Voice agent pipelines deadlock or lose tool call results when users interrupt during function execution
Real-time voice agents (built on Pipecat, LiveKit, OpenAI Realtime API) fail silently or hang permanently when a user interrupts speech while the agent is mid-tool-call. Pipecat's GitHub documented at least 4 distinct bug classes in this exact scenario (queue recreation, deadlock, frame-drop, race condition) across 8 public issues closed March-May 2026. The deadlock case requires three concurrent conditions: pause_frame_processing=True, interruption during TTFB (200-700ms window), and a queued function call result -- leaving the bot permanently unresponsive. The function_call_timeout_secs default changed from 10.0 to None in v1.0.0 with no deprecation warning, causing production code to hang indefinitely on failure. The HN Show thread on sub-500ms voice agent (570 points, 153 comments, March 2026) confirms developers encounter additional production complexity: tool calling under load, webhook execution race conditions, and multi-tenant routing issues that are entirely separate from the latency problem. The OpenAI gpt-realtime community forum thread for the May 8 2026 launch also shows SIP/telephony integration underdocumented with zero docs for the /accept/ endpoint parameters and developers hitting echo cancellation self-interruption bugs at any VAD threshold.
An SDK that preserves voice-agent session state across mid-call interrupts and cross-agent handoffs
5.8k ▲Score Breakdown
Social Proof 1 sources
Gap Assessment
Pipecat and LiveKit are frameworks, not reliability layers. Vapi and Retell handle telephony routing but not mid-call tool execution correctness under interruption. No product does: interrupt-aware tool execution queuing with guaranteed result delivery + production test harness for voice agent tool calls. Gap is specific to the interrupt-during-function-call scenario at scale -- a coordination layer between the streaming pipeline and tool execution that survives barge-in.