Skip to Content
LYNKORIADocs

v1 implementation review (2026-11-12)

Code-grounded v1 status, fixes and limits of local verification.

Superseded

This is the 2026-11-12 baseline review, kept as history. Its open launch gates and limits moved to blockers and the tracker; the current review is the 2026-11-13 platform audit.

Reviewed on 2026-11-12 against lynkoria-implementation-plan-v1.md and the actual repository. The plan is a blueprint; this page records implemented behavior and explicit limits. Local automated checks do not constitute a production certification.

2026-11-13 platform extension

Added native SDK cores and media layers, push delivery, recording/Egress lifecycle, the independent malware scanner, console password sign-in with CLI provisioning, and load tooling. Verified scope, test results and open items are tracked in implementation progress and testing completion criteria; local load results are in load testing. The 2026-11-12 notes below remain accurate except where those pages supersede them (attachments are now scanned; console auth is no longer Access-only).

Issues addressed

AreaCorrection
Multi-tenant administrationOrganization-wide roles, project changes, key management, role grants and audit reads now remain inside the owning organization; disabled organizations are rejected.
Environment updatesPATCH contracts no longer apply defaults to omitted feature/quota/policy fields and silently reset existing settings.
AuthenticationAccess assertions require expiry/issued-at/email; messaging scope is checked; socket events and idle sessions revalidate project/user/token status; refreshed tokens replace the stored session token.
CredentialsSecret-bearing admin responses are non-cacheable; cross-origin admin mutations are rejected; expired keys cannot rotate; previous JWT keys cannot shadow the current signing key.
Webhook destination securityProduction destinations require HTTPS and an operator-approved origin; the delivery worker rechecks approval and does not follow redirects.
Dependency outagesRedis rate limits fail closed, offline queuing is disabled, and messaging/media issuance share token quotas. PostgreSQL CA verification survives URL SSL parameters.
MessagingChanges reread state under the conversation lock; membership writes share that lock; history snapshots/watermarks use repeatable-read; malformed cursors return 400; conversation/audit pagination preserves timestamp ties and PostgreSQL sub-millisecond dates.
Retention and filesExpiry emits ordered deletion tombstones, removes reactions, and retries object cleanup; user-deleted files and abandoned uploads are cleaned even without a retention policy; completion rechecks membership.
PresenceLease expiry atomically rechecks renewal, preventing an old sweep from marking an active device offline; subscription-limit failure no longer corrupts existing subscriptions.
Calls and usageFailed room creation can be retried; idempotency checks tenant; media teardown failure rolls back ended status; webhook marker/state commit together; reordered participant events and duplicate closure do not double-count usage. Reconciliation discovers missed joins.
SDKsServer-disconnected sockets can reacquire tokens/reconnect; replay handles gaps arriving mid-fetch; invalid JSON responses yield typed errors; React retains early edit/delete snapshots, cancels superseded/unmounted joins and isolates typing state.
OperationsRuntime LISTEN has a separate verified-TLS runtime connection; shutdown waits for worker tasks. Deployment uses separate migration credentials, environment-specific firewall tags, immutable images and secret-free rollback artifacts.
DocumentationFumadocs renders root docs/ with static search, responsive navigation, themes and highlighting; deployment, configuration, integrations, maintenance and verification share one content source.

Dependency and configuration review

The broad OpenTelemetry auto-instrumentation bundle pulled unused cloud-detector dependencies with two moderate advisories. It was replaced with explicit HTTP/PostgreSQL/Redis instrumentation. The ESM loader hook was added after a real export check showed that the preload alone emitted no HTTP spans; a local-collector regression test verifies export and excludes the authorization-header canary. The production dependency audit now reports no known vulnerabilities. No blanket transitive-major overrides were added.

The image scan separately found vulnerable npm tooling and Debian zlib in the pinned Node base. The runtime now excludes npm and installs the fixed, pinned zlib package; the rebuilt image passes the CI-equivalent high/critical scan with unfixed findings excluded.

Deployment, static-site and SDK-release workflows use reviewed revisions; SDK publishing now waits for successful CI. Oxlint/Oxfmt remain the lint/format tools. Non-fatal upstream warnings remain for TypeScript 7.0 declaration generation, Fumadocs Webpack cache dependency analysis and the demo playground’s media bundle size. They do not establish production media performance.

v1 implementation coverage

Calls, livestream role grants/moderation, device identities, service credentials/scopes, messaging tokens, project/environment management, console roles/audit/daily usage, ordered messaging/replay, receipts/reactions/threads, attachment authorization/retention and tenant-scoped presence are implemented. Automated tests exercise real local PostgreSQL, Redis/Valkey, LiveKit and S3-compatible storage, with local signed Access assertions for console authentication.

The source includes CI, image scanning, dependency audit, secret scanning, SDK packaging and release workflows. Their presence is not proof that hosted CI, credentials, registry publication or cloud deployments have run.

Validation performed

CheckResult and boundary
Tests92 passing: 83 package/API tests, 3 documentation checks and 6 infrastructure tests. Fresh Turbo run used no cached test results. All 55 API tests also pass against Valkey.
Static checksOxlint, Oxfmt and workspace TypeScript checks pass, including SDK documentation examples.
Production buildsAll workspace builds and the API Docker image build pass. The docs export verifies 42 HTML pages, internal links, the search index and OpenAPI.
Documentation browser checksDesktop and 390/320-pixel mobile layouts, search results/navigation, mobile sidebar, deployment contents navigation, dark theme and syntax highlighting checked locally; no browser errors observed.
Container smokeNon-root runtime, migrations, idempotent owner bootstrap, readiness, admin assertions/credentials, actual local LiveKit create/token/end and private metrics pass. Local services use development transport; managed TLS is a cloud gate.
Database permissionsSeparate migration/runtime roles tested in a temporary PostgreSQL database: migrations and runtime DML/LISTEN pass; runtime DDL is denied.
Deployment configurationProvisioning mock covers first creation and reruns; Bash/Compose validation, Actionlint, App Platform schema validation (including the monitoring worker), and Prometheus alert-rule validation pass. Account-specific deployment validation remains outstanding.
Security checksProduction dependency audit reports no known advisories; rebuilt image has no fixable high/critical findings under the CI scan policy; Gitleaks finds no leaks in the working-source snapshot. This is not a historical Git or penetration-test certification.
Release artifactsSDK package dry runs and built-asset secret checks pass. Packages have not been published.

These are local observations from this review, not ongoing guarantees. Repeat CI on the committed revision and the production verification runbook against the deployed environment.

Required launch gates outside this checkout

  • Provision and verify DigitalOcean private networking, least-privilege accounts, Spaces, Cloudflare Access, DNS and managed certificates with real credentials.
  • Exercise browser media on separate networks and mobile hardware, including UDP/TCP/TURN, screenshare, reconnect and viewer publish denial.
  • Measure target room/viewer concurrency, socket fanout, database contention, latency and transfer cost; the 300-viewer target is unverified locally.
  • Test backup/PITR restoration, secret restoration, API/media rollback and actual monitoring notifications.
  • Integrate each product’s user/tenant authorization, block/contact policy and incoming-call UX. Presence visibility is tenant-wide; service keys are environment-wide trusted credentials.
  • Decide and deploy Fidescript isolation before sensitive traffic. Per-environment LiveKit keys are not SFU isolation; one API configuration currently targets one LiveKit endpoint, so dedicated media uses a separate API/media stack.
  • Compile and device-test the native Android, iOS and React Native media layers; native media compatibility is not claimed until then.

Deliberate limits and future work

Keep one API replica; cross-node Socket.IO fanout/revocation is not implemented. Single API/media nodes are not HA. Media and PostgreSQL cannot commit atomically; orphan-room/late-webhook reconciliation and configured timeouts limit failures but do not make usage billing-grade. Buffered operational counters can lose the final 15-second interval on an abrupt crash. High-volume retention/event archival needs measured tuning.

Attachments are malware-scanned when the scanner is configured (required in production since 2026-11-13). Media removal is not a blanket self-hosted token revocation guarantee; short TTLs and ending/recreating a room are required for hard bans. Multi-device exclusion removes connected sessions but does not serialize already issued media tokens. These constraints need product/security acceptance.

Ingress, HLS/CDN delivery, Lynkoria-managed E2EE keys, multi-region, AR effects and durable offline send queues are not added speculatively. Recording/Egress, push and native SDKs were added on 2026-11-13 with the verification limits recorded in implementation progress. Their need and rollout gates remain in the plan/ADRs.

On this page