dsh-insight

A running DeepSeek Harness profile is assembled from layers you never see side by side: each bundle's cordis.patch.yml, your profile-level patch, settings.yaml, .credentials.yaml, and environment variables. Which pieces actually survive is decided by patch semantics (insert, update, disable, keyed by id) and by what the cordis loader managed to start. No single file gives you the full picture, so questions like why a plugin is not running or where a model came from become a hunt across four files and a runtime you cannot inspect.

The insight panel lays out that derivation along six axes. By config shows every layer in application order and what each did to the tree. By plugin gives a complete dossier: origin layer, full entry id, package and path, services provided and consumed, transitive blast radius, and effective settings stacked against defaults. By service maps the real edges between plugins. By tool lists every callable name with its registrant and sibling count. By model traces each row to its provider route, activation method, and exact settings path. By preset reveals the agent-plane composition mounted at session start, live session counts, and origin classification. Every axis labels whether data was observed at runtime or inferred from build output, and ambiguous short ids are left unattributed rather than guessed.

It is built for operators and power users who layer plugins from multiple sources, run cross-provider model routes, author custom presets, and need a single pass to see who depends on whom, who overrode whom, and what breaks if something is turned off. The panel is read-only by design; its sole write path is a double-confirm disable or enable action fenced to the profile patch file, rejects ambiguous ids outright, and never reads credential bodies.