Add OpenAI GPT-6 Astra

[myzie]

OpenAI published gpt-6-astra today. The provider watcher found it in the models index, pricing page, and changelog. The account API still does not list it and a direct call returns model_not_found — access is gated to the Trusted Access Program, with broader API access "in the coming days". Rechecked since: /v1/models still returns 130 models with no gpt-6*, /v1/models/gpt-6-astra and /v1/responses still answer model_not_found, and the docs are unchanged.

So this lands the support we can land from documentation and the SDK, and marks the parts that need a live probe before they can be trusted. It also carries a provider-wide retry fix that came out of reading the same changelog.

What's here

Catalog and constants — openai.ModelGPT6Astra (gpt-6-astra), 1.05M context, Responses API only. Deliberately not recommended and not default: the CLI should not offer a model that currently answers model_not_found.

Capabilities (providers/modelcaps) — reasoning effort low → max, with none excluded and temperature refused. This is documented, not probed. It is recorded rather than left Unverified because both exclusions are stated outright in the release notes, and passing them through would send a request already known to fail.

Pricing — the first OpenAI entry to use the long-context tier: above 272K input tokens, input and cache reads bill at 2x and output at 1.5x.

openai-go v3.56.0 — OpenAI cut the SDK release the same day ("add gpt-6-astra and related features"). gpt-6-astra is the only model id it adds, and it matches the entry written from the docs. It also brings the types behind the rest of the release: Async on function and custom tool params, ResponseConfigurationUpdateItem, ResponseSteerErrorCode / ResponseSteerPendingReason, and the misalignment error objects. Dive does not use any of them yet — this is just the dependency they need.

Watcher fix — a model linked as <id>.md in an upstream index was filed as a second, phantom missing model beside the real one. Today's report showed gpt-6-astra, gpt-6-astra.md, and gpt-6 as three separate gaps. Re-running the audit after this change drops all three (the bare gpt-6 clears on its own once Astra is catalogued).

Retry handling

Two problems, both from OpenAI's Sep 2 error-code change and both provider-wide rather than Astra-specific:

Retry-After was ignored. grep -rn "Retry-After" across the repo returned nothing. Providers retried on the status code alone with their own backoff schedule, so Dive came back before the provider was ready and prolonged the condition it was waiting out. OpenAI now sends the header with 429 slow_down (traffic ramped too fast) and 503 server_is_overloaded. The header is now honored in place of the exponential backoff for the next attempt, capped at the policy's MaxWait so a provider cannot park a caller for ten minutes. Both header forms RFC 9110 allows are parsed — a delay in seconds and an HTTP-date.

Terminal 429s were retried. An exhausted credit balance, an organization or project spend limit, and an ordinary rate limit all arrive as 429; only the last is worth another attempt. Dive burned its full attempt budget on the others and delayed an error the caller has to act on. The error code now separates them (insufficient_quota, credit_balance_exhausted, and the three spend/usage limits are permanent).

providers.RetryPolicy replaces the retry options each provider assembled inline — Anthropic, Google, OpenAI Responses, OpenAI Chat Completions, and the retrying stream iterator now share one policy. providers.NewError grows WithErrorCode and WithErrorHeader options, and ProviderError exposes Code() and RetryAfter(). Codes reach the classifier two ways: parsed out of the JSON envelope where the provider hands NewError a raw body, and passed explicitly where an SDK has already parsed it. Envelope members are decoded one at a time, because a provider that sends "error" as a bare string — xAI does — would otherwise fail the whole unmarshal and lose a top-level "code" beside it.

Twelve tests cover this in providers/providers_test.go, including the delay actually used, the MaxWait cap, and the fallback to exponential backoff.

Known gaps

• Nothing is API-verified. The effort ladder, the temperature refusal, and the prices are what OpenAI published. Re-probe when access opens; providers/modelcaps/tables.go carries the note.
• Long-context cache writes are understated. The published rate is $25.00/1M; llm.PricingInfo has no field for it, so a cache write on a >272K request is costed at the standard $12.50. TestGPT6AstraLongContextPricing pins this so it is visible rather than silent.
• Explicit prompt-cache breakpoints stay gated to gpt-5.6. Astra documents prompt_caching and a separate cache-write rate — the same two signals that gate 5.6 — so it is a likely candidate. Left off because guessing wrong fails every Astra request, while leaving it off only forgoes an optimization. Comment added at the gate.

New API surface, not implemented here

The release ships four Responses features that Dive has no binding for. Each is its own piece of work:

• Async tool calling — async: true on a function or custom tool lets the model keep working while the application runs the tool; results come back later against the original call_id.
• Mid-turn steering — send user instructions over a WebSocket while a response is in flight, preserving completed work.
• configuration_update — change reasoning effort mid-conversation without rewriting the prompt prefix, so the cache survives. Astra-only, and it cannot be combined with automatic compaction or truncation.
• Misalignment monitoring — asynchronous safety checks that can raise alerts or stop a conversation for review.

Separately noticed, not addressed here

• GPT-5.6 prices are stale. The audit flagged pricing_changed: Sol $5.00/$30.00 → $4.00/$20.00, Terra $2.50/$15.00 → $2.00/$12.00, Luna $1.00/$6.00 → $0.20/$1.20. A live cost-reporting inaccuracy, but a distinct change.
• OpenAI fast mode is unreachable. providers/openai/provider.go maps ServiceTier to auto, default, and flex only, and errors on anything else — so fast and priority cannot be requested even though the SDK defines them. Dive also never decodes the response's service_tier into Usage.Speed, so the fast-pricing branch at provider.go:139 can never fire, and the OpenAI catalog's fast_text pricing table is empty. Pre-existing, but Astra's fast tier ($20.00/$100.00) makes it more visible.
• The docs-index watch source may be dead. catalog.json points at https://developers.openai.com/llms.txt with discovery patterns for /api/docs/(models|pricing|deprecations|changelog). That page is now a hub that links to /api/llms.txt and lists none of those paths, so the discovery source likely matches nothing. The four explicit document sources still work, which is why this went unnoticed.
• Two stale CLI dependencies dropped. make tidy-all removes go-runewidth and uax29 from experimental/cmd/dive; neither is referenced by any Go file there. Pre-existing drift, swept up because the openai-go bump left five modules failing go mod tidy -diff.

Verification

make check passes. go build, go vet, go test, and go mod tidy -diff are clean in every module: root, providers/{openai,google,grok,meta}, a2a, otel, experimental/{mcp,cmd/dive}, examples, and demos/{colosseum,noodleville}.

🤖 Generated with Claude Code

https://claude.ai/code/session_01L21T5Bt9gsaCvdFX6GXRqa

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

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

[myzie]

Superseded by #286, which was stacked on this branch and carried its five commits. The squash-merge of #286 (5834b38) landed all of this PR's content in main alongside its own, so nothing here is outstanding — GitHub just couldn't auto-close it because the squash rewrote the commit SHAs.

Verified in main: ModelGPT6Astra and the catalog entry, the Retry-After / providers.RetryPolicy refactor across Anthropic, Google, OpenAI Responses, OpenAI Completions and retrying_stream.go, and the watcher .md phantom-gap fix. Diffing main against this branch tip shows only removals of #286's additions — main is a strict superset.

Closing and deleting the branch.