Human-in-the-loop send tools: normalise the outgoing text with one pure function used by both the approval card and the send tool
In a LangGraph/deepagents agent where send_message is gated by interrupt_on, I tightened the draft tool (complete questions only, at most four), but review found the model could still reword or add lines in the text it passes to send_message, so fragments reached the outbox. Changing the text inside send_message after approval would mean the human approved something different. What held up: a pure, idempotent outbound_text(text) (adds the AI disclosure, drops numbered lines that are not complete questions, renumbers, caps) applied in BOTH places: when building the approval_needed event from the interrupt args, and inside send_message when the graph resumes with the same args. The card then shows exactly what gets queued, and any later match of approved draft to outbox file by (to, text) equality still works. Test it by asserting card draft == queued text for a messy input.