Migrated host: agent-only composer (no classic composer) — video t2v --project fails with UiSelectorDriftError, .settings-trigger-button hidden, no agent-mode-chip

Environment

gflow-cli 0.73.1 built from ai4U23/gflow-cli@3387861 (PR fix: migrated-host auth verification + i2v picker confirm + run-unique upload #793) — the official 0.73.2 cannot log in on this account (auth: login verification never passes for accounts migrated to flow.google.com (labs NextAuth absent → GOOGLE_SESSION_ONLY forever) #791)
Windows 11, Google Chrome (profile created with gflow auth login --browser chrome), Python 3.13
Account migrated to flow.google.com, Google UI locale: ru
Auth: gflow auth status → verified via the fix: migrated-host auth verification + i2v picker confirm + run-unique upload #793 migrated-host fallback

Command

gflow video t2v "<prompt>" --project <existing flow.google.com project id> --model veo-lite --duration 4 --aspect 16:9 --count 1 --json

(Without --project it fails earlier with 401 at labs.google/fx/api/trpc/project.createProject, as expected per #639.)

Result

Exit 23, before submit (no credits spent):

migrated host: the settings trigger (.settings-trigger-button) did not become visible within 30s on https://flow.google.com/project/<id> (host=migrated)
  55 × locator resolved to hidden <button hidden="" ... aria-label="Settings trigger" class="... settings-trigger-button ...">

_exit_agent_mode returned absent — no button.agent-mode-chip[aria-pressed='true'] on the page.
Incident bundle: ui.json has all tag counts = 0 and the screenshot is blank white (same symptom as #792, but landscape 16:9 here).

What the project page actually shows (checked manually in a normal browser)

There is no Agent toggle/chip anywhere.
The only composer is the right-hand agent panel: session title ("Untitled session"), greeting "What do you want to create?", suggestion cards, and a prompt box at the bottom.
Next to the prompt box there is a "Settings" button. It opens an "Agent settings" pane with:
- "Confirm before generating": Always / Never
- Default image generation: aspect 16:9 / 4:3 / 1:1 / 3:4 / 9:16, count x1–x4, model (Nano Banana 2)
- Default video generation: aspect 16:9 / 9:16, count x1–x4, model dropdown (Omni 1.1 Flash)
- "Save" button
So on this cohort the classic migrated composer that MigratedComposer.ensure_editor waits for does not exist; model/aspect/count are agent defaults, not per-request settings.

Expected

Either a driver for the agent-only composer (set video defaults in Agent settings → type prompt → approve/"Never confirm" → wait → download), or a typed, pre-submit error that names this cohort instead of generic selector drift.

This is a third cohort, and it is not the one we already handle. We have two migrated-host arms on record (#749): a classic composer, and an agent composer reached by button.agent-mode-chip[aria-pressed='true']. When that chip is pressed, .settings-trigger-button gets a bare hidden attribute — exactly the DOM you captured. So #751's recovery was built for a state that looks identical to yours from the outside. It un-presses the chip. You have no chip to un-press.

That is the important part of your report: yours isn't a stuck mode we can recover from, it's a composer that has no classic arm at all. Your observation that model / aspect / count live in Agent settings as defaults rather than per-request controls is the same finding from the other side — there is nowhere for --aspect 16:9 to go on that page.

Two things follow.

The generic exit 23 is our bug, and it's the cheap half. ensure_editor waits 30 s for a control that is structurally absent and reports selector drift. It should detect the agent-only composer before submit and fail with a typed error that names the cohort. That's next on our list and doesn't need your account.

A driver for the agent-only composer is the real feature and is a larger piece of work. We're not going to promise a date for it, but it's tracked here.

On your incident bundle — the all-zero ui.json and white screenshot are a separate bug of ours, not a property of your page: the driver parked the page on about:blank in a finally before the incident was captured, so we screenshotted the blank page. Fixed in #794, shipped in v0.73.2. That won't help you directly while #791 keeps you on a fork of 0.73.1; if you want a real bundle before 0.74.0, cherry-picking #794 onto your branch is a small, self-contained commit.

And to state the obvious constraint: you are blocked by #791, which is why you're running someone else's fork at all. That's the one we're fixing first. We're deliberately not cutting 0.74.0 until #791 and #795 are in it, because a release before then reaches neither you nor @ai4U23.

fix(799): close the council's findings, and measure the live controls

fix(799): narrow the retryable contract instead of widening six const…

fix(799): name the agent-only composer instead of calling it selector…

@Cstanish127

A driver for the agent-only composer is up for review in #824.

It was built and verified on an account in this cohort (Google AI Pro, en locale, no agent-mode-chip):

gflow image t2i — 3:4 × 2 images, and the MCP gflow_generate_image twin
gflow video t2v — a 4 s 9:16 clip, downloaded as MP4

How it works: aspect / count / model are set in Agent settings for the run and restored afterwards; the request goes to the agent in plain language; completion is read from the page. "Confirm before generating" follows a new GFLOW_CLI_AGENT_CONFIRM=account|always|never (default leaves your account's setting alone).

Not covered yet on this composer (exit 36 before anything is clicked): image-to-image, image-to-video, reference-to-video, character references, Agent instructions, and more than one video per request.

@Cstanish127 — one thing we could not measure: the request text is English, and your account is ru. If you can try the branch, whether the agent accepts it there would be very useful to know.

@ffroliva

Thanks @ffroliva for the thorough reply, and @WAOmaster for #824 — very welcome.

On trying the branch on a ru account: I can't run it as is. #824 is based on develop, and my account still can't pass login verification there (#791 is open), so gflow stops before the composer. If there is a way to try it with the #793-style fallback on top — or once #791 lands — I'm happy to run the t2i / t2v e2e on ru and report back.

Meanwhile I've been using the agent-only composer daily through small page-side JS snippets (no gflow), since 2026-09-14. Account: Google AI Pro, UI locale ru, ~60 Veo 3.1 Lite clips (1–3 image ingredients on most of them), a few Nano Banana 2 images, several projects. Below is what might be useful for #824 and its review. It's one account, so treat it as data points, not a spec.

Directly relevant to the review / open questions

English request text works on a ru account. Every one of my ~60 clips was requested in English on the Russian UI, and the agent accepted them. My wording was different from compose_video_directive, though: "Generate exactly one video clip using the default video settings. Do not generate images first and do not ask follow-up questions." followed by the description. I never tried "in a 16:9 aspect ratio" in the prose. I avoided it on purpose: my notes say an agent can read "16:9" as a count. That's a hunch, not a measurement on this composer.
"Confirm before generating" radios, DOM order on ru: input[type=radio] value 1 = «Всегда» (Always) first, value 2 = «Никогда» (Never) second — matches _CONFIRM_INDEX = {"always": 0, "never": 1} (review item 3), on one more account and locale.
The gate is not guaranteed in either mode. With «Никогда» no gate appears at all — the submit itself spends credits. With «Всегда» I once got an image generated without any gate. The "at most one confirmation" logic probably has to treat "zero gates" as a normal path.
Agent settings are per project, not per account. A new project came with «Всегда» and Omni 1.1 Flash, while my older project had «Никогда» and Veo 3.1 - Lite. Snapshot and restore should probably be keyed to the project (relevant to review item 1).
Requested count is not always honoured. With image default x2, the agent returned 1 image once. That's the under-count mirror of review item 2.

Outcomes the driver may need to recognise

Generation error tile: flow-error-tile («Ошибка Не удалось сгенерировать аудио…» — failed to generate audio) appears instead of a video tile. No pending tile is left and no new uuid comes, so a "pending == 0 and a new uuid" loop would wait until timeout. The count of error tiles grows by one. But old error tiles also mount lazily after a page load, so I only trust growth once nothing is pending.
Content refusal without any tile: the agent just replies («Похоже, этот контент нарушает наши правила» — this content seems to violate our rules). No pending tile, no error tile, no media. Seen repeatedly for scenes with children in a swimming pool, even with a long-sleeved rash guard. For video I stop when there is no pending tile, ≥20 s have passed since submit, and the last reply matches such a pattern — similar to IDLE_S for images.
Upload rejected: some image uploads fail with only a toast «Не удалось загрузить .» (could not upload), no reason given. In my case it was photos of a real child; a photo of an adult in the same format uploaded fine. Relevant once references are ported. Worth a typed, non-retryable error.

Result detection — pitfalls I hit

Lazy virtual grid (.scroll-viewport). Newest tiles don't render while the grid is scrolled down, and older tiles mount after the baseline was taken. "Not in the baseline" once made me re-download an old clip. What has held since: the grid is scrolled to top, the first tile is unknown, and no flow-pending-tile is left. That is the same hole as the surviving _tile_key(head) != baseline_head mutation.
Tiles without any media src. After an extension reconnect, 38 flow-video-tile elements had no img/video src at all. The finished clip was only visible as flow-a2ui-video-option img.video-thumbnail in the chat, which options in feat(migrated): drive the agent-only composer for text-to-image and text-to-video (#799) #824 already reads. Good call.
Hidden tab. With document.visibilityState === 'hidden' (minimised window, locked screen), the ingredients popover stopped applying selections, lazy tiles lost their posters and page JS stalled. A visibility check before submit could turn that into a clear error instead of a timeout.
Collapsed agent panel. After saving Agent settings the panel sometimes collapsed to a short composer with no flow-agent-panel. It came back via the expand button (aria «Развернуть» on ru — I only have the label, not a structural anchor).

Ingredients in the agent-only composer (for when i2v / r2v are ported)

The flow that works for me:
- «Добавить ингредиенты в окно запроса» (add ingredients) opens flow-add-menu-popover-content;
- project assets are listed as [role=option] in flow-add-menu-asset-list;
- clicking a recent asset attaches it straight away as flow-image-ingredient-chip in the prompt box;
- an item from the «Все» (all) section is only selected and needs the «Добавить в запрос» (add to prompt) button;
- the chip's img src carries the uuid.
I count chips after every attach, because a stray click can leave an extra one. Popover thumbnails load lazily, so matching by asset name was more reliable than by uuid. Uploads: «Загрузить» creates <input type=file multiple> and calls .click() on it. The first upload in a project shows flow-upload-consent-dialog.
«Начать новый сеанс» (new session) empties the chat (flow-chat-empty-state) and clears attached chips. I start every clip with it so earlier turns don't leak into the next request.

Timings on AI Pro

Timings:
- Veo 3.1 Lite: the agent says "queued due to high demand", and the clip is ready in about 60–140 s.
- Nano Banana 2: ~35–50 s.
- Output via the tile menu, «720p Исходный размер» (original): h264 1280×720 24 fps + AAC, 8.0 s. The file is named after the agent's title.

Happy to run a specific probe or capture something on this account if it helps.

spike(799): drive the agent-only composer end to end — image and video

feat(config): GFLOW_CLI_AGENT_CONFIRM for the agent-only composer's c…

refactor(migrated): ensure_editor reports the composer kind; agent-on…

feat(migrated): drive the agent-only composer for t2i and t2v (ffroli…

fix(migrated): refuse up front only what neither composer takes (ffro…

fix(migrated): wait for aria-checked after an Agent-settings radio cl…

fix(migrated): agent-only video readiness from the pending tile, not …

docs+test(799): agent-only t2i/t2v documented, BDD-bound e2e for CLI,…

docs(799): agent-facing skill and USAGE exit-25 row describe the driv…

feat(config): GFLOW_CLI_AGENT_CONFIRM for the agent-only composer's c…

refactor(migrated): ensure_editor reports the composer kind; agent-on…

feat(migrated): drive the agent-only composer for t2i and t2v (ffroli…

fix(migrated): refuse up front only what neither composer takes (ffro…

fix(migrated): wait for aria-checked after an Agent-settings radio cl…

fix(migrated): agent-only video readiness from the pending tile, not …

docs+test(799): agent-only t2i/t2v documented, BDD-bound e2e for CLI,…

docs(799): agent-facing skill and USAGE exit-25 row describe the driv…

fix(migrated): discard the Agent-settings draft when applying default…

docs(799): plan — branch review and PR approval recorded

chore(plugin): regenerate the shipped gflow-cli skill after the ffrol…

Merge upstream PR ffroliva#824: agent-only composer driver (ffroliva#799

Merge pull request #989 from ffroliva/feature/799-agent-only-composer

Fixed by #989 (v0.86.0): gflow now drives the agent-only composer for image t2i and video t2v. Reading the credit gate before approving it is tracked in #988. i2i, i2v and r2v on that composer exit 36 as "not ported yet".

Navigation Menu

Repository navigation

Migrated host: agent-only composer (no classic composer) — `video t2v --project` fails with UiSelectorDriftError, `.settings-trigger-button` hidden, no `agent-mode-chip` #799

Environment

Command

Result

What the project page actually shows (checked manually in a normal browser)

Expected

Related

Directly relevant to the review / open questions

Outcomes the driver may need to recognise

Result detection — pitfalls I hit

Ingredients in the agent-only composer (for when i2v / r2v are ported)

Timings on AI Pro

5 remaining items

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Migrated host: agent-only composer (no classic composer) — video t2v --project fails with UiSelectorDriftError, .settings-trigger-button hidden, no agent-mode-chip #799

Description

Environment

Command

Result

What the project page actually shows (checked manually in a normal browser)

Expected

Related

Activity

ffroliva commented on Sep 13, 2026

WAOmaster commented on Sep 15, 2026

Cstanish127 commented on Sep 15, 2026

Directly relevant to the review / open questions

Outcomes the driver may need to recognise

Result detection — pitfalls I hit

Ingredients in the agent-only composer (for when i2v / r2v are ported)

Timings on AI Pro

5 remaining items

ffroliva commented on Oct 24, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions

Migrated host: agent-only composer (no classic composer) — `video t2v --project` fails with UiSelectorDriftError, `.settings-trigger-button` hidden, no `agent-mode-chip` #799