- Shipped
- July 18, 2026 at 7:04 PM UTC
- Author
- Kamo
- Commit
- 027a0dc
Found by probing production: org 1188756624575266822's vision calls resolved to gpt-4o-mini-tts-2025-03-20 — a TEXT-TO-SPEECH model — so every vision call returned 502. Document extraction and any other vision feature have been failing for that tenant. Two independent causes, both fixed: gpt-4o-mini-tts, gpt-4o-mini-transcribe and gpt-4o-mini-search-preview off one prefix, so speech models were recorded as vision-capable. EXCLUDED_PREFIXES did not catch them because it matches startsWith and these are suffixes. ModelModality checks hyphen- delimited modality TOKENS instead, and is a deny-list so a newly-released chat model keeps working. 2. Routing took findFirst() on an unordered query, so which vision model an org got was arbitrary — and it trusted supportsVision, which can be wrong (auto-discovered, and member-editable in the UI). It now rejects non-chat modalities outright and picks deterministically, so the same org resolves to the same model every call. Applied to the chat and text fallbacks too: those also did models.get(0), which could hand a conversation to an embedding or speech model. Also logs the upstream cause on a vision/text failure. It was discarded, so a 502 said nothing about WHY — a wrong model, a rejected key and a network fault looked identical. That cost real time diagnosing this. Existing rows carrying the bad flag still need correcting; discovery no longer writes it.