Treat any tag/markup coming from user input, LLM output, or external commands as untrusted. Validate and sanitize it at the earliest boundary before embedding into JSON or rendering in the UI—e.g., strip/escape disallowed tags so they can’t break parsers or visuals. Example (shell boundary sanitization): ```sh # sanitize_tags.sh # Remove/neutralize disallowed tags before downstream UI/JSON handling sanitize_tags() { # Example: drop ... sed -E 's#<\/?thinking[^>]*>##g' } INPUT="$1" SAFE_INPUT="$(printf '%s' "$INPUT" | sanitize_tags)" # Use $SAFE_INPUT for any JSON generation or UI rendering ``` Apply this by: (1) defining allowed vs. disallowed tags, (2) sanitizing/escaping before JSON construction and before UI templating/rendering, and (3) adding automated tests for payloads that previously broke visuals/JSON (e.g., special tags).