Description

Bug Report

Description

When using Claude Opus 4.7 via Amazon Bedrock, OpenCode can replay assistant reasoning/thinking blocks through the interleaved reasoning transform in packages/opencode/src/provider/transform.ts.
That transform is intended for OpenAI-compatible providers (reasoning_content / reasoning_details style fields), but it can also run for Bedrock models if the model catalog marks them as interleaved.
For Bedrock Claude models, this is invalid: native Anthropic/Bedrock thinking blocks must be replayed as content blocks with their original metadata/signature preserved. Rewriting them into providerOptions.openaiCompatible[...] strips the native structure and can cause Bedrock validation failures.

Error

Example failure:

undefined: The model returned the following errors:
{"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0.thinking.signature: Field required"}}
``` txt
Root cause
In packages/opencode/src/provider/transform.ts, normalizeMessages() has this path:
```txt
if (typeof model.capabilities.interleaved === "object" && model.capabilities.interleaved.field) {
  // collect reasoning text
  // remove reasoning blocks from content
  // move text to providerOptions.openaiCompatible[field]
}

This path is too broad. It is not restricted to @ai-sdk/openai-compatible, so Bedrock models can be routed through logic that is only valid for OpenAI-compatible providers.
For Anthropic models on Bedrock, reasoning blocks should remain native reasoning/thinking content with provider metadata preserved, not be rewritten into openaiCompatible.reasoning_content.
Why this matters for Opus 4.7
Opus 4.7 uses adaptive thinking and Bedrock expects prior thinking blocks to be replayed unchanged. If OpenCode rewrites them into an OpenAI-compatible shape, the request no longer matches the Bedrock/Anthropic format and signatures can no longer validate correctly.
Proposed fix
Restrict the interleaved rewrite to OpenAI-compatible providers only:

if (
  model.api.npm === "@ai-sdk/openai-compatible" &&
  typeof model.capabilities.interleaved === "object" &&
  model.capabilities.interleaved.field
) {
  ...
}

This keeps the interleaved behavior for providers that actually use reasoning_content / reasoning_details, while preserving native reasoning blocks for Bedrock Claude models.
Related issues / PRs
This appears related in symptom space, but not identical in root cause:

• fix: reasoning text trimEnd() in processor.ts invalidates Bedrock extended thinking signature #18078
• Error on session compaction on Opus 4.6 #17911
• fix: preserve thinking block signatures and fix compaction headroom asymmetry #14393
• fix(provider): skip empty-text filtering for assistant messages in normalizeMessages (#16748) #16750
• fix(provider): preserve assistant message content when reasoning blocks present #21370
  Those focus on signature preservation, empty text blocks, trimming, or compaction. This issue is specifically about the wrong provider transform being applied to Bedrock Claude replay.
  Environment
• OpenCode: current dev
• Provider: amazon-bedrock
• Model: Claude Opus 4.7

Plugins

oh-my-openagent

OpenCode version

1.14.20

Steps to reproduce

1. Configure OpenCode to use Amazon Bedrock.

2. Select Claude Opus 4.7 as the model.

3. Use a reasoning-heavy variant such as xhigh or max so adaptive thinking is likely to be emitted.

4. Start a fresh session.

5. Send a prompt that reliably triggers extended thinking, for example:
   Compare three approaches for migrating a large TypeScript monorepo from one provider SDK to another, including risks, rollout strategy, and rollback plan.

6. After the assistant responds, send a follow-up in the same session that depends on the previous turn, for example:
   Now turn that into a concrete phased migration plan with checkpoints and failure criteria.

7. If it does not fail on the first follow-up, continue with 1–2 more follow-ups in the same session so OpenCode replays prior assistant thinking blocks.

8. Observe the next Bedrock request fail with an error like:
   undefined: The model returned the following errors:
   {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0.thinking.signature: Field required"}}

Screenshot and/or share link

Operating System

Ubuntu 22.04 on WSL

Terminal

bash