test(e2e), docs: Open WebUI's image generation and code interpreter with Gemini

[owndev]

Why

#171 proposes moving Gemini image generation into a separate Tool, and discussion #80 asks whether Open WebUI's code interpreter works with the Gemini manifold. Open WebUI 0.11 has both on its own:

• a Gemini image engine (Imagen via predict, Gemini image models via generateContent, plus a Gemini edit engine);
• the code interpreter (pyodide in the browser, or Jupyter).

In Native mode it offers them as the built-in tools generate_image, edit_image and execute_code. Since pipeline 1.18.0, these reach Gemini like any other tool. This PR verifies that end to end and documents it. No pipeline change.

Change

• E2E: new gemini group owuitools (10 checks, suites/_gemini_owuitools.py). Each check proves the mock answered (mock_ok=True).

  Check	What it shows
  declared	All three tools are declared with the toggles on, none with the toggles off; the Pyodide note is added.
  declared-noedit	With image editing off, only generate_image is declared.
  api	API clients get no built-in tools.
  generate / generate-predict	generate_image round trip with a Gemini image model and with Imagen. The engine request is recorded, the image is saved with the answer, and its URL goes back to Gemini in the continuation (function call / response pairing, thought signature).
  edit / edit-followup	edit_image on an uploaded image and on an image generated in the previous turn.
  image-model	A pipe image model gets no tools and makes the image itself.
  code / code-jupyter	execute_code with pyodide (the test browser answers execute:python, and a printed PNG is uploaded and linked) and with Jupyter (Jupyter mock, no browser call).

  Against google_gemini.py 1.17.0, 6 of the 10 checks fail. That version ran the tools inside the pipe through automatic function calling, and the follow-up request broke the function call / response pairing.

• Mocks / harness: :predict in mock_gemini.py, a Jupyter mock in mock_tools.py, and python_answer / python_calls for execute:python in the browser session. The group restores the image and code execution settings at the end.

• Docs: new section in docs/google-gemini-integration.md → "Open WebUI's image generation and code interpreter":

  • setup with environment variables;
  • when Open WebUI offers the tools;
  • limits: the built-in engine uses a Gemini API key only (no Vertex AI login), and the pipeline's 🎨 image models stay for direct use.

  docs/testing.md covers the group, the layout and the counts.

Verification

Run	Result
All suites, v0.11.4-slim, --strict-known (on main ecac3fb)	541 PASS, 0 FAIL, 0 KNOWN (gemini 159, azure 198, n8n 51, infomaniak 32, filters 101), 1138 s
gemini, v0.11.3-slim and v0.11.4-slim, --strict-known	159 PASS, 0 FAIL, 0 KNOWN (each)
Meta-test (E2E_MOCK_FAULT=500)	57 PASS, 380 FAIL, 0 KNOWN; all 10 owuitools checks fail there, none is hidden
ruff 0.11.10 check + format	clean

Not tested: whether the real Gemini/Imagen APIs accept the requests that Open WebUI's image engine sends. Those requests come from Open WebUI, not from this pipeline.

Refs #171, #80

🤖 Generated with Claude Code