dsh-token-tracker

After a long-context inference run in dsh, the question most developers struggle to answer at a glance is: how many tokens did this conversation actually consume, and what does that cost in CNY? dsh-token-tracker is a dsh web purpose-built for exactly that. It reads provider usage fields from the durable session log, buckets tokens per session and per turn, attributes each message to the model named by the most recent request header, and prices everything against a peak/off-peak rate table denominated in CNY per million tokens, anchored to Beijing local time.

On the UI side, the plugin weaves cost information into several natural touchpoints of the dsh web GUI: a header badge with a peak/off-peak period tag, a one-line summary in the composer dock, a closed-turn tail after each exchange, and a dedicated Token tab inside conversation.view that provides a per-turn breakdown. A standalone overview page and a JSON API are also available for scripting or embedding the data in external dashboards.

By default, Beijing hours 09:00–12:00 and 14:00–18:00 are treated as peak (2× the off-peak rate) for deepseek-v4-flash and deepseek-v4-pro. To apply your own rates, drop a token-pricing.json file into the workspace root or a session working directory; the override is picked up within a few seconds via a short TTL cache.

Who it is for: individual developers running long-context inference in dsh who want an at-a-glance cost readout without leaving the browser; team engineers who need per-turn token exports for cost allocation or audit trails; and anyone who wants to reconcile API bills against actual usage directly inside the dsh web interface.