fix(fireworks_ai): send developer input items as system messages on the native responses path

litellm/llms/fireworks_ai/responses/transformation.py

@@ -4,6 +4,7 @@ from typing import TYPE_CHECKING, Final
 from urllib.parse import unquote
 
 import httpx
+from openai.types.responses import EasyInputMessageParam, ResponseInputItemParam
 
 from litellm.llms.fireworks_ai.common_utils import (
     resolve_fireworks_api_key,
@@ -30,6 +31,18 @@ def _session_params(litellm_params: GenericLiteLLMParams) -> Mapping[str, object
     )
 
 
+def _developer_item_as_system(item: ResponseInputItemParam) -> ResponseInputItemParam:
+    if "role" not in item or item["role"] != "developer":
+        return item
+    return EasyInputMessageParam(role="system", content=item["content"], type="message")
+
+
+def _developer_items_as_system(input: str | ResponseInputParam) -> str | ResponseInputParam:
+    if isinstance(input, str):
+        return input
+    return [_developer_item_as_system(item) for item in input]
+
+
 class FireworksAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
     @property
     def custom_llm_provider(self) -> LlmProviders:
@@ -65,7 +78,7 @@ class FireworksAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
     ) -> dict:  # mutable-ok: overrides the base class signature
         return super().transform_responses_api_request(
             model=resolve_fireworks_resource_name(model),
-            input=input,
+            input=_developer_items_as_system(input),
             response_api_optional_request_params=response_api_optional_request_params,
             litellm_params=litellm_params,
             headers=headers,

tests/test_litellm/llms/fireworks_ai/responses/test_fireworks_ai_responses_transformation.py

@@ -160,6 +160,26 @@ def test_responses_call_forwards_previous_response_id_and_store() -> None:
     assert body["input"][0]["call_id"] == "call_abc123"
 
 
+def test_responses_call_sends_developer_items_as_system_messages() -> None:
+    client: Final = _mock_http_client(_fireworks_response("accounts/fireworks/models/kimi-k3"))
+    with patch(HTTPX_CLIENT_FACTORY, return_value=client):
+        litellm.responses(
+            model="fireworks_ai/accounts/fireworks/models/kimi-k3",
+            input=[  # mutable-ok: the Responses API takes input as a JSON list
+                {"role": "user", "content": "Hi there"},
+                {"role": "developer", "content": "Answer with exactly one word."},
+                {"role": "user", "content": [{"type": "input_text", "text": "What is the capital of France?"}]},
+            ],
+            api_key="fw-test-key",
+        )
+    _, _, body = _sent_request(client)
+    assert tuple(body["input"]) == (
+        {"role": "user", "content": "Hi there"},
+        {"role": "system", "content": "Answer with exactly one word.", "type": "message"},
+        {"role": "user", "content": [{"type": "input_text", "text": "What is the capital of France?"}]},
+    )
+
+
 def test_responses_call_sends_session_affinity_for_caller_session_id() -> None:
     client: Final = _mock_http_client(_fireworks_response("accounts/fireworks/models/kimi-k3"))
     with patch(HTTPX_CLIENT_FACTORY, return_value=client):