[kvz]

Why

The MCP Apps result widget only rendered the tool result it was opened with. When ChatGPT showed it for transloadit_create_assembly (or a Builtin shortcut) while the Assembly was still running, the card said "Processing · Result files appear when the Assembly completes" forever, even after the Assembly had finished (seen in production with 0.7.1 on 2026-10-19). Demos only looked right when the model happened to call transloadit_wait_for_assembly, which rendered a second widget. Users should see their results no matter how the model waits.

What

• While the Assembly is uploading, executing or replaying (the same states the SDK's awaitAssemblyCompletion polls on), the widget calls the read-only transloadit_get_assembly_status through the host's tool bridge: MCP Apps tools/call when the host declares serverTools, otherwise ChatGPT's window.openai.callTool. It passes assembly_ssl_url (or the Assembly ID) from the result it already has, so no new data reaches the widget.
• Checks back off from 2 s by 1.5x to every 10 s, for at most 15 minutes of waiting. Each call times out after 30 s. Checking stops on a terminal state (completed, failed, canceled, aborted), on ui/resource-teardown, and pauses while the document is hidden (it checks again soon after it is back in view).
• After 15 minutes, or 3 failed checks in a row, a "Check again" link joins the busy note (same line, so nothing else moves) and restarts the checks. Hosts without a tool bridge keep today's text and get no button.
• A refreshed status replaces only the Assembly in the host's result, so Step labels and Console/Template links stay. The create call's errors and warnings are dropped once the Assembly has finished, since they no longer apply. Identical repeats of the host's original result (ChatGPT re-sends toolOutput on every theme or height change) no longer bring the old "Processing" state back.
• transloadit_get_assembly_status now declares openai/widgetAccessible: true (ChatGPT's opt-in for callTool) and an explicit ui.visibility: ['model', 'app'].
• The widget never fetches the Assembly status URL itself. Its CSP only lists result file origins, and the tool bridge reuses the caller's credentials.

Tests

• Unit (happy-dom, widget-handshake.test.ts): running to completed over tools/call with the exact 2 s / 3 s / 4.5 s delays; stops on failed, canceled and aborted; caps at 10 s and offers "Check again" after the 15 minute budget, which then finishes the job; "Check again" after 3 failures; no calls and unchanged text without a bridge; stops on teardown; pauses while hidden; the ChatGPT callTool path, including a later openai:set_globals not reverting to "Processing".
• tool-surface.test.ts checks the new status tool metadata.
• Playwright widget states: the fake host can now proxy tools/call. New states refreshed-to-completed (executing, then completed through the bridge) and check-again, plus a stability test that the transition happens without any new host result, keeps the upload preview node and fetches nothing twice. Baselines rendered in the official Playwright image (test:widget-visual); existing baselines are unchanged.

Closes #589

_{🤖 Claimed by kvz on framework1/node-sdk since 2026-10-19 · resume: claude --resume 26eaa88e-b3fa-4cc1-bcce-0c7ff5b6bab7}