Skip to content

Self-improving AI Employees for everyone

fix(llm): log where a failed call went on agent paths too

trohitg033543cfbe7d
diff --git a/server/services/llm/unifier.py b/server/services/llm/unifier.py
index 08830595..925bb98f 100644
--- a/server/services/llm/unifier.py
+++ b/server/services/llm/unifier.py
@@ -123,16 +123,18 @@ class ChatUnifier:
)
except LLMError as error:
# Raised by provider logic rather than the SDK, e.g. a 2xx that
- # carried an error body instead of a completion.
+ # carried an error body instead of a completion. Logged whether
+ # or not it is translated: agent steps take it untranslated, and
+ # only this layer knows where the call went (RFC-0003 D10).
+ self._log_failure("LLM provider request failed", error, entry, provider)
if not translate_errors:
raise
- self._log_failure("LLM provider request failed", error, entry)
raise NodeUserError(error.user_message) from error
except spec.sdk_exception_types as e:
error = LLMError.from_exception(provider, e)
+ self._log_failure("LLM provider request failed", error, entry, provider)
if not translate_errors:
raise error from e
- self._log_failure("LLM provider request failed", error, entry)
raise NodeUserError(error.user_message) from error
except (ValueError, TypeError, OSError) as e:
# Only normalize generic configuration/transport failures raised
@@ -152,14 +154,14 @@ class ChatUnifier:
category=category,
retryable=category == LLMErrorCategory.CONNECTION,
)
- if not translate_errors:
- raise error from e
logger.warning(
"LLM client construction failed",
provider=error.provider,
category=error.category.value,
retryable=error.retryable,
)
+ if not translate_errors:
+ raise error from e
raise NodeUserError(error.user_message) from error
finally:
if entry is not None:
@@ -182,11 +184,11 @@ class ChatUnifier:
)
models = await entry.client.fetch_models(api_key)
except LLMError as error:
- self._log_failure("LLM model-list request failed", error, entry)
+ self._log_failure("LLM model-list request failed", error, entry, provider)
raise NodeUserError(error.user_message) from error
except spec.sdk_exception_types as e:
error = LLMError.from_exception(provider, e)
- self._log_failure("LLM model-list request failed", error, entry)
+ self._log_failure("LLM model-list request failed", error, entry, provider)
raise NodeUserError(error.user_message) from error
except (ValueError, TypeError, OSError) as e:
if entry is not None:
@@ -430,18 +432,20 @@ class ChatUnifier:
) from None
@staticmethod
- def _log_failure(event: str, error: LLMError, entry: Optional[_ClientEntry]) -> None:
+ def _log_failure(event: str, error: LLMError, entry: Optional[_ClientEntry], provider: str) -> None:
"""One WARN line per provider failure, naming where the call went.
- ``url`` is redacted (no userinfo, query or fragment) and
- ``url_source`` says which setting produced it, so a routing mistake
- is distinguishable from a model failure (RFC-0003 D10). The key is
- never logged.
+ ``provider`` is the reference the call was made with, so a named
+ endpoint keeps its slug even when the provider raised under its
+ generic name. ``url`` is redacted (no userinfo, query or fragment)
+ and ``url_source`` says which setting produced it, so a routing
+ mistake is distinguishable from a model failure (RFC-0003 D10).
+ The key is never logged.
"""
client = entry.client if entry is not None else None
logger.warning(
event,
- provider=error.provider,
+ provider=provider,
category=error.category.value,
retryable=error.retryable,
status_code=error.status_code,
diff --git a/server/tests/llm/test_protocol_guard.py b/server/tests/llm/test_protocol_guard.py
index cbca3464..4002dd3e 100644
--- a/server/tests/llm/test_protocol_guard.py
+++ b/server/tests/llm/test_protocol_guard.py
@@ -85,6 +85,36 @@ class TestTwoHundredWithoutChoices:
model="m",
)
+ async def test_an_untranslated_failure_still_logs_where_the_call_went(self):
+ # Agent steps take errors untranslated (translate_errors=False); the
+ # failure log must not depend on translation (D10).
+ error = LLMError(message="raw", provider="openai_compatible", category=LLMErrorCategory.PROTOCOL)
+ client = MagicMock()
+ client.chat = AsyncMock(side_effect=error)
+ client.endpoint_url = LEAKY_URL
+ client.url_source = "proxy"
+ spec = ProviderSpec(
+ name="openai_compatible", factory=MagicMock(return_value=client), sdk_exception_refs=("openai:OpenAIError",)
+ )
+ auth = MagicMock()
+ auth.get_api_key = AsyncMock(return_value=LEAKY_URL)
+ unifier = ChatUnifier(defaults={"providers": {}}, auth_service=auth)
+
+ with patch("services.llm.unifier.get_provider", return_value=spec), patch("services.llm.unifier.logger") as log:
+ with pytest.raises(LLMError):
+ await unifier.chat(
+ provider="openai_compatible:home",
+ api_key="k",
+ messages=[Message(role="user", content="hi")],
+ model="m",
+ translate_errors=False,
+ )
+
+ logged = log.warning.call_args.kwargs
+ assert logged["provider"] == "openai_compatible:home"
+ assert logged["url"] == "http://host:1234/"
+ assert logged["url_source"] == "proxy"
+
class TestTheStoredKeyIsSentAsStored:
def test_openai_compatible_client_keeps_the_key_with_a_proxy(self):