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

CHANGELOG.md

@@ -7,6 +7,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+
+- **On `flow.google.com`, an agent-only composer now exits `25` (`FlowAgentUiError`) where
+  it previously exited `23` (`UiSelectorDriftError`)** — and reports `retryable: false`
+  rather than inheriting that class's retryable default. Scripts branching on `23` for
+  this failure must add `25`. Nothing else moved: a trigger missing from the DOM is still
+  exit 23. See the `### Fixed` entry below for why
+  ([#799](https://github.com/ffroliva/gflow-cli/issues/799)).
+
 ### Fixed
 
 - **Migrated-host i2v: the Frames picker is confirmed when it does not commit on the
@@ -61,6 +70,27 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   `credits` itself is still unavailable on migrated accounts; the balance surface for that
   cohort has not been located. #795 stays open for it.
 
+- **The migrated agent-only composer is named before submit instead of reported as
+  selector drift (#799).** Google has put some accounts on a `flow.google.com` composer
+  that has **no classic arm at all** — the prompt box is the agent panel, and aspect,
+  model and count live in Agent settings as defaults rather than per-request controls.
+  gflow waited 30 s for a control that is structurally absent and raised
+  `UiSelectorDriftError` (exit 23), which reads as *our* frontend bug and invites a retry
+  that cannot work.
+
+  It now raises `FlowAgentUiError` — **exit 25, `retryable: false`** — naming the cohort
+  and saying plainly that no flag or profile change helps. The discriminator is the chip:
+  the DOM is identical to #749's recoverable agent mode (settings trigger present under a
+  bare `hidden`), and what separates them is that a recoverable account still has a
+  `button.agent-mode-chip` to turn off while this one has none. A trigger that has left
+  the DOM entirely is unchanged — that is a renamed selector, our bug, and still drift.
+
+  `retryable` moved from `FlowAppError` to the error base to make that possible — its own
+  comment set the condition, *"move it up if, and only if, a second class needs it"*, and
+  this is the second. No existing error's retryability changed.
+
+  Reported with the DOM evidence that made it diagnosable by **@Cstanish127**. gflow-cli
+  still has no driver for that composer; #799 stays open for it.
 ## [0.73.2] — 2026-09-12
 
 ### Fixed

KNOWN_ISSUES.md

@@ -1247,6 +1247,40 @@ continue as a seeded I2V generation — tracked as backlog.
 
 ## Mitigated
 
+### Migrated accounts can have an agent-only composer, which gflow cannot drive
+
+Some accounts on `flow.google.com` are served a composer with **no classic arm at all**:
+the only prompt box is the agent panel, there is no `agent-mode-chip` to turn off, and
+aspect / model / count are Agent-settings **defaults** rather than per-request controls.
+Every control the migrated driver reaches for is structurally absent, so `gflow video`
+and `gflow image` cannot run there at all.
+
+**Mitigation (v0.74.0):** the state is named before submit — `FlowAgentUiError`, exit 25,
+`retryable: false`, $0 spent — instead of the generic `UiSelectorDriftError` (exit 23)
+that reads as a gflow frontend bug and invites a doomed retry. The discriminator is the
+chip, because the DOM is otherwise identical to the *recoverable* agent mode of
+[#749](https://github.com/ffroliva/gflow-cli/issues/749): a hidden settings trigger **with**
+a chip is recoverable and gflow turns it back itself; a hidden trigger with **no chip
+anywhere** is this cohort.
+
+**No workaround inside gflow.** No flag, `--ui-mode`, or profile change reaches it — the
+composer is a property of the Google account. Generating from the Flow web UI still works.
+A driver for the agent panel is not implemented;
+[#799](https://github.com/ffroliva/gflow-cli/issues/799) stays open for it.
+
+**Prevalence is unmeasured.** One reporter, one account (Windows 11, ru locale), whose DOM
+capture is what made this diagnosable. No account available to the maintainers is in this
+cohort, so the **positive** case is verified against the reporter's captured markup driven
+by a real Chromium (`tests/api/transports/test_agent_only_composer.py`) and **not** against
+live Flow — that needs an account in the cohort and is the named blocker on #799.
+
+The **negative** controls were measured live at $0 on 2026-09-13, which is what makes the
+discriminator more than a guess: a healthy migrated composer carries
+`button.agent-mode-chip` **present and un-pressed** with its settings trigger visible. So a
+migrated account normally has a chip, and having none is the anomaly this keys on.
+
+---
+
 ### Flow can pin the agentic cohort server-side for hours
 
 The classic↔agentic editor arm is server-assigned per page load and normally

docs/MCP.md

@@ -141,15 +141,19 @@ marks a transient failure a scheduler can re-run without operator intervention 
 WAF/reCAPTCHA bounce (`WafRejectionError`), rate-limit (`RateLimitError`),
 transport timeout (`TransportTimeoutError`), network blip (`NetworkError`), a
 dropped browser session (`BrowserSessionClosedError`), a Flow web-app crash
-(`FlowAppError`), an agentic-cohort flap (`FlowAgentUiError`), an unreachable
+(`FlowAppError`), an agentic-cohort flap (`FlowAgentUiError` — but see the
+agent-only exception in § 6.1, which overrides it to `false`), an unreachable
 UI arm (`UiModeUnavailableError`), and a partially-completed sync
 (`SyncPartialError`). That list is `errors.RETRYABLE_ERRORS`, but it is no longer
 the whole answer: `errors.is_retryable` consults the **instance** first, so a raise
-site can override its class. One does today — Flow's `/about` redirect raises
+site can override its class. Two do today — Flow's `/about` redirect raises
 `FlowAppError` with `retryable: false`, and since 2026-09-11 that is a **measurement**, not a
 preserved default: caught during a live occurrence, 5/5 consecutive attempts over ~3 minutes
 landed on `/about` again, on the account's own project with a healthy session — so a retry is
 doomed and costs ~35 s each ([#756](https://github.com/ffroliva/gflow-cli/issues/756)). The
+second is the migrated **agent-only composer**, which raises `FlowAgentUiError` with
+`retryable: false` because which composer an account gets is server-assigned and does not
+flap ([#799](https://github.com/ffroliva/gflow-cli/issues/799)). The
 *cause*, and whether it ever clears, remain unmeasured.
 Read the flag off the envelope; never re-derive it from the class list.
 Everything
@@ -350,8 +354,9 @@ gflow credits user         # does the Bearer path still work for this account?
 | verified | fails, `"credits endpoint returned 401"` | migrated; a token exists but aisandbox-pa rejects it | Same outcome, different raise site. Since v0.74.0 this one names the cause too ([#795](https://github.com/ffroliva/gflow-cli/issues/795)); through v0.73.2 it read `"aisandbox-pa returned 401 after token refresh"`. Measured on a migrated account, 2026-09-13 |
 | "Signed in to Google, but not to the Flow app" **forever** | — | migrated; labs no longer mints a Flow session at all | Login cannot complete on the released build. Tracked in [#791](https://github.com/ffroliva/gflow-cli/issues/791) |
 
-In every migrated row, **generation over the migrated composer still works** — only the
-aisandbox REST reads fail.
+In every migrated row above, **generation over the migrated composer still works** — only
+the aisandbox REST reads fail. The one exception is the agent-only composer described
+below, where there is no composer for gflow to drive at all.
 
 > **A `credits` failure does not mean your cookies are stale.** Two raise sites produce it
 > — labs answering with no token, and aisandbox-pa rejecting the token labs did issue —
@@ -362,11 +367,19 @@ aisandbox REST reads fail.
 > believe `auth status`.** Through v0.73.2 the second site carried the class-default
 > SAPISID advice instead — see the CHANGELOG for why.
 
-The migrated composer itself also comes in more than one shape. If `gflow_generate_video`
-fails pre-submit with a selector-drift envelope naming `.settings-trigger-button` as
-present-but-hidden, and no `agent-mode-chip` exists on the page, the account is on the
-**agent-only composer**, which has no classic composer at all — there is nothing to drive
-yet ([#799](https://github.com/ffroliva/gflow-cli/issues/799)).
+The migrated composer itself also comes in more than one shape. Since v0.74.0 an account
+whose composer is **agent-only** — no classic composer at all — is named as such before
+submit: `FlowAgentUiError`, **exit-25-equivalent**, `retryable: false`, saying the settings trigger is
+present but hidden and that no `agent-mode-chip` exists to turn off. Through v0.73.2 that
+same account got a generic selector-drift envelope (exit 23), which reads as *our* bug and
+invites a retry that cannot work ([#799](https://github.com/ffroliva/gflow-cli/issues/799)).
+
+The distinction that matters, because the DOM is identical either way: a **pressed chip**
+means the classic arm exists and gflow turns it back on itself
+([#749](https://github.com/ffroliva/gflow-cli/issues/749)); **no chip at all** means there
+is no classic arm to return to, aspect / model / count are Agent-settings defaults rather
+than per-request controls, and no flag or profile change helps. Generating from the Flow
+web UI still works.
 
 ### 6.2 Envelope → cause → what to do
 

docs/USAGE.md

@@ -1789,7 +1789,7 @@ shell scripts can branch on the failure mode without parsing stderr.
 | `22` | `UpscaleUnavailableError` | 4K upscale is gated to Flow **Ultra** accounts (HTTP 403) | Use `--scale 2k`, or upgrade the Flow plan                |
 | `23` | `UiSelectorDriftError` | A Flow editor control could not be located — Google changed the frontend (issues #183, #493) — **or** a blocking announcement overlay survived dismissal, so no control below it can be clicked (probe `overlay_close_button`, #593). On the migrated host a missing submit control is checked against the wallet first and reported as **37** when Flow swapped in its insufficient-credits warning, so a short balance no longer arrives here | Update gflow-cli; file a bug with the probe name + the diagnostics JSON / debug screenshot referenced in the error message, plus the incident bundle's `report.md` when one was written. For `overlay_close_button`: open the project once in Chrome and dismiss the announcement — the dismissal persists on your account |
 | `24` | `BrowserEngineUnavailableError` | `GFLOW_CLI_BROWSER_ENGINE=patchright` but the engine is not installed | `pip install 'gflow-cli[patchright]'`, or unset `GFLOW_CLI_BROWSER_ENGINE` |
-| `25` | `FlowAgentUiError`    | The profile is on Flow's Agentic UI cohort and the classic media panel is unrecoverable for this operation | Rare since v0.38.0 (#332): the mode controller reliably recovers agentic→classic, so first retry with `--ui-mode classic`; if it persists, see KNOWN_ISSUES on the agentic cohort |
+| `25` | `FlowAgentUiError`    | The profile is on Flow's Agentic UI cohort and the classic media panel is unrecoverable for this operation — **or**, on `flow.google.com`, the account's composer is agent-only and there is no classic arm at all ([#799](https://github.com/ffroliva/gflow-cli/issues/799)) | On labs this is rare since v0.38.0 (#332): the mode controller reliably recovers agentic→classic, so first retry with `--ui-mode classic`. The migrated agent-only cohort reports `retryable: false` — no flag or profile change helps; use the Flow web UI. See KNOWN_ISSUES |
 | `26` | `MediaAttributionError` | Generated media could not be reliably attributed to this request (issue #281) | Re-run; a dedicated project with fewer pre-existing assets avoids the ambiguity |
 | `27` | `MediaUploadRejectedError` | Flow's upload endpoint refused the input file (`uploadImage` 4xx, issue #287) | Re-encode the image (`ffmpeg -q:v 2 -map_metadata -1`), or reference the asset by its media UUID |
 | `28` | `UiModeUnavailableError` | The Flow UI arm this command required (`--ui-mode`/`GFLOW_CLI_UI_MODE`; `-i` forces agentic for images; **video always requires classic** — no agentic video driver exists) couldn't be reached after a switch attempt; aborted before submitting — no credits spent (issue #299) | Retry (the cohort flaps per load); try another `--profile`; for images you can also relax `GFLOW_CLI_UI_MODE` — for video there is nothing to relax |

skills/gflow-cli/SKILL.md

@@ -280,6 +280,7 @@ Documented errors agents commonly make — negative examples for the SkillOpt tr
 | Python: `client = FlowApiClient(...)` then method calls | Must use `async with FlowApiClient(...) as client:` — it's an async context manager |
 | Python: `from gflow_cli import FlowApiClient` | `from gflow_cli.api.client import FlowApiClient` |
 | `gflow video t2v`/`i2v`/`r2v` without `--project` on an account Google moved to `flow.google.com` (exit 11), or an unported UUID/entity reference or model (exit 36) | Pass `--project <id>` — migrated hosts support video t2v, local-file i2v/r2v, and image t2i/i2i (local refs) only; exit 36 is non-retryable, `GFLOW_CLI_FLOW_HOST=labs.google` is the kill switch (see USAGE § gflow video t2v / i2v / r2v and image sections) |
+| Telling a user on `flow.google.com` whose run exits **25** to retry, switch profile or pass `--ui-mode classic` | Their account's composer is **agent-only** — no classic arm exists, so aspect/model/count are Agent-settings defaults and gflow has no driver for it (`retryable: false`, $0, pre-submit). Nothing in gflow reaches it; the Flow web UI still works ([#799](https://github.com/ffroliva/gflow-cli/issues/799)). Exit 25 on `labs.google` IS the retryable A/B cohort — check the host before advising |
 | Suggesting a native `batch` subcommand under `gflow video` | It doesn't exist — that stub never worked and was removed. Loop `gflow video t2v`/`i2v` from the shell for multi-clip runs (`gflow image batch manifest.tsv\|json` is the real, working batch command, but it's image-only) |
 
 ## Disclaimer

src/gflow_cli/api/transports/migrated_composer.py

@@ -64,6 +64,7 @@
 )
 from gflow_cli.errors import (
     ConfigurationError,
+    FlowAgentUiError,
     FlowHostMigratedError,
     InsufficientCreditsError,
     MediaUploadRejectedError,
@@ -114,6 +115,11 @@
 #: Accounts, measurements and the transition inventory:
 #: docs/superpowers/spikes/2026-09-08-migrated-composer-agent-mode-hides-settings.md
 AGENT_MODE_CHIP = "button.agent-mode-chip[aria-pressed='true']"
+#: The same chip in ANY state. :data:`AGENT_MODE_CHIP` matches only a PRESSED chip, so it
+#: cannot tell "this account has no agent/classic split" from "the split exists and is
+#: currently off" — the distinction #799 turns on. Never click this one: it is not
+#: self-guarding, and a click would toggle a healthy composer INTO agent mode.
+AGENT_MODE_CHIP_ANY = "button.agent-mode-chip"
 #: How long the classic composer gets to come back after the chip is clicked. The swap is
 #: a local Angular re-render, not a navigation — measured well under a second on both
 #: accounts — so this is headroom, not an expectation.
@@ -726,6 +732,39 @@ async def ensure_editor(self, page: Page, project_id: str, *, timeout_s: float =
             # healthy run nothing — no extra query is issued unless the gate has failed.
             state, click_error = await self._exit_agent_mode(page)
             if state == "absent":
+                # #799. Before calling this drift, separate the third cohort from our own
+                # bug, because the DOM they leave is the same one: the trigger present
+                # under a bare `hidden`. With a pressed chip that is agent mode and the
+                # branch above recovers it. With NO chip in the page at all, there is no
+                # classic arm to return to — the account's only composer is the agent
+                # panel, so model/aspect/count live in Agent settings as defaults and
+                # nothing gflow drives is on the page. A trigger that has LEFT the DOM is
+                # the opposite finding, a renamed selector, and stays drift.
+                if await self._is_agent_only_composer(page, trigger):
+                    log.info("migrated.agent_only_composer", issue_ref="#799", url=page.url)
+                    raise FlowAgentUiError(
+                        detail=(
+                            f"migrated host: this account's composer is agent-only — the "
+                            f"settings trigger ({READY_ANCHOR}) is in the page but hidden, "
+                            f"and no agent-mode chip ({AGENT_MODE_CHIP_ANY}) exists to turn "
+                            f"off, so there is no classic composer to drive on {page.url} "
+                            f"(host=migrated). Not selector drift, and not the recoverable "
+                            f"agent mode of #749. Readiness gate: {e}"
+                        ),
+                        remediation_hint=(
+                            "Google has put this account on Flow's agent-only composer, "
+                            "where aspect, model and count are Agent-settings defaults "
+                            "rather than per-request controls. gflow-cli has no driver for "
+                            "it yet, so no flag or profile change helps and a re-run will "
+                            "not either — this is tracked in issue #799. Generating from "
+                            "the Flow web UI still works."
+                        ),
+                        # Not retryable: #749's chip flips per click, but which composer an
+                        # account gets is server-assigned per account, so a retry lands the
+                        # same page. Same reasoning that keeps FlowHostMigratedError out of
+                        # RETRYABLE_ERRORS; FlowAgentUiError is in it for the labs A/B.
+                        retryable=False,
+                    ) from e
                 raise UiSelectorDriftError(
                     detail=(
                         f"migrated host: the settings trigger ({READY_ANCHOR}) did not "
@@ -813,6 +852,34 @@ async def _agent_chip_pressed(page: Page) -> bool | None:
             log.warning("migrated.agent_mode_probe_failed", error=str(e)[:200])
             return None
 
+    @staticmethod
+    async def _is_agent_only_composer(page: Page, trigger: Any) -> bool:
+        """True when the page is #799's agent-only cohort rather than selector drift.
+
+        Two facts, and it takes both. The readiness anchor is **in the DOM but not
+        visible** — a trigger that is simply gone is a renamed selector, our bug, and
+        must keep saying so. And **no agent-mode chip exists at all**, pressed or not:
+        the chip is what makes a hidden trigger recoverable (#749), so its absence is
+        what says this account has no classic arm to go back to.
+
+        An un-pressed chip is deliberately excluded from the claim, and measuring it is
+        what makes the rest of this worth doing: on a live healthy migrated composer
+        (ffroliva, 2026-09-13, $0) the chip IS present and un-pressed while the trigger
+        is visible — so a normal account on this host has a chip, and having none is the
+        anomaly. An un-pressed chip beside a HIDDEN trigger is the state never observed;
+        it contradicts the measured mechanism (pressed ⇒ hidden) and falls through to
+        drift rather than being asserted as a cohort.
+
+        Fail-closed: an unreadable page is not evidence of a cohort.
+        """
+        try:
+            if not await trigger.count() or await trigger.is_visible():
+                return False
+            return not await page.locator(AGENT_MODE_CHIP_ANY).first.count()
+        except Exception as exc:  # noqa: BLE001 - an unreadable page is not an answer
+            log.warning("migrated.agent_only_probe_failed", error=str(exc)[:200])
+            return False
+
     @classmethod
     async def _exit_agent_mode(cls, page: Page) -> tuple[AgentModeExit, Exception | None]:
         """Turn Flow's agent mode off if it is on.