gatewaysdk 0.3.3__tar.gz → 0.3.4__tar.gz

{gatewaysdk-0.3.3 → gatewaysdk-0.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: gatewaysdk
-Version: 0.3.3
+Version: 0.3.4
 Summary: Gateway SDK for Python - tracing, worlds, evaluation and the gateway CLI runtime
 Project-URL: Homepage, https://withgateway.ai
 Project-URL: Documentation, https://withgateway.ai/docs

{gatewaysdk-0.3.3 → gatewaysdk-0.3.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "gatewaysdk"
-version = "0.3.3"
+version = "0.3.4"
 description = "Gateway SDK for Python - tracing, worlds, evaluation and the gateway CLI runtime"
 readme = "README.md"
 requires-python = ">=3.10"

{gatewaysdk-0.3.3 → gatewaysdk-0.3.4}/src/gatewaysdk/__init__.py

@@ -109,7 +109,7 @@ if _TYPE_CHECKING:
     from .users import UserEvent, UserProfile, UserProfileError, UsersClient  # noqa: F401
     from .build import define_tool, register_artifact, register_prompt, register_skill  # noqa: F401
 
-__version__ = "0.3.3"
+__version__ = "0.3.4"
 
 
 # module: public names. Keep this explicit: wildcard ownership made additions

{gatewaysdk-0.3.3 → gatewaysdk-0.3.4}/src/gatewaysdk/benchmark_hub/evals.py

@@ -166,6 +166,10 @@ class GatewayEvalsClient:
         rollouts_per_example: Optional[int] = None,
         environment_ids: Optional[List[str]] = None,
         metadata: Optional[Dict[str, Any]] = None,
+        # When the rollouts actually began (ISO 8601). An engine that harvests
+        # after the fact creates the evaluation minutes after its trials ran;
+        # without this the evaluation's duration reads as the harvest time.
+        started_at: Optional[str] = None,
     ) -> str:
         body: Dict[str, Any] = {"versionId": version_id}
         if dispatch_shard is not None:
@@ -190,6 +194,8 @@ class GatewayEvalsClient:
             body["environmentIds"] = environment_ids
         if metadata is not None:
             body["metadata"] = metadata
+        if started_at is not None:
+            body["startedAt"] = started_at
         evaluation_id = self._post(_BASE, body)["evaluationId"]
         _OPEN_EVALUATIONS[evaluation_id] = self
         return evaluation_id
@@ -304,12 +310,16 @@ class GatewayEvalsClient:
         status: str = "COMPLETED",
         metrics: Optional[Dict[str, Any]] = None,
         cost: Optional[Dict[str, Any]] = None,
+        # When the last rollout finished (ISO 8601); see `started_at` above.
+        completed_at: Optional[str] = None,
     ) -> str:
         body: Dict[str, Any] = {"status": status}
         if metrics is not None:
             body["metrics"] = metrics
         if cost is not None:
             body["cost"] = cost
+        if completed_at is not None:
+            body["completedAt"] = completed_at
         result = self._post(f"{_BASE}/{evaluation_id}/finalize", body)["status"]
         _OPEN_EVALUATIONS.pop(evaluation_id, None)
         return result