sheep-solver

# Sheep Solver ## Core Workflow Work in the solver repo, usually `D:\Agent\tmp\auto-clicker\sheep-solver`. 1. Check current files: ```powershell Get-ChildItem -Force ``` 2. Capture or reuse a screenshot: ```powershell python scripts/run.py --capture ``` Use `python scripts/detect_occupancy.py` directly when `images/_game.png` already exists. 3. Detect the board: ```powershell python scripts/detect_occupancy.py ``` Expected outputs: `board_grid.json`, `data/board.json`, `sheep_candidates.json`, `images/_occ_axis_rect.png`, `images/_grid_labels.png`. 4. Inspect the overlay before trusting the solver: - `images/_occ_axis_rect.png`: rectified board overlay, best for checking cell occupancy and facing. - `images/_grid_labels.png`: rectified board with A/B/C column labels, 1/2/3 row labels, and sheep ids. - `sheep_candidates.json`: kept/dropped candidates and per-end scores. 5. Solve: ```powershell python scripts/solve_board.py ``` `scripts/solve_board.py` draws the click order on `images/_occ_axis_rect.png` and writes `images/_solution.png`. For larger boards it uses weighted A* first, then beam search, then greedy fallback. ## Recognition Rules - Treat `data/grid_params.json` as the source of truth for perspective calibration. If the grid is offset, fix calibration before tuning masks. - The detector uses a rectified `rows x cols x 64px` board via `scripts/board_grid.py`; do not duplicate perspective math in ad hoc scripts. - Sheep are two-cell pieces. `cells[0]` is rump, `cells[1]` is head, and `facing` is `rump -> head`. - Head/tail is primarily a shape decision: the head is the pointier end with less white body mask and lower distance-transform radius. Warm face/ear pixels are only a weak tie-breaker because horns, ears, and feet can appear near the rump. - Prefer fixing `scripts/detect_occupancy.py` scoring or masks over hand-editing `data/board.json`; generated JSON should be reproducible. ## Debugging When direction is wrong: 1. Open `images/_occ_axis_rect.png` and identify the sheep id. 2. Open `sheep_candidates.json`; match `kept[id].source_id` to the raw candidate. 3. Compare the two endpoint metrics in `raw[].metrics`: `white`, `face`, `dt_mean`, and `hist`. 4. Adjust scoring in `scripts/detect_occupancy.py` so the rule generalizes, then rerun: ```powershell python scripts/detect_occupancy.py python scripts/solve_board.py ``` When occupancy conflicts appear: - Check `候选/保留/丢弃` and `冲突` counts from `python scripts/detect_occupancy.py`. - Inspect `sheep_candidates.json` and `images/_occ_axis_rect.png` for over-splitting, merged sheep, or wrong cell pairs. - Keep the non-overlap resolver conservative: a no-conflict board is more useful than an over-eager board with duplicate cells. For more detail, read `references/workflow.md`. ## Validation Run after detector or solver edits: ```powershell python -m compileall -q scripts python -m pytest -q tests/test_solver.py python scripts/detect_occupancy.py python scripts/solve_board.py ``` If `images/_game.png` is absent, run `python scripts/run.py --capture` first or ask for/provide a screenshot.