captureBeyondViewport was capturing the entire page in one giant image
(2804x20746), making text unreadable. Now captures per-viewport with
80% overlap, producing ~2400x1992 screenshots that vision can read.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
SKILL.md now specifies exact JSON structure the model must output
after reading screenshots. Weight in kg, dimensions in cm, omit nulls.
Ready for downstream API integration.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Remove all CSS selectors, regex parsers, and structured extraction.
Instead, capture full-page screenshots (scrolling) and download detail
images. The model reads these directly with vision to extract logistics
data — no fragile DOM dependencies.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- CDP discovery uses localhost instead of 127.0.0.1 (Chrome binds IPv6)
- Only overwrite logistics fields when parsing succeeds, preventing
later unparseable keys from nullifying valid parsed values
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
register-skill-release / register (push) Successful in 14sDetails
- Change default CDP port from 9222 to 18800
- Extract 包装信息 section (packaging type, box weight/dims, units per box)
- Extract 商品件重尺 table (per-piece weight/dimensions/volume)
- Backfill logistics from pieceWeightSize/packageInfo when attributes missing
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
register-skill-release / register (push) Successful in 15sDetails
The shared-actions/register-skill workflow requires auth-runtime
client-key integration. Restores auth-cli.ts and auth-rt auto-install.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
register-skill-release / register (push) Failing after 24sDetails
Scrapes 1688 product pages via Chrome browser to extract logistics data
(weight, dimensions, volume) from attributes, variants, and detail images.
Zero npm dependencies — uses raw CDP over WebSocket.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>