Before you start
Problem summary
在 subagent 执行过程中，如果同一个 graph step 内有多个 sandbox 工具并发懒初始化 sandbox，LangGraph报错
Affected area(s)
Agents / LangGraph (graph, prompts, langgraph.json)
What happened?
问题描述
在 subagent 执行过程中，如果同一个 graph step 内有多个 sandbox 工具并发懒初始化 sandbox，LangGraph 可能报错：
langgraph.errors.InvalidUpdateError: At key 'sandbox': Can receive only one value per step. Use an Annotated key to handle multiple values.
实际日志示例：
Subagent general-purpose async execution failed
langgraph.errors.InvalidUpdateError: At key 'sandbox': Can receive only one value per step. Use an Annotated key to handle multiple values.
Expected behavior
同一步内多个工具写入相同的 sandbox_id 时，应通过 merge_sandbox 幂等合并，不应失败。
如果写入不同 sandbox_id，则继续由 merge_sandbox 抛出冲突错误。
Steps to reproduce
Relevant logs
How are you running DeerFlow?
Docker (make docker-start)
Operating system
Linux
Platform details
No response
Python version
No response
Node.js version
No response
pnpm version
No response
uv version
No response
Git state
No response
Additional context
根因分析
SandboxMiddleware.wrap_tool_call/awrap_tool_call 会在懒初始化 sandbox 后返回：
Command(update={"sandbox": {"sandbox_id": sandbox_id}, "messages": [...]})当同一个 graph step 内多个 sandbox 工具都懒初始化并写入相同的 sandbox_id 时，理论上应该由 merge_sandbox 幂等合并。
但是 SandboxMiddlewareState 自己声明了 sandbox 字段，且没有 Annotated[..., merge_sandbox]。LangGraph 合并 middleware state schema 时，这个局部 schema 可能覆盖 ThreadState.sandbox 上的 reducer，导致 sandbox 退化成 LastValue channel，于是同一步多个写入触发：
Can receive only one value per step
建议修复
让 SandboxMiddlewareState.sandbox 保留和 ThreadState.sandbox 一致的 reducer：
from typing import Annotated, NotRequired, override
from deerflow.agents.thread_state import SandboxState, ThreadDataState, merge_sandbox
class SandboxMiddlewareState(AgentState):
sandbox: Annotated[NotRequired[SandboxState | None], merge_sandbox]
thread_data: NotRequired[ThreadDataState | None]
Before you start
main.Problem summary
在 subagent 执行过程中，如果同一个 graph step 内有多个 sandbox 工具并发懒初始化 sandbox，LangGraph报错
Affected area(s)
Agents / LangGraph (graph, prompts, langgraph.json)
What happened?
问题描述
在 subagent 执行过程中，如果同一个 graph step 内有多个 sandbox 工具并发懒初始化 sandbox，LangGraph 可能报错：
实际日志示例：
Expected behavior
同一步内多个工具写入相同的 sandbox_id 时，应通过 merge_sandbox 幂等合并，不应失败。
如果写入不同 sandbox_id，则继续由 merge_sandbox 抛出冲突错误。
Steps to reproduce
Relevant logs
How are you running DeerFlow?
Docker (make docker-start)
Operating system
Linux
Platform details
No response
Python version
No response
Node.js version
No response
pnpm version
No response
uv version
No response
Git state
No response
Additional context
根因分析
SandboxMiddleware.wrap_tool_call/awrap_tool_call会在懒初始化 sandbox 后返回：当同一个 graph step 内多个 sandbox 工具都懒初始化并写入相同的
sandbox_id时，理论上应该由merge_sandbox幂等合并。但是
SandboxMiddlewareState自己声明了sandbox字段，且没有Annotated[..., merge_sandbox]。LangGraph 合并 middleware state schema 时，这个局部 schema 可能覆盖ThreadState.sandbox上的 reducer，导致sandbox退化成 LastValue channel，于是同一步多个写入触发：建议修复
让
SandboxMiddlewareState.sandbox保留和ThreadState.sandbox一致的 reducer：