代理式迴圈 Hook
建立一個 CustomLogger callback,攔截模型回應、在伺服器端完成工具呼叫,並重新執行模型——對呼叫端而言是透明的。
支援的呼叫類型
- 僅
async(同步呼叫不會觸發此 hook) - 僅限非串流(串流回應無法檢查工具呼叫)
- 適用於
/v1/messages與/v1/chat/completions
實作 callback
在 CustomLogger 上覆寫兩個方法:
from litellm.integrations.custom_logger import CustomLogger
from litellm.types.integrations.custom_logger import AgenticLoopPlan, AgenticLoopRequestPatch
MY_TOOL = "my_tool"
class MyToolCallback(CustomLogger):
async def async_should_run_agentic_loop(
self, response, model, messages, tools, stream, custom_llm_provider, kwargs
):
# Return (True, context_dict) if there are tool calls to handle
content = getattr(response, "content", None) or []
calls = [b for b in content if isinstance(b, dict)
and b.get("type") == "tool_use" and b.get("name") == MY_TOOL]
if not calls:
return False, {}
return True, {"tool_calls": calls}
async def async_build_agentic_loop_plan(
self, tools, model, messages, response,
anthropic_messages_provider_config,
anthropic_messages_optional_request_params,
logging_obj, stream, kwargs,
):
calls = tools["tool_calls"]
results = [f"result for {c['input']}" for c in calls] # your logic here
follow_up = messages + [
{"role": "assistant", "content": [
{"type": "tool_use", "id": c["id"], "name": c["name"], "input": c["input"]}
for c in calls
]},
{"role": "user", "content": [
{"type": "tool_result", "tool_use_id": c["id"], "content": results[i]}
for i, c in enumerate(calls)
]},
]
return AgenticLoopPlan(
run_agentic_loop=True,
request_patch=AgenticLoopRequestPatch(messages=follow_up),
)
對於 /v1/chat/completions,請改為覆寫 async_build_chat_completion_agentic_loop_plan——概念相同,optional_params 取代 anthropic_messages_optional_request_params。
註冊它
import litellm
litellm.callbacks = [MyToolCallback()]
或在 config.yaml 中:
litellm_settings:
callbacks: ["my_module.MyToolCallback"]
AgenticLoopPlan 欄位
| 欄位 | 效果 |
|---|---|
run_agentic_loop=True + request_patch | 使用已修補的請求重新執行模型 |
response_override | 直接將此值回傳給呼叫端(不重新執行) |
terminate=True | 停止迴圈,回傳目前回應 |
run_agentic_loop=False(預設) | 略過;接著檢查下一個 callback |
AgenticLoopRequestPatch 接受:model、messages、tools、max_tokens、optional_params、kwargs。
迴圈安全性
- 預設最大重跑次數:
3——可透過kwargs["max_agentic_loops"]針對單一請求覆寫 - 相同的工具呼叫指紋會自動中止迴圈
- 目前深度位於
kwargs["_agentic_loop_depth"]
本倉庫中的範例
litellm/integrations/compression_interception/handler.pylitellm/integrations/websearch_interception/handler.py