fix(responses): stop agentic follow-up from passing request params twice

litellm/llms/custom_httpx/llm_http_handler.py

@@ -5573,17 +5573,19 @@ class BaseLLMHTTPHandler:
 
         internal_keys: Final = {"litellm_logging_obj"}
         kwargs_for_followup: Final = {
-            k: v
-            for k, v in kwargs.items()
-            if not is_interception_internal_key(k, prefixes=NON_CODE_INTERPRETER_INTERCEPTION_INTERNAL_PREFIXES)
-            and k != "_code_interpreter_interception_converted_stream"
-            and k not in internal_keys
-            and k not in optional_params
+            **{
+                k: v
+                for k, v in kwargs.items()
+                if not is_interception_internal_key(k, prefixes=NON_CODE_INTERPRETER_INTERCEPTION_INTERNAL_PREFIXES)
+                and k != "_code_interpreter_interception_converted_stream"
+                and k not in internal_keys
+                and k not in optional_params
+            },
+            **{k: v for k, v in patch.kwargs.items() if k not in optional_params},
+            "_agentic_loop_depth": depth + 1,
+            "max_agentic_loops": max_loops,
+            "_agentic_loop_fingerprints": fingerprints + [fingerprint],
         }
-        kwargs_for_followup.update(patch.kwargs)
-        kwargs_for_followup["_agentic_loop_depth"] = depth + 1
-        kwargs_for_followup["max_agentic_loops"] = max_loops
-        kwargs_for_followup["_agentic_loop_fingerprints"] = fingerprints + [fingerprint]
 
         try:
             response: ResponsesAPIResponse | BaseResponsesAPIStreamingIterator = await litellm.aresponses(

tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py

@@ -3713,3 +3713,47 @@ def test_image_edit_handler_keeps_the_sync_transform():
     assert config.transform_calls == ["sync"]
     assert captured["body"] == {"transformed_by": "sync"}
     assert response.data[0].b64_json == "sync"
+
+
+@pytest.mark.asyncio
+async def test_responses_agentic_followup_does_not_repeat_request_params_from_plan_kwargs(monkeypatch):
+    """A plan whose kwargs repeat a request param must not crash the Responses follow-up with a duplicate keyword"""
+    from litellm.integrations.custom_logger import CustomLogger
+    from litellm.types.integrations.custom_logger import AgenticLoopPlan, AgenticLoopRequestPatch
+
+    followup_calls: list[dict[str, object]] = []
+
+    async def fake_aresponses(**kwargs: object) -> str:
+        followup_calls.append(kwargs)
+        return "followup-response"
+
+    monkeypatch.setattr(litellm, "aresponses", fake_aresponses)
+    request_kwargs: Final = {"prompt_cache_key": "thread-1", "metadata": {"user": "u1"}}
+    plan: Final = AgenticLoopPlan(
+        run_agentic_loop=True,
+        request_patch=AgenticLoopRequestPatch(
+            model="gpt-5",
+            messages=[{"role": "user", "content": "x"}],
+            optional_params={"prompt_cache_key": "thread-1"},
+            kwargs=dict(request_kwargs),
+        ),
+    )
+
+    response: Final = await BaseLLMHTTPHandler()._execute_responses_agentic_plan(
+        plan=plan,
+        model="gpt-5",
+        response_api_optional_request_params={"prompt_cache_key": "thread-1"},
+        logging_obj=Mock(litellm_call_id="call-1"),
+        kwargs=dict(request_kwargs),
+        depth=0,
+        max_loops=3,
+        fingerprints=[],
+        fingerprint="fp",
+        callback=CustomLogger(),
+    )
+
+    assert response == "followup-response"
+    assert len(followup_calls) == 1
+    assert followup_calls[0]["prompt_cache_key"] == "thread-1"
+    assert followup_calls[0]["metadata"] == {"user": "u1"}
+    assert followup_calls[0]["_agentic_loop_depth"] == 1