dsh-usage-cost

Watching code stream into your editor and then hunting through a billing dashboard to figure out how many tokens that session actually cost — if that small pang of anxiety sounds familiar, dsh-usage-cost was built for you. The moment a conversation starts, a compact pill in the status bar begins ticking in real time: model name, input and output token counts, and running session cost, all at a glance. During streaming the numbers scroll live with a tilde mark flagging estimates; once the response finishes, authoritative usage figures returned by DeepSeek silently replace the estimates for an exact final record.

Under the hood the plugin works in three layers. While tokens stream, a local gpt-tokenizer estimates counts with throttled updates every 50 chunks or 100 ms. When the request completes, DeepSeek-prompt_tokens, completion_tokens, prompt_cache_hit_tokens, and prompt_cache_miss_tokens calibrate the numbers to their exact values. Cost is then computed against a configurable price table (defaults: deepseek-chat 0.5 / 2 / 8 and deepseek-reasoner 1 / 4 / 16 CNY per million tokens), and daily, monthly, and per-model accumulators are persisted locally so they survive restarts. Clicking the pill opens a detail panel with tokens/s, latency, cache-hit ratio, and data status (estimating / calibrated / incomplete).

Everything is computed and stored on your machine; nothing is sent anywhere. The plugin only observes session events and provider usage payloads — it injects no instructions, tools, or request fields visible to the model. If you call DeepSeek frequently, juggle multiple sessions, or simply want to keep an eye on API spend without leaving your editor, this is a drop-in, zero-config companion that starts working the instant your first token streams in.