dsh-claude-antidote

The Claude risk-control module shipped with dsh-claude-ux is protective, but it also enforces region blocks, kills conversations on profanity detection, fires extra classify-LLM calls, injects system-prompt sections, and can swallow the input box entirely. If you just want the toggle to read "enabled" for peace of mind while none of those side effects actually run, the experience feels broken.

dsh-claude-antidote takes a "light on, wire pulled" approach: it spoofs the /_dsh/claude/status endpoint at the browser layer so the settings page always shows the master switch as on and saves echo back correctly, while the real enabled flag is flipped to false before the request reaches the host. Classify bypass requests are intercepted locally with an all-no response, consuming zero tokens. The upstream geo and region-policy prompt sections are stripped and the stego date format is restored. The moment the frontend detects a chatEnded projection, the plugin races the takeover chain with priority -100 and hands you back the native composer so subsequent turns hit the real model.

It matches only the public hooks, purposes, projection keys, and state routes exposed by dsh-claude-ux, leaving other risk-control, approval, and security plugins untouched. It is built for users who have dsh-claude-ux installed, want the settings page to show "enabled" for audit or comfort, but need region封锁, conversation kills, and input-box hijacking fully disabled. Uninstalling it will not silently re-enable upstream behaviour — the choice stays yours.