[svasakorn]

Summary

Re-introduces the i2i local-file ref-attach from #50 (reverted in #57) and fixes the #56 root cause by making the media-dialog upload selectors locale-agnostic. Branched off develop @ 5345c1b as you suggested.

Root cause (#56)

It turned out to be a locale leak, not an intermediate-menu / nesting difference. Your ffroliva dump (#issuecomment-4528075496) shows the same popover with the same "Upload media" button at the same level as mine — just rendered in pt-BR (Enviar mídia). The old UPLOAD_MEDIA_BUTTON = button:has-text('Upload media') didn't match the pt-BR label, so the click missed → expect_file_chooser never fired → the ~34s silent timeout.

Flow renders the dialog in the Chrome profile language, which --lang=en-US cannot override on an already-configured profile (matches your "server-side locale overrides the launch arg" observation; same class as #51 / #24).

Changes

src/gflow_cli/api/transports/ui_automation_video.py:

• UPLOAD_MEDIA_BUTTON → the locale-free upload icon ligature, :text-is('upload') (exact — so it doesn't also grab the Uploads tab whose icon is drive_folder_upload), scoped to [role='dialog'][data-state='open']. This is the selector you recommended, with the exact-match refinement.
• Fallback tier — the original English-text selector (has-text('Upload media')), tried only if the icon misses. Covers the narrow case where Google changes the icon ligature but keeps the English label. Short per-tier timeouts so two misses can't recreate a long silent hang.
• 'Add to Prompt' has no icon (confirmed via DOM), so it's selected structurally as the lone iconless button in the open dialog (.filter(has_not=<i.google-symbols>)) instead of by localized text.
• Fail-loud guard — if neither tier opens a chooser, screenshot + a clear RuntimeError. The message explicitly tells the operator to set the Chrome profile language to English (chrome://settings/languages) and that the Google account language is not enough — that distinction is genuinely easy to get wrong.

tests/e2e/test_transports_e2e.py:

• Adds the C2-style i2i e2e variant (test_e2e_i2i_local_ref_attach), GFLOW_CLI_E2E_PROFILE-gated, ui-automation transport only (the REST transports can't drive the add-media dialog). Generates a tiny PNG inline — no asset dependency.

The shared _upload_via_open_dialog is also used by I2V/R2V, so this hardens those upload paths too.

Verification

Real CLI, English-rendering profile, this branch:

11:22:55.614  image_mode_entered
11:23:00.372  count_setter_completed
11:23:13.986  image_uploaded  status=200   <- upload fired (~13.6s, not the 34s hang)
11:23:15.670  reference_attached
11:23:16.318  prompt_submitted
11:23:50.935  batch_response_captured status=200  -> image returned

Honest caveat: I can only verify on an English-rendering profile, where both the icon and text selectors would match — so this proves the end-to-end path works, but not the pt-BR path specifically. The icon selector is locale-independent by construction and your ffroliva dump confirms the same upload ligature there. Could you run the gated e2e (or your capture_image_add_media_dom.py from #59) on ffroliva to confirm before closing #56? I deliberately wrote "Refs", not "Closes".

Not fixed here (focused PR — flagging per your preference)

FRAME_SLOT_BY_LABEL (I2V Start/End slots, div[aria-haspopup='dialog']:has-text('{label}')) is still text-based and will hit the same locale issue on a non-English profile. It's the I2V frame-slot path, separate from this upload fix, and belongs with #24 — happy to follow up separately rather than widen this PR.

Test plan

• check_repo_hygiene.py — clean (255 files)
• ruff check src tests — clean
• ruff format --check src tests — clean (103 files)
• pyright src — 0 errors
• pytest -q -m "not e2e and not live" — 806 passed, 6 skipped (single local interpreter)
• Real CLI gflow image i2i --ref <png> on an EN profile — image returned (timeline above)
• CI matrix 3.11 / 3.12 / 3.13 (fork PR → action_required; needs your approve-and-run)
• e2e i2i variant on a logged-in profile (GFLOW_CLI_E2E_PROFILE=...)

DCO-signed (both commits). Refs #56, #24.

[github-actions]

Thanks for the contribution. This PR has been marked as an external contribution and routed for maintainer review.

Before merge, please make sure:

• The PR targets develop.
• Commits use a real Git identity or GitHub noreply email.
• External commits include a DCO sign-off: git commit -s.
• The PR does not include secrets, cookies, account tokens, signed URLs, or private captured data.
• Focused tests/docs are included for behavior changes.

GitHub Copilot code review may be requested as an advisory first pass, but maintainer approval is still required.

[ffroliva]

Verification on ffroliva (pt-BR) ✅ — production code passes. But the test has a bug that needs a 1-line fix before merge.

Verification ledger (5 layers)

Layer	Result	Evidence
1. Exit code	0	pytest PASSED after 58s
2. File count	N/A	Test asserts on the API contract (media_name, fife_url), doesn't download — fine, the structlog covers the upload step
3. Pillow dims	N/A	(no file downloaded)
4. structlog invariants	✅ all green	image_mode_entered → count_setter_completed → image_uploaded status=200 target=ref0 (upload fired — NOT the 34s silent hang from before) → reference_attached index=0 → prompt_submitted (via locale-stable arrow_forward icon) → batch_response_seen/captured status=200
5. Gallery confirm	pending repo owner	Flow project becce409-a99b-4006-80bb-5ac516a562e1

The locale-agnostic upload icon ligature selector works on a pt-BR account. Mission accomplished on the production code — your fix direction was correct, and the exact-match :text-is('upload') refinement to avoid the Uploads tab confusion was a nice touch.

Test transport bug — please amend before merge

# tests/e2e/test_transports_e2e.py:169
async with _make_client("evaluate_fetch", profile) as client:  # ← REST
    image = await client.generate_image(req=req)

…but the docstring three lines up explicitly says:

UI-automation transport ONLY — the REST transports (bearer/sapisidhash) cannot drive the add-media dialog, so they never invoke _attach_references.

So the test as committed routes through evaluate_fetch (REST), which silently drops ref_paths (REST has no media-dialog code path). I had to manually patch this to "ui_automation" on a local checkout to actually exercise your fix — that's the run whose evidence is above.

This means your own EN-profile verification was a false positive (sorry to be the bearer): your test sent a REST request with the prompt, Flow generated a text-only image, the media_name + https:// assertion passed → green. The ref was never attached, the locale-agnostic selectors were never tested. On ffroliva we got a different false signal — AuthExpiredError: evaluate_fetch HTTP 401 because the REST session token is stale — but again, never reached the actual fix code.

Only after patching to "ui_automation" did the test do what its docstring promises.

One-line amend

- async with _make_client("evaluate_fetch", profile) as client:
+ async with _make_client("ui_automation", profile) as client:

(Could optionally inline a _make_client call with transport=None to use the registered default — same effect since _DEFAULT_TRANSPORT = "ui_automation" — but the explicit name is clearer.)

Optional follow-up (separate PR / your call)

The assertion image.media_name and image.fife_url.startswith("https://") would pass even if ref_paths were silently ignored (i.e. it doesn't detect the very class of bug we just lived through). A stronger assertion would be useful — e.g. inspect the captured wire request payload for an imageInputs[] entry, or assert the generated image dimensions match Aspect.PORTRAIT (text-only Narwhal vs ref-conditioned Narwhal usually differ in framing). Not blocking — just flagging.

State of PRs from this session

• i2i ref attach (#50) fails with TimeoutError on ffroliva account; image-mode add-media flow likely differs from video R2V #56 (this issue): reopened post-revert, still tracking; will close on merge of fix(image): re-introduce i2i ref-attach (#50) with locale-agnostic media-dialog selectors (#56) #60
• Revert i2i ref-attach (PR #50, commit 529d335); keep teardown guard. Closes #56 #57 (the revert) — merged
• chore(dev): add capture_image_add_media_dom.py for issue #56 #59 (capture script scripts/dev/capture_image_add_media_dom.py) — open, CI green, separate; not needed for fix(image): re-introduce i2i ref-attach (#50) with locale-agnostic media-dialog selectors (#56) #60 but might pair well as a future tripwire

Thanks again for the speed on this — the focused scope, the :text-is('upload') refinement, the iconless structural selector for "Add to Prompt", the fail-loud RuntimeError with the Chrome profile vs Google account language hint, and the honest scoping of FRAME_SLOT_BY_LABEL to #24 are all great. Ready to merge as soon as the test transport is fixed.

[svasakorn]

Thanks for merging #60! Heads-up: my test-transport fix (evaluate_fetch → ui_automation) landed a moment after the squash, so it missed the merge — the e2e on develop still routes through REST (the false positive you flagged). Small follow-up in #61. It also adds the reference_attached assertion so the test can't silently pass on a text-only image again.