Imported from stefanogebara/twin-me (.agents/skills/twinme-dev/SKILL.md). Install upstream with npx skills add stefanogebara/twin-me --skill twinme-dev. Copyright stays with the author.

TwinMe Development Patterns

Backend Service Pattern

All backend services follow this structure:

// api/services/newService.js
import { supabaseAdmin } from './database.js';
import { complete, TIER_ANALYSIS } from './llmGateway.js';

export async function myFunction(userId) {
  // 1. Always validate userId
  if (!userId) throw new Error('userId required');
  
  // 2. Use supabaseAdmin for server-side queries
  const { data, error } = await supabaseAdmin
    .from('table_name')
    .select('*')
    .eq('user_id', userId);
  
  if (error) throw error;
  return data;
}

API Route Pattern

// api/routes/newRoute.js
import { Router } from 'express';
import { authenticateUser } from '../middleware/auth.js';

const router = Router();
router.use(authenticateUser);

router.get('/', async (req, res) => {
  try {
    const userId = req.user.id; // From JWT (payload.id || payload.userId)
    const data = await myFunction(userId);
    res.json({ success: true, data });
  } catch (error) {
    console.error('[Route] Error:', error.message);
    res.status(500).json({ success: false, error: error.message });
  }
});

export default router;

Frontend API Client Pattern

// src/services/api/newAPI.ts
import { authFetch } from './apiBase';

// IMPORTANT: Paths start with / (no /api prefix - VITE_API_URL already includes it)
export const newAPI = {
  getItems: async (): Promise<Item[]> => {
    const response = await authFetch('/items');
    if (!response.ok) throw new Error(`Failed: ${response.statusText}`);
    const data = await response.json();
    return data.data ?? [];
  },
};

Memory Stream Integration

When a new service needs to store observations:

// addMemory takes POSITIONAL args (userId, content, memoryType, metadata) —
// passing an options object stores "[object Object]" as the memory content.
import { addPlatformObservation } from './memoryStreamService.js';

await addPlatformObservation(
  userId,
  'Natural language observation about what happened',
  'platform_name',
  { ingestion_source: 'on_demand', raw_data: structuredData }
);
// Lower-level alternative: addMemory(userId, content, 'observation', metadata)

Twin Chat Context Integration

To add a new context source to twin chat:

Add a parallel fetch in api/services/twinContextBuilder.js inside Promise.all
Add the section to the system prompt in api/routes/twin-chat.js
Track in contextSources object for debugging

LLM Calls

ALL LLM calls go through the gateway:

import { complete, TIER_ANALYSIS, TIER_EXTRACTION, TIER_CHAT } from './llmGateway.js';

const result = await complete({
  tier: TIER_ANALYSIS,  // Choose appropriate tier
  messages: [{ role: 'user', content: prompt }],
  temperature: 0.3,
  maxTokens: 1000,
  userId,
  serviceName: 'goal_suggestion' // For cost tracking
});
// Streaming variant: stream({ ... }) from the same module.

Database Migrations

Location: database/migrations/YYYYMMDD_description.sql — the canonical tree for ALL NEW migrations. Do not add files to database/supabase/migrations/: it is a read-only archive of migrations that WERE applied to prod (base schema + audit-track mirrors — e.g. its 20260514 file is annotated "Applied to prod... Mirrored here"), NOT unapplied/diverged wholesale; it just is no longer where new work goes. Beware: at least one table (user_transactions, two 20260420 files) has conflicting duplicate definitions across the two trees, so neither file alone is authoritative for that table — check the live DB. supabase/migrations/ (14 ancient files) and api/migrations/ are DEAD legacy dirs; never apply them (see the README.md in each migration dir, audit 2026-07-04).
Apply via Supabase MCP: mcp__supabase__apply_migration (pass name = the filename stem; the recorded version auto-generates as a 14-digit timestamp and will NOT match the file's date prefix — that is expected, not a bug)
FK constraints MUST reference public.users(id) not auth.users(id)
Always enable RLS and create policies for user isolation
Always add service_role full access policy

Platform Data Extraction

Platform data flows through Nango proxy for token management:

Platform API -> Nango -> extractionOrchestrator -> observationIngestion -> Memory Stream

Metric types for goal tracking: sleep_hours, recovery_score, hrv, meeting_count, focus_time, listening_hours

Testing Checklist

Generate JWT with {id: 'user-uuid'} (not userId)
Test with curl against http://localhost:3004/api/
Verify RLS policies work (user can only see own data)
Check server logs for errors
Use /test-api slash command for quick endpoint testing

twinme-dev

About