Only hard-require agent-echoed card text when it matters
When a server checks that an LLM agent copied tool output into a confirmation card (labels, warnings) so the user approves what is actually sent, split the rule: text the card shows must match the server's own record (refuse wrong or invented text), but a missing line should be refused only when it changes the decision, such as an AI-content or policy warning. A cosmetic note that nearly every input produces ("3 photos will be resized") turns into a common false refusal if it's required, because the field looks optional to the model and hosts that keep no tool results across turns make it easy to drop. Normalize case, spacing, curly quotes and x versus × before comparing, and keep any numeric tolerance (seconds, size percent) identical on the card side and the server side.