fix(session): strip reasoning blocks for Claude Code compatibility

[romancircus]

Summary

Strip reasoning blocks from message history when converting to model format, enabling seamless switching between reasoning and non-reasoning models.

Problem

When using a reasoning model (e.g., Claude with extended thinking) and then switching to Claude Code, the session fails with an "Invalid signature" error. This happens because Claude Code's API doesn't recognize the reasoning blocks that were added by the previous model.

Solution

Filter out reasoning type parts from messages in convertToModelMessages() before sending to the model API. This ensures compatibility regardless of which models were used earlier in the session.

Changes

• packages/opencode/src/session/message-v2.ts - Filter reasoning parts in message conversion

Testing

• Verified typecheck passes
• Manual testing: Start session with reasoning model, switch to Claude Code, continue conversation without errors

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Related PRs Found:

1. fix: strip incompatible thinking blocks when switching to Anthropic models #6748 - fix: strip incompatible thinking blocks when switching to Anthropic models

   • fix: strip incompatible thinking blocks when switching to Anthropic models #6748
   • Why related: This PR appears to address a very similar issue - stripping incompatible blocks (thinking blocks) when switching between models. This is essentially the same problem domain as your PR (fix(session): strip reasoning blocks for Claude Code compatibility #7035), which strips reasoning blocks for Claude Code compatibility.

2. Feature/OpenAI compatible reasoning #5531 - Feature/OpenAI compatible reasoning

   • Feature/OpenAI compatible reasoning #5531
   • Why related: This PR deals with reasoning models and OpenAI compatibility, which relates to the cross-model reasoning compatibility issue you're fixing.

3. fix(session): prevent GPT-5.2 resume reasoning crash #6114 - fix(session): prevent GPT-5.2 resume reasoning crash

   • fix(session): prevent GPT-5.2 resume reasoning crash #6114
   • Why related: Also addresses reasoning-related crashes in session management, though specific to GPT-5.2 resume scenarios.

Recommendation: PR #6748 appears to be the most closely related - it may have already implemented a solution for a similar issue with thinking/reasoning blocks. You should check if that PR's approach can be unified with your fix, or if they should be consolidated.

[rekram1-node]

this will not work at all, namely any stateful reasoning model will completely break, not to mention that anthropic models DO require reasoning to be sent back during the current loop and I think this would error:

[romancircus]

Thanks for the quick feedback @rekram1-node! You're absolutely right - stripping reasoning blocks unconditionally would break stateful reasoning models that need these blocks sent back during the conversation loop.

I see PR #6748 addresses this more correctly at the provider transform layer. I'll close this PR and defer to that approach.

Appreciate the review!