# Agent calls weather API
weather_data = call_weather_api("San Francisco")

# Log the tool call
sb.sessions.add_turn(
    session_id=session.id,
    input={"type": "tool_call", "content": "get_weather(city='San Francisco')"},
    output={"type": "tool_result", "content": json.dumps(weather_data)},
    reasoning="Weather API call succeeded",
    metadata={
        "tool_name": "weather_api",
        "latency_ms": 450,
        "success": True
    }
)

try:
    result = risky_operation()
except Exception as e:
    # Log the error
    sb.sessions.add_turn(
        session_id=session.id,
        input={"type": "text", "content": user_message},
        output={"type": "error", "content": str(e)},
        reasoning=f"Operation failed: {type(e).__name__}",
        metadata={
            "error_type": type(e).__name__,
            "error_message": str(e),
            "stack_trace": traceback.format_exc()
        }
    )

# Before agent processes input
state_before = sb.sessions.get(session.id).state

# Agent processes and updates state
response = agent.process(user_message)
sb.sessions.update_state(session.id, new_state)

# Log turn with state snapshots
sb.sessions.add_turn(
    session_id=session.id,
    input={"type": "text", "content": user_message},
    output={"type": "text", "content": response},
    state_before=state_before,
    state_after=new_state,
    reasoning="Updated state based on user input"
)

# Get all turns for a session
turns = sb.sessions.list_turns(session_id=session.id, limit=100)

# Calculate metrics
total_turns = len(turns)
tool_calls = [t for t in turns if t.metadata.get("tool_used")]
errors = [t for t in turns if t.output.type == "error"]
avg_latency = sum(t.metadata.get("latency_ms", 0) for t in turns) / total_turns

print(f"Total turns: {total_turns}")
print(f"Tool calls: {len(tool_calls)}")
print(f"Errors: {len(errors)}")
print(f"Avg latency: {avg_latency:.0f}ms")