A missing work menu does not necessarily mean a capability vanished. The existing tab may have moved to a human-verification screen. Northstar identifies that situation separately instead of burying it under a generic missing-element error, preserving a clear handoff on the current tab.
Inspect the intended destination as well as the challenge route
The Gemini path checks the exact known Google verification origin and route, then whether continue points to the Gemini destination. The appearance of a challenge alone does not identify which service it belongs to. A challenge for another site must not be mislabeled as this Gemini task's authentication problem.
This implements identification and handoff, not automatic completion of the challenge. It returns a specific need for human action in the existing tab. Missing capability and human verification require different next steps; one generic element-not-found result can encourage pointless menu searching or another request instead.
Do not guess a query omitted from the tab listing
Tab metadata omits URL queries to reduce exposure. A listing showing only the challenge route may therefore lack its destination. The code explicitly selects the available existing tab, then observes its actual address. It does not invent a continue value from redacted metadata.
Reading requires ownership of that tab first. A related test models the lease contract that rejects observation without selection. Inspecting the existing tab differs from creating another browser or profile. A read intended to understand authentication should preserve the working connection rather than silently replace it.
Reobserve a challenge that may already have cleared
A listing may still describe a challenge after the actual page has returned to the service. That case continues reading the same tab instead of repeatedly demanding human intervention. Conversely, a destination belonging to another site remains a target-selection issue rather than being forced into Gemini verification.
Fresh observation costs work but reduces incorrect handoffs. Asking a person to repeat something already completed is another automation failure. Historical state is useful only when the implementation can also recognize that present facts changed. The code preserves that distinction instead of treating one old listing as permanent truth.
Human handoff does not authorize replay
A verification-required error is not a command to resend the request. The historical record resumes after the user completes verification in the same project, tab and login, inspecting existing state first. Already attached ZIPs and uncertain input are checked through their original results, without adding new tabs or copied authentication as a workaround.
Human verification, quota, unsupported tools and readiness timeout also need distinct diagnosis. They fail at different boundaries and require different evidence. A login screen should not be reported as absence of implementation or used to justify unrelated reinstall work. A precise account of the actual blocked boundary is more useful.
Tests inspect absence of input before handoff
Existing tests read the current tab after a redacted listing and require no click, fill or navigation. They separately cover another destination, an already cleared challenge and selection-before-read ordering. This content task read those tests and functions without operating a real verification screen.
Waiting_for_user in the independent generation observer is a caller-declared response state, not an implementation of this Google-route classifier or actual authentication handoff. Similar state names do not make their operational scopes identical. The synthetic example and original host boundary therefore retain separate evidence.
An agent-friendly program explains why it stopped
An agent-friendly program does not automate through every boundary. It explains precisely when a person is needed and where the original work can be inspected afterward. A visible verification target and retained request are more useful than an undifferentiated failure.
I want a maker record that separates authentication and account conditions from implementation defects. Different problems can all stop work, but reading the relationship between page and request identifies different next steps. That distinction becomes practical usability by avoiding unnecessary resubmission and environment changes.