Add MkDocs documentation: architecture, deployment, development, decisions

This commit is contained in:
brutsalvadi
2026-05-08 10:07:20 +02:00
parent fd2d81328f
commit ccec23c3a3
10 changed files with 591 additions and 102 deletions
+115
View File
@@ -0,0 +1,115 @@
# Authentication & Invites
## Shared database
BincioWiki shares its user database with `bincio_activity`. Both services read from and write to the same SQLite file at `/var/bincio/data/instance.db`. There is no separate user store for the wiki.
The relevant tables:
### `users`
```sql
CREATE TABLE users (
handle TEXT PRIMARY KEY,
display_name TEXT NOT NULL DEFAULT '',
password_hash TEXT NOT NULL,
is_admin INTEGER NOT NULL DEFAULT 0,
created_at INTEGER NOT NULL,
wiki_access INTEGER NOT NULL DEFAULT 1,
activity_access INTEGER NOT NULL DEFAULT 0
);
```
`wiki_access = 1` grants access to BincioWiki. `activity_access = 1` grants access to BincioActivity. The two flags are independent; a user can have one, both, or neither.
### `sessions`
```sql
CREATE TABLE sessions (
token TEXT PRIMARY KEY,
handle TEXT NOT NULL REFERENCES users(handle) ON DELETE CASCADE,
created_at INTEGER NOT NULL,
expires_at INTEGER NOT NULL
);
```
Sessions expire after **30 days**. The cookie is `bincio_session`, HTTP-only, `SameSite=Lax`, domain `.bincio.org` (shared across all subdomains).
### `invites`
```sql
CREATE TABLE invites (
code TEXT PRIMARY KEY,
created_by TEXT NOT NULL REFERENCES users(handle) ON DELETE CASCADE,
used_by TEXT REFERENCES users(handle) ON DELETE SET NULL,
created_at INTEGER NOT NULL,
used_at INTEGER,
grants_activity INTEGER NOT NULL DEFAULT 0,
grants_wiki INTEGER NOT NULL DEFAULT 0
);
```
Wiki invites are created with `grants_wiki = 1, grants_activity = 0`. The `grants_wiki` column was added by a migration in `edit/server.py` (`_migrate_db()` runs at sidecar startup).
## Session flow
```
POST /api/auth/login → bcrypt.checkpw → INSERT sessions → Set-Cookie: bincio_session
GET /api/me → SELECT sessions JOIN users → 200 + user JSON | 401
POST /api/auth/logout → DELETE sessions → Clear cookie
```
Every protected Astro page calls `GET /api/me` on load via inline JavaScript. A 401 or 403 response triggers an immediate `window.location.replace('/login/')`.
## Security model
This is **client-side enforcement only**. The static HTML is served publicly by nginx without any HTTP-level authentication. A user without JavaScript, or one who crafts direct HTTP requests, can read the raw HTML.
This is an intentional, documented tradeoff. The content is community memories (cycling routes, gear notes, trip reports) — not financial or medical data. The goal is keeping search engine crawlers and casual visitors out, not resisting determined attackers. All members of the community are trusted.
For future pages that require stricter enforcement, the architecture supports adding an Astro SSR mode with server-side cookie validation before any HTML is rendered. This would require migrating from `output: "static"` to `output: "server"`.
## Invite system
### Creating an invite
Any wiki user can generate an invite link from `/invites/`. The sidecar creates a single-use token:
```
POST /api/invites
→ { "code": "abc123def456ghi7" }
```
The invite URL is `https://wiki.bincio.org/join/?code=<token>`.
### Registration
The invitee visits the URL and fills in a handle, optional display name, and password. On submit:
```
POST /api/auth/register
{ code, handle, display_name, password }
```
The sidecar:
1. Validates the code exists and `used_by` is null.
2. Checks `grants_wiki = 1`.
3. Validates the handle (lowercase, 220 chars, `[a-z][a-z0-9_-]{1,19}`).
4. Validates password length ≥ 8 characters.
5. Checks handle uniqueness.
6. Inserts the user (`wiki_access=1, activity_access=0`) and marks the invite as used — in a single transaction.
7. Creates a session and sets the cookie. The user is immediately logged in.
### Revoking an invite
A user can revoke any of their own unused invites via the `/invites/` page. Used invites cannot be revoked.
### User limits
The `settings` table stores a `max_wiki_users` value (currently `100`). The `/api/auth/register` endpoint reads this value and rejects registration if the current wiki user count has reached the limit.
```sql
SELECT value FROM settings WHERE key = 'max_wiki_users';
-- 100
```
+83
View File
@@ -0,0 +1,83 @@
# Content Model
## Collections
Astro 6 content collections are defined in `site/src/content.config.ts`. All three load files from **outside** the `site/` submodule — Vite is configured with `server.fs.allow: ['..']` to allow this.
| Collection | Source directory | URL pattern |
|------------|-----------------|-------------|
| `entries` | `pages/*.md` | `/entries/{slug}/` |
| `blog` | `blog/*.md` | `/blog/{slug}/` |
| `index` | `config/*.md` | internal only (wikibonsai tree) |
### `entries` — wiki pages
Plain Markdown with minimal frontmatter:
```markdown
---
title: Nome della pagina
---
Contenuto in markdown...
```
Slugs are derived from the filename minus extension. Subdirectory slugs are supported: `pages/bincio-tech/gps-e-ciclocomputer.md` → slug `bincio-tech/gps-e-ciclocomputer` → URL `/entries/bincio-tech/gps-e-ciclocomputer/`.
A `generateId` override in `content.config.ts` strips the `.md` extension and preserves slashes, so Astro does not apply any githubSlug mangling.
### `blog` — stories and posts
Blog frontmatter is richer:
```markdown
---
title: Titolo del post
description: Breve descrizione
pubDate: 2024-05-01
heroImage: /assets/foto.jpg # optional
---
```
### `index` — wikibonsai semantic tree
`config/i.bonsai.md` defines the hierarchical semantic tree used by the wikibonsai integration. It is not rendered as a page; Astro reads it to resolve `[[WikiLink]]` relationships and generate forward/back reference data.
## Sections
`config/sections.json` defines the top-level wiki sections (e.g. BincioTech, BincioOfficina) and their subsections. This file is imported by:
- The Astro pages/entries index to group and display pages
- The `PageEditor.svelte` component to populate the section picker when creating a new page
Current sections: `bincio-tech`, `bincio-officina`, `bincio-tour`, `bincio-corsa`, `bincio-abbigliamento`.
## WikiLinks and wikibonsai
Pages can reference each other with `[[WikiLink]]` syntax. The remark pipeline resolves these at build time:
1. `remark-caml` parses CAML attribute syntax in frontmatter.
2. `remark-wikirefs` resolves `[[target]]` to `/entries/{slug}/` links using `resolveHtmlHref` and `resolveHtmlText` from `site/src/wikibonsai/wikirefs.ts`.
3. `generateForeRefsRemarkPlugin` collects all forward references so they can be stored in the content store.
4. Back-references (pages that link to the current page) are computed at build time in `site/src/wikibonsai/backrefs.ts` and rendered by the `BackRefs.astro` component.
Invalid wikilinks (targets that don't exist) are rendered as dimmed grey text rather than broken links.
## Assets
User-uploaded images live in `assets/` at the repo root. This directory is:
- **Gitignored** — not versioned, not part of any git commit.
- **Served** by FastAPI's `StaticFiles` mount at `/assets/{filename}` (nginx proxies `/assets/` to FastAPI in production).
- **Uploaded** via `POST /api/assets` (multipart, images only, max 10 MB). The editor inserts `![name](/assets/filename)` into the Markdown on upload.
On the VPS, assets are separate from the git push/pull cycle and must be managed independently if ever migrated.
## Delete behaviour
Deleting a page via the editor calls `DELETE /pages/{slug}` or `DELETE /stories/{slug}`. The sidecar:
1. Unlinks the file.
2. Commits the deletion with `git rm`.
In Astro dev mode, a `contentDeleteWatcher` Vite plugin watches `pages/` and `blog/` for `unlink` events. When a `.md` file is deleted it clears `.astro/data-store.json` and restarts the Astro dev server, forcing a full re-scan. Without this, Astro's persisted data store would keep serving the deleted entry.
+129
View File
@@ -0,0 +1,129 @@
# Edit Sidecar
The edit sidecar (`edit/server.py`) is a FastAPI application that handles all write operations. It runs alongside the static Astro build and is the only process that modifies content on disk.
## API endpoints
### Auth
| Method | Path | Auth | Description |
|--------|------|------|-------------|
| `GET` | `/api/me` | session | Return current user info |
| `POST` | `/api/auth/login` | — | Create session, set cookie |
| `POST` | `/api/auth/logout` | session | Destroy session, clear cookie |
| `POST` | `/api/auth/register` | — | Register via invite token |
### Wiki pages
| Method | Path | Description |
|--------|------|-------------|
| `GET` | `/pages` | List all page slugs |
| `GET` | `/pages/{slug}` | Get page content + `base_hash` |
| `POST` | `/pages/{slug}` | Save page (with 3-way merge) |
| `DELETE` | `/pages/{slug}` | Delete page |
### Blog / stories
Same as pages, with `/stories` prefix.
### Invites
| Method | Path | Description |
|--------|------|-------------|
| `GET` | `/api/invites` | List own invites |
| `POST` | `/api/invites` | Create a new wiki invite |
| `DELETE` | `/api/invites/{code}` | Revoke an unused invite |
### Other
| Method | Path | Description |
|--------|------|-------------|
| `POST` | `/api/assets` | Upload image (multipart) |
| `POST` | `/rebuild` | Trigger Astro rebuild + rsync |
| `GET` | `/api/log` | Last 50 web-editor commits |
## Git attribution
Every page save or delete is committed to git with the editor's handle as the author:
```
git add pages/bincio-tech/gps-e-ciclocomputer.md
git commit -m "brut: edited gps-e-ciclocomputer" \
--author="brut <brut@bincio.wiki>"
git update-ref refs/heads/main HEAD
```
The committer is always `bincio-wiki <wiki@bincio.wiki>`. The author field carries the human attribution.
`GIT_COMMITTER_NAME` and `GIT_COMMITTER_EMAIL` are set in the environment passed to each subprocess. On the VPS, `GIT_DIR` points to the bare repo at `/opt/bincio-wiki-repo.git` and `GIT_WORK_TREE` to `/opt/bincio_wiki`.
### The `update-ref` fix
The VPS post-receive hook runs `git checkout -f <SHA>` to check out the new tree. This detaches `HEAD` in the bare repo: subsequent sidecar commits advance the detached HEAD pointer but not `refs/heads/main`. On the next `git push` from a developer's machine, those sidecar commits become unreachable and are eventually garbage-collected.
The fix: after every `git commit`, the sidecar immediately runs:
```
git update-ref refs/heads/main HEAD
```
This keeps `refs/heads/main` in sync with the detached HEAD, so sidecar commits are always reachable via the branch and survive a future push.
### Serialisation
All git operations are serialised with a module-level `asyncio.Lock()`. This prevents two concurrent saves from interleaving `git add` + `git commit` calls and producing a corrupted index or commit history.
## Optimistic concurrency
When a page is opened for editing, the sidecar returns the current HEAD commit hash alongside the content:
```json
{ "slug": "gps-e-ciclocomputer", "content": "...", "base_hash": "a85a2ee" }
```
The editor holds this hash. On save, it sends it back:
```json
{ "content": "...", "base_hash": "a85a2ee" }
```
The sidecar compares `base_hash` against the current HEAD for that file. If they differ, another edit landed while the user was writing.
## 3-way merge
Rather than rejecting the save with a 409, the sidecar attempts a 3-way merge using `git merge-file`:
```
base = content at base_hash (what the user started with)
current = content at HEAD (what's on disk now)
user = content from request (what the user wrote)
```
If the edits touched different lines, `git merge-file` produces a clean merged result and the save proceeds transparently. If the same lines were changed by both, `git merge-file` exits with a non-zero code and embeds conflict markers (`<<<<<<< attuale / ======= / >>>>>>> handle`). The sidecar returns a 409 with the conflict-marked content so the user can resolve manually.
```python
# exit code 0 → clean merge; > 0 → conflicts present
proc = await asyncio.create_subprocess_exec(
"git", "merge-file",
"-L", "attuale", "-L", "base", "-L", handle,
str(p_current), str(p_base), str(p_user),
...
)
await proc.wait()
merged_content = p_current.read_text()
has_conflict = proc.returncode > 0
```
## Rebuild
`POST /rebuild` does three things:
1. Deletes `.astro/data-store.json` — Astro's persisted content cache. Without this, changes to files in `pages/` and `blog/` (which live outside `site/`) are not picked up by a cached build.
2. Runs `npm run build -- --force` in `site/`. The `--force` flag tells Astro to ignore any remaining stale cache.
3. If `WIKI_WEBROOT` is set, rsyncs `site/dist/` to the webroot so nginx serves the updated build immediately.
The rebuild is triggered automatically by the `PageEditor.svelte` after every successful save.
## Slug validation
Slugs are validated against `^[a-zA-Z0-9_][a-zA-Z0-9_\-/]*$` and checked for path traversal before any file operation. The resolved path must be a child of the configured base directory.
+57
View File
@@ -0,0 +1,57 @@
# System Overview
## Components
```
Browser
│ HTTPS (443)
nginx ─── static files ──► /var/www/bincio-wiki/ (Astro build output)
│ proxy_pass (127.0.0.1:4042)
FastAPI sidecar (edit/server.py)
├── reads/writes ──► pages/ (wiki markdown)
├── reads/writes ──► blog/ (blog markdown)
├── reads/writes ──► assets/ (uploaded images)
├── git add + commit ──► /opt/bincio-wiki-repo.git (bare repo)
└── reads/writes ──► /var/bincio/data/instance.db (shared SQLite)
```
## Request flow
1. Browser requests `wiki.bincio.org/entries/some-page/`.
2. nginx serves the pre-built static HTML from `/var/www/bincio-wiki/`.
3. The page's inline JavaScript calls `GET /api/me` on load. nginx proxies this to FastAPI.
4. If FastAPI returns 401/403 (no valid session), JS redirects to `/login/`.
5. On a page edit, the browser `POST /pages/{slug}` with new Markdown content. FastAPI writes the file, runs `git commit`, then `POST /rebuild` triggers `astro build --force`.
6. nginx serves the rebuilt static output immediately on next request.
## Technology stack
| Layer | Technology |
|-------|------------|
| Frontend framework | [Astro 6](https://astro.build) (static output) |
| CSS | Tailwind CSS |
| Interactive editor | Svelte (PageEditor component) |
| Markdown extensions | `remark-wikirefs`, `remark-caml` |
| Wikibonsai integration | Custom TypeScript (`site/src/wikibonsai/`) |
| API / sidecar | FastAPI + uvicorn |
| Python packaging | uv |
| Database | SQLite (shared with `bincio_activity`) |
| Passwords | bcrypt |
| VCS | Git (two bare repos on VPS) |
| Web server | nginx + Let's Encrypt (Certbot) |
| Process manager | systemd |
| VPS | Hetzner, Debian 12 |
## Port map
| Port | Service |
|------|---------|
| 443 | nginx (HTTPS, wiki.bincio.org) |
| 4042 | bincio-wiki FastAPI (production) |
| 8001 | bincio-wiki FastAPI (local dev) |
| 4321 | Astro dev server (local dev only) |
| 4041 | bincio_activity FastAPI |