Blaze Livekit plugin integration

[HoangPN711]

No description provided.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

Devin Review found 2 new potential issues.

View 16 additional findings in Devin Review.

[devin-ai-integration]

🔴 Tool serialization references non-existent callable_function attribute on Tool

In _build_tools_param, the code checks hasattr(tool, "callable_function") at llm.py:244, but no Tool subclass in the framework (FunctionTool, RawFunctionTool, ProviderTool) has a callable_function attribute. FunctionTool uses _info (a FunctionToolInfo with .name and .description) and _func. As a result, the if branch at line 244 is dead code — every tool falls through to the else at line 258, which only serializes {"type": "function", "function": {"name": tool.id}}, omitting the function's description and parameters. This makes function calling effectively non-functional since the LLM backend receives no schema information about what the tools do or what arguments they accept.

Prompt for agents

The _build_tools_param method in llm.py references a non-existent `callable_function` attribute on Tool objects. In the livekit agents framework, tool information is accessed differently depending on the tool type:

- FunctionTool (from livekit.agents.llm.tool_context) has:
  - `.id` -> returns `self._info.name`
  - `._info` -> a `FunctionToolInfo` with `.name` and `.description`
  - The function's parameters can be obtained by inspecting the function signature or using the framework's built-in serialization utilities.

- RawFunctionTool has:
  - `.id` -> returns `self._info.name`  
  - `._info` -> a `RawFunctionToolInfo` with `.name` and `.raw_schema` (dict containing name, description, parameters)

The code should check `isinstance(tool, FunctionTool)` or `isinstance(tool, RawFunctionTool)` and access the appropriate info attributes. Look at how other LLM plugins (e.g., livekit-plugins-anthropic, livekit-plugins-google) serialize tools using the `_provider_format` utilities in `livekit.agents.llm._provider_format` for the canonical approach.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🔴 Inconsistent JSON request body format between tool and non-tool LLM requests

At llm.py:323, the request body format changes depending on whether tools are present: json=messages if not tools_param else body. Without tools, the HTTP body is a bare JSON array ([{"role":"user","content":"..."}]). With tools, it's a JSON object ({"messages":[...],"tools":[...]}). These are structurally different — a REST API endpoint will expect one format or the other. If the API expects a bare array, the tools-enabled path will fail; if the API expects an object with a "messages" key, the non-tools path will fail.

Suggested change

	json=messages if not tools_param else body,
	json=body,

---

Was this helpful? React with 👍 or 👎 to provide feedback.