dsh-session-cost

When you run inference on the DeepSeek API, the scariest part is rarely a crash but the monthly bill that reveals tokens quietly burned through a chunk of budget. dsh-session-cost surfaces that number in real time: a cumulative cost badge sits to the right of the DSH session title bar, and a floating line under every completed reply shows exactly what that turn cost, so every inference call is priced at a glance.

The pricing logic goes beyond a flat tokens-times-rate figure. The plugin splits each model call into cache-hit input, cache-miss input, and output tokens, then automatically applies peak (Beijing 09:00-12:00 and 14:00-18:00, at 2x off-peak) or off-peak unit prices based on request time. Built-in rates are calibrated against actual DeepSeek usage billing as of 2026-08-17, covering DeepSeek-V4-Flash and V4-Pro. A single reply aggregates every model step in the turn, including multi-step reasoning around tool calls, sharing the same turn boundary as the adjacent duration, first-token, and tok/s metrics; when streaming and final usage belong to the same step the final value replaces the stream to avoid double counting. Opening a historical session triggers a full re-fold from the complete session log so old data is still priced correctly.

It is a pure plugin registered through the conversation.chat.assistant-actions slot and touches zero lines of DSH core frontend. Uninstalling it makes both the title-bar total and per-reply badges vanish immediately with no rollback required. It is built for developers who run DeepSeek inference daily through DSH Web or Desktop and want a running cost meter before the invoice arrives.