metis app

A dependency-free reverse proxy for the Python backend, built on Node's http module (apps/metis/api-gateway, default port 3000). MetisApiGateway in src/app.ts composes an onion-model middleware chain — logging, rate-limiting, JWT auth-gating with per-method least-privilege permissions, response cache, and camel/snake transform — then matches against domain route tables (DEFAULT_ROUTES in src/config.ts: courses, assessments, tutoring, content, users, analytics, admin, search, recommendations, notifications). It forwards upstream through a BackendProxy with a real circuit breaker (src/proxy/backend-proxy.ts, CLOSED/OPEN/HALF_OPEN states) and exposes /health, /ready, and /api/v1/gateway/status plus a WebSocket relay (src/realtime/websocket-proxy.ts). It handles CORS, graceful shutdown, and body-size limits. Each middleware and route has a co-located *.test.ts. Fully implemented edge service.

The background-job runtime (apps/metis/worker). MetisWorkerService in src/main.ts orchestrates a priority JobQueue (src/queues/job-queue.ts), a cron-style JobScheduler (src/queues/scheduler.ts with DEFAULT_SCHEDULES), a processor registry, a concurrency-bounded polling loop, heartbeat and memory-pressure monitoring, graceful shutdown, and an HTTP health server (/health, /ready, /stats). It ships processors for course export, single and batch notifications, analytics aggregate/snapshot, cleanup, reports, and content transcode/thumbnail/search-reindex/assessment-generation (src/jobs/*). A deliberate honesty seam runs through it: processors that need a real data source fail closed rather than fabricate — e.g. course-export.ts refuses to export without an injected MetisCourseLoader ("refusing to fabricate course content"), and the analytics/notification processors report failure rather than fake metrics or delivery until a real backend transport is wired (see the comments in registerDefaultProcessors). The orchestration is fully implemented and tested; the default-wired processors are intentionally inert until their backing stores are injected.

The Expo / React Native learner app (apps/metis/mobile). app/_layout.tsx and app/index.tsx are the Expo Router entry; src/MetisMobileHome.tsx renders the learner home against a typed learning model in src/learning-model.ts. That model defines 19 required mobile surfaces (browse, course detail/play, study, tutor, assessment, authoring, saved, queue, resume, profile, goals, pace, preferences, reminders, exports, downloads, sharing) and five run-states, with resolver functions (resolveMetisMobileResumeState, …SyncState, buildMetisMobileSurfaceInventory, coverage/gap finders) and tests. It composes @aja/domain-motion-pipelines for embodied-instruction hooks. Honest scope note: the screen is driven by a hand-authored demo learner state (metisMobileLearnerState, learner "Maya") rather than live backend data — the surface inventory, navigation actions, and offline/grounding semantics are real and tested, but the data is a fixture, and there is a single home screen plus a Maestro smoke flow (e2e/flows/smoke.yaml) rather than the full navigable app.

The learner-facing PWA (apps/metis/web, Next.js App Router, port 3020). It has the full learner journey as real routes — courses, course play, assessments, tutoring, create/authoring, progress, study, saved, notifications, settings, login (src/app/*) — backed by ~30 domain libraries under src/lib (course-data, assessment-data, profile-data, progress-data, tutoring-data, grounded-sources, plus genuine client logic like the kalika-exploration math solver and iris-study-continuity resume planner), each with co-located tests. Server routes (src/app/api/metis/[...path], src/lib/server/metis-proxy.ts) proxy to the Python backend and forward Isis course-generation and Lilith pedagogical-tone control headers; the data libs parse the backend's snake_case responses into client types. It is a real PWA (public/sw.js, manifest.json, PwaBootstrap), has Lighthouse budgets and a high-value-route performance test, an extensive Playwright suite, and composes @kalika/*, @aja/*, @metis/models, and @metis/integrations. Fully implemented.

The operator console (apps/metis/admin, Next.js App Router, port 3021). It is a deep moderation / oversight / observability surface: routes for analytics, audit trail, complaints, content moderation, courses, incidents, monitoring, operations, oversight, review queues, source-rights, and user management (src/app/*/page.tsx). All backend access flows through AdminApiClient (src/lib/admin-client.ts, ~2,400 lines) which calls a /api/admin route that server-side-proxies to the Python backend (src/lib/server/admin-proxy.ts), and maps richly-typed snake_case API payloads (concept-graph validation, runtime rollout decisions, retrieval inspections, cohort mastery heatmaps, source-rights postures) into camelCase view models. This is a fully implemented app with a large Playwright e2e suite (e2e/ — moderation, incident command, audit trail, WCAG AA signoff, keyboard-focus paths) and per-page vitest tests, not a scaffold.