Close the three deferred GPT-6 Astra gaps

[myzie]

Builds on #285 and carries its five commits, because CI (go-test.yml) and CodeRabbit only run against main — targeting the feature branch got this zero checks. If #285 merges first, this shrinks to the one commit on top; if it does not, this is the whole change. #285 landed Astra with three gaps recorded in its "Known gaps" section, each blocked on the model being reachable. All three close here.

Explicit prompt-cache breakpoints

The gate was left at gpt-5.6 because guessing wrong fails every Astra request, while leaving it off only forgoes an optimization. It is no longer a guess: OpenAI's prompt-caching guide draws the line at "GPT-5.6 and later" — that cohort accepts developer-placed breakpoints, GPT-5.5 and earlier take implicit ones only. Astra's published rates follow the same formula the guide states for the cohort (a cache write at 1.25× input, $12.50 against $10.00; a read at 0.1×, $1.00).

supportsExplicitPromptCaching now checks a list of prefixes instead of one. Enumerated rather than expressed as "5.6 or newer": a version comparison would opt in every future model id on faith, and gpt-5.2-pro and gpt-5.3-chat in providers/modelcaps are the standing counterexamples to a family name predicting what a variant accepts.

Long-context cache writes

Above 272K input tokens Astra bills cache writes at $25.00/1M. llm.PricingInfo had no field for it, so a write on a long request was costed at $12.50 — half of what it cost. The new LongContextCacheWritePrice runs through the catalog schema, the Python generator, CostOf, and Scaled.

The field is optional, and that is the interesting case: a model with a flat write rate and a long-context tier for everything else (nothing in the catalogs today, but the shape is legal) keeps its standard rate past the threshold rather than dropping to zero. TestPricingInfoCostOf_LongContextCacheWrite pins both branches. Astra is the only priced model with both a long-context threshold and a cache-write rate, so no other entry changes. Parse rejects a long-context write rate with no standard one beside it.

TestGPT6AstraLongContextPricing now asserts the tier is applied instead of pinning the understatement. Both cases size the whole request — cache-creation tokens count toward TotalInputTokens, which is what selects the tier, so the original 1M-token write in that test would have crossed the threshold on its own.

CLI recommendation

The catalog withheld recommended from a model the CLI could not reach. Astra now leads the OpenAI list at cli_order 1 with the 5.6 family shifted down; gpt-5.6-sol stays default, so nothing changes for anyone who does not pick Astra. The gpt- prefix already routes it to the Responses provider, so selecting it works with no other wiring.

What did not change

• The capability entry. Still low through max, still documentation-derived, and its comment now says that without claiming a probe that has not happened — this session has no OpenAI key, so nothing here is API-verified. The docs are unambiguous on both exclusions ("GPT-6 Astra does not support none reasoning effort"; no custom temperature, top_p, or logprobs), which is why the entry is recorded rather than Unverified. A live re-probe is still worth doing.
• Responses-only. The reasoning guide still says "Chat Completions does not support function calling with GPT-6 Astra", so the adapter omission from Add OpenAI GPT-6 Astra #285 stands.
• The unimplemented API surface. Async tool calling, mid-turn steering, configuration_update, and misalignment monitoring remain separate pieces of work. configuration_update is Astra-only and the most likely next one.
• Stale GPT-5.6 prices, flagged in Add OpenAI GPT-6 Astra #285 as a distinct change.

Verification

make check passes except TestListDirectoryTool_PermissionDenied, which fails identically on the parent commit — it asserts that an unreadable directory is unreadable, and this container runs as root. go build, go vet, and go test are clean in every module, and go mod tidy -diff is clean where dependencies could have moved.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HdkAnjp3sZqRGdCxwhaRL5

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Essentials

Run ID: a9d365f0-c367-49b9-9b68-bae6086287fd

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[myzie]

Heads up: no CI has run on this PR, and none will until the next push to the branch.

This PR opened against feat/openai-gpt-6-astra. go-test.yml triggers on pull_request: branches: [main], so it did not fire, and CodeRabbit skipped for the same reason (its comment above). I retargeted the base to main, but a base change is a pull_request.edited event, which is not in the default trigger types — so it did not dispatch either. list_workflow_runs for this branch is still empty at 568ec84; there is nothing red here, just nothing run.

I am not pushing an empty commit or reopening the PR to kick it, and I have no code change left to make that would justify a push. The next real push to this branch will run the full suite.

In the meantime, everything go-test.yml does was run locally on 568ec84:

• make provider-catalog-check — clean (the generated files match catalog.json)
• go test ./... (root)
• cd experimental/cmd/dive && go test ./...
• go test ./... in providers/{openai,google,grok,meta}

plus go vet, go build, and go mod tidy -diff across every module including a2a, otel, experimental/mcp, examples, and both demos.

One pre-existing failure, unrelated to this change: TestListDirectoryTool_PermissionDenied fails identically on the parent commit 3bf561a. It chmods a directory unreadable and asserts the read fails; the container runs as root, for which the mode bits do not apply. It passes in CI, which runs as a normal user.

The gap CI would still cover that I could not: the workflow has provider API keys in its environment, so any integration test gated on OPENAI_API_KEY was skipped here. That matters for this PR specifically — the gpt-6-astra capability and prompt-caching entries are derived from OpenAI's published docs, not from a live probe, and a live re-probe is still worth doing before this is treated as verified.

---

Generated by Claude Code