dsh-profile-stats

A personal usage profile page for the DeepSeek Harness Web UI (dsh). It adds one entry to the left sidebar — Profile / 个人主页 — with the numbers people actually ask about:

• Cumulative tokens with an input / output / cache-hit split
• Peak day (plus peak step and peak session)
• Longest task (longest turn/start → turn/end span)
• Longest streak / current streak of active days
• Token activity heatmap for the last 12 months (GitHub-contribution style, quartile buckets, hover shows the day's tokens, the exact number and the day's request count)
• Provider breakdown (today / all time) — which billing route each token went through

Everything is folded from this machine's own session logs. No network calls, no telemetry, no credentials, no upload.

What each number means

Card	Definition
Cumulative tokens	Per assistant/message step: max(totalTokens, input + output) summed over all sessions. totalTokens from the provider already equals input + output + cache, so cache is counted once.
Peak day	Highest single-day total. Peak step / peak session are recorded too and are available in the snapshot.
Longest task	Longest span between a turn/start and its turn/end (turns interrupted without turn/end are not counted).
Streaks	Active day = a local calendar day with at least one token. Current streak ends today (or yesterday, if today has no activity yet).
Token mix	Input / output / cache-hit totals over the selected scope.
Provider breakdown	Per-step data.message.source.provider, e.g. wincode, arkcli-coding-plan, myapi, deepseek-account.

Why provider breakdown matters

Gateway consoles often show input + cache-hit + output as a single “total tokens” figure. Since cache-hit is a subset of input, that sum counts the cache twice. On a real machine (measured 2026-10-30) the gateway card said 340.15M for one token while input + output was 173.02M — and this plugin's number for that same provider was 172.21M (0.5% apart, explained by the scan being a few minutes older). Use the provider breakdown to compare one billing route at a time.

Requirements

• DeepSeek Harness with the Web UI (desktop app or dsh web), dsh >= 0.2.1-rc.5 < 0.3.0.
• No other plugin is required. Light/dark themes both supported (theme tokens only).

Install

Three equivalent ways — use whichever your setup allows:

1. From this repository (works today, no npm account or marketplace entry needed):

plugin_manager(action: "install_bundle", target: "github:Lovenndme/dsh-profile-stats")

which is the same as dsh plugin add github:Lovenndme/dsh-profile-stats. pnpm fetches the repo and installs the packed files; there is no build step and no dependency to approve.

2. From npm, once published:

plugin_manager(action: "install_bundle", target: "dsh-profile-stats")

3. From the marketplace: Plugins → Discover, search dsh-profile-stats, install.

After installing, a Profile row appears in the left sidebar (person icon). Click it; the page renders immediately from a cached snapshot and rescans in the background.

Who is shown (name / handle / avatar)

The identity block is read from this machine, not from any account. Precedence:

1. $DSH_HOME/dsh-profile-stats/profile.json — optional, written by you:

   { "name": "Mike", "handle": "@lovenndme", "avatar": "M" }
   

   handle gets a leading @ if you omit it, avatar is optional (defaults to the first letter of name). The file is re-read on every snapshot request, so editing it needs no restart.

2. The operating-system account name (os.userInfo().username), handle @<lowercased name>.

3. Fallback text DSH 用户 / @local when the account name is unavailable.

The snapshot reports identity.source (file or os) and identity.file (the exact path), so you can always tell which rule applied.

Uninstall

Disable or remove the bundle in Plugins. The plugin only writes two files; delete them if you want a clean slate:

$DSH_HOME/dsh-profile-stats/cache.json      # default: ~/.dsh/dsh-profile-stats/cache.json
$DSH_HOME/dsh-profile-stats/profile.json    # optional display-name override

Data flow

Host half (index.js):

• GET /api/dsh-profile-stats/snapshot — current snapshot; scanning: true carries progress
• POST /api/dsh-profile-stats/refresh — request a background rescan
• Reads sessions through ctx.sessionQuery (listSessions, readSession), folds them per session and caches the fold. Only sessions that are still running or never folded get re-read; finished sessions reuse the cache.
• Fork/seed sessions copy their source events into the log header (inheritedEventCount). Those copies are excluded — unless the source session is gone from this machine, in which case the copies are the only record and are counted.

Client half (client.js): registers into sidebar.panellist (id profile) and the main slot (key profile), styles with theme tokens, and routes all visible text through the locale service (en, zh).

Snapshot fields:

totals:  tokens, inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens, calls,
         sessions, activeDays, firstDay, lastDay,
         peakDay {date, tokens}, peakStepTokens, peakSessionTokens, longestTurnMs,
         longestStreak, currentStreak, byProvider {}
daily[]: date, tokens, calls, inputTokens, outputTokens, cacheTokens, byProvider {}

Limitations

• Local machine only: usage from other devices or other tools never enters these logs.
• Numbers move while a session is active (the log has no fixed end); expect sub-1% drift between two scans a few minutes apart.
• A day is bucketed by the host's local timezone; moving machines shifts historical buckets.
• A session whose log cannot be read is counted in failedSessions and skipped, never guessed.

Development

• Host changes need a fresh module generation: toggle include:dsh-profile-stats off and on via plugin_manager, or restart Harness.
• Client changes only need a page refresh.
• index.js exports apply / inject only; client.js registers through window.__ModuleLoader__.

License

MIT © 2026 Lovenndme

---

中文说明

给 DeepSeek Harness Web UI 加一个个人主页面板：左侧图标栏多一个「个人主页」入口，主区显示

• 累计 Token 数（输入 / 输出 / 缓存命中拆分）
• 单日峰值（另记单次调用峰值、单会话峰值）
• 最长任务用时（同一 turn 内 turn/start → turn/end 的最大跨度）
• 最长连续天数 / 当前连续天数（活跃日按本机时区）
• Token 活动热力图（近 12 个月，按当日用量四分位分档；悬浮显示日期、当天用量、精确数字、当天请求数）
• 供应商拆分（今日 / 累计），按每一步的 data.message.source.provider 分组

数据全部来自本机会话日志：不联网、不上报、不碰任何凭据。删掉缓存文件只会让下次重扫，数字不会和日志不一致。

为什么需要「供应商拆分」

很多网关面板的「总 Token」是 输入 + 缓存命中 + 输出 三列相加，而「缓存命中」本身就包含在「输入」里，等于把缓存算了两次。 实测（2026-10-30）某网关对单个令牌显示 3.4 亿，而 输入 + 输出 只有 1.73 亿；本插件对同一 provider 的统计是 1.72 亿（差 0.5%，来自扫描时刻不同）。 所以对账时请按 provider 逐家比较，不要直接和「总 Token」相减。

安装

在 Harness 里打开 插件 → 发现，搜索 dsh-profile-stats 安装；或用 plugin_manager(action: "install_bundle", target: "dsh-profile-stats")。 装好后左侧出现「个人主页」一行，点开即用（先出缓存快照，再后台补扫）。

要求

DeepSeek Harness Web UI，dsh >= 0.2.1-rc.5 < 0.3.0。无其它插件依赖，明暗主题均支持（只用主题 token）。

已知边界

• 只统计本机日志：其它设备、其它工具的用量不在这里。
• 会话进行中数字会动，两次扫描间隔几分钟通常有 1% 内的差异。
• 按主机本地时区分日，换时区会让历史分档平移。
• 日志读不出来的会话会计入 failedSessions 并跳过，不会瞎猜。