A switch beside the permission control in the composer: while it is on, that session mounts no skill - the skill tool and its catalog are gone, and turning it back off restores both.
Install
# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)
dsh plugin --profile web add github:owenKu1010/dsh-pure-mode
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
English | 简体中文
A tiny DSH bundle that puts one switch beside the permission control in the composer. While it is on,
that session mounts no skill at all: the skill tool is gone and so is the persistent skill catalog
that would otherwise ride along in every request.
Use it for the sessions where skills are pure overhead — plain Q&A, writing, a one-off command — and for sessions where you do not want the model deciding on its own that some listed skill applies.
[ + ] [ ⛨ Full access ▾ ] [ A|B ] [ (leaf) Pure mode ] [ ✦ Experts ▾ ] [ ⏻ PUA ] [ model ▾ ]
↑ this one
| State | What you see | What the model gets |
|---|---|---|
| Off (default) | 1px accent-blue border, no fill, blue leaf, normal label | everything, exactly as before |
| On | same bordered chip flooded with the accent, inverted label, filled leaf + a dot | no skill tool, no skill catalog |
Turning it back off restores both, from the next step onward.
Not related to dsh-purge
The names sound adjacent — "purge", "pure" — so, plainly: this plugin has nothing to do with
dsh-purge and does not replace, extend or depend on it. They solve different problems:
| dsh-purge | Pure mode (this plugin) | |
|---|---|---|
| What it touches | instruction / prompt authority: patches, shim injection, override files | tool visibility of one live session |
| Where the change lives | files in the profile, applied at composition time | one tool restriction in that Agent's context |
| When it takes effect | after a restart of the composition | on the next step, no restart |
| How you undo it | revert the patches / shim / overrides | release the restriction, tool returns immediately |
| Scope | the whole profile | a single session |
It writes no files outside $DSH_HOME/pure-mode/, reads none of dsh-purge's state, and changes no prompt
text. Both can be installed in the same profile and neither notices the other.
Why
Two things are paid for on every request in every session:
- the
skilltool schema, and - a durable
<available_skills>user message listing every model-invocable skill with a capped description.
That catalog is worth its tokens in a session that uses skills. In a session that does not, it is a recurring tax plus a standing invitation for the model to go load something. Pure mode is the switch that turns the whole subsystem off for one session without touching a patch, a preset, or another plugin.
Install
This is a normal DSH bundle (dsh.bundle.patch in package.json), installed into a profile:
- Clone this repository anywhere on disk:
git clone https://github.com/owenKu1010/dsh-pure-mode.git - In DSH open Plugins → Add plugin and point it at the cloned folder (local directory install).
- Confirm the
@local/dsh-pure-moderow is enabled, then reload the page.
The row resolves as a bundle, inserts one plugin row named pure-mode, and needs a page reload once for
the client half to appear in the composer.
Requires the shipped skill tool to be mounted (@deepseek-ai/dsh-tool-skill, part of the default base
composition). Nothing else.
Use
- The switch is per session. Each session keeps its own choice.
- A session that has never been toggled follows the value the last toggle left behind, so a fresh session inherits your most recent decision.
- Choices persist at
$DSH_HOME/pure-mode/state.json. Delete that file to reset everything to off. - Turning it on is cheap (one small empty-catalog message). Turning it off republishes the full catalog — that is the direction that costs tokens, so decide before the first message when you can.
$DSH_HOME/pure-mode/client.logrecords which client bundle each page actually loaded (build=v4), so a stale browser tab is visible instead of guessed at.
Configuration
By default only skill is hidden. To take the rest of the skill surface with it — for example the
create_skill tool registered by the skills manager — restate the row's config in the profile patch:
- id: pure-mode
name: '@local/dsh-pure-mode'
config:
denyTools: [skill, create_skill]
Names must be global tool names the agent can see. A wrong name does not fail silently: the switch reports
"hiding the tool failed" and the Host log carries the tools.restrict() message listing known tools.
Companion: Skills Manager (recommended)
Pure mode is an all-or-nothing switch. For per-skill control — enable/disable individual skills, turn whole sources on or off, browse and install skills from repositories — pair it with DSH Skills Manager by MichengAI:
github.com/MichengAI/dsh-skills-manager
@michengai/dsh-skills-manager· Apache-2.0 · actively maintained
They operate on different layers and compose cleanly:
| Skills Manager | Pure mode | |
|---|---|---|
| Scope | one skill or one source | the whole skill tool |
| Mechanism | per-skill invocation policy in the ctx.skills registry |
agent.ctx.tools.restrict() |
| When it is off | each skill keeps its own state | everything is restored |
The manager reports a disabled skill as { modelInvocable: false, userInvocable: false }, so it never
reaches the catalog, the skill tool or the /name gesture in the first place. Pure mode only hides the
tool, and therefore never re-enables what the manager disabled:
- pure mode on → nothing is mounted (including skills the manager allows);
- pure mode off → exactly what the manager allows; anything you disabled there stays gone;
- toggling skills in the manager while pure mode is on still records normally and applies once pure mode is off.
How it works
One restriction on the Agent's own context, released when the Agent or the plugin goes away:
agent.ctx.effect(
() => agent.ctx.tools.restrict({ deny: ['skill'] }),
'pure-mode: hide the skill tool for this session',
)
@deepseek-ai/dsh-tool-skill compares the skill definition it registered with the definition the calling
agent actually resolves. A restricted-away tool therefore removes both the model-facing schema and the
durable catalog; a session that never published one never gets one at all.
The client half holds no authority: it reads and writes the stored mode over an authenticated same-origin
route (/api/pure-mode) registered through ctx.connection.fetch, and the Host decides whether to apply
the restriction. The UI is plain React registered into the conversation.input.left slot, styled with
--dsw-alias-* theme tokens only.
Notes and caveats
- History is append-only. A catalog that was already published stays in the conversation and keeps costing tokens until compaction; the saving is real from the start of a session, not retroactive. Toggling on→off→on appends a replacement each time.
- Tool-level change costs cache. The tool list sits at the front of the request, so every toggle can invalidate prompt-cache reuse from that point.
/skill-namegestures still work for skills the manager allows. Typing a skill name is a deliberate user act, not a mount, and the Host does not intercept it.- Scope is the tool, not the registry. Other tools, MCP servers and plugins are untouched.
- Restriction needs a restrictable registration. If a composition registers the skill tool in the
Agent's own layer instead of an ancestor layer,
tools.restrict()refuses; the switch then reports the failure instead of pretending it worked. - Tested against DSH
0.2.2-rc.3on Windows (Electron 44 / Chromium 152). Nothing version-specific is used beyond the public Services (tools,connection,agents), but other releases are unverified.
Repository layout
package.json bundle manifest: dsh.bundle.patch + dsh.client
cordis.patch.yml inserts the `pure-mode` row
index.js Host half: state, reconcile, tools.restrict, /api/pure-mode
client.js Client half: the composer switch
locale/{en,zh}.json plugin-list title/description + switch strings
icon.svg plugin icon
docs/DETAIL.md the in-app detail page
License
MIT. See LICENSE.
Links
More in this category
tw93/Mole#mole★ 75286
Official agent skill for the open-source Mole CLI (mo), a disk cleanup tool for macOS and Windows: read disk usage, preview and run cleanup, inspect installed apps, and read system health and cleanup history from Mole's JSON output. It previews with --dry-run and asks before removing anything, and leaves app uninstalls and privileged steps to the user's own terminal. The plugin ships only the skill; the CLI is installed separately.
zhu1090093659/dsh-web#packages/dsh-skill-explorer★ 10458
Skill center for the dsh web GUI: browse all loaded skills grouped by source, enable or disable model invocation, create new skills, and delete into a recoverable trash.
QingYunA/answer-me-with-html★ 6365
Answers a question with a one-page visual HTML explainer: the model writes a short Markdown draft and the bundled `am` CLI builds the page, with flow, sequence, tree, timeline and ER diagrams, tables and callouts.
iflytek/skillhub#dsh★ 5432
Browse and search a SkillHub registry and manage local skill installation, updates, and removal inside DeepSeek Harness. Targets DSH 0.2.0-rc.2 and 0.2.1-rc.2; not yet compatible with 0.2.1-rc.3.
hyhmrright/brooks-lint★ 1578
Provides six code quality skills for code review, architecture audits, technical debt, test quality, health dashboards, and remediation.
GanyuanRan/Aegis★ 1443
Software-engineering method pack for coding agents, with skills for baseline-first planning, systematic debugging, prompt hygiene, verification before completion, and repair/retirement tracking.
Community comments
Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.