deepseek-harness-wallet-patched

DeepSeek Harness Control Center

DeepSeek Harness monitoring, alerts, recharge, and session control center.

Balance ¥5.89 · Session ¥0.72 · Official 18.8M | Third-party 800K · ↗ Recharge

English · 简体中文 · Install · Compatibility · Changelog

A local-first companion that keeps account status, per-conversation usage, completion reminders, official recharge, flexible layout, and host-gated session controls beside the DSH composer.

Version: v0.3.13.

If DeepSeek Harness Control Center helps you, please consider leaving a ⭐ Star. Thank you!

About this repository A personal patched fork of feibi-mochi/deepseek-harness-control-center. Upstream has not been updated since 2026-09-09, so a working version is kept here.

• My time is limited: I do not guarantee timely attention to this repository, and it may stop being updated at any time
• Issues are welcome, though replies may be slow
• If upstream resumes maintenance, I will prefer to return to it

What it does

The UI follows the host language setting, with built-in Simplified Chinese and English wording. DeepSeek balances come from the official API; cost accounting is a local estimate.

余额 ¥5.89 · 本场 ¥0.72 · 官 18.8M | 三方 800K · ↗充

• Official DeepSeek — live balance (60s global refresh with fast boot retries), an estimated current-session cost (not an official bill) locked to the price active for each usage event, including the 2026-08-17 peak/off-peak rollout, and token breakdown.
• Vision model accounting — deepseek-v4-flash-vision-exp is priced like V4 Flash; image tokens reported by the Harness are included with text tokens.
• v4 peak/off-peak ring clock — a resident 24-hour sidebar footer widget for v4-flash, v4-pro, and v4-flash-vision-exp. Weekday peak windows are 09:00–12:00 and 14:00–18:00 Beijing time. After Friday 18:00 the card previews “weekend all-day off-peak”; Saturday and Sunday name the current all-day off-peak rule; Monday before 09:00 shows the time remaining to enter peak. Optional notifications treat Friday 18:00 through Monday 09:00 as one continuous off-peak period.
• Official pricing sync — periodically checks the official DeepSeek pricing page and applies only a fully validated table. Network failures retain the last validated rule (or the built-in rule before the first successful sync); an unrecognized page structure is marked for review instead of silently changing billing.
• Z.ai Coding Plan quotas — a generic official-plan adapter monitors configured Global and China plans without exposing credentials. It separates the 5-hour model-token window from monthly MCP-tool usage, leads with quota remaining from 100% down while keeping usage as secondary context, retains the last successful snapshot on failure, and never converts subscription quota into CNY balance.
• Provider-aware composer surfaces — the chip and sidebar clock follow the session's selected provider/model. Z.ai—including transparent vision-toolkit- variants—replaces DeepSeek balance, recharge, and peak pricing with plan-window summaries; unrelated providers show only their own session tokens, and DeepSeek V4 restores the wallet and peak clock.
• Local usage ledger (up to 365 days / 20,000 events) — Wallet settings keeps the heatmap visible, while compact wallet panels keep it collapsible. Stable request identities are deduplicated, official cost is locked at usage time, official and third-party data stay separate, and prompts or responses are never stored. Collection begins after upgrading to v0.3.2; older aggregate counters have no trustworthy dates and are not backfilled.
• Third-party total — current-session tokens (input / cache read / output) remain available with zero configuration.
• Custom third-party pricing — enter an exact Provider/model route, currency, and base per-million input/cache-read/cache-write/output rates, then optionally add multiple IANA-timezone, weekday, and cross-midnight pricing windows. Current-session and 365-day ledger costs are recalculated locally by occurrence time and clearly labeled as user-defined estimates, never as provider balances or invoices.
• Provider classification — observed wrapper routes appear in the settings page; opted-in routes join the official token/cost bucket for subsequent calls and are priced with the official table. Existing history is not retroactively reclassified.
• Click the chip to open the detail panel: correctly formatted per-currency balances, cost and token splits, a freely editable low-balance threshold for the active account and currency (two decimals, persisted per account; alerts never mix currencies), manual refresh, and a jump to the official recharge page (first click shows the domain for confirmation — anti-phishing).
• Move, dock, and scale — drag the chip freely, preview nearby snap targets, use compact horizontal or vertical layouts, adjust its scale from the control panel, and show official or third-party data independently. The peak clock background can be explicitly set to transparent (solid on hover) or solid; there is no automatic mode. The choices are remembered locally.
• Composer-label controls and skin compatibility — independently show/hide the composer label or reduce it to the primary remaining value (DeepSeek balance, Z.ai five-hour quota) without disabling reminders, settings, plan monitoring, or history. The standard chip also resists broad skin button rules and aligns with maid-atelier's 38px navy-and-gold composer toolbar.
• Floating window mode — detach the detail panel into a draggable window with a remembered position, or minimize it directly to a freely movable dot; the dot turns red below the threshold.
• Completion reminders — optionally notify when a conversation finishes, with persistent or timed modes, queueing and deduplication for simultaneous completions, cross-tab coordination, and an in-page fallback when system notifications are unavailable.
• Low-balance alert — below the threshold the chip turns red with a breathing animation and fires one desktop notification; it resets automatically once the balance recovers.
• Theme-native UI — uses DSH --dsw-alias-* variables with safe fallback colors, so light and dark themes both render correctly; the panel closes when you click outside and flips open-direction near screen edges.
• Clear current-session wallet data — one button clears only the open conversation's token/cost records; it does not delete the conversation, and every other conversation is untouched. Historical ledger clearing is a separate action.

Multi-account

• Open the wallet panel → Account Management（账户管理） to add accounts (name + API key), switch the active one, or remove them.
• Adding the first account attempts to synchronize its key with the host credential store. It activates only if that write succeeds; a refused write retains the saved account without marking it active for billing.
• Switching prompts a confirmation because it changes LLM billing for subsequent requests: the switch writes the account key into the credentials seam (credentials.set('DEEPSEEK_API_KEY', ...)), and since the llm-deepseek provider route resolves that reference per request, the very next LLM call is billed with the new account — no restart needed.
• Account keys are encrypted at rest in $DSH_HOME/storages/accounts.json: Windows uses the current user's DPAPI; other platforms use an owner-only AES-GCM key file. An encrypted .bak recovers a missing, corrupt, or undecryptable primary file; if neither copy can be read, writes are locked instead of overwriting account data. The UI only shows masked keys.
• Session usage estimates follow the active account's currency: USD-settled accounts show 本约 $x, converted from the CNY price table using a fixed estimate of 7.25 CNY per USD (neither a live exchange rate nor the official USD tariff); CNY accounts show 本场 ¥x. These are local estimates, not an official invoice.
• If DEEPSEEK_API_KEY is supplied by the launching environment, switching is refused with a clear error (the credentials provider rejects shadowed writes) — unset it in your shell to enable switching.

Project overview

DeepSeek Harness Wallet is for users who want to check balances during a conversation, compare model costs, or manage multiple accounts. A chip beside the composer shows the selected model’s balance, remaining plan quota, or token usage; open it for details or detach it into a floating panel. Settings brings together encrypted accounts, low-balance and completion reminders, a usage heatmap, and fixed or time-of-use prices for third-party APIs. DeepSeek balances and Z.ai quotas come from their respective APIs. Costs are estimated locally from host-reported usage, helping you track consumption without replacing provider invoices.

Install

From npm:

dsh plugin --profile web add deepseek-harness-wallet-patched

or from GitHub main:

dsh plugin --profile web add github:ThinkofRain1213/deepseek-harness-wallet-patched

Restart dsh web, then hard-refresh the page.

Quick use

1. Click the wallet or peak/off-peak card to open its control panel; open the Harness settings card for health and compatibility checks.
2. The peak card supports horizontal/vertical layout and 100%–120% scaling. The wallet chip uses a separate scale: 100%–105% in the composer and up to 125% when docked or floating.
3. Turn the official recharge button off when you need a smaller card; official and third-party rows can also be shown independently.
4. Drag the card to any open area. If it is hard to find after a layout change, use Reset/Dock（归位/停靠） in the panel to return it to the sidebar.
5. The card follows the host light/dark theme. A hard refresh after upgrading makes sure the new client bundle is loaded.

Update

dsh plugin --profile web update deepseek-harness-wallet-patched

Remove

dsh plugin --profile web remove deepseek-harness-wallet-patched

Browser, desktop, and OS compatibility

See 0.3.13 compatibility evidence. The official 0.1.5-alpha.1 host does not provide this plugin’s permanent-delete capability, so that switch stays disabled; the old source integration patch is not a drop-in update.

The client contains no operating-system-specific feature branch; it checks the Web and host capabilities it needs. That makes the same code portable, but portable code is not the same as real-device verification:

The last row describes intended compatibility, not a claim that every browser/OS/wrapper combination was physically tested. If system notifications are unavailable or denied, reminders fall back to an in-page notice; if Web Locks are unavailable, a renewable local-storage lease coordinates reminder ownership across tabs. CSS zoom also has a transform fallback. Core wallet data, controls, dragging, docking, scaling, and visibility settings use these shared paths rather than an OS name check.

Electron, Tauri, and other DSH desktop wrappers can run the wallet when they expose the normal DSH Web plugin loader, slots, wallet HTTP endpoints, DOM, and fetch. A wrapper that restricts native notifications, persistent storage, or external links may define one optional adapter before the plugin bundle loads:

window.__DSH_WALLET_ADAPTER__ = {
  // All fields are optional. Keep storage synchronous and localStorage-compatible.
  storage: { getItem, setItem, removeItem },
  notify({ title, body, tag, requireInteraction, onClick, onClose }) {
    // May return a notification-like handle, Promise, or nothing.
    // Call the supplied onClick/onClose callbacks for native events.
  },
  requestNotificationPermission() { return 'granted' },
  openExternal(url) { return true },
  capabilities: { permanentDelete: true },
}

notify() may return a notification-like handle, a Promise for one, or nothing for fire-and-forget native APIs. The payload also includes onClick / onClose callbacks so Electron IPC, Tauri notification actions, and other desktop bridges can return events without copying wallet logic; returning false asks the wallet to use its browser fallback. requestNotificationPermission() is optional for hosts such as Tauri and macOS that require a native permission request. Returning false from openExternal() likewise asks the wallet to try the browser fallback. Declare permanentDelete only when the host actually implements the wallet preference and session-menu action; compatible hosts advertise it automatically, while unsupported hosts show a disabled control instead of a switch that has no effect. Platform adaptations are intentionally confined to createCompatibilityAdapter() in src/client/core.js, so an Agent can add a new wrapper without editing wallet accounting or UI logic.

For buildable DSH hosts, the repository includes a versioned Agent-assisted permanent-delete integration kit with a Chinese guide, complete Agent prompt, read-only preflight, compatibility manifest, upstream notice, and an exact-baseline reference patch. The patch is not a universal installer: a different DSH commit must be inspected and adapted by semantics, and closed or non-rebuildable desktop applications remain unsupported.

Optional host integration

Permanent session deletion is implemented by the host, not the wallet. Its switch remains disabled on the official 0.1.5-alpha.1 host. For a source integration, see the guide and Agent adaptation prompt; the old reference patch is not a drop-in update for newer hosts.

Data & trust

Client development now uses five readable files under src/client/ and a committed lib/client.js artifact; run npm ci, npm run build:client, and npm run check:client after source edits. Installation runs no build scripts. Exact-version smoke checks are documented in 0.3.10 compatibility evidence.

Pricing timeline

CNY per 1M tokens, curated from official announcements (cache writes are not billed):

• Since 2025-02-09 — deepseek-chat 2/8 (cache read 0.5), deepseek-reasoner 4/16 (cache read 1)
• Since 2026-04-24 — v4-flash 1/2 (cache read 0.02), v4-pro 3/6 (cache read 0.025)
• Since 2026-08-17 00:00 Beijing — peak/off-peak pricing for the v4 models (peak windows Beijing 09:00–12:00 / 14:00–18:00; off-peak is half the peak rate):
  • v4-flash (off-peak / peak): cache read 0.05 / 0.10, input 1.5 / 3, output 4.5 / 9
  • v4-pro (off-peak / peak): cache read 0.15 / 0.30, input 4.5 / 9, output 13.5 / 27
• Since 2026-08-21 — v4-flash-vision-exp launched with the V4 Flash peak/off-peak table: cache read 0.05 / 0.10, input 1.5 / 3, output 4.5 / 9.
• Since 2026-08-23 00:00 Beijing — Saturday and Sunday are no longer split into peak/off-peak windows; weekend calls use the off-peak rates all day. Weekday peak windows remain 09:00–12:00 and 14:00–18:00.

Historical deepseek-chat and deepseek-reasoner records retain their original flat-rate table; this is not a claim that those legacy model names remain currently available. Each usage event is priced when it arrives; upgrading from 0.1.2 migrates legacy counters once using the then-current rate. Costs are estimates; the API-returned balance is authoritative.

Roadmap

• 365-day Token heatmap and rebuildable local usage ledger
• Z.ai Coding Plan Global/China monitoring on a generic official-plan adapter contract
• User-defined third-party Provider/model base and time-of-use pricing
• Additional provider price/balance adapters only after real-account validation

License

MIT