DSHEssentialTools

DET · dsh-essential-tools — the plugin that turns DeepSeek Harness into a usable engineering cockpit. Let the AI drive your own logged-in browser · Branch, edit and retry any conversation · See a picture precisely, not approximately · Pay only for the tools this turn actually uses

Quick start · ALV · ADU · Feature map · 用户指导手册 · Changelog · Security

中文一句话：DET 是 DSH 的永久插件 —— 让 AI 开你自己已登录的浏览器、任意消息开分支改问、跑/快照/回退真工程、管所有插件与每一笔 token；2.10 起再加两件事：ALV 给图片造坐标系（栅格·辅助线·边线坐标·本机 OCR），ADU 让工具按需装载（启动只装 5 个元工具）。

😤 The problem

If you hack on real code with an AI agent in the loop, you already know these five:

Pain · What everyone does today · What DET does 🧠 One wrong turn kills the thread — you want to re-ask from message #12, not start over · Copy-paste into a fresh chat and lose all context · VTD: edit / retry any message → a real hidden branch, switch back with `, and the **workspace code rolls back with you** 🌐 **The AI can't see or click what you see** — your logged-in dashboards, internal pages, file://` pages · Screenshot, paste, hand-hold, retry · DET Browser Extension: the model drives your already-logged-in Edge/Chrome through a localhost bridge you gate yourself (off / read / write / on) 👁 "I think the number is 41.00 but the crop is blurry" — vision models guess coordinates · Re-ask, re-screenshot, hope · ALV: grids, rulers, guide lines, edge positions and on-device OCR with per-word boxes — *then* let the model look 🛠 Build-run-look-check is a manual loop · Alt-Tab to the IDE and terminal 40 times an hour · Right toolbar: ▶ run · 🗎 file tree + preview/edit · 🕘 program snapshots & rollback 💸 You have no idea what a session costs — or which tools cost you context · Find out at the end of the month · Live balance + per-model price + per-turn cost, plus ADU: only the tools this turn needs are in the schema budget

DET (the plugin dsh-essential-tools) is the permanent DeepSeek Harness plugin that fixes all five — and adds a full plugin manager, a conversation-memory layer, and a security audit on top. It lives in your web profile, survives restarts, and shows up in Settings → Plugin inventory.

🎁 Zero config: without project paths, every conversation / token / plugin / image feature works out of the box. Add paths only when you want ▶ run.

🆕 What changed in 2.10 / 2.11

Three answers to three long-standing problems:

Problem · Answer A vision model looks at *a picture*, not *a coordinate system* · ALV — 8 tools that build the scaffolding: grid, ruler, guide lines, edge positions, on-device OCR, tiling, annotation, enhancement Every turn paid for 30+ tool schemas while using one or two · ADU — boot with 5 meta tools, mount modules on demand (adu_list → adu_describe → adu_load) Failures only produced one word: browser-timeout · Per-command budgets + structured failures + a failure ledger (_diagPush → detHealth panel)

A real ALV round trip, on a table screenshot:

alv_info   → 900×600, RGB, mean 241, dark 3.9%
alv_edges  → horizontalPositions: y=81(4) y=141(4) y=201(4) y=261(4) y=321(4) y=381(4)
alv_grid   → cell=60, guides drawn at y=470/540, x=430 → returns gridded copy + toSource mapping
alv_ocr    → 型号数量 / ALV-100 12 18.50 222.00 / … / TOTAL 471.00   (with per-line boxes)
alv_annotate→ 2 boxes + 1 cross drawn at exact coordinates, saved and re-read

🔍 ALV — seeing a picture precisely

ALV is geometry, not a better model: no ML, no network, no API key. You (or the model) state coordinates and thresholds; ALV draws and measures exactly that. Everything is offline; OCR uses the OS engine (Windows.Media.Ocr) so results are reproducible and carry word boxes.

Tool · What it gives you alv_info · Size / mode / DPI / brightness / dark-pixel ratio / 16-bin histogram (optional noise). Decide *whether* to crop, upscale or binarize before wasting a look. alv_grid · The coordinate system itself: grid + ruler labels in the margin + guide lines at any position (guides:{horizontal:[470],vertical:[430]}). Returns a new image and the toSource mapping. alv_crop · Exact box crop, optional upscale (2–4×) for small print; omit box to auto-trim uniform borders. Returns the mapping back to the original. alv_edges · Hardcoded edge-position extraction: Canny + probabilistic Hough → line segments *and* merged row/column positions (horizontalPositions:[{y,votes}]), plus an annotated copy labelled y= / x=. This is how you answer "which row line is where" without eyeballing. alv_ocr · Hardcoded OCR: offline, word boxes. mode=lines merges glyphs into visual lines and de-spaces CJK (型 号 数 量 → 型号数量); mode=cells OCRs cell by cell and returns a table matrix — feed it the positions from alv_edges. alv_tile · Overlapping tiles + upscale, for images with too much text to read in one look. alv_annotate · Draw boxes / lines / arrows / crosses / circles / text at exact coordinates — show the user *where you think it is*, then verify by looking. alv_enhance · Gray, auto-contrast, adaptive threshold, unsharp, upscale (auto = all of them). A blurry scan becomes legible before OCR.

Engine ships in the package: py/alv.py (Pillow + numpy + OpenCV on the python312 environment) and py/alv_ocr.ps1 (WinRT bridge, pure-ASCII source, UTF-8 JSON out). ALV also injects a system-prompt section telling the model: *don't guess coordinates — build the scaffolding, then look again.*

🧩 ADU — load once, remember on demand

Registering every tool up front means every request carries every schema. ADU makes loading an explicit protocol — but loading is sticky: ADU never unloads anything on its own, not even across restarts.

boot        → only the 6 adu_* meta tools are registered
adu_list    → module table: which tools each module brings, its cost/risk, whether it is mounted
adu_describe id=alv  → sub-capabilities, per-tool purpose, side effects, when to use it
adu_load    id=alv   → mount (sticky: it stays mounted, and comes back after a restart)
adu_rememory         → re-inject the tool index into context (after compaction, when you forgot what you have)
adu_unload  id=alv   → unmount — rarely needed, and usually not what you want

Module · Tools it brings · Notes plugins · 19 · global plugin library (5 tiers), GitHub install ×2, CDM, TCT, MDA, MMS — the widest permission surface browser · 1 · det_browser; read / write / run, gated by the extension's tier web · 5 · web_inspect / web_act / web_focus / web_human_search / web_insite_search dlt · 6 · env table, run, build, PDF/Word/Excel/CSV read-write-convert alv · 8 · the image suite above security, project · — · always-on listeners / endpoints, not mount units

Why adu_rememory exists

Long conversations get compacted. Tool *schemas* survive, but the model's working memory of "which tool does what, with which parameters" gets summarized away — and re‑mounting does not help, because nothing was unmounted. adu_rememory returns a compact index (one line per tool: name(params) — what it does) built from the same table that drives the module list, so the model can pick up exactly where it left off without paying schema budget twice.

工具记忆（ADU 按需模式 · 2 组 · 14 个工具）

## alv · ALV 辅助线视觉
   alv_grid(path, cell?, majorEvery?, guides?, out?) — 加栅格+标尺+辅助线，建立像素坐标系
   alv_edges(path, box?, low?, high?, minLineLen?, mergeTol?, scale?) — 边线提取：线段 + 合并后的行/列位置
   …

Four rules keep it honest:

1. The module table is the single source of truth (lib/adu.js); the meta tools only validate and report — they hold no state.
2. Mounting reuses the very same disposer pipeline as the master switch and the UI switches, so there is never a "second state". Bookkeeping lives in _aduGroups (id → [disposer]).
3. Sticky by design: mounts persist to det.adu.mounted and are restored at boot. Only adu_unload, the master switch, or unloading the plugin removes them.
4. Turning ADU off is the escape hatch: flip ADU 按需装载 off (or set aduBoot: "eager") and you are back to load-everything at boot — applied immediately, no restart.

🗺 Feature map

Settings → DET 管理器 groups the switches by concern. Two new groups in 2.10 (identifier in parentheses):

Group · Switches 界面注入 · 文件视图 (file) · 运行按钮 (run) · 版本控制 (ver) 对话与版本 · VTD (vtd) — branch / edit / retry + message micro-versions 分组与插件 · MDA 分组 (mda) · 插件管理 (plugins) 成本优化 · MMS 混合模型 (mms) DLT 集成 · 文档工具 · 右栏预览 · 环境表 · 执行与编译 · 每轮成本 · 草图导入 (dltDocs … dltDraft) 识图与工具预算 🆕 · ALV 辅助线视觉 (alv) · ADU 按需装载 (adu) 安全与审批 · 代码修改审批 · AI 命令审计 · Prompt 攻击防御 (approve / secCmdAudit / secPromptDefense)

Plus the master switch (master): off = stock DSH, with only this settings page (and the switch) left behind. Every switch takes effect immediately, persists in ~/.dsh/storages, and unloads its UI along with its tools.

🚀 Quick start (60 seconds)

Prerequisites — DSH 0.2.1-rc.3+ (Web profile), Node 22+, Windows for the browser bridge / Office / OCR paths.

# Option A — installer from the clone (recommended: installs + registers the plugin)
git clone https://github.com/LLYlab/DSHEssentialTools.git
cd DSHEssentialTools
.\install.ps1

# Option B — manual, exactly equivalent
#   1) junction the package into the profile's node_modules
#   2) add an `insert` block for id `dsh-essential-tools` in
#      ~/.dsh/profiles/web/cordis.patch.yml  (a bundle-layer install does this for you)
#   3) restart DSH, then check Settings → Plugin inventory

# Option C — straight from GitHub (no npm needed)
dsh plugin --profile web add github:LLYlab/DSHEssentialTools

Restart DSH, open Settings → DET 管理器, and you should see the switch groups, the DET toolbar on the right edge, and (in ADU mode) five adu_* tools in the model's tool list.

Browser control needs two things: install the MV3 extension from browser-extension/, then set 网络权限 → 使用用户浏览器 and pick a tier in the extension popup (off / read / write / on).

🏗 Architecture

┌──────────────────────── DSH host (Node) ────────────────────────┐
 cordis.patch.yml ──▶ │  lib/index.js  EssentialToolsService (typert Remote)             │
   (bundle layer)     │    ├─ VtdStore        conversation tree / micro-versions         │
                      │    ├─ GlobalPluginStore  5-tier plugin library + source stores   │
                      │    ├─ MdaStore        grouping / model cooperation               │
                      │    ├─ BrowserBridge   WS 127.0.0.1:9123 ◀── MV3 extension         │
                      │    ├─ DltBridge       6 dlt_* tools → py/dlt_docs.py             │
                      │    ├─ AlvBridge       8 alv_* tools → py/alv.py + alv_ocr.ps1    │
                      │    └─ AduBridge       5 adu_* meta tools → _aduGroups bookkeeping │
                      │       adapt.js (service names) · host.js (capability probe)      │
                      └───────────────────────────────┬─────────────────────────────────┘
                                                      │ typert endpoints (~80 methods)
                      ┌───────────────────────────────▼─────────────────────────────────┐
                      │  lib/client.js  (web bundle, no relative imports)               │
                      │   toolbar · VTD tabs · sidebar registry · status box · panels   │
                      └─────────────────────────────────────────────────────────────────┘

Design rules that this repo actually enforces with tests:

• One table, one owner. Feature keys live in normalizeFeatures; their labels/grouping live in DET_FEATURE_GROUPS; the client-structure suite fails if they drift apart (每个功能键恰好归组一次).
• Never break the whole for one part. Every registration goes through safe(...) + disposers; a single failing sub-capability cannot take the plugin down.
• Facts vs. intent. globallyEnabled is an *intent* record; the loader is the *fact*. They are reported side by side and never overwrite each other.
• Loadable modules are not a second state. ADU mounts through the same disposers the master switch uses.
• Diagnostics never throw. _diagPush and every probe is try/catch-wrapped; a diagnostic bug must not become an outage.

🔒 Permission & safety model

Layer · Control Network · 5 tiers (off → official API → fetch → use your browser). DET gates its own requests; the tier is also written into the model's system prompt. Browser · Extension-side tiers off / read / write / on, enforced in the extension (DSH is read-only towards it). Write actions are refused in read mode; write mode strips page content from replies; run needs on. Code changes · approve requires a diff preview + approval before saving or rolling back. Audit · secCmdAudit runs an independent model audit on tool calls; secPromptDefense screens inputs (including command output) for injected instructions. Global plugins · 5 tiers — always / ai-auto / ai-approve / frozen / disabled; frozen and disabled refuse every new enable, including the user's. Global plugin code runs with the host's real permissions — that is why tier control and the code scanner exist.

🩺 Troubleshooting

Symptom · Cause · Fix browser-timeout · Page too heavy for the command budget, or the extension is stuck · Retry with a larger timeoutMs; prefer read_text over screenshot; web_focus the tab first (background tabs are throttled by the browser) browser-disconnected · Extension socket dropped mid-command · Reload the extension, then retry — DET now fails fast instead of waiting 45 s ext-mode-off / ext-read-denied-in-write-mode · Extension tier · Open the extension popup and raise the tier Tools missing after a restart · ADU is in lazy mode · adu_list → adu_load , or turn ADU 按需装载 off in DET 管理器 alv_ocr returns empty text · Image too small / low contrast · alv_crop scale=2..3 then alv_enhance mode=auto, then OCR again ALV errors with "引擎不存在" · Package tree incomplete · Reinstall; lib/alv.js resolves py/alv.py relative to the package root A permanent plugin "comes back" after restart · Intent vs. layer mismatch · Disable it in the patch layer (disabled: true) — the runtime flag alone is re-applied by the next composition Everything vanished · Master switch off · Settings → DET 管理器 → 总开关

The 健康与失败诊断 panel (Settings) shows boot health, plugin-library intent↔fact mismatches, DLT/ALV/ADU mount facts, and the recent failure ledger.

🧪 Tests

11 suites, 631 assertions, no profile / no network / no server required:

node tests\unit.test.mjs              # pure functions + Config neutrality      128
node tests\host.test.mjs              # real host-half mount (fake Cordis ctx)   60
node tests\adapt.test.mjs             # service access layer + architecture guard 23
node tests\client-structure.test.mjs  # information architecture + token guard   47
node tests\client-load.test.mjs       # executes the client bundle for real      17
node tests\dlt-host.test.mjs          # DLT integration: tools/prompt/endpoints  37
node tests\dlt-integration.test.mjs   # client mounting matrix                    22
node tests\draft.test.mjs             # sketch draft: live model judgement        40
node tests\gp-state.test.mjs          # permanent-plugin intent model + guards    71
node tests\pause.test.mjs             # master-switch pause/resume scope          10
node tests\alv-adu.test.mjs           # ALV/ADU + sticky/warm/self-heal      176

Run them from inside the package (ESM resolves the real @deepseek-ai/* peers there). Exit code 0 = all green.

📚 Docs

• docs/GUIDE.md — 用户指导手册：逐功能详解、权限模型、扩展装载、FAQ
• docs/SECURITY.md — 威胁模型与边界
• docs/DET功能.md — 功能总览
• ARCHITECTURE.md — 内部结构、约定与测试矩阵
• CHANGELOG.md — 每一版的「为什么」

🤝 Contributing

Issues and PRs welcome. Two house rules: every new switch must be grouped (the structure test will fail otherwise), and every claim in the UI must have a fact behind it — if the panel says "applied", something must have read it back.

License

MIT — see LICENSE.