dsh-plugin-token-usage

You have been running DeepSeek Harness for a while. Every session you start writes token counts, request numbers, and cache-hit ratios straight into the local session logs under ~/.dsh/sessions/. The data is right there, yet there has never been a visual entry point to answer where your tokens actually went and which model burns the most. dsh-plugin-token-usage turns that quiet log stream into a GitHub-style 30-day usage heat map with a five-level colour scale, Sunday-first columns, month labels, and instant hover tooltips showing daily tokens, request count, and cache-hit percentage. Toggle between day, week, and month views with one click; the colour scale renormalises to match whichever granularity you pick.

Beneath the map sits a granularity-linked per-model table that aggregates exact in-window figures through dual session-by-day and model-by-day indexes — no approximations, no sampling. Models that logged requests but produced zero actual tokens are filtered out automatically so your totals stay honest. A sidebar hover badge refreshes the daily token count on a silent 30-second poll and rolls over at midnight without prompting. Or skip the panel entirely: ask your agent to call the usage_report tool and get the same numbers mid-conversation.

Everything is read-only parsing of local logs with streamed zstd decompression at eight-way concurrency. No network requests, no disk writes, no telemetry — your usage data never leaves the machine. It is a good fit for solo developers doing periodic cost self-audits, team leads who want a one-glance usage trend before a stand-up, and any Harness user who wants to turn how much spent into exactly where it went.