OCR d’écriture manuscrite

# Handwriting Image OCR

Extract and review handwritten text from an image URL with <integration-tag>direct:deepseek_ocr_tools</integration-tag>.

## Workflow

1. Require one reachable image URL. If the user supplied a local file, first use the runtime's approved asset-upload flow to obtain a short-lived URL; never expose a caller-local path to a hosted Action.
2. Confirm the requested language, reading order, and whether printed text mixed with handwriting should also be retained.
3. Inspect the live Action schema. Call it once with `imageUrl`.
4. Present the extracted text before summarizing or correcting it.
5. Mark uncertain characters with `[unclear]` instead of silently guessing. Preserve line breaks, labels, and list structure when the returned text supports them.

## Review checklist

- Compare names, dates, amounts, and symbols against the visible image when it is available to the agent.
- Distinguish transcription corrections from content edits.
- If the extraction is empty or visibly incomplete, report that limitation and ask for a clearer crop before retrying.
- Do not claim support for PDFs, local paths, cursive accuracy, layout coordinates, or a particular language unless the live Action establishes it.

## Attribution

This native Loadout re-authoring narrows the original GLM handwriting workflow to the live DeepSeek OCR image-URL contract. GLM-specific credentials, scripts, PDF support, layout details, and accuracy claims were removed. See [UPSTREAM.md](UPSTREAM.md) and [LICENSE.txt](LICENSE.txt).