Skip to content
Python

Route native OpenAI (blank endpoint) to the Responses provider

Devika Verma9d3f6d389d08
diff --git a/coworker/providers/__init__.py b/coworker/providers/__init__.py
index 6c35ef95..5245a15a 100644
--- a/coworker/providers/__init__.py
+++ b/coworker/providers/__init__.py
@@ -10,6 +10,7 @@ from .base import (
from .capabilities import capabilities_for
from .gemini_provider import GeminiProvider
from .openai_provider import OpenAIProvider, resolve_api_key
+from .openai_responses import OpenAIResponsesProvider
from .registry import (
ProviderDescriptor,
ProviderField,
@@ -34,6 +35,7 @@ __all__ = [
"BedrockProvider",
"GeminiProvider",
"OpenAIProvider",
+ "OpenAIResponsesProvider",
"VertexProvider",
"resolve_api_key",
"capabilities_for",
diff --git a/coworker/providers/registry.py b/coworker/providers/registry.py
index a07ec316..6c147705 100644
--- a/coworker/providers/registry.py
+++ b/coworker/providers/registry.py
@@ -6,8 +6,9 @@ GUI, same `to_dict()` shape connectors use) and a `build(profile, secrets)` fact
a `ProviderClient`. The `ProviderRouter` selects a descriptor by the `provider:` prefix of a
model string and builds (and caches) its client from the matching SecretStore profile.
-Today: `openai` (the default, with an optional custom endpoint that covers Azure OpenAI's
-`/openai/v1` and any OpenAI-compliant gateway), `anthropic` (native Messages API via
+Today: `openai` (the default — native models via the Responses API; an optional custom
+endpoint covering Azure OpenAI's `/openai/v1` and any OpenAI-compliant gateway keeps the
+Chat Completions path), `anthropic` (native Messages API via
`AnthropicProvider`), `gemini` (native Google GenAI API via `GeminiProvider`), `bedrock`
(models in the user's own AWS account — Claude natively, everything else via Converse),
`vertex` (the user's own GCP project — Gemini and Claude natively, open-weight via the
@@ -25,6 +26,7 @@ from .base import ProviderClient
from .bedrock_provider import BedrockProvider
from .gemini_provider import GeminiProvider
from .openai_provider import OpenAIProvider
+from .openai_responses import OpenAIResponsesProvider
from .vertex_provider import VertexProvider
DEFAULT_OLLAMA_URL = "http://localhost:11434"
@@ -111,11 +113,15 @@ def _normalize_ollama_url(url: Optional[str]) -> str:
def _build_openai(profile: dict[str, Any], secrets: Any) -> ProviderClient:
- # Key resolution stays in OpenAIProvider/resolve_api_key (explicit → env → SecretStore),
- # so we just hand it the SecretStore. An optional custom endpoint (Azure OpenAI /openai/v1,
- # OpenRouter, vLLM, …) comes from the stored profile.
+ # Key resolution stays in resolve_api_key (explicit → env → SecretStore), so we just
+ # hand over the SecretStore. Stock OpenAI (no custom endpoint) speaks the Responses
+ # API — the only wire with reasoning + tools on GPT-5.6+. A custom endpoint (Azure
+ # OpenAI /openai/v1, vLLM, any OpenAI-compliant gateway) keeps Chat Completions,
+ # which is what compat servers implement.
base_url = ((profile or {}).get("base_url") or "").strip() or None
- return OpenAIProvider(secrets=secrets, base_url=base_url)
+ if base_url:
+ return OpenAIProvider(secrets=secrets, base_url=base_url)
+ return OpenAIResponsesProvider(secrets=secrets)
def _build_anthropic(profile: dict[str, Any], secrets: Any) -> ProviderClient:
diff --git a/tests/test_openai_responses.py b/tests/test_openai_responses.py
index b7fe3b7a..9e871d57 100644
--- a/tests/test_openai_responses.py
+++ b/tests/test_openai_responses.py
@@ -556,3 +556,29 @@ def test_ensure_client_without_key_raises(monkeypatch):
monkeypatch.delenv("OPENAI_API_KEY", raising=False)
with pytest.raises(RuntimeError, match="No model API key"):
OpenAIResponsesProvider()._ensure_client()
+
+
+# -- registry routing ----------------------------------------------------------------
+
+
+def test_registry_routes_blank_endpoint_to_responses():
+ from coworker.providers import OpenAIProvider
+ from coworker.providers.registry import build_provider_client
+
+ assert isinstance(
+ build_provider_client("openai", {}, None), OpenAIResponsesProvider
+ )
+ assert isinstance(
+ build_provider_client("openai", {"base_url": " "}, None),
+ OpenAIResponsesProvider,
+ )
+ # A custom endpoint (Azure, vLLM, any compat gateway) keeps Chat Completions…
+ custom = build_provider_client(
+ "openai", {"base_url": "https://my.azure.example/openai/v1"}, None
+ )
+ assert isinstance(custom, OpenAIProvider)
+ # …and so do Ollama and every compat vendor (their own descriptors).
+ assert isinstance(build_provider_client("ollama", {}, None), OpenAIProvider)
+ assert isinstance(
+ build_provider_client("deepseek", {"api_key": "sk-x"}, None), OpenAIProvider
+ )
diff --git a/tests/test_provider_router.py b/tests/test_provider_router.py
index 81e53d30..d5c124a9 100644
--- a/tests/test_provider_router.py
+++ b/tests/test_provider_router.py
@@ -409,12 +409,16 @@ def test_provider_builders(monkeypatch):
with pytest.raises(RuntimeError, match="Gemini"):
build_provider_client("gemini", {}, None)._ensure_client()
- # OpenAI custom endpoint (Azure /openai/v1, OpenRouter, vLLM, …) passes through
+ # OpenAI custom endpoint (Azure /openai/v1, OpenRouter, vLLM, …) passes through and
+ # keeps Chat Completions; a blank endpoint means stock OpenAI → the Responses API.
+ from coworker.providers import OpenAIResponsesProvider
+
o = build_provider_client(
"openai", {"base_url": "https://my.azure.example/openai/v1"}, None
)
+ assert isinstance(o, OpenAIProvider)
assert o._base_url == "https://my.azure.example/openai/v1"
- assert build_provider_client("openai", {}, None)._base_url is None
+ assert isinstance(build_provider_client("openai", {}, None), OpenAIResponsesProvider)
def test_anthropic_gemini_capabilities():