dsh-ocr

dsh-ocr (alpha.1)

The agent can finally read a picture. Core DSH has no OCR: read_image returns the picture to a model that has eyes, and a text-only model gets nothing from it, while read refuses a binary file. The PDF row's OCR needs tesseract and a rasterizer installed. So a screenshot, a photo of a page, a scanned TIFF - or a scanned PDF on a machine with neither - was unreadable. This is one host row with one tool, ocr, and it works on a stock Windows machine with nothing installed at all, because Windows has shipped a real OCR engine since Windows 10.

What it adds

• One tool, ocr - the path to a file, and the text comes back. lang accepts either spelling (eng/por for tesseract, en-US/pt-PT for Windows, eng+por for two at once); pages chooses a PDF's pages ("3", "1-5", "1,4-6", default every page up to 20); psm is tesseract's page-segmentation mode for an odd layout; dpi (50-400) is the resolution a PDF page is drawn at before it is read.
• Two engines, chosen by capability, never installed. tesseract from PATH on any platform - preferred for an image, because it is explicit about language and segmentation, and it is not offered a PDF, since a PDF is not a raster image and dsh-pdf already owns that route (pdf_scan rasterizes with poppler/mutool/Ghostscript and calls tesseract). And on Windows, Windows.Media.Ocr, which needs no install and can also rasterize a PDF page through Windows.Data.Pdf - so a scanned PDF is readable on a machine with no tesseract, no poppler and no Ghostscript.
• The answer says what actually read the file: the engine, the language tag it used, the pixel size, how many pages, and how long it took. A language this machine has no data for is refused with the list of what it can read.
• It is honest about what OCR is. Every answer that contains text carries the transcription caveat and points at the tool that extracts text instead (pdf_read for a PDF with a text layer, read for anything textual). An engine that found nothing says the page may be blank or the type too small, and offers dpi.
• A mistake is named as one. A page number past the end is "does not exist in this 3-page document", not "no text found"; pages on an image is "that is a single image, drop pages"; a text file is refused by what it is - its own first bytes, not its extension.

How it plugs in

Piece	Value
id / kind	dsh-ocr / ocr
rows	one: ocr (host-only - no dsh.client, no route, no tab)
injects	tools
engines	tesseract from PATH; lib/win-ocr.ps1 through powershell.exe (5.1) on Windows
deps	none - zero npm dependencies, no vendored engine, no download, no network egress

The tool's path policy is the pack's own, duplicated because a bundle may not reach into another bundle's files: a relative path resolves inside the conversation workspace through realpath (a symlink out is refused), an absolute path is read as given - which is the door a chat attachment under $DSH_HOME/attachments/v1/files/... and a screenshot in Downloads come through.

Rules and limits

• lib/win-ocr.ps1 is pure ASCII, and that is load-bearing. Windows PowerShell 5.1 reads a .ps1 with no BOM as ANSI, so one accented character in a literal there would be decoded as two wrong ones before the script ran. Everything that comes back from the engine - accented text, a translated error - is written to a file as UTF-8, never echoed to a console whose code page is not ours; stdout carries exactly one compact JSON verdict.
• PowerShell 5.1, not 7, on Windows. PowerShell 7 cannot project the Windows.Media.Ocr type without an extra SDK assembly; the 5.1 inside Windows has the projection built in. DSH_OCR_POWERSHELL overrides the path.
• The engine's own limits are respected, not tripped. A page is rasterized at the requested dpi unless that would exceed OcrEngine.MaxImageDimension (10000 px on the machine this was built on), in which case it is scaled to fit; an oversized image is decoded straight to Bgra8 with its EXIF orientation respected. Lines are read from OcrResult.Lines, never OcrResult.Text - the flat text joins an invoice's lines into one sentence, which is how layout is lost.
• Recognized text comes back whole or says it was cut, at 200 000 characters, and one call recognizes at most 20 pages.
• It reads raster images and PDFs only: PNG, JPEG, GIF, BMP, TIFF, WebP, and HEIC/AVIF where Windows has the codec. A HEIC is offered to the Windows engine alone, because what reads one is a codec and not a parser. A multi-page TIFF is read as its first page by the Windows engine, and the count is read out of the TIFF's own IFD chain so the answer says so - tesseract reads every page of one, and the two would otherwise be indistinguishable.
• Not claimed: office documents, spreadsheets or archives (convert to an image or a PDF first), and handwriting, which neither engine is trained for. Nothing here writes to the file it reads, and the engine's temporary directory is removed afterwards.

Verify

node --check packages/dsh-ocr/lib/engines.js
node --check packages/dsh-ocr/lib/tools.js
node --check packages/dsh-ocr/lib/index.js
npm run check            # in packages/dsh-ocr

checks/check-ocr-node.mjs is this package's own tracked check, additional to the pack-wide ones. It drives the pure decisions (language matching across both tag spellings, page specs, content sniffing, the no-engine and blank-page sentences), then the engines end to end on inputs it builds itself: a 24-bit BMP whose text is drawn from a 5x7 bitmap font the check carries, and a minimal three-page PDF - no fixtures, no image library, no dependencies. Then the ocr tool through the real buildTools: a workspace-relative path, an absolute one, a missing file, a text file, pages on an image, and a PDF page range that must not return another page's text. With no usable engine it skips loudly and exits 0, because a green run there is not evidence.

Install

Both installers pick the package up from packages/; it is a new bundle, so the profile learns about it once:

scripts\setup\install.bat -Force        # Windows
./scripts/setup/install.sh -Force       # macOS / Linux

Then restart npx @deepseek-ai/dsh web. On Windows this row works immediately with no install; on macOS and Linux, install tesseract (with the traineddata for the languages to read) and it is found on PATH.