Problem with AWS Bedrock

[franjorub]

Description

I have a configuration to use two different models from Bedrock, Opus for plan, Sonnet for build.

if I use the same model in both agents, its fine.

but if I use them separately, when I chat with one model, and continue the conversation in the other I get this error

undefined: The model returned the following errors: messages.1.content.0.thinking.signature: Field required

I have tested this methodology of different models with plan/build in other code assistants compatible with Bedrock and the problem doesn't happen

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Bedrock Claude Opus 4.7 replays thinking blocks through interleaved OpenAI-compatible transform #23926: Bedrock Claude replays thinking blocks through wrong transform, causing the same thinking.signature: Field required error
• fix: thinking block signature lost when model differs, breaking multi-turn with extended thinking #22813: Thinking block signature lost when switching models mid-session, also causing signature validation failures on Bedrock/Anthropic

If your issue is not fully covered by those, please add details (OpenCode version, OS, terminal) so it can be tracked separately.

[shakedlokits]

I think it might be possible to resolve this with something like:

export const plugin: Plugin = async () => {
  let isAnthropicBedrock = false;

  return {
    "chat.params": async (input, _output) => {
      isAnthropicBedrock =
        input.model.providerID === "amazon-bedrock" &&
        input.model.api.id.includes("anthropic");
    },
    "experimental.chat.messages.transform": async (_input, output) => {
      if (!isAnthropicBedrock) return;
      for (const msg of output.messages) {
        msg.parts = msg.parts.filter((part) => part.type !== "reasoning");
      }
    },
  };
};

[rekram1-node]

ill try fixing this today

[Dulani]

We hit the same messages.1.content.0.thinking.signature: Field required error when switching models mid-session (Azure OpenAI GPT-5.4 → AWS Bedrock Claude Opus) and after investigating we've identified the same root cause you're encountering with Plan(Opus) ↔ Build(Sonnet) on Bedrock.

Root Cause

packages/opencode/src/session/message-v2.ts in toModelMessages() replays assistant parts when building conversation history. For type: "reasoning" it currently does:

assistantMessage.parts.push({
  type: "reasoning",
  text: part.text,
  ...(differentModel ? {} : { providerMetadata: part.metadata }),
})

When differentModel is true (either different provider or different model), it strips providerMetadata but keeps the reasoning part. The downstream Anthropic/Bedrock path then serializes that part as a Claude thinking block without a valid signature, and Bedrock rejects the request with exactly the error you're seeing.

For your Plan(Opus) ↔ Build(Sonnet) case, differentModel is true (same provider, different model), so Claude's own signature gets stripped on replay — even though it would have been valid to pass through unchanged.

Fix Applied

We introduced differentProvider and split the behavior:

• different provider → drop the reasoning part entirely (can't translate foreign reasoning into Anthropic thinking safely)
• same provider, different model → preserve the reasoning part and its providerMetadata, so signatures from one Claude model can round-trip into the next

const differentProvider = model.providerID !== msg.info.providerID
// ...
if (part.type === "reasoning") {
  if (differentProvider) continue
  assistantMessage.parts.push({
    type: "reasoning",
    text: part.text,
    providerMetadata: part.metadata,
  })
}

Verification

• Two-turn repro (GPT-5.4 → Bedrock Opus) reproduced the failure deterministically before the fix, passes cleanly after
• Switching a live session with existing Claude thinking blocks also works
• Added focused tests for both cross-provider drop and same-provider preservation

Related Issues

This issue appears related in symptom to #22813 (which focuses on Anthropic thinking block signature validation) and #23926 (which identifies a different code path in transform.ts). All three converge on the same family of errors when mixing providers or models with extended thinking enabled.

Ready to submit as a PR if helpful.

[rekram1-node]

@Dulani im going to fix it, this solution u propose "works" but it is poor. Also please dont paste giant AI messages

[Dulani]

@rekram1-node - Apologies for the "giant AI message." I do most of my patching/PRs with LLMs and the larger messages are super helpful for giving the LLM the context it needs to do the job (I don't really have the Typescript expertise needed to write the code all on my own, but I still want to be helpful).

I will keep them shorter in the future (wish I could attach more context instead). In any case, thanks for your GREAT work on this harness, it's been a pleasure to work in/with.

[xer0x]

Thanks for working on this. As a Bedrock user it's been a frustrating surprise to get stuck on this. Running /compact on this history is my workaround at the moment.

[josechirivella]

I'm planning with Opus, and my executor is with Sonnet, but when I do the switching, that's when I get the error. My solution at the moment is to get the planning in Markdown and copy-paste to a new chat with Sonnet, but I've been getting mixed results.