[Bug]: Optional 500 ms preview metadata timeout disconnects the automation host

Before submitting

• I searched existing issues and did not find a duplicate of the specific optional-metadata timeout problem described here.
• I included enough detail to reproduce or investigate the problem.

Related: #12146 covers host eviction without re-registration. This report identifies a specific server-side eviction trigger: a best-effort metadata lookup with a 500 ms deadline disconnects the host even though its failure is caught and omitted from the original tool result. Fixing re-registration would aid recovery; optional metadata should also avoid evicting the host in the first place.

Area

apps/server

Steps to reproduce

Observed sequence (intermittent; the exact cause of the delayed status response is not yet isolated):

1. Connect the Mac T3 Code desktop app to a Linux workspace through T3 Connect and open the thread's Browser panel.
2. Open a private HTTPS web preview. Successfully use browser tools to navigate, fill forms, evaluate page state, and record flows.
3. Continue normal browser automation. One captured occurrence followed a sign-out action; another occurred around recording completion. These are observed contexts, not established deterministic triggers.
4. An optional post-action broker status request exceeds its 500 ms deadline. The trace below records the broker disconnect at that deadline.
5. Subsequent navigate, evaluate, recording-stop, open and status calls report no available host. Restarting the Mac app and reconnecting restores control.

Suggested deterministic regression reproduction, not yet executed: let the primary tool operation succeed, delay its secondary metadata status response beyond 500 ms while the transport remains connected, then invoke another browser operation. Assert that the metadata timeout does not remove the registered host.

Expected behavior

The optional lookup for toolIcon.pageUrl can fail without invalidating the browser automation host. Return the primary action result without metadata and keep subsequent browser operations usable. A slow metadata response should not require a desktop restart.

Actual behavior

Browser automation repeatedly became unavailable during normal use. The optional metadata request timed out after 500 ms and the broker disconnected the host; subsequent tools failed with No preview automation host is available.

Both machines remained online, and the private preview URL still returned HTTP 200 from Linux. The failure interrupted authenticated browser verification and prevented exporting one active recording.

Read-only inspection of the bundled Linux server, resources/app.asar/apps/server/dist/bin.mjs, found:

• PreviewToolkit.invoke adds a secondary broker status invocation after most operations to obtain toolIcon.pageUrl. It sets timeoutMs: 500 and updateCurrentTab: false, and catches lookup failure by returning null.
• status, open, navigate, and snapshot are excluded from this extra lookup.
• PreviewAutomationBroker.awaitResponse handles timeout by calling disconnect(connection.clientId, connection.queue) before raising PreviewAutomationTimeoutError.
• Therefore, catching the optional metadata lookup error in the toolkit does not undo its host-disconnection side effect. This provides a concrete mechanism by which a successful-looking action result can be followed by persistent no-host failures. This matches the captured timeout/disconnect sequence; it does not establish what delayed the status response.

In this installed build, the relevant sections are around bundled lines 198290–198310 (metadata lookup) and 164680–164700 (timeout/disconnect). These are generated-bundle locations, not asserted upstream source line numbers. The T3 installation was not patched or restarted on Linux.

Suggested fix: distinguish optional metadata failure from host-liveness failure. Add coverage for a successful primary action followed by slow metadata, a response arriving after its deadline, and subsequent operations on the same host. Host re-registration is separately discussed in #12146.

Impact

Major degradation or frequent failure

Version or commit

T3 Code Alpha 0.0.42 on Mac and Linux. Linux version verified from the installed app package; findings refer to that installed bundle, not an assumed current upstream revision.

Environment

T3 Connect with workspace/agent/server on Linux and the embedded browser on macOS. Mac preview user agent: Chrome 152.0.7977.65, Electron 44.1.0. Exact macOS version not collected. Production-mode SvelteKit preview over valid private Tailscale HTTPS. No TLS or authentication bypass. Observed September 17, 2026.

Logs or stack traces

The Linux server trace shows this sequence at approximately 2026-09-17 16:23:02 UTC:

16:23:02.090  PreviewAutomationBroker.invoke / awaitResponse (status)
16:23:02.590  PreviewAutomationBroker.disconnect / closeConnection
             PreviewAutomationTimeoutError:
             Preview automation status timed out after 500ms.
16:23:02.600  PreviewAutomationNoAvailableHostError:
             No preview automation host is available for navigate in environment <id>.
16:23:02.603  Same no-host failure for evaluate
16:23:02.614  Same no-host failure for recordingStop
16:23:06.709  PreviewAutomationBroker.respond completes successfully

The later respond span shows subsequent response handling, but its trace does not identify the request in the displayed fields; it is not proof of which request replied late. Focus-host RPC spans also succeeded while normal operations remained unavailable.

Screenshots, recordings, or supporting files

Relevant sanitized trace and installed-code findings are included above. Full traces and application recordings are omitted because they contain unrelated session/application data. No credentials, token-bearing URLs or private hostnames are included.

Workaround

Restarting T3 Code on the Mac and reconnecting restored automation. The final verification session completed using explicit navigation tools and form submissions separated from document navigation, without modifying T3. That successful session does not establish a reliable workaround.

Start by locating the upstream implementations corresponding to PreviewToolkit.invoke and PreviewAutomationBroker.awaitResponse, using the bundled sections around lines 198290–198310 and 164680–164700 as references. Add regression coverage for a successful primary action followed by delayed optional metadata, a late response, and subsequent operations on the same host; done means the metadata timeout no longer disconnects the host.