[WIP] LangGraph-based functional Kaggle solver

[YuanmingLeee]

Implement a refactored Kaggle workflow using LangGraph's functional API, providing a modular design for easier readability and maintenance. Each task is encapsulated as a callable node, streamlining collaboration and functionality swapping.

This change improves flexibility and scalability for managing Kaggle competitions while maintaining seamless orchestration.

Closes #314

Demo

1. Download dataset

mle bench prepare -c aerial-cactus-identification

2. Start vLLM Server

vllm serve Qwen/Qwen3-4B-AWQ --enable-auto-tool-choice --tool-call-parser hermes --reasoning-parser deepseek_r1

3. Run kaggle competition workflow

python ./kaggle_solver.py -c aerial-cactus-identification --model Qwen/Qwen3-4B-AWQ test

Demo logs with vllm+qwen3-4B-AWQ model backend:

Step1 Setup env, collect basic metadata:

Step2 Advisory Summary:

Step3 Planer:

Step4 Code for each planned tasks sequentially

... (nothing to change for the last step) Kaggle workflow completed!

TODO

• Project setup with a venv (simple sandbox)
• Coder output to a file
• Debuger actually run the code, rather than just view it, and collect the debugging logs
• Add plan: output a submission csv

Before submitting this PR, please make sure you have:

• confirmed all checks still pass OR confirm CI build passes.
• verified that any code or assets from external sources are properly credited in comments and/or in
  the credit file.

[HuaizhengZhang]

Looks good

[HuaizhengZhang]

working dir is all you need 😏

[YuanmingLeee]

Previous prompt flows were too complex for integration with vLLM, particularly when involving tool calling.

So I copy and simplified previous agent workflows to exp folder; add some env collection steps for better enforcement with 8B Qwen model (seems like this is the most suitable model for balancing perf and speed with one 4090)

More simplification can be achieved with language graph message passing. 2-3 more days needed for refinement

[HuaizhengZhang]

integration with vLLM, particularly when involving tool calling.

I feel chances.

[YuanmingLeee]

@HuaizhengZhang A workable example using Qwen3-4B, the coding procedure seems neat and minimal. The debugger agent needs to be added to ensure the code is executable without error and produces submission.csv.

Do not review exp/decorators.py; it's for experimental purposes and not yet in effect to simplify the agent code.