Skip to content

AGENTS.md

Latest commit

 

History

History
24 lines (18 loc) · 3.13 KB

AGENTS.md

File metadata and controls

24 lines (18 loc) · 3.13 KB
Raw

Agent Rules

Online Screenshot Testing / 在线截图测试

When the user asks to "截张图测一下", "在线测一下", "拿当前页面跑一下", or otherwise test code against the live emulator/device screen, treat it as an online screenshot test.

Use this workflow:

  1. Capture one live screenshot first, save it under logs/ with a timestamp, and reuse that exact frame for all locate/OCR/extract calls in the test. Do not mix multiple implicit screenshots unless the test is explicitly about UI transitions.
  2. Use the project virtualenv Python: .venv\Scripts\python.exe. In PowerShell, prefer UTF-8 output ($env:PYTHONIOENCODING='utf-8' or python -X utf8) so Chinese item names and OCR text are not misread as file corruption.
  3. Put temporary test scripts, crops, overlays, and result JSON/text in a new logs/<topic>_<timestamp>/ folder unless the user asks to modify source. Do not change production code for a pure online test.
  4. Record the screenshot size and coordinate convention before interpreting boxes. For Box, confirm whether values mean (x, y, w, h) or (x1, y1, x2, y2) in the current helper.
  5. For crop-sensitive work, save debug crops or an overlay image with boxes drawn. First prove the crop is correct, then judge recognition quality. If a result is wrong, classify it as crop/coordinate failure, recognition failure, or post-processing failure.
  6. Prefer batch calls for OCR/extraction over per-box loops. Preserve input shape: Box -> value, list[Box] -> list, list[list[Box]] -> nested list. Empty or unreadable slots should stay as None, not be deleted or compacted.
  7. Reuse the same screenshot frame through APIs that support it, such as passing screenshot=frame to locate-style calls and screenshot_frame=frame to extract/OCR-style calls. This avoids UI drift between separate captures.
  8. Measure timings with time.perf_counter(). Report capture time, recognition/code time, and total time separately when useful. If there is model warm-up, say which run is warm-up and which run is measured.
  9. For irreversible game actions, do not silently guess from low-confidence or inconsistent OCR. Return None, pause, or ask for confirmation instead. When there is no oracle, compare independent evidence such as row stitching, per-cell crops, and raw OCR/confidence.
  10. In final results, include the saved screenshot/debug folder, the exact boxes or grid used, the recognized values with missing values preserved, elapsed time, and the accuracy basis if visible/manual ground truth was checked.

Useful project-specific notes:

  • Prefer package imports such as from AutoScriptor.utils.box_grid import make_box_grid, indexof; top-level tools can be shadowed by third-party packages after PaddleOCR imports.
  • For inventory counts, keep recognition semantics separate from business semantics: a missing item can become 0 at the inventory layer, while a present item with no numeric badge can become 1; the OCR layer should still return None for an empty unreadable badge.
  • If the user says not to peek at a newly added difficult screenshot before implementation, do not inspect it until after the implementation is complete.