voice-agent-expert

name	voice-agent-expert
description	This skill is a practical, 'use-it-while-debugging' reference for getting a LiveKit + Letta voice agent working reliably.

Voice Agent Expert – Debugging Cheat Sheet (Skill)

This skill is a practical, "use-it-while-debugging" reference for getting a LiveKit + Letta voice agent working reliably.

---

Core rules (fix 90% of issues)

1. Override the right LiveKit hook

   • The framework calls llm_node, not generate_reply.
   • If you override the wrong method, your changes won't run and your agent may "look alive" but never route text correctly.

2. Only ONE voice-agent process can run at a time

   • Duplicate processes commonly cause:
     • timeouts
     • audio cutting out
     • inconsistent behavior (you're looking at logs from one process while another is actually serving)

3. Use dev mode for local testing

   • start can sit waiting for dispatch and look "broken."
   • If you're actively testing locally, dev mode is the fast path.

---

Nuclear reset (when it's just stuck)

Run this when things are wedged, timeouts are happening, or audio is cutting out. It kills old processes, restarts LiveKit in dev mode, then restarts the Letta voice agent.

pkill -9 -f letta_voice_agent.py
pkill -f livekit-server
sleep 2

cd /home/adamsl/ottomator-agents/livekit-agent
nohup ./livekit-server --dev --bind 0.0.0.0 > /tmp/livekit.log 2>&1 &

cd /home/adamsl/planner/a2a_communicating_agents/hybrid_letta_agents
/home/adamsl/planner/.venv/bin/python3 letta_voice_agent.py dev > /tmp/letta_voice_agent.log 2>&1 &

---

Is it alive? checks (fastest signals)

1. LiveKit responding (not just "listening")

   curl -v http://localhost:7880/ 2>&1 | head -20
   

   If this hangs or times out, LiveKit is stuck → restart it.

2. Confirm you don't have duplicate voice-agent processes

   ps aux | grep letta_voice_agent | grep -v grep
   

   You should see exactly one.

3. Confirm Letta routing is actually being used Watch your logs for your 🎤 marker (or equivalent logging) that proves llm_node is being hit.

---

Common gotchas

WebSocket double-slash problem

If the browser tries:

ws://127.0.0.1:7880//rtc

Use instead:

ws://localhost:7880

Reason: The client appends /rtc, and some configs end up producing a double-slash.

Code changes not taking effect

• Edits don't apply until you kill + restart the running process.
• If you edited code and nothing changed, assume you're running an old process.

---

If messages aren't showing up in Letta Desktop

Most common causes:

• You overrode the wrong method (generate_reply instead of llm_node)
• An old process is still running (your edits aren't live)
• The session LLM is bypassing Letta routing

Recommended fix flow

1. Search your code for llm_node and confirm your override is in place.
2. Kill all running agent processes.
3. Restart cleanly using the Nuclear reset section above.
4. Tail logs and confirm you see your 🎤 marker.