l3_live_context.py

"""L3 - Agent + live context.

Upgrades your agent with read-only SQL tools from an MCP server and opens a
conversation. Ask, run `python inject.py` in a second terminal, then ask again:
the answer changes.

    python l3_live_context.py
"""

from common import (
    agent_params,
    chat,
    ensure_agent,
    ensure_environment,
    load_config,
    load_layer,
    mcp_vault_ids,
    open_session,
    orca_client,
    run_main,
    state_for,
)

QUESTION = "Which accounts look like an account takeover right now?"


def main() -> None:
    config = load_config(["ORCA_BASE_URL", "ORCA_MODEL"])
    client = orca_client(config)
    state = state_for(config)
    environment_id = ensure_environment(client, state, f"hello-env-{config.participant}")

    # The same agent, next version: agent/<stack>/l3-live-context.json adds the MCP server.
    layer = load_layer("l3-live-context", config.stack)
    agent = ensure_agent(client, state, agent_params(layer, config))
    print(f"{agent.name} v{agent.version}: {layer['summary']}")

    # StreamNative Cloud's MCP server needs a credential. It goes in a vault, never
    # in the prompt. The MCP server on your laptop takes none, so there is no vault.
    vault_ids = mcp_vault_ids(client, state, config)

    session = open_session(client, state, environment_id, agent, "L3: live context", vault_ids=vault_ids)
    print("Tip: after the first answer, run `python inject.py` in another terminal and ask again.\n")
    chat(client, session.id, QUESTION, send_first=config.stack == "local")


if __name__ == "__main__":
    run_main(main)