agentengine-sdk-python 0.6.4__py3-none-any.whl → 0.7.0__py3-none-any.whl

ksadk/runners/remote_runner.py

@@ -0,0 +1,792 @@
+"""
+RemoteRunner - 远程 Agent 运行时
+
+与 InteractionLoop 配合使用，提供和本地 Runner 一致的接口
+"""
+
+import json
+import os
+from typing import Any, AsyncIterator, Dict, Mapping, Optional, Sequence
+
+from ksadk.runners.base_runner import BaseRunner
+
+
+class RemoteRunner(BaseRunner):
+    """远程 Agent 运行时
+
+    通过 HTTP 调用远程部署的 Agent，兼容 OpenAI API 格式
+    """
+
+    def __init__(
+        self,
+        endpoint: str,
+        api_key: Optional[str] = None,
+        session_id: Optional[str] = None,
+        insecure: bool = False,
+        model: Optional[str] = None,
+        api_format: str = "chat_completions",
+        responses_session_header: Optional[str] = None,
+    ):
+        # 不调用父类 __init__，因为不需要 detection_result
+        self.endpoint = endpoint.rstrip("/")
+        self.api_key = api_key
+        self.session_id = session_id
+        self.insecure = insecure
+        self.model = model
+        self.api_format = self._normalize_api_format(api_format)
+        self.responses_session_header = (
+            str(
+                responses_session_header or os.environ.get("KSADK_RESPONSES_SESSION_HEADER") or ""
+            ).strip()
+            or None
+        )
+        self._agent = None  # 兼容 BaseRunner
+        self._responses_tool_names: dict[str, str] = {}
+        self._responses_tool_args: dict[str, str] = {}
+        self._responses_streamed_item_keys: set[str] = set()
+        self._responses_text_streamed = False
+        self._responses_reasoning_streamed = False
+        self._observed_model: str | None = None  # 流式回包里观察到的真实模型名
+        self.available_models: list[dict[str, Any]] | None = None  # ListAgentModels 拿到的可选模型列表
+
+    @staticmethod
+    def _normalize_api_format(api_format: Optional[str]) -> str:
+        normalized = str(api_format or "chat_completions").strip().lower()
+        if normalized in {"responses", "response", "openresponses", "open_responses"}:
+            return "responses"
+        return "chat_completions"
+
+    def load_agent(self) -> None:
+        """远程 Runner 不需要加载 Agent"""
+        pass
+
+    def prepare_for_request(self, model: Optional[str]) -> None:
+        normalized = self.normalize_requested_model(model)
+        if normalized is None:
+            return
+        self.model = normalized
+
+    def _get_client_kwargs(self) -> dict:
+        """获取 httpx 客户端配置"""
+        is_local = any(x in self.endpoint for x in ["localhost", "127.0.0.1", "0.0.0.0"])
+        kwargs = {"timeout": 120, "trust_env": not is_local}
+        if self.insecure:
+            kwargs["verify"] = False
+        return kwargs
+
+    @staticmethod
+    def _build_responses_input(user_input: Any) -> Any:
+        """Use the OpenAI Responses-compatible simple string shape when possible.
+
+        OpenClaw accepts `input` as a string or item array, but rejects a Chat-style
+        message object whose `content` is a bare string. For remote chat/TUI calls
+        we only need the current user turn, so the string form is the safest common
+        denominator and matches OpenClaw's documented examples.
+        """
+        if isinstance(user_input, Mapping):
+            if user_input.get("type"):
+                return [dict(user_input)]
+            if RemoteRunner._is_chat_style_message(user_input):
+                return RemoteRunner._responses_message_text(user_input)
+            return dict(user_input)
+        if isinstance(user_input, Sequence) and not isinstance(user_input, (str, bytes, bytearray)):
+            items = list(user_input)
+            if any(isinstance(item, Mapping) and item.get("type") for item in items):
+                return [dict(item) if isinstance(item, Mapping) else item for item in items]
+            chat_messages = [
+                item
+                for item in items
+                if isinstance(item, Mapping) and RemoteRunner._is_chat_style_message(item)
+            ]
+            if chat_messages:
+                latest_user = next(
+                    (
+                        item
+                        for item in reversed(chat_messages)
+                        if str(item.get("role") or "").strip().lower() == "user"
+                    ),
+                    chat_messages[-1],
+                )
+                return RemoteRunner._responses_message_text(latest_user)
+            return items
+        return str(user_input or "")
+
+    @staticmethod
+    def _build_responses_conversation_history(history: Any, current_input: Any) -> list[dict[str, Any]]:
+        if not isinstance(history, Sequence) or isinstance(
+            history, (str, bytes, bytearray)
+        ):
+            return []
+
+        messages: list[dict[str, Any]] = []
+        current_text = RemoteRunner._responses_message_text(
+            {"role": "user", "content": current_input}
+        ).strip()
+        for item in history:
+            if not isinstance(item, Mapping):
+                continue
+            role = str(item.get("role") or "").strip().lower()
+            if role == "model":
+                role = "assistant"
+            if role not in {"user", "assistant"}:
+                continue
+            text = RemoteRunner._responses_message_text(item).strip()
+            if not text:
+                continue
+            if role == "user" and current_text and text == current_text:
+                continue
+            messages.append(
+                {
+                    "role": role,
+                    "content": [{"type": "input_text", "text": text}],
+                }
+            )
+        return messages
+
+    @staticmethod
+    def _responses_conversation_name(input_data: Mapping[str, Any], session_id: Optional[str]) -> str:
+        if not session_id:
+            return ""
+        platform_context = input_data.get("platform_context")
+        agent_id = ""
+        if isinstance(platform_context, Mapping):
+            agent_id = str(platform_context.get("agent_id") or "").strip()
+        if agent_id:
+            return f"agentengine:{agent_id}:{session_id}"
+        return f"agentengine:{session_id}"
+
+    @staticmethod
+    def _responses_conversation_value(value: Any) -> str:
+        if isinstance(value, Mapping):
+            return str(value.get("id") or "").strip()
+        return str(value or "").strip()
+
+    def _build_responses_payload(
+        self,
+        input_data: Mapping[str, Any],
+        *,
+        stream: bool,
+    ) -> dict[str, Any]:
+        user_input = input_data.get("input", "")
+        session_id = input_data.get("session_id") or self.session_id
+        previous_response_id = input_data.get("previous_response_id")
+
+        if self.responses_session_header:
+            payload = {
+                "input": self._build_responses_input(user_input),
+                "stream": stream,
+            }
+        else:
+            payload = {
+                "input": self._build_responses_input(user_input),
+                "stream": stream,
+            }
+            history_enabled = bool(input_data.get("responses_conversation")) and not previous_response_id
+            if history_enabled:
+                history = self._build_responses_conversation_history(
+                    input_data.get("history"),
+                    user_input,
+                )
+                if history:
+                    payload["conversation_history"] = history
+            conversation = self._responses_conversation_value(input_data.get("conversation"))
+            if conversation and not previous_response_id:
+                payload["conversation"] = conversation
+            elif (
+                input_data.get("responses_conversation")
+                and session_id
+                and not previous_response_id
+            ):
+                conversation = self._responses_conversation_name(input_data, str(session_id))
+                if conversation:
+                    payload["conversation"] = conversation
+
+        if previous_response_id:
+            payload["previous_response_id"] = str(previous_response_id)
+        tools = self._builtin_response_tool_schemas(input_data)
+        if tools:
+            payload["tools"] = tools
+        return payload
+
+    @staticmethod
+    def _builtin_response_tool_schemas(input_data: Mapping[str, Any] | None = None) -> list[dict[str, Any]]:
+        try:
+            from ksadk.toolsets import builtin_tool_descriptors_for_runtime, builtin_tools_mode, describe_agentengine_tools
+
+            mode = builtin_tools_mode(default="off")
+            descriptors = builtin_tool_descriptors_for_runtime(mode=mode)
+        except Exception:
+            return []
+        deferred_names = RemoteRunner._deferred_tool_names(input_data)
+        if deferred_names:
+            try:
+                direct_descriptors = describe_agentengine_tools(include=deferred_names, profile="coding")
+            except Exception:
+                direct_descriptors = []
+            direct_descriptors = [
+                item
+                for item in direct_descriptors
+                if str(item.get("execution") or "builtin") != "external"
+            ]
+            existing_names = {str(item.get("name") or "") for item in descriptors}
+            descriptors = [
+                *descriptors,
+                *[
+                    item
+                    for item in direct_descriptors
+                    if str(item.get("name") or "") not in existing_names
+                ],
+            ]
+        return [RemoteRunner._response_tool_schema(descriptor) for descriptor in descriptors]
+
+    @staticmethod
+    def _deferred_tool_names(input_data: Mapping[str, Any] | None) -> list[str]:
+        if not isinstance(input_data, Mapping):
+            return []
+        raw_names = input_data.get("deferred_tool_names")
+        if not isinstance(raw_names, Sequence) or isinstance(raw_names, (str, bytes, bytearray)):
+            return []
+        names: list[str] = []
+        seen: set[str] = set()
+        for item in raw_names:
+            name = str(item or "").strip()
+            if not name or name in seen:
+                continue
+            seen.add(name)
+            names.append(name)
+        return names
+
+    @staticmethod
+    def _response_tool_schema(descriptor: Mapping[str, Any]) -> dict[str, Any]:
+        args = descriptor.get("args")
+        properties = dict(args) if isinstance(args, Mapping) else {}
+        required = [
+            name
+            for name, schema in properties.items()
+            if isinstance(schema, Mapping) and schema.get("required") is True
+        ]
+        parameters: dict[str, Any] = {
+            "type": "object",
+            "properties": properties,
+        }
+        if required:
+            parameters["required"] = required
+        return {
+            "type": "function",
+            "name": str(descriptor.get("name") or ""),
+            "description": str(descriptor.get("description") or descriptor.get("name") or ""),
+            "parameters": parameters,
+        }
+
+    @staticmethod
+    def _is_chat_style_message(value: Mapping[str, Any]) -> bool:
+        return not value.get("type") and ("role" in value or "content" in value)
+
+    @staticmethod
+    def _responses_message_text(message: Mapping[str, Any]) -> str:
+        content = message.get("content")
+        if isinstance(content, str):
+            return content
+        if isinstance(content, Sequence) and not isinstance(content, (str, bytes, bytearray)):
+            parts: list[str] = []
+            for part in content:
+                if isinstance(part, str):
+                    parts.append(part)
+                elif isinstance(part, Mapping):
+                    text = part.get("text")
+                    if isinstance(text, str):
+                        parts.append(text)
+            return "".join(parts)
+        return str(content or "")
+
+    def _get_headers(self, session_id: Optional[str] = None) -> dict:
+        """获取请求头"""
+        headers = {"Content-Type": "application/json"}
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+        if self.api_format == "responses" and session_id and self.responses_session_header:
+            headers[self.responses_session_header] = session_id
+        return headers
+
+    @staticmethod
+    def _response_usage_payload(data: Mapping[str, Any]) -> dict[str, Any]:
+        usage = data.get("usage")
+        if isinstance(usage, Mapping):
+            return dict(usage)
+        response = data.get("response")
+        if isinstance(response, Mapping) and isinstance(response.get("usage"), Mapping):
+            return dict(response["usage"])
+        return {}
+
+    async def invoke(self, input_data: Dict[str, Any]) -> Dict[str, Any]:
+        """非流式调用远程 Agent"""
+        import httpx
+
+        user_input = input_data.get("input", "")
+        session_id = input_data.get("session_id") or self.session_id
+
+        if self.api_format == "responses":
+            url = f"{self.endpoint}/v1/responses"
+            payload = self._build_responses_payload(input_data, stream=False)
+        else:
+            url = f"{self.endpoint}/v1/chat/completions"
+            payload = {
+                "messages": [{"role": "user", "content": user_input}],
+                "stream": False,
+            }
+        if session_id and self.api_format != "responses":
+            payload["session_id"] = session_id
+        if self.model:
+            payload["model"] = self.model
+
+        async with httpx.AsyncClient(**self._get_client_kwargs()) as client:
+            response = await client.post(url, json=payload, headers=self._get_headers(session_id))
+            response.raise_for_status()
+            data = response.json()
+
+        usage = self._response_usage_payload(data)
+
+        if self.api_format == "responses":
+            result = {"output": self._extract_responses_output_text(data) or str(data)}
+            if usage:
+                result["usage"] = usage
+            return result
+
+        # 提取 OpenAI Chat Completions 格式响应
+        try:
+            content = data["choices"][0]["message"]["content"]
+        except (KeyError, IndexError):
+            content = str(data)
+
+        result = {"output": content}
+        if usage:
+            result["usage"] = usage
+        return result
+
+    async def stream(self, input_data: Dict[str, Any]) -> AsyncIterator[Dict[str, Any]]:
+        """流式调用远程 Agent"""
+        import httpx
+
+        # 工具 item/call_id 只在单次 response 内有效。RemoteRunner 会被 TUI 多轮复用，
+        # 若保留上一轮的追踪状态，runtime 复用 id 时 response.completed 会被误判为重放。
+        self._responses_tool_names.clear()
+        self._responses_tool_args.clear()
+        self._responses_streamed_item_keys.clear()
+        self._responses_text_streamed = False
+        self._responses_reasoning_streamed = False
+
+        user_input = input_data.get("input", "")
+        session_id = input_data.get("session_id") or self.session_id
+
+        if self.api_format == "responses":
+            url = f"{self.endpoint}/v1/responses"
+            payload = self._build_responses_payload(input_data, stream=True)
+        else:
+            url = f"{self.endpoint}/v1/chat/completions"
+            payload = {
+                "messages": [{"role": "user", "content": user_input}],
+                "stream": True,
+                "stream_options": {"include_usage": True},
+            }
+        if session_id and self.api_format != "responses":
+            payload["session_id"] = session_id
+        if self.model:
+            payload["model"] = self.model
+
+        async with httpx.AsyncClient(**self._get_client_kwargs()) as client:
+            async with client.stream(
+                "POST", url, json=payload, headers=self._get_headers(session_id)
+            ) as response:
+                response.raise_for_status()
+
+                event_name = ""
+                accumulated_text = ""
+                final_sent = False
+                async for line in response.aiter_lines():
+                    if not line:
+                        event_name = ""
+                        continue
+
+                    if line.startswith("event:"):
+                        event_name = line.split(":", 1)[1].strip()
+                        continue
+
+                    if line.startswith("data: "):
+                        data_str = line[6:].strip()
+                        if data_str == "[DONE]":
+                            break
+
+                        try:
+                            data = json.loads(data_str)
+
+                            # 记录回包里的真实模型名：
+                            # chat completions 顶层 data.model；responses 格式在
+                            # data.response.model（response.completed 事件的 response 对象）。
+                            observed = data.get("model") if isinstance(data, Mapping) else None
+                            if not observed and isinstance(data, Mapping):
+                                resp = data.get("response")
+                                if isinstance(resp, Mapping):
+                                    observed = resp.get("model")
+                            if observed and not isinstance(observed, (list, dict)):
+                                self._observed_model = str(observed)
+
+                            if self.api_format == "responses":
+                                async for item in self._iter_responses_stream_events(
+                                    data, event_name=event_name
+                                ):
+                                    yield item
+                                continue
+
+                            # 兜底：api_format 标成 chat 但 runtime 实发 responses 格式
+                            # 事件（event_name 或 data.type 以 response. 开头）→ 按 responses
+                            # 解析，否则 response.reasoning.delta 会被 chat 路径当顶层 text 显示。
+                            _ev = event_name or (str(data.get("type")) if isinstance(data, Mapping) else "")
+                            if _ev.startswith("response."):
+                                async for item in self._iter_responses_stream_events(
+                                    data, event_name=event_name
+                                ):
+                                    yield item
+                                continue
+
+                            # 解析 OpenAI Chat Completions 流式格式
+                            choices = data.get("choices", [])
+                            usage = data.get("usage")
+                            if isinstance(usage, Mapping):
+                                final_chunk = {"type": "final", "usage": dict(usage)}
+                                metadata = data.get("metadata")
+                                if isinstance(metadata, Mapping):
+                                    final_chunk["metadata"] = dict(metadata)
+                                yield final_chunk
+                                final_sent = True
+                            if choices:
+                                delta = choices[0].get("delta", {})
+                                content = delta.get("content", "")
+                                reasoning = delta.get("reasoning_content", "")
+
+                                if reasoning:
+                                    yield {"delta": reasoning, "type": "thinking"}
+                                if content:
+                                    accumulated_text += content
+                                    yield {"delta": content, "type": "text"}
+                            else:
+                                # 非标准简化流（runtime 直发顶层 delta，无 choices 包装）：
+                                # 兜底提取顶层 delta 作为正文，与 _extract_content 口径一致。
+                                top_delta = data.get("delta")
+                                if isinstance(top_delta, str) and top_delta:
+                                    accumulated_text += top_delta
+                                    yield {"delta": top_delta, "type": "text"}
+
+                        except json.JSONDecodeError:
+                            pass
+                if self.api_format != "responses" and accumulated_text and not final_sent:
+                    yield {"output": accumulated_text, "type": "final"}
+
+    @staticmethod
+    def _extract_responses_output_text(data: Dict[str, Any]) -> str:
+        output_text = data.get("output_text")
+        if output_text:
+            return str(output_text)
+        output = data.get("output") or []
+        for item in output:
+            if not isinstance(item, dict):
+                continue
+            for content in item.get("content") or []:
+                if isinstance(content, dict) and content.get("text"):
+                    return str(content["text"])
+        return ""
+
+    @staticmethod
+    def _stringify_responses_payload(value: Any) -> str:
+        if value is None:
+            return ""
+        if isinstance(value, str):
+            return value
+        if isinstance(value, (dict, list)):
+            return json.dumps(value, ensure_ascii=False, indent=2)
+        return str(value)
+
+    @staticmethod
+    def _responses_error_message(data: Dict[str, Any]) -> str:
+        response = data.get("response") if isinstance(data.get("response"), dict) else data
+        error = response.get("error") if isinstance(response, dict) else data.get("error")
+        if isinstance(error, dict):
+            return str(error.get("message") or error.get("code") or "Agent 运行失败")
+        if error:
+            return str(error)
+        return "Agent 运行失败"
+
+    @staticmethod
+    def _responses_item_key(item: Dict[str, Any], data: Dict[str, Any]) -> str:
+        return str(
+            item.get("id")
+            or item.get("item_id")
+            or item.get("call_id")
+            or data.get("item_id")
+            or data.get("call_id")
+            or data.get("output_index")
+            or ""
+        )
+
+    def _remember_responses_tool(
+        self, key: str, item: Dict[str, Any], name: str, args: str
+    ) -> None:
+        if key:
+            self._responses_streamed_item_keys.add(key)
+        if key:
+            self._responses_tool_names[key] = name
+            self._responses_tool_args[key] = args
+        call_id = str(item.get("call_id") or "")
+        if call_id:
+            self._responses_streamed_item_keys.add(call_id)
+            self._responses_tool_names[call_id] = name
+            self._responses_tool_args[call_id] = args
+
+    def _responses_tool_name(self, key: str, item: Dict[str, Any], fallback: str = "tool") -> str:
+        call_id = str(item.get("call_id") or "")
+        return str(
+            item.get("name")
+            or item.get("tool_name")
+            or (self._responses_tool_names.get(key) if key else "")
+            or (self._responses_tool_names.get(call_id) if call_id else "")
+            or fallback
+        )
+
+    @staticmethod
+    def _responses_item_text(item: Dict[str, Any]) -> str:
+        for field in ("output_text", "text", "summary_text", "summary", "delta"):
+            value = item.get(field)
+            if isinstance(value, str) and value:
+                return value
+        content = item.get("content")
+        if isinstance(content, list):
+            parts: list[str] = []
+            for part in content:
+                if isinstance(part, str):
+                    parts.append(part)
+                elif isinstance(part, dict) and isinstance(part.get("text"), str):
+                    parts.append(part["text"])
+            return "".join(parts)
+        if isinstance(content, str):
+            return content
+        return ""
+
+    async def _iter_responses_output_item(
+        self,
+        data: Dict[str, Any],
+        *,
+        status: str,
+        replay_completed_text: bool = False,
+        replay_completed_reasoning: bool = False,
+    ) -> AsyncIterator[Dict[str, Any]]:
+        item = data.get("item") or data.get("output_item") or data
+        if not isinstance(item, dict):
+            return
+
+        item_type = str(item.get("type") or "").strip()
+        key = self._responses_item_key(item, data)
+        if item_type == "function_call":
+            name = self._responses_tool_name(key, item)
+            args = self._stringify_responses_payload(
+                item.get("arguments")
+                if "arguments" in item
+                else item.get("args", item.get("input"))
+            )
+            self._remember_responses_tool(key, item, name, args)
+            yield {"type": "tool_call", "tool_name": name, "tool_args": args, "status": status, "call_id": key}
+            return
+
+        if item_type == "function_call_output":
+            name = self._responses_tool_name(key, item)
+            if key:
+                self._responses_streamed_item_keys.add(key)
+            call_id = str(item.get("call_id") or "")
+            if call_id:
+                self._responses_streamed_item_keys.add(call_id)
+            output = self._stringify_responses_payload(
+                item.get("output") if "output" in item else item.get("result", item.get("content"))
+            )
+            yield {"type": "tool_result", "tool_name": name, "tool_output": output, "call_id": call_id or key}
+            return
+
+        if item_type == "mcp_approval_request":
+            name = str(item.get("name") or "approval")
+            args = self._stringify_responses_payload(item.get("arguments") or item.get("args"))
+            approval_request_id = str(item.get("id") or item.get("approval_request_id") or "")
+            yield {
+                "type": "tool_call",
+                "tool_name": name,
+                "tool_args": args,
+                "status": "paused",
+                "approval_request_id": approval_request_id,
+            }
+            yield {
+                "type": "interrupt",
+                "interrupt_info": {
+                    "id": approval_request_id,
+                    "approval_request_id": approval_request_id,
+                    "name": name,
+                    "server_label": str(item.get("server_label") or ""),
+                },
+            }
+            return
+
+        if item_type in {"reasoning", "reasoning_summary", "reasoning_summary_text"}:
+            text = self._responses_item_text(item)
+            if text and (status != "completed" or replay_completed_reasoning):
+                self._responses_reasoning_streamed = True
+                yield {"delta": text, "type": "thinking"}
+            return
+
+        if item_type == "message":
+            text = self._responses_item_text(item)
+            if text and (status != "completed" or replay_completed_text):
+                self._responses_text_streamed = True
+                yield {"delta": text, "type": "text"}
+            return
+
+    async def _iter_responses_stream_events(
+        self,
+        data: Dict[str, Any],
+        *,
+        event_name: str = "",
+    ) -> AsyncIterator[Dict[str, Any]]:
+        event_type = str(data.get("type") or event_name or data.get("_event") or "")
+        if event_name == "response.reasoning.delta":
+            delta = data.get("delta")
+            if delta:
+                self._responses_reasoning_streamed = True
+                yield {"delta": str(delta), "type": "thinking"}
+            return
+        if event_type in {
+            "response.reasoning.delta",
+            "response.reasoning_text.delta",
+            "response.reasoning_summary.delta",
+            "response.reasoning_summary_text.delta",
+        }:
+            delta = data.get("delta") or data.get("text")
+            if delta:
+                self._responses_reasoning_streamed = True
+                yield {"delta": str(delta), "type": "thinking"}
+            return
+        if event_type in {"response.output_text.delta", "response.output_text.done"}:
+            text = data.get("delta") or data.get("text")
+            if text and (
+                event_type == "response.output_text.delta" or not self._responses_text_streamed
+            ):
+                self._responses_text_streamed = True
+                yield {"delta": str(text), "type": "text"}
+            return
+        if event_type == "response.output_item.added":
+            async for item in self._iter_responses_output_item(data, status="running"):
+                yield item
+            return
+        if event_type == "response.output_item.done":
+            async for item in self._iter_responses_output_item(data, status="completed"):
+                yield item
+            return
+        # ksadk/runtime 特有 tool 事件（对标 hosted UI responses-stream.js:176-186）
+        if event_type in {"response.tool_call", "response.ksadk.stage_tool_call"}:
+            name = str(data.get("name") or data.get("tool_name") or "tool")
+            args = self._stringify_responses_payload(data.get("args") if "args" in data else data.get("arguments"))
+            cid = str(data.get("call_id") or data.get("item_id") or data.get("run_id") or "")
+            yield {"type": "tool_call", "tool_name": name, "tool_args": args, "status": "running", "call_id": cid}
+            return
+        if event_type in {
+            "response.tool_result",
+            "response.ksadk.tool_result",
+            "response.ksadk.stage_tool_result",
+        }:
+            name = str(data.get("name") or data.get("tool_name") or "tool")
+            output = self._stringify_responses_payload(data.get("output") if "output" in data else data.get("result"))
+            cid = str(data.get("call_id") or data.get("item_id") or data.get("run_id") or "")
+            yield {
+                "type": "tool_result",
+                "tool_name": name,
+                "tool_output": output,
+                "call_id": cid,
+            }
+            return
+        if event_type == "response.function_call_arguments.delta":
+            key = str(data.get("item_id") or data.get("call_id") or "")
+            name = self._responses_tool_name(key, data)
+            args = f"{self._responses_tool_args.get(key, '')}{str(data.get('delta') or '')}"
+            if key:
+                self._responses_tool_args[key] = args
+            yield {"type": "tool_call", "tool_name": name, "tool_args": args, "status": "running", "call_id": key}
+            return
+        if event_type == "response.function_call_arguments.done":
+            key = str(data.get("item_id") or data.get("call_id") or "")
+            name = self._responses_tool_name(key, data)
+            args = self._stringify_responses_payload(
+                data.get("arguments") or self._responses_tool_args.get(key, "")
+            )
+            if key:
+                self._responses_tool_args[key] = args
+            yield {"type": "tool_call", "tool_name": name, "tool_args": args, "status": "running", "call_id": key}
+            return
+        if event_type == "response.completed":
+            response = data.get("response") if isinstance(data.get("response"), dict) else data
+            output = response.get("output") if isinstance(response, dict) else None
+            if isinstance(output, list):
+                replayed_item_keys = set(self._responses_streamed_item_keys)
+                replay_completed_text = not self._responses_text_streamed
+                replay_completed_reasoning = not self._responses_reasoning_streamed
+                for item in output:
+                    if isinstance(item, dict):
+                        key = self._responses_item_key(item, {"item": item})
+                        call_id = str(item.get("call_id") or "")
+                        if (key and key in replayed_item_keys) or (
+                            call_id and call_id in replayed_item_keys
+                        ):
+                            continue
+                        async for projected in self._iter_responses_output_item(
+                            {"item": item},
+                            status="completed",
+                            replay_completed_text=replay_completed_text,
+                            replay_completed_reasoning=replay_completed_reasoning,
+                        ):
+                            yield projected
+                chunk = {
+                    "type": "responses_output",
+                    "output": output,
+                    "response_id": response.get("id"),
+                }
+                usage = response.get("usage")
+                if isinstance(usage, Mapping):
+                    chunk["usage"] = dict(usage)
+                metadata = response.get("metadata")
+                if isinstance(metadata, Mapping):
+                    chunk["metadata"] = dict(metadata)
+                yield chunk
+            return
+        # response.content_part.delta：glm 等模型把 reasoning 走这个事件，
+        # partType 含 "reasoning" → thinking（不显示），否则 text（对标 hosted UI）。
+        if event_type == "response.content_part.delta":
+            part = data.get("part") if isinstance(data.get("part"), dict) else {}
+            delta = data.get("delta")
+            part_type = str(part.get("type") or (delta.get("type") if isinstance(delta, dict) else "") or data.get("content_type") or "")
+            text = ""
+            if isinstance(delta, dict):
+                text = str(delta.get("text") or delta.get("content") or "")
+            elif isinstance(delta, str):
+                text = delta
+            if not text:
+                text = str(data.get("text") or "")
+            if not text:
+                return
+            if "reasoning" in part_type:
+                self._responses_reasoning_streamed = True
+                yield {"delta": text, "type": "thinking"}
+            else:
+                self._responses_text_streamed = True
+                yield {"delta": text, "type": "text"}
+            return
+        if event_type == "response.failed":
+            yield {"type": "error", "message": self._responses_error_message(data)}
+            return
+        if event_type == "response.incomplete":
+            yield {"type": "error", "message": "Agent 响应未完成"}
+            return
+        # 未知事件默认丢弃（对标 hosted UI responses-stream.js: 未知事件 return []）。
+        # 之前默认把 data.delta 当 text 显示，会导致 reasoning 走未识别事件时漏进正文。
+        # 正文已由 output_text.delta / content_part.delta 显式处理，未知事件不该当正文。