Deepseek custom model fails with MCP tool calls due to array-based message.content

When using a custom Deepseek chat model in AI Toolkit with MCP enabled, each time a tool call (e.g. browser open) occurs, the subsequent model request fails:

Error:
400 Failed to deserialize the JSON body into the target type: messages[1]: invalid type: sequence, expected a string

Root Cause:
The request payload uses an OpenAI "content parts" array for messages:
"content": [
{ "type": "text", "text": "Open the browser" },
{ "type": "text", "text": "Tool result ..." }
]
Deepseek API expects:
"content": "Open the browser\nTool result ..."

Requested Fix:
Add model-specific serialization or a compatibility flag to flatten multi-part content arrays into a single string when invoking non-OpenAI providers (e.g., Deepseek).

Reproduction Steps:

1. Add agent with MCP (e.g., browser tool)
2. Configure custom model pointing to Deepseek chat endpoint
3. Trigger tool invocation
4. Observe 400 error above

Environment:

• AI Toolkit version: x.y.z
• VS Code version: ...
• OS: ...
• Deepseek endpoint: ...
  Please let me know if additional logs are needed.