●Change8

v1.84.0-rc.1

📦 litellmView on GitHub →
✨ 11 features🐛 26 fixes🔧 26 symbols

Summary

This release introduces Docker image signature verification via cosign and adds several new features, including support for gpt-image-2 and AIHubMix provider. Numerous bug fixes address issues related to caching, logging, Vertex AI, and proxy stability.

Migration Steps

  1. If using the /ui/chat page, note that it has been removed and functionality may need migration.
  2. If relying on specific behavior related to model IDs, note that the deprecated Bedrock Claude 3.7 Sonnet model ID was replaced.

✨ New Features

  • Added support for Docker image signature verification using cosign.
  • Added --timeout_worker_healthcheck flag for uvicorn worker triage in LiteLLM Proxy.
  • Added cleanup job for expired LiteLLM dashboard session keys.
  • Added retry settings for generic API logger.
  • Added support for gpt-image-2 model.
  • Added AIHubMix as an OpenAI-compatible provider.
  • Added support for Azure Entra discovery endpoint.
  • Added durable agent workflow run tracking via /v1/workflows/runs endpoint.
  • Added LiteLLM headers on Google native generateContent routes.
  • Added support for Bedrock retrievalConfiguration passthrough in vector stores.
  • Added opt-in short-ID tool prefix for MCP tool names to keep them under the 60-character limit.

🐛 Bug Fixes

  • Cached GCP IAM token in Redis to prevent async event loop blocking.
  • Fixed NOMA v2 deepcopy crashing in build scan payload.
  • Used stored-credentials endpoint for tools fetch on MCP edit page.
  • Supported CircleCI rerun failed tests for local_testing jobs.
  • Updated GPT-5.5 Pro Pricing.
  • Preserved prompt_tokens_details through embedding cache round-trip.
  • Fixed backfill of streaming hidden response cost.
  • Reused Anthropic messages config instances for Vertex AI.
  • Fixed Vertex AI to preserve items on array branches in anyOf with null and de-flaked a test.
  • Replaced deprecated Bedrock Claude 3.7 Sonnet model ID.
  • Cached LiteLLM_Config param reads in DualCache and batch.
  • Unified cost calculation in success_handler dict and typed branches.
  • Fixed ERROR log assertion scope to LiteLLM logger in test_model_alias_map.
  • Translated OpenAI file content on tool-result path for Bedrock and Anthropic.
  • Added optional TCP SO_KEEPALIVE support to aiohttp's TCPConnector.
  • Self-healed Prisma read paths and hardened reconnect state machine in Proxy.
  • Redacted spend logs error message.
  • Reconnected Prisma DB without blocking the event loop in Proxy.
  • Dropped sensitive locals from re-raised error messages.
  • Set is_redirect=False on mocked retrieve response in Vertex batches tests.
  • Redacted credentials in vector store list/info/update responses; gated update by per-store access.
  • Substituted alias for master key on UserAPIKeyAuth.
  • Tightened public-route detection and OAuth2 fallback gating in MCP.
  • Fixed team member null budget fallback.
  • Inherited caller identity in passthrough batch managed-object in Proxy.
  • Tightened guardrail modification permission check in Proxy/Auth.

Affected Symbols

redisnoma v2uvicorn worker triageCircleCI local_testing jobsGPT-5.5 Proembedding cachegeneric API loggerVertex AIBedrock Claude 3.7 SonnetDualCachesuccess_handlertest_model_alias_mapBedrockAnthropicaiohttp's TCPConnectorPrismaspend logsAzure entra discovery endpoint/v1/workflows/runsGoogle native generateContent routesBedrock retrievalConfigurationMCP tool namesUserAPIKeyAuthpublic-route detectionOAuth2 fallback gatingmanaged-object