The dsh security mode plugin achieves a zero third-party plugin startup by constructing a blank profile.
Install
# from npm (prebuilt)
dsh plugin --profile web add dsh-safemode-profile
# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)
dsh plugin --profile web add github:jinsiyu/dsh-safemode-profile
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
One job: keep dsh --profile safemode clean — force-restore on startup,
continuously guard at runtime, so the profile only ever contains the
whitelisted core bundles.
dsh --profile safemode # boot with zero third-party plugins (suggest: dsh --profile safemode --port 3081)
What it does (two phases)
1. On startup: force-restore (regardless of whether the profile exists or was modified)
Every time DSH boots and loads this plugin row, it unconditionally rewrites
$DSH_HOME/profiles/safemode/ back to the whitelist template:
| File | Forced content |
|---|---|
package.json |
dsh.profile.bundles = whitelist (default @deepseek-ai/dsh-base + @deepseek-ai/dsh-web-app), dependencies emptied |
cordis.patch.yml |
empty user patch layer ([] template) |
pnpm-workspace.yaml |
same pnpm settings as an official profile |
If the files already match the template, nothing is written (avoids a self-triggering watch loop); anything that drifted is rewritten.
2. At runtime: continuous guard (drift checking)
While the plugin row is alive, two channels watch the profile for modification:
- fs.watch realtime: any change to
package.json/cordis.patch.yml/pnpm-workspace.yaml→ debounced 300ms → immediate restore; - 30s polling fallback: full
detectDrift()comparison (covers unreliable watch mounts and files replaced wholesale externally).
On drift (bundle added/removed, patch layer written to, dependencies added)
the guard restores the profile and logs a warning:
safemode profile was modified (...); restored to whitelist template.
Activation (no build scripts — installs cleanly in one shot)
This package has no postinstall / no build scripts of any kind — pnpm
never blocks it, dsh plugin add succeeds with exit code 0 and reconcile runs
normally. No approve-builds / allowBuilds configuration needed.
All work is done by the plugin row (lib/index.js → lib/guard.js):
every DSH start force-restores once, then the guard stays resident.
Optional manual tool: node scripts/ensure-safemode.mjs force-restores
immediately (handy without a DSH restart, or for troubleshooting).
Install
Prerequisites: DSH (@deepseek-ai/dsh) and Node.js (>=18) installed.
Method A (recommended, one command from npm) — the package is published on
npm with a dsh.bundle manifest; install directly with the official plugin
command:
dsh plugin --profile web add dsh-safemode-profile
Method B (from GitHub) — install straight from the source repository
(commit-ish pinned to the default branch main, works with both npm and pnpm):
dsh plugin --profile web add github:jinsiyu/dsh-safemode-profile#main
Method C (local tarball / unpublished) — pack and install a local build:
# pack inside the plugin directory
npm pack # → dsh-safemode-profile-0.3.5.tgz
# install into a target profile (e.g. web)
dsh plugin --profile web add .\dsh-safemode-profile-0.3.5.tgz
Restart DSH; once the plugin row is active, safemode enters "force-restore + resident guard" mode. On startup the plugin prints a short usage banner (guard status, start command, whitelist, repo link).
Custom whitelist
The whitelist is the only customization entry point — set it via the environment variable (do not hand-edit the safemode profile files; any edit is restored by the guard):
# pure CLI only (no GUI)
$env:DSH_SAFEMODE_BUNDLES = "@deepseek-ai/dsh-base"
The variable is read from process.env by both the guard and the restore
logic, and takes effect the next time dsh starts with it set.
Hardening: make the three managed files read-only
Recommended: set the three managed files under
~/.dsh/profiles/safemode/ read-only to block
dsh plugin --profile safemode add <pkg> at the filesystem level — pnpm
fails when it cannot write them, so the plugin never gets in (DSH startup
is unaffected: it only reads them and writes cordis.yml):
package.json(dependencies + bundles — the choke point of any install)cordis.patch.yml(the user patch layer)pnpm-workspace.yaml(pnpm-managed file)
⚠️ Never lock the whole directory: DSH rewrites
cordis.ymlon every boot (prepareProfileunconditionally writeFileSync), so a read-only directory or read-onlycordis.ymlmakes safemode fail to start (verified: exit 1). Windowsattrib +R <dir> /Srecursively locks every file includingcordis.yml— do not use it.
# Windows: lock only the three files (no /S recursion)
attrib +R "$env:USERPROFILE\.dsh\profiles\safemode\package.json" `
"$env:USERPROFILE\.dsh\profiles\safemode\cordis.patch.yml" `
"$env:USERPROFILE\.dsh\profiles\safemode\pnpm-workspace.yaml"
# POSIX: the three files to 444 (the directory still needs r-x and must stay writable)
chmod 444 ~/.dsh/profiles/safemode/package.json \
~/.dsh/profiles/safemode/cordis.patch.yml \
~/.dsh/profiles/safemode/pnpm-workspace.yaml
To unlock:
attrib -R "$env:USERPROFILE\.dsh\profiles\safemode\package.json" `
"$env:USERPROFILE\.dsh\profiles\safemode\cordis.patch.yml" `
"$env:USERPROFILE\.dsh\profiles\safemode\pnpm-workspace.yaml" # Windows
chmod 644 ~/.dsh/profiles/safemode/package.json \
~/.dsh/profiles/safemode/cordis.patch.yml \
~/.dsh/profiles/safemode/pnpm-workspace.yaml # POSIX
Notes:
- Let the plugin create/restore the profile first, then lock — the content must match the whitelist template before locking (the guard only writes when it detects drift; template-matching content is left untouched, so locking coexists with the guard). If drift is found while locked, the guard's restore fails with a warning — that is the expected behavior of a lock: locked means no drift should exist.
- Unlock before changing the whitelist. The guard itself is unaffected — it
reads
package.jsonto detect drift and only writes when a drift is found.
Deeper protection: ACL (Windows) and chattr +i (Linux)
The attrib +R / chmod 444 above is "basic read-only": the two
platforms treat it differently, and neither can stop deletion (deleting
a file depends on the parent directory's permissions; root ignores all
permission bits anyway). The following are "hard locks", pick as needed:
Windows: ACL deny (icacls) — precise control over write vs delete
# Deny the current user write access to the three files (W = write data)
icacls "$env:USERPROFILE\.dsh\profiles\safemode\package.json" /deny "$env:USERNAME:(W)"
icacls "$env:USERPROFILE\.dsh\profiles\safemode\cordis.patch.yml" /deny "$env:USERNAME:(W)"
icacls "$env:USERPROFILE\.dsh\profiles\safemode\pnpm-workspace.yaml" /deny "$env:USERNAME:(W)"
# Stronger: also deny deletion (DE = delete; WD = write data / create file)
icacls "...\package.json" /deny "$env:USERNAME:(WD,DE)"
To unlock (remove that user's deny entries):
icacls "...\package.json" /remove:d "$env:USERNAME"
Key points:
(W)blocks content modification,(DE)blocks deletion,(WD)blocks write + create — more precise than attrib: you can block writes while leaving admins able to delete;- Still does not stop administrators: Administrators / SYSTEM have Full
Control (F) by default and can take ownership or clear the deny. To block
them too, deny
BUILTIN\Administratorsas well — but admins can still bypass via "take ownership"; - icacls requires permission to modify the file's ACL (the file owner has it by default).
Linux: immutable attribute (chattr +i) — root cannot touch it either
# Lock: file cannot be modified, deleted, or renamed (root included)
sudo chattr +i ~/.dsh/profiles/safemode/package.json \
~/.dsh/profiles/safemode/cordis.patch.yml \
~/.dsh/profiles/safemode/pnpm-workspace.yaml
# Inspect
lsattr ~/.dsh/profiles/safemode/package.json # output containing "i" = locked
# Unlock
sudo chattr -i ~/.dsh/profiles/safemode/package.json \
~/.dsh/profiles/safemode/cordis.patch.yml \
~/.dsh/profiles/safemode/pnpm-workspace.yaml
Key points:
+i(immutable) is the strongest lock: nobody (root included) can modify, delete, or rename — you must-ifirst for any operation, harder than chmod or ACL;- Requires
sudo(root), and the filesystem must support the attribute (ext4/xfs do; some network/container filesystems do not); - Side effect for safemode: the guard normally rebuilds a deleted file
(detectDrift finds "missing" → force-restore), but
+iblocks even that; upgrading the plugin or changing the whitelist also needs-ifirst. Use it only when "no account on this machine (root included) may touch the safemode config" is a real requirement; basic read-only (attrib / chmod 444) is enough for everyday use.
Three levels compared
| Option | Blocks content edit | Blocks delete | Blocks root | Requires | Impact on guard self-heal |
|---|---|---|---|---|---|
attrib +R / chmod 444 (basic) |
✅ for normal users | ❌ | ❌ | none | none (guard skips writes when clean) |
| icacls deny (ACL) | ✅ | ✅ (with DE) | ❌ (admins can bypass) | file owner | none |
chattr +i (immutable) |
✅ | ✅ | ✅ | sudo | ⚠️ blocks guard rebuild; -i first |
Suggestion: basic read-only is enough for everyday use (pnpm runs as a
normal user, 444 already blocks installs completely); add ACL (DE) if you
want to guard against accidental deletion; only choose chattr +i when
"no account (root included) may touch the safemode config" is a hard
requirement.
Notes
- Port: safemode also carries the webServer, default 3080. Running it
alongside the web profile conflicts (
EADDRINUSEfails the boot) — use--portto separate them:dsh --profile safemode --port 3081. - Sessions/credentials are NOT isolated: sessions,
settings.yamland.envare home-level shared; safemode only isolates plugins. - Home patch leaks in:
~/.dsh/cordis.patch.ymlapplies to every profile — don't mount plugins there (the guard only watches the safemode directory itself, not the home patch). - Don't hand-edit
cordis.yml: DSH rewrites it on every boot; editcordis.patch.ymlinstead — but safemode's patch layer is restored to empty by this plugin, so customize through the whitelist env var.
Links
More in this category
Tencent/WeKnora#dsh-weknora★ 37764
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★ 37055
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★ 8452
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★ 5921
Provides bash, PowerShell 7, and high-performance Rust tools for DSH on Windows to reduce token usage.
sixtysevenlf/dsh-blender-plugin★ 2194
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★ 1751
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.