灵犀学习协作平台，将私聊、Study Rooms 和 Labs 集成到浏览器工作台中，并通过 LingxiOS 执行持久化 Agent 任务。
Part of LingXi · 灵犀. The organization repository is LingXi-Org/LingxiLoop.
- LingxiOS: the independent agent execution runtime consumed through its published package.
- LingxiLit: the organization's OpenLIT-based AI observability project.
LingxiLoop uses the published @lyyzka/lingxios@3.2.4 runtime. The Web process owns authenticated ingress and control operations; the Worker is the only process that claims LingxiOS work.
LingxiLoop is a Web learning-collaboration product with direct messages, Study Rooms, and Labs.
The browser Web app is the only supported release surface. Electron remains available only for local development; it is never published, auto-updated, offered for download, or tested in CI.
Browser Web ──> LingxiLoop Web/API ──> PostgreSQL / Redis / WuKongIM / Open Notebook
│
LingxiLoop Worker ───────┘
- WuKongIM is the authoritative durable message store.
- PostgreSQL stores product state, audit, and the append-only LLM ledger.
- Redis carries ephemeral coordination.
- Vendored Open Notebook/SurrealDB owns its independent knowledge schema lifecycle.
- Web and Worker use the same server image but are independently scalable processes; Web never starts background jobs.
Requirements: Node.js 22, PostgreSQL 16 with pgvector, and Redis 7.
The scoped LingxiOS package uses GitHub Packages. Root and server npm configuration read NODE_AUTH_TOKEN from the process environment; supply a GitHub token with read:packages through your local secret manager. CI uses LINGXIOS_PACKAGES_TOKEN when configured, otherwise GITHUB_TOKEN (the package must grant this repository Actions access). Docker builds accept --secret id=npm_token,env=NODE_AUTH_TOKEN; never pass the token as a build argument or commit its value.
Web serves the official control plane on LINGXIOS_CONTROL_HOST:LINGXIOS_CONTROL_PORT (loopback port 5182 by default). Workers connect through LINGXIOS_CONTROL_URL using the same random LINGXIOS_SERVICE_TOKEN of at least 32 characters. Keep this listener on the private service network. Production shares SSE previews through a dedicated nonpersistent LINGXIOS_REALTIME_REDIS_URL; PostgreSQL remains authoritative when previews are unavailable.
Production requires a separate private LINGXIOS_R2_BUCKET and bucket-scoped LINGXIOS_R2_ACCESS_KEY_ID / LINGXIOS_R2_SECRET_ACCESS_KEY, using the existing R2_ENDPOINT. Artifacts and per-Python-cell workspace checkpoints use S3 reads with size/hash verification. Checkpoints restore files and directories, not interpreter memory. Before upgrading, stop admission and drain workers, back up the database and old homes, explicitly migrate the schema and import any existing files with the published SDK's migrateLocalRuntimeFiles; missing source files must block cutover. Keep backups until both workers and cross-node streams have been verified.
For SiliconFlow deepseek-ai/DeepSeek-V4-Flash, work-level accounting uses measured tokens with fixed standard rates of CNY 3 input, CNY 0.3 cached input and CNY 9 output per million tokens. There is no time-of-day switching. SILICONFLOW_USD_CNY_RATE converts these estimates to the existing USD ledger (development default: 7; production must configure it). CNY amounts and the conversion are retained alongside each measured call. Native budget admission conservatively reserves at the uncached rate; absent cache usage is treated as uncached and marked as unavailable. Provider invoices remain authoritative.
npm ci
Copy-Item .env.local.example .env.local
# Fill the required database, Redis, OpenAI, WuKongIM, identity, and R2 values.
npm run dev:migrate
npm run dev:previewOpen http://localhost:5180. For direct process development, run npm run dev:all. Electron can be run locally with npm run electron:dev; every package command is fixed to --publish never.
PostgreSQL starts from 0001_v1_baseline.sql and evolves only through new numbered migrations. npm run db:migrate takes an advisory lock and verifies names and checksums. The native runtime cutover is atomic: migration 0010_lingxios_native_runtime.sql rejects any old runtime data, removes the retired schema, installs the exact package schema, and records its version and SHA-256 manifest. Application startup only checks migration and package readiness; Web and Worker never execute DDL.
For the packaged service topology:
Copy-Item .env.example .env
# Fill required product and secret values; image tags are managed by CI.
npm run mvp:upCompose runs the one-shot db-migrate service before Web and Worker.
Run only the commands for the changed surface: Web uses the unprefixed lint/typecheck/test/build commands; Admin, Control, and Server use their matching prefixes; Agent Eval uses eval:check; integration accepts owning files through --file.
CI classifies changed paths, runs only their checks, and publishes only affected images to ghcr.io/<repository-owner-lowercase>/ with immutable commit-SHA tags. CI does not install or run a browser.
Production deployment and migration requirements are in docs/RELEASE.md. The current domain model is in docs/DOMAIN_MODEL.md, and Agent Eval is documented in docs/agent-eval.md.
Licensed under MIT.