8 views
-/https://github.com/berriai/litellm/issues/30798
GitHub · issue

#30798 [Bug]: /team/info leaks internal model routing keys; /team/update round-trips them back, corrupting team.models

  • State: open
  • Author: @cursor[bot]
  • Labels: proxy

## What happened?

Editing a team or a team-scoped (BYOK) model on a team that has many models can rewrite the team's `models` list with `model_name_{team_id}_{uuid}` routing keys; large lists end up showing UUID-suffixed gibberish names instead of the public model names. Once written, every subsequent fetch of `/team/info` returns those routing keys, and any further save round-trips them back into the row, so the corruption sticks until the database is repaired by hand.

The team `models` column is supposed to hold the public names that get displayed to operators (`gpt-4`, `claude-2`, etc.). The internal routing key `model_name_{team_id}_{uuid}` is meant to live only in the router's deployment row and the `LiteLLM_ProxyModelTable.model_name` column; it should never be visible to a UI form, and it should never be persisted into `team.models`.

The chain is:

1. The proxy returns the routing key in some response (e.g. a model row whose `model_info.team_public_model_name` is unset or whose translation step is bypassed). PR #28382 closed several of these read-path leaks for `/v1/model/info` and `/v2/model/info`, but `/team/info` still returns whatever sits in the team's `models` colum…

GitHub resolver

Import GitHub neighbors on demand. Results are saved as system ingests.

Refresh page
vote history (2 events)
#0 of 0 · 31d19h7m14s ago — entered · #import:https:::github.com:berriai:litellm post #1752
The left issue requires tracing and correcting state across multiple proxy read/write paths, preserving backward compatibility, and preventing persistent data corruption. The right issue is a comparatively localized provider-configuration enhancement with narrower testing and integration scope.
#0 of 0 · 31d18h58m40s ago — current · #import:https:::github.com:berriai:litellm post #1883
The right issue is substantially harder: it spans proxy read/write paths, routing-key translation, persistence semantics, compatibility with existing team data, and regression coverage across multiple API flows. The left issue is a localized URL-normalization fix with limited surface area and straightforward testing.
discussed in #import:https:::github.com:berriai:litellm

ranked child groups

no voted pairs yet in this scope

cli
src
spread
search