ensure performance on all models

[TeenSpirit1107]

because we claimed that our agent works on multiple models.

• gemini
• deepseek (problematic)
• chatgpt (too expensive, cannot try)

[TeenSpirit1107]

Results:

1. cannot use GPT, because: (1) it's too costly; (2) it strictly limits the "token per minute" for non-paying users, and our agent failed due to this.

2. Gemini is significantly better than Deepseek. In our latest version (commit 88be4d1), Gemini-2.5-flash performs well on our sample task.

3. Deepseek's performance remains problematic. Whether it's something intrinsic to the model or can it be solved by us?

   • implement better error-handling mechanism, sinces Deepseek usually fails
   • test multiple rounds of conversation with Deepseek, see whether it can fix its own error in acceptable number of rounds
   • baseline: send the code directly to Deepseek in chat, to see whether it can solve the problem. (if not, this indicates that it may not have the ability, despite improvements on our agent)