wenzetan/dsh-quota-panel

English | 中文

dsh-quota-panel is a provider quota / balance status widget for the DeepSeek Harness (DSH) web surface (dsh web). It sits in the bottom-right corner of the product UI, watches every AI provider whose API key you have configured, and tells you at a glance how much balance / quota is left — DeepSeek, OpenRouter, SiliconFlow, Moonshot, StepFun, xAI, Zhipu GLM, OpenCode Go, Volcengine Ark (Agent/Coding Plan), plus one-api / new-api style aggregators, and the coding plans (智谱 GLM Coding, Z.AI, Kimi Coding, MiniMax Coding global/CN) with 5-hour / weekly usage windows and MCP monthly quota. xAI, Zhipu GLM, OpenCode Go, ChatGPT subscription (Plus/Pro via Codex login), plus one-api / new-api style aggregators, and the coding plans (智谱 GLM Coding, Z.AI, Kimi Coding, MiniMax Coding global/CN) with 5-hour / weekly usage windows and MCP monthly quota.

Since v0.5 it is a dual-face plugin with a built-in provider catalog and auto discovery: install it, restart dsh web, and every provider whose key resolves automatically appears on the panel — zero configuration. It needs no npm dependencies and asks for no allowBuilds authorization.

Screenshots (real browser rendering)

Collapsed capsule, light theme:

Expanded card, light theme:

Settings panel (⚙), light theme:

Collapsed capsule, dark theme:

Expanded card, dark theme:

Settings panel (⚙), dark theme:

Supported features

• Auto discovery — the host half ships a catalog of well-known providers; each entry names the provider's standard credential references, and every provider whose key resolves ($DSH_HOME/.credentials.yaml / .env / environment variables) appears on the panel automatically, with zero config. Remove the key and the row disappears. No credential enumeration API exists in DSH, so the catalog is probed each refresh cycle.
• Coding-plan usage windows — GLM / Z.AI / Kimi / MiniMax coding plans render as usage rows: 5-hour window, weekly pool and (GLM/Z.AI) the MCP monthly lane, each with its own reset countdown; windows the plan does not carry show — instead of a fabricated 0%.
• Two sizes — collapsed: a minimal capsule with one independent "status dot + value" pair per account (● ¥58.36 · ● 45%); expanded: a full card with a row per provider (status dot, name, primary value, secondary info, progress bar for usage-kind providers).
• Auto refresh — follows the configured interval (default 60 s), paused while the page is hidden; the refresh button spins during a fetch and repeated clicks never fire concurrent requests.
• Per-account status — balance rows are graded by tier (critical <= warn <= healthy), usage rows by percent (error >= warn); the offending dot/value alone recolors, others stay calm. Usage percentages use battery-style three-color grading, independent of the status dots.
• Settings panel (⚙) — per-provider visibility, refresh interval, per-provider warning thresholds, per-provider HTTP(S) proxy URL, the capsule display mode (auto = highest window — the default — / 5h window / weekly window / highest), and "restore defaults". All local settings apply immediately, persist to browser localStorage, and are never written to the profile or uploaded.
• Per-row HTTP(S) proxy — configure a proxy for providers that cannot be reached directly from your network (see below).
• One-api / new-api aggregators — the built-in openai-billing format adapts aggregator dashboards.
• Theming — driven entirely by Harness design tokens (--dsw-alias-*, --dsw-static-*, --dsw-shadow-*, --dsw-font-*) with sensible fallbacks, so it follows the product theme (light/dark) and ships no palette of its own.
• Security by construction — API keys never reach the browser; the browser talks only to this plugin's methods inside DSH's own authenticated /api channel and receives only normalized views (see "How it works").

Not supported (yet)

• Usage-only providers — OpenAI, Anthropic, Together, Groq, Mistral, Cohere, DashScope, Baichuan expose no public "remaining balance" endpoint, only usage/cost queries (usually admin keys + time windows, "spent" semantics rather than "remaining"). Planned as a separate usage row kind showing monthly spend (Anthropic Admin API and OpenAI usage API first).
• Cookie / CLI-only coding plans — the quota pages of Qwen Token Plan (Bailian console), Xiaomi MiMo Token Plan, and Qoder expose no API-key quota endpoint: they require web cookies, the arkcli CLI, or chat-endpoint rate-limit probes (per CodexBar research). Volcengine Ark / Doubao Agent Plan and Coding Plan are now supported via the AK/SK-signed OpenAPI (see the catalog table above). Other providers on this list cannot be wired in until an API-key endpoint appears.
• socks5 proxies — only HTTP/HTTPS proxies are accepted (a socks URL is rejected with a clear per-row error).
• Custom adapters — new upstream formats cannot be plugged in from the profile; a format value outside the built-in set fails loud at mount.
• Multi-page placements — the widget lives in the shell.overlay slot only (bottom-right corner), not in sidebars, headers, or the status bar.

Requesting a new provider

Missing a provider? Open an issue with:

1. the provider id you want (^[a-z0-9-]+$, e.g. together), using a -cn suffix for the China site of a dual-site provider (cf. siliconflow / siliconflow-cn);
2. the balance API URL — a public endpoint that answers the provider's standard API key with remaining balance/quota (e.g. GET https://api.provider.com/v1/user/info, Bearer auth), plus the response shape if you can paste it.

That is all the catalog needs: an id whose standard credential reference resolves, an endpoint, and a format adapter for the response. Providers with only cookie/CLI quota pages (see above) cannot be supported until they expose an API-key endpoint.

How it works

┌─────────────── browser (lib/client.js) ───────────────┐
│  shell.overlay slot → capsule / card / settings panel  │
│  localStorage: visibility · interval · thresholds ·    │
│                proxy URLs (frontend settings)          │
└──────────────┬─────────────────────────────────────────┘
               │ Connection /api channel (browser-session fenced):
               │   POST /api/dsh-quota-panel/specs (render hints)
               │   POST /api/dsh-quota-panel/fetch-all { proxy: {...} }
┌──────────────▼────────────── host (lib/index.js) ──────┐
│  ctx.credentials → API keys (never leave the host)      │
│  catalog probe → auto discovery (15 built-in providers) │
│  per-row fetch → proxy engine (CONNECT tunnel /         │
│                  absolute-URI) → upstream JSON          │
│  normalization → {balance | usage | info} view models   │
└─────────────────────────────────────────────────────────┘

• Host half (lib/index.js) mounts one exact Fetch route per endpoint on the Connection service's authenticated /api channel, under the dsh-quota-panel method namespace:
  • POST /api/dsh-quota-panel/specs — the resolved rows with render hints only (id, label, row kind, currency, threshold tiers, window labels, configured proxy name). No credentials, no endpoints.
  • POST /api/dsh-quota-panel/fetch-all — fetches every visible row, normalizes each upstream response into a generic view model (balance / usage / info), and returns {rows: [{id, view} | {id, error}], fetchedAt}. Raw upstream JSON stays host-side like the keys; one failing row never affects the others.
  • POST /api/dsh-quota-panel/chatgpt-auth-status / chatgpt-login-start / chatgpt-login-cancel / chatgpt-logout — the optional ChatGPT subscription device-login flow. Requests are fenced by DSH's own /api route (trusted host + browser session), so the endpoints are not reachable from another machine or from a client without the page session.
• Auto discovery — because DSH's credential store has no enumeration API, the host half probes the catalog entries' standard refs each fetch cycle; every entry whose key resolves joins the panel, and entries with unresolvable keys are skipped (a missing key yields a clear per-row error only when it was explicitly configured via providers).
• Proxy engine — zero-dependency hand-rolled proxiedGetJson: https targets go through an HTTP CONNECT tunnel (TLS over the tunnel), http targets via absolute-URI forwarding. 15 s per-row timeout, 1 MB body cap. Proxy selection precedence: frontend settings panel > profile config > direct.
• Threshold judgement happens client-side from the specs hints, so local threshold overrides apply without refetching; profile thresholds ship in specs and the frontend settings override them locally.
• Config validation — the exported Config schema (vendored schemastery) declares structure and defaults; cross-field constraints (id uniqueness, critical <= warn <= healthy, proxy references, catalog override keys) are validated host-side at mount and fail loud.
• DOM safety — the card builds DOM exclusively with createElement/textContent; API values never touch innerHTML; technical errors (401, timeout, missing credential, refused proxy) surface only in title tooltips or inline row text.
• A broken proxy can never take the host down — every socket and request the proxy engine opens funnels its error event into the row's result, so an unreachable proxy (e.g. clash stopped, ECONNREFUSED 127.0.0.1:7890) shows as a per-row error instead of an unhandled EventEmitter error.

Configuration

Out of the box: nothing. Install, restart, and any provider whose key resolves appears automatically. The table below is only for tuning.

All keys are optional — the structure and defaults live in the exported Config schema, so profile patches may omit every defaulted field.

Key	Meaning	Default
auto	probe the built-in catalog; providers with a resolvable key join the panel	true
hide	row ids to drop (catalog and explicit rows alike)	[]
proxies	named proxy definitions {<name>: "http://host:port"}, HTTP(S) only	{}
catalog	partial overrides for auto-discovered rows {<catalog-id>: {...}}	{}
refreshMs	auto-refresh interval	60000
providers	explicit rows; a same-id entry replaces the catalog row wholesale	[]

Each catalog override may set: label / endpoint / format / proxy / refs (credential references to probe, UPPER_SNAKE) / secretRefs (a second credential reference — required for the Volcengine AK/SK pair; the row is only discovered when BOTH refs and secretRefs resolve) / region (Volcengine OpenAPI region; default cn-beijing) / currency (balance rows: symbol like $ or US$) / balanceTiers / warnPercent / errorPercent / windowLabels.

Explicit providers fields:

Field	Meaning	Default
id	row id (RPC rows align by id), ^[a-z0-9-]+$	required
label	provider name shown on the card	required
credential	credential reference ($DSH_HOME/.credentials.yaml or environment)	required
secretCredential	second credential reference (Volcengine volcengine-agent-usage / volcengine-coding-usage: the SK)	—
endpoint	quota JSON endpoint; base URL for openai-billing	required
format	row adapter (see table below)	deepseek-balance
proxy	a proxy name defined in proxies; absent = direct	—
region	(volcengine-agent-usage / volcengine-coding-usage) Volcengine OpenAPI region	cn-beijing
currency	(balance rows) currency symbol, overrides the format default	format default
balanceTiers	(balance rows) {critical, warn, healthy}	{10, 20, 50}
lowBalance	legacy alias for balanceTiers.warn	—
windowLabels	(usage-kind formats) labels for the usage windows	{滚, 周, 月}
warnPercent / errorPercent	(usage rows) thresholds	70 / 90

Built-in provider catalog (auto discovery)

Provider	Credential refs probed	Endpoint	Row kind
DeepSeek	DEEPSEEK_API_KEY	api.deepseek.com/user/balance	¥ balance
OpenRouter	OPENROUTER_API_KEY	openrouter.ai/api/v1/credits	$ balance (purchased − used)
SiliconFlow (global)	SILICONFLOW_API_KEY	api.siliconflow.com/v1/user/info	$ balance
SiliconFlow (CN)	SILICONFLOW_CN_API_KEY	api.siliconflow.cn/v1/user/info	¥ balance
Moonshot / Kimi	MOONSHOT_API_KEY	api.moonshot.cn/v1/users/me/balance	¥ balance
MiniMax Coding (global)	MINIMAX_API_KEY	www.minimax.io/v1/token_plan/remains	5h prompt usage %
MiniMax Coding (CN)	MINIMAX_CN_API_KEY	api.minimaxi.com/v1/token_plan/remains	5h prompt usage %
StepFun	STEP_API_KEY / STEPFUN_API_KEY	api.stepfun.com/v1/accounts	¥ balance (hover: cash/voucher)
xAI	XAI_API_KEY	api.x.ai/v1/billing/credits	$ balance
Zhipu GLM	ZHIPU_API_KEY / GLM_API_KEY	open.bigmodel.cn/api/monitor/usage/quota/limit	text row (quota remaining/total; no public balance API)
智谱 GLM Coding	ZAI_CODING_CN_API_KEY	open.bigmodel.cn/api/monitor/usage/quota/limit	coding-plan windows (5h tokens / weekly / searches)
Z.AI GLM Coding	ZAI_API_KEY	api.z.ai/api/monitor/usage/quota/limit	coding-plan windows (5h tokens / weekly / searches)
Kimi Coding	KIMI_API_KEY	api.kimi.com/coding/v1/usages	usage % (5h rate limit + weekly request pool)
OpenCode Go	OPENCODE_GO_API_KEY	opencode.ai/zen/go/v1/usage	three-window usage %
Volcengine Ark Agent Plan	VOLC_ACCESS_KEY + VOLC_SECRET_KEY	open.volcengineapi.com (OpenAPI, signed)	usage % (5h / weekly / monthly, GetAFPUsage)
Volcengine Ark Coding Plan	VOLC_ACCESS_KEY + VOLC_SECRET_KEY	open.volcengineapi.com (OpenAPI, signed)	usage % (session / weekly / monthly, GetCodingPlanUsage)

Volcengine Ark has two separate subscriptions — Agent Plan and Coding Plan — shown as two independent rows (like two providers) that share the same AK/SK pair. Each row queries only its own plan's API and never falls back to the other, so a plan the account has not subscribed to shows a "not subscribed" message on that row instead of the other plan's numbers. Volcengine authenticates with an AccessKey ID / SecretAccessKey pair using HMAC-SHA256 request signing — not a Bearer token. The inference ARK_API_KEY (shaped ark-...) cannot query usage; only the AK/SK pair has OpenAPI permission. Full setup is in the next section.

Volcengine Ark setup

Ark Agent Plan / Coding Plan usage comes from the control-plane OpenAPI, which requires an AK/SK pair with read-only access. Three steps:

1. Create an AccessKey

Open https://console.volcengine.com/iam/keymanage (Volcengine console → Identity and Access Management → Access Keys) and click New access key. Prefer creating a sub-user key dedicated to this plugin rather than using the primary account key. Save the AccessKey ID and SecretAccessKey when shown — the SecretAccessKey is displayed once.

2. Grant Ark read-only permission

Attach the ArkReadOnlyAccess policy to the sub-user (or role) that owns the key:

• Open IAM → Users, pick the sub-user, choose Permissions → Add permissions;
• In the Search policy name and remarks box, type ArkReadOnlyAccess;
• The result whose service source is "Volcengine Ark" is the one you need — check it. There are only two same-named policies in the search results, and checking both is harmless too.

ArkReadOnlyAccess alone is sufficient — GetAFPUsage and GetCodingPlanUsage are read-only actions; ArkFullAccess or account-level billing permissions are not required.

3. Save the credentials

Add two lines to $DSH_HOME/.credentials.yaml (by default C:\Users\<you>\.dsh\.credentials.yaml on Windows or ~/.dsh/.credentials.yaml on Linux/macOS):

VOLC_ACCESS_KEY: AKLTxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
VOLC_SECRET_KEY: xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

You can also use the VOLC_ACCESS_KEY / VOLC_SECRET_KEY environment variables (DSH's credential resolver falls back to the environment). Restart dsh web; the "Volcengine Agent" and "Volcengine Coding" rows appear in the bottom-right panel automatically (whichever plan the account subscribes to shows data; both rows share the same AK/SK pair) — no providers: block required.

Verifying permissions

After restart, check the panel:

• The Agent row shows three percentages (5h / weekly / monthly) and the Coding row shows three (session / weekly / monthly) → AK/SK and ArkReadOnlyAccess are both working (if only one plan is subscribed, the other row shows a "not subscribed" message);
• volcengine SignatureDoesNotMatch: ... → the SK was copied wrong (watch the trailing =);
• volcengine AccessDenied: ... → the policy is not attached or the wrong source policy was selected;
• No active Volcengine Ark Agent/Coding Plan subscription → the signature worked but the account has no subscription to that plan (typical for pay-as-you-go accounts; the two rows report independently — hide the unsubscribed one from the ⚙ settings panel).

Migration (from ≤ 0.9.1, if you pinned the old row): the single catalog row id volcengine and the format id volcengine-usage were replaced by volcengine-agent (volcengine-agent-usage) and volcengine-coding (volcengine-coding-usage). Auto-discovered setups need no changes; a hand-written catalog: override or providers: entry that still references the old ids makes the plugin refuse to load with a validation error listing the valid ids — update it to the two new ids.

Security note: an AK/SK pair can read all Ark usage data for the account. Redact it before pasting into chats, tickets, or screenshots, and rotate it from the key management page when no longer needed. | ChatGPT subscription (Plus/Pro) | in-plugin login or ~/.codex/auth.json (no API key) | chatgpt.com/backend-api/wham/usage | weekly usage % (Pro includes a 5h window) |

An additional openai-billing format adapts one-api / new-api style aggregators: set endpoint to the aggregator base URL and the host half requests {base}/v1/dashboard/billing/subscription (hard_limit_usd) plus {base}/v1/dashboard/billing/usage (total_usage); remaining = limit − used ($). Aggregator domains differ per deployment, so this format is explicit-config only.

Dual-site provider ids (custom id → site mapping)

Some providers run separate international and China sites with different endpoints, credential references and currencies. The catalog models each site as its own provider id, so configuring the matching key is all it takes — and an explicit providers: entry reusing one of these ids replaces the catalog row wholesale (same fields, your endpoint/label/currency):

provider id	Site	Endpoint	Credential ref	Currency
siliconflow	SiliconFlow global	api.siliconflow.com/v1/user/info	SILICONFLOW_API_KEY	$
siliconflow-cn	SiliconFlow China	api.siliconflow.cn/v1/user/info	SILICONFLOW_CN_API_KEY	¥
minimax	MiniMax Coding global	www.minimax.io/v1/token_plan/remains	MINIMAX_API_KEY	— (usage %)
minimax-cn	MiniMax Coding China	api.minimaxi.com/v1/token_plan/remains	MINIMAX_CN_API_KEY	— (usage %)
zai	Z.AI GLM Coding global	api.z.ai/api/monitor/usage/quota/limit	ZAI_API_KEY	— (usage %)
zai-coding-cn	智谱 GLM Coding China	open.bigmodel.cn/api/monitor/usage/quota/limit	ZAI_CODING_CN_API_KEY	— (usage %)

Both sites of one provider can be on the panel at the same time (configure both keys); hide: ["siliconflow"] drops either row individually.

The currency symbol for balance-kind rows comes from the format by default (siliconflow-balance renders ¥) and can be overridden per row: catalog rows carry currency (the global SiliconFlow row sets $), a catalog: override may set it, and explicit providers: entries accept a currency field (e.g. "US$").

ChatGPT subscription (Plus/Pro)

ChatGPT subscription usage is not API billing — there is no public balance/usage API. This plugin calls the same internal usage endpoint Codex uses, with a ChatGPT OAuth token, and shows the weekly window (and the 5-hour window on Pro) as a used-percentage row. Two login methods are supported, pick whichever you prefer:

⚠️ Experimental. The endpoint (chatgpt.com/backend-api/wham/usage) is an undocumented internal API used by the Codex CLI; its response shape may change. The plugin only performs read-only queries.

Option A: in-plugin login (recommended — no Codex CLI install required)

1. Restart dsh web and open the panel settings (gear icon) in the bottom-right corner;
2. In the "ChatGPT account" section at the top, click "Log in to ChatGPT";
3. The plugin starts an OAuth device-code flow and displays a one-time code plus the sign-in URL https://auth.openai.com/codex/device;
4. Open the URL in your browser, sign in with your ChatGPT Plus/Pro account, and enter the code;
5. Once authorized, a "ChatGPT" row appears automatically (no restart). Hover to see plan: plus/pro and weekly: N% (Pro also shows a 5h window).

Tokens are stored in $DSH_HOME/dsh-quota-panel/chatgpt-auth.json (C:\Users\<you>\.dsh\dsh-quota-panel\ on Windows, file mode 0600). When the access token expires the plugin refreshes it with the refresh token and writes it back. Click "Log out" in the same section to delete the local token.

Option B: reuse a Codex CLI login

If you have already logged in via the Codex CLI (ran codex and completed the browser sign-in), the plugin automatically reads ~/.codex/auth.json (or $CODEX_HOME/auth.json) — no extra configuration needed. With this method, refreshed tokens stay in the plugin process memory only and are never written back to auth.json (that file is owned by the Codex CLI).

When both are present, the in-plugin login takes precedence. When neither exists the ChatGPT row is hidden. If the token is invalidated by a login elsewhere, the row shows an error — run Option A again, or codex login, to recover.

Built-in formats

format	Row kind	Upstream response shape
deepseek-balance	¥ balance	{ balance_infos: [{ currency, total_balance, granted_balance, topped_up_balance }] }
openrouter-credits	$ balance	{ data: { total_credits, total_usage } }
siliconflow-balance	balance (¥ by default, per-row currency override)	{ data: { balance, chargeBalance, totalUsage } }
moonshot-balance	¥ balance	{ data: { total_balance } }
minimax-remains	usage %	{ base_resp, model_remains: [{ model_name, current_interval_total_count, current_interval_usage_count, current_interval_remaining_percent, end_time, current_weekly_total_count, current_weekly_usage_count, weekly_end_time }] } — coding model row (MiniMax-M*) preferred; counts are remaining-side (used = total − count); weekly window only when current_weekly_total_count > 0
stepfun-accounts	¥ balance	{ balance, total_cash_balance, total_voucher_balance }
xai-credits	$ balance	{ total: { val } } (cents → dollars)
openai-billing	$ balance	aggregator dashboard/billing endpoints
zhipu-quota	text	{ code: 200, data: { limits: [{ remaining, number }] } } (limits without remaining fall back to percentage)
opencode-usage	usage %	`{ usage: { rolling
zai-coding-quota	usage %	{ code: 200, data: { limits: [{ type: TOKENS_LIMIT | TIME_LIMIT | CREDIT_LIMIT, unit, number, percentage, currentValue, usage, remaining, nextResetTime }] } } — semantic mapping (glm-plan-usage2, issue #2): TOKENS_LIMIT unit=3 → 5h window, unit=6 → weekly, TIME_LIMIT → MCP monthly lane; unknown units fall back to nextResetTime ordering; every window prefers the percentage field. Credit-package plans (issue #7) answer with CREDIT_LIMIT rows that carry the same unit/number declaration, so both row kinds share one unit match (unit=3 → the 5h credit window, e.g. 2000 credits; unit=6 → the weekly pool, e.g. 10000 credits) and either kind may fill a lane the other left empty; only rows that declare no unit fall back to nextResetTime ordering. Credit lanes are tagged in the hover title as [CREDIT_LIMIT u3n5 left remaining/usage], or no unit when the declaration is absent
kimi-coding-usage	usage %	{ usage: { limit, used, remaining, resetTime }, limits: [{ window: { duration, timeUnit }, detail: { limit, used, remaining, resetTime } }] } — 5h = the duration=300 window, weekly = duration=10080 (fallback: top-level usage); used = limit − remaining
volcengine-agent-usage	usage %	Dispatched inline in fetchRow (not through adaptRow): signs and calls the Volcengine Ark OpenAPI GetAFPUsage, parsing Result.AFPFiveHour / AFPWeekly / AFPMonthly (Agent Plan 5h/weekly/monthly; AFPDaily skipped per the console).
volcengine-coding-usage	usage %	Dispatched inline in fetchRow (not through adaptRow): signs and calls the Volcengine Ark OpenAPI GetCodingPlanUsage, parsing Result.QuotaUsage[].Level ∈ {session,weekly,monthly} (Coding Plan session/weekly/monthly, percentages only). Independent from the Agent row — no fallback between them.
chatgpt-subscription	usage %	{ plan_type, rate_limit: { primary_window: { used_percent, reset_at, limit_window_seconds }, secondary_window? } } — read via the OAuth token in ~/.codex/auth.json against Codex's internal usage endpoint; windows are classified by limit_window_seconds (18000s ≈ 5h → rolling, 604800s ≈ 7d → weekly), falling back to the typical positional layout (primary = 5h session, secondary = weekly pool). Experimental

Proxy (providers that cannot be reached directly)

Configure per-provider proxies in the frontend settings panel (⚙ → 代理): fill an HTTP(S) proxy URL (e.g. http://127.0.0.1:7890, user:pass allowed), saved to browser localStorage, effective immediately — leave it empty to fall back to the profile config or a direct connection. Requests still run host-side: the browser sends each row's proxy URL in the fetch-all payload, the host validates it (http/https only, socks rejected) and fetches through it — keys never reach the browser, but the proxy itself can observe them (see Known issues & risks under Security).

The profile proxies map + row-level proxy / catalog.<id>.proxy remain available as default proxies (used when the frontend field is empty). Precedence: frontend settings > profile config > direct.

# example profile-level default proxy (the ⚙ panel can override per row)
- id: quota-panel
  name: 'dsh-quota-panel'
  config:
    proxies:
      home: http://127.0.0.1:7890     # local proxy http port (clash / v2rayN …)
    catalog:
      openrouter:
        proxy: home                    # OpenRouter via proxy by default
    providers:
      - id: my-agg
        label: My aggregator
        credential: AGG_API_KEY
        endpoint: https://agg.example  # base URL for openai-billing
        format: openai-billing
        proxy: home

Threshold defaults

DeepSeek balance (balanceTiers {critical: 10, warn: 20, healthy: 50}):

Balance	Status	Secondary info
<= 10	error (red dot + red value)	建议充值
10 < x <= 20	warn (amber)	余额紧张
20 < x <= 50	ok	余额正常
> 50	ok	余额充足

OpenCode usage (high = max(rolling, weekly, monthly)):

Usage	Status
< warnPercent	ok (green dot, DeepSeek-blue bar)
>= warnPercent	warn (amber dot + bar)
>= errorPercent	error (red dot + bar)

Compatibility

One DSH host line at a time. This package supports exactly the host line its peerDependencies pin — currently @deepseek-ai/dsh@0.1.7-rc.1 (npm next). The five seam packages it actually talks to are declared as exact peers:

Seam package (peer, exact)	Used for
@deepseek-ai/dsh-client-connection	host connection.fetch RPC routes + browser connection.rpc
@deepseek-ai/dsh-credentials	host-side credentials.resolve
@deepseek-ai/dsh-client-ui-renderer	browser slots service (shell.overlay registration)
@deepseek-ai/dsh-cordis-client-runner	browser timer service (ctx.interval)
@deepseek-ai/dsh-client-locale	browser locale service (dictionaries)

Because those peers are exact versions, a DSH 0.1.7-or-newer host's plugin/host compatibility gate (evaluatePluginCompatibility) refuses to install or boot this package on any other host line. That is intentional: old host lines are served by old plugin versions. Keep running the release that pinned your DSH version — its tag records which one — and upgrade the plugin together with DSH. (Host lines older than 0.1.7 have no such gate and simply ignore the declaration; they are not tested or supported.)

Verification for the current line lives in the testbed and the CI boot gate; docs/2026-09-24-dsh-0.1.7-rc.1-assessment.md records the seam-by-seam diff against the previous line.

Versioning and tags

The version string carries the DSH host line plus a local revision,

0.1.7-rc.1-v0.1
└────┬────┘ └┬┘
DSH line      └─ local revision (this repo's counter)
(peer pin)

• package.json#version is <dsh-line>-v<local> (for example 0.1.7-rc.1-v0.1).
• The git tag is v + that version: v0.1.7-rc.1-v0.1.
• The npm channel mirrors DSH's own dist-tags. The release job reads npm view @deepseek-ai/dsh dist-tags and publishes the plugin under the tag that currently names the declared host line: a release for dsh latest (today 0.1.5-rc.3) goes to npm latest; a release for dsh next (today 0.1.7-rc.1) goes to npm next. A line no dist-tag points at falls back to next. GitHub releases follow the same split: latest → normal release gated by the production environment's reviewer, everything else → pre-release.
• Releases for a maintenance line are cut from that line's branch (release/0.1.5-rc.3) by pushing the tag by hand or dispatching the CI workflow with release: true; main only ever auto-tags its own line.
• A human can override the mapping for a release they have confirmed: dispatch the CI workflow with promote_tag (e.g. v0.1.7-rc.1-v0.1) to move npm latest and the GitHub Latest flag to that release, even when its host line is still dsh next.

Because the host line is pinned in peerDependencies, one DSH line gets one plugin line: old host lines keep their old plugin release, and the plugin is upgraded together with DSH.

Install

Install a released version — not the main branch. main receives unverified work-in-progress; only tagged releases have passed the CI gates (check + boot) and — for the shipped line — the human approval gate.

Recommended — the release that matches your DSH line:

# dsh npm `latest` (0.1.5-rc.3 today) → plugin release tagged
# v0.1.5-rc.3-v0.1, published as npm `latest`:
dsh plugin --profile web add dsh-quota-panel

# dsh npm `next` (0.1.7-rc.1 today) → plugin release tagged
# v0.1.7-rc.1-v0.1, published as npm `next`:
dsh plugin --profile web add dsh-quota-panel@next

# Pin the tag instead (checked on the Releases page):
dsh plugin --profile web add "github:wenzetan/dsh-quota-panel#v0.1.5-rc.3-v0.1"

# Restart `dsh web` (bundle layer and client module graph apply at boot)

Check npm view @deepseek-ai/dsh dist-tags to see which tag your DSH line rides, then install the matching plugin channel — the version rule is in Versioning and tags.

Avoid bare github:wenzetan/dsh-quota-panel (no #tag) — it tracks main HEAD, which is the testing branch: it may carry unreleased work, fail CI, or break. Only developers iterating on the plugin itself should install from main.

Refresh the browser page once after installing. Zero npm dependencies (the schema library — schemastery + cosmokit, both MIT — is vendored under src/vendor/ with relative-path imports), no allowBuilds authorization needed.

Release channels & npm publishing (maintainer)

The release identity and channel rules live in Versioning and tags. In short: package.json#version is <dsh-line>-v<local>, the tag is v<that version>, and the publish channel mirrors @deepseek-ai/dsh's own dist-tags — latest for the shipped line (GitHub release, production environment reviewer), that line's dist-tag for everything else (GitHub pre-release).

package.json version	Host line it pins	Channel	Gate	GitHub Release	npm dist-tag
0.1.7-rc.1-v0.1	dsh next	that line's tag	CI only (check + boot)	flagged pre-release	next
0.1.5-rc.3-v0.1	dsh latest	that line's tag	CI + human approval	normal release	latest

Workflow:

1. Release the current line (automatic) — bump package.json#version (e.g. 0.1.7-rc.1-v0.1) and push main. CI runs the full gates, auto-tags v0.1.7-rc.1-v0.1 and publishes it under the dist-tag that @deepseek-ai/dsh itself currently carries for 0.1.7-rc.1 (next today). The classify job re-derives that mapping at publish time, so nothing is hardcoded.
2. Release another host line — check out that line's branch (release/0.1.5-rc.3), bump the version there (0.1.5-rc.3-v0.1), and either push the tag by hand or run the CI workflow with the release input on that branch. The release job publishes under the tag that dsh itself carries for that line — latest for 0.1.5-rc.3 — and the release-latest job waits in the production environment for a human approval before creating the GitHub Release and publishing.
3. Promote (optional, human) — once you have confirmed the newer line works for real, dispatch the CI workflow with promote_tag set to its tag (e.g. v0.1.7-rc.1-v0.1). The promote job moves npm latest and the GitHub Latest flag there, and demotes every other normal release to a pre-release — so exactly one release (the current latest) is a normal release. It runs behind the production environment.
4. Verify (human) — install the pinned tag (dsh plugin --profile web add "github:wenzetan/dsh-quota-panel#v0.1.5-rc.3-v0.1", or dsh-quota-panel@latest / @next from npm) and test it for real.

One-time setup:

• npm token — create an Automation (or granular) token with publish rights to dsh-quota-panel (name free as of this writing) and add it as the repository secret NPM_TOKEN (Settings → Secrets and variables → Actions). Without it, GitHub Releases still ship; npm steps are skipped.
• stable gate — Settings → Environments → New environment → production → Required reviewers → add yourself. This is what makes "no stable release without human confirmation" enforced rather than conventional. (Without the reviewer configured, the stable channel publishes without pausing — same as before.)

The package declares dsh.bundle.patch (host half auto-activates as a profile layer) and the dsh.client manifest (browser half auto-joins the __DSH_BOOT__ module graph, immediately: true prefetched with the shell).

Acknowledgments

This plugin builds on community work — thanks to:

• yingjunnan/dsh-deepseek-quota — the original bottom-right DeepSeek balance card for the DSH Web page (auto-refresh + manual refresh); the capsule/card interaction model is directly inspired by it.
• Ghost011118/dsh-balance-meter — DeepSeek account balance and session cost readout for the DSH Web GUI; its panel design informed the expanded card layout.
• 0xsline/awesome-deepseek-harness — the community plugin catalog that surfaced the projects above and the broader DSH plugin ecosystem.
• hanmumuHL/check_balance — endpoint research (DeepSeek balance API) that informed the catalog.
• steipete/CodexBar — its provider docs (z.ai/GLM coding-plan windows, Kimi Code usage API, MiMo / Qwen / Qoder / Doubao auth research) shaped the coding-plan adapters and the not-supported list.
• zwen64657/glm-plan-usage2 — Rust GLM usage tracker whose monitor-API research (docs/api-research.md real-world samples) pinned the semantic window mapping: TOKENS_LIMIT unit=3 → 5h, unit=6 → weekly, TIME_LIMIT → MCP monthly, percentage as the authoritative field; its Kimi (window.duration 300/10080, limit − remaining) and MiniMax (coding-model row, weekly lane) clients informed the matching adapters here (issue #2).
• PowerUserZ/OpenTokenUsage — documented the MiniMax token_plan/remains response quirks and the Kimi Code usage endpoint.
• schemastery and cosmokit (both MIT) — the schema library vendored under src/vendor/.

Changelog

• 0.1.7-rc.1-v0.1 (and its maintenance twin 0.1.5-rc.3-v0.1) — Adopts the DSH-line version scheme: package.json#version is <dsh-line>-v<local>, the git tag is v<that version>, and the npm channel mirrors @deepseek-ai/dsh's own dist-tags (0.1.7-rc.1 → next, 0.1.5-rc.3 → latest). The five seam packages are declared as exact-version peers, so a DSH 0.1.7+ host's plugin compatibility gate only loads the release built for its own host line (old lines keep their old plugin release). Client bundle URLs in the boot graph are document-relative since 0.1.7 — the testbed probes accept both forms. Every release before this entry was unpublished from GitHub and deprecated on npm.
• v0.9.2-rc.4 — Fixes the swapped GLM Coding Plan 5h/weekly lanes on credit packages (issue #7): CREDIT_LIMIT rows now share the TOKENS_LIMIT window declaration (unit=3 → the 5h credit window, unit=6 → the weekly pool) instead of being placed by nextResetTime order alone — the weekly pool resetting before the 5h window inverted the two lanes every time. Either row kind can now also fill a lane its sibling left empty (a lone TOKENS_LIMIT row used to discard the CREDIT_LIMIT row and lose the 5h lane entirely). Rows that declare no unit keep the reset-order fallback, and the hover title spells out the declaration and quota ([CREDIT_LIMIT u3n5 left 1900/2000], no unit when absent). Existing V1 (5h-only) and V2 (5h/weekly/MCP) shapes are unchanged.
• v0.9.2-rc.3 — Fixes a host crash in the proxy engine: the CONNECT tunnel request had no error listener, so an unreachable proxy (ECONNREFUSED 127.0.0.1:7890 — proxy not running) surfaced as Node's Emitted 'error' event on ClientRequest instance and killed dsh web. Every socket and request now funnels error into the row result, so a dead proxy degrades to a per-row message.
• v0.9.2-rc.2 — Restores the host half on @deepseek-ai/dsh@0.1.5-rc.1 and newer. connection.rpc.handle() no longer works for third-party plugins there: its route disposer reads owner.webServer, and that owner resolves to the Connection plugin's own fiber, which never holds webServer — the registration throws inside a child fiber, so boot stays silent while the channel disappears. The host half now mounts one exact Fetch route per endpoint through connection.fetch.register() (which needs only owner.effect), under DSH's own authenticated /api channel: POST /api/dsh-quota-panel/<endpoint>. This supersedes rc.1, which remains merged but must not be promoted because its RPC endpoints are missing on current DSH.
• v0.9.2-rc.1 — Volcengine Ark Agent Plan and Coding Plan are now shown as two rows at once (previously a single row queried Agent Plan first and fell back to Coding Plan). The two subscriptions render as two independent providers sharing the same AK/SK pair: the Agent row (volcengine-agent, GetAFPUsage, 5h/weekly/monthly) and the Coding row (volcengine-coding, GetCodingPlanUsage, session/weekly/monthly) each query only their own API with no cross-fallback; a plan the account has not subscribed to reports its own "not subscribed" message on that row. The Coding row's hover title labels its rolling window session: (it is a session limit, not a 5h window), and a migration note for the replaced row/format ids (volcengine / volcengine-usage) was added to the Volcengine troubleshooting section.
• v0.8.1-rc.6 — layout fixes from issue #1, reworked: the panel is now draggable — grab the collapsed capsule or the expanded card header and move it anywhere (pointer capture, 5px move threshold so click-to-expand still works, viewport-clamped so it stays grabbable, position persisted to localStorage with the other settings, restored positions re-clamped on resize, "restore defaults" clears it). The default anchor stays right/bottom 18px until the first drag. Also injects [class*="overlayLayer"]{z-index:1150 !important} so the shell overlay layer (z-index 20) is no longer covered whole-layer by body-mounted third-party fixed panels (z-index 1000+) — the widget stays inside the React tree, keeping event delegation intact. Both style tags are removed on plugin unload. (rc.5 was briefly auto-released with a fixed 60px bottom offset instead of dragging; its tag/release were rolled back — the npm version is an orphan pre-release.)
• v0.8.1-rc.4 — capsule display mode (issue #2 follow-up): the settings panel gains 胶囊显示 / "capsule display" (auto = highest window — the default, unchanged behavior / 5h window / weekly window / highest). In rolling/weekly modes the collapsed capsule's value, status dot, progress bar and 100% caption all follow the chosen window instead of the highest one, so a 5h capsule no longer glows warn because the weekly pool sits at 40%; a plan without the chosen window falls back to the highest. The expanded card always shows every window.
• v0.8.1-rc.3 — coding-plan adapter fixes (issue #2, cross-checked with glm-plan-usage2): zai-coding-quota maps windows semantically (TOKENS_LIMIT unit=3 → 5h, unit=6 → weekly, TIME_LIMIT → the MCP monthly lane; unknown units fall back to nextResetTime ordering) instead of the size heuristic that swapped 5h/weekly on plans returning both rows, prefers the percentage field for every window, and relabels the third slot 搜索 → 月; kimi-coding-usage matches windows by window.duration (300 = 5h, 10080 = weekly) instead of blind limits[0] and computes used as limit − remaining (the old code read a nonexistent detail.used, so the 5h window silently dropped); minimax-remains prefers the MiniMax-M* coding model row over whatever comes first and adds the weekly window (current_weekly_total_count > 0, remaining-side counts).
• v0.8.1-rc.1 — first pre-release on the automatic rc pipeline: 100% usage caption appends the reset time (当前已使用 100% 等待重置 …); CI reworked (reference dsh-llm-newapi): pre-releases auto-tag + publish to npm next with a latest reclaim guard; stable versions require the manual rc_tag promote workflow.
• v0.8.0 — first stable release on the dual-channel pipeline: same code as v0.7.3 (which already passed check + boot) plus the install guidance rework (released tags / npm latest + next instead of bare main).
• v0.7.3 — no unconfigured provider rows: cordis.patch.yml no longer ships explicit example rows (deepseek / opencode-go), so the settings panel lists exactly the providers whose credential resolves (auto discovery). The CI boot gate now asserts both directions: the seeded key appears, and unconfigured providers do not.
• v0.7.2 — web i18n: the panel follows the shell's language setting (通用设置 → 语言, locale.preference; zh / en) through the ctx.locale service — all copy (capsule, card, settings panel, errors, aria labels, usage windows) ships as zh/en dictionaries registered under the quota-panel namespace; provider labels are proper nouns kept as-is (GLM, MiniMax, Kimi Coding…) with Chinese brand names romanized (智谱 → ZhiPu); host catalog labels normalized accordingly (SiliconFlow CN, MiniMax Coding CN, ZhiPu GLM). Also: usage reset times now show absolute 24h timestamps (下次重置 2026-08-15 14:00, dictionary key nextReset); usage rows drop the weekly segment when the plan has no weekly window and render the search/MCP lane as -% when it is unknown (no fabricated 0%); the usage caption reads 当前已使用 X%.
• v0.7.1 — dual-site SiliconFlow: catalog id siliconflow now maps to the global site (api.siliconflow.com, $), new id siliconflow-cn maps to the China site (api.siliconflow.cn, ¥, ref SILICONFLOW_CN_API_KEY); balance rows gained a per-row currency override (catalog rows, catalog: overrides, and explicit providers: entries); README documents the dual-site provider id → endpoint/currency mapping.
• v0.7.0 — adopted the org TypeScript tool-bundle template (dsh-plugin-check compliant, zero waivers): sources moved to src/*.ts compiled into lib/ by npm run build (tsc + vendored runtime copy), committed artifacts verified current by CI; new dsh-plugin-check CI gate (any error or warning fails — currently verdict=pass, 0 error / 0 warning); CI check job now installs dev dependencies and builds before testing.
• v0.6.0 — coding-plan support: new catalog rows 智谱 GLM Coding (ZAI_CODING_CN_API_KEY), Z.AI GLM Coding (ZAI_API_KEY), Kimi Coding (KIMI_API_KEY), MiniMax Coding global/CN (MINIMAX_API_KEY / MINIMAX_CN_API_KEY); new zai-coding-quota (5h/weekly token windows + search lane) and kimi-coding-usage (5h + weekly request pool) adapters; minimax-remains rewritten for the real model_remains response (now a usage row); zhipu-quota shows percentage when a limit carries no remaining; usage rows render missing windows as — (labels from windowLabels, no longer hardcoded rolling/weekly/ monthly).
• v0.5.0 — built-in provider catalog + auto discovery (probes credential refs; 9 providers on board with zero config); 8 new format adapters (incl. one-api/new-api openai-billing); fetch-all contract switched to host-side normalized views (balance / usage / info), upstream JSON no longer shipped; per-row HTTP(S) proxy (CONNECT tunnel / absolute URI, zero-dependency), configured in the ⚙ settings panel (localStorage, takes precedence over profile proxies / proxy); new auto / hide / proxies / catalog config keys.
• v0.4.0 — dual-face refactor: host half moved to a loopback Connection RPC channel (specs / fetch-all) + Config schema; browser half moved into the dsh.client manifest + shell.overlay slot (React); added the ⚙ settings panel (visibility / refresh interval / thresholds, localStorage-persisted).
• v0.3.0 — two sizes: collapsed minimal capsule (independent status dot + battery-style value per account), click to expand the full card.
• v0.2.0 — Harness-native card: design tokens, balance tier thresholds, usage progress bar.
• v0.1.0 — initial floating panel: server-side quota proxy + page badge.

Security

• API keys are resolved host-side via ctx.credentials and used only fo

…