fix: keep selected prompt and parameter loads out of fallback cache

py/src/braintrust/logger.py

@@ -2066,6 +2066,7 @@ def load_prompt(
     :returns: The prompt object.
     """
     effective_environment = None if version is not None else environment
+    should_fall_back_to_cache = version is None and effective_environment is None
 
     if id:
         # When loading by ID, we don't need project or slug
@@ -2102,7 +2103,7 @@ def compute_metadata():
             if not _is_loader_cache_fallback_error(server_error):
                 raise
             # If environment or version was specified, don't fall back to cache
-            if effective_environment is not None or version is not None:
+            if not should_fall_back_to_cache:
                 raise ValueError(f"Prompt not found with specified parameters") from server_error
 
             eprint(f"Failed to load prompt, attempting to fall back to cache: {server_error}")
@@ -2138,6 +2139,9 @@ def compute_metadata():
                 )
             resp_prompt = response["objects"][0]
         prompt = PromptSchema.from_dict_deep(resp_prompt)
+        # Selected versions/environments must not replace the default fallback entry.
+        if not should_fall_back_to_cache:
+            return prompt
         try:
             if id:
                 _state._prompt_cache.set(
@@ -2322,6 +2326,8 @@ def load_parameters(
         )
 
     parameters = RemoteEvalParameters.from_function_row(response["objects"][0])
+    if not should_fall_back_to_cache:
+        return parameters
     try:
         if id:
             _state._parameters_cache.set(parameters, id=id, cache_namespace=cache_namespace)

py/src/braintrust/test_logger.py

@@ -705,6 +705,52 @@ def test_load_prefers_version_over_environment(load, endpoint, response, lookup,
     assert endpoint_mock.call_args_list == [expected_call]
 
 
+@pytest.mark.parametrize(
+    ("load", "response_factory", "resource", "cache_name", "cache_type"),
+    [
+        (braintrust.load_prompt, _prompt_response, "prompts.get_prompt", "_prompt_cache", PromptCache),
+        (
+            braintrust.load_parameters,
+            _parameters_response,
+            "functions.get_function",
+            "_parameters_cache",
+            ParametersCache,
+        ),
+    ],
+    ids=["prompt", "parameters"],
+)
+@pytest.mark.parametrize("by_id", [False, True], ids=["by-slug", "by-id"])
+@pytest.mark.parametrize("selector", [{"environment": "staging"}, {"version": "v2"}], ids=["environment", "version"])
+def test_load_selected_values_do_not_replace_latest_cache(
+    load, response_factory, resource, cache_name, cache_type, by_id, selector
+):
+    simulate_login()
+    latest_response = response_factory("saved")
+    selected_response = copy.deepcopy(latest_response)
+    selected_response["objects"][0]["_xact_id"] = "v2"
+    lookup = {"project": "test-project", "slug": "saved"}
+    if by_id:
+        resource += "_id"
+        latest_response = latest_response["objects"][0]
+        selected_response = selected_response["objects"][0]
+        lookup = {"id": latest_response["id"]}
+
+    mock_api_client = MagicMock()
+    endpoint_mock = operator.attrgetter(resource)(mock_api_client)
+    endpoint_mock.side_effect = [latest_response, selected_response, _http_error(503)]
+    cache = cache_type(memory_cache=LRUCache(max_size=10))
+
+    with (
+        patch.object(logger._state, "api_client", return_value=mock_api_client),
+        patch.object(logger._state, cache_name, cache),
+    ):
+        assert load(**lookup).version == "v1"
+        assert load(**lookup, **selector).version == "v2"
+        assert load(**lookup).version == "v1"
+
+    assert endpoint_mock.call_count == 3
+
+
 def test_load_parameters_returns_remote_object():
     simulate_login()
     mock_api_client = MagicMock()