Bug: Switching from GitHub Copilot (GPT) to Bedrock Claude in same session fails with thinking.signature: Field required #24909

Summary

When switching models mid-session from GitHub Copilot (ChatGPT / GPT-x) to Bedrock Claude, the next request to Claude fails because the previously-generated assistant messages from the GPT model do not contain the thinking.signature field that Anthropic's API requires for thinking content blocks.

Environment

• Provider A (used first): github-copilot — a GPT model (ChatGPT)
• Provider B (switched to): amazon-bedrock — global.anthropic.claude-* (e.g. Claude Opus / Sonnet with extended thinking)
• Same OpenCode session / conversation (message history is reused)

Steps to Reproduce

1. Start an OpenCode session using a GitHub Copilot GPT model (e.g. GPT-5 / ChatGPT).
2. Have at least a few turns of conversation so the history grows (≥ 5–6 messages).
3. Without starting a new session, switch the model to a Bedrock Claude model that supports extended thinking.
4. Send any follow-up message.

Actual Behavior

Request to Bedrock Claude fails immediately with:

undefined: The model returned the following errors: {"type":"error","request_id":"req_hnoklttbjgqogsmlp7yqffgj4teomlda3ufeaf76qw2gslcadwxa","error":{"type":"invalid_request_error","message":"messages.5.content.0.thinking.signature: Field required"}}

The session becomes unusable on Claude until history is trimmed / a new session is started.

Expected Behavior

Switching providers/models within the same session should work seamlessly. Specifically, when replaying history to a Claude model, OpenCode should either:

1. Strip / convert non-Anthropic assistant blocks (including any synthetic thinking blocks) so Claude only sees plain text content, or
2. Properly sign thinking blocks (only possible if they were originally produced by Claude), or
3. Omit thinking blocks entirely from the replayed history when the originating model was not an Anthropic thinking model.

The current behavior passes through content that looks like an Anthropic thinking block but lacks the required signature field, which Anthropic rejects.

Additional Notes

• The same cross-provider switching pattern also occasionally triggers a different error when skills are involved in the session history. That one reproduces with much lower probability, but the root cause looks related: history produced under one provider is not normalized before being replayed to another.
• Error message prefix in the UI shows undefined: which suggests the provider/label isn't being set on the error surface — minor, but worth cleaning up.

Suggested Fix Direction

When constructing the request payload for Anthropic/Bedrock Claude, sanitize the message history:

• Drop thinking blocks that did not originate from a Claude thinking response (i.e. no valid signature).
• Alternatively, convert any foreign "reasoning"/"thinking" content to plain text before sending.