Skip to content

Migrated host (flow.google.com): settings trigger button relocated to agent-footer-actions, requires composer interaction first #749

Closed

Description

@lopezdu53

Summary

On an account migrated to flow.google.com (the "agentic" Flow UI cohort mentioned in #183/#174), READY_ANCHOR = ".settings-trigger-button" resolves to a <button> that is permanently hidden in the DOM. The real settings button now lives elsewhere in the page and only becomes interactive after the composer has been focused/used.

Environment

  • gflow-cli version: 0.71.0
  • OS: Windows 11
  • Commands affected: gflow video t2v, gflow image t2i
  • Account: migrated to flow.google.com (agentic UI cohort)

What I found

1. The old selector target is dead, not missing:

<button hidden="" matbutton="" flow-button="" cdkoverlayorigin=""
  aria-label="Settings trigger"
  class="mdc-button mat-mdc-button-base settings-trigger-button mdc-button--unelevated
         mat-mdc-unelevated-button mat-unthemed flow-button-primary flow-button-small">
  …
</button>

It's present in the DOM (57 locator matches in the wait_for retry loop) but hidden is set directly on the element — it will never become visible, so Locator.wait_for(state="visible") always times out at 30s.

2. The real settings button lives in agent-footer-actions, next to the prompt composer:

<button _ngcontent-ng-c1385090464="" flow-icon-button maticonbutton
  class="mdc-icon-button mat-mdc-icon-button mat-mdc-button-base mat-mdc-tooltip-trigger
         agent-action-button mat-unthemed flow-icon-button-transparent flow-button-small"
  mat-ripple-loader-class-name="mat-mdc-button-ripple" mat-ripple-loader-centered
  aria-label="Configuración">
  <span class="mat-mdc-button-persistent-ripple mdc-icon-button__ripple"></span>
  <mat-icon role="img" class="mat-icon notranslate flow-icon-m google-symbols mat-icon-no-color"
    aria-hidden="true" data-mat-icon-type="font">tune</mat-icon>
  …
</button>

Confirmed via playwright codegen against a real logged-in session — recording a manual click on this button produced:

page.get_by_role("button", name="Configuración", exact=True).click()

(Account locale is Spanish; presumably "Settings" in English locales — worth confirming across locales the way #24 handled locale-agnostic selectors.)

3. The new button is not available until the composer is touched.

In the codegen recording, the button only became clickable after this sequence:

  1. Click into the prompt composer ([contenteditable='true'])
  2. Type text into it
  3. Then the Configuración button responds to a click

I patched _open_pane to click the composer first (await page.locator(COMPOSER).first.click(timeout=5000) before locating READY_ANCHOR) and updated READY_ANCHOR to role=button[name='Configuración']. This got further but still timed out — I did not confirm whether composer focus alone is sufficient or whether it needs actual text content first (my test used an empty click, not a fill()). Wanted to flag this before going further since it touches apply_video_settings's call ordering relative to send_prompt, which looked like a bigger structural change than I was comfortable making blind.

Suggested fix direction

  • Update READY_ANCHOR for the migrated host to target the agent-action-button with aria-label="Configuración" / "Settings" (locale-aware, per Make UiAutomationTransport selectors language-agnostic (structural-first) #24's pattern) containing a mat-icon with text tune, rather than .settings-trigger-button.
  • Investigate whether apply_video_settings (called at line ~1528, before send_prompt) needs to be reordered to run after the composer has content, for the migrated/agentic host specifically — the classic host's flow (open settings pane first) may not hold on the agentic cohort.

Incident bundles

Happy to attach report.md / ui.json from my incident bundles if useful — redacting the account-identifying screenshot first.

Activity

  1. ffroliva commented on Sep 8, 2026

    @ffroliva
    Owner

    Reproduced, root-caused and fixed — thank you, the DOM dump made this quick. PR: #751.

    What it is: Flow's composer carries an agent-mode chip, button.agent-mode-chip[aria-pressed]. Pressed, the app swaps flow-prompt-box for flow-creative-agent-prompt-box — and that is what puts hidden on .settings-trigger-button. Your agent-footer-actions block is the replacement composer's own footer. It is not selector drift and not a cohort: it is a mode, and gflow's migrated driver had no idea the mode existed.

    I probed it on two of our accounts (scripts/dev/spike_migrated_composer_arms.py, $0 — navigation and DOM reads only). One click on the chip reproduces your DOM exactly:

    chip aria-pressed=false chip aria-pressed=true
    .settings-trigger-button hidden=false, display: flex, 147x32 hidden=true, display: none, 0x0
    component chain flow-prompt-box flow-creative-agent-prompt-box
    div.agent-footer-actions absent present
    button.agent-action-button + tune absent present

    Flow remembers the chip per account, which is why it looked permanent to you: one click in a browser, and every later gflow run lands in agent mode with nothing on the CLI to say so. The fix leaves agent mode before the readiness gate.

    Two notes on the direction you'd sketched, both meant as detail rather than correction of your diagnosis, which was right:

    1. Retargeting READY_ANCHOR would have broken the classic arm — the default state, and what both of our accounts render. It needed to be a recovery, not a swap.
    2. get_by_role("button", name="Configuración") can't be used. This project bans text-label selectors outright (AGENTS.md, locale-invariance) — you flagged the locale risk yourself, and it's real: our accounts render en/en-GB and would never match. The chip's component class plus aria-pressed carry the same identity everywhere.

    And the good news on the part you were unsure about: you were right to stop before reordering apply_video_settings — that turned out to be unnecessary. Your "the composer must be touched first" observation was about the agent prompt box; once the run leaves agent mode the question disappears. No structural change was needed.

    Also worth separating: you list gflow image t2i as affected. That one never reaches this code — it's refused earlier by a host guard (exit 36, client.py), which is a different problem. If you're seeing something other than a clean exit-36 refusal there, please open a separate issue.

    No bundle needed — the spike captured everything. Thanks for redacting before offering, though.

    Fix is on develop shortly. If you'd like to verify before it ships, the workaround today is to click the Agent chip off in a browser once on that account.

  2. ffroliva commented on Sep 8, 2026

    @ffroliva
    Owner

    Merged to develop in #751 — this will be in the next release.

    Thank you for this one. It was an unusually good report, and it's worth naming why, because it's what made a same-day fix possible:

    • You pasted the actual DOM of both the dead element and its replacement, rather than just the error. The hidden="" attribute and the agent-footer-actions wrapper were the whole diagnosis.
    • You noticed the element was present but hidden and counted the retries (57 matches). That single distinction is what separated "Google changed the frontend" from "the app is in a different mode" — and it's the distinction the fix is built on.
    • You stopped where your evidence stopped. You flagged the composer-interaction question as unconfirmed, said explicitly that you'd used an empty click rather than a fill(), and declined to make a structural change blind. That was the right call, and it turned out to save work: leaving agent mode makes the whole ordering question disappear, so no restructuring was needed.
    • You raised the locale risk on "Configuración" yourself, and you were right — it's a hard rule here, and your instinct matched it.

    For the record, the final anchor is button.agent-mode-chip[aria-pressed='true']. The 'true' is doing real work: it makes the locator self-guarding, so the recovery can only fire when there is something to undo and can never click a healthy composer into agent mode.

    Two things your report led to beyond the fix itself: a reusable $0 probe (scripts/dev/spike_migrated_composer_arms.py) that inventories both composer arms, and a written finding at docs/superpowers/spikes/2026-09-08-migrated-composer-agent-mode-hides-settings.md. Your incident bundles weren't needed in the end — but thank you for offering them redacted, that's the right habit.

    If anything still misbehaves on that account after the release, please reopen.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions