dsh-usage-insights

Working inside DeepSeek Harness leaves a lot of key questions unanswered out of the box: how many Agent turns actually completed, how many tool calls fired, whether tokens went into uncached input or cache reads, and whether caching is pulling its weight at all. Those signals are buried in raw session events, and log-parsing by hand is both slow and error-prone. dsh-usage-insights aggregates that scattered data into a local, read-only work-activity dashboard. Slice by today, the last 7 days, the last 30 days, or the full history, then cross-filter along workspace, provider, and model to inspect uncached input, cache reads, cache writes, output, and reasoning tokens side by side, while tracking metered requests, Agent Usage coverage, model and tool latency, TTFT coverage, cache reuse rate, and daily tool-failure trends. Five paginated breakdowns (workspace, provider, model, session, tool) all support search, sorting, and CSV export so the numbers feed straight into downstream spreadsheets.

The plugin stays deliberately restrained. It never displays monetary amounts, never calls a model or provider API, and never reads prompts, replies, tool parameters, or tool output. Every metric is derived locally from DSH session events: per-session atomic writes, watermark-based deduplication so no event is double-counted, a backfill buffer that merges real-time and replayed events by sequence number, and session-level isolation so a single unreadable record cannot block the rest. When the timezone or aggregation algorithm version changes, the old derived cache is discarded entirely and rebuilt from raw events rather than blended with a different statistical baseline. Removing a source session only cleans up the plugin's own derived records; original DSH conversation data is never modified or deleted.

If you are running Agent workflows across multiple workspaces, providers, and models in DeepSeek Harness and need to audit usage periodically, hunt down latency hotspots, or report cost structure to a team, this dashboard replaces the manual log-parsing ritual with a single page refresh. It will not make decisions for you, but it puts the numbers those decisions depend on clearly on the table.