## Is your feature request related to a problem? Please describe.

Yes. We run a multi-provider LiteLLM proxy (Mistral, DeepSeek, OpenAI, OpenRouter) behind an OpenWebUI frontend. Models are heterogeneous in their error returns: different body structure, different status codes, and some leak sensitive information (reasoning text, moderated content, internal details) in the error body that surfaces verbatim in the user-facing UI.

Concrete examples we hit in production:

- **Mistral 400 `extra_forbidden`** on `reasoning_details` — full reasoning text is echoed in the error `input` field (see #30835, #30882). Same root cause keeps resurfacing as per-field `drop_params` patches (#30850, #30884, #30909). - **Rate limiting / quota exhausted** — providers return different structures (Retry-After headers, JSON bodies, mid-stream SSE) and we want to normalize the message the end user sees without leaking internals. - **Content policy violations** — some providers include the flagged content in the error body; we want to strip it before it reaches the frontend. - **Auth / 404 / 5xx** — we want to map provider-specific status codes to consistent, user-friendly messages.

Today there is no cl…