Tool use trouble

[scudderfish]

I'm connecting to a model running on ollama on another machine on my network. If I ask a question such as 'what files are in this directory?' I get the command for the tool (as a chunk of json) sent into the chat rather than the tool being executed. Is there a piece of config I've missed?

[Doorman11991]

What LLM model are you using?

[scudderfish]

Sorry for missing that vital piece of information. qwen2.5-coder:14b

[Doorman11991]

Fix is up in #41. The OpenAI-compat endpoint on Ollama doesn't always populate structured tool_calls for qwen2.5-coder — sometimes the model emits the invocation as a <tool_call>...</tool_call> JSON blob inside message.content. The PR adds a defensive extractor that detects four shapes (tool_call tags, fenced JSON, bare JSON, OpenAI-style function objects) and lifts them into the proper tool_calls field before the agent loop reads it. No config change needed once merged — works automatically for qwen2.5-coder:14b and similar local models.