dsh-xray

Every plugin you mount quietly bills every single LLM request—prompt sections, tool schemas, skill catalog lines, tokens stacking up silently—yet no existing tool can tell you who registered what, what it costs per request, or whether disabling it is safe. dsh-xray is that missing X-ray: a tab sitting beside Chat and Trajectory inside a running DeepSeek Harness that itemizes the context bill down to the exact text of every entry, turning a number you once trusted by default into a fact you checked in three clicks.

Its capabilities span the full static-to-runtime spectrum. A per-plugin rollup ranks each contributor's request tax; a dedicated skill view prices every skill twice (resident catalog line versus per-load body); a per-request ledger tags system and tools byte-stability (KV-cache-friendly versus prefix-broken) with a delta-previous growth column, while separating compaction calls from chat. Static tools—layer attribution, declared-versus-actual diff, conflict detection, composition snapshot with drift alerting, service dependency graph, and a heuristic capability audit—work even when the harness cannot boot. All surfaces (native GUI tab, standalone /xray page, CLI) carry the same data, and the mounted agent tool lets an LLM answer what it has, who provides a given service, and why a feature is unavailable, about its own composition.

It is built for DSH users managing more than a handful of plugins who want to know exactly who is taxing their context; maintainers wiring CI guards around composition drift and declared-versus-runtime mismatches (every CI-relevant command exits 1 on detection); and any team where token budgets matter enough that feels-more-expensive must become this-line-costs-312-tokens, registered-by-foo, safe-to-disable. dsh-xray only reads: it never executes plugin code, never persists session text, and its audit is a pattern scan over source, not a verdict. It hands you a clear diagnostic image, not a surgery.