dsh-claude-ux

For DeepSeek Harness web profiles, replicating Claude's region-based risk control and autonomous conversation ending usually requires custom code or manual moderation. dsh-claude-ux brings these two Claude behaviors natively to Harness, letting administrators restrict access for target user groups and automatically end conversations when abuse or severe harmful requests occur.

The regional risk control works both ways: target cn to gate Chinese users, or switch to non-cn for reverse gating, blocking anyone who is not Chinese. It scores multiple signals including timezone, language, Chinese fonts, proxies, domain blacklist, public IP geolocation, and WebRTC IP consistency. On hit, it applies an escalation ladder: refusal messages with attempt counters, then conversation termination with persistent server-side rejection, and optionally model-level regional instructions injected into the system prompt. For autonomy, abuse detection combines instant wordlist matching with LLM contextual fallback; LLM classification requests are purpose-tagged, never enter the conversation log or model context, and are asynchronously pre-classified for near-zero added latency. Severe harmful content (e.g., CSAM, terrorism, mass violence) ends the conversation immediately without warning, while self-harm or harm-to-others messages never trigger termination. All warning and ending messages are independently customizable, with built-in defaults shown as gray placeholder text.

By default the plugin performs all detection locally and makes no external calls unless optional IP lookup or WebRTC checks are explicitly enabled, making it suitable for privacy-conscious self-hosters and deployments with compliance needs. The settings page provides an instant-apply master switch, risk target, thresholds, and text overrides, and the classification model can be picked from DSH's already configured model catalog. It is ideal for DeepSeek Harness administrators who want Claude-grade risk control and conversation governance from a single plugin.