dsh-plugin-oauth-subs

简体中文 | English

Use a ChatGPT / Codex, xAI Grok, Zhipu GLM, AWS Kiro, Google Antigravity, Cursor, or Ollama Cloud subscription inside DeepSeek Harness. Official OAuth, plus Kiro API keys, Cursor CLI/IDE reuse, and Ollama API keys.

A loopback proxy plus llm-pi-ai route sync. Each family picks one DSH api from the closed union openai-responses | openai-completions | anthropic-messages.

Install

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

Open Settings → OAuth subs. Icon tabs stay pinned at the top: Codex, Grok, Z.ai (GLM), Kiro, Antigravity, Cursor, Ollama, Models, About. Sign in more than once per family; one card per account, each with its own quota. Click a card to switch the chat account. GLM matches ZCode's welcome screen: Z.ai (global) and BigModel (China) OAuth, plus paste-an-API-key. Kiro stacks Social / GitHub / Google, Builder ID, Enterprise IdC, Entra / Azure AD, and ksk_ keys. Antigravity is Google login like the official IDE. Cursor is PKCE poll plus Import local Cursor (CLI Keychain / IDE state.vscdb / CURSOR_ACCESS_TOKEN) — user-owned login reuse, not a second OAuth. Ollama is ollama.com Cloud (paste API key / OLLAMA_API_KEY), not localhost:11434. About links the GitHub repo. Check for updates compares GitHub latest and, when newer, runs dsh plugin --profile web update dsh-plugin-oauth-subs. Restart dsh web to load the new module. Or mount the bundle patch by hand:

- insert:
    - id: oauth-subs
      name: dsh-plugin-oauth-subs

pnpm dsh web --patch ./cordis.patch.yml

Sign-in

Provider	Flow	Client	Upstream
ChatGPT Codex	PKCE on localhost:1455 (falls back to 1457); paste-callback supported	app_EMoamEEZ73f0CkXaXp7hrann	chatgpt.com/backend-api/codex/responses
xAI Grok	Device-code (default); PKCE on 127.0.0.1:56121 as fallback	b1a00492-073a-47ea-816f-4c329264a828	api.x.ai/v1/responses
Zhipu GLM · Z.ai (global)	ZCode CLI poll, provider: zai, then mint id.secret	client_P8X5CMWmlaRO9gyO-KSqtg	api.z.ai/api/anthropic (Completions leftover …/coding/paas/v4)
Zhipu GLM · BigModel (China)	Same CLI poll, provider: bigmodel; poll JWT is the bearer	zcode	open.bigmodel.cn/api/anthropic (Completions leftover …/coding/paas/v4)
AWS Kiro · Social	Portal PKCE at app.kiro.dev; callback ports 3128…53153	(none — portal)	Auth prod.us-east-1.auth.desktop.kiro.dev
AWS Kiro · Builder ID	AWS SSO OIDC device code; registers a public client each login	issued at login	https://view.awsapps.com/start
AWS Kiro · Enterprise / IdC	Same device code against the org Start URL	issued at login	https://oidc.{region}.amazonaws.com
AWS Kiro · Entra / Azure AD	Paste refresh token; public-client refresh_token grant	your Entra client id	*.microsoftonline.com token endpoint
AWS Kiro · API key	Paste ksk_…	—	Bearer, no refresh
Google Antigravity	Google OAuth on localhost:51121/oauth-callback; paste-callback supported	1071006060591-…apps.googleusercontent.com	daily-cloudcode-pa.googleapis.com/v1internal:streamGenerateContent (hub; prod is IDE fallback)
Cursor	PKCE poll at cursor.com/loginDeepControl; or Import local Cursor	Cursor CLI / IDE	Connect agentn.us.api5.cursor.sh AgentService/Run
Ollama Cloud	Paste API key; or Import OLLAMA_API_KEY	ollama.com/settings/keys	https://ollama.com/v1/chat/completions

Already signed in on this machine via Codex CLI, Grok CLI, Hermes, ZCode Desktop, Kiro IDE, kiro.rs, Antigravity CLI, CLIProxyAPI, or Cursor CLI/IDE? Use Import local session (Cursor button: Import local Cursor):

• ~/.codex/auth.json
• ~/.grok/auth.json
• ~/.hermes/auth.json
• ~/.zcode/v2/config.json (ZCode Desktop; also older ~/.zcode/cli/config.json / ~/.zcode/config.json)
• credentials.json (kiro.rs CWD dump)
• ~/.kiro/credentials.json
• ~/.aws/sso/cache/kiro-auth-token.json (IdC pairs the hashed client registration in the same folder)
• Paste in Settings: kiro-manager-lite kami / compact JSON / full backup / CSV, or a Social refresh token / ksk_…
• ~/.gemini/antigravity-cli/antigravity-oauth-token
• ~/.cli-proxy-api/antigravity-*.json
• macOS Keychain cursor-access-token / cursor-refresh-token (Cursor CLI)
• Cursor IDE state.vscdb keys cursorAuth/accessToken + cursorAuth/refreshToken (current OS user only; WSL does not walk other Windows profiles)
• CURSOR_ACCESS_TOKEN env (no refresh)
• OLLAMA_API_KEY env (Ollama Cloud; not ~/.ollama/id_ed25519.pub)

Tokens live at <profile>/data/dsh-plugin-oauth-subs/auth.json with mode 0600. Multiple accounts per family sit in that file as a vault; a legacy single-session file still loads. Enabled-model choices live in models.json next to it.

How it works

Settings (control plane)
  └─ OAuth login / import / logout, then model sync

DeepSeek Harness (call plane)
  └─ llm-pi-ai
       └─ http://127.0.0.1:8318/{codex,grok}/v1/responses
       └─ http://127.0.0.1:8318/glm/v1/messages
       └─ http://127.0.0.1:8318/{kiro,antigravity,cursor,ollama}/v1/chat/completions
            └─ refreshed subscription bearer against upstream

Codex and Grok are Responses (vendor-native). GLM is Anthropic Messages (ZCode Desktop default; Completions leftover stays at /glm/v1/chat/completions until the next sync). Kiro, Antigravity, Cursor, and Ollama stay Completions adapters because their native wires (AWS EventStream / generateContent / Connect protobuf / Ollama /api/chat) are none of the three. Ollama Cloud still hops Completions because https://ollama.com/v1/chat/completions accepts the same Bearer key. The 150% GLM Coding Plan boost is identity (ZCode Desktop UA), not a protocol claim — this plugin has not compared quota slope vs official Desktop.

This is not a second LLM adapter. After you close Settings, DSH still calls the loopback proxy through llm-pi-ai. The proxy binds loopback only and checks the local credential DSH_OAUTH_SUBS_API_KEY.

Stack, module tree, and the docs/error.md rule are in AGENTS.md. Host code is TypeScript under src/oauth and src/utils. Settings is React under src/ui. Do not edit compiled lib/.

src/
  oauth/codex/     Codex catalog, identity, Responses body
  oauth/grok/      Grok catalog, identity, device-code
  oauth/kiro/      Kiro Social / Builder ID / IdC / Entra / API key
  oauth/           proxy, PKCE, quota, models
  oauth/antigravity/ Google OAuth + cloudcode-pa fingerprint
  oauth/cursor/    Cursor PKCE + CLI/IDE import + AgentService/Run hop
  oauth/ollama/    Ollama Cloud API key + Completions passthrough
  ui/              React Settings (classic-script factory)
  utils/           jwt, pkce, fast/context, session analyzer

Reliability

The proxy is the cache-affinity and stream-retry path. Two contracts matter on a long Codex turn:

1. Cache shard (Codex). A Codex prompt_cache_key is forwarded as both session-id and x-client-request-id. Keys are sanitized to [A-Za-z0-9._:-] and clipped to 64 characters instead of dropped — a too-long session id must still pin the shard. Missing or illegal keys fall back to session_id. The clipped key is written back into the body so Codex does not 400 on a >64-character value. DSH session_id is then stripped from the upstream JSON — chatgpt.com rejects it (Unsupported parameter: session_id).
2. Cache shard (Grok). xAI stores prompt cache per server. The proxy writes the same sanitized key as Responses prompt_cache_key and sends x-grok-conv-id. Codex session-id / x-client-request-id are not copied — they do nothing on this backend. A later call that reuses <10% of the previous prompt, including xAI's 512-token block on the wrong shard, is an affinity miss.
3. Stable prefix. Codex matches the longest prefix of instructions then input. Duplicate leading developer/system items are stripped; extra plan or header text is parked at the input suffix so the conversation prefix can still hit. prompt_cache_retention is dropped (gpt-5.6 rejects it).
4. Commit gate. A silent pre-output break is retried before headers are committed, so llm-pi-ai does not see a clean EOF and fire five TRANSPORT retries.

Acceptance on the full session-772f7f3a-… SkillStar turn (oauth-codex / gpt-5.6-terra-fast, 211 calls, 71 min):

	2026-08-26 incident	After 0.0.14 affinity headers
Weighted cache hit	27.4%	95.6%
Prefix reuse (median)	—	99.6%
Affinity misses	47 / 90 zero-cache	0
Prefix rewrites	—	1 adapter rebuild + 9 compaction
TRANSPORT faults	29	0

The remaining uncached tokens are almost all new tool output (delta) plus expected prefix rewrites: leaving plan mode (step 55, 169k) and DSH compaction (330k). The next call after each rewrite reused ~99%. That is not a shard miss.

Healthy rule: weighted hit ≥ 80%, zero affinity misses, no TRANSPORT. Compaction / request/header rebuild zeros do not fail the session. Details: docs/error.md.

Diagnose a session

Export the DSH session.jsonl (or unzip the session archive) and score it:

npm run analyze -- path/to/session.jsonl
node --experimental-strip-types scripts/analyze-session.ts --json path/to/session.jsonl
node --experimental-strip-types scripts/analyze-session.ts --fail-below 80 path/to/session.jsonl

The analyzer reads assistant/message usage once per turn+step (the later assistant/chunk usage event is a duplicate). It labels each call cold_start / delta / compaction / rebuild / affinity_miss so a compacted session is not flagged as a shard regression. Tool errors are split into host_timeout / cascade_abort / invalid and are not TRANSPORT. glob/grep's 30s budget lives on dsh-tool-fs-search, not this proxy — this plugin cannot raise it. Import as dsh-plugin-oauth-subs/analyze-session.

Fast mode

On Codex it is Priority Processing, not a different model family. The proxy peels host-side -fast and asks the ChatGPT Codex backend the same way Codex CLI 0.149+ does: body service_tier: "priority" plus header x-codex-routing-hint: model=<id>;tier=priority. store is forced false (the subscription Responses API 400s otherwise).

Model	Fast
GPT-5.6 Sol / Terra / Luna, GPT-5.5, GPT-5.4	Yes. Pick the -fast sibling in the model list.
GPT-5.4 Mini, GPT-5.3 Codex Spark	No. Their catalog rows carry an empty service_tiers, so no -fast sibling exists. A stale *-fast id is peeled locally instead of forwarded.
Grok	No. Grok 4.6 accepts service_tier: "priority" on the wire but a 2026-08-30 interleaved run showed no speed gain (83.34 vs 82.80 tok/s, ratio 0.994). Older ids reject the field; the proxy strips it.

ChatGPT Codex often echoes created=auto / completed=default even when Priority is requested — that echo is not a confirmation (openai/codex#14204). A 2026-08-26 Luna run measured 88.3 vs 57.5 tok/s (1.54×); a 2026-08-30 interleaved rerun did not reproduce a stable lift (mean 1.33×, pair ratios 1.90 then 0.93). Throughput-only; TTFT and cache are unchanged.

Login, token refresh, chat, and quota use one official client identity: Codex pairs originator: codex_cli_rs with User-Agent: codex_cli_rs/<version>; Grok sends x-xai-token-auth: xai-grok-cli and User-Agent: grok-cli/<version>. GLM uses ZCode's CLI poll: global provider: zai (client client_P8X5CMWmlaRO9gyO-KSqtg, then api.z.ai/api/auth/z/login to mint id.secret); China provider: bigmodel (bigmodel.cn/login, poll JWT is the Coding Plan bearer). Chat hops as ZCode Desktop 3.10.1 (User-Agent: ZCode/3.10.1 ai-sdk/anthropic/3.0.81, X-ZCode-App-Version, X-ZCode-Agent: glm, Referer / X-Title: Z Code) to api.z.ai or open.bigmodel.cn /api/anthropic/v1/messages. Completions leftover /api/coding/paas/v4 stays until the next sync. Quota still hits /api/monitor/usage/quota/limit. CLI init/poll against zcode.z.ai stays a CLI-shaped ZCode/3.10.1 identity. Antigravity chat / loadCodeAssist hop as antigravity/hub/<ver> <os>/<arch> (installed Antigravity.app or 2.11.0) to daily-cloudcode-pa with body ideType: ANTIGRAVITY and User-Agent only. No TLS fingerprint impersonation.

Models

Settings → OAuth subs → Models lists every Codex, Grok, GLM, Kiro, Antigravity, and Cursor catalog id, including Codex -fast and -900k siblings. Each row is an on/off checkbox. All on / All off apply per family.

Kiro follows kiro.dev/docs/models (no Auto router): GPT-5.6 Sol / Terra / Luna, Claude Opus 5 / 4.8 / 4.7 / 4.6 / 4.5, Claude Sonnet 5 / 4.6 / 4.5 / 4, Claude Haiku 4.5, DeepSeek 3.2, MiniMax M2.5 / M2.1, GLM-5, Qwen3 Coder Next. Ids are Kiro native (claude-opus-5, claude-sonnet-4.6, gpt-5.6-sol). Claude and GPT-5.6 advertise image input; the open-weight rows are text.

GLM is three Coding Plan models: GLM-5.3 (text), GLM-5.3-Flash (image + text), GLM-5-Turbo (text). Flash is the only multimodal row — text-only GLM models do not advertise image input to the Harness picker.

Default is all on except 900K. Pick a Codex Fast sibling (gpt-5.6-sol-fast) for Priority Processing. The -fast suffix is host-side only — the proxy strips it and sends service_tier: "priority" plus x-codex-routing-hint. Grok has no Fast sibling. GPT-5.4 Mini and GPT-5.3 Codex Spark have no Fast sibling; a leftover gpt-5.4-mini-fast is peeled to gpt-5.4-mini.

GPT-5.6 Sol / Terra / Luna accept 872K and GPT-5.4 accepts 1M, well past their default window. Pick gpt-5.6-sol-900k (and the Terra / Luna / 5.4 twins) to opt in — the -900k suffix is a stable host-side id even though the real ceiling is per-model, and the proxy strips it before the upstream request. GPT-5.5, GPT-5.4 Mini and Spark have no large variant.

Both 900K and Fast spend quota faster.

Turning a model off removes it from the next llm-pi-ai sync — it disappears from the Harness picker. Choices persist in models.json. A catalog id added later stays on until you turn it off (900K ids stay off until you turn them on).

You can pre-select while signed out; the family applies on the next sign-in. Checking a model rewrites the live routes immediately.

Grok 4.6 thinking depth is low / medium / high / xhigh. Grok 4.5 is low / medium / high (no xhigh). Reasoning cannot be turned off; if you leave it unset the API uses high. Codex GPT-5.6 Sol / Terra / Luna add max on top of low / medium / high / xhigh. Other Codex models stop at xhigh. minimal is not offered: every Codex model rejects it.

GLM-5.3 and GLM-5.3-Flash thinking depth is low / high / max (default max). There is no medium, and thinking cannot be turned off — thinking.type: disabled 400s. GLM-5-Turbo has no depth control (thinking stays on by default). The session picker only lists levels the catalog declares.

Kiro GPT-5.6 thinking depth is off / low / medium / high / xhigh / max. Off sends Kiro's wire value none (DSH has no none key). Opus 5 / 4.8 / 4.7 and Sonnet 5 add xhigh; 4.6 family stops at max; Haiku and the open-weight rows have no depth control.

Set the level in the DeepSeek Harness session model menu → Reasoning. It is not on Settings → Models. Login, logout, and each checkbox already sync the picker.

Quota

After sign-in, each account card shows official remaining quota.

Subscription	Endpoint	Display
ChatGPT Codex	chatgpt.com/backend-api/wham/usage	Plan badge (Plus / Pro / Team …) plus 5-hour + weekly windows, remaining percent and reset time
ChatGPT Codex reset	…/wham/rate-limit-reset-credits + /consume	Banked weekly-window reset credits and expiry; one confirm button per credit on the Codex card
xAI Grok	cli-chat-proxy.grok.com/v1/billing?format=credits plus /v1/user?include=subscription	Plan badge (SuperGrok / X Premium+ …) plus period usage, prepaid balance, product split
Zhipu GLM	api.z.ai or open.bigmodel.cn monitor/usage/quota/limit	Plan badge (Lite / Pro / Max) plus Coding Plan credit windows; host follows the active account
Google Antigravity	daily-cloudcode-pa loadCodeAssist + fetchAvailableModels (prod only on 5xx / transport)	Plan badge (Pro / Ultra / Free / Standard) plus SkillStar model-group remaining bars and reset time
Cursor	api2.cursor.sh DashboardService/GetCurrentPeriodUsage	Plan badge (Free / Pro / Pro+ / Ultra …) plus cycle remaining percent

Quota refreshes about once a minute, or immediately from Refresh quota. A failed read does not block chat.

After sign-in the account title shows a Plan badge. Codex reads JWT chatgpt_plan_type and usage plan_type (pro → Pro 20x / $200, prolite → Pro 5x / $100). Grok reads JWT tier and billing / user subscription_tier.

Bars interpolate green → yellow → red with remaining percent (hsl(remaining × 1.2, 78%, 38%)).

ChatGPT / Codex Plus and Pro may bank extra weekly-window resets. When the account has unused credits, the Codex card nests a Reset credits box and draws one button per credit, labeled with that credit’s expiry. Clicking Reset opens the DeepSeek Harness risk-confirmation dialog (warning icon, checkbox acknowledgement, then confirm). Confirm, then the plugin POSTs chatgpt.com/backend-api/wham/rate-limit-reset-credits/consume with { redeem_request_id } plus idempotencyKey. That spend refreshes the weekly window. Grok has no equivalent.

Options

Option	Default	Notes
port	8318	Loopback proxy port
provider	oauth	llm-pi-ai route prefix (oauth-codex / oauth-grok / oauth-glm / oauth-antigravity)
dataDir	profile data dir	auth.json, models.json, and proxy-key
grokLogin	device	device or pkce

Develop

npm test
npm run analyze -- path/to/session.jsonl

See CONTRIBUTING.md.