fix: keep native structured output for Claude Opus 5.5

[anandpant]

Summary

Claude Opus 5.5 rejects forced tool use, so the structuredOutputMode: 'jsonTool' that the Anthropic adapter otherwise requests fails with HTTP 400 through the Foundry proxy. Request the provider's native output format for that model instead.

• Add requiresAnthropicNativeStructuredOutput next to the catalog entries and thread it through createFoundryAnthropic, matching on both the alias and the RID. Every other Claude model keeps the existing JSON-tool mode.
• Set structuredOutputMode: 'outputFormat' explicitly rather than omitting the option. The SDK resolves its default 'auto' from the model ID, and this adapter addresses models by Foundry RID, so 'auto' takes the unknown-model path. On @ai-sdk/anthropic 3.0.64, 3.0.95 and 4.0.5 — all inside the declared peer range — that path reports supportsStructuredOutput: false and still sends the forced JSON tool. 'outputFormat' short-circuits that check on every version from 3.0.64 through 4.0.63.
• The adapter's value is spread after the caller's anthropic options, so a caller passing jsonTool cannot downgrade Opus 5.5 back into the failing path.
• @ai-sdk/anthropic 4.0.60 added an equivalent fallback upstream, but it keys on the claude-opus-5-5 model ID, which a RID never matches. The fix has to live here.

Without this, Opus 5.5 ships GA with standalone JSON and combined JSON + tools both broken by an option this package sets on the caller's behalf.

Validation

New anthropic.test.ts drives createFoundryAnthropic through the real @ai-sdk/anthropic peer with a stubbed fetch and asserts the outgoing Anthropic body: output_config.format present and no tools/tool_choice for the Opus 5.5 alias and raw RID, the forced json tool for Claude Opus 5, and no caller downgrade. Reverting to the earlier omit-based approach makes the downgrade case fail, so the coverage is load-bearing. Note that the pinned dev version 4.0.52 has a generic claude- fallback that masks the peer-floor behaviour, so the 3.0.64/4.0.5 path was verified by reading those releases' compiled getModelCapabilities directly.

Affected unit tests, lint, and typecheck passed, and pnpm run verify passed at the top of the stack. A focused live rerun of Opus 5.5's JSON and JSON + tools probes on 4.0.52 passed with output identical to the previous run, so no results table changed.

[anandpant]

• docs: publish September 24 live results #79
• fix: keep native structured output for Claude Opus 5.5 #78 👈 (View in Graphite)
• feat: refresh Foundry model catalog #77
• main

---

How to use the Graphite Merge Queue

Add either label to this PR to merge it via the merge queue:

• merge - adds this PR to the back of the merge queue
• fast - for urgent changes, fast-track this PR to the front of the merge queue

You must have a Graphite account in order to use the merge queue. Sign up using this link.

An organization admin has enabled the Graphite Merge Queue in this repository.

Please do not merge from GitHub as this will restart CI on PRs being processed by the merge queue.

This stack of pull requests is managed by Graphite. Learn more about stacking.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Repo admins can enable using credits for code reviews in their settings.

[graphite-app]

Merge activity

• Sep 25, 5:53 AM UTC: anandpant added this pull request to the Graphite merge queue.
• Sep 25, 5:55 AM UTC: Merged by the Graphite merge queue.