本次围绕 Eino 流式与续跑做了三类改动：一是在 multi_agent / eino_single_agent 里合并多段 Run 的 MCP execution id，并在前端合并 SSE 中的 id、补全渗透详情芯片，避免「中断并继续」后工具卡片只剩最后一轮。二是在 eino_adk_run_loop 流式路径把合并后的 tool_calls 写入 runAccumulatedMsgs，避免落库轨迹缺 assistant(tool_calls) 导致 RepairOrphan 删掉全部 tool 结果、模型「失忆」。三是新增「模型入参」快照中间件，在每次调用大模型前（summarization / reduction 等之后）深拷贝 Messages，写 last_react_input 时优先用该快照而非事件流单调追加，使继续对话与中断续跑更对齐上下文压缩后的真实状态；Plan-Execute 仅挂在 Executor，Deep / Supervisor 仅主编排。
Changes fall into three areas: merging MCP execution IDs across multi-segment runs (interrupt-and-continue) on the backend plus frontend ID merging and incremental detail chips so tool badges are not truncated to the last segment; appending streamed merged tool_calls into runAccumulatedMsgs so persisted traces keep assistant↔tool pairing and tool outputs are not stripped by orphan repair; and a new pre-model snapshot middleware that deep-copies state.Messages after summarization/reduction (etc.) and prefers that JSON for last_react_input over the append-only event trace—so follow-up turns and resume align with the compressed model-facing context—with wiring on Eino single, Deep/Supervisor orchestrators, and Plan-Execute’s executor only.