●Change8

v1.93.0-dev.1

📦 litellmView on GitHub →
✨ 9 features🐛 28 fixes🔧 32 symbols

Summary

This release introduces new features like entra_obo token exchange profile and enhanced Docker image signing verification via cosign. Numerous bug fixes address issues across streaming, guardrails, MCP operations, and spend reporting.

Migration Steps

  1. If verifying Docker images, use the pinned commit hash for the strongest verification: `cosign verify --key https://raw.githubusercontent.com/BerriAI/litellm/0112e53046018d726492c814b3644b7d376029d0/cosign.pub ghcr.io/berriai/litellm:vX.Y.Z`.

✨ New Features

  • Added entra_obo profile to the token_exchange (OBO) arm.
  • Added separate ITPM/OTPM deployment rate limits for the router.
  • Added cost optimization feedback banner to models page in the UI.
  • Support user-defined volumes and volumeMounts in the Helm microservices chart.
  • Added custom_technical_keywords configuration for complexity_router.
  • Added support for JWT team memberships as a fallback when JWT has no team claims.
  • Added Google Cloud Speech-to-Text Chirp 3 transcription support for vertex_ai.
  • Added Azure data-zone and long-context pricing for gpt-5.4/5.5.
  • Added experimental dashboard pages flagging on the draft deprecation list in the UI.

🐛 Bug Fixes

  • Headroom guardrail now logs real token/compression stats instead of "allow".
  • Forward provider response headers on streaming /v1/messages responses for anthropic_messages.
  • Emit SSE error event when Bedrock Messages stream ends without message_stop.
  • Register batch + rust OCR deployments via /model/new.
  • Merge websearch tool parameters.
  • Surface in-body error payloads on OpenAI-compatible streams.
  • Reflect persisted "Store Prompts in Spend Logs" toggle state on UI load.
  • Filter /global/spend/report by team_id when group_by=team.
  • Send only new messages since last assistant turn to CrowdStrike AIDR guardrails.
  • Resolve os.environ/ refs for all AWS auth params in DB-sourced models in the proxy.
  • Forward verbosity param to chat completion providers.
  • Stop CacheCodec from dropping null fields on cache round-trip.
  • Hash Bearer-prefixed API keys in spend logs for security.
  • Build responses input_items url with path before query string for Azure.
  • Decrypt response ids for input_items follow-ups for responses_id_security.
  • Map upstream 4xx on cancel to client error instead of 500 for responses.
  • Bump wolfi-base digest for glibc 2.43-r10 in Docker.
  • Stamp oauth2_flow=authorization_code when persisting a DCR client registration in MCP.
  • Classify allm_passthrough_route as async to prevent duplicate success callbacks in logging.
  • Forward short OAuth state upstream, keep session in a cookie for MCP.
  • Honor AWS auth params in realtime handler for Bedrock.
  • Scan MCP tool calls for pre_mcp_call / during_mcp_call modes in model_armor.
  • Make test_priority_reservation deterministic across minute rollover for dynamic_rate_limiter.
  • Startup backfill stamping oauth2_flow on legacy null rows in MCP.
  • Stamp completion_start_time on first chunk for /v1/messages and /v1/responses streaming endpoints.
  • Keep tool names whole in filter response header for mcp_semantic_filter.
  • Price anthropic passthrough message batches correctly in batch cost job.
  • Trigger gateway fallbacks on local rate limiting in proxy.

Affected Symbols

token_exchange/v1/messagesbedrock/model/newwebsearch toollangfuse callbacks-in-db e2e testrouterOpenAI-compatible streamsspend logs toggle/global/spend/reportCrowdStrike AIDR guardrailshelm chartAWS auth paramschat completion providersCacheCodeccomplexity_routerBearer-prefixed API keysazure responses input_items urlresponse idsupstream 4xx on cancelwolfi-baseDCR client registrationallm_passthrough_routeOAuth staterealtime handlermodel_armordynamic_rate_limitervertex_aispend read surfacemcp_semantic_filteranthropic passthrough message batchesgateway fallbacks