Tool calls with empty/null arguments silently dropped (affects vLLM, Ollama, and other OpenAI-compatible providers)

[alpsencer]

pipecat version

Latest main branch (also verified in 0.0.76+)

Python version

3.12

Operating System

Linux (CI) / macOS

Issue description

When an LLM returns a streaming tool call with no arguments (arguments=null in the delta chunks), the tool call is silently dropped — no error, no warning. The tool handler is never invoked.

This happens in BaseOpenAILLMService._process_context (base_llm.py):

1. Line 544: if tool_call.function and tool_call.function.arguments: — when arguments is None, this check is falsy, so nothing is accumulated. The local arguments variable stays as "".

2. Line 561: if function_name and arguments: — the empty string "" is falsy, so the entire tool call execution block is skipped. run_function_calls is never called.

OpenAI masks this bug because it always sends arguments: "{}" (empty JSON object string) even for tools with zero parameters. But other OpenAI-compatible providers behave differently:

• vLLM (with Qwen, Llama, Mistral, etc.) omits arguments entirely when the tool schema has no required properties
• Ollama and other local providers may exhibit the same behavior

This is a significant issue because:

• Tools with no required parameters (e.g., end_call, hang_up, get_status) are common
• The failure is completely silent — no error logged, no exception raised
• It's hard to debug because the LLM does correctly identify and call the tool; pipecat just ignores it

Reproduction steps

1. Serve any model via vLLM with --enable-auto-tool-choice --tool-call-parser hermes
2. Register a tool with no required parameters:

   tools = [{"type": "function", "function": {"name": "end_call", "description": "End the call", "parameters": {"type": "object", "properties": {}, "required": []}}}]

3. Send a prompt that triggers the tool call (e.g., "End the call")
4. The streaming response includes tool_calls with function.name = "end_call" and finish_reason = "tool_calls", but no arguments chunks
5. The tool handler is never invoked

Expected behavior

Tool calls with no arguments should be treated as having arguments = {} (empty dict) and the registered handler should be invoked.

Actual behavior

The tool call is silently dropped. run_function_calls is never called. No error or warning is logged.

Suggested fix

Three lines in base_llm.py:

Line	Before	After
535	arguments_list.append(arguments)	arguments_list.append(arguments or "{}")
561	if function_name and arguments:	if function_name:
564	arguments_list.append(arguments)	arguments_list.append(arguments or "{}")

This defaults empty arguments to "{}" so json.loads produces an empty dict. No behavior change for OpenAI (which already sends "{}"). Fixes all OpenAI-compatible providers.