dsh-toolkit

A DeepSeek Harness enhancement pack — four plug-and-play tools in one: paste-image auto-vision, one-click workspace open plus path copy, a message-history rail, and an archived-sessions panel. One command installs everything.

国内镜像 / Mirror: also hosted on gitee.com/chill109/dsh-toolkit — Gitee is a mainland-China Git host (faster access from mainland China; use it if GitHub is slow).

中文版见 README.zh.md。

Part 1 — If you are a human

This part is written for people who just want everything installed.

What the toolkit does

One install gives you four enhancements:

1. vision-bridge — paste an image, get an answer. Paste any image into the chat box and send it, even with a text-only model: the harness hands the agent the image as a local path and the agent automatically calls its available vision tool (vision_chat from qwen-mm-plugins-api, or any other vision tool you configured) to describe/OCR/answer. No more "current model does not support images".

2. workspace-launcher — open the workspace, or copy its path. Hover a workspace in the sidebar, click "...": Open in File Explorer / Open in VS Code / Open in terminal / Copy absolute path / Copy relative path. Every action shows a result toast.

3. msgrail — a message-history rail. A thin vertical rail on the left of the chat, one bar per message you sent in the current conversation (all of them, loaded or not). Hover a bar for a preview, click to jump to that message.

4. archives — archived sessions, visible again. An "Archived" seat at the sidebar foot lists archived sessions grouped by workspace; click to restore or fork.

Screenshots

Screenshots are taken from the standalone component projects; components without captures show a description only.

vision-bridge — no screenshot available yet. Paste an image into the chat box and send it (even with a text-only model): the agent automatically recognizes it through your configured vision tool and answers.

workspace-launcher — the workspace "..." menu with the five added actions:

msgrail — the message-history rail docked on the left of the chat; hover a bar for a preview, click to jump:

archives — an "Archived" seat at the sidebar foot opens the archived-sessions panel, grouped by workspace; each row shows the title and last-active time, with fork and restore actions:

Requirements

• DeepSeek Harness web UI running (dsh web --host 127.0.0.1 --port 3080), profile layout 0.1.0-rc.6
• git (applies the patches; LF-safe invocation)
• node (msgrail's layout patch)
• vision-bridge additionally needs a vision tool — configure one yourself (e.g. register the qwen-mm-plugins-api MCP server with a working API key); any tool that takes an image path and returns text works. If you did not specify one, ask your AI to recommend a compatible option.

Manual deployment

One-command install

.\install.ps1          # Windows: installs all four components, then restarts the harness

./install.sh           # macOS / Linux: installs all four components

Then hard-refresh the browser (Ctrl+F5) after the harness is back.

Installing components individually

Each component is also usable on its own — see components/<name>/ and the "For AI" part below for the exact per-component steps (patch files, plugin packages, cordis entries).

Let an AI deploy it

Copy/paste this whole paragraph into your AI assistant:

Deploy the dsh-toolkit from https://github.com/chou109/dsh-toolkit (mainland-China mirror: https://gitee.com/chill109/dsh-toolkit) to my local DeepSeek Harness web profile.

1. Windows: run .\install.ps1 in the repo root — it installs all four components (vision-bridge patches to dsh-host-apiproxy + dsh-llm-pi-ai; workspace-launcher plugin + dsh-client-ui-workspace patch; msgrail plugin + dsh-client-ui-layout patch; archives plugin), registers the cordis entries, and restarts the harness. macOS/Linux: ./install.sh (restart dsh web yourself).
2. If the script is unavailable, follow the README "For AI" part, or install components one by one from components/.
3. Note: vision-bridge itself does NOT include a vision model — configure one separately (e.g. mcp-qwen-mm-plugins-api / vision_chat); if I did not specify one, recommend a compatible option.
4. After the harness restarts, verify each component: (a) paste an image and send — the agent must describe it automatically; (b) hover a workspace, open the "..." menu — the five workspace-launcher actions must work (Explorer/VS Code/terminal/copy); (c) the msgrail rail must appear on the left of a conversation with bars for my messages; (d) the "Archived" seat must appear at the sidebar foot.

Usage

• vision-bridge: paste (Ctrl+V) or drag an image into the chat box; type a question if you like; send. The agent recognizes it automatically.
• workspace-launcher: hover a workspace in the sidebar → "..." → pick an action.
• msgrail: open a conversation; the rail is docked on the left; hover a bar to preview, click to jump.
• archives: click "Archived" at the sidebar foot; click a row to restore/fork an archived session.

Uninstall

.\install.ps1 -Uninstall     # Windows

./install.sh --uninstall     # macOS / Linux

Then restart the harness and hard-refresh.

Part 2 — If you are an AI

This part is written for AI agents that install, debug, or extend the toolkit. It assumes you can run shell commands and read the dsh packages in node_modules.

What this is (facts)

The toolkit bundles four independent enhancements, each touching different parts of dsh:

Key design points:

• No rebuilds: dsh serves client bundles at runtime from disk (per-bundle content hashes; client-modules HMR), so patched/plugin bundles take effect on browser refresh.
• No conflicts: the four components touch disjoint areas (LLM serialization, workspace rows, app layout, sidebar foot).
• Pinned versions: patch context is exact for the 0.1.0-rc.6 packages (dsh-host-apiproxy, dsh-llm-pi-ai, dsh-client-ui-workspace, dsh-client-ui-layout). A different version breaks git apply — see Debugging.

Deployment (exact steps)

The unified installer does all of this; the manual equivalent:

$profiles = "$env:USERPROFILE\.dsh\profiles"          # or $env:DSH_HOME\profiles
$repo = "<this repo>"

# 1. vision-bridge
$d = "$profiles\node_modules\@deepseek-ai\dsh-host-apiproxy\lib"
git -c core.autocrlf=false apply --unsafe-paths --directory="$d" "$repo\components\vision-bridge\patches\dsh-host-apiproxy.patch"
$d = "$profiles\node_modules\@deepseek-ai\dsh-llm-pi-ai\lib"
git -c core.autocrlf=false apply --unsafe-paths --directory="$d" "$repo\components\vision-bridge\patches\dsh-llm-pi-ai.patch"

# 2. workspace-launcher: plugin + patch
Copy-Item -Recurse "$repo\components\workspace-launcher\plugin\dsh-workspace-launcher" "$profiles\node_modules\"
$d = "$profiles\node_modules\@deepseek-ai\dsh-client-ui-workspace\lib"
git -c core.autocrlf=false apply --unsafe-paths --directory="$d" "$repo\components\workspace-launcher\patches\dsh-client-ui-workspace.patch"

# 3. msgrail: plugin + layout patch
Copy-Item -Recurse "$repo\components\msgrail\plugin\dsh-msgrail" "$profiles\node_modules\"
node "$repo\components\msgrail\scripts\patch-layout.mjs"   # default target resolves via DSH_HOME/~/.dsh

# 4. archives
Copy-Item -Recurse "$repo\components\archives\plugin\dsh-archives" "$profiles\node_modules\"

Cordis registration — append to $profiles\web\cordis.patch.yml (idempotent; the installer does this automatically):

- insert:
    - id: workspace-launcher
      name: 'dsh-workspace-launcher'
- insert:
    - id: msgrail
      name: 'dsh-msgrail'
- insert:
    - id: archives
      name: 'dsh-archives'

Then restart the harness and hard-refresh the browser.

Verification

# vision-bridge patches live
Select-String "$profiles\node_modules\@deepseek-ai\dsh-host-apiproxy\lib\index.js" -Pattern 'MODEL_DOES_NOT_SUPPORT_IMAGES'   # -> no match
Select-String "$profiles\node_modules\@deepseek-ai\dsh-llm-pi-ai\lib\index.js" -Pattern 'projectImageBlocksToText'            # -> match

# workspace-launcher endpoint + patch
Invoke-RestMethod -Uri http://127.0.0.1:3080/workspace-open/open -Method Post -ContentType 'application/json' -Body '{"path":"D:\\","app":"explorer"}'   # -> {"ok":true} and Explorer opens
Select-String "$profiles\node_modules\@deepseek-ai\dsh-client-ui-workspace\lib\client.js" -Pattern 'open-explorer'           # -> match

# msgrail layout
Select-String "$profiles\node_modules\@deepseek-ai\dsh-client-ui-layout\lib\client.js" -Pattern 'shell.history'              # -> match

Browser checks: paste an image and send (auto-described); workspace "..." menu shows the 5 actions; msgrail rail appears left of a conversation; "Archived" seat at the sidebar foot.

Debugging

Operations

• Restart harness: taskkill /F /T /PID <node dsh web pid> then npx -y @deepseek-ai/dsh web --host 127.0.0.1 --port 3080 (install.ps1 does this automatically).
• Uninstall: install.ps1 -Uninstall / install.sh --uninstall (reverses patches, removes plugins + cordis entries; the msgrail layout patch needs git apply -R-style manual restore or a package reinstall).
• Idempotency: the installer skips already-patched files and existing cordis entries; running it twice is a no-op.

Extra — how it works (and why it's shaped this way)

• dsh's client-module runtime (@deepseek-ai/dsh-client-modules): each client plugin's browser half is served from node_modules at /plugins/<id>/client.js?rev=<hash> and evaluated in the browser — patching a shipped bundle or adding a plugin is a hot update, no rebuild/republish.
• vision-bridge's trick: a text-only LLM cannot consume image bytes, so images become paths; the placeholder doubles as an instruction ("use whatever vision tool you have"), which keeps the agent decoupled from any specific vision vendor.
• workspace-launcher's host half resolves the workspace path per click from the client (row.cwd), so the server never needs to track "the current workspace".
• msgrail's layout patch is the only structural change: a zero-width grid column (var(--dsh-history-width, 0px)) that costs nothing when the plugin is absent.
• One installer, four owners: the unified script keeps every component's patch/plugin/cordis handling isolated and idempotent, so a failure in one component never blocks the others.

FAQ

• Q: Do the four components conflict? No — they modify disjoint parts (LLM serialization, workspace rows, app layout, sidebar foot).
• Q: Does vision-bridge send my image somewhere extra? It only changes who reads the image (your configured vision tool instead of the chat model); storage stays local.
• Q: Can I install only some components? Yes — use the per-component manual steps in Part 2, or copy the plugin + apply only the patches you want.
• Q: Why does the repo not ship a LICENSE? MIT is included; adjust the copyright line if needed.

Made for DeepSeek Harness users who like their workflows one command away.