fix: thinking block signature lost when model differs, breaking multi-turn with extended thinking

[bainos]

Description

When using an Anthropic model with extended thinking enabled, the following error appears mid-conversation:
The model returned the following errors: messages.N.content.1: thinking or redacted_thinking blocks in the latest assistant message cannot be modified.

Root cause is in packages/opencode/src/session/message-v2.ts. When reconstructing conversation history, reasoning parts strip providerMetadata when the message was produced by a different model:

...(differentModel ? {} : { providerMetadata: part.metadata }),

This is correct for text and tool parts — it prevents provider-specific data (e.g. OpenAI itemId) from leaking into other providers. But Anthropic's thinking blocks carry a cryptographic signature inside providerMetadata. Anthropic verifies this signature on every subsequent turn. If it is missing, the block is considered modified and the request is rejected.
The fix is to always preserve providerMetadata for reasoning parts regardless of the differentModel flag.

---

Steps to reproduce

1. Enable extended thinking on a Claude 3.7+ model
2. Have a conversation long enough to accumulate thinking blocks in history, or switch models mid-session and switch back to Anthropic
3. Continue the conversation — the error fires on any turn that replays a thinking block without its signature

---

OpenCode version: 1.4.6
OS: WSL
Terminal: Alacritty

Plugins

No response

OpenCode version

1.4.6

Steps to reproduce

1. Enable extended thinking on a Claude 3.7+ model
2. Have a conversation long enough to accumulate thinking blocks in history, or switch models mid-session and switch back to Anthropic
3. Continue the conversation — the error fires on any turn that replays a thinking block without its signature

Screenshot and/or share link

No response

Operating System

Win 11 (WSL)

Terminal

Alacritty

[github-actions]

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

• Thinking blocks in assistant messages are modified/dropped, causing API error on multi-turn conversations #18254: Same error (thinking blocks modified/dropped causing API error in multi-turn conversations with extended thinking)
• Claude Opus 4.5 (latest) eventually fails: thinking block cannot be modified #13286: Same "thinking block cannot be modified" error on multi-turn Claude sessions with extended thinking
• Anthropic subagents fail with reasoning.effort due to thinking signature replay in multi-turn sessions #16246: Same thinking signature replay failure in multi-turn sessions with Anthropic extended thinking
• fix: reasoning text trimEnd() in processor.ts invalidates Bedrock extended thinking signature #18078: Related — a different root cause (trimEnd() invalidating the signature) but same error and same general area (providerMetadata/signature not being preserved correctly)