The agent that grows with you
Python aiai-agentai-agentsanthropicchatgptclaudeclaude-codecodexhermeshermes-agentllmnous-researchopenai
"""A standalone gateway (``multiplex_profiles`` off) keeps resolving the launch profile's credentials
after a native hosted room activated the process-wide secret guard (#112878).
``tui_gateway.launch_profile_policy.activate_multi_profile_hosting`` runs inside the messaging
gateway process when a hosted room serves a second profile; ``get_secret`` then fails closed for
every unscoped read. The standalone gateway's turns and handler entry points must bind the launch
profile's OWN scope (``.env`` over the env frozen at activation) — not skip binding because the
config flag is off, and not rebuild from ``.env`` alone (systemd / ``op run`` injection has no file).
"""
import asyncio
from contextlib import nullcontext
from unittest import mock
import pytest
from agent import secret_scope
from agent.secret_scope import UnscopedSecretError, current_secret_scope, get_secret
from gateway.config import GatewayConfig
from gateway.run import GatewayRunner
from tui_gateway import launch_profile_policy
INJECTED = "HOSTEDROOM_TEST_INJECTED_KEY"
@pytest.fixture
def standalone(tmp_path, monkeypatch):
launch = tmp_path / "launch"
launch.mkdir()
(launch / ".env").write_text("OPENAI_API_KEY=launch-dotenv-key\n", encoding="utf-8")
secondary = tmp_path / "profiles" / "roomie"
secondary.mkdir(parents=True)
(secondary / ".env").write_text("OPENAI_API_KEY=secondary-key\n", encoding="utf-8")
monkeypatch.setenv("HERMES_HOME", str(launch))
monkeypatch.setenv(INJECTED, "launch-env-injected") # systemd / `op run` style injection
monkeypatch.setattr(secret_scope, "_MULTIPLEX_ACTIVE", False)
monkeypatch.setattr(launch_profile_policy, "_snapshot", None)
runner = GatewayRunner.__new__(GatewayRunner)
runner.config = GatewayConfig(multiplex_profiles=False)
return runner, secondary
def _keys():
return get_secret("OPENAI_API_KEY"), get_secret(INJECTED)
def test_standalone_turn_binds_launch_profile_scope_after_hosted_activation(standalone):
runner, secondary = standalone
from tui_gateway.server import _session_profile_runtime_scope
# A native hosted room ran a second profile: real activation, real secondary scope entered and left.
launch_profile_policy.activate_multi_profile_hosting()
with _session_profile_runtime_scope({"profile_home": str(secondary)}):
assert get_secret("OPENAI_API_KEY") == "secondary-key"
assert get_secret(INJECTED) is None # the launch env never leaks into a secondary
assert secret_scope.is_multiplex_active()
source = mock.MagicMock(profile=None)
with runner._profile_scope_for_source(source):
assert _keys() == ("launch-dotenv-key", "launch-env-injected")
runner._handle_message = mock.AsyncMock(side_effect=lambda event: _keys())
assert asyncio.run(runner._primary_message_handler()(mock.MagicMock(source=source))) == (
"launch-dotenv-key", "launch-env-injected")
with pytest.raises(UnscopedSecretError): # the guard itself is not weakened
get_secret("OPENAI_API_KEY")
def test_standalone_without_hosted_activation_stays_unscoped(standalone):
runner, _secondary = standalone
source = mock.MagicMock(profile=None)
assert isinstance(runner._profile_scope_for_source(source), nullcontext)
runner._handle_message = mock.AsyncMock(side_effect=lambda event: current_secret_scope())
assert asyncio.run(runner._primary_message_handler()(mock.MagicMock(source=source))) is None