feat: support Claude models on the Bedrock Mantle endpoint

[ktoulgaridis]

Issue for this PR

Closes#43230

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Lets Claude models be used on the bedrock-mantle endpoint.

Mantle serves Claude only through the Anthropic Messages API at /anthropic/v1/messages, and nothing bundled reaches that:

• @ai-sdk/amazon-bedrock/mantle is OpenAI-compatible only, and Mantle rejects Anthropic IDs there — The model 'anthropic.claude-opus-4-8' does not support the '/openai/v1/responses' API
• @ai-sdk/amazon-bedrock/anthropic speaks Messages, but hardcodes buildRequestUrl to the bedrock-runtime/model/{id}/invoke path and moves the model ID into an anthropic_version body field
• @ai-sdk/anthropic speaks Messages and takes a baseURL, but authenticates only with a key, so profile/SSO users would need a Bedrock API key refreshed every 12h

So this adds @ai-sdk/amazon-bedrock/mantle-anthropic: AnthropicMessagesLanguageModel pointed at the Mantle base URL, with a fetch that SigV4-signs the request (service bedrock-mantle) using the credential provider the bedrock loader already builds from region/profile. Signing in fetch is the only request-time auth hook the AI SDK offers; it wraps the fetch opencode already puts in options, so the existing timeout and SSE-abort handling still runs. Bearer tokens (AWS_BEARER_TOKEN_BEDROCK or options.apiKey) keep their current precedence and skip signing.

Two fields the SDK sends by default are rejected by Mantle, both verified live:

• output_config.format → 400 Extra inputs are not permitted, which broke every generateObject call (e.g. agent create). supportsNativeStructuredOutput: false routes structured output through the SDK's JSON tool instead.
• a tool-level strict field → 400 tools.0.custom.strict: Extra inputs are not permitted, so supportsStrictTools: false is set too.

Both are model-config-only settings, so the model is built from the declared @ai-sdk/anthropic/internal subpath rather than createAnthropic — the same way @ai-sdk/google-vertex/anthropic builds its model. One visible consequence: requests no longer carry the ai-sdk/anthropic/<version> User-Agent suffix (@ai-sdk/google-vertex/anthropic drops it too). provider.tools and the embedding/image model stubs go with it; nothing here uses them.

The key also has to be threaded through the Anthropic paths in the transform layer, alongside @ai-sdk/google-vertex/anthropic — sdkKey (otherwise per-model providerOptions are silently dropped), the empty-content guard, usesAnthropicAutomaticCaching, variants/reasoningEffort/reasoningBudget (otherwise reasoning is silently off), supportsMediaInToolResult, and the v1 config lowerers. The Responses-API-only lists correctly exclude it.

Base URL is derived from region, so no api override is needed — that also keeps it working on v2, where ${AWS_REGION} is not substituted (#40075). Model IDs pass through unprefixed, since Mantle has no cross-region inference profiles.

{
"provider": {
"amazon-bedrock": {
"options": { "region": "us-east-1", "profile": "my-sso-profile" },
"models": {
"anthropic.claude-opus-4-8": {
"provider": { "npm": "@ai-sdk/amazon-bedrock/mantle-anthropic" }
}
}
}
}
}

Related: #36208 (native llm layer design — this is the AI SDK path), #39325 (Mantle selector keyed to the amazon-bedrock provider id).

How did you verify your code works?

• bun typecheck clean in packages/core and packages/opencode; full suites green (packages/core 1101 pass, packages/opencode 3322 pass)
• New tests cover SigV4 signing, streaming, the API-key path, static env credentials, the missing-credentials error, JSON-tool routing, and the absence of tool strict; plus the transform-layer edits and both the v1 and v2 wirings. Every production edit was mutation-tested — reverting any one of them fails a test.
• Live against us-east-1 with an SSO profile, on a binary from ./packages/opencode/script/build.ts --single: anthropic.claude-opus-4-8 and anthropic.claude-opus-5 answer over Mantle, tool calls execute and their real output is used, --variant high --thinking streams reasoning, and agent create (the path that previously 400'd) writes a schema-valid agent file. openai.gpt-5.6-sol still works, so the existing Mantle OpenAI route is unaffected.
• Capability limits were probed directly against the endpoint and are documented: fixed-budget thinking.type: "enabled" is rejected on Opus 4.8/Opus 5/Sonnet 5 (accepted on Haiku 4.5), while adaptive thinking with output_config.effort is accepted.
• Reviewer repro: aws sso login --profile <profile>, add the config above, then opencode run -m amazon-bedrock/anthropic.claude-opus-4-8 "hi"

Note: packages/core/test/plugin/provider-amazon-bedrock.test.ts already fails on dev when run as a single file (TDZ on AmazonBedrockPlugin via a circular import); it passes as part of bun test test/plugin. Untouched here.

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[Enough1122]

AI code review — automated review for reference; please use your judgment.

• packages/core/src/plugin/provider/amazon-bedrock.ts:117 — createBedrockMantleAnthropic(options) receives the user's options (which per the docs include profile), but BedrockMantleAnthropicSettings has no profile handling and no default credential-chain resolution (env vars / SSO / IRSA). Without an explicit credentialProvider, SigV4 mode always throws "requires either an API key or AWS credentials". The docs claim "profiles, SSO, and IRSA work as they do elsewhere in Bedrock" — that looks untrue as implemented. Either resolve credentials from profile/env here or fix the docs.
• packages/core/src/amazon-bedrock/mantle-anthropic.ts:60 — body: init?.body as string: if the AI SDK ever passes a non-string body (stream/Uint8Array), SigV4 signing will produce an invalid payload hash. Consider asserting/handling non-string bodies explicitly.
• packages/core/src/amazon-bedrock/mantle-anthropic.ts:33 — placeholder apiKey: "sigv4" is clever but means a leaked-looking dummy key header is constructed on every request before deletion; a comment exists — good. Just make sure no retry path bypasses the wrapping fetch.
• packages/core/test/amazon-bedrock/mantle-anthropic.test.ts — unit tests cover signing and bearer paths well; consider adding a test asserting the helpful error when neither key nor credentials exist.
• Docs addition is clear, including the no-inference-profile warning.

[ktoulgaridis]

Thanks — 4 is fair, added. On the rest:

1. The credential chain isn't meant to live in mantle-anthropic.ts; the callers inject it, exactly as for @ai-sdk/amazon-bedrock and /mantle. packages/core/src/plugin/provider/amazon-bedrock.ts gates on evt.package, the new package is in that allowlist, and it then sets credentialProvider = fromNodeProviderChain(profile ? { profile } : {}) from options.profile ?? AWS_PROFILE. Same on v1 through the amazon-bedrock loader. Verified live with an SSO profile and no key.

There was a real gap next door, now closed: unlike createAmazonBedrock, this had no fallback to static AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY when nothing is injected. Fixed and tested. The other half — v1 wiring the chain only to the literal id amazon-bedrock — is pre-existing and identical for /mantle; that's #39325.

2. Unreachable today: @ai-sdk/anthropic only calls postJsonToApi, which always stringifies. aws4fetch throws rather than mis-hashing if handed anything else, and an absent body hashes to the correct empty-payload digest. Leaving the cast.

3. Retries re-enter doGenerate/doStream, so each attempt re-signs — required anyway, since SigV4 signatures expire; provider-utils has no retry of its own. The x-api-key header is deleted before signing, which now also covers a caller-supplied one via options.headers.

4. Added.

Your review pointed at the right file, though, and digging there turned up two real 400s I've now fixed:

• output_config.format was still being sent, so any generateObject call — e.g. agent create — failed with output_config.format: Extra inputs are not permitted. The model config now sets supportsNativeStructuredOutput: false, so structured output takes the SDK's JSON-tool path, the same thing @ai-sdk/google-vertex/anthropic does.
• A tool-level strict field failed the same way (tools.0.custom.strict), so supportsStrictTools: false is set too.

Reaching those flags means building AnthropicMessagesLanguageModel from @ai-sdk/anthropic/internal rather than createAnthropic, since AnthropicProviderSettings can't express them.

Separately, and bigger than any of the above: the new npm key was invisible to the transform layer, so reasoning/effort variants were never generated, providerOptions weren't remapped to the anthropic key, and the empty-content and tool-result-media checks skipped it. All wired now, with tests, and every edit is mutation-tested.