ScanSlicer/AGENTS.md

# ScanSlicer (fork) — Project Guide

Purpose of this file: give any LLM working in this repo enough context to be
effective without re-reading everything. Keep it current as the code changes.

## What this project is

ScanSlicer is a desktop Qt application (Python/PySide6) that detects multiple
photos in a flatbed scan image, lets the user correct each box manually, then
exports perspective-corrected crops. Inspired by
[PhotoSlicer](https://github.com/francescovannini/photoslicer) (Francesco
Vannini). Runs on Linux over Wayland or X11.

This repo is a **local fork** of
[macrodev8/ScanSlicer](https://github.com/macrodev8/ScanSlicer) (upstream owner
inactive). License: GNU GPL v3 (must be preserved). Attribution: keep upstream
credit and the "inspired by PhotoSlicer" note.

## Goals of this fork

1. Fix a mirrored-export orientation bug and add an explicit left/right
   orientation workflow (see below). This is the primary patch.
2. Keep the app usable on the maintainer's own photo scans; afterwards only
   minimal maintenance is expected.
3. Push to the maintainer's Forgejo instance at `code.dgerdesmann.de` (no
   GitHub fork). Keep licensing, attribution, and repository description proper.

The maintainer is primarily an R user and new to Python/Git. Changes should be
small, self-checked, clean, and secure; do not assume the maintainer can vet
Python code.

## The orientation bug (primary work item)

- Exported slices can come out mirrored (left/right swapped).
- Manual workaround: drag-switch the two corners of the top edge so the red
  edge (edge 0) becomes the true top.
- Problem: `squarize()` and `reset_rotation()` re-run `normalize_top_left()`,
  which re-derives "top" from geometry and silently overwrites the manual fix.

Desired outcome:

1. A visual indication of left vs. right edges/corners (in addition to the
   existing red top edge).
2. A "swap left/right" button for the active slice.
3. The corrected orientation must persist through `squarize()` /
   `reset_rotation()`.

## Key implementation facts

- `PhotoSlice.bbox` is a 4×2 numpy array in the order
  `[top-left, top-right, bottom-right, bottom-left]`.
- `PhotoSlice.normalize_top_left()` (in `slices.py`) picks a top-left corner and
  resolves winding so edge 0→1 is the "most horizontal top" edge. This is the
  source of the ambiguity.
- `PhotoSlice.squarize()` and `reset_rotation()` both call
  `normalize_top_left()` at the end, discarding any manual orientation.
- `Autoslicer.save_slice()` (in `autoslicer.py`) builds the perspective
  transform assuming the `[TL,TR,BR,BL]` order; mirroring arises upstream of
  this from a wrongly "normalized" quad.
- Edge 0 is drawn red in `CanvasRenderMixin.paintEvent()` (`canvas_render.py`).

A persistent orientation flag on `PhotoSlice` (respected by
`normalize_top_left`/`squarize`/`reset_rotation`) plus a swap action is the
likely shape of the fix. Do not change the `[TL,TR,BR,BL]` contract of
`save_slice` without updating every consumer.

## Repository structure

```
scanslicer/                 application source (package)
  entry.py                  CLI parsing, Wayland/X11 backend selection
  main.py                   QtScanSlicer main window, theme, app bootstrap
  controller*.py            UI action handlers (status, canvas, detection,
                            image I/O, params)
  ui_builders*.py           widget construction (controls, preview, shared)
  canvas*.py                interactive canvas (interaction/render/state mixins)
  slices.py                 PhotoSlice geometry model + polys_iou
  autoslicer.py             detection/export engine wrapper (save_slice)
  autoslicer_detection.py   detection pipelines, contour→box extraction
  autoslicer_feedback.py    auto-tune profiles, scoring, feedback persistence
  params.py                 parameter model, defaults, presets
packaging/appimage/         AppImage recipe + build scripts (preferred deploy)
Dockerfile                  container runtime image
docker-compose.yml          GUI container wiring
pyproject.toml              package metadata + dependencies (uv.lock pinned)
docs/                       user-guide, architecture, parameters, autotune,
                            runtime-config, appimage, development
```

## Architecture in brief

- `QtScanSlicer` (`main.py`) is a `QMainWindow` composed from UI builder mixins
  and controller mixins.
- `QtSlicingCanvas` (`canvas.py`) = interaction + render + state mixins; holds
  the image, zoom/pan state, and the list of `PhotoSlice` objects.
- Background work (detection, auto-tune) runs on a single-worker `QThreadPool`;
  results are delivered back to the UI via Qt signals.
- Export flow: `save_all()` (`controller_image_io.py`) loops over locked slices
  and calls `Autoslicer.save_slice(bbox, path, params)` for each.

## Run / build

```bash
uv sync --frozen
uv run scanslicer /path/to/scans     # or: python -m scanslicer
```

- AppImage (preferred): `bash packaging/appimage/build-appimage.sh`
- Docker: `docker compose up --build`
- Dependencies (from `pyproject.toml`): PySide6, opencv-python, numpy, Pillow.
- Python >= 3.10; dev env is Python 3.12 in `.venv`.

## Config / persistence

Runtime files live in `~/.config/scanslicer/` (or `$XDG_CONFIG_HOME/scanslicer/`)
by default: `defaults.json`, `presets.json`, `autotune-feedback.json`,
`status.log`. Paths are overridable via `SCANSLICER_*_PATH` env vars. In the
Docker setup these land in `./.scanslicer/` on the host.

## Notes

- No automated test suite exists; validate changes by running the app on sample scans (startup, detection, save, auto-tune).