Part of #596. Frontier: ready now; implementation independent of #597 page layout.
Provide an authenticated, bounded, harmless real tool request/result roundtrip through the intended ChatGPT Web integration. ADR-0019 explicitly says doctor flags, text smoke, synthetic receipts and merely opening connector settings do not prove coding capability. Specify the probe's exact no-side-effect tool, client/runtime seam, correlation, timeout, and sanitised success/failure evidence in code/docs before marking completion.
Acceptance:
- A public Runtime Settings endpoint (or existing compatible endpoint) initiates and observes one real tool request/result roundtrip. It does not execute arbitrary caller-supplied commands or expose credentials. Same-origin settings session, bounded timing, duplicate clicks/cancellation and stale outcomes are handled.
- Success is scoped to current account/configuration/runtime generation; changing account, Tunnel, connector or restarting invalidates stale success. Failure leaves coding setup incomplete with actionable category.
- Normal status flags never impersonate probe success. Secret/error redaction retained.
- Deterministic seam tests prove success, missing connector/tunnel, timeout, stale/duplicate response and no side-effect behavior; separately record whether a real live run was possible.
Part of #596. Frontier: ready now; implementation independent of #597 page layout.
Provide an authenticated, bounded, harmless real tool request/result roundtrip through the intended ChatGPT Web integration. ADR-0019 explicitly says doctor flags, text smoke, synthetic receipts and merely opening connector settings do not prove coding capability. Specify the probe's exact no-side-effect tool, client/runtime seam, correlation, timeout, and sanitised success/failure evidence in code/docs before marking completion.
Acceptance: