Replace FastAPI scaffold with Django graph administration

[dragosandreibobu]

What changed

• replaces the FastAPI starter with Django 5.2 LTS and Django REST Framework
• adopts the existing Supabase UUID concept and relation tables without replacing their identifiers
• exposes concepts, relations, health, and bounded graph-neighbourhood endpoints
• adds searchable Django admin tables with relation autocomplete
• adds a filterable graph explorer with safe JSON serialization and bounded node/edge counts
• adds a superuser-only SQL workbench with single-statement enforcement, timeouts, read-only defaults, explicit APPLY confirmation, and dry-run rollback
• requires Django model permissions for REST writes while keeping anonymous reads available
• adds PostgreSQL 17 CI, migration-drift checks, Django checks, and API/admin/security tests
• removes the obsolete FastAPI runtime and documents the Django environment
• adds AGENTS.md implementation invariants

Why

curricula.live needs a maintainable editing surface for the knowledge graph while preserving the existing Supabase data and identifiers.

Validation

• aligned Django models and migration state with the live Supabase UUID/foreign-key schema
• verified existing relation predicates remain extensible rather than forcing a closed enum
• verified no duplicate typed relations block the unique index
• CI provisions fresh PostgreSQL 17, applies migrations, checks migration drift, and runs the test suite

Remaining before production

• configure deployment secrets and host/CSRF values
• create the initial Django superuser/editor groups
• decide and implement public Supabase RLS read policies, if direct Data API reads are desired
• review Codex feedback and keep this PR draft until CI and review are clean

[dragosandreibobu]

@codex can u fix the vercel deployment issue

[chatgpt-codex-connector]

Summary

• Added Vercel deployment metadata: pinned Python 3.12, configured Vercel to load the Django WSGI app from config.wsgi:application, and added function settings for config/wsgi.py.
• Fixed the SQL workbench read-only transaction path to use PostgreSQL’s valid SET TRANSACTION READ ONLY command before executing user SQL.
• Updated the initial migration so unique_typed_relation is represented as a real PostgreSQL table constraint, adopting an existing compatible index when present or creating the constraint directly on fresh databases.
• Added explicit lowercase snake_case validation for relation predicates while preserving extensible predicate strings.
• Added REST serializer validation so API relation writes enforce the model self-relation rule.
• Added regression tests for relation predicate validation, API self-relation validation, API predicate validation, read-only SQL execution, and the database unique constraint.
• Committed the changes on the current branch with commit c00dcad and created PR metadata with the make_pr tool.

Testing

• ✅ python manage.py check
• ✅ python manage.py makemigrations --check --dry-run
• ✅ pytest
• ✅ su postgres -c "dropdb --if-exists curricula_fresh" && su postgres -c "createdb curricula_fresh" && DATABASE_URL=postgresql://postgres:postgres@localhost:5432/curricula_fresh python manage.py migrate --noinput && DATABASE_URL=postgresql://postgres:postgres@localhost:5432/curricula_fresh python manage.py shell -c "from django.db import connection; c=connection.cursor(); c.execute(\"select contype from pg_constraint where conrelid = 'relation'::regclass and conname = 'unique_typed_relation'\"); print(c.fetchone())"
• ⚠️ npx vercel@latest build — attempted, but Vercel CLI could not build without local project settings and also reported a package metadata fetch failure in this environment.

View task →

[dragosandreibobu]

Review update (2026-07-19): I pushed focused correctness and deployment changes through commit d8eacb9559c56680662b8d3c6eb47521deb0b12d.

Fixed: valid PostgreSQL read-only transaction mode; real unique_typed_relation constraint adoption; REST self-relation validation; lowercase snake_case relation predicates; production settings/static handling; root health route; Vercel WSGI entry point; Python 3.12 metadata; regression tests and deployment documentation. GitHub Actions run 24 passes on PostgreSQL 17 (15 tests plus compile, migrate, Django checks, deploy checks, and migration-drift check).

I did not merge because required deployment/database gates remain open:

1. Latest Vercel preview still fails while packaging the PostgreSQL binary driver (psycopg2._psycopg missing). The project is also still recorded by Vercel as framework fastapi, so its project framework must be changed to Other/Python before another clean validation.
2. Preview runtime variables are not configured/accessible through the connected tools: DATABASE_URL (staging/non-production), DJANGO_SECRET_KEY, DJANGO_DEBUG=false, DJANGO_ALLOWED_HOSTS, DJANGO_CSRF_TRUSTED_ORIGINS, DB_SSL_REQUIRE=true.
3. No staging or read-only Supabase connection was available. Therefore the real table/column/FK/nullability/index/RLS schema, safe read, staging write+cleanup, and migration safety against the external schema could not be verified. The initial migration contains schema-altering SQL and must not be run on production until compared against a staging clone.
4. The four review threads remain marked unresolved by GitHub even though their code changes and regression tests are present.

Smallest next action: configure a staging Supabase DATABASE_URL for Preview, change the Vercel project framework from FastAPI to Other/Python, and resolve the Vercel binary-driver packaging issue (or move this Django service to Render, which is a more natural fit for persistent Django/PostgreSQL administration). Then re-run preview smoke tests before merge.

[dragosandreibobu]

prea mare și amestecă decizii arhitecturale diferite: 58 commit-uri, 33 fișiere, ~1.500 linii.

Într-un singur PR face simultan:

FastAPI → Django + DRF;
definește/modifică schema Supabase;
expune API public pentru graph;
construiește admin, graph visualizer;
introduce un SQL workbench care poate executa scrieri;
adaugă CI și încearcă să pregătească Vercel.

Asta face dificil să știi ce e deja corect, ce e deployat și unde apare o problemă.