gewu-tools

DSH's session brain is usually a text-only model. It cannot read a rendered page, so questions like "does this proposal layout hold together, are the diagram connections correct, is the cross-page density consistent" simply go unanswered. Calling a vision subagent directly does not solve the problem either: the subagent lacks session context and mistakes a business badge such as ⭐ for a decorative emoji, while OCR-level misreads combined with screenshot-boundary page drift can corrupt critical numbers. gewu-tools hardens a three-step pipeline, converged on a 24-page pre-sales deck, into two plug-in tools that close the "can see but cannot trust" gap.

The pipeline runs in three stages: screenshot, briefing contract, and ground-truth verification. gewu_prep pages through the HTML with headless Chrome and, in parallel, assembles a context-rich briefing covering client background, a focus checklist, an artifact disclaimer, and output discipline, so the vision subagent knows that ⭐ means "core capability," not a decorative glyph. This systematically eliminates context-blind false positives. gewu_locate then takes every key figure or claim the subagent reports and locates it line by line in the source, returning the enclosing section, page number, and a tag-stripped excerpt. Claims that fail verification are excluded before they reach the user. Together the three steps drive false positives down, keep critical numbers error-free, and make narrative flow, density, and cross-page consistency auditable.

A custom preset lets you inject your own design conventions and acceptance criteria into the briefing, so the subagent reviews against your yardstick; invalid presets degrade gracefully to the community baseline with a clear error message. index.js also runs standalone outside DSH, letting Trae, Codex, Kimi, or any other host obtain the same screenshot, briefing, and verification artefacts via CLI. Whether you plug in a steady-tier or an insight-tier vision model, the verification step guarantees that the output is trustworthy and auditable—this is precisely why gewu-tools locks verification in as the third step.

Ideal for any DSH workflow that demands "look carefully before concluding": pre-sales layout acceptance, multi-page documentation density and consistency review, and any quality gate that pairs visual inspection with source-level proof.