OpenSplit Rules

Stack

Backend: FastAPI + SQLAlchemy + PostgreSQL (Supabase). No Flask, no Django, no MongoDB.
Frontend: Next.js 14 App Router + TypeScript + Tailwind. No Pages Router.
Auth: JWT + Supabase Auth. Guest sessions use is_guest: true JWT claim.

Routes → Services → DB. No business logic in routes. No direct DB calls in routes.
Agent logic in backend/agents/. All LLM calls via backend/agents/llm_client.py only.
Background jobs in backend/tasks/ via Celery.

Python: snake_case files/functions, PascalCase classes, UserModel for DB models, UserCreateSchema for Pydantic.
TypeScript: PascalCase components, useHookName hooks, kebab-case.tsx component files.

Money = integers only (store in cents/smallest unit). Never floats.
Currency stored as original. FX conversion at display time only. Rates cached 1hr in Redis.
Every table: id (UUID), created_at, updated_at, deleted_at (soft delete).
All routes return: { "data": {}, "message": "success", "error": null }
Mobile-first. 360px min width. 48px min touch targets. Dark mode default.
No any in TypeScript. No inline styles.
Never call external APIs from frontend — always go through backend.
No secrets in code. Use .env only.
All DB migrations via Alembic. Never edit existing migrations.

For non-trivial changes, query Brain first before editing code:
- find-symbol, who-calls, trace-flow, impact-analysis, semantic-search
Apply edits only after Brain context is collected and cross-checked.
Prefer doctor/health checks before and after larger refactors.
Use index --force after parser/analysis upgrades so stale hashes do not hide new findings.

If the user prompt includes BRAIN_AUTO, Brain analysis is mandatory before editing.
If the user prompt uses aliases like use brain workflow, brain workflow, use brain-first, or run brain first, treat it as BRAIN_AUTO.
Run at minimum: find-symbol + who-calls + impact-analysis + semantic-search.
Provide a short findings summary, then perform only scoped edits derived from those findings.