Context

This PR removes the forced context compression logic for volces.com (Ark) endpoints.

Previously, requests to volces.com were automatically converted to a simplified text format using convertToSimpleMessages. This was likely a workaround for older limitations. However, volces.com now fully supports multi-turn tool calling. The forced compression strips away structured information (like tool calls and results), causing the model to hallucinate or fail to understand previous tool interactions.

Implementation

• Removed the ark variable which detected volces.com in the model URL in src/api/providers/openai.ts.
• Removed the condition ark || from the logic that triggers convertToSimpleMessages.
• Now volces.com endpoints will use the standard OpenAI message format (including convertToOpenAiMessages) unless enabledLegacyFormat is explicitly enabled by the user.

Screenshots

How to Test

1. Configure a Custom OpenAI provider with a volces.com Base URL.
2. Start a task that requires multiple steps or tool usage (e.g., "Check the weather in two different cities").
3. Verify that the model can successfully execute multiple tool calls in sequence without losing context or hallucinating.
4. Ensure that the conversation history is maintained correctly.

Get in Touch