# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Project Overview

Single-file interactive toolkit (`manga.py`) that downloads manga from m.happymh.com, stores images in Cloudflare R2 as WebP, and writes metadata to PostgreSQL. Runs as an arrow-key TUI backed by a persistent Chrome session.

## Commands

```bash
pip install -r requirements.txt   # playwright, boto3, psycopg2-binary, Pillow, python-dotenv, simple-term-menu
python manga.py                   # launch the TUI (no CLI args)
```

No tests, no lint config, no build step. Requires Google Chrome or Chromium installed. The script auto-detects from `CHROME_CANDIDATES` (macOS/Linux/Windows paths). R2 and DB credentials load lazily — see `.env` section below.

## Architecture

### Anti-bot: real Chrome + CDP + persistent profile

Cloudflare fingerprints both the TLS handshake and the browser process. The anti-detection chain matters — changing any link breaks downloads:

1. **`subprocess.Popen(CHROME_PATH, ...)`** launches the user's real Chrome binary, not Playwright's Chromium. This gives a genuine TLS fingerprint.
2. **`connect_over_cdp`** attaches Playwright to Chrome via DevTools Protocol. Playwright never *launches* Chrome — only sends CDP commands to a separately-running process.
3. **Persistent `--user-data-dir=.browser-data`** preserves `cf_clearance` cookies between runs. After the user solves Cloudflare once (Setup menu), subsequent runs skip the challenge.
4. **Single session (`_session_singleton`)** — Chrome is lazy-started on first operation and reused across all commands in one `python manga.py` run. Closed only on Quit. `with_browser(func)` catches "target closed" / "disconnected" errors, resets the singleton, and retries once.
5. **`hide_chrome()`** runs `osascript -e 'tell application "System Events" to set visible of process "Google Chrome" to false'` after launch so the window doesn't steal focus. No-op on non-macOS.

**Do not switch to headless mode.** Tried — Cloudflare blocks it because the fingerprint differs from real Chrome. **Do not parallelize chapter work across threads** with Playwright's sync API — each thread would need its own event loop and crashes with "no running event loop".

### Cloudflare handling

`wait_for_cloudflare(session)` polls `page.title()` and `page.url` for the "Just a moment" / `/challenge` markers. Recovery is manual: the user is shown the browser window and solves CAPTCHA. The Setup menu (`cmd_setup`) is the dedicated flow for this. During sync/check-missing, if the reading API returns 403, the script prints "CF blocked — run Setup" and stops.

### Navigation: `page.goto` vs JS assignment

- **Manga listing page** (`/manga/<slug>`) uses `page.goto(..., wait_until="commit")`. Works because Cloudflare on this route is lenient.
- **Reader page** (`/mangaread/<slug>/<id>`) uses `page.evaluate("window.location.href = '...'")` — bypasses CF's detection of CDP `Page.navigate` for the stricter reader route.

### Image pipeline (happymh)

Per chapter (in `_try_get_chapter_images`):
1. Register a response listener that matches `/apis/manga/reading` **AND** `cid=<chapter_id>` in the URL **AND** validates `data.id` in the response body matches. Drops pre-fetched neighbouring chapters.
2. Navigate the reader URL via `window.location.href` assignment.
3. DOM-count sanity check: `[class*="imgContainer"]` total minus `[class*="imgNext"]` gives the current chapter's actual page count. Trim captured list if it includes next-chapter previews.
4. `fetch_image_bytes(page, img)` runs `fetch(url)` via `page.evaluate` inside a `page.expect_response(...)` block. The body is read via CDP (`response.body()`) — zero base64 overhead. Fallback strips the `?q=50` query if the original URL fails.
5. `fetch_all_pages(page, images, max_attempts=3)` retries each failed page up to 3 times with 2s backoff between rounds. Returns `{page_num: bytes}` for successful fetches.

### R2 + DB write ordering

**Page rows are inserted into the DB only after the R2 upload succeeds.** This prevents orphan DB records pointing to missing R2 objects. Every `INSERT INTO "Page"` includes `width` and `height` read from the JPEG/WebP bytes via PIL (`Image.open(...).width`).

### Storage layouts

```
# Local (download command)
manga-content/<slug>/detail.json       # title, author, genres, description, mg-cover URL
manga-content/<slug>/cover.jpg         # captured from page load traffic
manga-content/<slug>/<N> <chapter>/<page>.jpg

# R2 (upload / sync)
manga/<slug>/cover.webp
manga/<slug>/chapters/<N>/<page>.webp
```

Chapter order is the API's ascending index (1-based). Chapter names can repeat (announcements, extras) so the DB `Chapter.number` column uses this index, not parsed chapter titles.

### Menu actions

- **Setup** (`cmd_setup`) → brings Chrome to front, user solves CF, validates `cf_clearance` cookie.
- **Download** (`cmd_download`) → picks URL from `manga.json`, optional chapter multi-select; saves JPGs locally.
- **Upload** (`cmd_upload` → `upload_manga_to_r2`) → converts local JPGs → WebP, uploads to R2, writes DB rows.
- **Sync** (`cmd_sync`) → combined download+upload via RAM (no local files), refreshes `Manga` row metadata, only inserts chapters missing from DB.
- **R2 / DB management** submenu (`tui_r2_manage`):
  - **Status** — single-pass R2 object count grouped by slug, plus DB row counts
  - **Edit manga info** (`tui_edit_manga`) — title/description/genre/status/coverUrl
  - **Delete specific manga** — R2 prefix + cascade DB delete
  - **Delete specific chapter** (`tui_delete_chapter`) — multi-select or "All chapters"
  - **Check missing pages** (`tui_check_missing_pages`) — for each chapter: if site page count ≠ R2 count, re-upload **inline** (browser still on that reader page); if counts match but DB `width`/`height` are NULL or 0, fix by reading WebP bytes from R2 (no re-upload)
  - **Clear ALL (R2 + DB)**
  - **Recompress manga** (`r2_recompress`) — re-encodes every WebP under `manga/<slug>/` at quality=65, overwrites in place

### WebP encoding

`_to_webp_bytes(img, quality=WEBP_QUALITY=75, method=6)` — method=6 is the slowest/smallest preset. Covers use quality 80 via `make_cover` (crops to 400×560 aspect, then resizes). Resize-during-encode was explicitly removed — page originals' dimensions are preserved.

### ESC to stop

`EscListener` puts stdin in cbreak mode (POSIX `termios`+`tty`) and runs a daemon thread listening for `\x1b`. Download/Upload/Sync check `esc.stop.is_set()` between chapters and cleanly exit. Restores terminal mode on `__exit__`. No-op on Windows (no termios) and when stdin isn't a TTY.

### Lazy config loading

`_ensure_config()` is called at the start of each R2/DB helper. It reads required env vars and constructs the boto3 client on first use. If env vars are missing, it prints the missing list and `sys.exit(1)` — no KeyError traceback on import. `s3`, `BUCKET`, `PUBLIC_URL`, `DATABASE_URL` are module globals set by that call.

## Environment variables (.env)

```
R2_ACCOUNT_ID=        # cloudflare account id
R2_ACCESS_KEY=
R2_SECRET_KEY=
R2_BUCKET=
R2_PUBLIC_URL=        # e.g. https://pub-xxx.r2.dev (trailing slash stripped)
DATABASE_URL=         # postgresql://user:pass@host:port/dbname
```

Missing any of these produces a friendly error on first R2/DB operation, not on import.

## DB schema expectations

The script reads/writes but does **not** create tables. Create them externally:

```sql
CREATE TABLE "Manga" (
  id            SERIAL PRIMARY KEY,
  slug          TEXT UNIQUE NOT NULL,
  title         TEXT NOT NULL,
  description   TEXT,
  "coverUrl"    TEXT,
  genre         TEXT,                        -- comma-joined list of all genres
  status        TEXT NOT NULL,               -- PUBLISHED | DRAFT | HIDDEN
  "createdAt"   TIMESTAMPTZ NOT NULL,
  "updatedAt"   TIMESTAMPTZ NOT NULL
);

CREATE TABLE "Chapter" (
  id          SERIAL PRIMARY KEY,
  "mangaId"   INTEGER NOT NULL REFERENCES "Manga"(id),
  number      INTEGER NOT NULL,              -- 1-based index from the API order
  title       TEXT NOT NULL,
  UNIQUE ("mangaId", number)
);

CREATE TABLE "Page" (
  id          SERIAL PRIMARY KEY,
  "chapterId" INTEGER NOT NULL REFERENCES "Chapter"(id),
  number      INTEGER NOT NULL,              -- 1-based page number
  "imageUrl"  TEXT NOT NULL,
  width       INTEGER,
  height      INTEGER,
  UNIQUE ("chapterId", number)
);
```

Column identifiers are camelCase with double quotes — matches Prisma default naming.

## Where to change what

| Task | Location |
|---|---|
| Add a new site | Extract happymh-specific bits: `fetch_chapters_via_api`, `fetch_chapters_from_dom`, `fetch_metadata`, `_try_get_chapter_images`, the `/mcover/` cover capture in `load_manga_page`, the reader URL shape. Keep Chrome/R2/DB/TUI as common. |
| New menu item | Add to `show_menu` list in `main` and dispatch in the `if idx == N:` ladder. For R2/DB ops, add to `tui_r2_manage`. |
| Tweak CF detection | `wait_for_cloudflare` / `_wait_for_cf_on_page` — edit the title/URL heuristics carefully, both ops check the same signals. |
| Change image quality | `WEBP_QUALITY` at top of file; cover quality is hard-coded 80 in `make_cover`. |
| Add a new Page-table column | Update all three `INSERT INTO "Page"` sites (`upload_manga_to_r2`, `cmd_sync`, `tui_check_missing_pages` re-upload branch) and the `SELECT ... FROM "Page"` in the dim-check query. |
| Change parallelism | `UPLOAD_WORKERS` for R2 uploads; do **not** introduce chapter-level threading (sync Playwright breaks). |

## Future: multi-site support

Current code is happymh-specific (selectors, API paths, URL patterns). To generalise, a site module would implement `fetch_chapters(page, slug)`, `get_chapter_images(page, slug, chapter_id)`, and `fetch_metadata(page)`, keeping the Chrome/R2/DB/TUI layer common.