Thanks — you want to store logs so that after each turn you can see the complete context (including the system prompt). Yes — you can do that. Here are the practical options and a minimal example for each so you can pick what fits your setup. Which part do you want to use: logging the context before the agent responds, after the agent responds, or saving snapshots to session storage for later inspection? Options • Log before the agent responds: use the on_user_turn_completed lifecycle hook to read

turn_ctx.messages

(includes system instructions) and log/store them. • Log after the agent responds: listen for conversation events (e.g.,

conversation_item_added

) and inspect

session.chat_ctx.messages

when an assistant message is added. • Persist snapshots: write a copy of the full

chat_ctx

messages into

session.userdata

(or your own storage) each turn so you can review complete context history later. • For tests: the

run

method returns a result with

session.chat_ctx.messages

you can examine. Minimal examples 1) Before agent replies (Python-style):

async def on_user_turn_completed(self, turn_ctx: ChatContext, new_message: ChatMessage):
    # Full context (system + history + new user message)
    snapshot = [{ "role": m.role, "content": m.content } for m in turn_ctx.messages]
    print("Context before reply:", snapshot)
    # Or store snapshot in persistent storage

2) After agent replies (session event):

@session.on("conversation_item_added")
def on_conversation_item(item):
    if item.type == "message" and item.role == "assistant":
        # Full context after assistant response
        print("Context after reply:", session.chat_ctx.messages)

3) Save snapshot to session.userdata:

self.session.userdata["last_full_context"] = [
  {"role": m.role, "content": m.content} for m in session.chat_ctx.messages
]

Notes •

ChatContext

/

session.chat_ctx.messages

contains system instructions, user messages, agent replies, and tool calls — so it gives the complete prompt context. • Use on_user_turn_completed to capture the prompt the LLM will see next; use conversation events to capture the state after the agent response. Tell me which option you prefer and whether you want a complete code snippet for Node.js or Python, and I’ll provide it. Sources: Agents (Node.js) | livekit/livekit_composite | DeepWiki | Testing and evaluation | LiveKit docs | Pipeline nodes and hooks | LiveKit docs