bug: STT errors close AgentSession immediately, lacking the tolerance counter that LLM and TTS errors have #5665

Bug Description

AgentSession._on_error applies an max_unrecoverable_errors tolerance counter for llm_error and tts_error, but stt_error has no such counter — it falls through and immediately closes the session on the first non-recoverable STT error.

Root Cause

In livekit-agents/livekit/agents/voice/agent_session.py, the _on_error method:

def _on_error(self, error: llm.LLMError | stt.STTError | tts.TTSError | ...) -> None:
    if self._closing_task or error.recoverable:
        return

    if error.type == "llm_error":
        self._llm_error_counts += 1
        if self._llm_error_counts <= self.conn_options.max_unrecoverable_errors:
            return  # ← tolerance applied
    elif error.type == "tts_error":
        self._tts_error_counts += 1
        if self._tts_error_counts <= self.conn_options.max_unrecoverable_errors:
            return  # ← tolerance applied

    # stt_error: no branch → falls through → session closes immediately ❌

    self._closing_task = asyncio.create_task(
        self._aclose_impl(error=error, reason=CloseReason.ERROR)
    )

_llm_error_counts and _tts_error_counts are initialised and reset in multiple places (__init__, _aclose_impl, _update_agent_state), but there is no equivalent _stt_error_counts.

Real-World Impact

When an STT provider returns transient errors (e.g. HTTP 403/429 from Groq after exhausting per-utterance retries in the RecognizeStream._main_task loop), the session closes immediately on the very first non-recoverable STT error. There is no opportunity for application-level error handlers (registered via session.on("error", ...)) to perform graceful recovery such as transferring to a human agent.

The per-utterance retry loop in RecognizeStream._main_task already handles transient network blips. The session-level tolerance is meant to handle the case where retries are truly exhausted — but currently STT gets zero tolerance at the session level while LLM and TTS each get max_unrecoverable_errors (default 3).

Expected Behaviour

stt_error should receive the same tolerance as llm_error and tts_error: allow up to max_unrecoverable_errors non-recoverable STT errors before closing the session, giving application error handlers a chance to act (e.g. fall back to another STT, transfer to human, speak an acknowledgement).

Proposed Fix

Add _stt_error_counts counter mirroring the existing LLM/TTS pattern:

# __init__ and reset sites:
self._stt_error_counts = 0

# _on_error:
elif error.type == "stt_error":
    self._stt_error_counts += 1
    if self._stt_error_counts <= self.conn_options.max_unrecoverable_errors:
        return

And reset _stt_error_counts alongside the other counters in _aclose_impl and _update_agent_state (when state transitions to "speaking").

Environment

• livekit-agents (latest main)
• STT plugin: livekit-plugins-groq (wraps livekit-plugins-openai STT)
• Reproduced with any STT provider that returns HTTP 4xx errors after retries are exhausted