Human-in-the-loop workflows

Add approval gates to agent-generated email so a human signs off before anything goes out.

The canonical pattern

1. Agent researches the recipient and composes a draft via POST /v1/inboxes/{id}/drafts. The response has a draft id.
2. Your app notifies a reviewer (in-app banner, Slack DM, email to an approver mailbox, whatever you use).
3. Reviewer opens the draft in your UI. Your UI calls GET /drafts/{id} to show the current state.
4. Reviewer edits inline — PATCH /drafts/{id} — or clicks Send — POST /drafts/{id}/send — or discards — DELETE /drafts/{id}.

Why drafts not "pending sends"

We deliberately don't have a "queue a send for later approval" endpoint. A draft IS the approval object. This keeps the API surface small and the state model unambiguous: a message either exists as a draft (not yet sent) or as a message (delivered to the relay). There's no third "pending approval" limbo state that can get wedged if your approval service is down.

Audit trail

Pair human-in-the-loop with an HTTP webhook on message.sent. Every send produces a webhook call you can log in your audit system, correlated by draft id → message id. You can also store reviewer metadata on the draft via the inbox.metadata field, if you need per-draft notes.