The agent that grows with you
Python aiai-agentai-agentsanthropicchatgptclaudeclaude-codecodexhermeshermes-agentllmnous-researchopenai
"""A down managed local model server is reported as itself, never as a credential problem.
With the main model on the managed llama.cpp runtime and no server serving, the readiness probe
and a turn's resolution failure must say "local server down" in a structured way, so clients keep
the app open with that error instead of opening provider sign-in or offering "Update API key".
"""
from __future__ import annotations
import json
import pytest
from hermes_constants import get_hermes_home
@pytest.fixture
def llamacpp_home(monkeypatch):
from hermes_cli.local_runtime import detect
def write(local_runtime: dict) -> None:
home = get_hermes_home()
home.mkdir(parents=True, exist_ok=True)
config = {"model": {"provider": "llamacpp", "default": "local-qwen"}, "local_runtime": local_runtime}
(home / "config.yaml").write_text(json.dumps(config), encoding="utf-8")
# No external llama-server on this machine may answer for the managed one.
monkeypatch.setattr(detect, "detect_server", lambda extra_ports=(): None)
return write
def _resolution_error():
from hermes_cli.runtime_provider import resolve_runtime_provider
with pytest.raises(ValueError) as caught:
resolve_runtime_provider(requested="llamacpp", target_model="local-qwen")
return caught.value
@pytest.mark.parametrize("enabled", [False, True])
def test_runtime_check_names_a_down_local_server(llamacpp_home, enabled):
from hermes_cli.runtime_readiness import readiness_failure
llamacpp_home({"enabled": enabled})
answer = readiness_failure(_resolution_error())
assert answer["ok"] is False
assert answer["error_code"] == "local_runtime_unavailable"
assert "local model server" in answer["error"]
def test_runtime_check_leaves_other_failures_uncoded():
from hermes_cli.runtime_readiness import readiness_failure
assert readiness_failure(ValueError("No usable credentials found for openrouter.")) == {
"ok": False, "error": "No usable credentials found for openrouter."}
@pytest.mark.parametrize("enabled", [False, True])
def test_turn_failure_is_an_endpoint_failure_not_a_rejected_key(llamacpp_home, enabled):
from agent.error_surface import build_error_surface_from_result
from gateway.config import Platform
from gateway.run_turn_runner import _resolution_failure_reply
llamacpp_home({"enabled": enabled})
exc = _resolution_error()
text, reason, retryable = _resolution_failure_reply(exc, Platform.TELEGRAM)
# Switched on but not serving yet is worth a retry; switched off is not.
assert (reason, retryable) == ("local_runtime_unavailable", enabled)
assert str(exc) in text
surface = build_error_surface_from_result(
{"failed": True, "error": str(exc), "failure_reason": reason, "failure_retryable": retryable},
provider="llamacpp")
assert surface["layer"] == "endpoint"
assert surface["code"] == "local_runtime_unavailable"
assert "auth_kind" not in surface
@pytest.mark.parametrize("enabled", [False, True])
def test_chat_text_names_the_local_server_and_no_desktop_only_pane(llamacpp_home, enabled):
"""The messaging reply and the TUI/dashboard turn text both name the local model server, and
neither points at Settings → Providers → Local models, a pane only the Desktop app has."""
from agent.error_surface import build_error_surface_from_result
from gateway.config import Platform
from gateway.run_turn_runner import _resolution_failure_reply
from tui_gateway.user_messages import turn_error_text
llamacpp_home({"enabled": enabled})
exc = _resolution_error()
reply, reason, retryable = _resolution_failure_reply(exc, Platform.TELEGRAM)
surface = build_error_surface_from_result(
{"failed": True, "error": str(exc), "failure_reason": reason, "failure_retryable": retryable},
provider="llamacpp")
tui = turn_error_text(str(exc), surface, recoverable=retryable)
title = tui.splitlines()[0]
assert "local model server" in title
assert "custom model endpoint" not in tui
for text in (reply, tui):
assert "local model server" in text
assert "Settings" not in text
assert "/model" in text
@pytest.mark.parametrize("enabled", [False, True])
def test_chat_text_suggests_retry_only_while_the_server_can_come_up(llamacpp_home, enabled):
"""A switched-off server never comes up on its own: the TUI/dashboard text (built from the
error surface, as a turn builds it) names where to start it, never /retry."""
from agent.error_surface import build_error_surface_from_result
from tui_gateway.user_messages import turn_error_text
llamacpp_home({"enabled": enabled})
exc = _resolution_error()
surface = build_error_surface_from_result(
{"failed": True, "error": str(exc), "failure_reason": exc.error_code, "failure_retryable": exc.retryable},
provider="llamacpp")
hint = turn_error_text(str(exc), surface).splitlines()[-1]
assert ("/retry" in hint) is enabled
if not enabled:
assert "hermes model" in hint and "/model" in hint
# Surface-neutral: the backend sentence names the terminal path too, not only the Desktop app.
assert "hermes model" in str(exc)