Skip to content

Releases: shinegeer/dsh-vision-window

Release list

v1.2.0

v1.2.0 Latest
Latest

Choose a tag to compare

View all tags
@shinegeer shinegeer released this 15 Aug 02:24
v1.2.0
9e16f5a

Seven zero-config local pixel/OCR tools join the existing vision tool: vw_crop, vw_pixel_diff, vw_colors, vw_trace, vw_ocr, vw_html_screenshot, vw_extract_foreground. No provider, no key, no Python required; artifacts land in .dsh-vision-window/artifacts (configurable artifactsDir, master switch localTools).

  • vw_pixel_diff: diff ratio + worst 8×8 grid + red heatmap PNG + JSON report; size mismatches are resized and noted.
  • vw_trace: hard 1 MP input cap, enforced even when downscale is off.
  • vw_html_screenshot: system Chrome/Edge via puppeteer-core; all non-file: requests are blocked.
  • vw_ocr: system tesseract with clear install guidance when missing.
  • Bilingual README: English default + 中文 (README.zh.md), with a DSH Web demo screenshot.

Verified:

  • npm run check and 23 tests pass.
  • Real-machine headless zero-config: vw_colors succeeded under a preset without any credential; settings restored unchanged.
  • Real-machine vw_pixel_diff / vw_crop / vw_ocr error guidance / vw_html_screenshot (1200×720).
  • Temporary web profile confirmed the client bundle contains vw_* / localTools / artifactsDir.

Install:

dsh plugin --profile web add github:shinegeer/dsh-vision-window
dsh plugin --profile headless add github:shinegeer/dsh-vision-window

Restart dsh web after installation.

npm audit --omit=dev reports five moderate findings from a single phin advisory (transitive via potrace → jimp). The affected URL-loader path is never exercised: the plugin passes only locally validated buffers to potrace.

Assets 2
Loading

v1.1.0

v1.1.0

Choose a tag to compare

View all tags
@shinegeer shinegeer released this 15 Aug 02:24
v1.1.0
686cec0

v1.1 turns the plugin into a self-contained, credential-safe vision stack:

  • DSH credentials: config stores only reference names (OPENCODE_API_KEY / XIAOMI_API_KEY); values are resolved per call and never logged or returned to the browser.
  • Provider presets: OpenCode Go, OpenCode Zen (free), Xiaomi MiMo, plus a custom OpenAI-compatible provider.
  • Automatic failover with classified errors and 429 Retry-After backoff.
  • In-process answer cache keyed by image content hash + question + provider chain.
  • Pre-call sharp downscale, <think> / reasoning stripping, a compact collapsible settings panel, and headless profile support.

Install:

dsh plugin --profile web add github:shinegeer/dsh-vision-window
dsh plugin --profile headless add github:shinegeer/dsh-vision-window

Restart dsh web after installation.

Loading