windows-ocr

DeepSeek Harness (dsh) plugin that lets text-only models accept attached images: every image is recognized locally with the built-in Windows OCR engine (Windows.Media.Ocr) and only the recognized text is sent to the model API. Image bytes never leave your machine.

• No configuration changes to your models — no input: [text, image] hacks in settings.yaml.
• Works with any provider/model in dsh; OCR applies only to text models.
• Genuine vision models (declared image capability) pass images through untouched by default.
• Fail-closed: if the plugin is not loaded, models stay text-only and image attachments are refused — nothing can silently leak.

Quick install via an AI agent

Hand this repository to any AI agent, or paste the instruction below, and the agent will install and verify the plugin for you:

Please install the dsh plugin in this repository by following https://github.com/maxwell-feng/dsh-windows-ocr/blob/main/agents-install.md. Run every preflight check, choose an install mode, then complete the mandatory verification: attach an image to a text-only model session and confirm the model answers with the recognized text.

agents-install.md is a step-by-step guide written for AI agents: preflight checks, both install modes (permanent profile patch / temporary --patch overlay), mandatory functional verification, and troubleshooting for the failure modes you are likely to hit. Manual install instructions are below.

Why a plugin (not a skill)

dsh skills are Markdown instruction files injected into the model context — they cannot execute code, cannot hook the request pipeline, and cannot stop an image from being serialized. This feature needs exactly that, so it is a cordis plugin that hooks two public seams of the llm service:

1. Capability shim — ctx.llm.resolveModelInfo (also listModels). The host gates image attachments on inputModalities.includes("image") at three places: message admission, model switching, and the read_image tool. The shim answers "yes", so text models admit images.
2. Request rewrite — registration.adapter.stream (the single choke point both ctx.llm.stream and prepareCall().stream funnel through). Every image content block is replaced with an OCR text block before the adapter serializes the request, so the adapter's own image check never fires, no attachment bytes are read for the wire, and no image_url is ever built.

you attach an image
  → admission asks ctx.llm.resolveModelInfo (shimmed: "image" ✓)
  → image stored in the local attachment store (session log, UI preview)
  → agent builds the request → adapter.stream (wrapped)
  → image block read locally (ctx.attachments.readImage) → Windows OCR
  → block replaced with <image_ocr>…text…</image_ocr>
  → adapter serializes a text-only request → provider

Requirements

• Windows 10/11 (Windows PowerShell 5.1+ ships with the OS; no install needed)
• A Windows OCR-capable language pack for your language (Settings → Time & language → Language). English is usually present; Chinese requires the Chinese language pack (OCR-capable).
• dsh with a profile (tested against dsh 0.1.0-rc.6)

Install

Installing via an AI agent

agents-install.md in this repository is a step-by-step installation guide written for AI agents (and careful humans). Give it to an agent — e.g. "install this plugin per agents-install.md from https://github.com/maxwell-feng/dsh-windows-ocr" — and the agent can perform the preflight checks, install, verification, and troubleshooting on its own. The guide covers both install modes, the mandatory functional verification (attach an image → model answers with the OCR text), and the failure modes you are likely to hit.

Manual install

Two official ways to load this plugin, both referencing the plugin file by absolute path (see docs/user/develop/basic). On Windows the path must be a file:// URL — a bare C:/... path is parsed as the c: URL scheme and the loader rejects it.

Permanent: profile patch layer

Append to your profile's cordis.patch.yml (e.g. ~/.dsh/profiles/web/cordis.patch.yml):

- insert:
    - id: windows-ocr
      name: 'file:///C:/absolute/path/to/windows-ocr/lib/index.js'
      config:
        language: ''
        passthrough: true

Then restart dsh web. Remove the rows to uninstall — nothing else is touched.

Temporary: --patch overlay

Put the same rows in an overlay file and boot with it; your profile stays untouched:

dsh --profile web --patch C:/path/to/overlay.yml

Notes

• dsh web fails with EADDRINUSE on port 3080 when an older instance is still running: find it with netstat -ano | findstr :3080 and stop it (taskkill /PID /F) before starting a new one.
• For a packaged install (npm / tarball / github:user/repo), package the plugin as a bundle (dsh.bundle + cordis.patch.yml, see docs/user/develop/basic/publish); a git install additionally needs a prepare build script and pnpm allowBuilds consent.

To verify the plugin loaded, look for windows-ocr in the boot logs, or check the OCR smoke test below.

Configuration

All settings live in the patch row windows-ocr (cordis.patch.yml here) and can be overridden from your profile's cordis.patch.yml:

Key	Default	Meaning
language	""	BCP-47 tag for Windows OCR, e.g. zh-Hans, en-US. Empty = user profile languages.
passthrough	true	true: genuine vision models receive images untouched; false: OCR everything.
ocrScript	bundled lib/ocr.ps1	Absolute path override for the PowerShell OCR script.
timeoutMs	60000	Per-image OCR timeout.
maxCacheEntries	200	Bound on the per-run OCR cache (keyed by attachment id).

Example override in ~/.dsh/profiles/web/cordis.patch.yml:

- update:
    - id: windows-ocr
      config:
        language: zh-Hans

How the model sees the image

Each image block becomes a text block:

<image_ocr name="photo.png">
…recognized lines…
</image_ocr>

Recognition text is cached per attachment id for the lifetime of the dsh process, so repeated turns do not re-run OCR.

Temp-file hygiene

Every OCR run writes its input image and output text into a fresh temporary directory (windows-ocr-* under the system temp dir). The directory is removed automatically in finally — on success, on OCR error, and on timeout — so no per-run script, image, or output file survives. At plugin start, any orphaned windows-ocr-* directories left behind by a previously crashed process are swept as well. Nothing is written outside the plugin's own temporary directory and the dsh attachment store.

Smoke test (no dsh needed)

# 1x1 PNG — exercises WinRT loading, language availability, recognition
powershell.exe -NoProfile -ExecutionPolicy Bypass -File lib/ocr.ps1 -ImagePath test.png -OutFile out.txt
Get-Content out.txt

Exit code 0 with an empty/whitespace out.txt means the OCR engine works (a 1×1 image has no text). Exit 2/3 means a language pack is missing.

Verification inside dsh

1. Attach an image to a text-model session and send a message — the model should answer using the recognized text.
2. Confirm the image never goes out: open DevTools → Network in the web UI, inspect the request to your provider base URL, and verify the payload contains only text content parts (no image_url / data URI).

Limitations

• OCR language availability depends on installed Windows language packs (script exits 2/3 and the plugin degrades to a placeholder text).
• GIFs: Windows OCR recognizes the first frame.
• Cache is per process; a long-lived session keeps OCR text cached, bounded by maxCacheEntries.
• Hot reload (HMR) replaces adapters; the plugin re-wraps new adapters on llm/adapters-updated, but a full restart is the safe path after any dsh update.
• The model picker may show text models without an "image" badge (cosmetic only — listModels is shimmed consistently).
• If the OCR plugin is removed, image attachments to text models are refused again (fail-closed), not uploaded.

License

MIT