Tools are not called in agent mode

[eflavio]

Issue

No tools are called when using aider-ce in agent mode. What am I doing wrong?

aider-ce --agent 
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Aider v0.88.28
Model: ollama_chat/qwen3-coder:30b with agent edit format
Git repo: .git with 24 files
Repo-map: using 4096 tokens, auto refresh
MCP servers configured: local_tools
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
agent> Implement the following the task with task_id 1 unit_test_tasks.json.                                                                                                                                                                                        
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
{"name": "View", "arguments": {"file_path": "unit_test_tasks.json"}}

Tokens: 2.2k sent, 19 received.
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
agent> 

Version and model info

Aider v0.88.28
Model: ollama_chat/qwen3-coder:30b with agent edit format
Git repo: .git with 24 files
Repo-map: using 4096 tokens, auto refresh
MCP servers configured: local_tools

[dwash96]

Hmm, I will definitely investigate this and let you know. Ideally, I'll have a fix tonight! I know small models are middling at reliably calling tools but I feel like qwen coder was specifically trained for this

[eflavio]

I started suspecting it might be the model. Is there a way to log the requests being sent and received? Or to play around with aider's system prompt?

[eflavio]

Any recommended models for local llm use? It can be larger than 30b. I'm on a strix halo system with 128gb of unified memory.

[dwash96]

To your first question, if you install in a virtual env, by modifying these lines in aider/llm.py

        self._lazy_module = importlib.import_module("litellm")
        self._lazy_module.suppress_debug_info = True
        self._lazy_module.set_verbose = False
        self._lazy_module.drop_params = True
        self._lazy_module._logging._disable_debugging()

You'll be able to see the LiteLLM model output in it's insanely verbose glory (warning: it is A LOT) which is what I did to debug:
#180

To your second, for local LLMS, especially with 128GB of memory, gpt-oss-120b is likely to be your best bet for tool calling since it's pre-trained on the openai tool syntax which LiteLLM uses. I think folks in the discord might have better suggestions on local models than I do since I currently am not using any local ones

[dwash96]

For local/weaker LLMs I am going to add an agent_config that will be to the effect of "retry_on_empty_response", which should really just replay the last request at the local model so it can try again with calling the next tool

My eyes are kicking into gear, it's not that the response is empty, it's that the model replied as json directly, I see!

[dwash96]

So I did push something that should work for this specific type of tool call failure. I will close this for now, but definitely let me know if you encounter another variation