dsh-knowledge-graph

In DeepSeek Harness every capability is a plugin, so an environment that started small can quickly accumulate hundreds of plugins, skills, and tools. The consequences are familiar: you lose track of what is actually installed, a missing peerDependency silently breaks a feature with no visible clue as to where the chain snapped, each conversation turn is a black box regarding which tools were invoked and how the agent reasoned, and a growing stack of never-used items clutters the workspace. dsh-knowledge-graph puts the entire picture on a single interactive screen and lets you zoom into any individual turn.

The plugin renders a force-directed graph in which plugins, skills, and tools are nodes, dependency edges are lines, and bundle-contains groups are enclosed in convex-hull bubbles. A built-in turn selector highlights the tools a given turn called (with sequence badges and directional arrows) and draws the agent's inner reasoning and final reply as expandable speech bubbles. Every node carries live usage statistics—call count, last-used time, and the most recent turns it appeared in—while colored outer rings flag load failures or disabled plugins at a glance. With one click you can export a Markdown report that covers an overview, problem plugins, a dormant-items list, and a usage ranking. All data is stored locally, no telemetry is sent, and the plugin works entirely offline.

Use it when you need to audit a large plugin environment, debug invisible dependency chains, replay an agent's step-by-step tool calls and reasoning for any conversation turn, remove dead weight before it accumulates, or simply get a visual tour of what the DSH ecosystem can do.