Compatibility Issues with DeepSeek-R1 Model Response Format

[darielnoel]

Describe the Bug

The DeepSeek-R1 model outputs a response format incompatible with KaibanJS's default prompt engineering setup, which expects JSON-formatted responses. This mismatch causes workflows using the DeepSeek-R1 model to fail during execution. The reasoning-based response structure from the model needs specific adjustments to the parsing logic or prompt tuning in KaibanJS.

---

Steps to Reproduce

1. Configure an agent in KaibanJS to use the DeepSeek-R1 model:

   const reasoningAgent = new Agent({
     name: 'Reasoner',
     role: 'Analyze complex data patterns',
     llmConfig: {
       provider: 'deepseek',
       model: 'deepseek-r1',
     },
   });

2. Execute a workflow with the agent.

3. Observe that the workflow fails due to the model's reasoning-based response format.

---

Current vs. Expected Behavior

• Current Behavior: The agent fails to process the output from DeepSeek-R1 due to its non-JSON response format, leading to workflow interruptions.
• Expected Behavior: The agent should successfully process the model's responses and complete the workflow without issues.

---

Additional Notes

• DeepSeek-R1 provides a reasoning-based response format, as detailed in their API documentation: DeepSeek Reasoning Model Guide.
• This format requires adjustments to KaibanJS's response handling or prompt engineering to accommodate the reasoning model's output style.
• Reference issue in LangChain regarding similar compatibility concerns: LangChain Issue #29317.
• Potential solutions include extending the prompt engineering to support reasoning-style models or adapting KaibanJS's JSON parsing logic.

---