Skip to content

πŸ™Œ OpenHands: AI-Driven Development

fix: treat llm_base_url="" as explicit clear in store_llm_settings (#13471)

Abia75b576f1cf0
diff --git a/openhands/server/routes/settings.py b/openhands/server/routes/settings.py
index 62944d11ce..4affad1d61 100644
--- a/openhands/server/routes/settings.py
+++ b/openhands/server/routes/settings.py
@@ -123,10 +123,9 @@ async def store_llm_settings(
settings.llm_api_key = existing_settings.llm_api_key
if settings.llm_model is None:
settings.llm_model = existing_settings.llm_model
- # if llm_base_url is missing or empty, try to preserve existing or determine appropriate URL
- if not settings.llm_base_url:
- if settings.llm_base_url is None and existing_settings.llm_base_url:
- # Not provided at all (e.g. MCP config save) - preserve existing
+ if settings.llm_base_url is None:
+ # Not provided at all (e.g. MCP config save) - preserve existing or auto-detect
+ if existing_settings.llm_base_url:
settings.llm_base_url = existing_settings.llm_base_url
elif is_openhands_model(settings.llm_model):
# OpenHands models use the LiteLLM proxy
@@ -145,6 +144,9 @@ async def store_llm_settings(
logger.error(
f'Failed to get api_base from litellm for model {settings.llm_model}: {e}'
)
+ elif settings.llm_base_url == '':
+ # Explicitly cleared by the user (basic view save or advanced view clear)
+ settings.llm_base_url = None
# Keep search API key if missing or empty
if not settings.search_api_key:
settings.search_api_key = existing_settings.search_api_key
diff --git a/tests/unit/server/routes/test_settings_store_functions.py b/tests/unit/server/routes/test_settings_store_functions.py
index f51a5b506a..48bc79a280 100644
--- a/tests/unit/server/routes/test_settings_store_functions.py
+++ b/tests/unit/server/routes/test_settings_store_functions.py
@@ -211,9 +211,32 @@ async def test_store_llm_settings_partial_update():
assert result.llm_model == 'gpt-4'
# For SecretStr objects, we need to compare the secret value
assert result.llm_api_key.get_secret_value() == 'existing-api-key'
- # llm_base_url was explicitly cleared (""), so auto-detection runs
- # OpenAI models: litellm.get_api_base() returns https://api.openai.com
- assert result.llm_base_url == 'https://api.openai.com'
+ # llm_base_url="" is an explicit clear β€” must not be repopulated via auto-detection
+ assert result.llm_base_url is None
+
+
+async def test_store_llm_settings_advanced_view_clear_removes_base_url():
+ """Regression test for #13420: clearing Base URL in Advanced view must persist.
+
+ Before the fix, llm_base_url="" was treated identically to llm_base_url=None,
+ causing the backend to re-run auto-detection and overwrite the user's intent.
+ """
+ settings = Settings(
+ llm_model='gpt-4',
+ llm_base_url='', # User deleted the field in Advanced view
+ )
+
+ existing_settings = Settings(
+ llm_model='gpt-4',
+ llm_api_key=SecretStr('my-api-key'),
+ llm_base_url='https://my-custom-proxy.example.com',
+ )
+
+ result = await store_llm_settings(settings, existing_settings)
+
+ # The old custom URL must not come back
+ assert result.llm_base_url is None
@pytest.mark.asyncio