Skip to content
Python

fix(ollama): default to 127.0.0.1, not localhost, so Windows skips a 2 s wait per connection (OPE-229)

Devika Vermac621ad998528
diff --git a/coworker/providers/ollama_facts.py b/coworker/providers/ollama_facts.py
index 1d753522..0938c4c5 100644
--- a/coworker/providers/ollama_facts.py
+++ b/coworker/providers/ollama_facts.py
@@ -29,7 +29,8 @@ _lock = threading.Lock()
def native_base(base_url: Optional[str]) -> str:
- base = (base_url or "http://localhost:11434").strip().rstrip("/")
+ # The same default as `registry.DEFAULT_OLLAMA_URL` (127.0.0.1, OPE-229).
+ base = (base_url or "http://127.0.0.1:11434").strip().rstrip("/")
return base[: -len("/v1")] if base.endswith("/v1") else base
diff --git a/coworker/providers/openai_provider.py b/coworker/providers/openai_provider.py
index b9db639f..45db982c 100644
--- a/coworker/providers/openai_provider.py
+++ b/coworker/providers/openai_provider.py
@@ -229,7 +229,7 @@ class OpenAIProvider(ProviderClient):
# inject a `client` directly, bypassing all of this.
#
# `base_url` points the same OpenAI SDK at any OpenAI-compatible endpoint — used by the
- # provider router for Ollama (`http://localhost:11434/v1`, with a placeholder key) and,
+ # provider router for Ollama (`http://127.0.0.1:11434/v1`, with a placeholder key) and,
# later, other OpenAI-shaped backends. When None, behavior is identical to stock OpenAI.
#
# `http_client` is an optional httpx.Client handed to the SDK. Ollama uses it to carry
diff --git a/coworker/providers/registry.py b/coworker/providers/registry.py
index bfc94da3..25422ff5 100644
--- a/coworker/providers/registry.py
+++ b/coworker/providers/registry.py
@@ -32,7 +32,10 @@ from .openai_provider import OpenAIProvider
from .openai_responses import OpenAIResponsesProvider
from .vertex_provider import VertexProvider
-DEFAULT_OLLAMA_URL = "http://localhost:11434"
+# 127.0.0.1, not localhost: on Windows `localhost` tries ::1 first, where Ollama does not
+# listen, and every new connection waits about 2 s before falling back (OPE-229). An
+# address the user typed is used as typed.
+DEFAULT_OLLAMA_URL = "http://127.0.0.1:11434"
# The index deliberately lives in the SecretStore too: it names profiles which contain
# credentials, and keeping the two together makes a future secure-store backend a drop-in.
CUSTOM_PROVIDERS_PROFILE = "providers:custom"
diff --git a/coworker/server/manager.py b/coworker/server/manager.py
index 9aa5c4b3..652b9aab 100644
--- a/coworker/server/manager.py
+++ b/coworker/server/manager.py
@@ -92,7 +92,7 @@ from ..providers import (
verify_provider_key,
)
from ..providers.local_server import LOCAL_SERVERS
-from ..providers.registry import CUSTOM_PROVIDERS_PROFILE
+from ..providers.registry import CUSTOM_PROVIDERS_PROFILE, DEFAULT_OLLAMA_URL
from ..secrets import SecretStore, state_dir
from ..sessions import SessionRecord, usage_totals
from ..teams import Actor as TeamActor
@@ -4682,7 +4682,7 @@ class SessionManager:
if cached and now - cached[0] < 30:
return cached[1]
profile = self.secrets.get("provider:ollama") or {}
- base = (profile.get("base_url") or "http://localhost:11434").strip().rstrip("/")
+ base = (profile.get("base_url") or DEFAULT_OLLAMA_URL).strip().rstrip("/")
if base.endswith("/v1"):
base = base[: -len("/v1")]
try:
@@ -4697,7 +4697,7 @@ class SessionManager:
def _ollama_models(self) -> list[str]:
"""Live list of models pulled into the configured Ollama server (via its native
`/api/tags`), as `ollama:<name>` so they're directly selectable. Empty if Ollama is
- unreachable — best-effort, never raises. No stored profile → the default localhost
+ unreachable — best-effort, never raises. No stored profile → the default 127.0.0.1
endpoint, same as `_ollama_alive` (keyless Ollama needs no saved form)."""
from ..providers import ollama_facts
diff --git a/tests/test_provider_router.py b/tests/test_provider_router.py
index bb3a5d14..d0dfec33 100644
--- a/tests/test_provider_router.py
+++ b/tests/test_provider_router.py
@@ -52,12 +52,24 @@ def test_base_url_omitted_when_none(monkeypatch):
# -- ollama URL normalization ---------------------------------------------------
def test_normalize_ollama_url():
- assert _normalize_ollama_url(None) == "http://localhost:11434/v1"
+ # The default is 127.0.0.1 (OPE-229); an address the user typed is kept as typed.
+ assert _normalize_ollama_url(None) == "http://127.0.0.1:11434/v1"
assert (
_normalize_ollama_url("http://localhost:11434") == "http://localhost:11434/v1"
)
assert _normalize_ollama_url("http://h:1/v1/") == "http://h:1/v1"
- assert _normalize_ollama_url(" ") == "http://localhost:11434/v1"
+ assert _normalize_ollama_url(" ") == "http://127.0.0.1:11434/v1"
+
+
+def test_ollama_defaults_never_use_localhost():
+ """OPE-229: on Windows `localhost` tries ::1 first and every new Ollama connection
+ waits about 2 s. Every built-in default must be 127.0.0.1."""
+ from coworker.providers import ollama_facts
+ from coworker.providers.registry import DEFAULT_OLLAMA_URL
+
+ assert DEFAULT_OLLAMA_URL == "http://127.0.0.1:11434"
+ assert ollama_facts.native_base(None) == "http://127.0.0.1:11434"
+ assert ollama_facts.native_base("http://localhost:11434/v1") == "http://localhost:11434"
def test_build_ollama_client_uses_base_url(monkeypatch):