Let approval input flow through the LLM instead of being short-circuited in the REPL. The LLM calls plan_approve itself, which is cleaner (single input path) and more flexible (no hardcoded misspelling list).