Error: Tool use failure

[phillc73]

Describe the bug
I am using Dirac with a custom Inference Endpoint and model from Hugging Face. When using Dirac the following error is always returned:

Error: Tool use failure. Can potentially be mitigated with some user guidance (e.g. "Try breaking down the task into smaller steps").

To Reproduce
Steps to reproduce the behavior:

1. dirac auth to add the API key, HG inference endpoint and custom endpoint URL as an OpenAPI compatible custom model.
2. dirac to start the interactive CLI
3. Type any command. In my case I loaded a skill and asked the model to wite Hello World! code and save to disk.
4. See error. Full error output looks like:

 /ocaml-dev Write a file named main.ml with the content: let hello_world () = print_string "Hello, World!";;. 

   {
     "name": "write_to_file",
     "arguments": {
       "path": "src/main.ml",
       "content": "let hello_world () = print_string \"Hello, World!\";;."
     }
   }

   {
     "name": "attempt_completion",
     "arguments": {
       "result": "The main.ml file now contains the following content:",
       "command": ""
     }
   }

   {
     "name": "attempt_completion",
     "arguments": {
       "result": "The main.ml file now contains the following content:",
       "command": ""
     }
   }

   {
     "name": "attempt_completion",
     "arguments": {
       "result": "The main.ml file now contains the following content:",
       "command": ""
     }
   }
 
   {
     "name": "attempt_completion",
     "arguments": {
       "result": "The main.ml file now contains the following content:",
       "command": ""
     }
   }

 Error: Tool use failure. Can potentially be mitigated with some user guidance (e.g. "Try breaking down the task into smaller steps").

It seems to me that connection to the model is working fine itself, because the Hello World! code was output. The issue is calling the tool to write the file.

The error is not input specific. I can write any instruction and the same error is returned. Using the Hugging Face chat agent associated with this endpoint in the "Playground" works as expected. All instructions return expected content.

I have used Dirac successfully in the last days, so current v0.3.44, configured to use a Mistral model. No problem with making tool calls in this case. As far as I know, nothing has changed in the environment between the provider/model configuration change.

Expected behavior
Exected behaviour is that Dirac would write the model output to file as requested. Thus, Dirac should have access to the tools to do this.

Surface and Version:

• Cli/IDE
• Version v0.3.44

[dirac-run]

I think the model is populating 'command' as an empty string triggers the command run. Looking at the code, it seems there is also a bug there that doesn't process the completion correctly when the command is specificed

dirac/src/core/task/tools/handlers/AttemptCompletionHandler.ts

Line 112 in c4f1df6

if (command) {

Is this happening on a smaller or newer model?

[phillc73]

It's a smaller 1.5B model, fine tuned in Jan 2026.

It's this one I have deployed to a Hugging Face Inference Endpoint: https://huggingface.co/mradermacher/Qwen2.5-OCamler-1.5B-Instruct-GGUF

8 bit quantisation.

Using llama.cpp as the engine.

[dirac-run]

That sounds remarkably good for a 1.5B param model. What was the tool call error rate like?

To your original issue, I will fix it with after the current refactor. Won't be very long now.

[phillc73]

To be fair, it's the first time I've used a Hugging Face Inference Endpoint and I'm quite prepared to accept I'm doing something incorrect.

I haven't really got very far with using this model. Dirac throws the issue outlined here.

Hermes Agent pretty much refused to work with it because of a small context window, plus other API errors.

Goose seems to connect, but also returns some kind of empty JSON.

{"name": "execute_typescript", "arguments": {
"code": "async function run() {\n console.log("Hello World!");\n return { message: "Hello World!" }; // Simulating a response\n}",
"tool_graph": []
}}

I've managed to connect a custom provider with Mistral's Vibe CLI and on some level things work, I just can't make the model execute anything, regardless of the mode chosen. Returns sycophantic text replies, executes nothing. (again could be somehing I screwed up with the deployment).

Having said that, on the Hugging Face Interface Endpoints Overview page for this deployed model there is a "WebUI" button. Click this, launch the UI and action is all good.

Edit: I think the issue with Vibe CLI is also tool calls. Perhaps the error really isn't Dirac specific, but some nuance of the Hugging Face Inference Endpoint API.

[dirac-run]

Interesting! So yeah Dirac only supports the models with native tool calling and in my experience, smaller models while have gotten better, still make a lot of tool call emit errors.

I am more curious on this topic from a different angle: whether the model size has any correlation to tool call success rates.

[phillc73]

The Mistral Vibe CLI issue is that it doesn't support tool calling for anything other than their own models, even though adding a custom provider is possible.

The behavior you're seeing is due to how Vibe CLI handles different providers:

Le Chat Pro API (Mistral-medium-3.5)
Vibe CLI integrates natively with Mistral's own APIs (like Le Chat Pro). Even though the model runs remotely, Mistral's infrastructure allows Vibe CLI to execute local tools (e.g., read_file, write_file, grep) on your machine after receiving the model's response. This is a privileged integration specific to Mistral's official APIs.

Alternative Provider (e.g., Hugging Face Inference Endpoint)
When you connect Vibe CLI to a third-party or custom endpoint (like your Hugging Face Inference Endpoint), the model runs remotely without any special integration. Vibe CLI treats it as a pure text generation service—it sends prompts and receives text responses, but cannot execute local tools because the remote model has no access to your local environment. The tools require local execution, which is only enabled for Mistral's own APIs.

Why This Happens

Mistral's APIs: Vibe CLI is designed to bridge local tool execution with Mistral's remote models. The tool calls are processed locally by Vibe CLI after the model generates its response.
Third-party Endpoints: No such bridge exists. The model only returns text, and Vibe CLI cannot inject local tool execution for non-Mistral providers.

Maybe this one really is small enough to run locally. I'll try that, save some money. Still, it'd be good to see the issue somehow resolved, as I would quite like to fine tune a larger, newer model for the same use case, which I would struggle to run locally.

Quite liking Dirac, initially because of the hash anchors and batch dispatch, but overall it's a nice tool to use. Thanks for your work.