gflow-cli 0.80.0__tar.gz → 0.82.0__tar.gz

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "gflow",
-  "version": "0.80.0",
+  "version": "0.82.0",
   "description": "Reusable development and operations workflows for gflow-cli",
   "author": {
     "name": "Flavio Oliva",

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/.env.template

@@ -165,6 +165,14 @@
 # GFLOW_CLI_HEADLESS=false explicitly rather than relying on the default.
 # GFLOW_CLI_HEADLESS=false
 
+# Screen position X,Y of the headed generation browser. Default -30000,-30000 opens it
+# off-screen: still real headed Chrome (reCAPTCHA unaffected), just not over your
+# desktop. Set 0,0 (or a second monitor's origin) to watch or debug a run; set empty
+# for Chrome's own placement. It does NOT stop Chrome taking keyboard focus at launch.
+# Login windows are always visible: for an account chooser / consent screen, run
+# `gflow auth login --profile <name>` rather than making this window visible.
+# GFLOW_CLI_BROWSER_WINDOW_POSITION=-30000,-30000
+
 # Browser automation engine — playwright (default) | patchright
 # patchright is an OPT-IN, drop-in patched Playwright (Chromium) that avoids the
 # Runtime.enable CDP leak for stronger reCAPTCHA-Enterprise evasion on the HEADED

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/AGENTS.md

@@ -47,7 +47,7 @@ by construction, so Codex / Cursor / Aider / `agy` read exactly what Claude Code
 - Python CLI and MCP server for [Google Flow](https://labs.google/fx/tools/flow) — drives Veo (image-to-video, text-to-video) and Imagen (text-to-image) generations from the terminal by reverse-engineering Flow's private REST API at `aisandbox-pa.googleapis.com` — and, for accounts Google has moved to `flow.google.com`, that frontend's `batchexecute` wire (text-to-video, image-to-video from a local `--initial-frame`, reference-to-video from local `--ref` files, and text-to-image/local-file image-to-image, today; `GFLOW_CLI_FLOW_HOST`).
 - Python 3.11+ · `uv`-managed · `hatchling` builds · Playwright Chromium transport · `pyright` strict · `ruff` · `pytest`.
 - Single-package modular monolith. Top-level modules under `src/gflow_cli/`: `api/`, `auth/`, `data/`, `mcp/`, `services/`, `tools/`, `ui/`, `worker/`, `browser_manager.py`, `cli.py`, `_cli_helpers.py`, `diagnostics.py`, `json_output.py`, `media.py`, `profile_lease.py`, `redaction.py`, `storage.py`, `winsec.py`, `cli_project.py`, `cli_character.py`, `cli_credits.py`, `cli_data.py`, `cli_image.py`, `cli_instructions.py`, `cli_models.py`, `cli_movie.py`, `cli_run.py`, `cli_scene.py`, `cli_tools.py`, `cli_video.py`, `chain.py`, `chain_manifest.py`, `cli_doctor.py`, `cli_docs.py`, `docs_catalog.py`, `cli_update.py`, `composition.py`, `config.py`, `errors.py`, `file_integrity.py`, `flow_selectors/`, `update_check.py`, `exceptions.py`, `image_batch.py`, `movie_manifest.py`, `observability.py`, `paths.py`, `profile_store.py`.
-- Command surface: `gflow auth`, `gflow credits` (user/list — read-only Veo balance), `gflow image` (t2i/i2i/batch/upload/upscale), `gflow video` (t2v/i2v/r2v/chain/extend — `extend` continues an existing clip past Flow's 8s ceiling, server-seeded from the source so the join is continuous; no `batch` subcommand; the nonfunctional stub was removed, loop `gflow video t2v`/`i2v` from the shell for multi-clip runs), `gflow character` (create/list/show/rm/voices — reusable project-scoped Flow Character entities), `gflow scene` (create/show — Add Clip / Scenes, with `create --output` for credit-free server-side extended video), `gflow instructions` (persistent Agent-Mode brief cards — add/list/enable/disable/rm/apply/toggle-mode, credits-free, `--project` required), `gflow movie` (run/template — multi-scene manifest pipeline), `gflow tools` (list/show/run — prompt-rewriting tools, also `--tool` on generation commands), `gflow data` (catalog queries), `gflow docs` (browse the bundled documentation — bare lists every topic, `gflow docs <topic>` prints one as Markdown, `gflow docs --search TERM` searches all of them; there are **no** `list`/`print`/`search` subcommands — read-only, offline, deliberately no MCP twin), `gflow doctor` (read-only pre-flight diagnostic, exit 33 = findings present), `gflow update` (self-update through the installer that put it here — uv tool / pipx / pip; `--check` only reports; source installs refused, exit 11; deliberately no MCP twin), `gflow project`, `gflow models`, `gflow run`, `gflow mcp` (run/setup — stdio MCP server), and `gflow serve` (Streamable HTTP at `/mcp`; `--transport sse` is deprecated).
+- Command surface: `gflow auth`, `gflow credits` (user/list — read-only Veo balance), `gflow image` (t2i/i2i/batch/upload/upscale), `gflow video` (t2v/i2v/r2v/chain/extend — `extend` continues an existing clip past Flow's 8s ceiling, server-seeded from the source so the join is continuous; no `batch` subcommand; the nonfunctional stub was removed, loop `gflow video t2v`/`i2v` from the shell for multi-clip runs), `gflow character` (create/list/show/rm/voices — reusable project-scoped Flow Character entities), `gflow scene` (create/show — Add Clip / Scenes, with `create --output` for credit-free server-side extended video), `gflow instructions` (persistent Agent-Mode brief cards — add/list/enable/disable/rm/apply/toggle-mode, credits-free, `--project` required), `gflow movie` (run/template — multi-scene manifest pipeline), `gflow tools` (list/show/run — prompt-rewriting tools, also `--tool` on generation commands), `gflow data` (catalog queries), `gflow docs` (browse the bundled documentation — bare lists every topic, `gflow docs <topic>` prints one as Markdown, `gflow docs --search TERM` searches all of them; there are **no** `list`/`print`/`search` subcommands — read-only, offline, deliberately no MCP twin), `gflow doctor` (read-only pre-flight diagnostic, exit 33 = findings present), `gflow update` (self-update through the installer that put it here — uv tool / pipx / pip; `--check` only reports; source installs refused, exit 11; deliberately no MCP twin), `gflow project`, `gflow models`, `gflow run` (JSON image-batch config; a row may generate from an earlier row with `"ref": "batch:N"`, referenced in place, no upload, or from a local image file, uploaded once per run), `gflow mcp` (run/setup — stdio MCP server), and `gflow serve` (Streamable HTTP at `/mcp`; `--transport sse` is deprecated).
 - Works with any Google account that has Flow access. All generations bill against the user's own Google account.
 
 ## Headed-browser dependency (architectural reality)

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/CHANGELOG.md

@@ -7,6 +7,80 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.82.0] — 2026-10-01
+
+### Added
+
+- **A `gflow run --config` row can reference a local image file (#913).** Set `"ref"` to a
+  path, resolved against the config file's folder. The file is checked before the browser
+  starts (it must exist and be a real image). It is uploaded once per run, on first use;
+  every row naming the same file references that upload in place, so the project holds
+  one copy. See [docs/USAGE.md](docs/USAGE.md#referencing-a-local-file).
+- **A `gflow run --config` row can generate from an earlier row's image (#913).** Set
+  `"ref": "batch:N"`. The earlier image is already in the run's Flow project, so it is
+  referenced in place by the handle Flow returned. Nothing is re-uploaded, and the
+  project holds no duplicate. Rows run in file order, deferred only until
+  their parent has run. A failed parent skips its dependents with the reason instead of
+  submitting them without the reference. A referenced row must make one image. On
+  flow.google.com the image is found in the composer's `@` picker and chosen by its
+  thumbnail identity, not its caption (captions repeat). The submit is aborted before
+  Flow acts if it does not carry the reference. `gflow image batch` refuses references and
+  points at `gflow run`. See [docs/USAGE.md](docs/USAGE.md#referencing-an-earlier-row).
+- **`gflow run --config` and multi-prompt `gflow image t2i` now record their successful
+  generations in the local catalog.** They recorded failures only. A referencing row is
+  recorded as image-to-image with its parent as the input.
+
+### Fixed
+
+- **Manifest references are refused instead of silently ignored (#913).** A
+  `gflow run --config` or `gflow image batch` row with `ref` or `reference_entity` was
+  accepted, then ran as plain text-to-image with exit 0. v0.52.0 announced the fields
+  (#317), but nothing ever applied them. Both commands now refuse such a row before any
+  browser work, naming the row and field: exit 11 for `run`, a usage error (exit 2) for
+  `image batch`. The exceptions are `"ref": "batch:N"` and a local image file in
+  `gflow run --config`, which now work (see Added).
+
+- **The docs no longer suggest clearing a sign-in screen in the generation window
+  (#925).** 0.81.0's docs said to make the off-screen generation browser visible and
+  clear an account chooser or consent screen in it. A generation run gives up about 30 s
+  after it clicks the chooser row, so that was a race. It also contradicted the error
+  itself, which says to run `gflow auth login --profile <name>`. The login window is
+  always visible and waits for you, so that is the remedy the docs now give.
+  `GFLOW_CLI_BROWSER_WINDOW_POSITION` remains for watching or debugging a run.
+
+## [0.81.0] — 2026-09-30
+
+### Added
+
+- **The generation browser opens off-screen (#923).** Headed Chrome is required, so until
+  now every `image`/`video` run put a Chrome window over your desktop. It now opens at
+  `-30000,-30000`. It is still fully headed, so Flow and reCAPTCHA see the same browser.
+  Image and video generation were measured unchanged, and the page is not throttled.
+  `GFLOW_CLI_BROWSER_WINDOW_POSITION` takes any `X,Y` to place it where you want (to watch
+  a run or clear a consent screen by hand); empty restores Chrome's own placement. It
+  does not stop Chrome taking keyboard focus at launch. Login windows are unaffected. See
+  [docs/CONFIGURATION.md](docs/CONFIGURATION.md#gflow_cli_browser_window_position).
+  Thanks to @johngbl for the idea and the first implementation.
+
+### Security
+
+- **Locked `urllib3` 2.8.0 (CVE-2026-97687, CVE-2026-97688, CVE-2026-97689).**
+  Transitive only — via `requests` and, with the `s3` extra, `botocore`. No gflow
+  code imports it directly.
+
+### Fixed
+
+- **Windows profile write-denial is distinguished from browser contention.** When
+  the generation client's browser launch fails with a Chrome `ProcessSingleton`
+  error carrying Windows access-denied code 5, it now raises `ProfileAccessError`
+  (configuration exit code 11) instead of a misleading "another Chrome holds it"
+  `ProfileLockedError` or an unclassified error. The match is on the error code,
+  so it works in every Windows display language. The remediation points to
+  profile-directory write access, not to closing a competing Chrome. Not covered
+  yet: `gflow auth login` / auth verification, and the standalone and experimental
+  transport launches — those still surface Chrome's raw error.
+  Thanks @L1meSn0w for the diagnosis and the Windows reproduction (#919).
+
 ## [0.80.0] — 2026-09-29
 
 ### Security
@@ -2984,6 +3058,8 @@ completed exit 0, proving no regression. See
 ### Added
 
 - **Intra-batch reference support for image batches (#317).** `BatchPromptItem` and `gflow image batch` now support `ref` and `reference_entity` fields (e.g. `ref="batch:0"`), with topological dependency sorting and circular dependency validation.
+  **Correction (#913):** this never worked. The fields were parsed and silently ignored,
+  and the sorting was never called; see the `[Unreleased]` Fixed entry for #913.
 - **Character entity provenance recording and video CLI flag parity (#402).** Added `--reference-entity` and `--reference-entity-name` CLI options to `gflow video` commands (`t2v`, `i2v`, `r2v`) and verified character provenance recording in `operations.metadata_json`.
 
 ### Fixed
@@ -5708,7 +5784,9 @@ shell-script template that branches on these codes.
 
 First skeleton. Not functional end-to-end yet.
 
-[Unreleased]: https://github.com/ffroliva/gflow-cli/compare/v0.80.0...HEAD
+[Unreleased]: https://github.com/ffroliva/gflow-cli/compare/v0.82.0...HEAD
+[0.82.0]: https://github.com/ffroliva/gflow-cli/compare/v0.81.0...v0.82.0
+[0.81.0]: https://github.com/ffroliva/gflow-cli/compare/v0.80.0...v0.81.0
 [0.80.0]: https://github.com/ffroliva/gflow-cli/compare/v0.79.1...v0.80.0
 [0.79.1]: https://github.com/ffroliva/gflow-cli/compare/v0.79.0...v0.79.1
 [0.79.0]: https://github.com/ffroliva/gflow-cli/compare/v0.78.0...v0.79.0

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/KNOWN_ISSUES.md

@@ -390,6 +390,17 @@ listing endpoint (privacy-gated to `store` history mode), and `gflow doctor`
 (#542) surfaces the affected-row count. Freshly generated rows whose caption
 has not landed yet stay nameless until the next sync sweep.
 
+### `gflow run --config` references cannot resume (#913)
+
+- **Status:** Open · **Severity:** Low (a re-run regenerates; nothing is lost) · **Affected:** every `gflow run --config` with `"ref"` rows (`batch:N` or a local file)
+- **Tracked:** [#913](https://github.com/ffroliva/gflow-cli/issues/913)
+
+A run with `"ref": "batch:N"` rows has no resume: a re-run creates a new project and
+regenerates every row, parents included, and one early failure skips the whole chain under
+it. The handle that makes in-place referencing work (Flow's reply for the parent) exists
+only within the run that generated it. A local-file `ref` is uploaded again into the
+re-run's new project for the same reason.
+
 ### Video duration control is absent on some account cohorts
 
 - **Status:** Mitigated — fail-fast shipped in
@@ -1451,7 +1462,10 @@ failed assert left the profile unusable — it is usable, just possibly as the w
 
 Two further items on the same issue are unfixed and worth knowing about: a second chooser
 hop (chooser → consent → chooser) is not handled and degrades into the landing timeout, and
-that timeout is still an unmeasured number.
+that timeout is still an unmeasured number. To clear such a screen by hand, run
+`gflow auth login --profile <name>`: its window is always visible and waits for you. The
+generation browser opens off-screen since v0.81.0, and making it visible is not a remedy
+here, because the run gives up within 30 s (see [CONFIGURATION](docs/CONFIGURATION.md#gflow_cli_browser_window_position)).
 
 ### Auth verification depends on Google's NextAuth session endpoint
 
@@ -1485,6 +1499,8 @@ key — surfaces as a `RuntimeError` that `auth/cookies.py` normalizes to
 
 Chromium refuses to open two persistent contexts on the same `user-data-dir` simultaneously. Historically this surfaced as an unhelpful Chromium "ProcessSingleton: profile is locked" error partway through a run. As of the profile-lease hardening (production-readiness plan, slice D1/D3), gflow-cli enforces this itself: a cross-process advisory lock (`ProfileLease`, kernel `flock` on POSIX / `msvcrt.locking` on Windows) guards every profile directory. A second `gflow` invocation, `gflow serve` daemon task, or MCP call against an already-leased profile is rejected **immediately** by default — before any Chrome process starts — with a typed `ProfileLockedError` (**exit code 11**); it never silently corrupts the profile. Since #478, setting [`GFLOW_CLI_LEASE_WAIT_SECONDS`](docs/CONFIGURATION.md#gflow_cli_lease_wait_seconds) opts a waiter into a bounded wait that takes over as soon as the current holder finishes (holders always run to completion and are never asked to release early; same-process contention still fails fast — waiting on yourself would deadlock). Since #864 an MCP server waits 180 s by default, and queues its own calls on one profile rather than rejecting the second.
 
+**Not every `ProcessSingleton` failure is contention.** On Windows, a `ProcessSingleton` error carrying access-denied code `(0x5)` means Chrome cannot write the profile directory; at the generation client's launch gflow reports it as `ProfileAccessError` — see [CONFIGURATION § Profile-directory permissions at browser launch](docs/CONFIGURATION.md#profile-directory-permissions-at-browser-launch).
+
 **Workaround:** use different profiles for parallel work — different profiles acquire independent leases and run fully concurrently.
 
 ### gflow refuses to open my profile after a downgrade ("written by a newer Chromium", exit 11)

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: gflow-cli
-Version: 0.80.0
+Version: 0.82.0
 Summary: CLI and MCP server for Google Flow — drive Veo text-to-video, image-to-video and Imagen image generation from the terminal.
 Project-URL: Homepage, https://github.com/ffroliva/gflow-cli
 Project-URL: Documentation, https://ffroliva.github.io/gflow-cli/

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/docker/Dockerfile

@@ -45,7 +45,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 # Deliberately NOT pinned: google-chrome-stable. Chrome's apt repo keeps only the current
 # build, so a pinned version stops resolving within weeks. The trade-off is that this
 # layer caches indefinitely and a rebuild is what picks up a newer Chrome.
-ARG GFLOW_VERSION=0.80.0
+ARG GFLOW_VERSION=0.82.0
 RUN pip install --no-cache-dir "gflow-cli==${GFLOW_VERSION}"
 
 VOLUME ["/data"]

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/docs/CONFIGURATION.md

@@ -31,6 +31,22 @@ See [AUTHENTICATION § Session storage](AUTHENTICATION.md#session-storage) for t
 create it — a `tools/` directory of user-authored "My Tools" TOMLs
 (`<GFLOW_CLI_HOME>/tools/*.toml`, auto-loaded; see [TOOLS.md § My Tools](TOOLS.md)).
 
+#### Profile-directory permissions at browser launch
+
+On Windows, when a generation run's browser launch fails with a Chrome
+`ProcessSingleton` error carrying access-denied code 5, gflow reports
+`ProfileAccessError`, a configuration error (exit code 11), rather than suggesting
+contention. Ensure the process can write the complete persistent profile directory,
+or use a profile inside a writable `GFLOW_CLI_HOME`. Relocating application-level
+lease locks alone does not relocate Chrome's `ProcessSingleton`, cookies, or crash
+reports.
+
+This classification covers the generation client's launch only. `gflow auth login`,
+auth verification, and the standalone and experimental transports still
+surface Chrome's raw error for the same cause; look for `ProcessSingleton` plus
+`(0x5)` in it. Launch failures without the access-denied code keep their existing
+classification; this is not general cross-platform permission-error detection.
+
 ### `GFLOW_CLI_PROFILE`
 
 **What:** Default profile name used when `--profile` isn't passed on the CLI.
@@ -333,6 +349,56 @@ GFLOW_CLI_HISTORY_PROMPTS=redacted gflow image t2i "confidential brief"
 **Default:** `false` — **headed real Chrome is the production default**, not an opt-in fallback. The `ui_automation` transport (gflow-cli's only production transport) requires a headed browser: reCAPTCHA Enterprise rejects headless Chromium with an immediate 403, so `headless=true` is not a WAF workaround — it only exists for CI/CD environments running a non-`ui_automation` transport (e.g. `bearer`/`sapisidhash`, experimental).
 **WAF-sensitive runs:** set `GFLOW_CLI_HEADLESS=false` explicitly (it is already the default, but pin it in CI/CD env files or scripts that also set `headless=true` for a different transport, so a transport switch back to `ui_automation` doesn't silently regress to a rejected headless launch).
 
+### `GFLOW_CLI_BROWSER_WINDOW_POSITION`
+
+**What:** Where on screen the headed generation browser opens, as `X,Y` in pixels.
+**Values:** `X,Y` (integers, may be negative) | empty
+**Default:** `-30000,-30000`: off-screen. Generation needs a real **headed** Chrome (see
+[`GFLOW_CLI_HEADLESS`](#gflow_cli_headless)), and without this setting that window
+opens over whatever you are working on during every run. Off-screen it is still fully
+headed, so Flow and reCAPTCHA see exactly the same browser.
+
+| Set it to | Result |
+|---|---|
+| unset | off-screen (default) |
+| `0,0`, `1920,0`, … | the window opens there: watch a run, or park it on a second monitor |
+| empty (`GFLOW_CLI_BROWSER_WINDOW_POSITION=`) | Chrome's own placement, the behaviour before this setting existed |
+
+**When to bring it back on-screen:** to watch a run, debug a step, or record a demo.
+
+**Not for a screen that needs a person.** If a run stops on an account chooser, a Google
+consent or verification screen, the error tells you to run
+`gflow auth login --profile <name>`. Do that: the login window always opens where you
+can see it, and it waits for you. A generation run does not: after clicking the
+chooser row it gives up within 30 seconds, so clearing a screen in a visible
+generation window is a race. (Flow's one-time *rights to use this image* confirmation
+has its own remedy: see [KNOWN_ISSUES](../KNOWN_ISSUES.md).)
+
+**What it does not do:** it does not stop Chrome **taking keyboard focus** when it
+launches: the `--no-focus-on-init` switch we tried made no measurable difference. If you are typing when a run starts, the
+first keystrokes can land in the (invisible) browser. Login (`gflow auth login`) is
+unaffected and always opens where you can see it.
+
+**Measured (Windows 11, 2026-09-30):** off-screen and visible runs behaved the same.
+
+| | off-screen | visible |
+|---|---|---|
+| `gflow image t2i` | image saved, 34 s | image saved, 33 s |
+| `gflow video t2v` (`veo-lite`) | video saved, 53 s, no poll stalls | video saved, 65 s, no poll stalls |
+| page `visibilityState` | `visible` | `visible` |
+| animation frames / 100 ms timer ticks in 5 s | 300 / 50 | 295 / 50 |
+
+Chrome does not throttle an off-screen window: the page still counts as visible.
+Windows clamps very large offsets, so `-30000,-30000` lands at `-21845,-21845`, still
+past any monitor, on every launch measured. A visible position such as `100,100` opened
+exactly there on 8 of 10 launches. In the other two the window was found elsewhere, and
+once it was seen moving mid-run: a visible window can be moved like any other once it
+is open. Not yet measured on macOS or
+Linux, or for runs whose polling lasts longer than about five minutes.
+Measurement script: `scripts/dev/spike_offscreen_window.py`.
+
+**Malformed values** (`100`, `1,2,3`, `a,b`) fail at startup with a validation error.
+
 ### `GFLOW_CLI_BROWSER_ENGINE`
 
 **What:** Selects the browser-automation engine backing the Playwright API.
@@ -557,5 +623,7 @@ gflow image t2i "test idea" --profile experiments
 | `ValidationError: GFLOW_CLI_TIMEOUT_SECONDS must be a positive integer` | Bad `.env` value | Set to a number ≥ 1 |
 | `FileNotFoundError: $GFLOW_CLI_HOME/profile_default not found` | First run, no auth yet | `gflow auth login` |
 | `AuthExpiredError` | Cookies expired or revoked | `gflow auth login --profile <name>` |
+| A run stops on an account chooser, consent or verification screen (`FlowAccountChooserError`, exit 38) | Google wants a person to answer that screen | `gflow auth login --profile <name>` and clear it there; the login window is always visible. Making the generation window visible does not help: the run gives up within 30 s. See [`GFLOW_CLI_BROWSER_WINDOW_POSITION`](#gflow_cli_browser_window_position) |
+| `ProfileAccessError` (exit code 11) | Chrome cannot write the profile directory (Windows access denied) | Grant write access to the whole profile directory, or use a writable profile. See [Profile-directory permissions](#profile-directory-permissions-at-browser-launch) |
 | Output files don't appear where I expect | Flag > env > .env > default — check actual resolved path | `gflow image t2i ... --verbose` shows the resolved output path |
 | `ProfileLockedError` (exit code 11) | Two concurrent calls against the same profile — the cross-process `ProfileLease` fails fast (never waits) on same-profile contention, whether the second holder is another `gflow` process, the `gflow serve` daemon, or an MCP call | Wait for the first call to finish, or use `--profile other` — different profiles run fully in parallel, each with its own lease |

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/docs/INDEX.md

@@ -98,6 +98,7 @@ Slash commands for Claude Code, stored in `.claude/commands/gflow/`. All prefixe
 **"How much will this batch cost me in Veo credits?"** → [USER_GUIDE § Journey 10](USER_GUIDE.md#journey-10--budgeting-credits-before-a-batch-run)
 **"How many Veo credits remain on one or all profiles?"** → [USAGE § `gflow credits`](USAGE.md#gflow-credits)
 **"How do I feed gflow outputs into ffmpeg / a pipeline?"** → [USER_GUIDE § Journey 11](USER_GUIDE.md#journey-11--wiring-gflow-outputs-into-a-downstream-pipeline)
+**"How do I generate an image from an earlier image in the same batch, or from a local file, without re-uploading?"** → [USAGE § Referencing an earlier row](USAGE.md#referencing-an-earlier-row) · [Referencing a local file](USAGE.md#referencing-a-local-file) (`gflow run --config`)
 **"How do I chain clips into one continuous video (initial-frame I2V)?"** → [USAGE § `gflow video chain`](USAGE.md#gflow-video-chain)
 **"How do I stitch existing clips into one extended .mp4 (credit-free, no ffmpeg)?"** → [USAGE § `gflow scene`](USAGE.md#gflow-scene)
 **"How do I make one clip longer than Flow's 8-second cap?"** → [USAGE § `gflow video extend`](USAGE.md#gflow-video-extend) (costs credits per segment; a segment carries ~7s of content though Flow bills 8s — see [KNOWN_ISSUES](../KNOWN_ISSUES.md))
@@ -170,7 +171,7 @@ Slash commands for Claude Code, stored in `.claude/commands/gflow/`. All prefixe
 **"A command failed — where is the incident bundle and what's in it?"** → [DEBUGGING § Automatic incident bundles](DEBUGGING.md#automatic-incident-bundles) (layout, triggers, review-before-sharing); privacy boundaries: [SECURITY § Automatic incident bundles](SECURITY.md#automatic-incident-bundles-gflow_cli_incident_capture-default-on); disable via [CONFIGURATION § GFLOW_CLI_INCIDENT_CAPTURE](CONFIGURATION.md#gflow_cli_incident_capture)
 **"Flow's UI broke a selector — how do I diagnose it?"** → [DEBUGGING § Inspecting Flow's live UI](DEBUGGING.md#inspecting-flows-live-ui)
 **"What does each `ui_automation.*` log event mean?"** → [DEBUGGING § Listener & HTTP-layer debugging](DEBUGGING.md#listener--http-layer-debugging)
-**What was actually live-verified for the latest release?** → latest: [LIVE_VERIFICATION_v0.80.0](LIVE_VERIFICATION_v0.80.0.md) (**2026-09-29, 0 Veo credits.** `--resolution` 360p/720p proven by driving the real settings pane on `omni_flash` and refusing the credit-spending submit, on the CLI transport **and** the MCP twin; `nano2-lite` e2e 2 passed; #891 measured before/after — `gflow run` `[nano2, imagen4]` went from exit 1 `RecaptchaError` with a lost results table to exit 36 naming the model, plus 3 e2e (fresh, warm, MCP) with a spy asserting zero client mints; #896 `e2e_data` 2 passed, its failure path offline-only. Named blockers: rendered resolution (needs a billed clip), the `/about` re-check arm, labs-served accounts.) · previous: [LIVE_VERIFICATION_v0.79.1](LIVE_VERIFICATION_v0.79.1.md) (**the signed-media download survives a dropped connection, verified 2026-09-22 at $0.** All three download sites route through one new shared helper, and that helper was driven through a real Flow clip route and a real signed CDN transfer by `pytest -m e2e_data` — 2 passed in 20.74 s, `ftyp` magic, byte length equal to the size Flow reports, and **no `url=` on any log event**. The retry itself could not wait for an organic reset — 20/20 clean here against the reporter's 2-in-6 — so it was **measured against a real TCP RST with an A/B control**: with `max_retries` the transfer recovers on a second connection, with it disabled the identical fault kills the call, and three attempts at a 5 s timeout finish in 1.57 s rather than ~15 s, proving the backoff shares one budget. **Two sites are NOT live-verified and name their blockers:** the composer call site needs a billed generation and `gflow credits` answers 401 on this account (#795), so no credits were spent; and the labs download cannot be reached at all from here (HTTP 308, 6/6 profiles). Offline: 21 new tests, full sweep 2143 passed.) · earlier: [LIVE_VERIFICATION_v0.79.0](LIVE_VERIFICATION_v0.79.0.md) (**the migrated host stops echoing a model it never observed, verified 2026-09-18 at $0.** `model_name_type` is `null` on `flow.google.com` in a real `image t2i --json`, with every sibling field populated — which is what makes the null a result rather than an empty envelope; the first read of it was flat rather than into `images[]` and returned `None` for everything, and that instrument error was discarded rather than reported. Both surfaces ran: `test_e2e_t2i_runs_on_a_moved_account` and `test_e2e_mcp_i2i_runs_on_the_migrated_host`, 2 passed in 86 s. **`gflow data download` was tested on an image and failed** — 45 s, then a message blaming the project, the trash and your prompt; recovery is video-only and the release notes now say so (#877). **#803 is NOT live-verified and names its blocker:** reaching those raise sites needs a non-JSON or token-less aisandbox reply, and the labs route refuses first with a 404 — itself a misleading error, filed as #875. Offline: 4 new unit tests, 1775 passed.) · previous: [LIVE_VERIFICATION_v0.78.0](LIVE_VERIFICATION_v0.78.0.md) (**generating with no project on `flow.google.com` — and video from a start frame — verified 2026-09-17.** The release exists because labs.google's `project.createProject` is retired — 404 *"Flow RPCs have been deprecated and disabled"* with a labs token, 401 without, 0 successes in 12 across four profiles — and gflow reported the 401 as an expired login. Creation and rename are driven on flow.google.com and read back from Flow's own `jHPbke` / `o8DA4` replies; the CLI and the MCP twin are each exercised separately, and both video paths reach a real submit that is aborted before Flow can bill it. Two failures are recorded and re-tested rather than explained away — one MCP image reply that did not reproduce under a four-run A/B, and a Chrome launch starved by a concurrent coverage run that passed in 36 s alone. The 3:4 image aspect is re-measured on the live radiogroup, five radios where 2026-09-08 saw four. Three items are NOT verified and each names its blocker: no labs-served account, no entitled `/about` account, no non-`en` locale. **#860 is the one thing here that cost credits, deliberately:** the only proof that a start frame binds is a real generation, so ten i2v clips were generated and downloaded on a live migrated account — `migrated.frame_uploaded` → `frame_bound` → `submit_observed` → `download`, the middle event being the one that could not fire before this release. **#859 is recorded as UNOBSERVED, not verified** — its dismissal ran on every one of those runs and no promo modal ever appeared, so the $0 route-intercepted e2e is what covers it. `gflow docs` (#861) is verified from the built wheel, which is its whole contract: 127 topics, `docs/FILE.md:LINE` search hits, no network and no checkout.) · previous: [LIVE_VERIFICATION_v0.77.1](LIVE_VERIFICATION_v0.77.1.md) (**a four-fix patch release, verified 2026-09-17 at ZERO cost.** The headline is #846: the published package could not run a single interactive command on a clean Windows install, because `ConsoleRenderer(colors=True)` needs `colorama` and nothing declared it. Verified as a controlled A/B on a fresh Python 3.13 venv — `SystemError` before, exit 33 after — plus an arm with stderr piped that shows why the AUTO/JSON split hid this from every CI job, and the MCP twin run separately. #848's import probe is exercised against a **deliberately broken venv** and an intact one, including the contract that a probe which cannot run reports nothing rather than reporting breakage. #852 is re-measured on the release head: 40 000 base64-shaped chars fall from 12.07 s to 0.020 ms with identical output on Workspace and custom domains. **#849 is a NAMED BLOCKER, not an omission** — its end-to-end behaviour needs a fresh interactive Google sign-in, which no agent can perform; its live A/B is recorded on PR #851 and its e2e file re-ran 4/4 here. **Seven live e2e failures are recorded rather than hidden**, and controlled: the release changes exactly four source files and none of them is a failing test's module.) · previous: [LIVE_VERIFICATION_v0.77.0](LIVE_VERIFICATION_v0.77.0.md) (**the migrated-host login lockout (#791), verified as a controlled A/B on a real Google **Workspace** account at ZERO cost — same profile, same cookies untouched between arms, so the only variable is the code: exit 8 with the Chrome marker rolled back and `channel_for_profile` collapsed to `None`, then exit 0 with the marker intact and the account recorded. The refusal path is measured too — of 10 local profiles, 5 reproduce #791 and exactly one carries both required cookies: 4 refusals, 1 upgrade, 0 false positives. v0.76.0 recorded this fix as **withdrawn**; the approach that shipped is a different one, server-attested on the landing URL rather than on client-rendered DOM. An earlier revision gated auth on an `@gmail.com`-only pattern that silently declined every custom domain — the account used here is one, which is why this is its verification. **#796 is reproduced and prevented**, not merely reported: it was filed as *'not yet reproduced end-to-end'*. **Two rows are named blockers, not omissions:** the MCP Registry publish (#841) cannot be rehearsed before a real tag push, so this release is its first exercise; and the migrated-host generation e2e cannot run here at all — every account bounces to `/about`, A/B-controlled against `develop` so it is not mistaken for a regression.) · previous: [LIVE_VERIFICATION_v0.76.0](LIVE_VERIFICATION_v0.76.0.md) (**a diagnostics-and-distribution release, verified 2026-09-16 at ZERO cost — no Veo credits, no image generation, no Flow submit.** The decisive arm is #833's **negative** control, run first-hand on two real accounts rather than relayed: `flow-pinhole-unavailable-screen` is **0** on both `ffroliva` and `denon82` while `aisandbox-root` is **1**, so the guard does not fire on accounts that HAVE Flow access and the zero is a real absence rather than a page that never booted — a guard with only a positive arm is indistinguishable from one that is permanently on. Both accounts settled at **`/about`**, which is the state that broke this release's *withdrawn* #791 oracle, and the exit-39 signal reads a component that only exists on the unavailable screen, so `/about` cannot mimic it. mcpservers.org is verified with a **404 negative control** because the listing 403s to `curl` and a naive check would have called the link dead. **#791 is NOT in this release and was withdrawn rather than deferred**: a fix passed eight council dimensions and was then refuted by a five-day-old spike nobody had routed to it — Flow's `/about` hop is decided client-side with zero requests to Flow, and *"the backend grants access while the frontend declines to open it"* (#756), so the probe would have reported authenticated users as logged out, reintroducing #791 as its own cure; the e2e caught it by going red on a live account whose state moved mid-session. **Two rows are blocked by the platform rather than by effort:** the Docker sign-in needs a WSL-integrated Docker host, and the MCP Registry publish cannot be observed until this very tag reaches `main`, since both triggers resolve the workflow from the default branch.) · previous: [LIVE_VERIFICATION_v0.75.0](LIVE_VERIFICATION_v0.75.0.md) (**a distribution-and-hardening release, verified 2026-09-15 on `ffroliva` at ZERO cost — no Veo credits, no image generation, no Flow submit.** The `gflow serve` token fix is an **A/B against the released 0.74.0 wheel from PyPI**, not a diff read: the token was *set* in both arms, so the probe measures whether it is checked — before, `no token 200/200` with 13 tools listed and `wrong 200/200`; after, `401/401` with zero tools and `WWW-Authenticate: Bearer realm="gflow"`, while the correct token still gets 200. Loopback DNS-rebinding was re-confirmed at **421**/**403** because the fix forwards the bind host and getting that wrong would silently disable protection that works today — and an earlier draft of the accompanying advisory claimed that protection was *disabled*, a false claim caught by measuring rather than re-reading. #813's guard is verified against a **real wheel in a real venv installed without extras**, the arm first proving it is in that state (`av present: False  PIL present: False`); the decisive row is that a **nonexistent** manifest still exits **20** on the dependency error rather than file-not-found, which is what shows the guard precedes the manifest read and not merely the generation. The Rich escape fix is verified on the **console render separated from the structlog JSON** — matching the JSON line would have been a false pass, since `--json` was never affected — with a **neutered control** through the same Rich version reproducing the bug exactly: `UNESCAPED -> pip install 'gflow-cli'` against `ESCAPED -> pip install 'gflow-cli[chain]'`. Also: `gflow-cli.exe` exists in the built wheel so `uvx gflow-cli` resolves at all, and the wheel carries 138 entries with **zero** ZIP duplicates. **Three items recorded as NOT verified rather than omitted:** a successful multi-link `chain` run spends Veo credits and stays outstanding e2e debt on #813 — what changed is the refusal path, which *is* verified; `gflow credits` against a working balance is impossible here because the account still returns 401, pre-existing **#795** and unrelated to this release; and the SSE arm of the token fix was exercised on loopback only, a deliberate scope limit, since binding a LAN address to test it would put an authenticated daemon on a live network for the run. **Two rows are blocked by the platform rather than by effort:** the PyPI metadata and the MCP Registry listing are both unobservable until this release is uploaded — PyPI renders the last artefact's metadata, and `mcp-publisher` reads its ownership token out of the *published* README.) · previous: [LIVE_VERIFICATION_v0.74.0](LIVE_VERIFICATION_v0.74.0.md) (**the migrated driver's run-unique uploads and its Frames-picker confirm, plus the `credits` remediation, verified 2026-09-14 on `ffroliva` at ZERO cost — every arm is a read or an upload-and-stop.** The confirm branch is the interesting one: this cohort's picker closes on its own, so the branch would never execute against live Flow by accident, and the e2e drives `FRAME_COMMIT_GRACE_S` to **0.001** so the grace expires before Flow can auto-close and **every** run enters it — the difference between a named blocker and an unrun claim. A *genuinely stuck* picker still cannot be produced here and stays `Refs #792`, verifier the reporter. #795 was verified on the cohort it is for, because this account is now in it — `auth status` green, `credits user` exit **3** with `detail` `credits endpoint returned 401` and `route: credits`, where v0.73.2 printed the class-default *"SAPISID cookie missing… re-run `gflow auth login`"*, advice that is wrong in every word there and that can roll a profile's strategy marker back and start the #791 spiral; the `credits list --json` envelope carries the per-profile `remediation_hint` that used to be dropped. **Two items recorded as NOT verified rather than omitted:** #799's exit-25 raise site needs an account in the agent-only cohort — the maintainer's accounts are served `flow.google.com` but still have the recoverable agent chip, so the state is structurally unreachable on them, a named blocker and #799 stays open; and neither `r2v` nor `i2i` had its own attach path driven to completion — `r2v` was not attempted (its e2e spends Veo credits), while `i2i` **was** attempted twice and failed both times at exit 23, before reaching the code under test, because the `--project` UUID came from `gflow project list` (the aisandbox catalog) and those ids 404 on `flow.google.com` — a harness error, recorded rather than dropped, which left a side finding: gflow reports Flow's own `/404?reason=project` as `UiSelectorDriftError`, blaming our selectors for Flow's 404. **#791 is not fixed and was not attempted blind** — all nine local profiles were probed at $0 and none is in the #791 state (SAPISID + a live `flow.google.com` session + no labs NextAuth token); four print its exact symptom string and none of them is it, because that string is also what a stale profile says.) · previous: [LIVE_VERIFICATION_v0.73.2](LIVE_VERIFICATION_v0.73.2.md) (**the blank incident bundle, reproduced and then cured by a committed e2e test, 2026-09-12 on `ffroliva` at ZERO cost — every arm fails before submit.** The decisive layer is an **A/B run through the same instrument**, not an inspection: on `develop` the new test FAILS in 46.3 s and prints the bundle it got — `tag_counts.div = 0`, `title.length = 0`, `host_category = "other"`, a 4,254-byte screenshot — which is field-for-field the signature [#792](https://github.com/ffroliva/gflow-cli/issues/792) reported from a different account, a different OS and a different command; on the fix it PASSES in 37.8 s with `div` **25**, `title.length` **81**, `host_category` **`flow_app`**, `route` `/404` and a 38.4 KB screenshot. Both arms exit **23** on the same `.settings-trigger-button` drift at `flow.google.com/404?reason=project`, so the arms differ only in what was *captured*, never in what failed. **The reporter's own bundle contained the refutation of their diagnosis**: viewport 1280x720 is `client.py`'s client-owned page, not the 1920x1080 a stray transport tab would have — it was always the driven page, photographed one navigation too late, and the portrait correlation had no mechanism behind it. The existing quality benchmark could not have caught this: it captures deliberately on a **healthy** page, so it graded only bundles that were never at risk. **Recorded as NOT verified rather than omitted:** #796's `PROFILE_MARKER_MISSING` is offline-only because its live trigger is a cookie-store **decryption** failure, which does not occur on Windows/DPAPI — it is the macOS Keychain path of [#768](https://github.com/ffroliva/gflow-cli/issues/768), a **named external blocker**, and the issue stays open; a synthetic attempt to force it failed informatively, since a corrupt store raises `sqlite3.DatabaseError`, bypasses the marker gate and still prints "check network connectivity" — the same misdirection from a different trigger, untouched here. #795's remediation string is unit-asserted only, and the *other* half of #795 — `credits` being labs-only on migrated accounts — is **not fixed**.) · previous: [LIVE_VERIFICATION_v0.73.1](LIVE_VERIFICATION_v0.73.1.md) (**Google's `glue` consent bar covering the migrated composer, measured and then cured, 2026-09-11 on `ci-probe` for one image generation and no Veo credits.** The premise was nearly rejected: both prior sightings of that selector in this repo are labs.google, and the 2026-09-10 spike had read 159 DOM samples on this exact host and project and watched the click land 3/3 — it simply never looked for a cookie bar. Seventeen hours later the bar was up: `position: fixed`, `z-index: 1000`, spanning y 547-655 of a 655 px viewport, with `elementFromPoint` over the settings trigger AND over the image submit both returning `span#glue-cookie-notification-bar-1-label` in **0 of 5** rendered samples hit-testable. **The control arm is what makes it a cure rather than a coincidence**: the same probe re-run through the driver's own `_dismiss_cookie_bar` restored both controls to **4/4**, one attempt, no error. A full `gflow image t2i` then ran end to end on the merged code — `status: ok`, a 807,761-byte `ffd8ffe0` JPEG that Pillow reads as 1376x768 RGB, matching the requested 16:9 and the reply's own dimensions, with `migrated.editor_ready` at 2.0 s and **no** `cookie_bar_not_dismissed`. A side finding the control arm produced by accident: the bar's buttons are `__accept` then `__reject` in DOM order, so anything taking "the first button" accepts cookies for the operator — the shipped fix anchors on reject. **Three items recorded as NOT verified rather than omitted:** the cure against a live bar outside the browser, because this release's own control arm consumed the consent on the only profile that had it; **prevalence**, because the second profile never reached an editor (both arms redirected to `/about`) so it is one account blocked and one already-consented, and KNOWN_ISSUES says so rather than generalising from n=1; and the video path live, which shares the identical function on the identical control but spends Veo credits.) · previous: [LIVE_VERIFICATION_v0.73.0](LIVE_VERIFICATION_v0.73.0.md) (**four error paths that stopped lying about what went wrong, verified 2026-09-10 on `ci-probe` and `denon82` at $0 — no run below reaches a Veo submit.** The decisive arm is an **A/B**, not an inspection: the pre-fix run printed an `accounts.google.com` challenge URL together with the OAuth `state` and `client_id`, and the identical command after the fix gave the same exit 38 and the same landing named with **five secret matches down to zero** — reading the post-fix output alone would only have shown that *these* secrets were absent. A second arm **falsified a claim already written into PR #775's body**: it said the change fixed the RED canary, and the live run showed `denon82` had moved to `accounts.google.com` mid-run, a landing the classifier excluded **by design**, with a unit test encoding the wrong reason — the offline suite was green throughout and could not have caught it. #776's happy path is verified by a real `gflow image t2i` (exit 0, 367 135-byte JPEG, `image_settings_applied` and `prompt_typed` both present, so **three of the four converted click sites** clicked successfully through the new helper); its **failure** path is real Chromium driving a page we wrote, which proves the mechanism and explicitly **not** that Flow still produces that state. **Three items recorded as NOT verified rather than omitted:** the `/about` redirect stopped reproducing (**0/5**, reported as *unmeasured* rather than argued into a retry flag), the `/fx/api/auth/*` landing moved before the run reached it, and #764's *success* path — an external contribution whose error boundary this release verifies but whose auto-selection needs an account currently behind a Google password challenge. The fourth converted click site is the **video** submit, unrun because it spends Veo credits, named rather than implied.) · previous: [LIVE_VERIFICATION_v0.72.0](LIVE_VERIFICATION_v0.72.0.md) (**auto-closing `gflow auth login`, verified 2026-09-09 on a migrated account at $0 across six runs.** Four paths: auto-close fires (`probe=in_context` at `elapsed_s` **84.8**, reproduced 3x), a manual close still verifies from disk (exit 0, never exit 12), `--browser auto` selects and completes, and a window closed **mid-2FA** is noticed in **2 s** (exit 8) where it previously ran the full 600 s deadline with the session endpoint touched **zero** times. **Two runs were discarded rather than counted**, and that is the useful part: the first returned exit 0 at `elapsed_s: 0.2` because a throwaway profile had survived the previous session and nobody signed in — a cached session reading as a clean pass — and three attempts at the manual-close path all returned exit 0 while being the *auto-close* path in disguise, separable only by an **absent** `probe=in_context`. The manual path was obtained by widening the poll interval for one run so the competing branch could not fire; the 2FA path by killing the profile's own Chrome, because hand-timing lost every race on a host that completes passkey sign-in in ~15 s. **Recorded as NOT verified rather than omitted:** the migrated-host image slice (#692) was **re-run this cycle** — CLI `t2i` PASSED first-hand on a moved account, while the queued-MCP `i2i` twin failed with a typed non-retryable exit 27 whose **cause is unknown** — some dialog opened after the file was chosen and no `maseQ` left the page, on an account that has ALREADY consented, so the message's "most likely the one-time upload-terms dialog" does not fit; no incident bundle was written, so nothing names the dialog, and it is filed for a spike rather than guessed; the subprocess fallback is unit-tested only; and the OAuth-callback *mechanism* behind the poll fix is **inferred, not proven** — a host-only gate does not exclude NextAuth's callback, since it runs on the app's own origin, yet all four sign-ins passed under it, so #769 carries the spike that would settle it.) · previous: [LIVE_VERIFICATION_v0.71.1](LIVE_VERIFICATION_v0.71.1.md) (**two migrated-host error paths stopped blaming the wrong thing, 2026-09-08, $0 across three profiles.** The agent-mode recovery landed earlier in this same release (#749) and worked, but its first cut collapsed three outcomes into one message — chip clicked / chip found but click blocked / chip clicked and mode still on — so a modal eating the click sent users to toggle a chip that was never the problem, and genuine selector drift *after* the mode was left was filed under an account setting the driver had already changed. Verified by an e2e that drives the account INTO agent mode and carries its own neutered-selector control; run **twice**, because review fixes landed after the first run and it no longer covered the code. **The larger find:** an account's FIRST upload on this host is blocked by Flow's one-time *"Rights to use this image"* dialog, which renders *after* the file chooser hands the file over — so `_dismiss_dialog` never sees it and the driver spent 60 s waiting for a request the page had already declined to make, then advised re-encoding the image. Six runs on `ci-probe` close the chain: guard fires while the dialog is up → owner-authorised accept → uploads (`media_id 8914400f`) → fresh session uploads clean with `dialog: None`. The order was forced, because the dialog is one-off and accepting it destroys the failing state — so the guard was written and verified BEFORE anything was clicked, and its firing branch is now unrepeatable here. `maseQ` is confirmed **not** renamed (six uploads, three profiles), killing #719's top hypothesis. Also: Flow's queue listing is on **`Zzl0ze`**, not the `jwpduf`/`as29s` progress polls, and two `abra_r2v_8s` records sat at status 6 for ~20 h unseen by any CLI command (#741). Three documents are corrected rather than patched — a KNOWN_ISSUES entry that told #719's reader this dialog did not affect gflow-cli, and a v0.71.0 host label calling `ci-probe` *labs* when it is migrated, which helped a credit-based theory survive four runs. Recorded as NOT verified: #719's second failure shape — an upload request that leaves the page and is never answered, ~1 run in 4 — which is unfixed and keeps #719 open.) · previous: [LIVE_VERIFICATION_v0.71.0](LIVE_VERIFICATION_v0.71.0.md) (**`gflow character create --voice` verified end to end for the first time, 2026-09-07 on profile `ffroliva`.** Before this release a repo-wide grep for `--voice` across `tests/e2e/` matched **nothing** — every voice test was a unit test of the hardcoded `VOICES` constant and the one that looked live parsed a fixture, so a voice that silently failed to attach was invisible to the whole suite while the command exited 0. A live create + `character show` read-back settles a contradiction between two of our own documents: `sent='Charon' stored='Charon' identical=True`, so the Capitalized canonical form round-trips and `CHARACTER_RECON.md`'s "preset id is the lowercased name" does not describe today's wire. Both docs also now record that `personalityNotes` is **Agent-scoped** per Flow's own editor copy, not a control on the audio engine. **A claim that was in the code is overturned:** the `_unported_form` entity guard said the submit "never produces a reply" and that the backend refuses — in fact `MZZa6b` replies with a null payload and the generation is **accepted and queued** (Flow types it `abra_r2v_8s` and renders it); the failure is the *observer*, which waits out `SUBMIT_REPLY_BUDGET_S` (60 s, calibrated on 4.0–4.6 s replies against an idle queue) and exits 9 while the video is still rendering. The guard stays until the observer is fixed, because a timeout reported on a healthy generation is worse than an explicit refusal. Also verified: credit shortfall reports exit **37** with `_raise_if_out_of_credits` called from **both** give-up paths — and it is *short for the selected model*, not empty (the measured account held **50** and asked for `veo-quality`, which costs **100**); the "+ New project" CTA anchored structurally on `add` (Tier-1 matches **1** where every previous entry matched **0**, control 47 ligature nodes) after this same release first asserted the control did not exist and a $0 run disproved it in one click; and the incident-bundle DOM dump de-blinded — it queried `i.google-symbols` only, so **every bundle a migrated user sent carried an empty ligature list**, which is why #727 and #731 stayed invisible. One item recorded as NOT verified rather than omitted: whether a bound character's **voice reaches rendered audio** (#738) — attachment is proven, application is not, because no entity-bound generation has yet returned a file to measure. The blocker is no longer credits or account access but retrieval, and the measurement is pre-calibrated: plate-bound takes of one character gave **88 / 103 / 118 Hz** against a **4.3 Hz** engine noise floor, so an applied Algenib should sit near its sample's **163.3 Hz**.) · previous: [LIVE_VERIFICATION_v0.70.0](LIVE_VERIFICATION_v0.70.0.md) (**reference-to-video on the migrated `flow.google.com` host, verified 2026-09-06 on profile `denon82`.** Three r2v runs bound their references every time; the decisive one is semantic rather than structural — two hand-drawn reference images went in and the clip came back carrying the same marker strokes, face and paper texture, which is what separates "Flow bound the reference" from "Flow accepted and ignored it". A four-beat, two-character piece was then produced end to end and joined with the concat filter: 19.000s, 1280x576, `clip_qa` **ok**, A/V **+0.000s**, both actors recognisable across every cut from one face plate plus the canon repeated verbatim. Also verified as exit-code corrections on the real CLI: `-o <existing directory>` now exits **2 in 0.8s** instead of exit 1 after ~2 min **and a billed, orphaned clip**; `character create` on a moved account exits **36 immediately** instead of a bare `RuntimeError` exit 1 after a 20s wait. Four items recorded as NOT verified rather than omitted: #692's original failure could not be reproduced [the reporter's own re-run was on a build without the fix, so it shows the failure is intermittent, not that the fix works]; Flow CHARACTER entities were never exercised because `character create` cannot run on any account available here; `image`/`scene`/`movie`/`extend` all exit 36 on the migrated host and no unmoved account exists here; and the sign-in-interstitial fidelity question from #699 has no bundle evidence either way.) · previous: [LIVE_VERIFICATION_v0.68.0](LIVE_VERIFICATION_v0.68.0.md) (`gflow update` self-upgrading a real `uv tool` and a real `pipx` install from their own running `gflow.exe` to the PyPI release, Windows 11, zero credits — see its own row below) · previous: [LIVE_VERIFICATION_v0.67.0](LIVE_VERIFICATION_v0.67.0.md) (the migrated `flow.google.com` driver — see its own row below) · previous: [LIVE_VERIFICATION_v0.64.0](LIVE_VERIFICATION_v0.64.0.md) (**#626 `--model omni-flash --end-frame` — first+last interpolation on Omni 1.1 Flash, verified 2026-09-02.** The guard this release removes was evidence-gated, so it was retired on evidence: a route-aborted submit capture on **two distinct Google accounts** (`ffroliva`, `denon82`) fired `video:batchAsyncGenerateVideoStartAndEndImage` with `startImage` AND `endImage` non-null at **zero credits**, ruling out a single-account artifact. The decisive layer is semantic, not structural: the paid 4s generation's **last frame IS the supplied end image** and its first frame is the start image — two visibly different subjects — which is the only check that distinguishes "Flow bound the end frame" from "Flow accepted and ignored it", and which layers 1–4 would pass either way. Render: 4.01s / 720×1280 / h264+aac / 2.4 MB, `MEDIA_GENERATION_STATUS_SUCCESSFUL`. The static capability table is **deleted**, not corrected — it went stale silently once and would again; a post-submit route check (`_assert_i2v_route`) now fails a run whose end frame Flow dropped back to the `StartImage` route, catching a rollback on any account without anyone re-reading a support page. One item is recorded as NOT verified rather than omitted: `--duration 10` + end frame is **submit-verified only** (`duration_set seconds=10` on the correct route with both images bound), because the status poll returned HTTP 401 mid-run and a retry 401'd at `project.create` before submitting — [#561](https://github.com/ffroliva/gflow-cli/issues/561), pre-existing, so verification was stopped rather than spending further credits on an environmental fault.) · previous: [LIVE_VERIFICATION_v0.63.0](LIVE_VERIFICATION_v0.63.0.md) (**`gflow video extend` — continuing a clip past Flow's 8s ceiling, verified 2026-09-01 on profile `ffroliva`, 20 Veo credits.** The decisive observation is that segment 2's `source_media_id` **is** segment 1's `media_id` (`0c9364f3…` seeded from `b9458021…`, then `648f9291…` seeded from `0c9364f3…`) — tail-only chaining, which no mock can establish; had it re-seeded from the source, the output would have been two divergent continuations of the same moment instead of one continuous shot. `extend_model_resolved` logged `candidate_count=99` against `SERVICE_TIER_INTERMEDIATE`, proving the key came from the live capability listing rather than a constant, and plan cost matched actual spend (`20 credits total, balance 875` → `Extended — 2/2 segment(s), 20 credits`). Render: 23.02s / 1280×720 / 24fps / h264+aac. **The run found a defect the offline suite is structurally incapable of seeing:** an extend segment carries **7.000000s** of content, not the 8 Flow advertises and bills, so server-side concat pads every internal seam with a frozen frame and digital silence (15s at −75.1 dB against −29 neighbours, reproduced on a second independent render) — filed in KNOWN_ISSUES with the three questions that must be answered before any clamp, and the reason `--extend N` on t2v/i2v was deliberately NOT shipped. Seven items are recorded as NOT verified rather than omitted: portrait `9:16`, `--aspect` against a mismatched source, chains longer than 2 segments, `--resume-from` against a live partial scene, the insufficient-credits refusal, a live Ctrl+C, and `OperationKind.EXTEND` rows.) · previous: [LIVE_VERIFICATION_v0.62.1](LIVE_VERIFICATION_v0.62.1.md) (**#604 `--model omni-flash` selects again after Flow's `Omni Flash` → `Omni 1.1 Flash` rename, verified 2026-08-30 on profile `ffroliva` at ZERO Veo credits.** The production `_select_video_model` was driven against the real picker: the shipped selector resolves verbatim — `model_selected model=omni_flash via="[role='menuitem']:has-text('Omni'):has-text('Flash'):not(:has-text('[Lower Priority]'))"` — and all six cases came back as expected: the four offered tiers SELECT, a tier Flow does not offer REFUSES, and a deliberately ambiguous selector REFUSES rather than resolving `.first`. Selection happens before submit, so the whole matrix is free. The transport's own refusal diagnostic read the live menu back as `Omni 1.1 Flash / Veo 3.1 - Lite / Fast / Quality`, independently corroborating the fixture. Four items are recorded as NOT verified rather than omitted: the rename *direction* [the `Omni Flash` baseline is a different account, locale and date], an actual `Omni ... [Lower Priority]` entry [Flow has never offered one], a full credit-spending generation on omni-flash, and #539's absence question, which stays open.) · previous: [LIVE_VERIFICATION_v0.62.0](LIVE_VERIFICATION_v0.62.0.md) (**#595 `auto` ≡ `classic` for images, verified 2026-08-28 on profile `ci-probe` — an account Flow had moved to the agentic cohort the day before — at ZERO Veo credits.** With no flags and no env vars the run logged `ui_driver.ui_mode.attempt_exit_agent` → `ui_driver.bound mode=classic ui_mode=classic` and generated a real 768x1376 JPEG (exit 0); the day before, the identical command on the same account bound `auto`→`agentic` and failed. The same log independently closed **#183**: `mode_switch_trigger` and `image_mode_tab`, the selectors it reported as unfindable, both matched. #591 was proven against the real 500-row catalog — **0** occurrences of the string `"None"`, 31 proper JSON `null`s, over a catalog holding 119 NULL-bearing rows. #592 shows exactly one `client.account_locale_cached` per command. Three items are recorded as NOT verified rather than omitted: the exit-28 abort on a *pinned* agentic account, the #597 batch inter-prompt guard, and any fresh announcement modal — the last two because all three accounts have already acked the current changelog, so no modal can be raised until Google ships the next one.) · previous: [LIVE_VERIFICATION_v0.61.0](LIVE_VERIFICATION_v0.61.0.md) (**#539 video-model refusal + #586 image-model refusal and server-side attribution, verified 2026-08-26/27 on profile `denon82`, ONE Veo generation total** — the video fix was proven by a **zero-credit A/B against the stashed pre-fix source running on live Flow**: the old code returned SELECTED for `veo-lite-lp`, a model Flow does not offer to this account, and for a deliberately ambiguous selector matching 3 entries; the new code refuses both with exit 18 naming what Flow offered. Model selection happens before submit, so refusals cost nothing and the A/B was free. The happy path still generates: `--model omni-flash` → exit 0, a real 2.2 MB `ftypisom` mp4, catalog recording `omni_flash`. Two items are recorded as NOT verified rather than omitted: three of #584's four navigation-settle sites were never exercised [no bearer refresh occurred in these runs], and #582's canary self-re-exec has its first real exercise tonight. Also falsified: #539's recorded note that the video picker uses a different trigger — the two constants are byte-identical strings, and the repeated empty menu captures were the capture's own fault.) · previous: [LIVE_VERIFICATION_v0.60.0](LIVE_VERIFICATION_v0.60.0.md)
+**What was actually live-verified for the latest release?** → latest: [LIVE_VERIFICATION_v0.82.0](LIVE_VERIFICATION_v0.82.0.md) (**2026-10-01.** `"ref": "batch:N"` and local-file refs on `gflow run` verified by a BDD e2e on flow.google.com: in place, no re-upload, lineage recorded; 1 veo-lite clip for the #926 e2e) · previous: [LIVE_VERIFICATION_v0.81.0](LIVE_VERIFICATION_v0.81.0.md) (**2026-09-30, 3 veo-lite clips.** The off-screen generation browser (#923) measured unchanged for image and video, window placement read back from Chrome, focus-steal measured and documented; #919 verified at merge) · previous: [LIVE_VERIFICATION_v0.80.0](LIVE_VERIFICATION_v0.80.0.md) (**2026-09-29, 0 Veo credits.** `--resolution` 360p/720p proven by driving the real settings pane on `omni_flash` and refusing the credit-spending submit, on the CLI transport **and** the MCP twin; `nano2-lite` e2e 2 passed; #891 measured before/after — `gflow run` `[nano2, imagen4]` went from exit 1 `RecaptchaError` with a lost results table to exit 36 naming the model, plus 3 e2e (fresh, warm, MCP) with a spy asserting zero client mints; #896 `e2e_data` 2 passed, its failure path offline-only. Named blockers: rendered resolution (needs a billed clip), the `/about` re-check arm, labs-served accounts.) · previous: [LIVE_VERIFICATION_v0.79.1](LIVE_VERIFICATION_v0.79.1.md) (**the signed-media download survives a dropped connection, verified 2026-09-22 at $0.** All three download sites route through one new shared helper, and that helper was driven through a real Flow clip route and a real signed CDN transfer by `pytest -m e2e_data` — 2 passed in 20.74 s, `ftyp` magic, byte length equal to the size Flow reports, and **no `url=` on any log event**. The retry itself could not wait for an organic reset — 20/20 clean here against the reporter's 2-in-6 — so it was **measured against a real TCP RST with an A/B control**: with `max_retries` the transfer recovers on a second connection, with it disabled the identical fault kills the call, and three attempts at a 5 s timeout finish in 1.57 s rather than ~15 s, proving the backoff shares one budget. **Two sites are NOT live-verified and name their blockers:** the composer call site needs a billed generation and `gflow credits` answers 401 on this account (#795), so no credits were spent; and the labs download cannot be reached at all from here (HTTP 308, 6/6 profiles). Offline: 21 new tests, full sweep 2143 passed.) · earlier: [LIVE_VERIFICATION_v0.79.0](LIVE_VERIFICATION_v0.79.0.md) (**the migrated host stops echoing a model it never observed, verified 2026-09-18 at $0.** `model_name_type` is `null` on `flow.google.com` in a real `image t2i --json`, with every sibling field populated — which is what makes the null a result rather than an empty envelope; the first read of it was flat rather than into `images[]` and returned `None` for everything, and that instrument error was discarded rather than reported. Both surfaces ran: `test_e2e_t2i_runs_on_a_moved_account` and `test_e2e_mcp_i2i_runs_on_the_migrated_host`, 2 passed in 86 s. **`gflow data download` was tested on an image and failed** — 45 s, then a message blaming the project, the trash and your prompt; recovery is video-only and the release notes now say so (#877). **#803 is NOT live-verified and names its blocker:** reaching those raise sites needs a non-JSON or token-less aisandbox reply, and the labs route refuses first with a 404 — itself a misleading error, filed as #875. Offline: 4 new unit tests, 1775 passed.) · previous: [LIVE_VERIFICATION_v0.78.0](LIVE_VERIFICATION_v0.78.0.md) (**generating with no project on `flow.google.com` — and video from a start frame — verified 2026-09-17.** The release exists because labs.google's `project.createProject` is retired — 404 *"Flow RPCs have been deprecated and disabled"* with a labs token, 401 without, 0 successes in 12 across four profiles — and gflow reported the 401 as an expired login. Creation and rename are driven on flow.google.com and read back from Flow's own `jHPbke` / `o8DA4` replies; the CLI and the MCP twin are each exercised separately, and both video paths reach a real submit that is aborted before Flow can bill it. Two failures are recorded and re-tested rather than explained away — one MCP image reply that did not reproduce under a four-run A/B, and a Chrome launch starved by a concurrent coverage run that passed in 36 s alone. The 3:4 image aspect is re-measured on the live radiogroup, five radios where 2026-09-08 saw four. Three items are NOT verified and each names its blocker: no labs-served account, no entitled `/about` account, no non-`en` locale. **#860 is the one thing here that cost credits, deliberately:** the only proof that a start frame binds is a real generation, so ten i2v clips were generated and downloaded on a live migrated account — `migrated.frame_uploaded` → `frame_bound` → `submit_observed` → `download`, the middle event being the one that could not fire before this release. **#859 is recorded as UNOBSERVED, not verified** — its dismissal ran on every one of those runs and no promo modal ever appeared, so the $0 route-intercepted e2e is what covers it. `gflow docs` (#861) is verified from the built wheel, which is its whole contract: 127 topics, `docs/FILE.md:LINE` search hits, no network and no checkout.) · previous: [LIVE_VERIFICATION_v0.77.1](LIVE_VERIFICATION_v0.77.1.md) (**a four-fix patch release, verified 2026-09-17 at ZERO cost.** The headline is #846: the published package could not run a single interactive command on a clean Windows install, because `ConsoleRenderer(colors=True)` needs `colorama` and nothing declared it. Verified as a controlled A/B on a fresh Python 3.13 venv — `SystemError` before, exit 33 after — plus an arm with stderr piped that shows why the AUTO/JSON split hid this from every CI job, and the MCP twin run separately. #848's import probe is exercised against a **deliberately broken venv** and an intact one, including the contract that a probe which cannot run reports nothing rather than reporting breakage. #852 is re-measured on the release head: 40 000 base64-shaped chars fall from 12.07 s to 0.020 ms with identical output on Workspace and custom domains. **#849 is a NAMED BLOCKER, not an omission** — its end-to-end behaviour needs a fresh interactive Google sign-in, which no agent can perform; its live A/B is recorded on PR #851 and its e2e file re-ran 4/4 here. **Seven live e2e failures are recorded rather than hidden**, and controlled: the release changes exactly four source files and none of them is a failing test's module.) · previous: [LIVE_VERIFICATION_v0.77.0](LIVE_VERIFICATION_v0.77.0.md) (**the migrated-host login lockout (#791), verified as a controlled A/B on a real Google **Workspace** account at ZERO cost — same profile, same cookies untouched between arms, so the only variable is the code: exit 8 with the Chrome marker rolled back and `channel_for_profile` collapsed to `None`, then exit 0 with the marker intact and the account recorded. The refusal path is measured too — of 10 local profiles, 5 reproduce #791 and exactly one carries both required cookies: 4 refusals, 1 upgrade, 0 false positives. v0.76.0 recorded this fix as **withdrawn**; the approach that shipped is a different one, server-attested on the landing URL rather than on client-rendered DOM. An earlier revision gated auth on an `@gmail.com`-only pattern that silently declined every custom domain — the account used here is one, which is why this is its verification. **#796 is reproduced and prevented**, not merely reported: it was filed as *'not yet reproduced end-to-end'*. **Two rows are named blockers, not omissions:** the MCP Registry publish (#841) cannot be rehearsed before a real tag push, so this release is its first exercise; and the migrated-host generation e2e cannot run here at all — every account bounces to `/about`, A/B-controlled against `develop` so it is not mistaken for a regression.) · previous: [LIVE_VERIFICATION_v0.76.0](LIVE_VERIFICATION_v0.76.0.md) (**a diagnostics-and-distribution release, verified 2026-09-16 at ZERO cost — no Veo credits, no image generation, no Flow submit.** The decisive arm is #833's **negative** control, run first-hand on two real accounts rather than relayed: `flow-pinhole-unavailable-screen` is **0** on both `ffroliva` and `denon82` while `aisandbox-root` is **1**, so the guard does not fire on accounts that HAVE Flow access and the zero is a real absence rather than a page that never booted — a guard with only a positive arm is indistinguishable from one that is permanently on. Both accounts settled at **`/about`**, which is the state that broke this release's *withdrawn* #791 oracle, and the exit-39 signal reads a component that only exists on the unavailable screen, so `/about` cannot mimic it. mcpservers.org is verified with a **404 negative control** because the listing 403s to `curl` and a naive check would have called the link dead. **#791 is NOT in this release and was withdrawn rather than deferred**: a fix passed eight council dimensions and was then refuted by a five-day-old spike nobody had routed to it — Flow's `/about` hop is decided client-side with zero requests to Flow, and *"the backend grants access while the frontend declines to open it"* (#756), so the probe would have reported authenticated users as logged out, reintroducing #791 as its own cure; the e2e caught it by going red on a live account whose state moved mid-session. **Two rows are blocked by the platform rather than by effort:** the Docker sign-in needs a WSL-integrated Docker host, and the MCP Registry publish cannot be observed until this very tag reaches `main`, since both triggers resolve the workflow from the default branch.) · previous: [LIVE_VERIFICATION_v0.75.0](LIVE_VERIFICATION_v0.75.0.md) (**a distribution-and-hardening release, verified 2026-09-15 on `ffroliva` at ZERO cost — no Veo credits, no image generation, no Flow submit.** The `gflow serve` token fix is an **A/B against the released 0.74.0 wheel from PyPI**, not a diff read: the token was *set* in both arms, so the probe measures whether it is checked — before, `no token 200/200` with 13 tools listed and `wrong 200/200`; after, `401/401` with zero tools and `WWW-Authenticate: Bearer realm="gflow"`, while the correct token still gets 200. Loopback DNS-rebinding was re-confirmed at **421**/**403** because the fix forwards the bind host and getting that wrong would silently disable protection that works today — and an earlier draft of the accompanying advisory claimed that protection was *disabled*, a false claim caught by measuring rather than re-reading. #813's guard is verified against a **real wheel in a real venv installed without extras**, the arm first proving it is in that state (`av present: False  PIL present: False`); the decisive row is that a **nonexistent** manifest still exits **20** on the dependency error rather than file-not-found, which is what shows the guard precedes the manifest read and not merely the generation. The Rich escape fix is verified on the **console render separated from the structlog JSON** — matching the JSON line would have been a false pass, since `--json` was never affected — with a **neutered control** through the same Rich version reproducing the bug exactly: `UNESCAPED -> pip install 'gflow-cli'` against `ESCAPED -> pip install 'gflow-cli[chain]'`. Also: `gflow-cli.exe` exists in the built wheel so `uvx gflow-cli` resolves at all, and the wheel carries 138 entries with **zero** ZIP duplicates. **Three items recorded as NOT verified rather than omitted:** a successful multi-link `chain` run spends Veo credits and stays outstanding e2e debt on #813 — what changed is the refusal path, which *is* verified; `gflow credits` against a working balance is impossible here because the account still returns 401, pre-existing **#795** and unrelated to this release; and the SSE arm of the token fix was exercised on loopback only, a deliberate scope limit, since binding a LAN address to test it would put an authenticated daemon on a live network for the run. **Two rows are blocked by the platform rather than by effort:** the PyPI metadata and the MCP Registry listing are both unobservable until this release is uploaded — PyPI renders the last artefact's metadata, and `mcp-publisher` reads its ownership token out of the *published* README.) · previous: [LIVE_VERIFICATION_v0.74.0](LIVE_VERIFICATION_v0.74.0.md) (**the migrated driver's run-unique uploads and its Frames-picker confirm, plus the `credits` remediation, verified 2026-09-14 on `ffroliva` at ZERO cost — every arm is a read or an upload-and-stop.** The confirm branch is the interesting one: this cohort's picker closes on its own, so the branch would never execute against live Flow by accident, and the e2e drives `FRAME_COMMIT_GRACE_S` to **0.001** so the grace expires before Flow can auto-close and **every** run enters it — the difference between a named blocker and an unrun claim. A *genuinely stuck* picker still cannot be produced here and stays `Refs #792`, verifier the reporter. #795 was verified on the cohort it is for, because this account is now in it — `auth status` green, `credits user` exit **3** with `detail` `credits endpoint returned 401` and `route: credits`, where v0.73.2 printed the class-default *"SAPISID cookie missing… re-run `gflow auth login`"*, advice that is wrong in every word there and that can roll a profile's strategy marker back and start the #791 spiral; the `credits list --json` envelope carries the per-profile `remediation_hint` that used to be dropped. **Two items recorded as NOT verified rather than omitted:** #799's exit-25 raise site needs an account in the agent-only cohort — the maintainer's accounts are served `flow.google.com` but still have the recoverable agent chip, so the state is structurally unreachable on them, a named blocker and #799 stays open; and neither `r2v` nor `i2i` had its own attach path driven to completion — `r2v` was not attempted (its e2e spends Veo credits), while `i2i` **was** attempted twice and failed both times at exit 23, before reaching the code under test, because the `--project` UUID came from `gflow project list` (the aisandbox catalog) and those ids 404 on `flow.google.com` — a harness error, recorded rather than dropped, which left a side finding: gflow reports Flow's own `/404?reason=project` as `UiSelectorDriftError`, blaming our selectors for Flow's 404. **#791 is not fixed and was not attempted blind** — all nine local profiles were probed at $0 and none is in the #791 state (SAPISID + a live `flow.google.com` session + no labs NextAuth token); four print its exact symptom string and none of them is it, because that string is also what a stale profile says.) · previous: [LIVE_VERIFICATION_v0.73.2](LIVE_VERIFICATION_v0.73.2.md) (**the blank incident bundle, reproduced and then cured by a committed e2e test, 2026-09-12 on `ffroliva` at ZERO cost — every arm fails before submit.** The decisive layer is an **A/B run through the same instrument**, not an inspection: on `develop` the new test FAILS in 46.3 s and prints the bundle it got — `tag_counts.div = 0`, `title.length = 0`, `host_category = "other"`, a 4,254-byte screenshot — which is field-for-field the signature [#792](https://github.com/ffroliva/gflow-cli/issues/792) reported from a different account, a different OS and a different command; on the fix it PASSES in 37.8 s with `div` **25**, `title.length` **81**, `host_category` **`flow_app`**, `route` `/404` and a 38.4 KB screenshot. Both arms exit **23** on the same `.settings-trigger-button` drift at `flow.google.com/404?reason=project`, so the arms differ only in what was *captured*, never in what failed. **The reporter's own bundle contained the refutation of their diagnosis**: viewport 1280x720 is `client.py`'s client-owned page, not the 1920x1080 a stray transport tab would have — it was always the driven page, photographed one navigation too late, and the portrait correlation had no mechanism behind it. The existing quality benchmark could not have caught this: it captures deliberately on a **healthy** page, so it graded only bundles that were never at risk. **Recorded as NOT verified rather than omitted:** #796's `PROFILE_MARKER_MISSING` is offline-only because its live trigger is a cookie-store **decryption** failure, which does not occur on Windows/DPAPI — it is the macOS Keychain path of [#768](https://github.com/ffroliva/gflow-cli/issues/768), a **named external blocker**, and the issue stays open; a synthetic attempt to force it failed informatively, since a corrupt store raises `sqlite3.DatabaseError`, bypasses the marker gate and still prints "check network connectivity" — the same misdirection from a different trigger, untouched here. #795's remediation string is unit-asserted only, and the *other* half of #795 — `credits` being labs-only on migrated accounts — is **not fixed**.) · previous: [LIVE_VERIFICATION_v0.73.1](LIVE_VERIFICATION_v0.73.1.md) (**Google's `glue` consent bar covering the migrated composer, measured and then cured, 2026-09-11 on `ci-probe` for one image generation and no Veo credits.** The premise was nearly rejected: both prior sightings of that selector in this repo are labs.google, and the 2026-09-10 spike had read 159 DOM samples on this exact host and project and watched the click land 3/3 — it simply never looked for a cookie bar. Seventeen hours later the bar was up: `position: fixed`, `z-index: 1000`, spanning y 547-655 of a 655 px viewport, with `elementFromPoint` over the settings trigger AND over the image submit both returning `span#glue-cookie-notification-bar-1-label` in **0 of 5** rendered samples hit-testable. **The control arm is what makes it a cure rather than a coincidence**: the same probe re-run through the driver's own `_dismiss_cookie_bar` restored both controls to **4/4**, one attempt, no error. A full `gflow image t2i` then ran end to end on the merged code — `status: ok`, a 807,761-byte `ffd8ffe0` JPEG that Pillow reads as 1376x768 RGB, matching the requested 16:9 and the reply's own dimensions, with `migrated.editor_ready` at 2.0 s and **no** `cookie_bar_not_dismissed`. A side finding the control arm produced by accident: the bar's buttons are `__accept` then `__reject` in DOM order, so anything taking "the first button" accepts cookies for the operator — the shipped fix anchors on reject. **Three items recorded as NOT verified rather than omitted:** the cure against a live bar outside the browser, because this release's own control arm consumed the consent on the only profile that had it; **prevalence**, because the second profile never reached an editor (both arms redirected to `/about`) so it is one account blocked and one already-consented, and KNOWN_ISSUES says so rather than generalising from n=1; and the video path live, which shares the identical function on the identical control but spends Veo credits.) · previous: [LIVE_VERIFICATION_v0.73.0](LIVE_VERIFICATION_v0.73.0.md) (**four error paths that stopped lying about what went wrong, verified 2026-09-10 on `ci-probe` and `denon82` at $0 — no run below reaches a Veo submit.** The decisive arm is an **A/B**, not an inspection: the pre-fix run printed an `accounts.google.com` challenge URL together with the OAuth `state` and `client_id`, and the identical command after the fix gave the same exit 38 and the same landing named with **five secret matches down to zero** — reading the post-fix output alone would only have shown that *these* secrets were absent. A second arm **falsified a claim already written into PR #775's body**: it said the change fixed the RED canary, and the live run showed `denon82` had moved to `accounts.google.com` mid-run, a landing the classifier excluded **by design**, with a unit test encoding the wrong reason — the offline suite was green throughout and could not have caught it. #776's happy path is verified by a real `gflow image t2i` (exit 0, 367 135-byte JPEG, `image_settings_applied` and `prompt_typed` both present, so **three of the four converted click sites** clicked successfully through the new helper); its **failure** path is real Chromium driving a page we wrote, which proves the mechanism and explicitly **not** that Flow still produces that state. **Three items recorded as NOT verified rather than omitted:** the `/about` redirect stopped reproducing (**0/5**, reported as *unmeasured* rather than argued into a retry flag), the `/fx/api/auth/*` landing moved before the run reached it, and #764's *success* path — an external contribution whose error boundary this release verifies but whose auto-selection needs an account currently behind a Google password challenge. The fourth converted click site is the **video** submit, unrun because it spends Veo credits, named rather than implied.) · previous: [LIVE_VERIFICATION_v0.72.0](LIVE_VERIFICATION_v0.72.0.md) (**auto-closing `gflow auth login`, verified 2026-09-09 on a migrated account at $0 across six runs.** Four paths: auto-close fires (`probe=in_context` at `elapsed_s` **84.8**, reproduced 3x), a manual close still verifies from disk (exit 0, never exit 12), `--browser auto` selects and completes, and a window closed **mid-2FA** is noticed in **2 s** (exit 8) where it previously ran the full 600 s deadline with the session endpoint touched **zero** times. **Two runs were discarded rather than counted**, and that is the useful part: the first returned exit 0 at `elapsed_s: 0.2` because a throwaway profile had survived the previous session and nobody signed in — a cached session reading as a clean pass — and three attempts at the manual-close path all returned exit 0 while being the *auto-close* path in disguise, separable only by an **absent** `probe=in_context`. The manual path was obtained by widening the poll interval for one run so the competing branch could not fire; the 2FA path by killing the profile's own Chrome, because hand-timing lost every race on a host that completes passkey sign-in in ~15 s. **Recorded as NOT verified rather than omitted:** the migrated-host image slice (#692) was **re-run this cycle** — CLI `t2i` PASSED first-hand on a moved account, while the queued-MCP `i2i` twin failed with a typed non-retryable exit 27 whose **cause is unknown** — some dialog opened after the file was chosen and no `maseQ` left the page, on an account that has ALREADY consented, so the message's "most likely the one-time upload-terms dialog" does not fit; no incident bundle was written, so nothing names the dialog, and it is filed for a spike rather than guessed; the subprocess fallback is unit-tested only; and the OAuth-callback *mechanism* behind the poll fix is **inferred, not proven** — a host-only gate does not exclude NextAuth's callback, since it runs on the app's own origin, yet all four sign-ins passed under it, so #769 carries the spike that would settle it.) · previous: [LIVE_VERIFICATION_v0.71.1](LIVE_VERIFICATION_v0.71.1.md) (**two migrated-host error paths stopped blaming the wrong thing, 2026-09-08, $0 across three profiles.** The agent-mode recovery landed earlier in this same release (#749) and worked, but its first cut collapsed three outcomes into one message — chip clicked / chip found but click blocked / chip clicked and mode still on — so a modal eating the click sent users to toggle a chip that was never the problem, and genuine selector drift *after* the mode was left was filed under an account setting the driver had already changed. Verified by an e2e that drives the account INTO agent mode and carries its own neutered-selector control; run **twice**, because review fixes landed after the first run and it no longer covered the code. **The larger find:** an account's FIRST upload on this host is blocked by Flow's one-time *"Rights to use this image"* dialog, which renders *after* the file chooser hands the file over — so `_dismiss_dialog` never sees it and the driver spent 60 s waiting for a request the page had already declined to make, then advised re-encoding the image. Six runs on `ci-probe` close the chain: guard fires while the dialog is up → owner-authorised accept → uploads (`media_id 8914400f`) → fresh session uploads clean with `dialog: None`. The order was forced, because the dialog is one-off and accepting it destroys the failing state — so the guard was written and verified BEFORE anything was clicked, and its firing branch is now unrepeatable here. `maseQ` is confirmed **not** renamed (six uploads, three profiles), killing #719's top hypothesis. Also: Flow's queue listing is on **`Zzl0ze`**, not the `jwpduf`/`as29s` progress polls, and two `abra_r2v_8s` records sat at status 6 for ~20 h unseen by any CLI command (#741). Three documents are corrected rather than patched — a KNOWN_ISSUES entry that told #719's reader this dialog did not affect gflow-cli, and a v0.71.0 host label calling `ci-probe` *labs* when it is migrated, which helped a credit-based theory survive four runs. Recorded as NOT verified: #719's second failure shape — an upload request that leaves the page and is never answered, ~1 run in 4 — which is unfixed and keeps #719 open.) · previous: [LIVE_VERIFICATION_v0.71.0](LIVE_VERIFICATION_v0.71.0.md) (**`gflow character create --voice` verified end to end for the first time, 2026-09-07 on profile `ffroliva`.** Before this release a repo-wide grep for `--voice` across `tests/e2e/` matched **nothing** — every voice test was a unit test of the hardcoded `VOICES` constant and the one that looked live parsed a fixture, so a voice that silently failed to attach was invisible to the whole suite while the command exited 0. A live create + `character show` read-back settles a contradiction between two of our own documents: `sent='Charon' stored='Charon' identical=True`, so the Capitalized canonical form round-trips and `CHARACTER_RECON.md`'s "preset id is the lowercased name" does not describe today's wire. Both docs also now record that `personalityNotes` is **Agent-scoped** per Flow's own editor copy, not a control on the audio engine. **A claim that was in the code is overturned:** the `_unported_form` entity guard said the submit "never produces a reply" and that the backend refuses — in fact `MZZa6b` replies with a null payload and the generation is **accepted and queued** (Flow types it `abra_r2v_8s` and renders it); the failure is the *observer*, which waits out `SUBMIT_REPLY_BUDGET_S` (60 s, calibrated on 4.0–4.6 s replies against an idle queue) and exits 9 while the video is still rendering. The guard stays until the observer is fixed, because a timeout reported on a healthy generation is worse than an explicit refusal. Also verified: credit shortfall reports exit **37** with `_raise_if_out_of_credits` called from **both** give-up paths — and it is *short for the selected model*, not empty (the measured account held **50** and asked for `veo-quality`, which costs **100**); the "+ New project" CTA anchored structurally on `add` (Tier-1 matches **1** where every previous entry matched **0**, control 47 ligature nodes) after this same release first asserted the control did not exist and a $0 run disproved it in one click; and the incident-bundle DOM dump de-blinded — it queried `i.google-symbols` only, so **every bundle a migrated user sent carried an empty ligature list**, which is why #727 and #731 stayed invisible. One item recorded as NOT verified rather than omitted: whether a bound character's **voice reaches rendered audio** (#738) — attachment is proven, application is not, because no entity-bound generation has yet returned a file to measure. The blocker is no longer credits or account access but retrieval, and the measurement is pre-calibrated: plate-bound takes of one character gave **88 / 103 / 118 Hz** against a **4.3 Hz** engine noise floor, so an applied Algenib should sit near its sample's **163.3 Hz**.) · previous: [LIVE_VERIFICATION_v0.70.0](LIVE_VERIFICATION_v0.70.0.md) (**reference-to-video on the migrated `flow.google.com` host, verified 2026-09-06 on profile `denon82`.** Three r2v runs bound their references every time; the decisive one is semantic rather than structural — two hand-drawn reference images went in and the clip came back carrying the same marker strokes, face and paper texture, which is what separates "Flow bound the reference" from "Flow accepted and ignored it". A four-beat, two-character piece was then produced end to end and joined with the concat filter: 19.000s, 1280x576, `clip_qa` **ok**, A/V **+0.000s**, both actors recognisable across every cut from one face plate plus the canon repeated verbatim. Also verified as exit-code corrections on the real CLI: `-o <existing directory>` now exits **2 in 0.8s** instead of exit 1 after ~2 min **and a billed, orphaned clip**; `character create` on a moved account exits **36 immediately** instead of a bare `RuntimeError` exit 1 after a 20s wait. Four items recorded as NOT verified rather than omitted: #692's original failure could not be reproduced [the reporter's own re-run was on a build without the fix, so it shows the failure is intermittent, not that the fix works]; Flow CHARACTER entities were never exercised because `character create` cannot run on any account available here; `image`/`scene`/`movie`/`extend` all exit 36 on the migrated host and no unmoved account exists here; and the sign-in-interstitial fidelity question from #699 has no bundle evidence either way.) · previous: [LIVE_VERIFICATION_v0.68.0](LIVE_VERIFICATION_v0.68.0.md) (`gflow update` self-upgrading a real `uv tool` and a real `pipx` install from their own running `gflow.exe` to the PyPI release, Windows 11, zero credits — see its own row below) · previous: [LIVE_VERIFICATION_v0.67.0](LIVE_VERIFICATION_v0.67.0.md) (the migrated `flow.google.com` driver — see its own row below) · previous: [LIVE_VERIFICATION_v0.64.0](LIVE_VERIFICATION_v0.64.0.md) (**#626 `--model omni-flash --end-frame` — first+last interpolation on Omni 1.1 Flash, verified 2026-09-02.** The guard this release removes was evidence-gated, so it was retired on evidence: a route-aborted submit capture on **two distinct Google accounts** (`ffroliva`, `denon82`) fired `video:batchAsyncGenerateVideoStartAndEndImage` with `startImage` AND `endImage` non-null at **zero credits**, ruling out a single-account artifact. The decisive layer is semantic, not structural: the paid 4s generation's **last frame IS the supplied end image** and its first frame is the start image — two visibly different subjects — which is the only check that distinguishes "Flow bound the end frame" from "Flow accepted and ignored it", and which layers 1–4 would pass either way. Render: 4.01s / 720×1280 / h264+aac / 2.4 MB, `MEDIA_GENERATION_STATUS_SUCCESSFUL`. The static capability table is **deleted**, not corrected — it went stale silently once and would again; a post-submit route check (`_assert_i2v_route`) now fails a run whose end frame Flow dropped back to the `StartImage` route, catching a rollback on any account without anyone re-reading a support page. One item is recorded as NOT verified rather than omitted: `--duration 10` + end frame is **submit-verified only** (`duration_set seconds=10` on the correct route with both images bound), because the status poll returned HTTP 401 mid-run and a retry 401'd at `project.create` before submitting — [#561](https://github.com/ffroliva/gflow-cli/issues/561), pre-existing, so verification was stopped rather than spending further credits on an environmental fault.) · previous: [LIVE_VERIFICATION_v0.63.0](LIVE_VERIFICATION_v0.63.0.md) (**`gflow video extend` — continuing a clip past Flow's 8s ceiling, verified 2026-09-01 on profile `ffroliva`, 20 Veo credits.** The decisive observation is that segment 2's `source_media_id` **is** segment 1's `media_id` (`0c9364f3…` seeded from `b9458021…`, then `648f9291…` seeded from `0c9364f3…`) — tail-only chaining, which no mock can establish; had it re-seeded from the source, the output would have been two divergent continuations of the same moment instead of one continuous shot. `extend_model_resolved` logged `candidate_count=99` against `SERVICE_TIER_INTERMEDIATE`, proving the key came from the live capability listing rather than a constant, and plan cost matched actual spend (`20 credits total, balance 875` → `Extended — 2/2 segment(s), 20 credits`). Render: 23.02s / 1280×720 / 24fps / h264+aac. **The run found a defect the offline suite is structurally incapable of seeing:** an extend segment carries **7.000000s** of content, not the 8 Flow advertises and bills, so server-side concat pads every internal seam with a frozen frame and digital silence (15s at −75.1 dB against −29 neighbours, reproduced on a second independent render) — filed in KNOWN_ISSUES with the three questions that must be answered before any clamp, and the reason `--extend N` on t2v/i2v was deliberately NOT shipped. Seven items are recorded as NOT verified rather than omitted: portrait `9:16`, `--aspect` against a mismatched source, chains longer than 2 segments, `--resume-from` against a live partial scene, the insufficient-credits refusal, a live Ctrl+C, and `OperationKind.EXTEND` rows.) · previous: [LIVE_VERIFICATION_v0.62.1](LIVE_VERIFICATION_v0.62.1.md) (**#604 `--model omni-flash` selects again after Flow's `Omni Flash` → `Omni 1.1 Flash` rename, verified 2026-08-30 on profile `ffroliva` at ZERO Veo credits.** The production `_select_video_model` was driven against the real picker: the shipped selector resolves verbatim — `model_selected model=omni_flash via="[role='menuitem']:has-text('Omni'):has-text('Flash'):not(:has-text('[Lower Priority]'))"` — and all six cases came back as expected: the four offered tiers SELECT, a tier Flow does not offer REFUSES, and a deliberately ambiguous selector REFUSES rather than resolving `.first`. Selection happens before submit, so the whole matrix is free. The transport's own refusal diagnostic read the live menu back as `Omni 1.1 Flash / Veo 3.1 - Lite / Fast / Quality`, independently corroborating the fixture. Four items are recorded as NOT verified rather than omitted: the rename *direction* [the `Omni Flash` baseline is a different account, locale and date], an actual `Omni ... [Lower Priority]` entry [Flow has never offered one], a full credit-spending generation on omni-flash, and #539's absence question, which stays open.) · previous: [LIVE_VERIFICATION_v0.62.0](LIVE_VERIFICATION_v0.62.0.md) (**#595 `auto` ≡ `classic` for images, verified 2026-08-28 on profile `ci-probe` — an account Flow had moved to the agentic cohort the day before — at ZERO Veo credits.** With no flags and no env vars the run logged `ui_driver.ui_mode.attempt_exit_agent` → `ui_driver.bound mode=classic ui_mode=classic` and generated a real 768x1376 JPEG (exit 0); the day before, the identical command on the same account bound `auto`→`agentic` and failed. The same log independently closed **#183**: `mode_switch_trigger` and `image_mode_tab`, the selectors it reported as unfindable, both matched. #591 was proven against the real 500-row catalog — **0** occurrences of the string `"None"`, 31 proper JSON `null`s, over a catalog holding 119 NULL-bearing rows. #592 shows exactly one `client.account_locale_cached` per command. Three items are recorded as NOT verified rather than omitted: the exit-28 abort on a *pinned* agentic account, the #597 batch inter-prompt guard, and any fresh announcement modal — the last two because all three accounts have already acked the current changelog, so no modal can be raised until Google ships the next one.) · previous: [LIVE_VERIFICATION_v0.61.0](LIVE_VERIFICATION_v0.61.0.md) (**#539 video-model refusal + #586 image-model refusal and server-side attribution, verified 2026-08-26/27 on profile `denon82`, ONE Veo generation total** — the video fix was proven by a **zero-credit A/B against the stashed pre-fix source running on live Flow**: the old code returned SELECTED for `veo-lite-lp`, a model Flow does not offer to this account, and for a deliberately ambiguous selector matching 3 entries; the new code refuses both with exit 18 naming what Flow offered. Model selection happens before submit, so refusals cost nothing and the A/B was free. The happy path still generates: `--model omni-flash` → exit 0, a real 2.2 MB `ftypisom` mp4, catalog recording `omni_flash`. Two items are recorded as NOT verified rather than omitted: three of #584's four navigation-settle sites were never exercised [no bearer refresh occurred in these runs], and #582's canary self-re-exec has its first real exercise tonight. Also falsified: #539's recorded note that the video picker uses a different trigger — the two constants are byte-identical strings, and the repeated empty menu captures were the capture's own fault.) · previous: [LIVE_VERIFICATION_v0.60.0](LIVE_VERIFICATION_v0.60.0.md)
 **"Was the playwright upper bound raised, and on what evidence?"** → [LIVE_VERIFICATION_playwright_1.61](LIVE_VERIFICATION_playwright_1.61.md) (**`>=1.59.0,<1.60.0` → `>=1.61.0,<1.62.0` verified 2026-08-05, 1 Imagen + 1 Veo credit** — the 2026-08-03 regression that motivated the bound does NOT reproduce on 1.61.0: a live `i2v` drove the full chain with `image_uploaded status=200` → `frame_attached` → `generate_captured status=200` with `startImage` parsed, where **1.62.0 hung silently at exactly that upload step**; live `i2i` local-ref attach passed outright [96.7s]. **1.62.0 stays excluded** — never root-caused. Also found, A/B-proven pre-existing on BOTH 1.59.0 and 1.61.0: Flow dropped the duration-tab UI, so **`--duration` is currently broken for i2v** — `UiSelectorDriftError` fires correctly [refusing rather than silently accepting Flow's default, #288] but the selector needs re-deriving; that is why the i2v proof was driven via the CLI without `--duration`, and Flow's own server-side `PUBLIC_ERROR_VIDEO_GENERATION_TIMED_OUT` means no finished mp4 is claimed).
 **"Where is the reverse-engineered wire protocol for a feature?"** → the `*_RECON.md` design docs: [CHARACTER_RECON](CHARACTER_RECON.md) (Flow character entity protocol), [IMAGE_UPSCALE_RECON](IMAGE_UPSCALE_RECON.md) (`/v1/flow/upsampleImage` wire). Naming convention: one `<FEATURE>_RECON.md` per reverse-engineered surface, kept as the durable spec after the feature ships. Pre-capture research recons: [ASSET_TAGGING_RECON](ASSET_TAGGING_RECON.md) (`@`-mention asset tagging — shipped in v0.40.0, see [LIVE_VERIFICATION_v0.40.0](LIVE_VERIFICATION_v0.40.0.md)).
 **"What was live-verified for v0.66.1?"** → [LIVE_VERIFICATION_v0.66.1](LIVE_VERIFICATION_v0.66.1.md) — **the two migrated-origin fixes, verified 2026-09-03 on profile `ffroliva` at ZERO credits, on BOTH sides of what was then read as a flapping rollout** (superseded 2026-09-04: the handoff is a one-way per-account flag and exit 36 is non-retryable since). On a real `flow.google.com` load `get_ui_driver` raised `FlowHostMigratedError` in **0 ms** (was ~36 s: ~8 s detect window + ~24 s crop cascade + 4 s URL settle), exit 36, `retryable: true` (then; `false` since 2026-09-04), `ui_driver.migrated_host_bail` logged; `await_url_settled` returned `null` in **0 ms** (measured 4018 ms before); and the locale was recovered as `en` from `html lang=en-GB` where the URL gave `null`. The no-regression half is the stronger evidence: minutes later the same command landed on the OLD host and completed **exit 0** with a real 768x1376 JPEG in 42.2 s, proving the host guard is scoped and does not short-circuit the working path. Three items recorded as NOT verified rather than omitted: driving the migrated frontend (still impossible, #639 stays open), the pt-BR recovery live (measured `html lang=pt` on `denon82` but exercised live only on `en-GB`), and the `en-GB`→`en` region reduction against a locale where region is load-bearing (`zh-Hans`/`zh-Hant`) — only two locales observed.

{gflow_cli-0.80.0 → gflow_cli-0.82.0}/docs/LIVE_VERIFICATION_v0.52.0.md

@@ -25,7 +25,13 @@ This live verification run exercises recent core infrastructure, transport, and
 | **2. Artifact 0 (Prompt 0)** | `prompt_0_0.jpg` (862,548 bytes, `1:1` aspect, Nano Banana 2) | 🟢 PASS |
 | **3. Artifact 1 (Prompt 1)** | `prompt_1_0.jpg` (1,093,468 bytes, `16:9` aspect, Nano Banana 2) | 🟢 PASS |
 | **4. Structlog Invariants** | `ui_automation.image_mode_entered` → `ui_automation.image_model_selected` → `ui_automation.aspect_ratio_set` → `ui_automation.prompt_submitted` → `batch_jitter_sleep` (0.99s) | 🟢 PASS |
-| **5. DAG Intra-Batch Binding** | Prompt 1 (`16:9`) bound Prompt 0's generated asset as `ref: "batch:0"` via `batchGenerateImages` payload | 🟢 PASS |
+| **5. DAG Intra-Batch Binding** | Prompt 1 (`16:9`) bound Prompt 0's generated asset as `ref: "batch:0"` via `batchGenerateImages` payload | ~~🟢 PASS~~ **Withdrawn (2026-10-01, #913)** |
+
+> **Correction (2026-10-01, #913).** Row 5 could not have passed. At this tag the manifest
+> `ref` field was read only inside `resolve_batch_dependencies`, which nothing called
+> (`git grep item.ref v0.52.0 -- src`), and neither request builder attached a reference.
+> A live re-measurement on 2026-10-01 confirmed rows carrying `ref` ran as plain
+> text-to-image. See `docs/superpowers/spikes/2026-10-01-batch-ref-dropped.md`.
 
 ---
 

gflow_cli-0.82.0/docs/LIVE_VERIFICATION_v0.81.0.md

@@ -0,0 +1,84 @@
+# Live verification — v0.81.0
+
+> Evidence for the user-facing changes in this release, run against real Flow on
+> 2026-09-30. Where a check could not be run, the blocker is named. Each row is a run
+> that someone watched, not an inference.
+
+## Environment
+
+| | |
+|---|---|
+| Profile | `ci-probe` (real-browser Chrome strategy) |
+| Host Flow served | `flow.google.com` |
+| Transport | `ui_automation` → migrated composer |
+| OS | Windows 11 |
+| Date | 2026-09-30 |
+| Veo credits spent | 3 × `veo-lite` clips (two measurement arms, one final-code run) |
+
+## Summary
+
+| Change | Surface | Verified live? | Cost |
+|---|---|---|---|
+| Generation browser off-screen, `GFLOW_CLI_BROWSER_WINDOW_POSITION` (#923) | `FlowApiClient` launch (CLI and MCP); the standalone `UiAutomationTransport` launch is unit-tested only | ✅ Yes: image e2e, video run, window bounds, throttling | 3 images, 3 clips |
+| `ProfileAccessError` for a write-denied profile (#919) | generation client launch | ✅ at merge (PR #919): real Chrome on a write-denied profile raises `ProfileAccessError`, exit 11 | $0 |
+| `urllib3` 2.8.0 CVE lock (#920) | none (transitive) | n/a: no Flow surface; `pip-audit` clean in CI | — |
+
+## Off-screen generation browser — #923
+
+**Measured before adopting the change** (the contributor's branch, then the final code).
+
+| Check | Off-screen | Visible control |
+|---|---|---|
+| `gflow image t2i` (PR branch) | exit 0, image saved, 34 s | exit 0, image saved, 33 s |
+| `gflow video t2v --model veo-lite` (PR branch) | exit 0, mp4 saved, 53 s, 8 status polls, 0 stall events | exit 0, mp4 saved, 65 s, 11 polls, 0 stall events |
+| `document.visibilityState` | `visible` | `visible` |
+| animation frames / 100 ms timer ticks in 5 s | 300 / 50 | 295 / 50 |
+
+**Final code (merge `1f42b222`), default setting:**
+
+- `pytest -m e2e tests/e2e/test_json_output_e2e.py` → **3 passed, 1 skipped** (the image
+  t2i JSON shape among them). The skipped video test is covered below.
+- `gflow video t2v --model veo-lite --aspect 9:16 --json`, five layers:
+  1. **Count:** 1 mp4 in the output directory.
+  2. **Magic bytes:** `ftypisom` at offset 4.
+  3. **Shape:** h264, 720×1280, 8.0 s.
+  4. **Structlog:** `migrated.submit_clicked` 1, `migrated.status` 9, `migrated.result` 1;
+     zero stall, `bring_to_front` or reCAPTCHA-failure events.
+  5. **User-confirmable:** JSON `status: "ok"` with the clip's `media_id`.
+
+**Window placement, final setting, read back from Chrome (CDP `Browser.getWindowForTarget`):**
+
+| `GFLOW_CLI_BROWSER_WINDOW_POSITION` | Result |
+|---|---|
+| unset | `-21845,-21845` on every launch (Windows clamps -30000); off every monitor |
+| `100,100` | exactly `100,100` on 8 of 10 launches. Twice it was found elsewhere, once seen moving mid-run: a visible window can be moved once open |
+| empty | Chrome's own placement |
+
+**Focus:** Chrome took the foreground on 5 of 5 launches with `--no-focus-on-init` and on 3
+of 4 without it. The flag was dropped, and the docs state that Chrome can still take
+keyboard focus at launch.
+
+## Blockers (named, not waived)
+
+- **macOS / Linux:** no machine here. The flag is a standard Chromium switch; occlusion
+  behaviour there is unmeasured.
+- **Polls longer than ~5 minutes** (Chrome's intensive throttling): every clip here
+  finished in about a minute. The page reports `visible` off-screen, so throttling should
+  not engage, but it was not observed.
+- **Video e2e test:** `test_e2e_video_t2v_json_shape` passes `--duration 8`, which this
+  account's composer does not offer (KNOWN_ISSUES "Video duration control is absent").
+  It is refused before submit (exit 11), unrelated to this release. The same command
+  without `--duration` is the video row above.
+
+## Pre-tag gates
+
+- **PR #923:** CI 15/15 (SonarCloud skips fork PRs; it runs on the develop push).
+- **`/gflow:doc-review`:** mechanical sections PASS (version sites, links, website mirror, release artifacts).
+  - Council: Completeness **YELLOW → fixed**. The setting was unreachable from troubleshooting: added rows to the CONFIGURATION troubleshooting table and a pointer from the KNOWN_ISSUES chooser → consent gap. An unverified remedy for the first-upload rights dialog was removed in favour of the documented one. Tracked, not fixed: no error message names the setting yet.
+  - Council: Cross-reference **GREEN**. `ProfileAccessError` scope added in KNOWN_ISSUES and USAGE; an unmeasured causal claim about focus removed; this ledger now says the standalone launch is unit-tested only.
+  - Council: Drift **YELLOW → fixed**. The CHANGELOG and CONFIGURATION listed a "CDP launch" site that does not exist in `src/`.
+- **`/gflow:check`:** offline suite 4694 passed, 24 skipped, coverage 93%; ruff, format, pyright, hygiene, doc links, PII, council memory, website mirror all green.
+
+## Post-tag evidence
+
+_Filled after the tag: PyPI publish, GitHub Release, back-merge._