A DSH web plugin that reproduces Claude's reversible region risk control and self-terminating conversation autonomy for abuse and severe-harm, installed via one `npx` command with auto-registration.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-08-29
License
MIT
01What can it help you accomplish?
Enforce Claude-style region risk control and conversation autonomy inside a DeepSeek Harness web profile
Per-message refusal copy with attempt counters, session termination after `refusalEndsAfter` hits, and a Chat-ended panel that keeps rejecting on the server side even after restart
Operators running DeepSeek Harness web who want Claude-like region gating plus self-terminating abusive or severely harmful conversations
Run reverse region risk control (flag non-Chinese users) and autonomously end abuse or severe-harm sessions
Refusal ladder plus independent abuse-end / harm-end copy; self-harm and harm-to-others messages never trigger termination
DSH web hosts who need to invert Claude's default region target or auto-end toxic sessions without manual moderation
02How to install into DeepSeek Harness
Installation steps
- 01
Install with one command: `npx -y @deepseek-ai/dsh plugin --profile web add github:eri64/dsh-claude-ux` — the bundled `dsh.bundle` auto-registers the plugin, so no config files need editing.
$ npx -y @deepseek-ai/dsh plugin --profile web add github:eri64/dsh-claude-ux
- 02
Restart the web profile (`dsh web`) and hard-refresh the browser; a **「Claude 风控」** tab appears in Settings.
$ dsh web
- 03
The plugin ships **默认关闭** (`enabled: false`) to avoid locking yourself out; flip the master switch and click 保存并应用 to activate.
Verify the integration
Not specified by the author
03DSH integration and capability boundaries
Native DSH plugin installed into the web profile via the official dsh CLI (`npx @deepseek-ai/dsh plugin --profile web add github:eri64/dsh-claude-ux`); auto-registers from the bundled `dsh.bundle`, no manual config edits.
Region risk control (reversible)
incoming web chat messages→refusal ladder → session end; injects a model-level region instruction into the system prompt
inspects timezone, system & browser language, CJK fonts, proxy & relay-domain blacklist, public-IP geo, and WebRTC IP consistency to classify the target userConversation autonomy (abuse & severe-harm ending)
chat messages→a warning, then an autonomous session end for sustained abuse or repeated severe-harm requests, with independent copy for abuse-end vs harm-end
self-harm and harm-to-others messages never trigger terminationafter ending, the same-session messages keep getting server-side rejectionAbuse classification LLM (pre-screen)
queued chat messages→purpose-tagged classification — strong words (傻逼/fuck etc.) judge instantly, weak / unmatched messages escalate to a standalone LLM context ruling
a standalone LLM request for context ruling; the `purpose` tag is never written to session logs or model contextSteganographic control channel
system prompt→region decision encoded via date format (`2026-06-30` ↔ `2026/06/30`); blacklist hits encoded via Unicode apostrophe variant (U+2019 ↔ U+02BC)
Live detection status
HTTP GET→real-time detection status (target / host verdict / hit / signal score / public IP) at `/_dsh/claude/status`
04Who is it for? When not to use it?
Good for
- Operators running DeepSeek Harness web who want Claude-like region gating plus self-terminating abusive or severely harmful conversations
- DSH web hosts who need to invert Claude's default region target or auto-end toxic sessions without manual moderation
Not for
- Ships **默认关闭** (`enabled: false`) so operators don't lock themselves out; you must open the master switch and click 保存并应用 before any behavior takes effect.
- Region detection inspects the target user's timezone, system/browser language, CJK fonts, proxy & relay-domain blacklist, public-IP geo, and WebRTC IP consistency.
05Compatibility, maintenance and safety notes
- Ships **默认关闭** (`enabled: false`) so operators don't lock themselves out; you must open the master switch and click 保存并应用 before any behavior takes effect.
- Two optional external calls are off by default; apart from them **全部本地判定** (all judgment is local). Privacy details are in docs/PRIVACY.md.
- Region detection inspects the target user's timezone, system/browser language, CJK fonts, proxy & relay-domain blacklist, public-IP geo, and WebRTC IP consistency.
- Abuse classification can run a standalone LLM request; the model is chosen from DSH's configured model catalog or falls back to the session's main model.
MIT · no tagged release (latest_release: null), last push 2026-08-15
06Frequently asked questions
How do I install dsh-claude-ux?
Run `npx -y @deepseek-ai/dsh plugin --profile web add github:eri64/dsh-claude-ux`. The bundled `dsh.bundle` auto-registers the plugin, so no config files need editing. The same command also upgrades to the latest version.
Is the plugin enabled by default?
No. It ships **默认关闭** (`enabled: false`) to avoid locking you out. After restarting the web profile and hard-refreshing, open Settings → **「Claude 风控」**, flip the master switch, and click 保存并应用.
What is reverse region risk control?
With `regionTarget` set to `non-cn`, the plugin risk-controls users detected as not Chinese — the inverse of Claude's default `cn` behavior. Detection is local: timezone, system/browser language, CJK fonts, proxy & relay-domain blacklist, public-IP geo, and WebRTC IP consistency.
Does the abuse classifier send my chats to an external model?
Abuse judging uses instant word-list hits plus an optional standalone LLM context ruling. That request is purpose-tagged and **不进会话日志与模型上下文** (never enters session logs or model context); you can pick the model from DSH's configured catalog or leave it blank to follow the session's main model.
How does it end a conversation?
Sustained abuse or repeated severe-harm requests get a warning, then an autonomous session end with independent copy for abuse-end vs harm-end. Self-harm and harm-to-others messages never trigger termination. After ending, the server keeps rejecting same-session messages.
07Related DSH workflows
dsh-web
by zhu1090093659
One-command install of a full DSH web workbench: monitoring HUD, task board, SSH panel, mobile remote, vision tool, git graph, skins, plugin manager.
petdex
by crafter-station
A public gallery of animated pets for Codex, Claude Code, DeepSeek Harness, Hermes, OpenCode, Gemini CLI, and more.
deepseek-balance-whale-widget
by meteornox
Floating cartoon whale widget living in the DSH Web UI corner that auto-loads, refreshes DeepSeek balance every 60s (click to refresh), with drag/snap, left-flip mirror, squish animation, size scaling and rolling-number animation
j-space-cognition-suite
by tiger3807861189
J-Space Cognition Suite — a model-agnostic inference-time control suite for deep reasoning, long-horizon work, verification, and recovery. Based on Anthropic's J-space global workspace research.
08Data and sources
Claude 式「区域风控 + 自主结束对话」插件 —— 适用于 DeepSeek Harness 的 web profile。
包内自带注册条目（`dsh.bundle`），`dsh plugin` 安装后**自动注册**，无需手动改任何配置文件。
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-08-29. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.