dsh-cachescope

Every LLM call returns a usage field with scattered token counts—Cache Read, Cache Write, Prompt, Output—but those numbers sit orphaned in logs and answer none of the questions that matter: "How much of this call actually hit the prompt cache?" or "Why did the uncached portion grow compared to the previous turn?" CacheScope places provider-reported cache metrics side-by-side with the DSH-side diff between two adjacent logical inputs, giving you a single local panel that makes cache-hit behaviour and input drift visible at a glance.

The desktop dashboard records llm/stream attempts and surfaces a token-weighted Cache Read ratio, an arithmetic reconciliation of how the uncached bucket shifted (Delta-Prompt minus Delta-Cache Read minus Delta-Cache Write), a colour-coded input tree (stable prefix, first local difference, downstream regions, no-baseline regions), plus TTFT, total duration, retries, and filtering and sorting by Session, provider, model, and purpose. It does not modify model requests or register model-visible tools; all data lives in DSH process memory and vanishes when the process stops.

Ideal for developers and prompt engineers who iterate on System Prompts, tool schemas, or message history inside DSH and need a low-cost way to investigate dropping cache-hit rates, sanity-check token costs, or build reproducible evidence for a provider support ticket. No extra deployment required—just start DSH with the WebServer bound to loopback and open the panel.