Change8

v1.95.0-dev.2

📦 litellmView on GitHub →
5 features🐛 22 fixes🔧 19 symbols

Summary

This release introduces enhanced Docker image signature verification using cosign and includes numerous bug fixes and improvements across various components like Bedrock, SCIM, Proxy, and the UI. Several UI components have been migrated to shadcn for a more modern look and feel.

Migration Steps

  1. Verify Docker image signatures using cosign. Recommended method is to use the pinned commit hash for the `--key` argument.

✨ New Features

  • Added `only_scan_new_messages` for per-session incremental scanning in guardrails.
  • UI: Edit fallback chains from router settings.
  • UI: Give each Models + Endpoints tab its own path.
  • Added RESTful PATCH /v2/organization/{organization_id}.
  • MCP: Admit gateway DCR session bearers at the aggregate /mcp scope.

🐛 Bug Fixes

  • Bedrock: Emit Nova Sonic realtime session.created on connect and session.updated on session.update.
  • Proxy: Share CLI SSO login sessions across workers without enable_redis_auth_cache.
  • SCIM: Use members_with_roles as the source of truth for group membership.
  • SCIM: Sync team roster and dedup teams for existing-user email upsert.
  • Proxy: Raise dashboard session budget default to $1 and make it configurable in config and Admin UI.
  • SCIM: Prune deleted user from teams' members_with_roles.
  • Team: Make team member add atomic to prevent concurrent-add member loss.
  • UI: Surface SSO and SMTP settings supplied as process env vars.
  • SCIM: Parse membership id from filtered PATCH path when value omitted.
  • UI: Remove misleading os.environ tooltip from logging settings.
  • Anthropic: Strip uniqueItems + other unsupported array/object constraints from output_format schema.
  • MCP: Use official Google Drive streamable HTTP MCP server.
  • Tests: Remove importlib.reload of http_handler that breaks client injection in later tests.
  • MCP: Fall through to an ephemeral DCR mint when passthrough authorize has no client_id.
  • Responses: Preserve reasoning through prompt hooks.
  • Budget: Resolve word-form budget_duration so it no longer silently resets daily.
  • Bedrock: Include type in tool_choice disable_parallel_tool_use config for Converse.
  • Autoroute: Discover models via /v1/models so an AI-API-only key works.
  • Docker: Bake non_root prisma engines at /opt/prisma so migrations run offline for any uid.
  • Anthropic: Strip all remaining output_format schema keywords rejected by Anthropic.
  • SageMaker: Forward stream events as they arrive to cut TTFT.
  • MCP: Log actionable OAuth discovery failures for misconfigured server urls.

Affected Symbols