feat(gemini-web): expose image generation through /v1/images/generations (closes #10466)

[Abhishek4512009]

Summary

Adds a gemini-web image-generation path so a valid Gemini Web session can generate images (Nano Banana) through POST /v1/images/generations, following the existing chatgpt-web web-session precedent.

• Registry: gemini-web image provider entry (format: "gemini-web", cookie auth, alias gweb) with the nano-banana-web model. The -web suffix is deliberate: the bare nano-banana id is owned by adobe-firefly (operator decision 2026-07-31, pinned by the cheaperinference-image-models guard), and parseImageModel's bare-model scan walks providers in insertion order — a bare id here would have stolen that resolution. A regression test pins this.
• Executor (gemini-web.ts): new parseStreamResponseImages() extracts generated-image URLs from the StreamGenerate candidate extension block — inner[4][0][12][7][0], URL at entry[0][3][3] (handles both the string and list forms), dedupes across cumulative frames, upgrades to =s2048 full resolution, and deliberately skips web-search thumbnails at [12][1] so scraped images are never served as generated. Image mode (x_gemini_web_image_mode) captures every StreamGenerate frame (images can land in a later frame than the text), resolves on first image, and gets a 90s window. Chat mode is byte-for-byte unchanged — the original first-response capture path is preserved verbatim.
• Handler (handlers/imageGeneration/providers/geminiWeb.ts): drives the executor in image mode with an explicit generation-directive prompt (the web UI otherwise answers with web-search images instead of generating), caps n at 4 (each n is a separate ~30-60s web turn), returns URLs or b64_json (downloads the public googleusercontent asset), and surfaces the assistant's refusal text in the 502 when no image was produced. Injectable executor + image fetcher for tests.
• Dispatch: branch on format === "gemini-web" in handleImageGeneration.

Related Issues

• Closes feat(gemini-web): expose image generation through /v1/images/generations #10466

Validation

Change type: provider (new image-generation capability for an existing web-cookie provider).

• Focused unit suite: tests/unit/gemini-web-image-generation-10466.test.ts — 21 tests, all passing
• Fixtures built from the documented StreamGenerate frame layout for generated images (corroborated by gpt4free's Gemini provider and HanaokaYuzu/Gemini-API's _parse_candidate): string + list URL forms, cumulative-frame dedupe, web-search-image exclusion, size-directive handling, malformed-line tolerance
• Handler coverage: success (url + b64_json), refusal visibility, missing prompt/cookie, n-cap, executor error passthrough, sequential n>1
• Registry wiring: gemini-web/nano-banana-web + gweb/ alias resolution, and the bare nano-banana → adobe-firefly regression guard
• npx eslint on all touched files — clean (2 pre-existing any warnings in gemini-web.ts unchanged, verified against the unmodified file)
• npx tsc -p open-sse/tsconfig.json — 0 errors
• Adjacent regression suites all green: gemini-web (6 test files), chatgpt-web-image-silentdrop, image-generation-handler, image-generation-route, image-registry-gpt56, cheaperinference-image-models, adobe-firefly, freepik-image-handler, microsoft-designer-web, nanobanana-image-generation, image-generation-baseurl-3205 — 195 tests passing

Coverage Notes

• The Playwright browser transport is covered indirectly: handler behavior is validated via the injected fake executor; the frame parser is validated against realistic captured-shape fixtures. A live browser round-trip against a real Gemini session is the remaining verification step (no live session was available in this environment at PR time) — happy to run it once a session cookie is configured.

Reviewer Notes

• Protocol stability: the image frame layout is reverse-engineered (no official API), same category as the existing gemini-web chat path — behavior may change upstream. The layout is documented in the parser's docstring with its corroborating sources.
• v1 scope: text-to-image only; image-to-image editing through the session (upload path) is not implemented.
• Model id: nano-banana-web rather than bare nano-banana to avoid hijacking adobe-firefly's owned alias (see above + regression test).

[diegosouzapw]

Revisão do PR #10494: encontrei três pontos importantes antes do merge. Em cloudflare-playground.ts:309-310, o ramo “Attention Required” retorna após lançar o Chromium sem chamar close(), o que pode acumular processos em cada bloqueio. No streaming, o timeout fecha o transporte e o finally emite apenas [DONE] (439-442, 529-537), fazendo respostas vazias/parciais parecerem concluídas; isso também contraria a regra de não engolir erros em SSE. Na imagem Gemini, o handler propaga o status do executor, mas o fallback de credenciais só tenta outra conta para HTTP 401 (geminiWeb.ts:123-145; imageCredentialRetry.ts:95-98), então sessões expiradas/bloqueadas que caem nos caminhos 400/500 não fazem fallback como exige a issue #10466. Também recomendo separar o Cloudflare Playground/#10389 deste PR para reduzir o escopo. Os testes unitários e o diff hygiene probe foram revisados; não foi possível executar uma rodada browser-backed no head do PR, portanto as claims estão marcadas como PLAUSIBLE.

[diegosouzapw]

Pushed the 3 mandatory pre-merge fixes from review:

1. Browser/transport leak (cloudflare-playground.ts): the detected-Cloudflare-challenge branch now closes the launched browser on every non-success start() path (was only closed in the catch path before). Added a regression test asserting close() is called.
2. Streaming timeout misreported as success (cloudflare-playground.ts): a server-side chat timeout now emits an explicit timeout_error SSE chunk before [DONE] instead of silently completing. Added tests for both an empty-timeout stream and a partial-then-timeout stream.
3. Image-account fallback gap (geminiWeb.ts + imageCredentialRetry.ts): the underlying GeminiWebExecutor's expired/blocked-session failure modes (400/500, from its own Playwright timeout/catch-all branches) are now classified as retryable, so executeImageWithCredentialFallback advances to the next eligible account instead of only doing so on a plain 401. Added a multi-account regression test and an invalid-session test driving the real executor's actual status code.

Also merged current release/v3.8.50 into this branch — the PR had drifted ~100 commits behind (it's built on top of the still-open Cloudflare Playground PR #10442, so both feature sets travel together here per the existing bundling). All 51 merge conflicts were in auto-generated/regenerated artifacts (provider-count mirrors across docs/i18n/*/llm.txt, docs/reference/PROVIDER_REFERENCE.md, README.md/AGENTS.md/llm.txt/package.json, and 4 static SVG diagrams) — none in this PR's feature/fix code, which merged cleanly. Regenerated PROVIDER_REFERENCE.md via npm run gen:provider-reference and hand-updated the remaining provider-count mentions (341 → 342) to match the live catalog after this branch's Cloudflare Playground addition.

⚠️ base-red inherited: #9985 — typecheck:core/dashboard-typecheck now surface open-sse/services/compression/engines/omniglyphAdapter.ts TS2339, confirmed byte-identical to the current release/v3.8.50 tip (not touched by this branch) — a pre-existing base defect, not introduced here.

Full validation run: typecheck:core (1 pre-existing base-red, unrelated to this PR), ESLint clean on touched files, check-complexity/check-cognitive-complexity/check-test-discovery/check-dashboard-typecheck all within their ratchet baselines, check-changelog-integrity now clean post-merge. New/updated tests (19 across the two touched suites) all pass. One pre-existing, out-of-scope finding not touched: open-sse/config/imageRegistry.ts is 1017 lines (cap 1000) — from this PR's original diff, not the review findings above.