Tools enabled but no tool execution with OpenAI-compatible endpoint (request lacks tools, response contains only reasoning_content and empty content) — gpt-oss + llama.cpp

[dragonkz01]

When Tools are enabled (e.g., search, website) and Newelle is configured to use an OpenAI-compatible endpoint (llama.cpp server), the assistant often produces no visible response and no tool is executed. Captured HTTP traffic shows Newelle sends /v1/chat/completions WITHOUT the OpenAI tools: [...] parameter, while the system prompt embeds “Tools Usage Rules”. With gpt-oss on llama.cpp, the model replies with content: null/"" and only reasoning_content (e.g., “Use search tool.”), so Newelle never sees a parsable tool-call JSON nor a tool_calls structure to run. (Logs attached)

Suggested fixes

Implement true OpenAI tool calling for OpenAI providers/endpoints:

add tools: [...] to request payload when Tools are enabled

support tool_calls responses, run tools, and send results back via role: "tool" messages

Robust fallback for prompt-based tools:

if content is empty but reasoning_content indicates a tool directive, show an actionable error or retry with a different mode (or parse tool JSON if present)

Reasoning mode compatibility:

allow disabling/overriding reasoning_effort for models/endpoints where it breaks tool output, or detect the empty-content case and retry with lower reasoning

Notes/evidence

Manual curl tests against llama.cpp show that the model CAN return proper OpenAI tool_calls when tools + tool_choice are provided, but Newelle does not appear to send tools in its API requests.

[FrancescoCaracciolo]

Are you on 1.2.5? Generally if the model replies with a tool call (formetted in openai) it is still read from that version. (Tools schema is given in the prompt, if you enable it)
At the moment Newelle does not use fully native openai tool calling in order to have retrocompatibility with most models that do not support that format.

[Thomas-git]

I’m using local ollama endpoint. Search works, linux mcp server installed. Model is seeing all available tools.
When asked to get available system memory, it responds with
{"tool": "tool.get_memory_information", "arguments": {"host": null}}
But linux mcp server shows no request and Ai chat stop there.

Same problem with extensions providing tools: for example, basic calculator. LLM try to run tool, but nothing append.

I use gptoss20b as my LLM.

[FrancescoCaracciolo]

I’m using local ollama endpoint. Search works, linux mcp server installed. Model is seeing all available tools. When asked to get available system memory, it responds with {"tool": "tool.get_memory_information", "arguments": {"host": null}} But linux mcp server shows no request and Ai chat stop there.

Same problem with extensions providing tools: for example, basic calculator. LLM try to run tool, but nothing append.

I use gptoss20b as my LLM.

It sounds like the LLM is using the wrong format. Also, some extensions tools are not updated to the new tools format introduced to 1.2

[Thomas-git]

I suppose the LLM is using what the prompt tell it.
Here it the tool prompt:

Tools

Tools Usage Rules

You have access to the following tools.

When using a tool, you must:

1. Output only a single valid JSON object.

2. Use exactly this format:

   {
     "tool": "tool_name",
     "arguments": {
       "arg_name": "arg_value"
     }
   }

3. Ensure the JSON is valid (no comments, trailing commas, or extra text).

4. Use only the tools listed below and only their defined arguments.

5. Do not include any explanations, markdown, or additional text before or after the JSON.

After invoking a tool, you must immediately stop the message.

Available tools:

{TOOLS}

It seems to me the LLM it following rules.

[FrancescoCaracciolo]

tool.get_memory_information sounds like a wrong tool name, also provide console output because sometimes it might be recognized as tool but if it fails executing it gives you the codeblock

[Thomas-git]

I manage to have gptoss use tools by modifying system prompt

   {
     "tool": "name",
     "arguments": {
       "arg_name": "arg_value"
     }
   }

note "name" instead of "tool_name"