dsh-plugin-oauth-subs

简体中文 | English

[!NOTE] This is a backup mirror of xxww0098/dsh-plugin-oauth-subs (v0.0.110, MIT). The original repository has gone offline (404). All credit and copyright belong to xxww0098. This copy is preserved unchanged — the plugin's functionality is identical to the original v0.0.110.

Use a ChatGPT / Codex, xAI Grok, Zhipu GLM, AWS Kiro, Google Antigravity, Cursor, Ollama Cloud, Kimi Code Plan, GitHub Copilot, Devin Agent, Cline, or Claude subscription—or an OpenCode Go API key—inside DeepSeek Harness. The subscription families use a loopback proxy; OpenCode Go routes directly to its API. Model routes use the DSH api values openai-responses, openai-completions, and anthropic-messages.

Provenance

• This mirror was created on 2026-10-21 from the locally installed copy of dsh-plugin-oauth-subs@0.0.110 — the package is not on the public npm registry, and the upstream repository (xxww0098/dsh-plugin-oauth-subs) is offline (404), making this the only surviving copy of the artifact.
• All shipped files (lib/, scripts/, assets/, cordis.patch.yml, package.json, LICENSE, AGENTS.md) are byte-identical to that artifact. The only documentation-level changes: the attribution note above, this section, and de-linking docs/ references (those files never shipped with the package, so the links were dead).

Install

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

Open 订阅 / Subscriptions — the host home sidebar entry under 插件 Plugins. The workbench pins three top tabs — 额度 Quota, 模型 Models, 版本 Version — with a provider rail on the left and a scrollable pane on the right. One card per account (quota on every card; Ollama Cloud has no quota bars). The Version card's Installed shows the version captured when this process loaded the plugin; after a self-install it stays on the old running version until restart. The plugin is desktop-first: it never spawns dsh/npm and never restarts the host — the Version card compares the running version with the GitHub latest tag, and 安装更新 self-installs the tag tarball into the profile's node_modules (a 15-minute auto-update switch lives in the same card; on a local-directory link the switch stays, but its note says npm run build hot-reloads the plugin and drops the release-tag result). If profile node_modules is newer than the running process, the card lists On disk and flags the stale process. Or pnpm dsh web --patch ./cordis.patch.yml (id: oauth-subs).

Desktop

The desktop profile is managed exclusively by the Electron app — dsh plugin --profile desktop is rejected. Install through the UI instead:

1. DeepSeek Harness → 插件 (Plugins) → 添加插件
2. Paste https://github.com/xxww0098/dsh-plugin-oauth-subs → 安装
3. Toggle the plugin on; the component row should read 运行中

Data lives under ~/.dsh/profiles/desktop/data/dsh-plugin-oauth-subs/ — logins are not shared with the web profile. To migrate existing accounts, quit the app, then copy auth.json (and models.json for picker state) from ~/.dsh/profiles/web/data/dsh-plugin-oauth-subs/ into that directory and relaunch. Copy while the app is stopped: the plugin holds tokens in memory and may overwrite a hot edit.

The proxy port (8318 by default) is a global loopback bind — web and desktop profiles cannot run simultaneously (EADDRINUSE). Kill the other profile or set a different config.port under id: oauth-subs in the profile's cordis.patch.yml.

This plugin is specialized for Desktop: the host-lifecycle surface (DSH-CLI/npm update, restart-host) has been removed, not hidden — the Electron app owns the profile and process. Updates are self-installed: 检查更新 → 安装更新 downloads the release tarball and swaps the plugin dirs in place (or turn on Auto-update for 15-minute checks); restart the app to load the new copy. Your data/ directory survives either way. Manual fallback: 插件 → 卸载 → 添加插件 → reinstall the repo URL.

If the app exits instantly on launch, check launchctl getenv ELECTRON_RUN_AS_NODE — that variable in the user launchd environment forces every Electron app into plain Node mode; launchctl unsetenv ELECTRON_RUN_AS_NODE fixes it.

Families

Provider	Auth	DSH api	Upstream hop
ChatGPT Codex	PKCE localhost:1455 (1457 fallback); paste-callback; app_EMoamEEZ73f0CkXaXp7hrann	openai-responses	chatgpt.com/backend-api/codex/responses
xAI Grok	Device-code (default); PKCE 127.0.0.1:56121; b1a00492-073a-47ea-816f-4c329264a828	openai-responses	api.x.ai/v1/responses
GLM · Z.ai (global)	ZCode CLI poll provider: zai; mint id.secret; client_P8X5CMWmlaRO9gyO-KSqtg	anthropic-messages	api.z.ai/api/anthropic (Completions leftover …/coding/paas/v4)
GLM · BigModel (China)	Same CLI poll, provider: bigmodel; poll JWT is the bearer; client zcode	anthropic-messages	open.bigmodel.cn/api/anthropic (Completions leftover …/coding/paas/v4)
AWS Kiro	Social PKCE app.kiro.dev (3128…53153) / Builder ID / IdC / Entra / ksk_	openai-completions	q.<region>.amazonaws.com GenerateAssistantResponse
Google Antigravity	Google OAuth localhost:51121; paste-callback; 1071006060591-…apps.googleusercontent.com	openai-completions	daily-cloudcode-pa.googleapis.com/v1internal:streamGenerateContent
Cursor	PKCE poll cursor.com/loginDeepControl; or Import local Cursor	openai-completions	Connect agentn.us.api5.cursor.sh AgentService/Run
Ollama Cloud	Paste API key / import OLLAMA_API_KEY	openai-completions	https://ollama.com/v1/chat/completions
Kimi Code Plan	Device-code (no PKCE); import ~/.kimi-code/credentials/kimi-code.json; optional KIMI_API_KEY	openai-completions	https://api.kimi.com/coding/v1/chat/completions
GitHub Copilot	Device-code (no PKCE); import ~/.config/github-copilot/hosts.json; optional GITHUB_TOKEN	openai-completions	https://api.githubcopilot.com/chat/completions (tid= session)
Devin Agent	PKCE 127.0.0.1:59653; import ~/.local/share/devin/credentials.toml; paste devin-session-token$…	openai-completions	Connect server.codeium.com ApiServerService/GetChatMessage
Cline	WorkOS device-code (no PKCE); import ~/.cline/data/settings/providers.json	openai-completions	https://api.cline.bot/api/v1/chat/completions
Claude (Anthropic)	PKCE Claude.ai claude.com/cai or Console platform.claude.com (loopback /callback); read-only import of the local Claude Code login	anthropic-messages	api.anthropic.com/v1/messages (claude-code-20250219 + oauth-2025-04-20 beta identity)
OpenCode Go	Paste API key; optional Console cookie and workspace for quota	openai-completions / openai-responses	Direct https://opencode.ai/zen/go/v1

Import existing credentials

Path	Family
~/.codex/auth.json	Codex
~/.grok/auth.json, ~/.hermes/auth.json	Grok
~/.zcode/v2/config.json (also older cli/config.json / config.json)	GLM
~/.kiro/credentials.json; credentials.json (kiro.rs CWD); ~/.aws/sso/cache/kiro-auth-token.json	Kiro
Panel paste: kami / JSON / CSV / Social refresh / ksk_…	Kiro
~/.gemini/antigravity-cli/antigravity-oauth-token; ~/.cli-proxy-api/antigravity-*.json	Antigravity
macOS Keychain cursor-access-token / cursor-refresh-token; IDE state.vscdb (current OS user only); CURSOR_ACCESS_TOKEN	Cursor
OLLAMA_API_KEY env (not ~/.ollama/id_ed25519.pub)	Ollama Cloud
~/.kimi-code/credentials/kimi-code.json; read-only ~/.kimi/credentials/kimi-code.json; KIMI_API_KEY	Kimi
~/.config/github-copilot/hosts.json; OpenCode ~/.local/share/opencode/auth.json; COPILOT_GITHUB_TOKEN / GITHUB_TOKEN / GH_TOKEN	Copilot
~/.local/share/devin/credentials.toml ($XDG_DATA_HOME/devin/; Windows %LOCALAPPDATA%\devin\); DEVIN_API_KEY / WINDSURF_API_KEY	Devin
~/.cline/data/settings/providers.json	Cline
macOS Keychain Claude Code-credentials, then ~/.claude/.credentials.json (read-only)	Claude
Panel paste: OpenCode Go API key; optional Console cookie / workspace	OpenCode Go

Subscription tokens: <profile>/data/dsh-plugin-oauth-subs/auth.json (0600). OpenCode Go accounts: opencode-go.json in the same directory. Model selections: models.json.

How it works

Plane	Role
Subscriptions panel	Login / import / logout, then model sync
llm-pi-ai	DSH call plane; routes subscription families to the loopback proxy and OpenCode Go directly to its API
Loopback	http://127.0.0.1:8318/{codex,grok}/v1/responses, /glm/v1/messages (Completions leftover /glm/v1/chat/completions until the next sync), /{kiro,antigravity,cursor,ollama,kimi,copilot,devin,cline}/v1/chat/completions
Upstream	Refreshed subscription bearer or the active OpenCode Go API key

After the panel closes, DSH continues using the configured routes. The proxy binds only to loopback and uses DSH_OAUTH_SUBS_API_KEY; OpenCode Go bypasses it. Stack and module tree: AGENTS.md. Upstream references: docs/oauth.md.

Cache

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

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

Diagnose

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
npm run analyze -- --dir ~/.dsh/sessions --since 30d [--until ISO] [--json] [--compare base.json]

Directory mode aggregates every session (plain or zstd, one copy per session id) per provider and model: hit rate by call index, host retries, 300s idle timeouts, TTFB, stream silence, and pool idle; --compare prints per-1k-call deltas against a saved --json report.

The analyzer labels each call cold_start / delta / compaction / rebuild / affinity_miss so a compacted session is not flagged as a shard regression. Import as dsh-plugin-oauth-subs/analyze-session.

Fast / models / reasoning

Login and chat use official client identity; UA / fingerprint live in each src/oauth/<id>/README.md. The repos those fingerprints were copied from are listed in docs/oauth.md. The panel's Models tab: per-family checkboxes (default all on except 900K). Reasoning is set in the Harness session menu, not the Models tab. Fast and 900K spend quota faster.

Family	Fast	Window	Thinking
Codex GPT-6 Astra / Sol / Luna and GPT-5.6 Sol / Terra / Luna	Yes. -fast → Priority (service_tier: "priority" + x-codex-routing-hint; store: false)	258K default; -900k (872K)	low / medium / high / xhigh / max
Other Codex	GPT-5.5 only: Yes, -fast → Priority. GPT-5.4 / 5.4-mini / Spark retired (not supported when using Codex with a ChatGPT account)	258K (no -900k)	low–xhigh (no minimal)
Grok	No. 2026-08-30: 83.34 vs 82.80 tok/s (0.994). Older ids reject the field	—	4.6: low / medium / high / xhigh (unset = high); 4.5: no xhigh
GLM	—	—	5.3 / Flash: low / high / max (default max; no medium; disabled 400s). Turbo: on, no depth. Flash is the only GLM image row
Kiro	—	—	GPT-5.6: off / low / medium / high / xhigh / max (off → wire none). Opus 5 / 4.8 / 4.7 and Sonnet 5 add xhigh; 4.6 family to max; Haiku / OSS: none. Catalog: kiro.dev/docs/models (no Auto)
Antigravity	No	Cloud Code catalog; see family model notes	Per-model upstream support
Cursor	Per-model -fast variant	Live GetUsableModels + AvailableModels; 15-row offline fallback	Per-family registry parameters; see family model notes
Ollama Cloud	No	Live GET /api/tags (17-row Cloud snapshot fallback). Context from POST /api/show model_info.<family>.context_length. No quota bars	off / low / medium / high / max (off → wire none)
Kimi	No	Live GET /coding/v1/models (static kimi-for-coding / highspeed / k3 / k3-256k, 256k/32k). Prefix-hash cache	off / minimal / low / medium / high / xhigh / max → thinking.effort
Copilot	No	Live GET {api}/models (static floor refreshed from GitHub's official docs tables + models.dev github-copilot, 2026-09-23). Prefix-hash + X-Interaction-Id	live reasoning_effort when the catalog advertises it
Devin	Yes. -fast is a real backend variant (not Codex Priority), never through applyFastMode	Live GetCliModelConfigs (2026-09-23: 49 families / 81 picker rows; the static fallback mirrors them)	Mapped to backend chat_model_uid per family (defaultUid); thinking / fast / 1m become picker rows
Cline	No	Live GET /ai/cline/recommended-models (static feed snapshot fallback)	minimal / low / medium / high / xhigh / max → reasoning_effort (max→xhigh)
Claude	No	Static catalog pinned to pi-ai's anthropic.json (Claude Code set; cross-checked against Kiro rows)	Host-native mapping: adaptive ids (Opus 5 / 4.8 / 4.7, Sonnet 5 / 4.6, Fable) → output_config.effort (4.6 family to max); classic ids keep budget thinking
OpenCode Go	No	28 Completions + 6 Responses rows; see model audit docs/model-audit-2026-09-26.md	Depends on the model; DSH closed effort keys

Codex Priority echo created=auto / completed=default is not a confirmation (openai/codex#14204). 2026-08-26 Luna: 88.3 vs 57.5 tok/s (1.54×); 2026-08-30 interleaved mean 1.33× (1.90 then 0.93). Throughput-only; TTFT and cache unchanged.

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 + grok.com GetGrokCreditsConfig (gRPC-web)	Plan badge (SuperGrok / X Premium+ …), Grok Code access tag, period usage with cycle date range, product split; {val} cent fields become USD rows (included pool, pay-as-you-go cap, prepaid) when billed
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
AWS Kiro	q.<region>.amazonaws.com/getUsageLimits	Current cycle usage, limit, and any trial or bonus allowance
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 included-usage dollars (includedSpend/limit, cents → $used/$total) and per-split remaining percent
Ollama Cloud	No documented quota JSON	No quota bars
Kimi Code	api.kimi.com/coding/v1/usages + /me	Plan badge from /me.user_level_name plus remaining bars; no invented reset times
GitHub Copilot	api.github.com/copilot_internal/user	Plan badge (Free / Pro / Pro+ / Business / Enterprise) plus Premium remaining percent
Devin	server.codeium.com SeatManagementService/GetUserStatus	Plan badge (Pro / Max / Teams / Enterprise / Free / Trial) plus Prompt / Flow / Flex credit buckets (used / monthly grant, resets at plan_end), overage balance in USD, and daily + weekly remaining bars when the tier exposes them
Cline	api.cline.bot /users/me + /users/{id}/balance (micro-USD) + /users/me/plan; ClinePass adds /plan/usage-limits	Plan badge plus prepaid credit balance ($x.xx); ClinePass adds 5-hour / weekly / monthly bars. Credit accounts have no window bars
Claude (Anthropic)	api/oauth/usage (limits[]: session / weekly_all / weekly_scoped) + 1-token probe on /v1/messages for unified rate-limit headers	5-hour + weekly remaining bars with reset time; model-scoped weekly bars (e.g. Fable, via weekly_scoped / 7d_oi headers / seven_day_* fields); a 429 still reports utilization
OpenCode Go	Console /console/api/{orgs,go/status,billing/status,user}; legacy workspace fallback	Per-account Go usage, balance, and account email when the Console cookie is available

Refresh about once a minute, or Refresh quota. Bars: hsl(remaining × 1.2, 78%, 38%). Codex pro → Pro 20x / $200, prolite → Pro 5x / $100. Plus/Pro may bank weekly resets — one confirm button per credit on the Codex card (Harness risk dialog, then POST …/consume with { redeem_request_id } + idempotencyKey). That spend refreshes the weekly window. Grok has no equivalent. Ollama Cloud has no documented quota JSON (/api/quota 404); the card stays idle with no bars.

Options

Option	Default	Notes
port	8318	Loopback proxy port
provider	oauth	llm-pi-ai route prefix; every family lands at oauth-<id> (oauth-codex, oauth-grok, …)
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

Develop

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

See CONTRIBUTING.md.