AI SDK message conversion loses assistant reasoning/thinking content

[hannesrudolph]

Problem

Assistant reasoning/thinking content is not preserved when messages are converted through the AI SDK message conversion layer.

Impact

Providers that rely on the AI SDK cannot reliably round-trip reasoning across tool-call turns, and reasoning content may be lost in persisted history.

Expected behavior

Reasoning/thinking content should be preserved through conversion so it can be carried across turns and persisted without being dropped.

[roomote]

I've analyzed this issue and here's my proposed implementation plan:

Problem Analysis

The convertToAiSdkMessages function in src/api/transform/ai-sdk.ts (lines 127-163) handles assistant messages but only extracts text and tool_use content parts. It ignores:

1. reasoning content blocks (Roo Code's internal format stored in API history)
2. thinking content blocks (Anthropic's format with signature)
3. reasoning_content property at the message level (used by DeepSeek, Z.ai, etc.)

This causes reasoning/thinking content to be lost when messages are converted through the AI SDK layer for providers like xAI, Groq, Cerebras, Fireworks, etc.

Proposed Solution

Modify convertToAiSdkMessages to preserve reasoning content:

1. Extract reasoning from content blocks: When iterating through assistant message parts, also check for reasoning and thinking type blocks and collect them.

2. Extract reasoning from message-level properties: Check for reasoning_content property at the message level (similar to how r1-format.ts and zai-format.ts handle it).

3. Include reasoning in output: Add reasoning parts to the AI SDK message content array as { type: "reasoning", text: string } parts, which the AI SDK supports.

Implementation Details

// In the assistant message handling section:
const reasoningParts: string[] = []

for (const part of message.content) {
    if (part.type === "text") {
        textParts.push(part.text)
    } else if (part.type === "tool_use") {
        toolCalls.push(...)
    } else if ((part as any).type === "reasoning" && (part as any).text) {
        reasoningParts.push((part as any).text)
    } else if ((part as any).type === "thinking" && (part as any).thinking) {
        reasoningParts.push((part as any).thinking)
    }
}

// Also check for top-level reasoning_content
const messageWithReasoning = message as any
if (messageWithReasoning.reasoning_content) {
    reasoningParts.push(messageWithReasoning.reasoning_content)
}

// Include reasoning in output content
if (reasoningParts.length > 0) {
    content.unshift({ type: "reasoning", text: reasoningParts.join("") })
}

Files to Modify

• src/api/transform/ai-sdk.ts: Update convertToAiSdkMessages function
• src/api/transform/__tests__/ai-sdk.spec.ts: Add tests for reasoning preservation

This approach follows the same pattern used in r1-format.ts and zai-format.ts for preserving reasoning content.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.