Voice agent: no final response when max_tool_steps is reached

[IanSteno]

Bug summary

When max_tool_steps is reached during the voice agent tool-calling loop, the pipeline can return early without generating a final LLM response.

This can leave users with silence/no reply after one or more tool calls.

Affected area

• livekit-agents/livekit/agents/voice/agent_activity.py

Repro/observations

• In the tool loop, once step count exceeds configured max_tool_steps, current behavior logs a warning and returns
• Tool results are already available in context, but no final text-only response is forced
• End-user experience is an abrupt stop instead of a summarized answer

Why this is a bug

Hitting a tool-step limit should end tool execution, not user communication. The agent should produce a final response with gathered results rather than silently stopping.

Proposed fix

Tracked in PR:

• fix: generate final LLM response when max_tool_steps is reached #4747

That PR removes the silent early return behavior and forces a final non-tool LLM response when the limit is reached.

[weiguangli-io]

I have traced the control flow in agent_activity.py. The problem is at line ~2204:

if speech_handle.num_steps >= self._session.options.max_tool_steps + 1:
    logger.warning("maximum number of function calls steps reached")
    return  # <-- silent exit, no response to user

This return exits _pipeline_reply_task before the recursive call that would generate a final text response. The tool results from the last step are available in chat_ctx but never used.

The existing code already handles a similar case: when draining is true (scheduling paused), it sets tool_choice="none" to force a text-only response. The fix should follow the same pattern.

Proposed fix: Replace the early return with a flag, and include it in the tool_choice condition:

max_steps_reached = speech_handle.num_steps >= self._session.options.max_tool_steps + 1
if max_steps_reached:
    logger.warning("max tool steps reached, generating final response")

speech_handle._num_steps += 1
# ... existing tool output processing ...

tool_choice="none" if (draining or model_settings.tool_choice == "none" or max_steps_reached) else "auto"

This reuses the existing _pipeline_reply_task recursive path with tool_choice="none", so the LLM produces a text response using whatever tool results have been gathered. No new code paths, fully backward compatible.

I see PR #4747 takes this exact approach. The fix looks correct to me. Would the maintainers like any adjustments to that PR, or should I submit a separate one?

[davidzhao]

fixed in #4747