# CLAUDE.md ## Project AgentFeed is a self-hosted, multi-user Django + Django REST Framework web app where AI agents, scripts, or people publish artifacts and each logged-in user browses their own time-ordered, label-filterable feed. The core product principle is: agents publish content, not files. The server owns IDs, styling, storage layout, dates, and normalization into browser-native HTML, image, or PDF outputs. ## Current MVP Decisions - Each user has a personal ingest API token (DRF `Token` model), viewable/regenerable from Settings → Account. - Require `source` in every artifact payload — a free-text label identifying what posted it (agent-profile, script, or person name). Ownership of the resulting artifact comes from whichever user's token authenticated the POST, not from the `source` value itself. - `source` and `tags` together form one unified **label** system: every artifact shows all its labels next to the title, and a single `label` filter/query param matches against either. - Require login (Django session auth) for the browser feed/UI. The JSON API accepts a per-user bearer token for both ingest (POST/PATCH/DELETE) and artifact reads (GET list/detail), owner-scoped — an agent reads back with the same token it posts with; API requests never redirect to a browser login (JSON 403 instead). `GET /api/capabilities` and `GET /api/skill` are public (no user data), so agents can bootstrap before they have a key. - Exception to the above: an artifact can be explicitly opted into public sharing (per-artifact, owner-toggled from the feed UI) via a high-entropy `share_token`, served unauthenticated at `GET /s/{share_token}`. Not exposed via the ingest API yet — UI-only. Turning sharing off immediately invalidates the old link. Images embedded in a shared artifact are also served unauthenticated while it's shared, so inline content doesn't break for a link recipient. - User accounts can be created from Settings → Users (superuser-only in-app page) or via Django admin. Regular users are not staff and cannot access `/admin/`. Creating a user via Settings → Users also issues their API token immediately, shown once. - `TagColor` (label color overrides) stays global/shared across all users. - Deleting an artifact is a **soft-delete to Trash** (`deleted_at`), not a hard removal — recoverable via the "Deleted" sidebar folder (undelete), off-quota and eviction-protected while trashed, and purged permanently via "Empty Trash". `DELETE /api/artifacts/{id}` is soft too (idempotent); trashed artifacts are hidden from all API reads. A **Keep** sidebar folder is a view over `is_kept`. Both folders hide when empty. Trash auto-purges after a ~3-day retention window via a cron-less background thread triggered by deletes (`AGENTFEED_ASYNC_PURGE`, on by default); "Empty Trash" purges immediately. - Use random 8-character artifact IDs. - Start as a plain Django project; add Docker after core ingest and viewer work. - Preserve original markdown as `.md` alongside rendered `.html`. ## Architecture Notes - Planning source: `planning/daybook_planning.md`. - Django project: `daybook_project/`. - Main app: `artifacts/`. - Theme CSS: `static/themes/`. - Feed template: `templates/artifacts/feed.html`. - Settings templates: `templates/artifacts/settings_*.html`. - Stored artifact files live under `artifacts_store/{source}/{YYYY-MM-DD}/`. ## Implementation Guidelines - Prefer small, direct changes over broad abstractions. - Keep ingest normalization server-side. - The viewer should only need to render HTML, image, or PDF. - Do not make agents responsible for styling, filenames, or storage paths. - Preserve existing planning decisions unless the user explicitly changes them. - Avoid adding backwards-compatibility code before anything has shipped. ## Development And Deployment Workflow - Develop and test locally in this workspace by default. - Use the local Django commands for normal iteration: `python manage.py check`, `python manage.py migrate`, and `python manage.py runserver`. - Run the local development server on `127.0.0.1:8001`. Port `8000` may be reserved or blocked on this workstation. - Always use this same fixed port for the dev server. Never start a second instance on a different port; if `8001` is already serving, reuse it rather than starting another. - All artifact-posting tools (`skills/agentfeed-posting/scripts/agentfeed_post.py`, `scripts/seed_demo_artifacts.py`, and any future posting skill) must default `AGENTFEED_URL` to `http://127.0.0.1:8001` to match. - A regular (non-superuser) user `test` / password `testpwd` must always exist in the local dev database — it's the permanent account a sibling project (`AGENT_DAYBOOK__LLM_TEST`) uses for API testing. If the local `db.sqlite3` is ever reset, recreate this exact user (e.g. via Settings → Users) before assuming that project is broken. - Do not deploy to the Proxmox NUC during routine code iteration. - Treat the NUC container as the local live/staging target. Deploy there only at explicit stable checkpoints, such as an end-of-day build. - Before any NUC deployment, preserve `/opt/daybook/deploy/.env` and `/opt/daybook/deploy/data`. - After any NUC deployment or configuration change, run the documented health checks and update the NUC ops knowledge base plus `planning/daybook-container-handoff.md` if takeover facts changed. ## Repository Notes - `_archive_/` exists for retired or moved project docs and experiments that should not drive current implementation decisions. - Current deployment/takeover details live in `planning/daybook-container-handoff.md`; canonical NUC infrastructure details live in the NUC ops repo. ## Commands - Install dependencies: `pip install -r requirements.txt` - Run migrations: `python manage.py migrate` - Start dev server: `python manage.py runserver 127.0.0.1:8001` - Local app URL: `http://127.0.0.1:8001/` - Local API health check: `http://127.0.0.1:8001/api/capabilities` - Django checks: `python manage.py check` ## API Defaults - MVP ingest endpoint: `POST /api/artifacts` - MVP list endpoint: `GET /api/artifacts?label=&day=&kind=&q=` - Artifact view: `GET /a/{id}` - Auth: personal per-user token from Settings → Account, `Authorization: Bearer `. Get one by creating a superuser (`python manage.py createsuperuser`), logging in, and visiting Settings → Account. ## Claude Code / OpenCode This project uses `CLAUDE.md` for project instructions, compatible with both Claude Code and OpenCode. OpenCode is configured to load this file via a local, untracked `opencode.json` (it contains machine-specific MCP config, so it is not committed — recreate it locally with `{"$schema": "https://opencode.ai/config.json", "instructions": ["CLAUDE.md"], "permission": "allow"}` plus whatever MCP servers you want). After changing `opencode.json`, `.opencode/`, or this file, restart your AI coding assistant so the updated instructions are loaded.