From 931ab68f1f794f193f25c95fbe4c11a413934914 Mon Sep 17 00:00:00 2001
From: "hanzhi.421" <hanzhi.421@bytedance.com>
Date: Fri, 10 Apr 2026 14:48:55 +0800
Subject: [PATCH] fix: ark_llm apikey and enable response without caching

---
 veadk/agent.py          |  4 ++++
 veadk/models/ark_llm.py | 22 ++++++++++++++++++----
 2 files changed, 22 insertions(+), 4 deletions(-)

diff --git a/veadk/agent.py b/veadk/agent.py
index abfc668c..32320078 100644
--- a/veadk/agent.py
+++ b/veadk/agent.py
@@ -87,6 +87,8 @@ class Agent(LlmAgent):
         example_store (Optional[BaseExampleProvider]): Example store for providing example Q/A.
         enable_shadowchar (bool): Whether to enable shadow character for the agent.
         enable_dynamic_load_skills (bool): Whether to enable dynamic loading of skills.
+        enable_responses_cache (bool): Whether Ark Responses API should reuse
+            `previous_response_id` and caching for multi-turn continuation.
     """
 
     model_config = ConfigDict(arbitrary_types_allowed=True, extra="allow")
@@ -118,6 +120,7 @@ class Agent(LlmAgent):
     tracers: list[BaseTracer] = []
 
     enable_responses: bool = False
+    enable_responses_cache: bool = True
 
     context_cache_config: Optional[ContextCacheConfig] = None
 
@@ -194,6 +197,7 @@ def model_post_init(self, __context: Any) -> None:
                     model=f"{self.model_provider}/{self.model_name}",
                     api_key=self.model_api_key,
                     api_base=self.model_api_base,
+                    enable_responses_cache=self.enable_responses_cache,
                     **self.model_extra_config,
                 )
             else:
diff --git a/veadk/models/ark_llm.py b/veadk/models/ark_llm.py
index 3fd4d035..5dc02258 100644
--- a/veadk/models/ark_llm.py
+++ b/veadk/models/ark_llm.py
@@ -481,10 +481,17 @@ def _remove_caching(request_data: dict) -> None:
     request_data.pop("caching", None)
 
 
-def request_reorganization_by_ark(request_data: Dict) -> Dict:
+def request_reorganization_by_ark(
+    request_data: Dict, enable_responses_cache: bool = True
+) -> Dict:
     # 1. model provider
     request_data = get_model_without_provider(request_data)
 
+    if not enable_responses_cache:
+        request_data.pop("previous_response_id", None)
+        _remove_caching(request_data)
+        request_data.pop("store", None)
+
     # 2. filtered input
     request_data["input"] = filtered_inputs(
         request_data.get("input"),
@@ -672,7 +679,9 @@ async def aresponses(
     ) -> Union[ArkTypeResponse, AsyncStream[ResponseStreamEvent]]:
         # 1. Get request params
         api_base = kwargs.pop("api_base", DEFAULT_VIDEO_MODEL_API_BASE)
-        api_key = kwargs.pop("api_key", settings.model.api_key)
+        api_key = kwargs.pop("api_key", None)
+        if api_key is None:
+            api_key = settings.model.api_key
 
         # 2. Call openai responses
         client = AsyncArk(
@@ -689,6 +698,7 @@ class ArkLlm(Gemini):
     llm_client: ArkLlmClient = Field(default_factory=ArkLlmClient)
     _additional_args: Dict[str, Any] = None
     use_interactions_api: bool = True
+    enable_responses_cache: bool = True
 
     def __init__(self, **kwargs):
         # adk version check
@@ -699,12 +709,14 @@ def __init__(self, **kwargs):
                 "`pip install -U 'google-adk>=1.21.0'`"
             )
         super().__init__(**kwargs)
+        self.enable_responses_cache = kwargs.get("enable_responses_cache", True)
         drop_params = kwargs.pop("drop_params", None)
         self._additional_args = dict(kwargs)
         self._additional_args.pop("llm_client", None)
         self._additional_args.pop("messages", None)
         self._additional_args.pop("tools", None)
         self._additional_args.pop("stream", None)
+        self._additional_args.pop("enable_responses_cache", None)
         if drop_params is not None:
             self._additional_args["drop_params"] = drop_params
 
@@ -733,7 +745,7 @@ async def generate_content_async(
         # ------------------------------------------------------ #
         # get previous_response_id
         previous_response_id = None
-        if llm_request.previous_interaction_id:
+        if self.enable_responses_cache and llm_request.previous_interaction_id:
             previous_response_id = llm_request.previous_interaction_id
         responses_args = {
             "model": self.model,
@@ -786,7 +798,9 @@ async def generate_content_async(
     async def generate_content_via_responses(
         self, responses_args: dict, stream: bool = False
     ):
-        responses_args = request_reorganization_by_ark(responses_args)
+        responses_args = request_reorganization_by_ark(
+            responses_args, enable_responses_cache=self.enable_responses_cache
+        )
         if stream:
             responses_args["stream"] = True
             async for part in await self.llm_client.aresponses(**responses_args):