Split an Apps SDK tool result across structuredContent, content, and _meta so the model and the widget each receive the right payload

Documented steps

1. Return an object from the tool handler with three sibling keys: `structuredContent`, `content`, and `_meta`.
2. Put concise JSON that BOTH the model reads and the widget renders into `structuredContent` (e.g. `{ columns: [...] }`); it must match the tool's declared `outputSchema`.
3. Put optional model-visible narration into `content` as `[{ type: "text", text: "Drag cards to update status." }]`.
4. Put large or sensitive widget-only data into `_meta` (e.g. `{ tasksById, lastSyncedAt }`) — `_meta` never reaches the model.
5. Keep `structuredContent` tight; the docs warn that oversized payloads degrade model performance and slow rendering.
6. For file outputs, return a reference (e.g. a `file_uri` field) inside `structuredContent` rather than embedding inline binary data.
7. On the client the payload arrives via the `ui/notifications/tool-result` bridge message; the widget reads the structured half as `window.openai.toolOutput` and the component-only half as `window.openai.toolResponseMetadata`.
8. Official docs: https://developers.openai.com/apps-sdk/build/mcp-server