Model-agnostic visual-inspection pipeline for text-only agents: page-by-page HTML screenshots plus a ready-made vision-subagent briefing contract (gewu_prep), then source-code truth verification of every finding (gewu_locate); validated on mimo-v2.5 & qwen3.7-plus.
Install
# from npm (prebuilt)
dsh plugin --profile web add gewu-tools
# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)
dsh plugin --profile web add github:nyantused-cpun/gewu-tools
Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).
README
The Gewu Inspection Surface: let text-only models in DSH complete visual-inspection tasks through vision subagents.
格物 (Gewu): from the Great Learning — "investigate things to attain knowledge." Look carefully first, then conclude. 中文: README.md
Quick install (npm)
npm i gewu-tools
pwsh node_modules\gewu-tools\scripts\install-gewu-plugins.ps1 -Install # idempotent mount into DSH
# Or install directly through the DSH plugin manager (requires the dsh.bundle manifest)
dsh plugin add gewu-tools
# Restart the DSH session; gewu_prep / gewu_locate appear in the tool catalog. -Verify = 4 greens.
# Optional: inject a custom preset (defaults to community, zero injection; GEWU_BRIEF_PRESET also works)
pwsh gewu\scripts\install-gewu-plugins.ps1 -Install -Preset <preset-file-path>
Problem
The DSH session brain is usually a text-only model — it cannot see rendered pages, so it cannot answer "does this proposal page look good / crowded / misaligned?" Vision subagents (e.g. subagent_vision, mimo-v2.5, qwen3.7-plus) can see, but have two weaknesses:
- No session context: they don't know client ground rules or business-marker semantics (e.g. ⭐ = "core capability", not decoration), so intentional design gets reported as defects;
- They misread: OCR-level character errors, and page attribution drifted by screenshot viewport boundaries.
gewu-tools packages a three-step pipeline (validated on a 24-page proposal, 2026-08-16) into two tools:
① gewu_prep screenshot + briefing: renders an HTML document page-by-page and generates a
complete vision-subagent briefing contract (discipline / artifact / attribution rules)
② vision subagent dispatch: the main agent calls DSH's native vision subagent with the briefing
③ gewu_locate truth verification: locates a text/number from the review back into the HTML source
(page section id + line number); count=0 means OCR misread — do not trust it
Model-agnostic statement
Validated on vision subagents mimo-v2.5 (conservative) and qwen3.7-plus (more insightful, slightly more false positives). Finding: model tier does not determine the trustworthiness of review results —
- The briefing contract (context) eliminates context-free false positives: e.g. reporting business markers like ⭐ (core capability) as decorative emoji — observed systematically in the no-briefing plugin mode during testing;
- The truth-verification loop (
gewu_locate) blocks OCR misreads and page-attribution drift before acceptance: in testing, 2.5 of qwen3.7-plus's 3 core new claims failed source-code verification — all were caught by the verification step.
Tier differences only show up in review style (higher-tier = sharper insight, steadier = more conservative conclusions), not in correctness. With any vision model, the pipeline output stays trustworthy and auditable — which is exactly why verification is hard-wired as the third step.
Install
Prereqs: Windows + Chrome/Edge + Node 24+ (for the verify script only).
# Direct install through the DSH plugin manager (requires the dsh.bundle manifest)
dsh plugin add gewu-tools
# Or use the bundled installer for local junction mounting and custom presets
pwsh gewu\scripts\install-gewu-plugins.ps1 -Install # idempotent
pwsh gewu\scripts\install-gewu-plugins.ps1 -Verify # verify only
pwsh gewu\scripts\install-gewu-plugins.ps1 -Uninstall # remove
pwsh gewu\scripts\install-gewu-plugins.ps1 -Install -Preset <preset-file-path> # optional custom preset
-Preset injects a custom briefing preset. If omitted, the installer uses community (zero injection); you can also set the GEWU_BRIEF_PRESET environment variable instead.
The script appends a gewu-tools mount row to the DSH agent preset (default pre-sales, override with env GEWU_PRESET), pointing at <workspace>\.dsh\gewu-tools (a Junction to this project).
Restart the DSH session after install (host plugins have no HMR). Verification: install-gewu-plugins.ps1 -Verify expects 4 greens (preset mount row / junction chain / import load / preset loading).
Usage (main-agent view)
1. gewu_prep(html_path="...", client="<client>", background="<ground rules / marker semantics>")
→ { pages: [...], briefing: "..." }
2. Call a vision subagent with the briefing as its prompt → per-page issue list + conclusions
3. gewu_locate(html_path=..., needle=<text/number from the findings>) for every key finding
Custom preset injection
A preset is gewu-tools' extension point: put your own design standards or acceptance criteria into a preset file, and gewu_prep injects them into the briefing so the vision subagent reviews against the same yardstick. The default community preset injects nothing and keeps the 1.0.0 community baseline.
| Field | Type | Description |
|---|---|---|
name |
string | Non-empty preset name. |
version |
string | Non-empty preset version. |
defaultFocus |
string | null | Overrides the default focus list; pass null to keep the built-in default. |
sections |
string[] | Sections injected into the briefing (design standard / acceptance criteria); each section ≤ 2,000 characters, total ≤ 8,000 characters. |
dispatchHints |
string | null | Dispatch hints (e.g. how many pages to review per pass); nullable. |
requires |
object | Reserved extension dependency declaration; an invalid object causes degradation. |
Minimal example brief-preset.js:
export default {
name: "my-design-standard",
version: "1.0.0",
defaultFocus: "layout consistency, information density, chart readability",
sections: [
"Visual standard: primary color #123456, left-aligned titles, 24px card spacing.",
"Acceptance criteria: each page must have one core claim; key data must match business-marker semantics."
],
dispatchHints: null,
requires: {}
};
If a preset file is invalid (not an object, missing name/version, wrong field types, sections too long, etc.), gewu-tools falls back to community and returns preset_degraded: true with preset_error explaining why.
CLI usage (cross-host)
index.js works without DSH. Non-DSH hosts (Trae/Codex/Kimi, etc.) can use this CLI to get the same screenshot + briefing + verification artifacts as the plugin.
# Screenshot + briefing (same source as gewu_prep)
node gewu\index.js --prep <html> [--client <client>] [--background "..."] [--focus "..."] [--out-dir <dir>] [--width <n>] [--height <n>] [--nav-offset <n>]
# Truth verification (same source as gewu_locate)
node gewu\index.js --locate <html> --needle "<text/number>" [--max-matches <n>]
# Smoke alias (old positional form)
node gewu\index.js <html> [outDir]
Artifacts: --prep writes prep_result.json + briefing.txt to --out-dir (default _tmp_vision_test/<file>_shots); --locate writes locate_result.json to _tmp_vision_test/<file>_locate/.
Exit codes: 0 success; 1 runtime/business failure (HTML missing, screenshot failure, etc.); 2 usage error.
How it works
- Screenshots: reads
<section>anchors (prefersv2-pageclass), captures each page via headless Chrome viewport; sticky nav pushed out by iframe offset; falls back to a full-page shot when no anchors exist. Chrome is spawned from the host process, bypassing the restricted sandbox's named-pipe limitation. - Briefing contract (baked into
briefing):- discipline: subagent skips entry protocols, never asks the user, reads images via
read_image(prevents deadlocks); - artifact notice: fixed-viewport screenshots concatenate page boundaries — a next-section header at the bottom is an artifact, not a design flaw; attribution rule: content from the next section belongs to the next page;
- output spec: per-page "page|issue|severity|fix", visible issues only.
- discipline: subagent skips entry protocols, never asks the user, reads images via
- Truth verification: locates each
needleoccurrence in source, returning page section id + line number + tag-stripped snippet.
Test
node gewu\index.js <any HTML with section anchors> [outDir] # standalone smoke, no DSH needed
Directory structure
gewu/
├── index.js # plugin body (cordis + defineTool, standalone CLI entry)
├── package.json # gewu-tools (dsh.bundle manifest)
├── cordis.patch.yml # dsh plugin add mount patch
├── presets/ # built-in community preset and preset contract
├── README.md / README.en.md
├── CHANGELOG.md
├── LICENSE # MIT
├── docs/TESTING.md # testing methodology and results
└── scripts/install-gewu-plugins.ps1 # install script (Install/Verify/Uninstall/DryRun)
Compatibility
- DSH 0.1.0-rc.6 (cordis plugin + defineTool surface)
- Peer dependency:
@deepseek-ai/dsh-toolsonly (resolved via the workspace AiBridge junction) - Browser: auto-detects Chrome/Edge, no config
Related links
The following projects complement the gewu ecosystem.
- Lanting Folio: deep-context review engine
- dsh-vision-toolkit: general-purpose vision toolbox; complementary to this project — it lets models see, gewu makes reviews trustworthy
- awesome-dsh-plugin
License
MIT — see LICENSE. Version 1.1.0 (2026-08-17).
Links
More in this category
Tencent/WeKnora#dsh-weknora★ 37587
Four read-only tools over a WeKnora knowledge base: list knowledge bases, hybrid passage search, reassemble one document's chunks in order, and WeKnora's own cited RAG or ReAct-agent answer with a resumable session id.
DayuanJiang/next-ai-draw-io#dsh-plugin★ 37004
Draws draw.io diagrams in DeepSeek Harness through the next-ai-drawio MCP server: the link in the reply opens a live preview in the right sidebar, you can edit the diagram by hand, and each AI change is a version you can restore. Node.js 18+; npx downloads the MCP server on first use.
superdesigndev/treg★ 8181
Tool catalog for agents: search ~2,600 external endpoints (SEO and SERP, backlinks, social, people and company enrichment, ad libraries, scraping) by the task you want done, read each one's parameters and per-call price, then call it with the credential injected server-side. Ships the skill plus an MCP row that stays disabled until TREG_TOKEN is set.
T-Auto/dsh-ops★ 5867
Provides bash, PowerShell 7, and high-performance Rust tools for DSH on Windows to reduce token usage.
sixtysevenlf/dsh-blender-plugin★ 2153
Lets a DSH agent drive a local Blender over a direct TCP channel with 15 tools: viewport frames, custom-angle renders, Python execution, headless and job offload, transaction rollback, and a planning layer of 228 ops across 34 families.
helloHub1212/dsh-windows-mcp★ 1490
Bridges the CursorTouch/Windows-MCP desktop automation server into DSH through the official MCP client bridge, registering its screenshot, UI tree snapshot, mouse and keyboard input, clipboard, process and registry tools as mcp__windows__* tools for the model, and adds a windows_mcp_doctor diagnostic tool. Computer Use is off by default and is switched on from the plugin panel.
Community comments
Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.