fix: Open WebUI 0.12.0 compatibility (Azure 3.0.1, Gemini 1.19.1, e2e)

[owndev]

Why

Open WebUI v0.12.0 was released on 2026-10-10. On main, all suites passed on 0.11.3 and 0.11.4, but 0.12.0 gave 542 PASS / 5 FAIL. Two of the causes are in our functions; the rest are test assumptions that 0.12 changed. One further 0.12 incompatibility turned up while reading Open WebUI's diff.

Fixes in the functions

• Azure AI Foundry 3.0.0 → 3.0.1:
  • Cause: Open WebUI 0.12 renamed the metadata key that holds the typed prompt before its RAG template is added: user_prompt became base_user_prompt (middleware.py:3333; 0.11.4: :3231). The pipeline read only the old key.
  • Effect on 0.12: with attached files or chats, knowledge, or web search, Azure AI Search searched with the template text ("### Task: Respond to the user query…") instead of what the user typed.
  • Fix: the pipeline now reads base_user_prompt first and falls back to user_prompt. Nothing changes on 0.11.
  • New check azure.rag.query-text.metadata covers four cases: only the new key, only the old key, both, and a None new key.
• Google Gemini 1.19.0 → 1.19.1:
  • Cause: when Open WebUI 0.12 rebuilds a saved chat, it drops every assistant message without content and output. Before, it dropped only failed ones (middleware.py:2277; 0.11.4: :2241).
  • Effect on 0.12: after an empty answer (for example, stopped before its first token), the next image-edit turn was taken for a guided regeneration. The edit message's upload was then also read as history, and with IMAGE_DEDUP_HISTORY=false it was sent twice.
  • Fix: request and saved chat now skip empty answers by the same rule.
  • New check gemini.imgedit.empty-answer: it fails with 1.19.0 on 0.12.0 and passes with 1.19.1 on 0.12.0 and 0.11.4.

Test changes

• Module names in the log:
  • Open WebUI 0.12 loads each function as function_<id>_<uuid hex> (utils/plugin.py:277). Expected-error signatures such as ("function_azure:pipe", …) therefore stopped matching. That caused the azure, n8n and infomaniak server-log failures.
  • One azure "no traceback" check passed on 0.12 without testing anything, because it no longer found its error blocks.
  • harness/logs.py now normalizes the names to function_<id> everywhere it reads the log.
• gemini.imgedit.admin: Open WebUI 0.12 deliberately no longer lets an admin continue another user's chat. It answers 404, because writing needs write access to the chat (main.py:1554, models/chats.py:1903-1945). The check is version-gated: on 0.12 and later it expects the 404 and no upstream request. The admin reading another user's file stays covered by imgedit.api-foreign.
• Docs: docs/testing.md lists 0.12.0, 0.11.4 and 0.11.3 as tested, with 549 checks. CLAUDE.md notes both 0.12 pitfalls. DEFAULT_IMAGE stays v0.11.4-slim until the first 0.12 patch release.

Verification

Run	Result
All suites, --strict-known, v0.12.0-slim	549 PASS, 0 FAIL, 0 KNOWN (1046 s)
All suites, --strict-known, v0.11.4-slim	549 PASS, 0 FAIL, 0 KNOWN
All suites, --strict-known, v0.11.3-slim	549 PASS, 0 FAIL, 0 KNOWN
Meta-test (E2E_MOCK_FAULT=500)	57 PASS, 388 FAIL, 0 KNOWN
ruff 0.11.10 check + format	clean

Other 0.12 changes, not addressed here

• Shared chats with "allow replies" (new in 0.12): Gemini's history loading accepts only the chat owner or an admin, which is stricter than Open WebUI's own access rules. A participant who replies doesn't get the earlier generated images sent along. This touches access control and is left for a separate decision.
• Cosmetic:
  • Other participants of a shared chat see streamed text only at the end.
  • time_token_tracker now counts the <attached_files> tags, because 0.12 adds them before inlet filters.
• Valve types are now validated when function code is saved. Changing a valve's type in a later version would block saving the update.

🤖 Generated with Claude Code