tool_choice parameter persists across turns when passed to generate_reply()

[enrique-quadrivia]

Description

When calling session.generate_reply(tool_choice="none") during agent initialization (e.g., in on_enter), the tool_choice setting appears to persist for all subsequent conversation turns, preventing tools from ever being called even when the user responds.

Expected Behavior

The tool_choice parameter passed to generate_reply() should only affect that specific generation, not subsequent turns. Tools should be callable on subsequent turns when the user responds.

Actual Behavior

After calling generate_reply(tool_choice="none") once, extraction/function tools are never called on subsequent user turns, even though tool_choice is not passed on those turns.

Reproduction Steps

1. Create an agent with function tools (e.g., extraction tools)
2. In the agent's on_enter or initial greeting, call:

   self.session.generate_reply(tool_choice="none", allow_interruptions=True)

3. User responds with information that should trigger tool extraction
4. Observe that tools are never called on subsequent turns

Code Analysis

Looking at agent_activity.py, the logic at lines ~900-903 in _generate_reply:

model_settings=ModelSettings(
    tool_choice=tool_choice
    if utils.is_given(tool_choice) or self._tool_choice is None
    else self._tool_choice
),

This uses the persisted self._tool_choice when:

• tool_choice is NOT explicitly given, AND
• self._tool_choice has been previously set

The self._tool_choice is set via update_options() at line 352, but it's unclear how generate_reply() triggers this persistence path.

Environment

• livekit-agents version: 1.3.10
• Python version: 3.13
• LLM: OpenAI (non-realtime)

Workaround

Don't pass tool_choice="none" to generate_reply() even when you don't want tools called on the greeting. The worst case (LLM calls extraction tool on greeting before user responds) results in empty extraction with no side effects.

# Instead of:
if self.tools:
    self.session.generate_reply(tool_choice="none", allow_interruptions=True)
else:
    self.session.generate_reply(allow_interruptions=True)

# Use:
self.session.generate_reply(allow_interruptions=True)

Related Issues

• Enhanced control over instructions and chat context in LLM interaction #3820 - Reports similar issues with tool_choice='none' being ignored
• llm.FallbackAdapter doesn't call tools #1207 - FallbackAdapter doesn't call tools

[longcw]

are you using realtime model? it would be helpful if you can share a reproducible example.

[enrique-quadrivia]

are you using realtime model? it would be helpful if you can share a reproducible example.

I'm not using realtime model, currently using gpt-4o-mini to carry out my experiments. Will try to compile a minimal reproducible example soon.

Have you seen this before?

[longcw]

no, I cannot reproduce this issue on my end with the basic_agent