fix(agents): auto-title crash on metadata=None

.env.example

@@ -50,9 +50,15 @@ AGENTS_SECRET_KEY=
 # at startup and routes per-call telemetry. Per-call gating is governed by
 # the workspace's analytics_consent (off / errors_only / full). Leave blank
 # to disable tracing entirely.
+#
+# LANGFUSE_HOST is the canonical name (matches the LiteLLM / Langfuse SDK
+# convention). LANGFUSE_BASE_URL is accepted as an alias so a self-hosted
+# setup that uses the alternative naming still wires up tracing. If both
+# are set, LANGFUSE_HOST wins.
 LANGFUSE_PUBLIC_KEY=
 LANGFUSE_SECRET_KEY=
 LANGFUSE_HOST=
+# LANGFUSE_BASE_URL=  # alias for LANGFUSE_HOST
 
 # Agent invocation rate limits — operator-level (not per-workspace). Defaults
 # below are 10× the original spec. Override only if you need to throttle

README.md

@@ -258,7 +258,8 @@ BACKEND_CORS_ORIGINS=http://localhost:5173
 # Optional — Langfuse tracing for agent calls (per-workspace consent gates each call).
 LANGFUSE_PUBLIC_KEY=
 LANGFUSE_SECRET_KEY=
-LANGFUSE_HOST=
+LANGFUSE_HOST=          # canonical, matches LiteLLM/Langfuse SDK
+# LANGFUSE_BASE_URL=    # accepted as alias for LANGFUSE_HOST
 ```
 
 ### ⚠️ Required for AI agents: `AGENTS_SECRET_KEY`

backend/app/agents/llm.py

@@ -100,7 +100,7 @@ async def acompletion(
         tools: list[dict] | None = None,
         tool_choice: str | dict | None = None,
         response_format: dict | None = None,
-        metadata: LLMCallMetadata,
+        metadata: LLMCallMetadata | None,
         model_override: str | None = None,
         max_tokens: int | None = None,
         temperature: float | None = None,
@@ -465,7 +465,7 @@ def _safe_completion_cost(resp: Any) -> Decimal | None:
             return None
 
     def _build_langfuse_metadata(
-        self, call_meta: LLMCallMetadata
+        self, call_meta: LLMCallMetadata | None
     ) -> dict | None:
         """Build per-call metadata for the LiteLLM Langfuse callback.
 
@@ -474,6 +474,12 @@ def _build_langfuse_metadata(
         env vars at app startup by ``app/agents/tracing.py`` (task 013); this
         method only constructs the trace identifying info.
         """
+        # Some callers (e.g. the auto-title endpoint) intentionally pass no
+        # metadata — they don't belong to an agent invocation and must never
+        # show up in Langfuse. Treat that as "tracing off" to avoid an
+        # AttributeError on call_meta.analytics_consent.
+        if call_meta is None:
+            return None
         if call_meta.analytics_consent == "off":
             return None
         if not os.environ.get(_LANGFUSE_PUBLIC_KEY_ENV):

backend/app/core/config.py

@@ -1,4 +1,4 @@
-from pydantic import SecretStr
+from pydantic import AliasChoices, Field, SecretStr
 from pydantic_settings import BaseSettings
 
 
@@ -48,7 +48,14 @@ class Settings(BaseSettings):
     # convention and the langfuse/skills setup pattern.
     langfuse_public_key: SecretStr | None = None
     langfuse_secret_key: SecretStr | None = None
-    langfuse_host: str | None = None
+    # Accept either LANGFUSE_HOST (LiteLLM/SDK convention) or
+    # LANGFUSE_BASE_URL (a common typo / alternative naming) so a misnamed
+    # env var doesn't silently disable tracing. Both are documented in
+    # .env.example.
+    langfuse_host: str | None = Field(
+        default=None,
+        validation_alias=AliasChoices("LANGFUSE_HOST", "LANGFUSE_BASE_URL"),
+    )
 
     # Agent invocation rate limits — operator-level, not per-workspace.
     # Defaults are 10× the original spec defaults (which were 600/h, 6000/d,

backend/app/core/database.py

@@ -1,3 +1,4 @@
+import contextlib
 from collections.abc import AsyncGenerator
 
 from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
@@ -19,7 +20,21 @@ async def get_db() -> AsyncGenerator[AsyncSession]:
     async with async_session() as session:
         try:
             yield session
-            await session.commit()
-        except Exception:
-            await session.rollback()
+        except BaseException:
+            # SSE endpoints get cancelled when the frontend aborts (tab close,
+            # navigation, network blip). uvicorn raises asyncio.CancelledError
+            # which is a BaseException — not Exception — so a plain
+            # `except Exception` would miss it and the noisy "Exception
+            # terminating connection" + "Exception in ASGI application"
+            # tracebacks would land in prod logs on every disconnect.
+            #
+            # Catch BaseException, attempt a best-effort rollback, then
+            # re-raise so the framework still treats the request as cancelled.
+            # The rollback itself can also raise CancelledError if the scope
+            # is already torn down — suppress that too, SQLAlchemy still
+            # returns the connection to the pool on its own.
+            with contextlib.suppress(BaseException):
+                await session.rollback()
             raise
+        else:
+            await session.commit()

backend/tests/agents/test_llm.py

@@ -335,6 +335,13 @@ def test_langfuse_metadata_off_returns_none(client: LLMClient):
     assert client._build_langfuse_metadata(meta) is None
 
 
+def test_langfuse_metadata_none_returns_none(client: LLMClient):
+    # Auto-title and other ad-hoc LLM calls pass metadata=None because they
+    # don't belong to an agent invocation. Must not blow up with
+    # AttributeError on call_meta.analytics_consent.
+    assert client._build_langfuse_metadata(None) is None
+
+
 def test_langfuse_metadata_full_with_env_returns_dict(
     client: LLMClient, monkeypatch: pytest.MonkeyPatch
 ):

backend/tests/agents/test_tracing.py

@@ -128,6 +128,34 @@ def test_is_langfuse_configured_false_when_all_missing(
     assert tracing.is_langfuse_configured() is False
 
 
+# ---------------------------------------------------------------------------
+# Env-var alias: LANGFUSE_HOST is canonical, LANGFUSE_BASE_URL is accepted
+# as a fallback so a misnamed env var doesn't silently disable tracing.
+# ---------------------------------------------------------------------------
+
+
+def test_settings_picks_up_langfuse_base_url_alias(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    monkeypatch.delenv("LANGFUSE_HOST", raising=False)
+    monkeypatch.setenv("LANGFUSE_BASE_URL", "https://lf.example.test")
+    monkeypatch.setenv("LANGFUSE_PUBLIC_KEY", "pk-lf-x")
+    monkeypatch.setenv("LANGFUSE_SECRET_KEY", "sk-lf-x")
+    fresh = config_module.Settings()
+    assert fresh.langfuse_host == "https://lf.example.test"
+
+
+def test_settings_prefers_langfuse_host_over_base_url(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    monkeypatch.setenv("LANGFUSE_HOST", "https://canonical.example.test")
+    monkeypatch.setenv("LANGFUSE_BASE_URL", "https://alias.example.test")
+    monkeypatch.setenv("LANGFUSE_PUBLIC_KEY", "pk-lf-x")
+    monkeypatch.setenv("LANGFUSE_SECRET_KEY", "sk-lf-x")
+    fresh = config_module.Settings()
+    assert fresh.langfuse_host == "https://canonical.example.test"
+
+
 # ---------------------------------------------------------------------------
 # setup_litellm_callbacks
 # ---------------------------------------------------------------------------