메뉴
Use when user wants to Context for debugging OCR misreads - coordinate spaces, preprocessing pipeline, cache state, and common failure patterns in inventory_vision.py
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when user wants to Run the Arc Raiders AutoScrapper validation stack after edits. Use for linting, typing, tests, workflow checks, and broader repo validation.
Use when user wants to Update Metaforge snapshots and bundled default rules
Use when user wants to Validate OCR changes against failure corpus before shipping
Add a new OCR regression fixture from an ocr_debug image. Use when a scan misidentified an item and you want to lock in the correct result as a test case.
Use when user wants to Run full validation, push branch, and open a PR with appropriate context notes
Workflow for safely adding, removing, or renaming persisted config fields in config.py. Use whenever editing dataclass fields in src/autoscrapper/config.py. Covers incrementing CONFIG_VERSION, writing a migration function, and validating the round-trip.
| name | ocr-debug |
| description | Use when user wants to Context for debugging OCR misreads - coordinate spaces, preprocessing pipeline, cache state, and common failure patterns in inventory_vision.py |
| user-invocable | false |
Two spaces coexist. Mixing them is the #1 source of bugs:
preprocess_for_ocr(), passed to TesseractAny bbox returned from Tesseract over a 2x image must be halved before use in original-space operations. find_action_bbox_by_ocr handles this division internally - callers receive original-space coords.
Must be: BGR → grayscale → upscale 2x → Otsu binarization → morphological ops → Tesseract
Binarization must happen before morphological ops. Raw BGR must never go to Tesseract.
Module-level caches in inventory_vision.py:
_last_roi_hash - skips re-OCR if region unchanged_last_ocr_result - cached result for the current ROI_ITEM_NAMES - loaded item name list for fuzzy matchingreset_ocr_caches() clears all three. It is called at the start of scan_pages(). If OCR results seem stale or items are missed after a new scan starts, the cache was not reset.
_last_roi_hash cache hit on a bad frame, or low contrast regionTITLE_STRIP_* constantsslot_bgr.size == 0 guard missing in _find_first_empty_slotocr-unavailable skillocr_title_strip automatically retries with upscale=False when the first attempt returns an empty item name. If a debug image from a 2x run looks wrong (interpolation artefacts around thin strokes), the fallback will try the raw-resolution crop. You can trigger the same path manually by calling preprocess_for_ocr(roi, upscale=False).
OCR debug images land in ocr_debug/ (timestamped PNGs). Key files:
*_infobox_detect_overlay.png - infobox detection result*_ctx_menu_processed.png - context menu binarized crop*_infobox_action_sell_processed.png - sell/recycle button OCR regionWhen the title strip looks correct visually but OCR still misreads, the issue is almost always the binarization threshold or the upscale step being skipped.