dsh-knowledge-graph

DeepSeek Harness treats everything as a plugin, but once your inventory grows to dozens of plugins, skills, and tools, trouble follows: you cannot remember what is installed, silent peerDependencies break without warning, every conversation turn is a black box as to which tools were called and how the agent reasoned, and a pile of dormant items sits unused. dsh-knowledge-graph lays it all out on an interactive graph so you can audit the whole environment and drill into any single turn.

Core capabilities include a force-directed graph where plugins, skills, and tools appear as nodes, depends edges are drawn as lines, and bundle-contains relations are grouped in convex-hull bubbles; a turn selector highlights the tools called in a given turn with sequence badges and call arrows while reasoning and reply render as clickable bubbles that expand to full text; per-node usage statistics show call count, last-used timestamp, and recent turns; health rings flag failed or disabled plugins at a glance; and a one-click Markdown export covers overview, problem plugins, dormant items, and a usage ranking. All stats and history persist to a local file, there is no telemetry, and the plugin runs fully offline.

It is well suited for heavy users who need to audit and inventory dozens of installed packages, for developers debugging broken dependencies or silent failures, for prompt engineers who want to replay how the agent called tools and reasoned turn by turn, and for newcomers exploring the DSH ecosystem who want a quick visual map of available capabilities and tools.