J621/AGENTS.md

Use the venv on backend/venv/

DO NOT mess with the system's python

The original J621 (Django) app is kept for reference at /mnt/Disco/Proyects/Python3.13/J621-Django/ Consult it when matching original behavior (batch sizes for e621 lookups, per-page sets, upload flow) instead of guessing.

For Frontend work:

read both Frontend Design related Markdown files on the frontend/ folder

For anything touching e621 endpoints or payloads (posts, pools, tags, users, favorites, IQDB, blacklists, uploads), consult the e621 OpenAPI spec first: https://e621.wiki/openapi.yaml

Download it once per session (it is ~450 KB YAML), then grep it for the exact path/parameter/schema names instead of guessing:

curl -sL https://e621.wiki/openapi.yaml -o /tmp/opencode/e621-openapi.yaml
grep -n "^  /pools.json" /tmp/opencode/e621-openapi.yaml

Notes learned from the spec (verify before trusting):

• Index endpoints often return a bare array (e.g. GET /pools.json) while the show endpoint returns the object directly (e.g. GET /pools/{id}.json) — not always wrapped in a named key.
• Pools index supports search[order] (id_asc, id_desc, name, created_at, post_count), search[name_matches], search[category], search[is_active].
• Own settings (including user[blacklisted_tags]) are updated with PATCH /users/{id}.json as form-encoded data; /users/me.json returns the user object at the top level.

Project constraints (do not regress):

• Backend is Django 6 + DRF with token auth (no session auth on the API); MariaDB/Redis come from docker-compose.yml.
• Deployment will be Docker-based (compose); do not add systemd/cron unit files for scheduling — use the container setup for timers/workers.
• deploy/ is the deployment source of truth: J621-Frontend (SPA on static nginx) and J621-Backend (gunicorn + whitenoise, ffmpeg, migrations on start), three compose variants (both / frontend-only / backend-only) behind a shared nginx proxy service, and a Tailscale sidecar per compose. serve..json are the public Funnel variants (only ports 443 / 8443 / 10000); serve..tailnet.json + compose.tailnet*.yml are the same stacks without AllowFunnel (tailnet-only, no public exposure). Images are pushed to the Gitea registry with deploy/push_*.sh (multi-arch, :latest + :sha, GIT_HASH baked in for the version pill); deploy/gen_env.sh generates .env with openssl secrets (--update keeps SECRET_KEY, --force rotates it).
• Periodic commands (follow syncs, similarity cleanup, guest blacklist refresh) run in the composes' scheduler service — the backend image with the j621-scheduler entrypoint, intervals via J621_*_EVERY. No host cron.
• CI lives in .gitea/workflows: ci.yml runs on every push/PR (Django checks + the full backend suite against MariaDB/Redis service containers, frontend lint/type-check/build); publish.yml is manual and builds/pushes both images multi-arch. Jobs run on the user-scoped msi-mortar-ci runner (labels desktop + ubuntu-latest). Do not add actions/cache (cache: pip/npm) to these workflows: Gitea's cache service hangs the job on restore/save.
• Security/permission tests live in backend/apps/core/tests and need a one-time grant: GRANT ALL ON test_j621.* TO 'j621'@'%';

Dev environment (this machine):

• MariaDB/Redis run from the repo root: docker compose up -d (host ports 3307 and 6380 — other local projects own 3305/6379/3406).
• Backend dev server: backend/venv/bin/python manage.py runserver 0.0.0.0:8000. Django's system checks touch the database, so start MariaDB first; run it from backend/ with the venv python.
• Frontend dev server: npm run dev in frontend/ — Vite listens on [::1]:5173 (http://localhost:5173) and proxies /api to 127.0.0.1:8000.
• Restarting: kill dev servers by PID from pgrep -af; never pkill -f "manage.py runserver" (the pattern matches its own shell).
• Suite: backend/venv/bin/python manage.py test apps.core.tests (~80 s).
• Same-origin and cross-origin frontends both work, and the SPA can also run with no backend at all: a production build asks on first start (/setup) and stores the answer in localStorage — a URL, "" for same origin, or "none" for local mode. Local mode shows only the e621-facing pages (Online, Pools) with credentials kept in this browser (j621.e621) and a "Setup Backend" button in the shell. DEFAULT_BACKEND_URL in frontend/src/lib/backend.ts is the prefilled default; the backend allows extra origins via CORS_ALLOWED_ORIGINS (plus CSRF_TRUSTED_ORIGINS for the admin), and API media URLs are absolute (built from the request host), so signed files load cross-origin too. TRUST_PROXY_HEADERS=true is required behind a TLS-terminating proxy.
• No server-side media processing: the home server cannot handle it. Compression/optimization runs client-side (WebCodecs + WASM in a worker) and the server only applies the result via POST /api/files/J-x/optimize/.
• Do not use imgdd; perceptual hashing is imagehash server-side.
• Chat/messaging features are out of scope.

Security hardening (do not weaken):

• e621 API keys are stored Fernet-encrypted with a key derived from SECRET_KEY (apps/accounts/crypto.py); rotating SECRET_KEY invalidates them (and all signed media URLs), so users must re-enter the key.
• API throttles live in REST_FRAMEWORK (env-overridable): anon 120/min, user 600/min, login 5/min, register 20/hour, e621_proxy 60/hour. Signed media URLs (raw/thumbnail/staged-file/similarity-file actions) are exempt on purpose: /
• Only admins (superusers) may grant/revoke the staff role or delete staff/admin accounts; staff manage regular/uploader accounts only.
• Storage, duplicates, delete, temp-clear, uploads and downloads require upload rights (CanUpload), not merely authentication.
• Server-side fetching only happens for allowlisted e621 media hosts via services.validate_remote_url / open_remote (every redirect hop is re-validated); do not call requests.get on user-supplied URLs elsewhere.
• A repeatable audit harness (permission matrix, IDOR, guest visibility, signed URLs, SSRF, throttles) was used to verify this; formalising it as a test suite is still open in the roadmap.