返回目錄

dsh-cost-dashboard

精選清單收錄維護狀態: 活躍

mike-lee0120/dsh-cost-dashboard

DSH 成本儀表盤:彙總各會話 Token,按內建價格表計費,含每日趨勢與分模型表格。

前往 GitHub
$ dsh plugin add dsh-cost-dashboard

安裝

dsh 沒有統一的安裝指令 —— 把該外掛 README 中的設定行加入你的 profile / patch 設定,然後重啟即可。

查看安裝教學

5

星數

0

Fork

JavaScript

語言

2026-08-18

建立於

2026-09-25

最近推送

README

dsh-cost-dashboard

English | 中文

A cost-dashboard plugin for DeepSeek Harness (dsh): aggregates model input / output / cache token usage across all local sessions, prices it with a built-in table (including DeepSeek peak/off-peak time-of-day pricing), and renders a dashboard under Settings -> Cost Dashboard.

What you get

  • Two entry points: Settings -> Cost Dashboard (the settings nav icons are hardcoded by the dsh settings shell, so plugins cannot customize them), plus a sidebar footer icon button (data-grid style) that opens the same dashboard in an anchored panel
  • CNY only: every amount is shown in CNY (¥); entries listed in USD (imported catalogs, foreign models) convert at the configurable fx.cnyPerUsd rate (default 6.79)
  • Summary cards: total cost, today's cost, input (cache-miss) / cache-read rate / output tokens, and the session count of sessions you started (subagent logs are folded in, not counted)
  • Daily trend chart: ECharts smooth line charts (gradient area fill and hover tooltips); cost mode is a single CNY series, tokens mode splits into "input / output" and "cache read" charts on independent scales; selectable 1W / 1M / 3M ranges (default 1W)
  • By-model table: tokens, cost, share per model
  • By-session table: one row per session-model pair (a session that used several models appears on several rows, each with its own model, tokens and cost), with title, project directory, subagent badge for a session that spawned delegates. Sortable by cost or last-active time - click the Cost / When header, click again to reverse; the cap is applied per session, so a listed session never loses rows
  • Pricing editor: edit the pricing JSON (including the FX rate) in-page; saves to ~/.dsh/cost-dashboard.json, effective immediately
  • Auto-synced catalog: fills in models missing from builtin/overrides from the LiteLLM price JSON (24h TTL + disk cache, degrades on network failure); never overrides builtin or hand-written prices
  • Actual billing (optional): with read-only provider keys configured, shows DeepSeek/OpenRouter real balances and OpenAI/Anthropic real spend next to the estimate; domestic cloud vendors (Volcengine/Alibaba/Tencent) are not integrated - prices come from the config file
  • Auto refresh: polls every 15s while open; the host re-reads only changed log files (mtime + size validated)

Cost dashboard - trend & summary Cost dashboard - model & session breakdown

Install

dsh plugin --profile web add <spec>

<spec> may be a local path, an npm name, or a GitHub repo:

dsh plugin --profile web add /path/to/dsh-cost-dashboard
dsh plugin --profile web add github:mike-lee0120/dsh-cost-dashboard

dsh plugin add runs pnpm in the profile directory and automatically appends any dsh.bundle-declaring package to dsh.profile.bundles. Restart dsh web and refresh the page, then open Settings -> Cost Dashboard.

Remove with dsh plugin --profile web remove dsh-cost-dashboard.

Requires dsh 0.1.0-rc.7+ and Node >= 22.15 (the node:zlib zstd API the host itself relies on for session logs).

Data source and accounting

  • Read-only scan of the current log generation inside $DSH_HOME/sessions/*/*/ (session.jsonl.zstd, session.vN.jsonl.zstd, or plaintext .jsonl); the numerically highest published generation wins, exactly as the runtime reads it, so a migrated session is never billed from its frozen older generation. Nothing is written, no projection touched.
  • Accounting mirrors the official @deepseek-ai/dsh-token-meter tokenUsage projection:
    • assistant/chunk {type:'usage'} is an early sample that survives a later request failure;
    • assistant/message usage (or the last usage chunk of its data.stream) is the final sample for the same (turn, step) and replaces it instead of double counting, and llm/retry-started closes that slot so a retried attempt adds on top;
    • assistant/attempt settlements are billed even when no message followed;
    • four disjoint buckets: uncached input (DeepSeek prompt_tokens with cache hits subtracted), cache read, cache write, output. Cache write is still folded and still prices cacheWrite rates, but no DeepSeek model produces that bucket, so the dashboard and its tables do not display it.
  • Model attribution: assistant/message carries message.source.provider/model; a bare usage chunk (failed request) is attributed to the latest request/header model.
  • Session attribution: a subagent's log names the session that spawned it in its header (parentSession), so its usage is folded into that session - the session table lists real sessions, each marked with how many subagents it absorbed. A log whose parent is missing keeps its own row and is badged as a subagent.
  • Mid-session model switches are split correctly.
  • Run node scripts/verify-totals.mjs: it reconciles against the official session_projcache.json (a point-in-time snapshot, so newer logs are skipped and said so), folds every selected log with the token meter's semantics, checks that a migrated session resolves to its highest format generation, and pins DeepSeek's official CNY rates plus the Beijing-time weekday peak window (an actively-writing session may drift by a live-write race, which is expected).

Pricing

Built-in pricing for 24 entries - 21 models plus three retired DeepSeek names - per 1M tokens, checked 2026-09-24; "hit" = cache-read rate, "write" = cache-write rate, defaults to the cache-miss input rate when unset. DeepSeek rates are the official CNY list prices (模型 & 价格); every other entry is shown in CNY after conversion:

CNY-listed models

在 GitHub 閱讀完整 README →

DSH Plugins 是獨立的 DeepSeek Harness 外掛目錄,與 DeepSeek 官方無關,也不代表官方背書。第三方外掛未經安全稽核,安裝前請審查原始碼。

每週取得最新的 DeepSeek Harness 外掛,絕不濫發。