mnemosyneos
MnemosyneOS plugins for Claude Code.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Check it deeper on your machine
It has changed since its scan on 19 Nov 2026.
As a Claude Code plugin, install it and then ask Claude Code:
/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/MnemosyneOS/dsh-claude-code-cli with scanskillsafety
Or from a terminal, clone it and run it:
git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/MnemosyneOS/dsh-claude-code-cli'
It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.
Badge for your README
Markdown
[](/listings/mnemosyneos--dsh-claude-code-cli)HTML
<a href="/listings/mnemosyneos--dsh-claude-code-cli"><img src="/api/listings/mnemosyneos--dsh-claude-code-cli/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
These steps run a script from the network that findsafeskills does not scan.
Step 1 — Install the Claude Code CLI
This plugin **does not** ship the Claude Code binary. It finds yours and runs it, so install it first.
**Windows, native installer (recommended, no Node needed):**
irm https://claude.ai/install.ps1 | iex
**Windows, npm (Node.js 22 or newer; never use `sudo`):**
npm install -g @anthropic-ai/claude-code@latest
**macOS / Linux:**
curl -fsSL https://claude.ai/install.sh | bash
If you already have it, confirm with `claude --version`. The plugin also accepts an explicit path in **Settings → Advanced → claude CLI path**, which is the field to use when the CLI lives in a project-local `node_modules` (a common layout for bundled installs).
Step 2 — Install the plugin into DSH
The plugin is a normal DSH package. Link it into a profile and add it to that profile's bundle list:
cd ~/.dsh/profiles/<profile> npm install "link:<path-to-this-repo>"
Then add `"@local/dsh-claude-code-cli"` to `dsh.profile.bundles` in `~/.dsh/profiles/<profile>/package.json`, and restart DSH.
`<profile>` is `desktop` for the desktop app. Note that `desktop` is owned by the Electron process and **cannot** be started from the CLI — that is expected, and the plugin handles it by using the `web` profile (same bundle set) for its phone-access instance.
Step 3 — Pick a model
Open DSH → sidebar → **Claude Code Desktop** → **Settings → Model**. Pick any model DSH already has. Send a message.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 21 lines
- {
- "name": "mnemosyneos",
- "owner": {
- "name": "MnemosyneOS",
- "url": "https://github.com/MnemosyneOS"
- },
- "description": "MnemosyneOS plugins for Claude Code.",
- "plugins": [
- {
- "name": "dsh-claude-code-cli",
- "displayName": "DSH Claude Bridge",
- "source": "./integrations/claude-code",
- "version": "2.5.6",
- "description": "Teaches Claude Code how to drive the Claude Code Desktop workspace this repo ships for DeepSeek Harness: the loopback bridge, the model catalogue, the turn/poll event stream, and the approval flow. The DSH-side plugin itself installs by linking this repo into a DSH profile.",
- "homepage": "https://github.com/MnemosyneOS/dsh-claude-code-cli",
- "repository": "https://github.com/MnemosyneOS/dsh-claude-code-cli",
- "license": "MIT",
- "keywords": ["dsh", "deepseek-harness", "claude-code", "bridge", "local-models", "desktop"]
- }
- ]
- }
README.md · README · 333 lines
- # DSH-Claude Code CLI
- A graphical Claude Code workspace that lives inside **DeepSeek Harness (DSH)** — no terminal,
- no hand-edited `settings.json`, no lost sessions. It drives the real `claude` CLI as a child
- process, so every flag and every `settings.json` key still works, and it routes the CLI at
- whatever model you already have configured in DSH, local or remote.
- **English** · [简体中文](README.zh-CN.md)
- - 35 settings, generated from the CLI's own flags and `settings.json` keys
- - Tool approvals inside the panel — or from your phone on the same Wi-Fi
- - Attachments, a live thinking/tool chain, session history with archive and trash
- - Every model DSH already has, in the composer's model picker
- - One-click CLI upgrade that reads the real npm registry and respects how yours is installed
- - English and Chinese UI, switchable in settings
- ---
- ## Contents
- - [Why not just the terminal?](#why-not-just-the-terminal)
- - [How it works](#how-it-works)
- - [What you need](#what-you-need)
- - [Install](#install)
- - [Using it](#using-it)
- - [Tool approvals](#tool-approvals)
- - [Upgrading the CLI](#upgrading-the-cli)
- - [Phone access](#phone-access)
- - [Settings](#settings)
- - [Security model](#security-model)
- - [Troubleshooting](#troubleshooting)
- - [Repository layout](#repository-layout)
- - [Verification](#verification)
- - [License and trademarks](#license-and-trademarks)
- ---
- ## Why not just the terminal?
- The CLI is excellent and this plugin does not reimplement it. What it adds is the part a
- terminal cannot give you:
- | | Terminal | This plugin |
- | :- | :- | :- |
- | Every CLI flag | type them each time | one schema-driven settings page, remembered |
- | `settings.json` keys | hand-edit JSON | generated from the official key names, injected with `--settings` |
- | Tool approvals | `y` / `a` / `n` at the prompt | a card in the panel, or on your phone |
- | Thinking and tool calls | interleaved text | a timeline you can collapse |
- | Sessions | `--resume` and memory | a list with rename, pin, archive, restore, permanent delete |
- | Attachments | paste a path | drag and drop, becomes an `@path` the CLI can read |
- | Switching models | env vars and restarts | a picker in the composer |
- | Upgrading | npm and luck | button, with the real version diff and the npm log |
- ## How it works
- ```
- ┌──────────────────────────────────┐
- │ DeepSeek Harness (DSH) │
- │ ┌────────────────────────────┐ │
- │ │ Claude Code Desktop panel │ │ lib/client.js (slots: sidebar + main)
- │ └───────────┬────────────────┘ │
- │ │ HTTP │
- │ ┌───────────▼────────────────┐ │
- │ │ lib/index.js │ │ ctx.webServer.register() — 30+ routes
- │ └───────────┬────────────────┘ │
- └──────────────┼───────────────────┘
- │ spawns one process per turn
- ┌────────▼─────────┐
- │ claude CLI │ -p --output-format stream-json --settings <runtime file>
- └────────┬─────────┘
- │ ANTHROPIC_BASE_URL=http://127.0.0.1:<bridge>
- ┌────────▼─────────┐
- │ loopback bridge │ Anthropic wire format ⇄ DSH ctx.llm.stream()
- └────────┬─────────┘
- │
- ┌────────▼─────────┐
- │ your DSH model │ whatever you authenticated: DeepSeek, a gateway,
- └──────────────────┘ a local Qwen on :8080, anything
- ```
- Three consequences worth knowing:
- 1. **No Anthropic account is required.** The CLI is pointed at the bridge with
- `ANTHROPIC_BASE_URL` and `ANTHROPIC_MODEL`, so it never calls Anthropic unless your chosen
- model is an Anthropic one.
- 2. **One turn is one process.** MCP servers and background bash are rebuilt every turn.
- That is a cost, and it is why long-lived MCP connections are not reproduced here.
- 3. **The bridge speaks both wire formats.** The CLI talks Anthropic; almost every other
- provider talks OpenAI. The bridge converts per turn, including tool calls.
- ## What you need
- | | |
- | :- | :- |
- | OS | Windows 10 1809+ — verified here. macOS / Linux should work (the plugin is pure Node) but we have not run it there. |
- | DSH | **0.2.2-rc.3 or newer** (Node ≥ 24.9 ships with it) |
- | Claude Code CLI | any recent version — the binary only, no login needed |
- | A model | at least one already configured in DSH |
- | Network | internet for the install and for update checks; none after that unless your model is remote |
- ## Install
- ### Step 1 — Install the Claude Code CLI
- This plugin **does not** ship the Claude Code binary. It finds yours and runs it, so install
- it first.
- **Windows, native installer (recommended, no Node needed):**
- ```powershell
- irm https://claude.ai/install.ps1 | iex
- ```
- **Windows, npm (Node.js 22 or newer; never use `sudo`):**
- ```powershell
- npm install -g @anthropic-ai/claude-code@latest
- ```
- **macOS / Linux:**
- ```bash
- curl -fsSL https://claude.ai/install.sh | bash
- ```
- If you already have it, confirm with `claude --version`. The plugin also accepts an explicit
- path in **Settings → Advanced → claude CLI path**, which is the field to use when the CLI
- lives in a project-local `node_modules` (a common layout for bundled installs).
- ### Step 2 — Install the plugin into DSH
- The plugin is a normal DSH package. Link it into a profile and add it to that profile's
- bundle list:
- ```bash
- cd ~/.dsh/profiles/<profile>
- npm install "link:<path-to-this-repo>"
- ```
- Then add `"@local/dsh-claude-code-cli"` to `dsh.profile.bundles` in
- `~/.dsh/profiles/<profile>/package.json`, and restart DSH.
- `<profile>` is `desktop` for the desktop app. Note that `desktop` is owned by the Electron
- process and **cannot** be started from the CLI — that is expected, and the plugin handles it
- by using the `web` profile (same bundle set) for its phone-access instance.
- ### Step 3 — Pick a model
- Open DSH → sidebar → **Claude Code Desktop** → **Settings → Model**. Pick any model DSH
- already has. Send a message.
- ## Using it
- | Area | What you get |
- | :- | :- |
- | Composer | model picker (bottom right), permission mode, workspace picker, attachments, paste-anywhere |
- | Per reply | model name, timestamp, token speed, cache hit rate, copy, branch into a new session |
- | Context | a ring progress indicator in the composer, fed by the CLI's own `modelUsage` |
- | History | rename, pin, archive, restore, permanent delete, batch operations, copy ID, groups |
- | Thinking | a collapsible timeline, visually distinct from the answer, with a status icon per step |
- Diagnostics live at `GET /api/claude-code-parity/health` (bridge port, CLI path, version),
- plus `/cli`, `/models` and `/upgrade/status`.
- ## Tool approvals
- The CLI is started with `--permission-prompt-tool mcp__ccp-approval__approve` and a generated
- `mcp-config.json` that points at `lib/approval-mcp.js`. Every tool the model wants to run
- becomes a card in the panel: **allow once**, **always allow**, or **deny**.
- - Timeout defaults to 120 s (configurable, 10–600 s); on timeout the tool is denied.
- - **Settings → Tool permissions** can auto-approve the read-only family
- (`Read`, `Glob`, `Grep`, `WebFetch`, `WebSearch`, `NotebookRead`, `TodoWrite`).
- - The permission mode is passed straight through as `--permission-mode`, so
- `acceptEdits`, `plan`, `bypassPermissions` etc. behave exactly as in the CLI.
- ## Upgrading the CLI
- **Settings → Advanced → Upgrade Claude CLI** does three real things:
- 1. Reads the installed version from the CLI's own `package.json` (falling back to
- `claude --version`), then fetches `https://registry.npmjs.org/@anthropic-ai/claude-code/latest`
- and compares them semantically. No hard-coded strings.
- 2. Determines **how** your CLI is installed by walking up from the binary to
- `node_modules/@anthropic-ai/claude-code`. A project-local install is upgraded in its own
- root; only a genuine global install gets `npm i -g`.
- 3. Streams npm's stdout/stderr into `upgrade-state.json` line by line, and reports
- `before → after` plus whether anything actually changed.
- > This matters: `npm update -g @anthropic-ai/claude-code` silently does nothing for a
- > project-local install, which is exactly why "the update button did nothing" is a common
- > report. The plugin inspects the layout first.
- ## Phone access
- **Settings → Mobile access** lists the machine's real LAN IPv4 interfaces. Pick one and press
- **Start**. The plugin then:
- - starts a second DSH web instance bound to that LAN address
- (`--host <LAN-IP> --trusted-host <host:port>`) — DSH **rejects** `0.0.0.0`, so a concrete
- literal is required;
- - parses the tokenised URL from that instance's stdout;
- - opens a tiny **short-address redirect** on `port + 1`, so you can type
- `http://192.168.1.103:9241` on the phone instead of the full token URL.
- The phone gets the complete DSH UI including this plugin, sharing session data with the
- desktop. **Stop** tears down both the instance and the redirect.
- ## Settings
- 35 items in six groups. Every one of them is wired into the run path — the full round-trip is
- covered by the verification suite.
- | Group | Items (schema key) |
- | :- | :- |
- | General (6) | `default_cwd`, `permission_mode`, `permission_prompts`, `timeout`, `tool_policy`, `language` |
- | Appearance (5) | `theme`, `show_thinking`, `show_tool_output`, `transcript_view`, `stream_speed` |
- | Tools & Permissions (7) | `allow_read`, `allow_write`, `allow_bash`, `allow_glob`, `allow_web`, `approval_timeout`, `auto_approve_readonly` |
- | Model (6) | `model`, `provider`, `reasoning_effort`, `max_tokens`, `temperature`, `vision_mode` |
- | Session (5) | `auto_title`, `keep_history`, `history_limit`, `paste_doc_limit`, `export_format` |
- | Advanced (6) | `system_prompt`, `append_identity`, `strip_images_on_error`, `debug`, `cli_path`, `extra_flags` |
- Full descriptions and defaults: [docs/configuration.md](docs/configuration.md).
- `reasoning_effort` defaults to `auto` — meaning "do not send the parameter". Forwarding a
- level to a model that does not support it fails the whole turn, so `auto` is the safe default
- and the plugin migrates older configs to it once (`settings_version`).
- ## Security model
- - The bridge listens on **loopback only** and requires the fixed header `x-api-key: cc-bridge`.
- - DSH web access uses a launch token that is exchanged for an `HttpOnly`, `SameSite=Strict`
- cookie; the token is regenerated on every process start.
- - Phone access requires you to press **Start**; nothing is exposed until you do, and the
- short address only 302-redirects to the tokenised URL — it never serves content itself.
- - Credentials never enter this repository. The plugin reads whatever DSH already
- authenticated and passes provider errors through verbatim.
- ## Troubleshooting
- | Symptom | Cause | Fix |
- | :- | :- | :- |
- | Panel is blank, backend works | `__ModuleLoader__.load({id})` does not match `package.json` `name` | they must be byte-identical |
- | Second turn always fails | assistant messages missing `source` | fixed in 2.5.1; check you are not running an old build |
- | `does not support reasoning effort "medium"` | a level was forwarded to a model that rejects it | set reasoning effort to `auto` |
- | `message content cannot be empty` | tool-result-only user message | fixed in 2.5.1; 2.5.2 no longer needs a placeholder — such a message is simply not emitted |
- | `profile "desktop" is managed exclusively by the Electron application` | CLI cannot start the desktop profile | use the `web` profile |
- | Update button reports no change | CLI is installed project-locally, `-g` cannot reach it | fixed; the plugin detects the layout |
- | Plugin card shows an old name | the card reads `locale/<lang>.json` `meta.title`, not `package.json` | update the locale files and restart |
- | Every tool returns `No result provided` | tool results were packed as parts of a user message; DSH only reads text/image there | fixed in 2.5.2 — results travel as their own `role: 'tool'` message |
- | The transcript jumps to the bottom while reading | auto-follow keyed on "new output" instead of "already at the bottom" | fixed in 2.5.2; the round arrow at the top right jumps back down |
- | A reasoning model shows thinking but no answer | `blocks` merged into the streaming thinking row | fixed in 2.5.2; see the knowledge base entry ㉚ |
- | The thinking chain appears all at once, never while it happens | the CLI was not asked for deltas, the panel appended whole messages, and the row was collapsed to one line | fixed in 2.5.3 — `--include-partial-messages` + a typing pump + a row that opens while live; see entry ㉝ |
- | After an update the panel still behaves like the old build | the desktop process started before the files changed, so it runs the cached module | since 2.5.3 a notice above the transcript says the running code is old; closing the window is not quitting |
- | `dsh: fatal load failure: EPERM ... rename` | two instances sharing `~/.dsh/ccp-data` renamed the same job file | fixed in 2.5.3 — writes retry and never escape as an unhandled rejection; see entry ㉞ |
- | An image attachment returns `Cannot read properties of undefined (reading 'attachmentId')` / `Claude Code error` | Anthropic image blocks were forwarded with `source`, while DSH requires a durable `attachment` ref | fixed in 2.5.4 — images are admitted through `ctx.get('attachments')` and swapped in by reference; see entry ㉟ |
- | A long thinking chain stops mid-sentence | the expanded detail was sliced to 8,000 characters and the typing pump capped at 400 chars/s | fixed in 2.5.5 — full detail is rendered and the pump scales with backlog; see entry ㊱ |
- | A just-sent instruction vanishes once the model starts thinking | overlapping fire-and-forget history saves: a stale snapshot finished last and wrote a shorter message list | fixed in 2.5.6 — saves carry a monotonic revision and older writes are rejected; see entry ㊲ |
- | `exceed_context_size_error` / `request (131770 tokens) exceeds the available context size` | `--resume` replays the whole transcript; a long Chinese session exceeds the local 131k window while the CLI's default compact threshold follows a reported 1M context | fixed in 2.5.7 — `--autocompact 100000`, a CJK-aware hard budget in the bridge, and an actionable Chinese hint; tune `设置 → 会话 → 自动压缩阈值(Token)` |
- | Thinking stops appearing when the reply also runs tools | the typing pump counted text queued *behind* a tool marker, so the marker at the head was never released and later thinking was unreachable | fixed in 2.5.8 — head markers release immediately; backlog is measured only over the leading text run; see entry ㊴ |
- | Scrolling up inside a live thinking chain snaps back down | the detail box forced `scrollTop = scrollHeight` on every delta | fixed in 2.5.9 — follow only while already at the bottom, plus a ↓ jump button; see entry ㊵ |
- | A cancelled/switched task shows "check the provider's credentials/quota" | the job died with `Connection error.` / `terminated`, but every provider failure shared one generic advice string | fixed in 2.5.10 — dropped streams are translated into a Chinese connection hint; credentials/quota advice only applies to real auth failures; see entry ㊶ |
- More, with root causes and evidence, in [docs/architecture.md](docs/architecture.md) and the
- knowledge base shipped alongside this plugin.
- ## Repository layout
- ```
- ├── package.json DSH manifest (dsh.bundle.patch + dsh.client.inject)
- ├── patches/bundle.yaml loader patch — default config values
- ├── icon.svg locale/{en,zh}.json
- ├── lib/
- │ ├── index.js host: routes, job manager, CLI spawn, upgrade, mobile
- │ ├── client.js client: sidebar + main panel (single-file React)
- │ ├── bridge.js Anthropic ⇄ DSH LLM protocol bridge
- │ ├── approval-mcp.js the MCP server the CLI calls to ask for permission
- │ └── common.js persistence, settings schema, CLI history scan
- ├── docs/ architecture, configuration, mobile, publishing
- ├── integrations/claude-code/ Claude Code plugin wrapper (marketplace entry)
- ├── tools/verify/ the CDP + HTTP verification harness
- └── .claude-plugin/ marketplace manifest
- ```
- ## Verification
- Nothing here is claimed without being run. The harness in `tools/verify/` drives a real DSH
- instance and a real headless Chrome:
- ```bash
- # start an isolated instance (note the token it prints)
- dsh --profile <verify-profile> --port 39120 --host 127.0.0.1 --no-open
- node tools/verify/_verify_all.mjs "<token>" # 21 backend assertions
- node tools/verify/_verify_ui.mjs "<token>" # 9 real-browser assertions
- node tools/verify/_verify_chat.mjs "<token>" # two real turns through CLI + bridge
- node tools/verify/_verify_reply_unit.mjs # 15 event-merge checks, no instance needed
- node tools/verify/_verify_reply_ui.mjs "<token>" # 13 — the answer must be painted
- node tools/verify/_verify_reply_history.mjs "<token>" # 15 — damaged records recover
- node tools/verify/_verify_scroll_ui.mjs "<token>" # 22 — scroll position holds, arrow works
- node tools/verify/_verify_tool.mjs "<token>" # 13 — the model reads real tool output
- node tools/verify/_verify_image.mjs "<token>" # 8 — an uploaded image does not kill the turn
- node tools/verify/_verify_thinking_full.mjs "<token>" # 7 — long chains render end to end
- node tools/verify/_verify_split_ui.mjs "<token>" # 12 — split view + Start pane
- node tools/verify/_verify_send_ui.mjs "<token>" # 6 — 40 MB attachment round trip
- node tools/verify/_verify_mobile_port.mjs "<token>" 9260 # 9 — port fallback
- node tools/verify/_verify_typing_ui.mjs "<token>" # 18 — the chain types itself out
- node tools/verify/_verify_save_guard.mjs "<token>" # 5 — stale history saves are rejected
- node tools/verify/_verify_usermsg_persist.mjs "<token>" # 9 — sent messages survive later turns
- node tools/verify/_verify_context_overflow.mjs "<token>" # 17 — autocompact argv, history trim, friendly error (overflow + lost connection)
- node tools/verify/_verify_title_stable.mjs "<token>" # 9 — stable topic names, locked renames
- node tools/verify/_verify_pump_interleave.mjs # 8 — tools cannot freeze the thinking pump
- node tools/verify/_verify_thinking_scroll.mjs # 9 — reading an early thought holds position
- ```
- Baseline at 2.5.10: **21/21**, **9/9**, **15/15**, **13/13**, **15/15**, **22/22**, **12/12**,
- **12/12**, **6/6**, **9/9**, **13/13**, **18/18**, **8/8** (image), **10/10** (long thinking),
- **5/5** (save guard), **9/9** (message persistence), **17/17** (context overflow), **9/9** (title),
- **8/8** (pump interleave), **9/9** (thinking scroll),
- plus a two-turn conversation on one session id. The scroll check passed 22/22 on a clean rerun after
- one flaky full-suite pass (18/22) caused by browser timing.
- The typing check samples the real DOM every 200 ms: a one-shot render can produce at most one step
- from zero to full, so the run requires many distinct intermediate lengths (measured: 95 steps between
- 0 and 928 characters for the chain, 36 for the answer). The chat harness pins a working text model and restores your previous selection
- afterwards, so it never leaves your model picker moved.
- ## License and trademarks
- MIT — see [LICENSE](LICENSE) and [NOTICE](NOTICE). **Claude** and **Claude Code** are
- trademarks of Anthropic, PBC. This is an independent front end for a locally installed CLI,
- not affiliated with or endorsed by Anthropic. **DeepSeek Harness** is a product of DeepSeek;
- this is a third-party plugin for that host.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.