PROJECT_STATUS.md

Project Status

Where gflow-cli is in its lifecycle, by release. Updated on every signed tag.

Current release

v0.73.0 — alpha. Four error paths stopped lying about what went wrong.

Every fix in this release is the same shape: gflow knew something had failed, and blamed the wrong thing. None of them changed what the tool can do — they changed what it says when it cannot, which is the difference between a user filing a useful bug and re-running blind.

A click that never lands now reports what was actually true (#776). On the migrated host a run reached migrated.editor_ready and died five seconds later as a bare Playwright TimeoutError: exit 1, no locator, no cause, no file. The control was visible — the guard above it proves that — and the click expired. Two causes were live and neither could be measured: Flow's announcement overlay (measured on labs.google, never on this host — a spike read body{pointer-events} as auto in 159/159 samples, including while Flow's own pane was open) and a mid-run agent-mode flip. So the driver reads Playwright's four actionability conditions back and reports the ones that fired; when every reading is healthy it says so, eliminating three and pointing at the fourth rather than inventing one. The MCP surface gains more than the CLI, where the same failure had been arriving as detail: "sha256:…" — a hash, not even the exception class.

A known Flow landing page is no longer reported as selector drift (#756). flow_host_kind() classifies the origin, and /about, /project/<id> and /fx/api/auth/signin all share one — so a readiness wait that timed out had nothing left to blame but its own anchor, sending the operator to "check for a newer release, then file a bug" over a session state no release changes.

Google's auth URLs no longer reach error messages with their query intact (#777). Those messages are the artifact users are asked to paste into an issue, and Google's auth URLs carry state, code_challenge, client_id and challenge tokens. Measured, not theorised: a real run printed five secret matches before the fix and zero after, with the landing still named — knowing where the session stopped is the whole value of the message.

Google's post-migration account chooser no longer stalls a run (#763/#764, thanks @stgmt). gflow now auto-selects the profile's recorded account instead of stalling into an opaque exit 1 — with the row match anchored so the chooser's Remove and Sign out rows can never be clicked, and unselectable cases raising a typed exit 38.

Also shipped: the Bug Lane is now the documented route from symptom to fix (#774) — spike → debug → BDD → TDD → fix → e2e, written once and cited everywhere, with an offline guard that fails CI when a browser-only scenario has no e2e test bound to it.

See LIVE_VERIFICATION_v0.73.0.md for what was exercised against real Flow — and what was not. Four arms verified live, one in a real browser against a page we wrote, and three recorded as not verified with named reasons: /about stopped reproducing (0/5), the /fx/api/auth/* landing moved before it could be reached, and #764's success path needs an account currently behind a Google password challenge.

v0.72.0 — auth login closes the browser, and the migrated host generates images

v0.72.0 — alpha. gflow auth login closes the browser for you, and Flow's migrated host now generates images.

Sign-in no longer ends with an instruction. gflow drives your real Google Chrome through Playwright, watches for the completed Flow sign-in and closes the window itself; closing it yourself still verifies, because that is what three releases of docs told people to do. On a machine where Playwright cannot resolve a Chrome channel, or where Google rejects the browser anyway, login falls back automatically to the previous subprocess flow — there is no new flag and nothing to choose.

That rests on a measured retraction. The standing claim was that Google rejects Playwright's bundled Chromium; the 2026-09-08 spike found the discriminator is navigator.webdriver, not the browser binary — real Chrome without the stealth flags was rejected at /v3/signin/rejected in 17.5 s, while bundled Chromium with them signed in normally. The bare control arm is the only reason that concludes anything: with just the two passing arms it would have read as "the block is gone".

Four defects surfaced only by driving it live, none of which the offline suite could reach. The session poll was hitting Flow's NextAuth session endpoint every 3 s for the whole login — including while Google held the page for the OAuth callback — and it now stays off both Google's host and NextAuth's own routes, because the callback runs on the app's origin and a host check sails straight through it. Gating that poll on labs.google alone would have timed out on every migrated account. And page.is_closed() was reachable only from an exception handler, so a window closed during a 2FA challenge was never noticed: a full ten-minute deadline, session endpoint touched zero times, ending in the wrong error.

gflow image t2i and local-file i2i also arrive on the migrated flow.google.com host (#692, @arjhinety), driving Angular Image mode and the page-owned ogiZ0b wire — Nano Banana 2 / Pro, the four aspects measured there, counts 1–4, --project required, with the direct and queued MCP twins on one payload path. UUID/entity references, Agent instructions, Imagen 4, image batch and the 3:4 aspect stay unported there and are refused before submit with exit 36 rather than reported as selector drift.

See LIVE_VERIFICATION_v0.72.0.md for what was exercised against real Flow — and what was not. The OAuth-callback mechanism is inferred rather than proven; issue #769 carries the spike that would settle it.

v0.71.1 — two migrated-host failures stop blaming the wrong thing

v0.71.1 — alpha. Two migrated-host failures stopped blaming the wrong thing — and both were found by asking what the app was saying, which this driver had never done.

An account parked in Flow's agent mode gained a recovery earlier in this same release (#749), but that first cut collapsed three distinct outcomes into one message: the chip was clicked, the chip was found and the click was blocked, the chip was clicked and the mode is still on. All three read as "the chip was clicked to leave it … the mode may be pinned". So a modal eating the click sent the user to toggle a chip that was never the problem, and genuine selector drift after the mode was successfully left was filed under an account setting the driver had already changed — which is to say, not filed at all. _exit_agent_mode now returns a tri-state, a blocked click raises at once instead of waiting out the recovery window (worst case 55 s → 35 s), and a timed-out recovery re-reads aria-pressed before choosing between pinned and ordinary drift. _open_pane had the same defect one gate later — it guarded with count(), and agent mode leaves the trigger present-but-hidden — so a mode flip mid-run escaped as a bare Playwright timeout with no exit code; it waits on visibility now.

An account's first upload on the migrated host failed with "no maseQ reply within 60s … the upload never reached Flow or was dropped", and advised re-encoding the image. Neither the file nor the network was involved: Flow shows a one-time "Rights to use this image" confirmation after the chooser hands the file over, and sends nothing until a human accepts it — so the driver spent its whole budget waiting for a request the page had already declined to make. It is detected by counting dialogs across the upload rather than matching one: the dialog's two buttons carry no ligature and no data attribute, separable only by DOM order, and its copy is translated, so no anchor there satisfies the locale rule — whereas "a dialog appeared between the file being chosen and the wait expiring" is upload-related by construction and cannot rot when Angular renames a class. gflow does not accept it for you, and there is no flag to make it: it affirms that you hold the rights to what you upload, it is one-off per account, and the path already requires an interactive gflow auth login.

The driver also now watches the upload request, not just the response, so "nothing left the page" and "it left and Flow did not answer" are no longer the same message. That distinction is load-bearing — the second argues for a retry and the first argues against one.

Three documents were wrong and are corrected, not quietly patched. KNOWN_ISSUES.md said this upload dialog affected "the legacy worker, NOT gflow-cli itself … workaround: none needed" — true of the REST path, false since the migrated driver began using the editor's own upload, and it was the entry a user hitting #719 would find and be reassured by. LIVE_VERIFICATION_v0.71.0 labelled the ci-probe profile labs when it is migrated, contradicting v0.70.0 one day earlier; in a repo where "Flow's UI shows X" is not a fact until the host is named, that silently re-scoped every conclusion keyed to it — and it helped make a credit-based theory for #719 look plausible for four runs. Two $0 instruments were added to settle such things by measurement: a queue reader (Flow's listing is on Zzl0ze, not the jwpduf/as29s progress polls) and an upload-wire probe.

Verification: LIVE_VERIFICATION_v0.71.1 — the agent-mode recovery A/B run twice against a live account, and the consent guard's full six-run chain (guard fires → accept → uploads → never asked again). Recorded as not verified: #719's second failure shape (an upload request that leaves the page and is never answered, ~1 run in 4 on a consented account) is unfixed and the issue stays open; and the consent guard's firing branch is now unrepeatable here, because the dialog is one-off and all three available accounts have accepted it.

Unreleased — the migrated image slice of #639 is implemented and live-verified. gflow image t2i and local-file gflow image i2i drive the Angular Image mode and the page-owned ogiZ0b wire on moved accounts (Nano Banana 2 / Pro, the four aspects measured there — 16:9, 4:3, 1:1, 9:16 — and counts 1–4, with --project required); the direct and queued MCP twins share the same payload path. UUID/entity references, Agent instructions, Imagen 4, image batch and the 3:4 aspect remain unported on that host and are refused before submit.

v0.71.0 — character create --voice verified, and two retractions

v0.71.0 — alpha. gflow character create --voice is verified end to end for the first time, and two confident wrong diagnoses shipped and were retracted inside this one release.

Before v0.71.0 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 plus a character show read-back now closes that, and settles a contradiction between two of our own documents: sent='Charon' stored='Charon', so the Capitalized form round-trips and CHARACTER_RECON.md's "the preset id is the lowercased name" does not describe today's wire. Both docs also now record that personalityNotes is Agent-scoped — Flow's own editor says the agent uses it to craft scenes — so it is not a control on the audio engine.

A credit shortfall reports exit 37 instead of "file a frontend bug", and the actionable half is that it is short for the selected model, not empty: the measured account held 50 credits and asked for veo-quality, which costs 100. Incident bundles are no longer blind on the migrated host — the DOM dump 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. The "+ New project" CTA is anchored structurally on add rather than on English text.

Two mistakes and their retractions are both recorded, not just the corrected state. #739 asserted the migrated gallery renders no "+ New project" control gflow can drive; a $0 run created a project there in one click, and #740 reverts it. Separately the entity guard was given the reason "the backend rejects the generation" — entity-bound submissions are in fact accepted and queued (Flow types them abra_r2v_8s and renders them); what fails is the observer, because MZZa6b replies with a null payload and SUBMIT_REPLY_BUDGET_S (60 s, calibrated on 4.0–4.6 s replies against an idle queue) expires while the video is still rendering. The guard stays until that is fixed (#723), because a timeout reported on a healthy generation is worse than an honest refusal. Both were unproven negatives overturned by one cheap direct observation.

Recorded as not verified rather than omitted: whether a bound character's voice reaches rendered audio (#738). Attachment is proven; application is not, and the blocker is no longer credits or account access but retrieval.

v0.70.0 — character create on the migrated host

v0.70.0 — alpha. gflow character create works on the migrated flow.google.com host — it was never broken there; the driver was not driving.

The character editor is fully present on flow.google.com, backed by the same labs tRPC and aisandbox APIs. Only the view layer differs: labs renders React + Slate, the migrated host renders Angular + ProseMirror. Seven selectors missed, a 20 s readiness gate timed out, and the absence of a match was recorded as the absence of the feature — then hardened into a guard that aborted before probing the DOM, which made the claim unfalsifiable. A complete character now builds end to end on a moved account: name, personality, voice, portrait and body, with two distinct workflow ids, two distinct media ids and both images on disk, verified by independent read-back.

--model is now deterministic. The picker was best-effort — every failure logged a warning and let the generation run on whatever tier the editor happened to show, so --model nano2 could quietly return a Nano Banana Pro image. It now reads the visible menu, refuses an ambiguous match, and verifies by re-reading the chip rather than trusting the click. Aborting is free: the picker runs before submit, so a refusal costs no quota, while proceeding produces a paid artefact from the wrong model.

A failed character create no longer strands an "Untitled Character" in the project, and the rollback asks the backend before deleting — an empty local workflow list means only that this process failed to read a result, and on the migrated host a portrait generated fine while the client timed out. Each slot also takes its own primaryMediaId now; reusing the entity thumbnail across slots corrupted the body workflow through commit_workflow.

-o <existing directory> no longer costs you a clip: exit 2 in 0.8 s instead of exit 1 after ~2 min with a billed, orphaned mp4.

Spike is now Phase 0 of the workflow (/gflow:spike), with the CDP/HAR harness vendored in-tree and CONTRIBUTING telling contributors how to produce redacted evidence for a bug report.

Verification: LIVE_VERIFICATION_v0.70.0 — the complete character read back from the backend, --model determinism over four alternating runs, and the orphan-rollback A/B (1 → 2 entities without the fix, 2 → 2 with it). Recorded as NOT verified: the labs.google path for all seven changed anchors, because every account available here has been migrated by Google.

v0.69.0 — i2v on the migrated host, and credits

v0.69.0 — alpha. The migrated flow.google.com host gains image-to-video from a local start frame, and gflow can finally tell you what your balance is.

gflow credits user and gflow credits list report the current Veo balance and account tier for one saved profile or all of them, and gflow_get_credits is the MCP twin. The primary path is genuinely browser-free: saved cookies authenticate a labs.google session request, that client is closed before a separate cookie-free, Bearer-only call to the credits endpoint, so no cookie crosses an origin boundary. A browser is the fallback, not the route. Multi-profile inspection keeps partial successes — a stale profile degrades to its own row with a reason instead of failing the run — and --json is a stable automation contract. The balance funds Veo video only; image generation draws on separate per-model daily quotas, which the docs now say in three places (#671, @ziyedbe).

gflow video i2v --initial-frame <file> --project <id> now runs on the migrated host (#639, slice 1) — on a moved account, and by default on an unmoved one. The port is UI-driven and observed rather than replayed: the composer picks the Frames submode, uploads through the editor's own toolbar entry, reads the media id off the app's maseQ reply, finds the upload in the picker by file name (the picker exposes no media id in its DOM, and indexes a fresh upload late — hence up to three searches), and inspects the submit request as it leaves. An unbound chip is refused before the click, exit 23 at zero credits, because an empty Frames submit silently goes out as text-to-video. Not ported in this slice, and named in the exit-36 detail: --end-frame, a frame given by UUID or @Name, and r2v.

Also in this release: --model veo-lite-lp is drivable on the migrated host — the one tier its model map omitted, matched by the [Lower Priority] tag rather than a label (#669, @kittinan); moved accounts now exit 36 rather than a misleading RecaptchaError on image t2i/i2i/upscale/extend, because the token was minted on the pool's bootstrap page before any transport guard could fire (#673); and three migrated model-picker defects are fixed — a tier bound that the user did not ask for, a stacked overlay that broke the next run after a --model switch, and a short-circuit that logged nothing.

The video-production skill ships at SkillOpt epoch 1: a scored rollout picked veo-quality for a two-reference shot while correctly reciting that its reference cap is 0, because the skill stated the prohibition twice and the substitute nowhere. It now routes — and names video chain as the exception, since chains refuse omni-flash. The SkillOpt harness moved onto the project's own GFLOW_CLI_LLM_* settings, dropping a parallel provider configuration and both LLM SDK dependencies.

Verification: LIVE_VERIFICATION_v0.69.0 — i2v proven on two moved accounts in two locales, the decisive run being the CLI entrypoint on the Portuguese-locale account (exit 0, 64 s, 632,755 B byte-exact with the status record, catalog row present); credits re-tested on the maintainer machine that had originally caught it launching a browser, now status_code=200 with no browser and an e2e_auth test whose pass is the browser-free proof. Recorded as NOT verified: 5 of the 19 skill-benchmark tasks (blocked on a 20-request/day free-tier quota, not on the code), the MCP twin of the credits path live, and the items the i2v ledger already lists.

v0.68.0 — gflow update, and the once-a-day banner

gflow-cli keeps itself current: gflow update upgrades an install in place, and every command shows a banner when a newer release is on PyPI.

gflow update [--check] [--json] runs the package manager that installed gflow-cli, read off the install itself rather than guessed from PATH — uv-receipt.toml → uv tool upgrade, pipx_metadata.json → pipx upgrade, otherwise that venv's own python -m pip (a uv venv with no pip module is refused before anything spawns). It asks PyPI first and runs nothing when already current; if PyPI is unreachable the manager still runs. The outcome is read back from the venv, not from the manager's exit code: on Windows the running gflow.exe launcher holds its own file open, so uv tool upgrade installs the new wheel and then exits 1 copying the launcher — measured — and that is reported as an upgrade with a note, because the launcher only points at the venv's python and keeps working. After an upgrade the venv's Playwright version is re-read and the playwright install chromium hint appears only when it moved. Editable / local / VCS / source installs, a manager binary missing from PATH, and a manager run that leaves the version unchanged are exit 11. Deliberately no MCP twin: the manager would replace the venv under a live gflow mcp run and write onto its stdout channel (#668).

The once-a-day notice (#479) is now a bordered stderr panel on a terminal — new version, gflow update, release-notes link — and stays one plain line when stderr is piped, so 2>&1 | jq pipelines keep one line per event. Same cache, same gates (GFLOW_CLI_UPDATE_CHECK=0, CI, non-index installs).

Also in this release: CONTRIBUTING.md routes contributors — and their coding agents — through the same lifecycle AGENTS.md defines (phase → skill → artifact), the PR template carries a matching Lifecycle checklist, and the contributor quality-gate list is identical to AGENTS.md's again. And a contributor fix on the migrated host (#670, @ChandraLiuswanto): the composer now waits up to 5 s for the arrow_forward submit button to enable after the prompt is inserted — it flips roughly 100 ms after insert_text lands, and reading it synchronously made every video t2v on a fully migrated account exit 23 before submitting.

Verification: LIVE_VERIFICATION_v0.68.0 — a wheel lowered to 0.60.1 installed through uv tool and through pipx, then gflow update run from each install's own gflow.exe: exit 0 both times, venv at PyPI 0.67.0 afterwards, the uv run carrying the stale-launcher note. Recorded as NOT verified: plain-venv pip on Windows, macOS / Linux for any manager, a real interactive-terminal screenshot of the panel, and the first genuine banner (0.68.0 is the first release after the change).

v0.67.0 — Flow's migrated flow.google.com host driven for text-to-video

v0.67.0 — alpha. Flow's migrated flow.google.com host is driven for text-to-video, and it is now the default host for every request it can serve.

Google is moving accounts off labs.google one at a time (#639); both maintainer accounts moved during this cycle. The migrated frontend is the same product on Angular Material over a different wire (batchexecute, not aisandbox-pa): three $0 spikes and two real clips settled the DOM and the protocol (spike), and the new migrated composer drives it — option-group radios, model menu, contenteditable composer, arrow_forward submit — then observes the app's own YhhmEf / jwpduf / as29s replies and downloads the clip from its signed CDN URL. Routing (GFLOW_CLI_FLOW_HOST): under the default auto, flow.google.com takes every video t2v --project on moved and unmoved accounts alike; what it cannot serve yet (image, i2v/r2v, characters, scenes, extend, instructions, tools, project creation) keeps the labs driver on an unmoved account and exits 36 on a moved one. flow.google.com forces it, labs.google is the kill switch.

Also in this release: exit 36 is non-retryable — the handoff is a server-assigned per-account flag applied client-side, one-way, not the per-page-load flap the error text claimed (spike); and --duration accepts 4/6/8 s on the Veo 3.1 models where the account's cohort renders the control (#650, @stgmt).

Verification: LIVE_VERIFICATION_v0.67.0 — gflow video t2v on the moved account (default route) exit 0 in 49.9 s, on the unmoved Portuguese-locale account (forced route) exit 0 in 50.5 s, both mp4s byte-exact with the size Flow reported, recorder rows present; --model veo-fast --duration 6 aborts pre-submit with exit 11 at zero credits because this cohort renders no duration row for Veo. The positive Veo 4/6/8 path is recorded as NOT verified (cohort-external), and the labs-side guard can no longer be verified here at all — there is no labs account left.

v0.66.3 — the `<html lang>` hydration race

v0.66.3 — alpha. gflow was reading Flow's locale off the en shell it serves before the app rewrites it, so every account whose URL could not answer came back English.

Flow serves an en shell and sets the real <html lang> during hydration. The single early read added by #643 therefore returned en — for exactly the accounts #643 existed to help, the ones whose URL carries no locale segment. Reported by @maipmacrothorax-75 on a pt-BR account resolving en while both pages served <html lang="pt">, with the mechanism offered explicitly as a guess.

The guess was right, and it reproduces on the old host — so this was never a migration bug. Measured, two consecutive loads:

t (ms)	lang	readyState
887	en	interactive
1510	en	complete
2092	en	complete
2488	pt	complete

The useful part of that table is the negative result: readyState reaches complete a full second before the flip, so the obvious "wait for complete, then read" would have shipped the same bug with more code. DOM node count oscillates and does not discriminate either. Nothing cheap predicts the flip, so the resolver now observes it — read, bounded-wait for a change, re-read — and treats a timeout as an answer (the shell value was already correct) rather than a failure, logged distinctly so the field can tell the two apart.

Verification: LIVE_VERIFICATION_v0.66.3. Proven at zero credits: the helper captures the flip live (en → pt, +638 ms) and a real bootstrap on an en account resolves correctly at the cost of the 4 s bound. Those two are reported separately, because collapsing a component measurement into a user-facing claim is the mistake v0.66.1 made and v0.66.2 corrected.

v0.66.2 — the fast-fail that had never fired, and a locale cache answering the wrong question

v0.66.2 — alpha. The fast-fail v0.66.1 announced had never once fired, and the locale cache was answering a question nobody asked it.

v0.66.1 claimed exit 36 arrived in 0 ms on Google's migrated flow.google.com frontend (#639). It did not. That number came from calling get_ui_driver in isolation on a page already sitting on the migrated origin — a state no real run ever reaches, because project_editor_url only ever builds a labs.google URL and the hop to the new origin is a post-goto redirect that neither settle path waits for. The guard read a pre-redirect URL and declined. The reporter measured the truth on three consecutive runs: 57.0 / 57.1 / 58.3 s, through the slow selector-probe path, with the bail event absent from the timeline entirely.

Check the host where the run already blocks. A shared raise_if_migrated() now runs at get_ui_driver entry, on every detect_ui_mode poll tick, in _exit_agent_mode once the media panel is found absent, and inside mode_control's 8 s composer poll — the last of which was found by review after the first fix shipped, and covers both --ui-mode agentic and a redirect landing mid-ensure_media_mode. None of these adds a wait: page.url is a cached property and flow_host_kind is one parse plus a dict lookup. Three blanket except Exception handlers that would have demoted the abort to a warning now re-raise it.

Stop letting a lang attribute mean "this account redirects". Two locale defects, both shipped in v0.66.1. NOT_REDIRECTED was an absorbing state: it returned before the only site of the <html lang> recovery, so the profiles #643 was written for could never use it. And the recovered locale was then recorded as evidence of a redirect — which it is not, since every account declares one — switching the URL settle back on permanently and costing the full 4 s URL_SETTLE_TIMEOUT_MS on every bootstrap of any non-redirecting account. Measured on a real profile: 7.41 s → 2.66 s, locale still recovered (#643).

Two designs were rejected by measurement, not preference: a bounded settle (72/72 navigations across two accounts showed no post-goto URL change at all, so the bound would be pure dead time), and caching "this profile is migrated" to navigate straight to flow.google.com (it would have locked both maintainer accounts out of a working frontend the day it shipped).

Drivable for text-to-video since v0.67.0; #639 stays open for the rest of the matrix.

Verification: LIVE_VERIFICATION_v0.66.2. The old host is proven live at zero credits (exit 0, real 768x1376 JPEG, twice). The migrated path — which no maintainer account could reach, because the rollout flapped back mid-work — was verified by @maipmacrothorax-75 on a permanently-migrated account: 57 s → 4.08–4.26 s, ui_driver.migrated_host_bail present with at=mode_control on all nine runs, and the crop cascade and agent dismissal both gone. That measurement was posted to #639 before the tag was cut and was not read in time, so the shipped verification doc understated its own result; the correction is recorded there rather than quietly applied.

Found by that same verification and NOT fixed here: on a pt-BR account the locale resolves en while both pages serve <html lang="pt"> — the probe appears to read before the app sets the attribute. That is the population #643 was written for, so the <html lang> fallback can latch a locale the account does not use. Tracked separately; it is not a migration bug.

v0.66.1 — a fast-fail that never fired (superseded by v0.66.2)

v0.66.1 — alpha. Intended to make the migrated-origin failure instant and to stop discarding a learned locale. Both fixes were real but neither reached the path users take, and this entry is kept to record that rather than to claim it.

The 0 ms exit 36 reported here was measured on get_ui_driver in isolation, on an already-migrated page; the CLI route was unchanged at ~57 s. The <html lang> locale recovery was likewise exercised directly, while the client returned before it on any profile cached NOT_REDIRECTED. v0.66.2 fixes both, and additionally a settle regression this release introduced. See the correction block at the top of LIVE_VERIFICATION_v0.66.1.

What did hold: _check_logged_in accepting the migrated host, and the old-host no-regression result (exit 0 with a real JPEG), which was the load-bearing evidence in that release.

v0.66.0 — naming Flow's flow.google.com migration

v0.66.0 — alpha. Google started moving Flow to a new domain, and gflow blamed its own selectors for it.

On a migrated page load, labs.google/fx/tools/flow/project/<id> redirects to flow.google.com/project/<id> and serves a rewritten frontend containing zero <i> elements. Every gflow selector anchors on Material Symbols ligatures, so cohort detection and every mode control missed at once and the run died UiSelectorDriftError — exit 23, retryable: false. That error means "a selector rotted, file a bug", so it sent operators hunting where nothing was broken. The selectors are correct for the host they were written against; that host is being replaced under them (#639, reported by @maipmacrothorax-75 with measurements from both sides of the rollout).

v0.66.0 read the rollout as flapping per page load — the same account landing on the old host on one navigation and the migrated one on the next — and on that reading made FlowHostMigratedError (exit 36) retryable. The reading was wrong: the handoff is a server-assigned per-account flag that the labs.google app applies client-side with window.location.replace, one-way and not transient (settled 2026-09-04 — spike), so exit 36 is non-retryable since that fix. What v0.66.0 got right stands: gflow recognises the migrated origin and raises a distinct error, and CLI and MCP both inherit it — they share the raise site, and is_retryable is a single source of truth for the --json, MCP and worker envelopes.

Tracing the same root cause turned up three things the report did not name. _check_logged_in hard-required labs.google in the URL, so a valid authenticated session on the migrated host was read as logged-out — and that gate was a substring match, satisfied by any foreign URL merely carrying the string in a path or query. Separately, UiSelectorDriftError is an incident-capture trigger, so swapping in a new class would have silently switched off bundle capture for exactly the failure whose evidence is most wanted; a test now pins that invariant across all four arms of the raise site rather than a list of names.

This does not add support for the migrated frontend. It converts a confusing hard failure into a clearly named one. No retry helps — the flag is per account — and #639 stays open for the driver work, with the anchor recon it needs already recorded.

Verification: LIVE_VERIFICATION_v0.66.0 — proven against the real migrated frontend at zero credits (i_total: 0, flow_host_kind: "migrated", check_logged_in: true, exit 36, retryable: true as designed then — false since the 2026-09-04 fix), with a credit-free image t2i on the old host minutes earlier completing exit 0 to prove no regression.

v0.65.0 — the referenceEntity guard that had never fired

v0.65.0 — alpha. A safety net that had never once fired, and two crashes that spent your credits before failing.

The referenceEntity guard strips character entities the caller did not request, so a "poisoned" entity left in the Flow composer cannot smuggle itself into an unrelated generation. It had never run on any released version (#615, reported by @DioServis). The route glob could not match Flow's namespaced endpoint — and the video guard was dead for the same reason, which the report did not mention. It failed open, silently, because the guard logged only when it stripped something: "never ran" and "ran, nothing to strip" were identical silence.

That silence is why it hid for months, and fixing it is what made the fix provable. The guard now announces every intercepted request, so absence of the event is evidence (#620). The result is A/B-verified live at zero credits: without the fix the guard never fired and the e2e failed; with it, it fired and passed — same account, same prompt, one variable. That also settled the question that had held the fix for two days (Flow delegates to a dedicated Web Worker, so context.route suffices) and proved the request rewrite does not corrupt the body.

Separately, gflow video chain and gflow movie run historically died mid-spend on duration × model (#634) — after earlier links or scenes had already rendered and billed. Historically, manifests pairing unsupported durations or models caused late crashes — including an example manifest tested against obsolete assumptions (#635). Both surfaces now share the canonical duration check: Veo accepts 4/6/8s, omni-flash accepts 4/6/8/10s, and invalid combinations are refused before the first submit (and --dry-run refuses what the real run refuses). Single-clip i2v (#630) similarly exits 2 naming the model instead of exit 1 "Unexpected error".

Validation: a movie.toml scene duration accepts 4/6/8 for Veo 3.1 and 4/6/8/10 for omni-flash; invalid values are rejected during parsing (exit 11) before any spending occurs.

Known limitation: the guard covers browser-driven generation only. Direct-wire routes issued through Playwright's APIRequestContext are not routable at all and bypass it regardless of matcher — #619.

Verification: LIVE_VERIFICATION_reference_entity_guard.

v0.64.0 — i2v --end-frame on Omni 1.1 Flash

v0.64.0 — alpha. gflow video i2v --model omni-flash --end-frame — first+last interpolation on Omni 1.1 Flash (#626). Google shipped first-and-last-frame generation for Omni 1.1 Flash, so the guard that rejected that combination with exit 17 was enforcing a fact that had expired. omni-flash is also the only model exposing --duration, so this is the one route to a 10-second first+last interpolation.

The interesting part is what replaced the guard. gflow used to decide which models could carry an end frame from a hand-maintained mirror of Google's support matrix — which went stale silently, and could have gone stale in the permissive direction just as easily. That table is deleted, not corrected. gflow now checks the route Flow actually used after submit: a run carrying an end frame that comes back on batchAsyncGenerateVideoStartImage — Flow dropping the frame and billing a clip that was never interpolated — fails with WireFormatError rather than being reported as success. That catches a partial or staged rollback on any account, at any time, with nobody re-reading a support page. It also closes a narrower hole the old backstop had: it only fired when Flow dropped every frame to the T2V route.

Live-verified on two accounts at zero credits plus one paid render. The route-abort probe fired batchAsyncGenerateVideoStartAndEndImage with both images non-null on two distinct Google accounts, ruling out a staged rollout. The decisive layer is semantic, not structural: the paid 4s clip's last frame is the supplied end image and its first frame is the start image — the only check that distinguishes "Flow used the end frame" from "Flow accepted and ignored it", since exit code, HTTP status and file properties pass either way. Recorded as not verified rather than omitted: --duration 10 + end frame is submit-verified only, its status poll having hit the pre-existing 401 of #561. See LIVE_VERIFICATION_v0.64.0.md.

Breaking: scripts branching on exit 17 for omni-flash + --end-frame now see success. Exit 17 is unchanged for gflow video chain --model omni-flash, still rejected because chain-scale seeded i2v remains unverified.

Also in this release, contributor-facing: MCP↔CLI parity became a duty of every pipeline phase. This very change drifted — mcp/tools.py and docs/MCP.md went on telling agents omni_flash was rejected for i2v-with-frames — through green lint, types, the full suite and test_cli_parity.py, which is command-level and cannot see a docstring that lies. Each skill now owns a slice (scenario D13, plan task 6, pr-council-review D15, check step 1b, doc-review blocking on a false MCP claim); automating the mechanically checkable part is tracked in #628.

v0.63.0 — gflow video extend, past Flow's 8-second ceiling

v0.63.0 — alpha. gflow video extend — continue a clip past Flow's 8-second ceiling. Veo caps a single generation at 8 seconds. extend chains server-side continuations: each segment is seeded from the previous segment's media rather than an extracted still, so the join is continuous rather than a cut. The run lands as a Flow Scene, and -o/--output renders it to one file through the existing credit-free server-side concat.

The model key is resolved from the account's live capability listing rather than hardcoded (extend_model_resolved logs candidate_count, service_tier, unit_cost), which prevents a tier-403 instead of classifying one after the fact. A whole-run balance pre-flight aborts before the first submit rather than at segment 6 holding a half-length video, submissions are paced by the shared jitter resolver, and an interrupted run publishes its resume handle before the first submit so Ctrl+C reports real credits spent and --resume-from appends after the scene's true tail.

Live-verified at 20 credits, and the run found a defect the offline suite structurally could not: an extend segment carries 7.000s of content while Flow advertises and bills 8, so server-side concat pads every internal seam with a frozen frame and digital silence. That is filed in KNOWN_ISSUES.md with the three questions that must be answered before any clamp, and it is why --extend N on t2v/i2v was deliberately not shipped — a convenience wrapper whose default output is defective is worse than no wrapper. See LIVE_VERIFICATION_v0.63.0.md.

Also in this release: CLAUDE.md now @-imports AGENTS.md so the project's agent rules load rather than being politely requested, and AGENTS.md opens with a Skill Routing table making the /gflow: lifecycle the default workflow.

Releases v0.60.0 – v0.62.1 are not expanded below. This file drifted for five releases; rather than reconstruct their summaries after the fact, they are recorded accurately in CHANGELOG.md and in their own LIVE_VERIFICATION_v*.md evidence files. The gap is named here rather than hidden, so the next release does not inherit a silent hole.

v0.58.0 — catalog-name picker contract (#529) + r2v named-reference fixes

v0.58.0 — alpha. Catalog UUIDs resolve through Flow display names (#529) + r2v named references un-broken after Flow's picker redesign. The picker contract is now catalog UUID → workflows[].metadata.displayName → picker name search → exact UUID-in-thumbnail tile: the UI response collector preserves the sibling workflow name so generated-image catalog rows retain the search key (store-mode history only), image --ref <uuid> and I2V frame UUIDs are enriched with it, and UUID/UUID-stem/prompt-hint/grid-scroll fallbacks are gone from that path — a missing name falls through to the SHA-verified local-file upload or a typed error, never an unfiltered click. Live-verifying the r2v surface caught two silent UI drifts and fixed them: the picker dialog exposes no accessible tree (tiles are now text-matched with an anchored, locale-badge-tolerant regex), and clicking a result now attaches directly (include button demoted to legacy fallback). Four committed e2e tests pin all of it against real Flow, including a real r2v video generated from a catalog name. See LIVE_VERIFICATION_v0.58.0.md.

v0.57.1 — two long-open UI bugs root-caused from live evidence (#493, #451/#288)

v0.57.1 — alpha. Two long-open UI bugs root-caused from live evidence (#493, #451/#288). #493 was reported as an unrecognized "third editor variant"; it is not. Flow's expanded chat sidebar removes the classic composer entirely — no crop_* settings trigger and no Agent pill — which is one state producing both reported symptoms, and why the failure surfaced as UiSelectorDriftError (exit 23) rather than the retryable agentic error: with no agentic indicator on screen, the cohort detector matched nothing. Recovery hinged on a close button scoped to the sidebar's edit_square affordance, so a cohort lacking that ligature could never recover; ensure_media_mode now falls back to an unscoped close from the demonstrably stuck state, A/B-proven live. #451/#288 were never selector drift either: Flow's settings popover is model-conditional by cohort. While the historical negative matrix (denon82, 2026-08-14) found no duration row for Veo, a subsequent live capture (presentation-reels-google, labs.google, 2026-09-04, PR #650) confirmed that cohorts exposing duration controls offer 4/6/8s for Veo 3.1 and 4/6/8/10s for omni-flash. Both matrices remain valid for their respective cohorts. The CLI now validates the canonical model cap at the edge with exit 2 before any browser work. Also: --reference-entity no longer advertised on video i2v (its DTO always rejected it), a typed ReferenceNotFoundError (exit 32) replaces a bare Playwright timeout. (gflow models reported max_duration 0 for the Veo models under that rule; since #650 it reports 8, because the cohorts that render the row genuinely accept it.) The v0.64.0 note below described a duration users cannot set. Everything verified at zero credits. See LIVE_VERIFICATION_v0.57.1.md.

v0.57.0 — video joins the UI-mode policy + MCP truthfulness wave

v0.57.0 — alpha. Video joins the UI-mode policy (#299) + an MCP truthfulness wave (#496–#501). gflow video t2v/i2v and MCP gflow_generate_video now take --ui-mode/ui_mode and bind their driver through get_ui_driver after editor mount instead of a hardcoded classic bind; video has only a classic driver, so auto ≡ classic, an env-sourced agentic degrades with a warning, and an explicit --ui-mode agentic is refused with exit 2 before any browser work. When Flow serves the agentic editor and classic cannot be recovered, video now fails fast pre-submit (UiModeUnavailableError, exit 28, zero credits) instead of burning 30–40 s of doomed selector timeouts. The agentic direction was hardened to match the classic one: mode_control.ensure_agent_mode replaces _force_agent_mode (deleted) with real-click-first + aria-pressed verification, no tune-ligature check, unknown editor variants no-op with a warning, and the sanctioned reload carries an explicit 15 s timeout. On the MCP side: gflow mcp run --no-spend (#496) unregisters both generate tools so an agent cannot even see them; gflow_auth_status (#497) gives agents a credit-free pre-flight session probe; gflow_list_projects paginates honestly with offset/has_more/next_offset (#498); the stub gflow_list_characters — which always answered "no characters", an active lie — is gone (#499); and gflow://docs/known-issues is bounded to a small index plus a per-issue templated read (#501) instead of injecting ~70 KB on every read. On the supply-chain side the release finishes the Scorecard hardening tail: every workflow now runs with a least-privilege token and every uses: action is pinned to a full commit SHA, the remaining in-workflow package installs are hash- or digest-pinned (website/requirements.txt, Dockerfile.triage, pip-audit), CI enforces a test-count floor so "a green build that ran nothing" cannot pass, check_repo_hygiene.py fails on version disagreement between pyproject.toml/__init__.py/plugin.json, and a self-run OpenSSF Scorecard workflow publishes the score as a README/website badge with a SECURITY.md section explaining it. Everything in this release was verified at zero credits. See LIVE_VERIFICATION_v0.57.0.md.

Develop (unreleased, post-v0.59.0): nothing queued yet — see CHANGELOG.md [Unreleased] for the authoritative list. develop is the staging branch for the next release; this line is not a substitute for the changelog.

v0.56.0 — ops-hardening batch + honest mode-switch evidence (#477/#478/#479, #493)

v0.56.0 — alpha. Ops-hardening batch (#477/#478/#479) + honest mode-switch evidence (#493). A Chromium major-version downgrade guard refuses to open a persisted profile with an older bundled engine than last wrote it (ProfileEngineDowngradeError, exit 11 — pre-auth, pre-credits) instead of letting Chromium's downgrade cleanup shred the session store; GFLOW_CLI_LEASE_WAIT_SECONDS=N adds an opt-in bounded wait on same-profile lease contention (default keeps the historical fail-fast); a once-a-day cache-served PyPI check prints a one-line stderr notice when a newer gflow-cli exists; and the exit-23 mode-switch fall-through now states that no known Flow cohort matched (the third, unrecognized editor layout from #493) while the drift remediation names the artifacts runs actually produce (diag_mode_switch_miss.json, the referenced screenshot, the incident bundle) instead of a phantom "debug screenshot from this message". Guard, lease wait, and the mode-switch path were live-verified; #479's notice is deferred-with-reason to a post-release check (it needs a published newer version on PyPI). See LIVE_VERIFICATION_v0.56.0.md.

v0.55.0 — Tier-1 hardening batch + docs truth sweep (#471–#476)

v0.55.0 — alpha. Tier-1 hardening batch (#471–#476) + docs truth sweep. gflow auth status now proves the Flow session (cookie-jar probe of the live session endpoint, no browser, no credits) and exits 0/1 with remediation hints; gflow mcp setup is implemented (claude-desktop / cursor / vscode, non-destructive merge with a pristine one-time backup); incident bundles stage a pre-filled report.md bug-report template built from allowlisted manifest fields only; Windows profile dirs get a real restrict-to-current-user DACL at login plus a marker-gated upgrade sweep at browser launch; all 11 MCP tools route through one error funnel that masks raw exception text from clients; llm_api_key/daemon_token are SecretStr so a Settings dump cannot leak them. The false "requires a Google AI Ultra or Pro subscription" claim was removed everywhere — any Google account with Flow access works; only feature gates (4K upscale) are tier-bound. See LIVE_VERIFICATION_v0.55.0.md.

v0.54.0 — login close-guidance + supply-chain guards (#465, #470)

v0.54.0 — alpha. Clearer close-the-browser guidance in gflow auth login (#470); overlay/watermark detection hardened to pure structural selectors; Dependabot now ignores playwright minors and all patchright bumps after PR #465 proved a pyproject.toml bound does not gate the uv ecosystem (it rewrites it); timing-jitter entropy moved to secrets.SystemRandom (SonarCloud S2245). First mainline release after the orphaned v0.53.x tags. See LIVE_VERIFICATION_v0.54.0.md.

v0.53.0 / v0.53.1 — interaction humanization + release-modal detection (#315, #403)

v0.53.0 — alpha. Driver interaction delay humanization (_jitter_ms, #315) and locale-invariant Flow release-overlay detection for the watermark-toggle modal (#403). v0.53.1 is a version-bump re-release recording headed live-verification evidence. Both tags were published without reaching mainline at the time; their content and honest CHANGELOG sections were reconciled during the v0.54.0 cut.

v0.52.0 — intra-batch references + entity provenance parity (#317, #402, #451)

v0.52.0 — alpha. Intra-batch reference support for gflow image batch (ref="batch:0" DAG ordering, #317); --reference-entity/--reference-entity-name parity on the video commands with provenance recording (#402); video duration selector cascade fix (#451). See LIVE_VERIFICATION_v0.52.0.md.

v0.51.0 — dependency & audit hygiene release

v0.51.0 — alpha. pip-audit now covers every optional extra; the full Dependabot backlog cleared in one lock update; the playwright bound raised >=1.59.0,<1.60.0 → >=1.61.0,<1.62.0 after live A/B evidence (1.62.0 stays excluded — it hangs i2v at frame upload). See LIVE_VERIFICATION_v0.51.0.md.

v0.49.0 / v0.50.0 — omni-flash I2V + MCP Tasks extension (#125, #409)

v0.49.0 — alpha. omni-flash start-frame I2V (#125), playwright upper-bound pin, submission-stage watchdog, count-tab fail-closed logic (#404). v0.50.0 — alpha. MCP 2026-07-28 Tasks extension (SEP-2663, #409) with non-blocking task handles, -o/--output path hardening (#414/#415), PR-triage alert resilience (#428). See LIVE_VERIFICATION_v0.49.0.md / LIVE_VERIFICATION_v0.50.0.md.

v0.48.0 — predictable output paths (#411)

v0.48.0 — alpha. Predictable output paths (#411). Adds an explicit -o / --output flag to image t2i/i2i and video t2v/i2v: the asset lands at the exact local path you name (parents auto-created, -o beats --out/--out-dir), with deterministic _1, _2, … suffixes for multi-count image runs. Cloud (s3:///gs://) targets, r2v/chain, video multi-count suffixes, and an MCP-tool output param are tracked follow-ups (#414, #415 — the MCP param was cut pre-release when the audit showed the worker queue never reads it). See LIVE_VERIFICATION_v0.48.0.md.

v0.47.0 — MCP SDK 2.0.0 migration + entity provenance (#402, #407, #408)

v0.47.0 — alpha. MCP SDK 2.0.0 migration + dual-era protocol + entity provenance (#402, #407, #408). Bounds mcp dependency to >=2.0.0,<3, migrates server to MCPServer, supports both 2026-07-28 and legacy protocol eras via SDK, defaults gflow serve to Streamable HTTP at /mcp, adds resolve-drift CI job, and records entity_ids / entity_names in operations.metadata_json on generation operations. See LIVE_VERIFICATION_v0.47.0.md.

v0.46.0 — prompt tools on any OpenAI-compatible endpoint (#387)

v0.46.0 — alpha. BREAKING — the prompt tools drive any OpenAI-compatible endpoint (#387). --tool creative-director / reverse-engineer / storyboard now speak OpenAI Chat Completions instead of being hardwired to Google's native Gemini API, so OpenAI, gateways/proxies (OpenRouter, LiteLLM), local runtimes (Ollama, LM Studio), and Google's own compatibility endpoint all work. New config: GFLOW_CLI_LLM_BASE_URL, GFLOW_CLI_LLM_API_KEY (optional — omitted when unset so keyless local gateways work), GFLOW_CLI_LLM_MODEL. The removed GFLOW_CLI_GEMINI_API_KEY / GFLOW_CLI_GEMINI_MODEL trigger a loud one-time notice, never a silent no-op. The user-supplied endpoint is treated as a trust boundary: redirects declined, https (or loopback http) enforced, error bodies redacted. reverse-engineer degrades to the original input instead of expanding a file path as a prompt. See LIVE_VERIFICATION_v0.46.0.md.

v0.45.0 — reference + character binding fixes (#393/#395)

v0.45.0 — alpha. Reference + character binding fixes (#393/#395). gflow image i2i --ref <UUID> now attaches the catalog's recorded file when Flow's per-project media picker cannot reach the tile, instead of failing the run — the fail-loud contract (never generate without a requested reference) is unchanged and pinned by a live e2e. gflow character create binds portraits to the character again: overlay dismissal was pressing Escape on Flow's own composer ([role='dialog']/[role='alert'] matched the app itself), and the character route could bounce to the project page, sending the prompt to the project composer — both produced generations with no entityContext, which Flow filed as plain project images. Also: a Flow web-app crash on the character route is now the typed retryable FlowAppError (exit 31), character binding failures say what actually happened, and --format-prompt (#383) is live-verified. 2810 tests pass. See LIVE_VERIFICATION_v0.45.0.md.

v0.44.0 — dual-side project naming & management (#381)

v0.44.0 — alpha. Dual-side project naming & management (#381). Feature set includes gflow project subcommand family (list, show, rename, create), --project-name / --project-title options across gflow image (t2i, i2i) and gflow video (t2v, i2v, r2v) generation commands, prompt slugging for scratch projects, dual-side title sync updating Google Flow's tRPC server/UI and local SQLite catalog in lockstep, HTTP 429 adaptive backoff / RateLimitError recovery (#384), character body prompt composer fixes (#378), and a repo-local Codex plugin ($gflow:* skills). Verified via full test suite and live Playwright Chromium E2E transport test. See LIVE_VERIFICATION_v0.44.0.md.

v0.42.0 — content-safety classification + Antigravity coding agent (#359/#360/#361)

v0.42.0 — alpha. Content-safety classification + Antigravity migration (#359/#360/#361). Content-safety 400 responses from Flow are now classified as ContentPolicyError (exit 5) instead of the misleading WireFormatError, so callers can branch deterministically on a policy rejection. Corrected a false flow_operation_id invariant (veo-lite can emit a remote_started checkpoint with operation_id unset; media_id is the canonical handle used by polling, download, and lookup). Replaced the retired Gemini CLI with Antigravity (agy) as the supported Google coding agent across skills and docs — Antigravity auto-discovers AGENTS.md, so the dedicated GEMINI.md hub is removed and agy is the pinned high-tier llm-council external reviewer. See LIVE_VERIFICATION_v0.42.0.md.

v0.41.0 — production-readiness hardening (#357)

v0.41.0 — alpha. Production-readiness hardening ([#357]): queue safety (versioned payloads, atomic claims, checkpointed execution phases), cross-process profile lease (ProfileLockedError exit 11), cancellation-safe browser teardown, driver honesty (typed SupportsSendPrompt injection, frozen TransportSetup), mention-index fail-closed (MentionIndexUnavailableError exit 29), and external-CDP lifecycle removal. Removed nonfunctional manifest-driven video batch command (never worked end-to-end; loop gflow video t2v/i2v from the shell instead). Also: /gflow:live-verify skill for per-feature live-verification enforcement. 2513 tests pass; live-verified against real Flow (stale-session fail-fast, free image gen, paid veo-lite T2V). See LIVE_VERIFICATION_v0.40.0-production-readiness.md.

v0.40.0 — prompt @-mention resolution for asset tagging (#344)

v0.40.0 — alpha. Prompt @-mention resolution for asset tagging (#344). @Name in a t2i/i2i/video prompt resolves to a staged, taggable character entity via services/mentions.py's resolve_and_apply, shared by the image/video CLI paths, the async worker, and MCP tools. Media-asset (non-character) @-mentions also work, but on the image path only — video-path media mentions are Phase 3. A bare character with no reference images is rejected early with a clear error instead of failing deep in the UI attach. De-tagged prompts are persisted to the catalog. See REFERENCE_STRATEGIES for @-mention vs --reference-entity vs --ref. Verification: LIVE_VERIFICATION_v0.40.0 (gflow character create + @Zoro t2i passed live end-to-end against unmodified develop, ~2 Imagen credits; also records a same-cycle investigation dead end where a stale local WIP branch was mistaken for develop's real state).

v0.39.0 — failed-generation persistence + gflow data list errors

v0.39.0 — alpha. Failed generations are now persisted to the local catalog (#341). Every paid-generation path (video t2v/i2v/r2v, video chain, image t2i/i2i, multi-prompt t2i, gflow run, movie run, and the async worker) now records a terminal status="failed" operation row — with a stable error_type derived from the exception's RFC 9457 problem_type and a redacted error_detail — before the error propagates, so WAF-403 block onset, duration, and recovery windows are measurable instead of reconstructed from memory. The new gflow data list errors subcommand browses failed operations newest-first, and videos whose poll returns succeeded=false are now recorded as failed (previously mis-recorded as succeeded). Bounded retention + export deferred to #345. Verification: LIVE_VERIFICATION_v0.39.0 (a real image t2i HTTP-400 wire failure produced the first failed row in the production catalog, $0).

v0.38.1 — agentic-pin recovery (opt-in reload after a real toggle-off)

v0.38.1 — alpha. Agentic-pin recovery (#338). When a real, unforced Agent-toggle click lands but the classic media panel never mounts in place (the 2026-07-17 both-accounts pin), ensure_media_mode can now reload the page once — opt-in and sanctioned only from the pre-bind classic path — to re-roll the server's per-load cohort and mount the persisted isAgentModeToggled=false preference; a composer-render race that made the toggle unreachable was also fixed. Verification: LIVE_VERIFICATION_v0.38.1 (classic recovered on denon82 after ~2h of active server pin).

v0.38.0 — robust agentic↔classic mode control + i2i ref dedup

v0.38.0 — alpha. Robust agentic↔classic mode control + i2i reference dedup (#332, #314). --ui-mode is now driven by a state-aware mode controller reading Flow's Agent toggle aria-pressed state (locale-invariant), ending spurious "forced agentic — not recoverable" aborts caused by an icon heuristic that matched both modes; --ui-mode classic reliably reaches the classic editor. Repeated local --ref images in gflow image i2i are now attached by selecting the already-uploaded library tile via exact-filename picker search (with a virtualised-grid scroll fallback, #335) instead of re-uploading duplicates. Also ships the PR-triage autopilot implementation (#238/#333, host deployment staged separately) and a behavior-preserving cognitive-complexity refactor (#331). Verification: LIVE_VERIFICATION_v0.38.0 (agentic→classic recovery observed in a real run; dedup contract proven with an upload-then-dedup run pair; veo-fast t2v successful).

v0.37.0 — viewport 1920×1080 + agentic count enforcement

v0.37.0 — alpha. Viewport 1920×1080 + agentic count enforcement + FIPS-safe SAPISIDHASH (#313, #315, #329). UI-automation and auth-login viewports harmonized to 1920×1080 to blend with the most common desktop resolution; the agentic cohort's requested image count (-n) is reliably enforced via the Agent settings panel; the protocol-mandated SHA-1 in SAPISIDHASH is marked usedforsecurity=False so it works under FIPS-mode Python. Also bumps mcp to 1.28.1 (CVE-2026-59950). Verification: LIVE_VERIFICATION_v0.37.0.

v0.36.0 — diagnostic tooling + reference-entity-smuggling fix

v0.36.0 — alpha. Diagnostic tooling (GFLOW_CLI_HAR_PATH + GFLOW_CLI_DEBUG_TRACEBACK) + reference-entity-smuggling fix (#312/#316). Two opt-in, env-var-only debug knobs: GFLOW_CLI_HAR_PATH captures full Playwright network traffic to a HAR file (0600-hardened on POSIX); GFLOW_CLI_DEBUG_TRACEBACK prints the real exception + traceback for unhandled errors (console and --json) instead of the generic placeholder, while structured telemetry stays hashed unconditionally either way. Also fixes a poisoned character entity leaking referenceEntities into unrelated image i2i calls in the same project workspace. Also ships the llm-council skill (/gflow:llm-council), adding external CLI reviewers (codex/gemini, opt-in agy) alongside the internal council for high-stakes reviews. Verification: LIVE_VERIFICATION_v0.36.0 (live HAR capture against real Flow traffic; typed-error scoping confirmed live; unhandled path covered by 6 new tests, 30 total across both files).

v0.35.0 — multimodal reverse-engineering + storyboard tool

v0.35.0 — alpha. Multimodal reverse-engineering + Storyboard tool + Dynamic Token Budgeting (#305-follow-up). Adds gflow tools run storyboard to generate sequential visual prompts from single ideas, and integrates gflow tools run reverse-engineer with claude-video's watch.py script for frame extraction and multimodal deconstruction of video/URL references using Gemini. Token budgets now scale dynamically with character limits. Verification: LIVE_VERIFICATION_v0.35.0 (proven live storyboard expansion + frame extraction).

v0.34.0 — bidirectional UI cohort switching

v0.34.0 — alpha. Bidirectional UI cohort switching (#299). Introduces --ui-mode / GFLOW_CLI_UI_MODE to force classic or agentic Flow UI cohort layouts on the fly, with verification and an exit-28 fail-fast when a required layout cannot be reached. Verification: LIVE_VERIFICATION_v0.34.0.

v0.33.0 — anti-bot jitter and video i2v project name overrides

v0.33.0 — alpha. Anti-bot jitter and video i2v project name overrides (#241, #287). Configurable anti-bot jitter range via --jitter / GFLOW_CLI_JITTER_RANGE, lower default batch jitter (0.5–1.5 s), and --project-name overrides for resolving in-project assets on localized or virtualized project dropdowns. Verification: LIVE_VERIFICATION_v0.33.0.

v0.32.1 — browser teardown hardening

v0.32.1 — alpha. Browser teardown hardening and profile lock translation (#293, #283). Fixed Chrome process leaks on aborted context teardowns, translated launch failures to ProfileLockedError (exit 11), and fixed picker grid off-by-one scroll bounds. Verification: LIVE_VERIFICATION_v0.32.1.

v0.32.0 — in-project asset i2v frame selection

v0.32.0 — alpha. In-project asset i2v frame selection by UUID (#287, #288). Select existing assets for video initial/end frames by UUID in place without re-uploading, and add fail-fast for duration settings control presence. Verification: LIVE_VERIFICATION_v0.32.0.

v0.31.0 — wrong-media attribution defenses

v0.31.0 — alpha. Wrong-media attribution defenses and multi-ref picker scrolling (#281, #282). Added pre-download verification guards against ambiguous agentic-cohort image downloads (MediaAttributionError exit 26), and added viewport-scrolling fallback to resolve multiple sequential --ref selections in the virtualized picker grid. Verification: LIVE_VERIFICATION_v0.31.0.

v0.30.0 — agentic-cohort image path support

v0.30.0 — alpha. Agentic-cohort image path support and MCP video parameters (#258). Supported native 768x1376 still generations in the agentic cohort, added character-creation integrity guards, and mapped model/duration/count video parameters on the MCP server. Verification: LIVE_VERIFICATION_v0.30.0.

v0.29.0 — persistent gflow instructions CRUD

v0.29.0 — alpha. Persistent gflow instructions CRUD — see LIVE_VERIFICATION_v0.29.0.md and the CHANGELOG entry.

v0.28.0 — agent instructions (-i) steer agentic generation

v0.28.0 — alpha. Agent instructions (-i / --instruction) now actually steer agentic image generation (PR #263). Instruction cards sync to the project's Agent brief via PATCH …/agentInfo and the agent folds every enabled card into generation. Root causes fixed: conversational (not imperative) composer directive + the project_brief.enabled master switch. Verification: LIVE_VERIFICATION_v0.28.0 (crayon e2e GREEN).

v0.27.1 — v0.27.0 follow-up fixes + documentation sync

v0.27.1 — alpha. v0.27.0 release follow-up fixes and documentation sync (#239). Patch release wiring package version dynamically to build_handoff() and FastMCP server, escaping brackets in Rich console planning output, updating MCP agent guide, and adding gflow movie usage documentation. Verification: LIVE_VERIFICATION_v0.27.1 (credit-free baseline verification).

v0.27.0 — Global [style] block with named variants + prompt-aware resume for gflow movie

v0.27.0 — alpha. Global [style] block with named variants + prompt-aware resume for gflow movie (#239). A movie.toml can now express a visual style system once — prefix/suffix on [style] plus [style.variants.*] sub-tables — and select it per-scene via style_variant / style_suffix (deterministic composition, none reserved as the opt-out keyword). The handoff manifest records style_applied (variant/prefix/suffix/scene_suffix) per clip. Resume is now prompt-aware: completed scenes persist a style_hash; a scene whose composed prompt changed is regenerated instead of silently skipped, and dry-run marks it re-run (style changed). Carries forward v0.26.0 (i2i select-in-place by UUID). Verification: LIVE_VERIFICATION_v0.27.0 (credit-free CLI ledger).

v0.26.0 — image i2i references a generated image by UUID (select in place)

v0.26.0 — alpha. Reference a generated image in image i2i by its Flow UUID. A reference_images entry that is a media UUID is attached by selecting the already-existing asset in Flow's reference picker (located by UUID in the thumbnail URL, surfaced by display-name search when hidden) — no duplicate upload; local upload remains the fallback. Generated images also record their Flow display_name (credited @C1ph3r404). Verification: LIVE_VERIFICATION_v0.26.0 (live e2e GREEN).

v0.25.0 — remote-UUID i2v + silent-failure guards

v0.25.0 — alpha. video i2v from a generated image's UUID proven live (#237) — the picker-search attach was reworked to a local-upload path, producing a real 8s interpolation from a catalogued UUID. Home-.env config matrix (#240) verified live. Two silent failures made loud: video-as-image download rejection and rejected-upload fail-fast. Verification: LIVE_VERIFICATION_v0.25.0.

v0.24.0 — `--project` parity across CLI + MCP

v0.24.0 — alpha. --project parity across CLI + MCP. The video commands (video t2v/i2v/r2v) gain --project <id> to generate into an existing Flow project instead of a scratch one (#233/#234), matching image t2i/i2i; and the MCP gflow_generate_image / gflow_generate_video tools gain a matching project parameter (#235), so agent callers get the same capability. Both surface an already-wired worker capability (payload["project_id"]) and validate the id identically. Carries forward v0.23.0 (MCP generation live + macOS 401 fix). Verification: LIVE_VERIFICATION_v0.24.0.

v0.23.0 — MCP generation live + macOS 401 fixed

v0.23.0 — alpha. MCP generation goes live + macOS 401 fixed. The MCP server's gflow_generate_image / gflow_generate_video tools — previously non-functional stubs — are now wired end-to-end to the FlowWorker queue (background worker owns download + history recording), the tools prompt-expansion parameter is actually applied, and i2v/r2v require their frame/reference inputs at the tool boundary. The long-standing macOS generation 401 (#222) is resolved (#230, @gunalak): Flow cookies are read from the full jar by domain instead of a path-/ filter that dropped the /fx-scoped session token, and the headed context is seeded from a pre-launch snapshot when macOS can't decrypt the store. Carries forward v0.22.0 (Tools framework) + v0.21.0 (MCP server). Verification: LIVE_VERIFICATION_v0.23.0 (MCP wiring proven live; #222 reporter-verified e2e on macOS).

v0.22.0 — Tools framework ("Creative Director")

v0.22.0 — alpha. Tools framework ("Creative Director"). A TOML-defined prompt-tool system: creative-director rewrites a terse prompt into a vivid one via Google's five-component formula (public Gemini API, never-fatal), with 15 category-gated domain styles and deterministic banned-keyword stripping. Invoke it via the new gflow tools list/show/run group or the uniform -t/--tool option on every generation command (image t2i/i2i/batch, video t2v/i2v/r2v/chain), replacing the never-released -e/--expand. History records the original prompt, the submitted expanded_prompt, and metadata_json.tool provenance (redaction-honoring). "My Tools": user-authored TOMLs in <GFLOW_CLI_HOME>/tools/*.toml load automatically. MCP parity via gflow_list_tools + a tools array param; the legacy expand_prompt MCP prompt is deprecated. The Gemini expander gained an overall wall-clock budget. Carries forward v0.21.0 (MCP server over stdio + HTTP/SSE). Verification: LIVE_VERIFICATION_v0.22.0 (CI/automated complete; live owner-run pending).

Milestone history

Milestone	Status
Four error paths stop lying about what went wrong: a click that never lands reports the actionability condition that failed instead of a bare timeout (#776), a known Flow landing is named rather than blamed on the selector (#756), Google's auth URLs are stripped from error messages (#777), and the post-migration account chooser auto-selects instead of stalling (#763/#764)	✅ done (v0.73.0)
gflow auth login closes the sign-in browser itself, on a measured retraction — G12 blocks navigator.webdriver, not bundled Chromium (#767); gflow image t2i/local-file i2i driven on the migrated host (#692)	✅ done (v0.72.0)
Two migrated-host error paths stop blaming the wrong thing: Flow's agent mode (three distinct outcomes, not one message) and its one-time upload-terms dialog (#749/#752, #719 shape A)	✅ done (v0.71.1)
gflow character create --voice verified end to end for the first time; a credit shortfall reports exit 37; incident bundles no longer blind on the migrated host	✅ done (v0.71.0)
gflow character create driven on the migrated flow.google.com host; --model made deterministic by chip read-back; spike promoted to Phase 0 of the workflow	✅ done (v0.70.0)
Read-only credit balance in the CLI and MCP (gflow credits user / list, gflow_get_credits) over a browser-free HTTP path; image-to-video from a local start frame on the migrated flow.google.com host (#639 slice 1)	✅ done (v0.69.0)
gflow update self-update through the installing manager (uv tool / pipx / pip), venv-verified outcome; the update notice as a stderr banner; CONTRIBUTING routes contributors and agents through the AGENTS.md lifecycle	✅ done (v0.68.0)
Flow's migrated flow.google.com host driven for text-to-video; the default host for what it can serve (GFLOW_CLI_FLOW_HOST)	✅ done (v0.67.0)
Migrated-origin runs fail fast and keep their learned locale (v0.66.1's fast-fail never fired in the field; corrected in v0.66.2)	✅ done (v0.66.2)
Flow flow.google.com migration named as its own failure class (exit 36; retryable in v0.66.0, non-retryable since 2026-09-04)	✅ done (v0.66.0)
Repo scaffold, CI, license, README, disclaimer	✅ done
Auth login flow (one-time browser capture)	✅ done
Video: t2v / i2v / batch (Veo 3.1)	✅ done (v0.2.0a1)
Image generation (T2I/I2I, 1–4 per call, 5 ratios, 3 models)	✅ done (v0.3.0a1)
End-to-end smoke test against live Flow	✅ done
First public alpha release on PyPI	✅ done (v0.2.0a1)
Batch concurrency / per-worker Page pool (GFLOW_CLI_CONCURRENCY=N)	✅ done (v0.4.0a2)
Typed errors (RFC 9457 Problem Details) + per-class exit codes 3–7	✅ done (v0.4.0a2)
Retry / backoff + reCAPTCHA re-mint inside the retry loop	✅ done (v0.4.0a2)
Structured logs (structlog, JSON on pipe)	✅ done (v0.4.0a2)
Pluggable image transport + ui_automation default strategy	✅ done (v0.5.0a1)
gflow run --config <file> sequential JSON batches	✅ done (v0.5.0a1)
examples/ directory with runnable single-image + batch scripts	✅ done (v0.5.0a1)
Shell multi-prompt gflow image t2i (PROMPT..., --prompts-file, --stdin)	✅ done (v0.6.0a1)
Downstream-worker ergonomics (out_dir, health_check(), optional project_id, BrowserSessionClosedError)	✅ done (v0.7.0)
Signed-tag release verification + first stable (v0.7.0)	✅ done (v0.7.0)
gflow video t2v restored on ui_automation with first-class video download	✅ done (v0.7.0 unreleased → v0.8.0)
Image/video mode-switch symmetry + live verify on ffroliva (PR #40)	✅ done (v0.8.0)
README + AGENTS.md + llms.txt refresh, docs governance	✅ done (v0.8.1)
gflow video t2v model picker (5 Veo models) + --duration / --count	✅ done (v0.9.1)
gflow video i2v (start + optional end frame) on ui_automation	✅ done (v0.9.1)
gflow video r2v (reference-to-video, model-aware ref cap omni≤7 / veo≤3)	✅ done (v0.9.1)
gflow image t2i/i2i --model actually selects the model (was a no-op)	✅ done (v0.9.0)
Local SQLite catalog (data layer) recording every project / image / video / operation	✅ done (v0.9.0)
gflow data list {projects,images,videos,profiles} read CLI over the catalog	✅ done (v0.9.0)
ROADMAP.md published (themed milestones through v1.0)	✅ done (v0.9.0)
Locale-agnostic media-dialog upload selectors (fixes non-English Chrome profiles)	✅ done (v0.9.0)
Wheel-build fix (removed redundant force-include causing duplicate ZIP entries)	✅ done (v0.9.0 hotfix, PR #74)
--json machine-readable output across image t2i/i2i, video t2v/i2v/r2v, auth list + gflow models catalog	✅ done (v0.10.0)
Per-model reference-image caps for i2i / r2v (Veo 3.1 Quality rejects R2V)	✅ done (v0.10.0)
Google-account identity persisted per profile + auto-rename of first-run default (issue #92)	✅ done (v0.10.0)
External cloud storage (S3 / MinIO / GCS) via GFLOW_CLI_STORAGE_URI	✅ done (v0.10.0)
gflow data prune + aggregated asset listing (--all-copies) + cross-profile count fixes (#111, #113)	✅ done (v0.10.0)
Layered cost-stratified e2e test strategy (e2e_auth/e2e_image/e2e_video/e2e_batch/e2e_data/smoke)	✅ done (v0.10.0)
gflow video i2v routes to the Veo i2v endpoint (no silent T2V fallback) + veo-lite default (issue #125)	✅ done (v0.11.0)
Create-project generation works under Flow's "Agent" composer mode	✅ done (v0.11.0)
Image-model selection hardened for non-English Flow UIs (selector cascade, #94)	✅ done (v0.11.0)
gflow character rm — free character deletion (#150)	✅ done (v0.13.0)
Align I2V CLI flags with Flow UI Labels (--initial-frame) (#122)	✅ done (v0.13.0)
In-project governance (ruff T20, materiality Classifier)	✅ done (v0.13.0)
gflow movie — multi-scene, character-consistent video from a TOML manifest (entity reuse, resumable, handoff manifest)	✅ done (v0.14.0)
gflow image t2i/i2i — reference locked CHARACTER entities (--reference-entity) + --project for character-consistent stills	✅ done (v0.15.0)
gflow character — reusable Flow Character entities (create/list/show/voices), persist-before-spend saga (#145)	✅ done (v0.12.0)
gflow scene — Add Clip / Scenes compose + credit-free server-side extended video (runVideoFxConcatenation)	✅ done (v0.12.0)
gflow video chain — last-frame I2V chaining from a JSONL manifest (--dry-run/--max-links/--resume-from)	✅ done (v0.12.0)
gflow video extend — chained server-side Veo continuations past the 8s ceiling (tier-resolved model, whole-run balance pre-flight, resumable)	✅ done (v0.63.0)
i2v --model omni-flash --end-frame — first+last interpolation on Omni 1.1 Flash; static capability table replaced by a post-submit route check that fails a dropped end frame (#626)	✅ done (v0.64.0)
Create-project generation works under Flow's Agent docked chat panel	✅ done (v0.12.0)
Video status poll raises AuthExpiredError (exit 3) on mid-workflow 401 (#156) + Docker /dev/shm hardening	✅ done (v0.15.1)
Locale-free resource-picker include selectors — entity attach works on every account language (#170)	✅ done (v0.16.0)
gflow image upscale <mediaId> --scale 2k|4k — credit-free download-menu upscale, 4K Ultra-gated (#171)	✅ done (v0.16.0)
Cookie-store session verification fast path (verify_flow_profile, PR #168) + Playwright fallback	✅ done (v0.17.0)
Entity-attach exit-7 remediation hint + entity_attach_context drift telemetry (#174 interim)	✅ done (v0.17.0)
Agentic-UI exit-23 UiSelectorDriftError + out_dir wiring (#183)	✅ done (v0.18.0)
Patchright opt-in browser engine (GFLOW_CLI_BROWSER_ENGINE=patchright)	✅ done (v0.19.0)
Aspect-ratio overrides under Agentic & Classic cohorts + GFLOW_CLI_PREFER_CLASSIC (#193)	✅ done (v0.20.0 / v0.20.1)
MCP server (gflow mcp run stdio + gflow serve HTTP/SSE) + daemon/queue scaffolding	✅ done (v0.21.0)
Tools framework: gflow tools group + --tool + creative-director + "My Tools" + MCP parity	✅ done (v0.22.0)
MCP generation wired to FlowWorker (tool→queue→download→record) + tools applied + i2v/r2v boundary validation	✅ done (v0.23.0)
macOS generation 401 fixed — /fx cookie-path read + headed-context seed (#222/#230)	✅ done (v0.23.0)
--project <id> on video t2v/i2v/r2v + MCP project parameter (#233/#234/#235)	✅ done (v0.24.0)
video i2v from a generated image's UUID (#237) + home-.env matrix (#240) + silent-failure guards	✅ done (v0.25.0)
image i2i references a generated image by UUID — select in place, no duplicate upload + display_name capture	✅ done (v0.26.0)
movie.toml [style] block with named variants + prompt-aware resume (style_hash) (#239)	✅ done (v0.27.0)
Agent instructions (-i/--instruction) steer agentic generation — conversational directive + brief master switch (PR #263)	✅ done (v0.28.0)
Persistent gflow instructions CRUD + movie.toml instructions brief-sync + gflow_instructions_* MCP tools + CI-enforced MCP↔CLI parity (#192)	✅ done (v0.29.0)
Diagnostic tooling (GFLOW_CLI_HAR_PATH + GFLOW_CLI_DEBUG_TRACEBACK) + reference-entity-smuggling fix (#312/#316)	✅ done (v0.36.0)
Viewports 1920×1080 + agentic count enforcement + FIPS-safe SAPISIDHASH (#313/#315/#329)	✅ done (v0.37.0)
Robust aria-pressed agentic↔classic mode control (#332) + i2i ref dedup via picker filename search (#314)	✅ done (v0.38.0)
Agentic-pin recovery: opt-in reload after a real Agent-toggle click (#338)	✅ done (v0.38.1)
Failed generations persisted to the local catalog + gflow data list errors (#341)	✅ done (v0.39.0)
Prompt @-mention resolution for asset tagging (#344)	✅ done (v0.40.0)
Production-readiness hardening: queue safety, profile lease, cancellation-safe teardown (#357)	✅ done (v0.41.0)
Content-safety ContentPolicyError classification + Antigravity coding agent (#359/#360/#361)	✅ done (v0.42.0)
Private incident diagnostics (GFLOW_CLI_INCIDENT_CAPTURE)	✅ done (v0.43.0)
Dual-side project naming & management (gflow project family, --project-name/--project-title) (#381)	✅ done (v0.44.0)
--ref catalog-backed upload fallback + character-entity binding fixes (#393/#395)	✅ done (v0.45.0)
Prompt tools on any OpenAI-compatible endpoint (GFLOW_CLI_LLM_*) (#387)	✅ done (v0.46.0)
Classic count-setter: digit-keyed tab selection + typed drift error (#404)	✅ done (v0.46.1)
Predictable output paths: -o/--output on t2i/i2i/t2v/i2v (#411; MCP + cloud follow-ups #414/#415)	✅ done (v0.48.0)
Manifest-driven video batch runner on ui_automation	❌ removed — never worked end-to-end, shipped as a nonfunctional stub; see v0.41.0 changelog. For multi-clip video, loop gflow video t2v/i2v from the shell; gflow image batch remains supported for images.
Persistence layer (stay-mounted batch sessions across project boundaries)	⏳ Phase B
Provider abstraction for official Veo 3.1 API	⏳ planned
Signed-tag CI verification automation (no manual signing in CI yet)	⏳ planned

What's new in each release

For per-release deltas see CHANGELOG.md. Per-release evidence files (live verification, screenshots, smoke logs) live under docs/LIVE_VERIFICATION_*.md.

Lifecycle policy

• Alpha (0.x.y) — current. APIs may change between minor versions; breaking changes are noted in the changelog.
• 1.0.0 — stable surface. Breaking changes require MAJOR bump + migration notes.
• Patch releases — bug fixes, doc refreshes (like v0.8.1), and other backward-compatible changes.

See RELEASE.md for the full release protocol and the prerelease vs full-release policy.