Comprehensive plan for fixing the production 500 on POST /auth/sign-in/oauth2
and redesigning the frontend login page to be OIDC-aware with multi-method
authentication support.
Key areas covered:
- Backend: OIDC startup validation, auth config discovery endpoint, BetterAuth
error handling, PKCE, session hardening, trustedOrigins extraction
- Frontend: Multi-method login page, PDA-friendly error display, adaptive UI
based on backend-advertised providers, loading states, accessibility
- Security: CSRF rationale, secret leakage prevention, redirect URI validation,
session idle timeout, OIDC health checks
- 6 implementation phases with file change map and testing strategy
Created with input from frontend design, backend, security, and auth architecture
specialist reviews.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The genericOAuth plugin is conditionally loaded based on OIDC_ENABLED
env var. Without it, BetterAuth has no /sign-in/oauth2 route, causing
404 when the login button is clicked.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The global CsrfGuard blocks POST /auth/sign-in/oauth2 with 403 because
unauthenticated users have no session and therefore no CSRF token.
BetterAuth handles its own CSRF protection via toNodeHandler().
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Docker/load-balancer health probes hit GET /health every ~5s from
127.0.0.1, exhausting the rate limit and causing all subsequent checks
to return 429 — making the service appear unhealthy.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
BetterAuth defaulted basePath to /api/auth but NestJS controller routes
to /auth/* (no global prefix). The auth client also pointed at the web
frontend origin instead of the API server, and LoginButton used a
nonexistent GET /auth/signin/authentik endpoint.
- Set basePath: "/auth" in BetterAuth server config
- Point auth client baseURL to API_BASE_URL with matching basePath
- Add genericOAuthClient plugin to auth client
- Use signIn.oauth2({ providerId: "authentik" }) in LoginButton
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The constant-time comparison test used Date.now() deltas with a 10ms
threshold which is unreliable in CI. Replace with deterministic tests
that verify both same-length and different-length key rejection paths
work correctly. The actual timing-safe behavior is guaranteed by
Node's crypto.timingSafeEqual which the guard uses.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
BetterAuth expects Web API Request objects (Fetch API standard) with
headers.get(), but NestJS/Express passes IncomingMessage objects with
headers[] property access. Use better-auth/node's toNodeHandler to
properly convert between Express req/res and BetterAuth's Web API handler.
Also fixes vitest SWC config to read the correct tsconfig for NestJS
decorator metadata emission, which was causing DI injection failures
in tests.
Fixes#410
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Alpine (musl libc) is incompatible with matrix-sdk-crypto-nodejs native binary
which requires glibc's ld-linux-x86-64.so.2. Switched all Node.js Dockerfiles
to node:24-slim (Debian/glibc). Also fixed docker-compose.matrix.yml network
naming from undefined mosaic-network to mosaic-internal.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
pnpm 10 blocks build scripts by default. The matrix-bot-sdk requires
@matrix-org/matrix-sdk-crypto-nodejs which downloads a platform-specific
native binary via postinstall. Added to onlyBuiltDependencies so the
Alpine (musl) binary gets installed in Docker builds.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The 0.1.0 package was ESM-only, causing ERR_PACKAGE_PATH_NOT_EXPORTED
when loaded by NestJS (which compiles to CommonJS). Version 0.1.1 ships
dual ESM/CJS builds.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Root cause: migration 20260129235248_add_link_storage_fields dropped the
personalities table and FormalityLevel enum, but migration
20260208000000_add_missing_tables later references personalities in a FK
constraint, causing ERROR: relation "personalities" does not exist on any
fresh database deployment.
Fix 1 — 20260208000000_add_missing_tables:
Recreate FormalityLevel enum and personalities table (with current schema
structure) at the top of the migration, before the FK constraint.
Fix 2 — New migration 20260215100000_fix_schema_drift:
- Create missing instances table (Federation module, never migrated)
- Recreate knowledge_links unique index (dropped, never recreated)
- Add 7 missing @@unique([id, workspaceId]) composite indexes
- Add missing agent_tasks.agent_type index
Verified: all 27 migrations apply cleanly on a fresh PostgreSQL 17 database
with pgvector.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
matrix-bot-sdk depends on the deprecated `request` library which pulls
in vulnerable form-data (<2.5.4, critical: unsafe random boundary) and
qs (<6.14.1, high: DoS via memory exhaustion). Add pnpm overrides to
force patched versions since matrix-bot-sdk has no newer release.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Keep both Mosaic Telemetry section (from develop) and Matrix Dev
Environment section (from feature branch) in .env.example.
Regenerate pnpm-lock.yaml with both dependency trees merged.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The Docker build failed because pip couldn't find mosaicstack-telemetry
from the private Gitea PyPI registry. Copy pip.conf into the image so
pip resolves the extra-index-url during docker build.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Critical fixes:
- Fix FormData field name mismatch (audio -> file) to match backend FileInterceptor
- Add /speech namespace to WebSocket connection URL
- Pass auth token in WebSocket handshake options
- Wrap audio.play() in try-catch for NotAllowedError and DOMException handling
- Replace bare catch block with named error parameter and descriptive message
- Add connect_error and disconnect event handlers to WebSocket
- Update JSDoc to accurately describe batch transcription (not real-time partial)
Important fixes:
- Emit transcription-error before disconnect in gateway auth failures
- Capture MediaRecorder error details and clean up media tracks on error
- Change TtsDefaultConfig.format type from string to AudioFormat
- Define canonical SPEECH_TIERS and AUDIO_FORMATS arrays as single source of truth
- Fix voice count from 54 to 53 in provider, AGENTS.md, and docs
- Fix inaccurate comments (Piper formats, tier prop, SpeachesProvider, TextValidationPipe)
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Comprehensive documentation for the speech services module:
- docs/SPEECH.md: Architecture, API reference, WebSocket protocol,
environment variables, provider configuration, Docker setup,
GPU VRAM budget, and frontend integration examples
- apps/api/src/speech/AGENTS.md: Module structure, provider pattern,
how to add new providers, gotchas, and test patterns
- README.md: Speech capabilities section with quick start
Fixes#406
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Implements the SpeechSettings component with four sections:
- STT settings (enable/disable, language preference)
- TTS settings (enable/disable, voice selector, tier preference, auto-play, speed control)
- Voice preview with test button
- Provider status with health indicators
Also adds Slider UI component and getHealthStatus API client function.
30 unit tests covering all sections, toggles, voice loading, and PDA-friendly design.
Fixes#404
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Implements AudioPlayer inline component with play/pause, progress bar,
speed control (0.5x-2x), download, and duration display. Adds
TextToSpeechButton "Read aloud" component that synthesizes text via
the speech API and integrates AudioPlayer for playback. Includes
useTextToSpeech hook with API integration, audio caching, and
playback state management. All 32 tests passing.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Fix sendThreadMessage room mismatch: use channelId from options instead of hardcoded controlRoomId
- Add .catch() to fire-and-forget handleRoomMessage to prevent silent error swallowing
- Wrap dispatchJob in try-catch for user-visible error reporting in handleFixCommand
- Add MATRIX_BOT_USER_ID validation in connect() to prevent infinite message loops
- Fix streamResponse error masking: wrap finally/catch side-effects in try-catch
- Replace unsafe type assertion with public getClient() in MatrixRoomService
- Add orphaned room warning in provisionRoom on DB failure
- Add provider identity to Herald error logs
- Add channelId to ThreadMessageOptions interface and all callers
- Add missing env var warnings in BridgeModule factory
- Fix JSON injection in setup-bot.sh: use jq for safe JSON construction
Fixes#377
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Add SpeechGateway with Socket.IO namespace /speech for real-time
streaming transcription. Supports start-transcription, audio-chunk,
and stop-transcription events with session management, authentication,
and buffer size rate limiting. Includes 29 unit tests covering
authentication, session lifecycle, error handling, cleanup, and
client isolation.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Add SpeechController with POST /api/speech/transcribe for audio
transcription and GET /api/speech/health for provider status.
Uses AudioValidationPipe for file upload validation and returns
results in standard { data: T } envelope.
Includes 10 unit tests covering transcribe with options, error
propagation, and all health status combinations.
Fixes#392
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Add fallback-tier TTS provider using Piper via OpenedAI Speech for
ultra-lightweight CPU-only synthesis. Maps 6 standard OpenAI voice
names (alloy, echo, fable, onyx, nova, shimmer) to Piper voices.
Update factory to use the new PiperTtsProvider class, replacing the
inline stub. Includes 37 unit tests covering provider identity,
voice mapping, and voice listing.
Fixes#395
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Create AudioValidationPipe for MIME type and file size validation,
TextValidationPipe for TTS text input validation, and DTOs for
transcribe/synthesize endpoints. Includes 36 unit tests.
Fixes#398
MB-007 (Streaming AI responses) done in commit 93cd314.
20 new tests, 132 total bridge tests pass.
Launching MB-008 (E2E tests) and MB-009 (Docs) in parallel.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Add ChatterboxSynthesizeOptions interface with referenceAudio and
emotionExaggeration fields, and comprehensive unit tests (26 tests)
covering voice cloning, emotion control, clamping, graceful degradation,
and cross-language support.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
MB-005 (Matrix command handling) and MB-006 (Herald adapter) done.
Both committed in ad24720 (bundled by pre-commit hooks).
49 Matrix tests pass, 112 total bridge tests pass.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Extract KokoroTtsProvider from factory into its own module with:
- Full voice catalog of 54 built-in voices across 8 languages
- Voice metadata parsing from ID prefix (language, gender, accent)
- Exported constants for supported formats and speed range
- Comprehensive unit tests (48 tests)
- Fix lint/type errors in chatterbox provider (Prettier + unsafe cast)
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Fix prettier formatting for Tooltip formatter props (single-line)
- Fix no-base-to-string by using typed props instead of Record<string, unknown>
- Fix restrict-template-expressions by wrapping number in String()
Refs #375
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Fix Tooltip formatter/labelFormatter type overload conflicts
- Fix Pie label render props type mismatch
- Fix telemetry.ts date split array access type
Refs #375
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Add the BaseTTSProvider abstract class and TTS provider factory that were
part of the tiered TTS architecture but missed from the previous commit.
- BaseTTSProvider: abstract base with synthesize(), listVoices(), isHealthy()
- tts-provider.factory: creates Kokoro/Chatterbox/Piper providers from config
- 30 tests (22 base provider + 8 factory)
Refs #391
Add abstract BaseTTSProvider class that implements common OpenAI-compatible
TTS logic using the OpenAI SDK with configurable baseURL. Includes synthesize(),
listVoices(), and isHealthy() methods. Create TTS provider factory that
dynamically registers Kokoro (default), Chatterbox (premium), and Piper
(fallback) providers based on configuration. Update SpeechModule to use
the factory for TTS_PROVIDERS injection token.
Also fixes lint error in speaches-stt.provider.ts (Array<T> -> T[]).
30 tests added (22 base provider + 8 factory), all passing.
Fixes#391
- Add CHAT_PROVIDERS injection token for bridge-agnostic access
- Conditional loading based on env vars (DISCORD_BOT_TOKEN, MATRIX_ACCESS_TOKEN)
- Both bridges can run simultaneously
- No crash if neither bridge is configured
- Tests verify all configuration combinations
Refs #379
The mosaicstack-telemetry package lacks py.typed marker. Add type
ignore comment consistent with other import sites.
Refs #370
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Add matrix_room_id column to workspace table (migration)
- Create MatrixRoomService for room provisioning and mapping
- Auto-create Matrix room on workspace provisioning (when configured)
- Support manual room linking for existing workspaces
- Unit tests for all mapping operations
Refs #380
The mosaicstack-telemetry package is hosted on the Gitea PyPI registry.
CI pip install needs --extra-index-url to find it.
Refs #370
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- llm-cost-table.ts: Add undefined guard for MODEL_COSTS lookup
- llm-telemetry-tracker.service.ts: Allow undefined in callingContext
for exactOptionalPropertyTypes compatibility
Refs #371
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Install recharts for data visualization
- Add Usage nav item to sidebar navigation
- Create telemetry API service with data fetching functions
- Build dashboard page with summary cards, charts, and time range selector
- Token usage line chart, cost breakdown bar chart, task outcome pie chart
- Loading and empty states handled
- Responsive layout with PDA-friendly design
- Add unit tests (14 tests passing)
Refs #375
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Create comprehensive telemetry documentation at docs/telemetry.md
- Cover configuration, event schema, predictions, SDK reference
- Include development guide with dry-run mode and troubleshooting
- Link from main README.md
Refs #376
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Instrument Coordinator.process_queue() with timing and telemetry events
- Instrument OrchestrationLoop.process_next_issue() with quality gate tracking
- Add agent-to-telemetry mapping (model, provider, harness per agent name)
- Map difficulty levels to Complexity enum and gate names to QualityGate enum
- Track retry counts per issue (increment on failure, clear on success)
- Emit FAILURE outcome on agent spawn failure or quality gate rejection
- Non-blocking: telemetry errors are logged and swallowed, never delay tasks
- Pass telemetry client from FastAPI lifespan to Coordinator constructor
- Add 33 unit tests covering all telemetry scenarios
Refs #372
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Create PredictionService for pre-task cost/token estimates
- Refresh common predictions on startup
- Integrate predictions into LLM telemetry tracker
- Add GET /api/telemetry/estimate endpoint
- Graceful degradation when no prediction data available
- Add unit tests for prediction service
Refs #373
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Create LlmTelemetryTrackerService for non-blocking event emission
- Normalize token usage across Anthropic, OpenAI, Ollama providers
- Add cost table with per-token pricing in microdollars
- Instrument chat, chatStream, and embed methods
- Infer task type from calling context
- Aggregate streaming tokens after stream ends with fallback estimation
- Add 69 unit tests for tracker service, cost table, and LLM service
Refs #371
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Add MOSAIC_TELEMETRY_* variables to .env.example with descriptions
- Pass telemetry env vars to api service in production compose
- Pass telemetry env vars to coordinator service in dev and swarm composes
- Swarm composes default to production URL (https://tel-api.mosaicstack.dev)
- Dev compose includes commented-out telemetry-api service placeholder
- All compose files default MOSAIC_TELEMETRY_ENABLED to false for safety
Refs #374
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Add mosaicstack-telemetry>=0.1.0 to pyproject.toml dependencies
- Configure Gitea PyPI registry via pip.conf (extra-index-url)
- Integrate TelemetryClient in FastAPI lifespan (start_async/stop_async)
- Store client on app.state.mosaic_telemetry for downstream access
- Create mosaic_telemetry.py helper module with:
- get_telemetry_client(): retrieve client from app state
- build_task_event(): construct TaskCompletionEvent with coordinator defaults
- create_telemetry_config(): create config from MOSAIC_TELEMETRY_* env vars
- Add 28 unit tests covering config, helpers, disabled mode, and lifespan
- New module has 100% test coverage
Refs #370
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Add docker-compose.speech.yml with three speech services:
- Speaches (STT via Whisper + basic TTS) on port 8090
- Kokoro-FastAPI (default TTS) on port 8880
- Chatterbox TTS (premium, GPU-required) on port 8881 behind
the premium-tts profile
All services include health checks, connect to the mosaic-internal
network, and follow existing naming/labeling conventions. Makefile
targets added: speech-up, speech-down, speech-logs.
Fixes#399
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Add SpeechConfig with typed configuration and startup validation for
STT (Whisper/Speaches), TTS default (Kokoro), TTS premium (Chatterbox),
and TTS fallback (Piper/OpenedAI). Includes registerAs factory for
NestJS ConfigModule integration, .env.example documentation, and 51
unit tests covering all validation paths.
Refs #401
Database: 6 models in the Prisma schema had no CREATE TABLE migration:
cron_schedules, workspace_llm_settings, quality_gates, task_rejections,
token_budgets, llm_usage_logs. Same root cause as the federation tables.
CORS: Health check requests (Docker, load balancers) don't send Origin
headers. The CORS config was rejecting these in production, causing
/health to return 500 and Docker to mark the container as unhealthy.
Requests without Origin headers are not cross-origin per the CORS spec
and should be allowed through.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Added CSRF_SECRET to docker-compose.swarm.portainer.yml (the active
Portainer deployment) and both example compose files. Also added
ENCRYPTION_KEY to the example files where it was missing.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Both env vars were missing from the API service environment in
docker-compose.prod.yml and docker-compose.build.yml, causing the
CSRF_SECRET check to fail at startup even when set in .env.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Standalone Synapse + Element Web deployment for Docker Swarm/Portainer.
Separate infrastructure from Mosaic Stack (same pattern as Authentik).
Includes: Synapse, Element Web, dedicated PostgreSQL, optional coturn.
Traefik labels match existing Stack conventions.
Refs #387
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Federation is optional and should not prevent the app from starting
when DEFAULT_WORKSPACE_ID is not set. Changed from throwing (crash)
to logging a warning. The endpoint-level validation in the controller
still rejects requests when federation is unconfigured.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The NestJS tsconfig compiles to CommonJS (module: "CommonJS") but
package.json had "type": "module", causing Node.js v24 to treat the
CJS output as ESM and fail with "exports is not defined in ES module
scope" at startup.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Federation models (FederationConnection, FederatedIdentity,
FederationMessage) and their enums were defined in the Prisma schema
but never had CREATE TABLE migrations. This caused the
20260203_add_federation_event_subscriptions migration to fail with
"relation federation_messages does not exist".
Adds new migration 20260202200000 to create the 3 missing enums,
3 missing tables, all indexes, and foreign keys. Removes the
now-redundant ALTER TABLE from the 20260203 migration since
event_type is created with the table.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The base openbao image's docker-entrypoint.sh injects -dev-root-token-id
and -dev-listen-address flags when it sees 'server' as $1, causing the
server to exit immediately (code 0). Override entrypoint with dumb-init
and call bao directly to avoid the dev-mode flag injection.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
The CMD exec form drops everything after & in the healthcheck URL,
causing uninitcode=200 and sealedcode=200 params to be lost. Without
them, OpenBao returns 501 when uninitialized, healthcheck fails, and
Swarm kills the container before the init sidecar can reach it.
Switch to CMD-SHELL with single-quoted URL to preserve query params.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
npx is unavailable in production image since npm is removed.
Use ./node_modules/.bin/prisma directly instead.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
BullMQ requires noeviction to prevent silent job data loss. With
allkeys-lru, Valkey could evict keys BullMQ depends on for job tracking.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Add docker-entrypoint.sh that runs prisma migrate deploy before
starting the app, ensuring all tables exist on deployment
- Add "type": "module" to package.json to eliminate Node.js ESM
reparsing warning for eslint.config.js
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Pass BETTER_AUTH_SECRET through all 6 docker-compose files to API container
- Fix BullModule to parse VALKEY_URL instead of VALKEY_HOST/VALKEY_PORT,
matching all other Redis consumers in the codebase
- Migrate Prisma encryption from removed $use() middleware to $extends()
client extensions (Prisma 6.x compatibility), keeping extends PrismaClient
pattern with only account and llmProviderInstance getters overridden
Co-Authored-By: Claude Opus 4.6 <[email protected]>
AuthGuard used across federation controllers depends on AuthService,
which requires AuthModule to be imported. Matches pattern used by
TasksModule, ProjectsModule, and CredentialsModule.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
CsrfGuard is already applied globally via APP_GUARD in AppModule.
The explicit @UseGuards(CsrfGuard) on FederationController caused a
DI error because CsrfService is not provided in FederationModule.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
CredentialsController uses AuthGuard which depends on AuthService.
NestJS resolves guard dependencies in the module context, so
CredentialsModule needs to import AuthModule directly.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Woodpecker v3 ignores .woodpecker.yml when a .woodpecker/ directory
exists, reading only files from the directory. Since develop has
.woodpecker/codex-review.yml, the main build pipeline was invisible
to Woodpecker on develop. Move it into the directory as build.yml.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Remove temporary debug RUN layers that were added during initial
build troubleshooting. These add build time and leak directory
structure into build logs unnecessarily.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Enable OpenBao + init sidecar in Swarm compose (was commented out)
- Fix healthcheck to accept uninitialized/sealed vault states
(add ?uninitcode=200&sealedcode=200 to /v1/sys/health)
- Replace nc-based healthcheck with wget in dev compose
- Add ORCHESTRATOR_URL env var to API service in Swarm compose
- Uncomment OpenBao volumes in Swarm compose
The healthcheck was returning HTTP 501 for uninitialized vault,
causing Swarm to restart OpenBao before init sidecar could run.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Coordinator: install all dependencies from pyproject.toml instead of
hardcoded subset (missing slowapi, anthropic, opentelemetry-*).
API: FederationAgentService now gracefully disables when orchestrator
URL is not configured instead of throwing and crashing the app.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
