Platform audit (2026-11-13)
Features, SDKs, consistency, security, reliability and deployment readiness, with verification results and an honest readiness assessment.
Audit date: 2026-11-13. Scope: the whole repository (API, scanner, console, playground, docs site, SDKs, database, infrastructure templates, CI, documentation). Method: reading code and docs, targeted fixes with tests, a full local verification run and browser checks. No penetration test, deployed environment, physical device or real third-party provider was used. Open items are tracked in the implementation tracker; security findings in the register; styling in the CSS audit.
Summary
The codebase was already mature: tenant isolation, token handling, ordered messaging, presence leases, fail-closed scanning and deployment tooling were implemented and tested against real local infrastructure. The audit found and fixed one high-severity issue (scanned attachments could be swapped before download), two medium issues (spoofable client IP for rate limits; unbounded password hashing), several hardening gaps, an SDK gap (no typed participant, quality or device APIs), and documentation that had fragmented across overlapping status pages. Remaining work is mostly verification that needs real infrastructure, devices or product decisions.
Capability matrix
Status: Supported = implemented and tested against real local infrastructure; Preview = implemented, verification incomplete; Product = the product backend's responsibility by design; Gap = not implemented.
Calls (3.1)
| Requirement | Status | Evidence |
|---|---|---|
| 1:1 and group audio/video, concurrent sessions | Supported | POST /v1/calls, join tokens per device; maxConcurrentRooms quota with advisory lock |
| Participant join/leave events, mute/unmute, camera, device selection, connection quality | Supported (web) | LiveKit webhooks update participant sessions; SDK CallSession.participants, onParticipantsChange, listDevices, switchDevice (added in this audit) |
| Network reconnection | Supported | LiveKit reconnect states surfaced as reconnecting; messaging socket reconnects with replay |
| Adaptive bitrate, simulcast, SVC | Supported (client) | adaptiveStream and dynacast on by default; simulcast is LiveKit's camera default; SVC codecs via roomOptions |
| Room management, participant permissions, host/moderator roles | Supported | Roles derive grants server-side; participants/{pid}/permissions, /remove; maxParticipants per call |
| Waiting rooms | Gap (FEAT-01) | Products can gate join-token issuance; no lobby state in Lynkoria |
| Call lifecycle and reliable termination | Supported | call:invited/accepted/declined/ended signals, /end, reconciler closes calls LiveKit no longer has |
| Mobile and desktop | Desktop web supported; native Preview | Native media layers not compiled or device-tested (NAT-01..03) |
| Production LiveKit configuration | Supported (template) | auto_create: false, TURN/TLS, per-environment keys, webhook key; unverified on a real node |
Screen sharing (3.2)
Supported on desktop browsers: screen, window and tab (browser picker), publish permission derived from role, events through participant snapshots, cleanup on stopScreenShare and room leave. Mobile browsers generally cannot capture the screen; native screen share is Preview.
Livestreaming (3.3)
| Requirement | Status | Evidence |
|---|---|---|
| One-to-many and interactive streams, host/speaker/viewer roles | Supported | Viewers subscribe only; backend promotes speakers (ADR 011) |
| Audience management, moderation, viewer permissions | Supported | Permission updates, removal, suspension |
| Viewer limits, multiple concurrent streams | Supported | maxParticipants, environment room quota |
| Stream health monitoring | Partial | Media metrics and alerts (MediaHostCpuHigh, MediaEgressHigh); no per-stream health API |
| Hundreds of viewers | Unverified | 300 simulated viewers locally at medium quality without loss; not a deployed result (LOAD-01) |
| Egress (HLS/RTMP), CDN | Preview / Gap | Egress lifecycle tested with a stub; HLS/CDN delivery not built (ARCH-02) |
Messaging (3.4)
Supported: direct, group and room conversations; text; delivery and read receipts; typing; edit and delete; history with cursor pagination; attachments (images, video, files) with scanning; reactions; threads; per-conversation sequence ordering; reconnect and gap replay; duplicate prevention by clientMessageId; retention with ordered tombstones. Offline handling: history replay on reconnect is supported; a durable client-side send queue is a Gap (FEAT-02). Messaging is not coupled to LiveKit: it runs on Socket.IO (WebSocket only) and PostgreSQL. The Socket.IO design is sound for one replica; scaling out needs the Redis adapter and cross-node revocation (ARCH-01).
Presence (3.5)
Supported: online/away/busy/dnd/invisible, last seen, subscriptions and updates, multiple sessions per user, heartbeats (about 25 s) with 60 s leases, disconnect detection via lease expiry, a single sweeper across instances (Redis lock) and atomic Lua recompute. Custom presence metadata is not offered. Presence is visible tenant-wide; contact/block filtering is Product.
Recording (3.6)
Preview: ledger, start/stop, webhook reconciliation with out-of-order protection, private storage, 5-minute presigned downloads through the service key, deletion and retention. E2EE calls cannot be recorded (409). No real Egress job has run (REC-01). Recording access is the product backend's decision (service-key only), so recordings are never public.
File and attachment security (3.7)
Supported: declared type allowlist per environment, size limits, signed type/length on upload, magic-byte detection and declared-type match, archive bomb limits, ClamAV scanning, fail-closed on scanner errors, scan status tracking, rejection and deletion of malicious files, quarantine upload keys (new), membership-checked 5-minute download URLs with attachment disposition, lifecycle rules. Filenames are stored as metadata only (object keys are server-generated UUID paths) and encoded in Content-Disposition.
Platform capabilities (3.8)
| Capability | Status |
|---|---|
| Multi-tenant organizations, projects, environments; per-environment origins, features, quotas, policies | Supported |
| Service keys (scoped, hashed, rotatable, revocable immediately) and short-lived ES256 messaging tokens with key rotation | Supported |
| Console RBAC (owner, admin, developer, auditor), audit log, CLI-provisioned accounts | Supported |
| Webhooks: signed, timestamped, retried with backoff, dead-letter and console retry; inbound LiveKit webhooks verified and deduplicated | Supported |
| Rate limits (per key, user, console IP/email, socket events) failing closed; quotas | Supported |
| Usage metering (daily per environment) | Supported, operational grade only |
| Service health, metrics, alerts, traces | Supported (templates unexercised in a deployment) |
API versioning (/v1), OpenAPI with CI drift check, compatibility window | Supported |
Idempotency (clientMessageId, productCallId), retries, resilience | Supported |
| Storage quotas per environment | Gap: attachment size per file only |
| AR filters, masks, beauty effects | Product/client-side by design (approach) |
SDK audit (Phase 4)
| SDK | State | Notes |
|---|---|---|
@lynkoria/server (reference) | Supported | Typed errors and request IDs for all product routes; no automatic retries (callers retry idempotent requests such as productCallId creates) |
@lynkoria/client | Supported | ESM, lazy livekit-client; now exposes participants, quality and devices instead of requiring raw LiveKit events |
@lynkoria/react | Supported | useCall, useConversation, usePresence, useTyping, new useCallParticipants |
@lynkoria/react-native | Preview | TypeScript API tested with a mocked native module; native code not compiled |
Android (io.lynkoria:*), iOS (LynkoriaCore, Lynkoria) | Core tested, media Preview | Cores tested in Docker toolchains; media/CallKit need Android SDK/Xcode |
@lynkoria/scanner-client | Supported | Used by the API; available to other products (Fidescript) |
Consistency: shared contracts (@lynkoria/contracts) define errors and events for every JavaScript SDK. CallSession.room remains an intentional escape hatch for rendering with LiveKit components. No package has been published yet (PUB-01), so clean-project installs remain to be done (TEST-01).
Consistency between docs, code, console and SDKs (Phase 5)
- Console ↔ API: every console request maps to an implemented
/v1/adminroute in the OpenAPI document; no mock data, hardcoded metrics or placeholder actions were found (input placeholders only). Metrics come fromproject_usage_dailyand the operations endpoint. - OpenAPI: regenerated with no diff (46 paths).
- Docs ↔ code fixes: environment reference completed (
CLIENT_IP_HEADER, scanner and tooling variables were missing); stale statements corrected (recording/native "not available", console Access-only, attachments unscanned); new SDK APIs documented in calls, client and React references. - Playground: exercises messaging, presence, calls and livestreams against the real API; it is development-only by construction (demo backend exists only in the dev server).
Reliability and performance (Phase 7)
Reviewed: graceful shutdown waits for workers; Redis outages fail closed with 503 and no offline queue; outbox claims with SKIP LOCKED and LISTEN/NOTIFY with polling fallback; retention in bounded batches; reconciler for missed LiveKit events; database pool limits and connect timeouts. Added: role-level statement and idle-transaction timeouts (SEC-07, to apply). Kept deliberately: per-event socket revalidation (SEC-10). Not changed: single API replica and media node (no HA). Load evidence is local only (load testing); no concurrency claim is made for deployed infrastructure.
Deployment readiness (Phase 8)
The supported target is DigitalOcean App Platform plus a LiveKit Droplet, managed PostgreSQL/Valkey, Spaces, a private scanner host and Cloudflare Pages, with provisioning, rendering, migration, rollback and verification scripts. This audit added prerequisites, a complete environment reference, scanner and console-account steps in the deployment guide, a production checklist and a security checklist. A single-host Compose deployment for other providers was not added (DEP-01): it would be a second supported target that this audit could not verify end to end.
Verification
Run locally on 2026-11-13 against PostgreSQL 18, Redis, LiveKit, RustFS (S3) and ClamAV 1.4 containers.
| Check | Result |
|---|---|
pnpm format:check, pnpm lint | Pass |
pnpm typecheck | 18/18 tasks pass |
pnpm test:infra | 6/6 pass |
pnpm turbo run test --force (with real ClamAV) | All 15 tasks pass: api 82, scanner 14 (4 against real ClamAV), auth 10, client 7, contracts 6, server 5, react 5, react-native 5, scanner-client 2, docs content 5 |
| New regression tests | Attachment swap (fails on old code, passes now), client IP header, scrypt limiter, call participants/devices, useCallParticipants |
pnpm --filter @lynkoria/api openapi | No drift (46 paths) |
pnpm build | 11/11 build tasks pass; docs export verifies 86 pages, internal links, search and OpenAPI |
| Docker images | docker build of apps/api/Dockerfile (507 MB) and apps/scanner/Dockerfile (371 MB) succeed; not Trivy-scanned locally |
node scripts/check-bundle-secrets.mjs | 141 built files contain no secrets (an env-doc PEM example was caught and replaced) |
pnpm audit --prod | No known vulnerabilities; full audit has one dev-only moderate advisory (SEC-09) |
| Browser | Console, playground (including an in-call LiveKit session) and docs checked; docs served with its CSP without violations |
Not run: hosted CI (including the new scanner Trivy job), gitleaks and Trivy locally, any deployed environment, real Egress, real push providers, native toolchains, devices, staging load tests, restore drills.
Production-readiness assessment
Not Production Ready as a whole platform. The web platform's code (calls, livestreams, messaging, presence, attachments, console) is ready with external prerequisites: provisioning, credentials and applying the infrastructure settings in blockers. Production still requires verification that has not happened: media on real networks and TURN, capacity at the target viewer counts, backup restore and rollback drills, and alert delivery. Native SDKs, recording and push are Preview until compiled, device-tested and run against real providers. A successful build and test run is not evidence of production readiness.