Table Image to Markdown

# Table Image to Markdown

Extract text from a table image through <integration-tag>direct:deepseek_ocr_tools</integration-tag>, then reconstruct a reviewable Markdown table without inventing cells.

## Workflow

1. Obtain one reachable image URL through the approved asset flow.
2. Ask which region or table matters when the image contains multiple tables.
3. Inspect the live Action schema and call it once with `imageUrl`.
4. Preserve the raw extraction as evidence.
5. Reconstruct rows and columns only where headers and cell boundaries are supported by that extraction or by the visible image.

## Output

Return:

1. The Markdown table.
2. An `Uncertain cells` list with row, column, observed text, and reason.
3. Any text outside the table that may affect interpretation, such as units, footnotes, or date ranges.

Use empty cells for missing values and `[unclear]` for unreadable values. Do not merge repeated labels, calculate totals, or normalize units unless the user asks; when doing so, keep the original values alongside the normalized result.

## Boundaries

The live Action accepts an image URL and promises text extraction, not guaranteed table structure. Do not claim merged-cell recovery, spreadsheet export, multi-page PDF support, or pixel-perfect layout. For high-stakes financial, medical, or legal tables, require human verification against the source image.

## Attribution

This native Loadout re-authoring retains the source table-review method while narrowing execution to DeepSeek OCR and agent-side Markdown reconstruction. See [UPSTREAM.md](UPSTREAM.md) and [LICENSE.txt](LICENSE.txt).