🏗️ Tech Stack & Architecture

Tech Stack

Architecture

Routing

Lazy-loaded routes. Guards: authGuard (session/dashboard routes), noAuthGuard (/auth and /invite/:token), adminGuard, counselorGuard.

State Management

RxJS observables + Firestore as single source of truth. Services are providedIn: 'root' singletons. All messages encrypted at application layer (AES-256-GCM) before Firestore write.

Key Services

Project Structure

/solomon/
├── src/app/
│   ├── components/
│   │   ├── auth/          # Sign-in, sign-up, password reset
│   │   ├── counselor/     # Counselor dashboard, annotations, management
│   │   ├── terms/         # Terms of service
│   │   ├── admin/         # Admin panel
│   │   ├── dashboard/     # User dashboard
│   │   ├── profile/       # User profile management
│   │   └── session/       # Chat/counseling session UI
│   ├── services/          # 12 core services
│   ├── models/            # 10 data models
│   ├── guards/            # authGuard, adminGuard, counselorGuard
│   └── app.routes.ts      # Lazy-loaded routes
├── functions/src/index.ts  # AI counseling Cloud Functions
└── firebase.json           # Hosting targets (app + docs)

Counseling Modes

Both modes support counselor assignment (licensed counselor can view sessions, add annotations), guided prompt injection, and full encrypted session history.

Firestore Collections

Security: Messages encrypted at application layer (AES-256-GCM) before Firestore write. Decryption only in Cloud Functions and client.

Core Features

• AI Counseling — Claude Sonnet conversations grounded in Reformed theology, Westminster Standards, ESV; individual and couples modes
• Counselor Integration — Licensed counselors can view sessions, add pastoral annotations, inject guided prompt cards
• Couples Mode — Two-user sessions with speaker-labeled messages; partner invitation via email link
• End-to-End Encryption — All messages encrypted AES-256-GCM before storage
• User Profile Context — AI uses church role, marital status, spiritual disciplines, prayer preference, theological influences, and counseling goals
• Denominational Calibration — Vocabulary adapts to user's tradition without changing Reformed theological foundation
• Stripe Subscriptions — Counselors subscribe for platform access; unlimited clients per subscription
• PDF / DOCX Export — Session export via jsPDF and docx libraries
• iOS App — Capacitor 8 wrapper for native iOS deployment

Firebase Configuration

Cloud Function: chat

• Callable HTTP function (~2,500 line functions/src/index.ts)
• Input: sessionId, userMessage, mode (individual/couples), userId
• Safety pre-flight checks crisis and abuse keywords before calling Anthropic
• Uses claude-sonnet-4-6 as primary model with Reformed theology system prompt
• Fetches conversation history ordered by sequence number; decrypts messages
• Builds context from user profile + counselor annotations
• Couples mode prefixes messages with speaker name
• On content-filtering error, automatically retries with fallback model claude-haiku-4-5-20251001 (less likely to trigger infrastructure-level filters on biblical imagery); if the fallback also fails, surfaces a user-friendly invalid-argument HttpsError
• On HTTP 429 rate-limit error, waits 1,500 ms and retries once; if the retry also fails, surfaces a resource-exhausted HttpsError with a user-friendly message
• Other API failures return unavailable HttpsError immediately (no retry)
• Writes encrypted user + AI messages back to Firestore with sequence numbers

Cloud Function: synthesizeSpeech

• HTTP function proxying ElevenLabs streaming TTS audio to the client
• On ElevenLabs error, parses the upstream JSON to detect quota_exceeded and returns a structured { error, reason } JSON response (HTTP 502) instead of a plain text error
• Client (TtsService) reads reason and sets a ttsError signal, displayed as a dismissible banner in the chat UI

CI/CD Pipeline

5-job workflow on push to main (skipped if [skip ci] in commit message):

AI-Powered Release Notes

The release-notes job calls the Anthropic API directly via curl to generate human-readable release notes from the commit log:

The final release notes are written to release-notes.md and passed to gh release create in the create-release job.

Claude Code as Development Assistant

Beyond CI automation, Claude Code (Anthropic's AI coding assistant) is used interactively throughout development:

• Commit messages — authored by Claude Code following conventional commit format (feat:, fix:, docs:, etc.), ensuring consistent style and accurate version bumping
• Code changes — features, bug fixes, and refactors implemented with Claude Code in paired sessions
• Documentation — architecture docs, AGENTS.md, and this docs site maintained by Claude Code alongside code changes

Claude Code is configured via CLAUDE.md (Claude Code-specific instructions) and AGENTS.md (universal AI assistant instructions shared with other tools like Copilot and Cursor).

Key settings: npm install (not npm ci), NODE_TLS_REJECT_UNAUTHORIZED=0, Firebase Tools pinned to v13.15.2, WIF authentication.

Local Development

npx ng serve --configuration production --port 4202
# App: http://localhost:4202

With Firebase Emulators

firebase emulators:start
# Emulator UI: http://localhost:4000

Key File Reference