROI or production outcome is claimed. Existing source lookup already compensated for the old contradiction. No new durable eval/test or script behavior change was required.

Independent fresh read-only review (Opus 5.5 high): APPROVE, no blocking findings. It verified C1–C5, the exact two-line diff, source contracts and both before/after inspections; the Routes/tier exclusion is unambiguous. Remaining limits are shared-bullet drift and unchanged generic diagnostics outside the accepted scope. No runtime behavior was tested.

GitHub CI verification output: validate passed for head 1cf6ea35eaa7f62fde5f6fc2bfde9556b716d845 (run completed 2026-10-20T09:07:18Z); all script tests, package validation, trigger summary, syntax, whitespace and bytecode checks passed.

[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 2026-10-20T09:08:41.677869Z	1cf6ea3	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.