Back to home

Compare

Comparing: DeepSeek Harness Plugin Adds Multi-Session Control as Agents Shift to Products & DeepSeek Harness 插件补上多会话盲区,Agent 管理正从界面走向产品化

AEN
DeepSeekDeepSeek Harnessdsh-session-matrix·

DeepSeek Harness Plugin Adds Multi-Session Control as Agents Shift to Products

A plugin that does three things: displays four native sessions side by side, maps the relationships between main agents and subagents, and warns users before context compression. We have noticed that the bottleneck in DeepSeek Harness—a development environment for running large-model tasks and agent collaboration—is shifting from model speed to multitask management.

The dsh-session-matrix plugin released by developer Wanqiu12345 targets a specific problem: when running multiple Agents simultaneously, including code modification and research, users have to switch between windows repeatedly, leaving task progress, parent-child relationships, and context risks to be guessed manually.

What this is

The plugin provides a 2×2 four-pane workspace, with each pane preserving a native DSH session, including messages, tool-call records, approval buttons, and the input field. Subagent panes are read-only, and the plugin uses the parentSession field in sessions to draw a “workspace—project—main session—subagent” relationship map, avoiding reliance on project membership alone to infer connections.

It also calculates warnings based on DeepSeek’s official compression algorithm. With a context window of roughly 1 million tokens and 256,000 tokens reserved for output, the compression point is approximately 788,500 tokens, or 75.2% of the window; the plugin alerts users at 80% and 92% of that point. When context compression—replacing earlier conversation with a summary to free up space—occurs, users can have the Agent generate a handoff document first, recording objectives, progress, key files, and unresolved issues.

The plugin also estimates costs in RMB and calculates cache hits separately from standard reads. It is important to note that it does not expand the context window or prevent compression; so far, it has only been tested on Windows 11 and DSH 0.2.2-rc.x.

Industry view

The notable shift is that competition among AI tools is moving from “how good the answers are” to “whether multiple agents can be managed clearly.” Once companies assign research, coding, and execution tasks to different Agents, observability, permission control, and handoff capabilities become more closely aligned with real production requirements than the quality of any single answer. The ability of an interface to display call relationships, approvals, and costs also shows that Agents are taking on a product form similar to a development and operations console.

The risks are equally clear. This plugin mainly addresses visibility; it does not solve erroneous tasks, permission violations, or unreliable results. Costs remain estimates, and peak/off-peak pricing rules do not yet cover statutory holidays. More importantly, it depends on DSH internal fields and a specific version, and its author has tested it only on Windows. Whether it continues to work reliably after upgrades remains to be seen. Making the interface clear does not mean an Agent’s decisions are more reliable.

Impact on regular people

For enterprise IT: Multi-session management, approvals, and cost statistics will bring Agents closer to auditable internal tools, but enterprises still need separate mechanisms for permissions, data boundaries, and result review.

For individual professionals: A workspace like this could make it easier for nontechnical users to advance research, writing, and organizational tasks simultaneously. However, handoff summaries may still omit information, and final outputs still require human confirmation.

For the consumer market: For now, this is more of a developer plugin than a mass-market application. If similar capabilities enter mainstream AI products, users will see not just chat boxes, but collaborative interfaces closer to task boards.

Source: juejin.cn
BZH
DeepSeekDeepSeek Harnessdsh-session-matrix·

DeepSeek Harness 插件补上多会话盲区,Agent 管理正从界面走向产品化

一个插件只做三件事:并排显示四个原生会话、画出主代理与子代理关系、在上下文即将被压缩前发出提醒。我们注意到,DeepSeek Harness(用于运行大模型任务和智能体协作的开发环境)的瓶颈正在从模型速度转向多任务管理。

开发者 Wanqiu12345 发布的 dsh-session-matrix,针对的是一个具体问题:同时运行代码修改、资料调研等多个 Agent(能自主调用工具并推进任务的智能体)时,用户只能反复切换窗口,任务进度、父子关系和上下文风险都依赖人工猜测。

这是什么

插件提供 2×2 四格工作台,每一格保留 DSH 原生会话,包括消息、工具调用记录、审批按钮和输入框。子代理窗格为只读,并通过会话中的 parentSession 字段绘制“工作台—项目—主会话—子代理”关系图,避免仅凭项目归属猜测关联。

它还会按 DeepSeek 官方压缩算法计算预警。以约 100 万 token 的上下文窗口、预留 25.6 万 token 输出为例,压缩点约为 78.85 万 token,即窗口的 75.2%;插件在达到压缩点的 80% 和 92% 时分别提示。上下文压缩(把早期对话替换成摘要以腾出空间)发生时,用户可让 Agent 先生成交接文档,记录目标、进度、关键文件和未决问题。

插件还按人民币估算费用,并将缓存命中与普通读取分开计算。需要说明的是,它不扩大上下文,也不能阻止压缩;目前只在 Windows 11 与 DSH 0.2.2-rc.x 上完成实测。

行业怎么看

值得关注的变化,是 AI 工具的竞争正从“回答得好不好”转向“多个智能体能否被清楚管理”。当企业把调研、编码和执行任务交给不同 Agent 后,可观察性、权限控制与交接能力会比单次回答质量更接近实际生产需求。界面能够显示调用关系、审批和费用,也说明 Agent 正在形成类似开发运维控制台的产品形态。

但风险同样明确。这个插件主要解决可见性,不解决任务错误、权限越界或结果不可靠;费用仍是估算值,峰谷规则暂未覆盖法定节假日。更关键的是,它依赖 DSH 内部字段和版本,作者仅在 Windows 环境验证,升级后能否稳定运行仍需观察。把界面做得清楚,不等于 Agent 的决策更可靠。

对普通人的影响

对企业 IT:多会话、审批和费用统计会让 Agent 更接近可审计的内部工具,但企业仍需另外建立权限、数据边界与结果复核机制。

对个人职场:这类工作台可能让非技术用户更容易同时推进调研、写作和整理任务;不过交接摘要仍可能遗漏信息,最终成果仍需人工确认。

对消费市场:目前它更像开发者插件,而非大众应用。若同类能力进入主流 AI 产品,用户看到的将不再只是聊天框,而是更接近任务看板的协作界面。

Source: juejin.cn