fix: strip incompatible thinking blocks when switching to Claude

[coleleavitt]

Fixes #6418

Problem

When switching from models like GLM 4.7 or MiniMax to Claude with extended thinking enabled, users get API errors:

messages.X.content.0: Invalid 'signature' in 'thinking' block

or

messages.X.content.0.type: Expected 'thinking' or 'redacted_thinking', but found 'text'

This happens because:

1. Other models produce thinking blocks with signatures that are incompatible with Claude's signature validation
2. Claude requires the last assistant message to start with a valid thinking block when thinking is enabled
3. reasoning parts from other models don't have the signature format Claude expects

Solution

In normalizeMessages(), when the target model is Claude:

1. Convert thinking blocks with signatures to wrapped text - Preserves the reasoning content while removing the invalid signature. Wrapped in <assistant_thinking> tags.
2. Convert reasoning parts to text - Same approach, wrapped in <assistant_reasoning> tags.
3. Track which messages had thinking blocks converted - Using a Set to avoid incorrectly removing messages that we already handled.
4. Only remove last assistant message if safe - Only when it doesn't start with valid thinking, has no tool calls, AND didn't have converted thinking blocks.

This approach preserves tool_use/tool_result pairing while still handling the signature validation issues.

Changes

• packages/opencode/src/provider/transform.ts: Added thinking block conversion logic in Claude-specific section
• packages/opencode/test/provider/thinking-blocks.test.ts: New test file with comprehensive coverage
• packages/opencode/test/provider/transform.test.ts: Updated existing test expectation

Testing

1. Start a session with GLM 4.7 or MiniMax (models that generate thinking/reasoning)
2. Send a few messages, including one that triggers tool use
3. Switch to Claude with extended thinking enabled
4. Send a message - should no longer get signature errors
5. Tool results should still work correctly

Related

• Supersedes fix: strip incompatible thinking blocks when switching to Anthropic models #6748 (similar approach but that PR has been stale)
• Related to GLM 4.7 thinking process not properly formatted - missing leading tags #7779 (GLM 4.7 thinking format issues)

[github-actions]

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

Potential Duplicate Found

PR #6748: fix: strip incompatible thinking blocks when switching to Anthropic models
#6748

Why it's related:

• This is the exact same fix that the current PR supersedes (as mentioned in the PR description)
• Both address the same issue: stripping incompatible thinking blocks when switching to Claude/Anthropic models
• Both involve message normalization and signature validation errors
• PR fix: strip incompatible thinking blocks when switching to Anthropic models #6748 has been stale, while fix: strip incompatible thinking blocks when switching to Claude #8958 appears to be the updated approach

The current PR (#8958) explicitly states it "Supersedes #6748 (similar approach but that PR has been stale)", so you should check if #6748 needs to be closed or if this PR is intended to replace it.

[coleleavitt]

Summary

Fixes the Invalid 'signature' in 'thinking' block error when switching from models like GLM 4.7 or MiniMax to Claude with extended thinking enabled.

The Fix

1. Convert thinking blocks with signatures to wrapped text - Preserves reasoning content in <assistant_thinking> tags while removing incompatible signatures
2. Convert reasoning blocks to wrapped text - Same approach with <assistant_reasoning> tags
3. Track converted messages - Prevents incorrectly removing messages we already handled
4. Safe last-message removal - Only removes last assistant if it doesn't start with valid thinking, has no tool calls, AND didn't have converted thinking

This preserves tool_use/tool_result pairing while handling signature validation.

Tested

• MiniMax → Claude with extended thinking: ✅ No signature errors
• Tool use preserved across model switches: ✅

[d-init-d]

Detailed Code Review

Thanks for tackling this issue @coleleavitt! The approach is solid - converting incompatible thinking blocks to wrapped text is the right direction. However, I've identified some issues that should be addressed before merging.

🐛 Critical Bugs

1. Devstral support regression (line ~133 in patch)
The original code includes devstral in the Mistral check, but this PR removes it:

// Original:
if (model.providerID === "mistral" || model.api.id.toLowerCase().includes("mistral") || model.api.id.toLocaleLowerCase().includes("devstral"))

// PR:
if (model.providerID === "mistral" || model.api.id.toLowerCase().includes("mistral"))

This will break devstral model support!

2. Inconsistent tag naming
PR description mentions <assistant_reasoning> for reasoning parts, but code uses <assistant_thinking> for both. This inconsistency makes it harder to distinguish between converted thinking vs reasoning blocks when debugging.

⚠️ Edge Cases Not Handled

3. Empty content array after filtering
After converting and filtering parts with .filter((part: any) => part !== null), msg.content could become []. Claude API rejects empty content arrays with validation errors. Should add a check to filter out messages with empty content:

if (msg.content.length === 0) return undefined
// Then filter undefined messages at the end

4. Valid Claude thinking blocks could be incorrectly converted
Currently, ANY thinking block with a signature gets converted. What about valid Claude thinking blocks that have proper Anthropic signatures from a previous Claude turn in the same session?

Consider checking signature format:

const isValidClaudeSignature = part.signature?.startsWith('ErUB')
if (part.type === "thinking" && part.signature && !isValidClaudeSignature) {
  // Convert only invalid signatures
}

🔧 Suggested Improvements

5. Type safety
Multiple uses of any casts ((part: any), (options as any)) could be replaced with proper types for better maintainability.

6. Extract helper function
The Claude-specific block has grown significantly (from ~10 lines to ~50+ lines). Consider extracting to a dedicated function like normalizeClaudeMessages() for better readability.

7. Missing test cases in thinking-blocks.test.ts
Good test coverage overall, but consider adding:

• Empty thinking text (edge case)
• Multiple thinking blocks in single message
• Valid Claude thinking blocks (should NOT be converted)
• Session with mixed Claude + non-Claude turns

---

I'm working on an improved PR that addresses these issues. Happy to collaborate if you'd like to incorporate these fixes into your PR instead!

Files reviewed:

• packages/opencode/src/provider/transform.ts - Core logic changes
• packages/opencode/test/provider/thinking-blocks.test.ts - New test file
• packages/opencode/test/provider/transform.test.ts - Test expectation update