dsh-subagent-inspector

When orchestrating multi-layer sub-agents in DeepSeek Harness Web, the reasoning paths, tool-call details, and token flow of each node often remain buried in raw logs. dsh-subagent-inspector turns that opacity into a visual panel: a single click on the session header opens a right-side drawer that lays out the full sub-agent execution tree, with per-node status, mode, depth, elapsed time, and token spend visible at a glance.

The drawer drills down from Turn to Step, surfacing the reasoning Harness actually recorded, the reply, Todo items, tool calls with results, errors, and termination reasons. SSE keeps events streaming in real time with automatic reconnection and de-duplication; history is pageable for retroactive review, and a one-click jump takes you straight into the native sub-agent session. Tool payloads are collapsed by default and fully expandable, and all log content is rendered as plain text with no raw HTML injection.

The plugin is strictly read-only. It registers no model tools, does not alter sub-agent scheduling, model selection, permissions, or context, and never speculates about internal reasoning that Harness did not persist. It is well suited for developers running multi-agent workflows in Harness Web who need to pinpoint an anomalous tool call or an unexpected token spike in seconds, as well as engineering leads who want to walk a team through exactly how a complex task is decomposed and executed layer by layer.