fix(models): 净化 invalid_tool_call 与孤儿 ToolMessage，规避 DeepSeek 等接口报错

backend/package/yuxi/models/chat.py

@@ -162,8 +162,11 @@ def _standardize_content(self, message: AIMessage, raw: dict) -> None:
     def _get_request_payload(self, input_, *, stop=None, **kwargs):
         """支持推理的供应商在工具续答时接收完整原文。"""
         payload = super()._get_request_payload(input_, stop=stop, **kwargs)
-        if self.preserve_reasoning and "messages" in payload:
-            originals = self._convert_input(input_).to_messages()
+        if "messages" not in payload:
+            return payload
+        originals = self._convert_input(input_).to_messages()
+        _sanitize_wire_invalid_tool_calls(payload["messages"], originals)
+        if self.preserve_reasoning:
             for original, wire in zip(originals, payload["messages"], strict=True):
                 if isinstance(original, AIMessage):
                     content = wire.get("content")
@@ -249,5 +252,40 @@ def select_model(model_spec: str, **kwargs) -> LangChainChatAdapter:
     )
 
 
+def _sanitize_wire_invalid_tool_calls(messages: list[dict], originals: list) -> None:
+    """把 wire 消息里来自 invalid_tool_calls 的截断 function 转成失败反馈。
+
+    langchain_openai 把 AIMessage.invalid_tool_calls 序列化成 type:"function"、
+    arguments 为截断 JSON 的 tool_call，DeepSeek 等接口因此报参数解析失败。这里在
+    发送边界按 tool_call id 移除这些截断调用，并在 content 里给模型明确的 text 反馈，
+    而不是发出畸形参数。原始 checkpoint（LangChain 消息对象）不动。
+    """
+    for wire, original in zip(messages, originals, strict=True):
+        if not isinstance(original, AIMessage):
+            continue
+        invalid = list(original.invalid_tool_calls or [])
+        if not invalid:
+            continue
+        invalid_ids = {call.get("id") for call in invalid if call.get("id")}
+        tool_calls = wire.get("tool_calls")
+        if isinstance(tool_calls, list):
+            kept = [call for call in tool_calls if call.get("id") not in invalid_ids]
+            if kept:
+                wire["tool_calls"] = kept
+            else:
+                wire.pop("tool_calls", None)
+        feedback = "；".join(
+            f"[工具调用失败] {call.get('name') or 'unknown'}: {call.get('error') or 'arguments malformed or truncated'}"
+            for call in invalid
+        )
+        content = wire.get("content")
+        if isinstance(content, list):
+            content.append({"type": "text", "text": feedback})
+        elif content:
+            wire["content"] = [{"type": "text", "text": content}, {"type": "text", "text": feedback}]
+        else:
+            wire["content"] = [{"type": "text", "text": feedback}]
+
+
 if __name__ == "__main__":
     pass

backend/test/unit/models/test_chat_invalid_tool_call_sanitize.py

@@ -0,0 +1,178 @@
+"""invalid_tool_call 窄方案：在 ChatCompletionsAdapter 发送边界净化 wire 载荷。
+
+langchain_openai 把 AIMessage.invalid_tool_calls 序列化成 type:"function"、arguments
+为截断 JSON 的 tool_call，DeepSeek 等接口因此报参数解析失败。窄方案在
+ChatCompletionsAdapter._get_request_payload 里按 id 移除这些截断调用，并在 content
+里给模型明确的 text 反馈，覆盖同步/异步/流式/非流式四条路径；不改 Anthropic/Gemini，
+不删孤儿 ToolMessage（按根因单独处理），原始 checkpoint 消息对象不动。
+"""
+
+from __future__ import annotations
+
+import json
+import os
+
+import pytest
+from langchain_core.messages import AIMessage, HumanMessage, InvalidToolCall
+from pydantic import SecretStr
+
+os.environ.setdefault("OPENAI_API_KEY", "test-key")
+
+from yuxi.models.chat import ChatCompletionsAdapter, _sanitize_wire_invalid_tool_calls
+
+
+def _adapter() -> ChatCompletionsAdapter:
+    return ChatCompletionsAdapter(
+        model="deepseek-chat",
+        api_key=SecretStr("test-key"),
+        base_url="http://test.local/v1",
+    )
+
+
+def _invalid_call() -> InvalidToolCall:
+    return InvalidToolCall(name="kbs_search", args='{"query":', id="call-bad", error="Unterminated string")
+
+
+def _messages_with_invalid() -> list:
+    return [
+        HumanMessage(content="hi"),
+        AIMessage(content="", invalid_tool_calls=[_invalid_call()]),
+    ]
+
+
+# ---- 单元：wire 净化函数 ----
+
+
+def test_pure_invalid_call_is_removed_and_feedback_added():
+    wire = [
+        {"role": "user", "content": "hi"},
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {"type": "function", "id": "call-bad", "function": {"name": "kbs_search", "arguments": '{"query":'}}
+            ],
+        },
+    ]
+    originals = [
+        HumanMessage(content="hi"),
+        AIMessage(content="", invalid_tool_calls=[_invalid_call()]),
+    ]
+
+    _sanitize_wire_invalid_tool_calls(wire, originals)
+
+    assert "tool_calls" not in wire[1]
+    assert wire[1]["content"] == [{"type": "text", "text": "[工具调用失败] kbs_search: Unterminated string"}]
+
+
+def test_mixed_keeps_valid_call_and_adds_feedback():
+    wire = [
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "type": "function",
+                    "id": "call-good",
+                    "function": {"name": "kbs_search", "arguments": '{"query": "ok"}'},
+                },
+                {"type": "function", "id": "call-bad", "function": {"name": "kbs_search", "arguments": '{"query":'}},
+            ],
+        },
+    ]
+    originals = [
+        AIMessage(
+            content="",
+            tool_calls=[{"id": "call-good", "name": "kbs_search", "args": {"query": "ok"}, "type": "tool_call"}],
+            invalid_tool_calls=[_invalid_call()],
+        ),
+    ]
+
+    _sanitize_wire_invalid_tool_calls(wire, originals)
+
+    assert [call["id"] for call in wire[0]["tool_calls"]] == ["call-good"]
+    assert wire[0]["content"] == [{"type": "text", "text": "[工具调用失败] kbs_search: Unterminated string"}]
+
+
+def test_no_invalid_call_leaves_wire_unchanged():
+    wire = [
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "type": "function",
+                    "id": "call-good",
+                    "function": {"name": "kbs_search", "arguments": '{"query": "ok"}'},
+                }
+            ],
+        }
+    ]
+    originals = [
+        AIMessage(
+            content="",
+            tool_calls=[{"id": "call-good", "name": "kbs_search", "args": {"query": "ok"}, "type": "tool_call"}],
+        ),
+    ]
+    snapshot = json.loads(json.dumps(wire))
+
+    _sanitize_wire_invalid_tool_calls(wire, originals)
+
+    assert wire == snapshot
+
+
+# ---- 集成：真实 adapter + mock HTTP，捕获请求体 ----
+
+
+def _assert_body_has_no_invalid_tool_call(request) -> None:
+    body = json.loads(request.content)
+    assistant = body["messages"][1]
+    # 截断 function 已移除，换成 text 反馈
+    assert "tool_calls" not in assistant
+    assert assistant["content"] == [{"type": "text", "text": "[工具调用失败] kbs_search: Unterminated string"}]
+
+
+def _ok_response() -> dict:
+    return {
+        "choices": [{"message": {"role": "assistant", "content": "ok"}}],
+        "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+    }
+
+
+def _stream_response() -> str:
+    return 'data: {"choices":[{"delta":{"content":"ok"},"index":0}]}\n\ndata: [DONE]\n\n'
+
+
+def test_invoke_sanitizes_request_body(httpx_mock):
+    httpx_mock.add_response(url="http://test.local/v1/chat/completions", json=_ok_response())
+
+    _adapter().invoke(_messages_with_invalid())
+
+    _assert_body_has_no_invalid_tool_call(httpx_mock.get_request())
+
+
+@pytest.mark.asyncio
+async def test_ainvoke_sanitizes_request_body(httpx_mock):
+    httpx_mock.add_response(url="http://test.local/v1/chat/completions", json=_ok_response())
+
+    await _adapter().ainvoke(_messages_with_invalid())
+
+    _assert_body_has_no_invalid_tool_call(httpx_mock.get_request())
+
+
+def test_stream_sanitizes_request_body(httpx_mock):
+    httpx_mock.add_response(url="http://test.local/v1/chat/completions", text=_stream_response())
+
+    list(_adapter().stream(_messages_with_invalid()))
+
+    _assert_body_has_no_invalid_tool_call(httpx_mock.get_request())
+
+
+@pytest.mark.asyncio
+async def test_astream_sanitizes_request_body(httpx_mock):
+    httpx_mock.add_response(url="http://test.local/v1/chat/completions", text=_stream_response())
+
+    async for _ in _adapter().astream(_messages_with_invalid()):
+        pass
+
+    _assert_body_has_no_invalid_tool_call(httpx_mock.get_request())

docs/develop-guides/decisions/implemented/2026-09-10-model-input-invalid-tool-call-sanitize.md

@@ -0,0 +1,45 @@
+# 模型输入净化：invalid_tool_call 的截断 function 在发送边界降级为失败反馈
+
+状态：implemented
+类型：bug-fix
+Owner：backend/package/yuxi/models/chat.py
+
+## 问题
+
+模型生成无效工具调用（JSON 参数截断/格式错）时，LangChain 把它记在 `AIMessage.invalid_tool_calls`。`langchain_openai` 序列化时把它转成 `type:"function"`、`arguments` 为截断 JSON 的 `tool_call`，DeepSeek 等接口因参数不合法报错并中断整轮。
+
+## 根因澄清（对上一版假设的修正）
+
+初版把现象归为「`invalid_tool_calls` 序列化成 `invalid_tool_call` 变体」，并为此清空解析字段、收敛 `additional_kwargs`、删孤儿 `ToolMessage`、动态包装四个入口。实测 `langchain_openai` 1.6.0 的真实序列化推翻了该假设：
+
+- `AIMessage.invalid_tool_calls` 经 `_lc_invalid_tool_call_to_openai_tool_call` 转成 `type:"function"`，**不是** `invalid_tool_call` 变体；真正的问题是 `arguments` 截断导致参数解析失败。
+- content 数组里的 `{"type":"invalid_tool_call"}` block 在 `_convert_from_v1_to_chat_completions` 里已被丢弃，不会泄漏到 wire。
+- 孤儿 `ToolMessage` 的根因在历史裁剪/恢复/消息组装，不是发送边界能安全修复的。
+
+## 决策
+
+收窄为**只在 Chat Completions 发送边界**处理「截断 function」：`ChatCompletionsAdapter._get_request_payload` 里用原始消息（`invalid_tool_calls` 的 id）对账 wire 载荷，按 id 移除这些截断 `tool_calls`，并在 content 里追加 text 反馈「`[工具调用失败] name: error`」——给模型明确的失败反馈，而不是发出畸形参数。
+
+- 只影响 `ChatCompletionsAdapter`（OpenAI 兼容协议）；Anthropic/Gemini 有自己的消息格式，不处理。
+- 不改消息对象（原始 checkpoint 不动），净化只落在序列化后的 wire 载荷。
+- 同步/异步/流式/非流式四条路径都经 `_get_request_payload`，一处覆盖。
+- 不删孤儿 `ToolMessage`：id 存在不代表调用与响应顺序合法，按根因另行处理。
+
+## 替代方案
+
+- 动态 `_InvalidToolCallFilterMixin` 包装四个入口：覆盖所有供应商、范围过大，且掩盖消息链路本身的问题；拒绝。
+- 在 `_convert_message_to_dict` 打补丁：第三方内部实现，升级即失效；拒绝。
+- 全历史 id 集合清洗工具响应：会把「响应在前、调用在后」等非法序列保留，或误删实际执行结果；拒绝。
+
+## 后果
+
+发送给 DeepSeek 等接口的载荷不再含参数截断的 `function` 调用，改为明确的文本失败反馈。`invalid_tool_calls` 属性仍在 checkpoint 里原样保留（可观测、可追溯），仅在 wire 边界降级。非 OpenAI 兼容供应商不受影响。
+
+## 验证
+
+`backend/test/unit/models/test_chat_invalid_tool_call_sanitize.py`（真实 adapter + mock HTTP，断言最终请求体）：
+
+- 纯无效调用：`tool_calls` 移除，content 追加失败反馈；
+- 有效/无效混合：wire 只保留 `call-good`，追加失败反馈；
+- 无无效调用：wire 原样不变；
+- 四条路径（`invoke`/`ainvoke`/`stream`/`astream`）都净化。