dsh-provider-usage

Every provider's balance, one glance away. A DeepSeek Harness plugin that floats a small draggable usage ball over the Web GUI — no more logging into each provider console to check whether you are about to run out of credit mid-session.

Features

• Auto-detection — enumerates the provider routes registered in the current profile (ctx.llm); zero configuration for well-known routes.

• Per-kind wire adapters — routes with no public balance/usage API (Google, Mistral, Groq, Bedrock, Azure, Qwen Token Plan, …) are listed as unsupported instead of being silently dropped:

  kind	routes	query	shows
  deepseek	deepseek-official, deepseek	GET {baseURL}/user/balance	total / granted / topped-up balance
  moonshot	moonshotai-cn, moonshotai	GET {baseURL}/users/me/balance	available / voucher / cash balance
  kimi-coding	kimi-coding	GET {baseURL}/v1/usages	weekly usage + rate-limit windows, reset countdown
  openrouter	openrouter	GET {origin}/api/v1/credits	credits used / total
  github-copilot	github-copilot	GET api.github.com/copilot_internal/user	plan usage snapshots (paid) or monthly usage (free)
  openai-codex	openai-codex	GET {baseURL}/wham/usage	ChatGPT subscription 5h / weekly windows + credits + spend control (OAuth login, not an API key — see OpenAI Codex via OAuth)
  openai	openai	GET {origin}/v1/organization/costs	current-month spend (admin key required; a regular key fails with 403)
  anthropic	anthropic	GET {baseURL}/v1/organizations/cost_report	current-month spend (admin key required, x-api-key auth)
  minimax	minimax, minimax-cn	GET {origin}/v1/api/openplatform/coding_plan/remains	Coding Plan 5h / weekly remaining %
  zai	zai, zai-coding-cn	GET {origin}/api/monitor/usage/quota/limit	GLM Coding Plan windows (raw key in Authorization, no Bearer)
  opencode	opencode, opencode-go	GET {baseURL}/usage	Zen Go rolling / weekly / monthly windows
  vercel-ai-gateway	vercel-ai-gateway	GET {baseURL}/v1/credits	team credit balance
  xai	xai	GET {baseURL}/billing/credits	prepaid balance (USD)

• Credentials stay safe — API keys are resolved per request through the harness credentials service (environment variables / ~/.dsh/.credentials.yaml); never cached, never written to disk. For OAuth providers (OpenAI Codex) the plugin reads the grant record the sign-in flow stored, and refreshes it transparently when it is about to expire.

• Floating ball widget — a draggable floating ball opens the usage panel. Drop it anywhere in the viewport (position persisted); it docks by default at the bottom-left of the chat area with equal margins, and the panel-header home button sends it back. The panel scrolls when the provider list grows past its height, and its top edge is draggable to resize it taller or shorter (height persisted). The halo around the ball encodes the health of the provider in use — the focused session's own selection (its composer's model seat), tracked live on the client, so switching sessions re-highlights that session's provider immediately without re-selecting a model: green all good, amber when a usage window has under 30% left or a balance below the yellow threshold, red on query failure / missing key / usage ≥90% / balance below the red threshold. Idle providers running low do not color the ball — switch to another provider with enough headroom and the ball turns green again; the panel marks the in-use provider and still lists every provider's numbers.

• Version badge — the panel header shows the running plugin version next to the title, so it is obvious which release is loaded.

• Bilingual panel — built-in Chinese/English UI; follows the harness language by default, with a one-click toggle in the panel header (persisted in localStorage).

• Configurable refresh — adjustable in the panel (15s–30min, persisted in localStorage); the default comes from the plugin config.

• Configurable balance thresholds — balance-type providers (DeepSeek, Moonshot, Vercel AI Gateway, xAI) and the credits rows of usage-type providers (OpenRouter, OpenAI Codex) turn red below the red threshold and yellow below the yellow threshold, compared in the balance's own currency (defaults: red refresh: expires: accountId:

> The grant must land in the harness credential store (the record above). Codex clients that keep their own credential file (e.g. `dsh-codex`'s `$DSH_HOME/.openai-codex-auth.json`, or the Codex CLI's `~/.codex/auth.json`) do not write this record, so this plugin cannot see them.

### 3. Watch the usage

That's it. The route is auto-detected (`ctx.llm` lists `openai-codex`), and the plugin:

1. reads the grant record from the credential store on every poll (no caching),
2. refreshes the OAuth token automatically when it is within 30 s of expiry, deciding **and** rotating inside the store's exclusive lock — so a concurrent process that already rotated the grant is adopted instead of spending the same single-use refresh token twice, and a write that fails is reported rather than swallowed,
3. calls `GET https://chatgpt.com/backend-api/wham/usage` with `Authorization: Bearer ` and the `ChatGPT-Account-Id` header derived from the token's own JWT claim, retrying once after a refresh if the endpoint answers `401`.

The panel then shows your subscription's **5h limit**, **weekly** windows (used %, reset countdown) plus **credits** and **spend control** balances when the plan reports them. If the grant is missing the card reads "OAuth authorization missing (llm-pi-ai/openai-codex)" — sign in again with step 2. If upstream instead *rejects* the stored refresh token (`refresh_token_reused` / `invalid_grant`, i.e. the token was already spent or revoked and the rotation never landed here), the card reads "OAuth authorization expired — sign in to Codex again" with the raw upstream detail on hover, and the plugin stops calling the token endpoint for ten minutes instead of once per poll.

### 4. Troubleshooting: intermittent "Our servers are currently overloaded"

The Codex backend occasionally answers with `Codex error: Our servers are currently overloaded. Please try again later.`, and the harness fails the round with `PI_AI_ERROR`. This is **transient overload on OpenAI's side** (account, quota, and network are usually fine — `GET /wham/usage` confirms window headroom), yet the default configuration never retries, for two reasons:

1. pi-ai's Codex client recognizes `overloaded` as retryable, but its default retry count is 0 (`dsh-llm-pi-ai` explicitly passes `maxRetries: 0`);
2. once the error bubbles up, its text contains no `5xx` / `rate limit` / `timeout` keyword, so it is classified as the catch-all `PI_AI_ERROR` — which is **not** among `dsh-llm-retry`'s default retryable codes (`EMPTY_RESPONSE` / `RATE_LIMIT` / `SERVER` / `TIMEOUT` / `TRANSPORT`), so the round fails immediately.

Add a `retryPolicy` for this provider in `~/.dsh/settings.yaml` that includes `PI_AI_ERROR` in the retryable codes (takes effect on a new session):

```yaml
llm-pi-ai:
  providers:
    openai-codex:
      retryPolicy:
        mode: normal
        maxRetries: 10
        retryableCodes:
          - EMPTY_RESPONSE
          - RATE_LIMIT
          - SERVER
          - TIMEOUT
          - TRANSPORT
          - PI_AI_ERROR
        backoff:
          initialDelayMs: 2000   # first retry delay, doubled each attempt
          maxDelayMs: 60000      # per-attempt delay cap
          jitterRatio: 0.2       # ±20% jitter

Do not confuse this with the other, deterministic failure: some models are unavailable to ChatGPT-subscription accounts, and the backend answers 400 The '' model is not supported when using Codex with a ChatGPT account. (observed with e.g. gpt-5.3-codex-spark, gpt-5-codex). That is a permanent error — retrying won't help; switch to a model your account supports (e.g. the gpt-5.4 / gpt-5.5 / gpt-5.6 families).

Install

[!NOTE] Requires an existing DeepSeek Harness installation.

npm

dsh plugin --profile web add dsh-provider-usage@latest

Build from source

git clone https://github.com/lizhouai/dsh-provider-usage.git
cd dsh-provider-usage
pnpm install
pnpm build
pnpm pack   # produces dsh-provider-usage-.tgz
dsh plugin --profile web add ./dsh-provider-usage-.tgz

Install the tarball, not the repo directory: dsh plugin --profile web add . links the repo, whose own node_modules then shadows the harness's shared @deepseek-ai/cordis instance and the host half never registers (RPC 404). The link form is still handy for client-only UI iteration — the browser bundle is self-contained, so a rebuild + page refresh picks it up — but switch to the tarball (or the npm release) whenever you need the host half. If pnpm fails with EPERM ... symlink while replacing a linked install, delete the stale node_modules/dsh-provider-usage junction in the profile directory and retry.

Restart dsh web after changing the plugin set (a plugin add/remove requires a restart; afterwards, code changes only need a rebuild + re-add + page refresh).

Upgrade

dsh plugin --profile web add dsh-provider-usage@latest

Then restart dsh web and refresh the page. If the release you want was published very recently, your profile's supply-chain cooldown (minimumReleaseAge) may silently keep the older version — pin the exact version instead (dsh plugin --profile web add dsh-provider-usage@0.3.10) and dsh will exempt it automatically. The version badge in the panel header confirms which release is actually loaded.

Configuration

Defaults work out of the box: the plugin auto-detects every provider route of the active profile. A trusted profile can tune behavior in ~/.dsh/profiles/web/cordis.patch.yml — override the bundle's row by id (the package's own bundle patch already inserts it; a second insert of the same id fails the boot with duplicate loader entry id):

- id: provider-usage
  name: dsh-provider-usage
  config:
    refreshSeconds: 60          # suggested panel refresh interval (5–86400)
    balanceRedThreshold: 10     # balance below this turns red (per the balance's own currency)
    balanceYellowThreshold: 30  # balance below this turns yellow (per the balance's own currency)
    autoDetect: true            # enumerate provider routes from the llm registry
    queryTimeoutMs: 20000       # per-attempt query timeout; each retry gets a fresh timeout
    queryRetries: 2             # retries for transient errors (timeout/network/HTTP 408/425/429/5xx)
    queryRetryDelayMs: 2000     # base retry delay, doubled per attempt, capped at 10s
    providers: []               # manual specs; an id matching a detected route overrides it

Field	Type	Default	Description
refreshSeconds	number	60	Suggested widget refresh interval in seconds (5–86400)
balanceRedThreshold	number	10	Balance below this amount turns red, compared in the balance's own currency
balanceYellowThreshold	number	30	Balance below this amount turns yellow, compared in the balance's own currency
autoDetect	boolean	true	Enumerate live provider routes from the llm registry
queryTimeoutMs	number	20000	Per-attempt query timeout (1000–120000); each retry gets a fresh timeout
queryRetries	number	2	Retries for transient errors — timeout / network / HTTP 408, 425, 429, 5xx (0–10); permanent 4xx errors fail immediately
queryRetryDelayMs	number	2000	Base delay between retries (100–60000), doubled per attempt, capped at 10s
providers	array	[]	Manual provider specs: {id, kind, baseURL, apiKeyEnv, displayName?, enabled?}; kind is one of the adapter table above

The same fields can be hot-updated under the provider-usage: namespace in ~/.dsh/settings.yaml.

Adding a manual provider

config:
  providers:
    - id: my-deepseek-gateway
      kind: deepseek
      baseURL: https://my-gateway.example.com
      apiKeyEnv: MY_GATEWAY_KEY
      displayName: My Gateway

How it works

• Host half (src/index.ts): UsageService extends TypertRemoteService exposes usage/list via the Remote('list') marker, applied without decorator syntax (SRC mode, no codegen). Config is declared with schemastery, and the settings provider's installSection enables hot updates from the settings document.
• Client half (src/client/): a window.__ModuleLoader__.load({id, factory}) bundle (built by tsdown) mounts through the sidebar.footer.action slot (used purely as a mount point — the trigger itself is a floating ball portaled to document.body) and polls usage/list through ctx.connection.rpc.call('/api', 'usage/list', {args:{}}) on its own interval. The service stays stateless — every poll fetches live values.

License

MIT

3. Watch the usage

That's it. The route is auto-detected (ctx.llm lists openai-codex), and the plugin:

1. reads the grant record from the credential store on every poll (no caching),
2. refreshes the OAuth token automatically when it is within 30 s of