Change8

v1.86.0-rc.1

📦 litellmView on GitHub →
13 features🐛 12 fixes🔧 3 symbols

Summary

This release introduces Docker image signing via cosign for enhanced security and adds significant features like Weighted-Routing Failover and componentized service architecture. Numerous bug fixes address validation, rate limiting, and provider integrations across Vertex AI, Bedrock, and MCP.

Migration Steps

  1. Users running LiteLLM Docker images should verify image signatures using the provided cosign commands to ensure authenticity.

✨ New Features

  • All LiteLLM Docker images are now signed with cosign for enhanced security.
  • Added tool-calling support to LassoGuardrail.
  • Optimized chat completions fast path.
  • Emit native web_search_tool_result blocks for Anthropic clients (Claude Desktop / Cowork citations).
  • Exposed delegate_auth_to_upstream in MCP server list API.
  • Fixed vector store retrieve/list/update/delete operations in the proxy when no model is specified.
  • Added Weighted-Routing Failover capability.
  • Support for OTEL GenAI semconv.
  • Componentized gateway, ui-backend, and ui as separate services.
  • Added OTel-standard attributes (status code, route/path, preprocessing latency) on the proxy SERVER span.
  • Added Azure AI Foundry GPT-5.4 model metadata to the model catalog.
  • Set http.response.status_code on the success SERVER span for OTel tracing.
  • Added AWS ECS + GCP Cloud Run deployment stacks for the componentized gateway.

🐛 Bug Fixes

  • Tightened budget field validation and authorization checks.
  • Improved CrowdStrike AIDR input handling.
  • Stopped v3 rate limiter from leaking internal stash to provider body.
  • Allowed for allowlisted redirect URIs.
  • Fixed conversion of raw output_file_id to managed ID in CheckBatchCost poller for managed batches.
  • Fixed zero cost/usage reporting on completed Vertex AI batch jobs.
  • Delegated PKCE bypass for internal MCP servers.
  • Fixed Bedrock Mantle integration to use /anthropic/v1/messages path.
  • Ensured get_secret is imported at runtime.
  • Made /config/update env-var encryption idempotent in the proxy.
  • Sorted BYOK models by their displayed name in /v2/model/info endpoint.
  • Added PROXY_BASE_URL escape hatch and diagnostic logging for 'invalid_request' errors in MCP OAuth.

Affected Symbols