dsh-quick-screenshot

English | 中文

One-click screenshots of the DeepSeek Harness window, placed straight into the composer draft.

It is one self-contained DSH bundle: a Host half that captures the DSH window (or a window / the whole desktop) through the platform's own tooling and serves the PNG over one authenticated Fetch route, and a hand-written Client half served as the package's ./client bundle.

What it adds

Seat	Result
conversation.input.left	A camera control in the composer's left tool row, next to +

Gesture	Capture target
Click	The DeepSeek Harness window (largest visible window whose title contains DeepSeek Harness, falling back to the foreground window)
Shift+click	The whole desktop, every monitor
Alt+click	Whatever window is currently in the foreground
Ctrl+Shift+S	The same as a plain click, for the composer that currently holds the keyboard focus

Every gesture ends in the same place: the PNG becomes a draft attachment in the composer, so the next message carries the image. A click never steals the keyboard — mousedown is prevented, so the caret stays where it was.

The control reports what happened in place: a spinner while the Host captures, 已放入输入框 with the captured pixel size for a few seconds afterwards, or the Host's own refusal text when the capture failed. The tooltip carries the full gesture list and the last result.

How the image reaches the composer

The Client half follows the path the composer's own paste and drop intake take:

1. POST /api/quick-screenshot returns the PNG bytes (and a small metadata header).
2. ctx.conversation.createDrafts(sessionId, [file]) registers the browser file and returns its draft id.
3. The inputActions standard prop admits those ids with addAttachments(ids) — the composer's rail then shows the thumbnail and the send serializes the image with the message.

When those services are unavailable, the client falls back to dispatching a synthetic paste event carrying the file on [data-composer-input="true"], which is exactly what a manual paste does.

Host route

GET /api/quick-screenshot describes what this Host can capture right now (supported, platform, tool, modes, directory); POST /api/quick-screenshot takes an optional JSON body and answers with the PNG:

{ "mode": "window", "titlePattern": "DeepSeek Harness", "maxDimension": 0 }

Field	Meaning
mode	window (default), active, or screen
titlePattern	Window title matched by window mode; defaults to DeepSeek Harness
maxDimension	Optional cap on the longest side; 0 keeps the native capture size

The response is image/png with Cache-Control: no-store and one X-Screenshot header carrying percent-encoded JSON metadata (mode, method, window, width, height, bytes). Failures answer JSON with a stable code (CAPTURE_UNSUPPORTED, CAPTURE_TIMEOUT, CAPTURE_FAILED, METHOD_NOT_ALLOWED) and a message meant for the user. Both methods are registered through ctx.connection.fetch, so the shared /api bridge applies Host/Origin validation and browser-cookie authentication before the handler runs.

Capture backends

Platform	Tool	Notes
Windows	powershell.exe + System.Drawing	PrintWindow(PW_RENDERFULLCONTENT) first, so an occluded window still captures; a flat result falls back to reading the same rectangle off the screen after raising the window. Per-monitor-v2 DPI aware.
macOS	screencapture -x -o	Display capture; window/active both resolve to the active display.
Linux	first of gnome-screenshot, spectacle, grim, import, scrot	Display capture.

DSH_SCREENSHOT_POWERSHELL overrides the Windows shell (useful when only PowerShell 7 is installed).

Captures are staged in <tmpdir>/dsh-screenshot/, and the directory keeps only its newest 24 files, so the plugin owns no durable state and never grows unbounded. One capture at a time: concurrent clicks queue instead of racing.

Install

Run the DSH Plugin Manager's install_bundle against this directory (a local path spec), or add it to the profile's dsh.profile.bundles after installing it with pnpm. The bundle's patch inserts one row, id: screenshot, naming this package.

From a GitHub checkout, no build step is involved:

dsh plugin --profile web add github:JackEngineer/dsh-quick-screenshot

Publishing to the plugin market

The market lists plugins from the curated awesome-dsh-plugin list, not from its own repository — publishing means opening one entry-file PR there. PUBLISHING.md has the whole path (naming conflicts, the required dsh.bundle manifest, the dsh-plugin topic, the 1-day repository age rule, the entry YAML, npm/tarball options, and the CI checklist). screenshots.json already declares this plugin's storefront screenshots from inside the repository.

The Host half loads with the new Loader row. Changing lib/index.js afterwards needs a Host restart, because the Loader keeps Node's module cache for the life of the process; lib/client.js re-reads from disk on a row toggle (bundle bytes are revisioned by mtime, ctime and size).

Tests

Both suites run offline and need no browser:

node scripts/selftest.mjs        # the Host route and real captures
node scripts/client-smoke.mjs    # the Client bundle, control and attach flow

scripts/selftest.mjs drives the real route with Fetch Request/Response objects: capabilities, method rejection, a real window capture through PowerShell (PNG signature, IHDR size, metadata header, content length), a real capped screen capture, the failure mapping for a missing capture shell, and staging-directory pruning. Real captures run where a desktop session and a capture backend exist; elsewhere they are reported as skipped, or skipped on purpose with DSH_SCREENSHOT_SKIP_CAPTURE=1.

.github/workflows/test.yml runs both suites on every push and pull request with that flag set, so the repository keeps a green check on a machine that has no desktop.

scripts/client-smoke.mjs evaluates the bundle against a stub module loader, a small React stub that runs effects, and fake composer services: it asserts what apply registers, then drives captureInto through the admit path, a locked composer, a Host refusal, an empty image, and the paste fallback, and finally mounts the control and fires the shared Ctrl+Shift+S chord.

Known limitations

• Window capture excludes overlapping windows but not popups. PrintWindow renders the window's own content; an open context menu drawn in a separate popup window is not part of it.
• The screen rectangle fallback needs the window on top. It is only used when PrintWindow returns a flat image, and it raises the target window first.
• macOS and Linux capture the display, not one chosen window.
• Images need an image-capable model. A text-only model refuses the attachment at send time with the composer's own error banner.
• The chord is a web chord. A browser that reserves Ctrl+Shift+S (Save as) absorbs it before the page sees it; the button always works.
• The Electron desktop shell is served over http://127.0.0.1:<port>, so the Client half reaches /api/quick-screenshot as a same-origin request there as well.