Skip to content

Personal AI, On Personal Devices

Python

fix: keep local GPT-OSS models out of cloud routing (#1203)

Elliot Slusky86b83938e723
diff --git a/src/openjarvis/server/cloud_router.py b/src/openjarvis/server/cloud_router.py
index 53d4a9e7..0e8565dc 100644
--- a/src/openjarvis/server/cloud_router.py
+++ b/src/openjarvis/server/cloud_router.py
@@ -69,6 +69,10 @@ def _load_keys() -> dict[str, str]:
def get_provider(model: str) -> str | None:
"""Return the provider for a model name, or None if it's a local model."""
+ # Bare GPT-OSS names and Ollama tags are local, despite the "gpt-" prefix.
+ # Provider-qualified IDs (e.g. "openai/gpt-oss-20b") still route to cloud.
+ if model == "gpt-oss" or model.startswith(("gpt-oss:", "gpt-oss-")):
+ return None
if any(model.startswith(p) for p in _OPENAI_PREFIXES):
return "openai"
if any(model.startswith(p) for p in _ANTHROPIC_PREFIXES):
diff --git a/tests/server/test_cloud_router.py b/tests/server/test_cloud_router.py
index b9f26851..a055db41 100644
--- a/tests/server/test_cloud_router.py
+++ b/tests/server/test_cloud_router.py
@@ -1,4 +1,4 @@
-"""Regression tests for OpenRouter model ID normalization."""
+"""Regression tests for cloud provider detection and model ID normalization."""
from __future__ import annotations
@@ -8,6 +8,31 @@ from openjarvis.core.types import Message
from openjarvis.server import cloud_router
+ "model",
+ ["gpt-oss", "gpt-oss:20b", "gpt-oss:120b", "gpt-oss-20b", "gpt-oss-120b"],
+)
+def test_gpt_oss_is_local(model):
+ assert cloud_router.get_provider(model) is None
+ assert not cloud_router.is_cloud_model(model)
+
+
+ "model,provider",
+ [
+ ("gpt-4o", "openai"),
+ ("claude-sonnet-4-20250514", "anthropic"),
+ ("gemini-2.5-pro", "google"),
+ ("openai/gpt-oss-20b", "openrouter"),
+ ("openrouter/openai/gpt-oss-120b", "openrouter"),
+ ("atlascloud/openai/gpt-oss-120b", "atlascloud"),
+ ],
+)
+def test_cloud_models_keep_their_provider(model, provider):
+ assert cloud_router.get_provider(model) == provider
+ assert cloud_router.is_cloud_model(model)
+
+
def test_get_provider_detects_bare_openrouter_id():
assert cloud_router.get_provider("anthropic/claude-haiku-4.5") == "openrouter"
diff --git a/tests/server/test_routes.py b/tests/server/test_routes.py
index fa3219f2..410dc1f5 100644
--- a/tests/server/test_routes.py
+++ b/tests/server/test_routes.py
@@ -1526,6 +1526,55 @@ class TestIdentityPromptInjection:
class TestModelsEndpoint:
+ def test_local_gpt_oss_is_listed_alongside_other_local_models(self):
+ """Do not filter GPT-OSS out of the installed-model picker (#1187)."""
+ models = ["gpt-oss:20b", "qwen3:8b"]
+ engine = _make_engine(models=models)
+ engine.engine_id = "ollama"
+ app = create_app(engine, models[0], config=_test_config())
+
+ with patch(
+ "openjarvis.server.cloud_router.list_local_models",
+ new_callable=AsyncMock,
+ return_value=[],
+ ) as list_local_models:
+ resp = TestClient(app).get("/v1/models")
+
+ assert resp.status_code == 200
+ assert [item["id"] for item in resp.json()["data"]] == models
+ list_local_models.assert_not_called()
+
+ @pytest.mark.parametrize("multi", [False, True])
+ def test_local_gpt_oss_streams_through_ollama(self, multi):
+ """A local GPT-OSS request must never need a cloud API key (#1187)."""
+ from openjarvis.engine.multi import MultiEngine
+
+ model = "gpt-oss:20b"
+ ollama = _make_engine(models=[model])
+ ollama.engine_id = "ollama"
+ ollama.is_cloud = False
+ engine = MultiEngine([("ollama", ollama)]) if multi else ollama
+ app = create_app(engine, model, config=_test_config())
+
+ with patch(
+ "openjarvis.server.cloud_router.stream_cloud",
+ side_effect=AssertionError("Local model routed to cloud"),
+ ) as stream_cloud:
+ resp = TestClient(app).post(
+ "/v1/chat/completions",
+ json={
+ "model": model,
+ "messages": [{"role": "user", "content": "hello"}],
+ "stream": True,
+ },
+ )
+
+ assert resp.status_code == 200
+ stream_cloud.assert_not_called()
+ assert "Hello" in resp.text
+ assert '"engine": "ollama"' in resp.text
+ assert "Error during generation" not in resp.text
+
def test_list_models(self, client):
resp = client.get("/v1/models")
assert resp.status_code == 200