s08: compact 工具在替换历史后追加了孤立的 tool_result——下一次 API 请求报 400 并导致进程崩溃

[lys849]

您好，我在阅读s08的代码时发现488行处如果模型想要主动调用compact_history时results.append({"type": "tool_result", "tool_use_id": block.id, ...})仍然引用了已经删除的block.id，导致可能的崩溃。

具体复现步骤如下：

1. 运行 python s08_context_compact/code.py
2. 让模型调用 compact 工具（如"请调用 compact 工具压缩上下文"）
3. 观察到 > compact、transcript 落盘后崩溃
4. 实际报错信息有：anthropic.BadRequestError: Error code: 400 - {'error': {'message':
   'unexpected messages.1.content.0: tool_use_id found in tool_result blocks: call_00_8u7tPGlQ6bLAxnweyCk24860.
   Each tool_result block must have a corresponding tool_use block in the previous message.', ...}}

我注意到s20相应的地方已经修改，推测已经注意到这个地方的错误，只不过s8没有进行对应的修改？

[1678zrj]

我也注意到这个问题，下面是我用langchain改写的版本，不知道可以不

rounds_since_todo = 0
def agent_loop(messages: list[BaseMessage]) -> list[BaseMessage]:
    global rounds_since_todo
    reactive_retries = 0
    while True:
        if rounds_since_todo >= 3:
            messages.append(
                HumanMessage(content="<reminder>Update your todos.</reminder>")
            )
            rounds_since_todo = 0
        messages = tool_result_budget(messages)
        messages = snip_compact(messages)
        messages = micro_compact(messages)
        if estimate_size(messages) > CONTEXT_LIMIT:
            print("[auto compact]")
            messages = compact_history(messages)
        try:
            response: AIMessage = model.invoke([SystemMessage(content=SYSTEM)]+messages)
            reactive_retries = 0
        except Exception as e:
            if ("prompt_too_long" in str(e).lower() or "too many tokens" in str(e).lower()) and reactive_retries < 3:
                print("[reactive compact]")
                messages = reactive_compact(messages)
                reactive_retries += 1
                continue
            raise
        rounds_since_todo += 1
        messages.append(response)
        tool_calls: list[ToolCall] = response.tool_calls
        if not tool_calls:
            force = trigger_hooks(HookEvent.Stop, messages)
            if force:
                messages.append(
                    HumanMessage(content=force)
                )
                continue
            break
        has_compact = any(tool_call["name"] == "compact" for tool_call in tool_calls)
        if has_compact:
            messages = compact_history(messages)
            continue
        for tool_call in tool_calls:
            force = trigger_hooks(HookEvent.PreToolUse, tool_call)
            if force:
                messages.append(
                    ToolMessage(
                        tool_call_id=tool_call["id"],
                        content=force,
                        name=tool_call["name"]
                    )
                )
                continue
            tool:BaseTool = tools_handler.get(tool_call["name"])
            if tool_call["name"] == "todo_write":
                rounds_since_todo = 0
            try:
                tool_result = tool.invoke(tool_call["args"]) if tool else f"Unknown: {tool_call['name']}"
            except ValidationError as e:
                tool_result = (
                    f"【参数校验异常】你传入的参数不符合工具 Schema 规范，请检查并修正:\n"
                    f"{str(e)}"
                )
            except Exception as e:
                tool_result = f"工具执行出错: {str(e)}"
            trigger_hooks(HookEvent.PostToolUse, tool_call, tool_result)
            messages.append(
                ToolMessage(
                    tool_call_id=tool_call["id"],
                    content=tool_result,
                    name=tool_call["name"]
                )
            )

    return messages