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.