dsh-llm-bridge

---

Overview

dsh-llm-bridge turns paid personal AI accounts into first-class LLM providers for DeepSeek Harness.

Sign in with OAuth PKCE. The plugin pools multiple accounts per vendor, rotates on rate limits and quota, and keeps tokens on the host. Other plugins can call the in-process Cordis service ctx.subscriptions without tokens leaving memory.

graph LR
    subgraph DSHCore [DeepSeek Harness Session]
        Agent[DSH Agent] --> Router{Provider Router}
    end

    subgraph BridgeCore [dsh-llm-bridge]
        Router --> Pool{Multi-Account Vendor Pool}
        Pool -->|Account 1| Acc1[Primary: Active]
        Pool -->|Account 2| Acc2[Secondary: Standby]
        Pool -->|Account 3| Acc3[Fallback: Cooldown]
        Acc1 -->|HTTP 429 / Quota| Rotate[Quota and Cooldown Rotator]
        Rotate -->|Switches traffic| Acc2
    end

    subgraph VendorBridges [Upstream Vendors]
        Acc1 --> B1[ChatGPT / Codex]
        Acc1 --> B2[Claude Pro / Max]
        Acc1 --> B3[xAI / Grok]
        Acc1 --> B4[Google Cloud Code Assist]
    end

    subgraph EcosystemBridge [Cordis: ctx.subscriptions]
        Pool --> Other[Other DSH plugins]
    end

    style DSHCore fill:#1e1e2e,stroke:#89b4fa,stroke-width:2px,color:#cdd6f4
    style BridgeCore fill:#181825,stroke:#cba6f7,stroke-width:2px,color:#cdd6f4
    style VendorBridges fill:#11111b,stroke:#a6e3a1,stroke-width:2px,color:#cdd6f4
    style EcosystemBridge fill:#181825,stroke:#f38ba8,stroke-width:2px,color:#cdd6f4

---

Built-in vendors

Key	Plan	Protocol
codex	ChatGPT Plus / Pro	Codex streaming, tool calling, image drawing
claude	Claude Pro / Max	Native Messages protocol, usage tracking
grok	xAI / X Premium	Reasoning responses, billing checks, social search
antigravity	Google Cloud Code Assist	Antigravity (/v1/loadCodeAssist, /v1/streamGenerateContent)
kimi	Moonshot Kimi	OAuth device flow (auth.kimi.com), OpenAI-compatible chat
glm	Z.ai GLM Coding Plan	GLM coding endpoint (api.z.ai), live quota monitor
cursor	Cursor	Cursor backend (api2.cursor.sh), billing-period usage
kiro	AWS Kiro	Desktop OAuth (app.kiro.dev), streaming assistance
copilot	GitHub Copilot	GitHub device flow, Copilot chat completions
qwen	Alibaba Qwen (DashScope)	OpenAI-compatible endpoint, API key auth
ernie	Baidu ERNIE (Qianfan)	OAuth2 via API Key + Secret Key
spark	iFlytek Spark	OpenAI-compatible Spark HTTP API
jetbrains	JetBrains AI Assistant	JetBrains AI relay (api.jetbrains.ai)
perplexity	Perplexity Pro	Sonar model catalog (api.perplexity.ai)
replit	Replit Core	Replit AI API, connect-token auth
cody	Sourcegraph Cody Pro	Sourcegraph API, access-token auth

Register extra vendors at runtime with createVendorFromProfile.

---

Features

Multi-account rotation

• Attach several accounts per vendor (CODEX_OAUTH_1, CODEX_OAUTH_2, …).
• Fail over on HTTP 429, RATE_LIMIT, and QUOTA_EXCEEDED.
• Preemptive switch with switchAtRemaining before a window hits zero.
• Cooldown from Retry-After, x-ratelimit-reset, ISO dates, and epoch timestamps.
• A 429 on a reasoning model cools only that family on the account. Standard models on the same account stay available.

Credentials and login

• Tokens never appear in HTTP API responses or in the Web UI. The UI shows masked labels, health, and quota bars.
• Tokens live in encrypted $DSH_HOME/.credentials.yaml on the host.
• Loopback callback (autoLoopback, on by default): for loopback redirect URIs (Codex :1455, Grok :56121) the plugin catches the callback locally. Paste fallback stays available.
• Device-code login (Codex): on a headless host, start device login, open https://auth.openai.com/codex/device on any device, enter the code.
• Web-origin redirect (useWebCallback) and manual paste of the redirected URL or code remain available.
• Access tokens refresh in the background before they expire.

Per-account proxy

Each slot accepts http://, https://, or socks5://[user:pass@]host:port. OAuth refresh, vendor checks, and model requests use that proxy. Check proxy measures round-trip latency to the vendor base URL.

Privacy and diagnostics

• privacyMask hides emails as j***[email protected] in the UI and in API labels.
• Generate diagnostics report copies plugin/runtime versions, OS, per-vendor health, HTTP status counts, recent ≥400 errors, and non-secret settings. Tokens, emails, credential refs, and proxy URLs are excluded.
• The same block links to this repo’s issue tracker.

Local Ollama fallback

When Ollama is reachable at ollamaBaseUrl (default http://127.0.0.1:11434), it appears in the native model picker. If every account of a provider is exhausted and nothing has streamed yet, chat continues on ollamaFallbackModel (or the first model from /api/tags). Logged as kind: fallback.

Reasoning, verbosity, and fast mode

• Codex effort comes from the live catalog and is sent as reasoning.effort. Grok forwards effort with its own catalog filter.
• codexVerbosity: low / medium / high as text.verbosity. Empty uses the protocol default.
• codexFastMode sends service_tier: priority.
• Claude adaptive thinking (thinking: { type: "adaptive" }) and effort (output_config.effort) for Opus 4.6+ / Sonnet 4.6+. Unsupported models omit the fields so the vendor does not return 400.
• hideDeprecatedModels drops test / preview / dev / alpha / beta / legacy ids from the picker.

Quota UI

• Composer indicator (composerQuota): off / percent / bar / forecast. Forecast uses a 24h burn-rate window.
• Session header pill with pool health, plus a modal of every account.
• Draggable overlay gauge that docks to screen edges.
• Codex reset-credit cards with a 5-second confirm and a single-flight lock so a double click cannot spend two cards.

Cordis service

Other plugins can call vendors in-process:

const res = await ctx.subscriptions.request('codex', '/backend-api/codex/images/generations', {
  method: 'POST',
  body: JSON.stringify({ prompt: 'Cyberpunk landscape', size: '1024x1024' }),
})

Calls stay in memory. A path allowlist blocks SSRF.

---

Install

dsh plugin --profile web add github:jayden-dang/dsh-llm-bridge

Or from npm after publish:

dsh plugin --profile web add dsh-llm-bridge

Restart the DSH Web UI after install (systemctl --user restart dsh-web). Open Settings → Plugins → Plugin Settings → LLM Bridge and connect accounts.

Common options live on that settings card. OAuth client IDs, redirect URIs, API base URLs, and custom vendors go in settings.yaml.

---

Configuration

dsh-llm-bridge:
  switchAtRemaining: 1
  cooldownMs: 60000
  autoLoopback: true
  privacyMask: false
  ollamaBaseUrl: http://127.0.0.1:11434
  ollamaFallback: true
  ollamaFallbackModel: ''      # empty = first model from /api/tags
  hideDeprecatedModels: false
  codexVerbosity: ''           # low | medium | high
  codexFastMode: false
  composerQuota: 'off'         # off | percent | bar | forecast
  accounts:
    codex:
      - ref: CODEX_OAUTH_1
        label: "Work Pro Account"
      - ref: CODEX_OAUTH_2
        label: "Personal Plus Account"
    claude:
      - ref: CLAUDE_OAUTH_1
        label: "Claude Max"
    grok:
      - ref: GROK_OAUTH_1
        label: "X Premium"

Per-slot fields include expiresAt (ms epoch) and proxyUrl.

---

HTTP API

Route	Method	Purpose
/dsh-llm-bridge/diagnostics	GET	Anonymized diagnostics (no secrets, tokens, or proxy URLs)
/dsh-llm-bridge/proxy-check	POST	Latency check of a slot proxy against the vendor base URL
/dsh-llm-bridge/oauth/device/start	POST	Start Codex device-code login
/dsh-llm-bridge/oauth/device/poll	POST	Poll device-code authorization
/dsh-llm-bridge/update	GET/POST	In-place update from the npm registry

Update routes require a loopback client (127.0.0.1, ::1, localhost), the x-dsh-plugin-update header, and same-origin checks.

---

Development

Source is TypeScript 7 in src/ (server + Web UI client), compiled to lib/. The client emits the DSH ModuleLoader bundle at lib/client.js.

npm run typecheck
npm run build

Localization

The plugin ships English only.

---

License

MIT © Jayden Đặng