dsh-plugin-oauth-subs

Use the AI subscriptions you already pay for, inside DeepSeek Harness. Sign in once and your ChatGPT / Codex, Grok, GLM, Kiro, Antigravity, Cursor, Ollama Cloud, Kimi, Devin or Cline plan becomes a model you can pick in any session — with quota, models and usage in one workbench.

Subscription workbench — quota tab with per-account cards and provider rail

More pages — 模型 Models · 用量 Usage · 榜单 Leaderboard · 设置 Settings

模型 Models

Models tab — model catalog with per-family enable switches

用量 Usage

Usage tab — token, cache-read, call and estimated-cost panels

榜单 Leaderboard

Leaderboard tab — model ranking by net improvement

设置 Settings

Settings tab — version, repository, auto-update and proxy

Quick start

1 · Install the plugin

dsh plugin --profile web add https://github.com/xxww0098/dsh-plugin-oauth-subs
dsh web

On the desktop app, use 插件 Plugins → 添加插件 and paste the same repo URL instead — the command above is web-only.

2 · Open the workbench

Sidebar → 插件 Plugins → 订阅 Subscriptions.

3 · Add an account

Pick your service and sign in. Most take one click — the browser opens, you approve, you are back. Already signed in on this computer? Choose Import local … and the plugin picks up the login you already have.

4 · Use your models

Open the 模型 Models tab, switch on the families you want, then pick a model from the model menu in a Harness session.

The workbench

Tab	What it does
额度 Quota	One card per account — plan badge, usage and reset bars. Refreshes about once a minute, or tap 刷新额度.
模型 Models	The model list. Switch families on or off, and raise a model's input window.
用量 Usage	Tokens, cache hits, calls and estimated cost — by model or by session (per-model latency reads on each row).
榜单 Leaderboard	Daily model ranking snapshot.
设置 Settings	Updates, proxy, and where the plugin comes from.

Supported services

Service	How you sign in
ChatGPT / Codex	Sign in with ChatGPT, or import your local Codex login
xAI Grok	Device code — approve it in the browser
Zhipu GLM	Z.ai / BigModel login, or import your local ZCode login
AWS Kiro	Social, Builder ID, IdC or Entra sign-in, an `ksk_` key, or a pasted JSON/CSV
Google Antigravity	Google sign-in, or import your local Antigravity token
Cursor	Sign in, or Import local Cursor
Ollama Cloud	Paste an API key
Kimi Code Plan	Device code, or import your local Kimi login
Devin Agent	Sign in, or import your local Devin credentials
Cline	Device code, or import your local Cline login
Command Code	Paste an API key, or import your local login
OpenCode Go	Paste an API key — add your console workspace to see quota

Good to know

Accounts stay on your machine. Everything is stored in your local DSH profile; the only traffic is to the service you signed into.
-fast models cost more quota. They answer sooner and spend your plan faster — reach for them when latency matters.
Reasoning effort is picked in the Harness session menu, not in the Models tab.
Updates install themselves. The 设置 Settings tab can also check and install on demand.
Cost is an estimate, worked out from published prices — not from your bill.

If something looks wrong

What you see	What to try
The tab still shows an old version	设置 Settings → 检查更新 → 安装更新, then restart
A model is missing	Open 模型 Models and check its family is switched on
Quota looks stale	Tap 刷新额度 — it also refreshes on its own about once a minute
The desktop app closes the moment it opens	Run `launchctl unsetenv ELECTRON_RUN_AS_NODE` in Terminal, then start it again

Advanced — configuration, families, internals and diagnostics

Families and routes

Provider	Login / import	DSH api
ChatGPT Codex	PKCE `localhost:1455` (paste-callback ok); import `~/.codex/auth.json`	`openai-responses`
ChatGPT (Sign in with ChatGPT)	Official app flow on the loopback; ID token verified against JWKS	`openai-responses`
xAI Grok	Device-code (default) or PKCE; import `~/.grok/auth.json`, `~/.hermes/auth.json`	`openai-responses`
Zhipu GLM (Z.ai / BigModel)	ZCode CLI login poll; import `~/.zcode/v2/config.json`	`anthropic-messages`
AWS Kiro	Social PKCE / Builder ID / IdC / Entra / `ksk_` key; paste JSON · kami · CSV; import `~/.kiro/credentials.json`	`openai-completions`
Google Antigravity	Google OAuth `localhost:51121` (paste-callback ok); import `~/.gemini/antigravity-cli/` token	`openai-completions`
Cursor	PKCE, or Import local Cursor (macOS Keychain / `state.vscdb`)	`openai-completions`
Ollama Cloud	Paste API key / `OLLAMA_API_KEY`	`openai-completions`
Kimi Code Plan	Device-code; import `~/.kimi-code/credentials/kimi-code.json`; `KIMI_API_KEY`	`openai-completions`
Devin Agent	PKCE `127.0.0.1:59653`; import `~/.local/share/devin/credentials.toml`; paste `devin-session-token$…`	`openai-completions`
Cline	Device-code; import `~/.cline/data/settings/providers.json`	`openai-completions`
Command Code	Paste API key / `COMMAND_CODE_API_KEY` / import `~/.commandcode/auth.json`; or studio browser login	`openai-completions`
OpenCode Go	Paste API key (+ optional Console cookie/workspace for quota); direct, no proxy	`openai-completions` / `openai-responses`

Per-family design, pinned client versions, and upstream hops: docs/oauth.md and each src/oauth/<id>/README.md / src/apikey/<id>/README.md.

How it works

Plane	Role
Subscriptions panel	Login / import / logout, then model sync
llm-pi-ai	DSH call plane; routes families to the loopback proxy (OpenCode Go direct)
Loopback proxy	`http://127.0.0.1:8318/{codex,grok}/v1/responses`, `/glm/v1/messages`, `/<family>/v1/chat/completions`
Upstream	Refreshed subscription bearer or the active API key

The proxy binds only to loopback and requires DSH_OAUTH_SUBS_API_KEY. After the panel closes, DSH keeps using the configured routes.

Models and quota

The Models tab has per-family switches (all on by default); a row's window badge opens a dialog to raise that row's input window up to its ceiling. Reasoning effort is set in the Harness session menu, not the Models tab. -fast variants exist where the vendor supports them (Codex Priority; Devin backend variant). Fast and larger windows spend quota faster.

The Quota tab shows a plan badge plus usage/reset bars for every account (Ollama Cloud: session/weekly usage), refreshed about once a minute or via Refresh quota.

Per-family model catalogs, efforts, windows, quota endpoints, and pricing sources: docs/models.md, docs/oauth.md, and the family READMEs.

Where data lives

Tokens live in <profile>/data/dsh-plugin-oauth-subs/auth.json (0600); OpenCode Go accounts in opencode-go.json; model selections in models.json — all in that data directory.

Options

Option	Default	Notes
`port`	`8318`	Loopback proxy port
`provider`	`oauth`	llm-pi-ai route prefix; every family lands at `oauth-<id>`
`dataDir`	profile data dir	`auth.json`, `models.json`, and `proxy-key`
`grokLogin`	`device`	`device` or `pkce`
`proxyUrl`	settings / env	Outbound HTTP(S) proxy for model / quota / login hops
`cursorProxy`	—	Cursor upstream proxy (`http://` or `socks5://`) for region-gated models

Diagnose

npm run analyze -- path/to/session.jsonl              # one session
npm run analyze -- --dir ~/.dsh/sessions --since 30d  # aggregate every session

Healthy: weighted cache hit ≥ 80%, zero affinity misses. Calls are labeled cold_start / delta / compaction / rebuild / affinity_miss, so compaction is not flagged as a shard regression. Details: CONTRIBUTING.md and docs/error.md.

Contributing

Issues and pull requests are welcome. Build, test and release commands live in CONTRIBUTING.md and docs/development.md.

dsh-plugin-oauth-subs

Installer

README