●Change8

v1.94.0-dev.2

📦 litellmView on GitHub →
✨ 10 features🐛 16 fixes🔧 25 symbols

Summary

This release introduces enhanced Docker image signature verification using cosign and includes numerous bug fixes and new features across various components like proxy, guardrails, and UI. It also updates dependency constraints for Python and ddtrace.

Migration Steps

  1. Verify Docker image signatures using cosign. Recommended method is to use the pinned commit hash for cryptographic immutability.
  2. Update `ddtrace` to the 4.x line.

✨ New Features

  • Added streaming text transformation in generic_guardrail_api.
  • Added resource-less InvokeGuardrailChecks (detect-only) mode for Bedrock guardrails.
  • Added ability to resolve auto-router routing plugins from proxy YAML config.
  • Added GPT-5.6 sol/terra/luna to the Bedrock model cost map.
  • Moved Caching out of Experimental into Developer Tools in the UI.
  • Added Compresr guardrail for query-aware context compression.
  • Forwarded optional metadata on POST /guardrails/apply_guardrail.
  • Added reusable BetaBadge to the UI and used it for the Projects sidebar item.
  • Show exact license expiration date in usage cards in the UI.
  • Added passthrough matrix row for the big-3 clouds and Anthropic API in e2e/claude_code tests.

🐛 Bug Fixes

  • Routed master key to team-scoped models in the proxy.
  • Routed xai.grok-4.3 via the /openai/v1 frontier path for Bedrock.
  • Intercepted web search on the Responses API.
  • Dropped empty content_block_delta events for Anthropic adapter.
  • Persisted discovered OAuth endpoints and kept the last known good on failed re-discovery for MCP.
  • Surfaced actionable CLI SSO errors when CLI and proxy versions skew.
  • Capped per-user OAuth token cache TTL at the token's own lifetime for MCP.
  • Preserved callback order in get_combined_callback_list.
  • Redacted assistant tool call arguments in spend logs.
  • Honored messages request timeout for Anthropic.
  • Applied sanitized prompt returned by moderation API to the request for llm_guard.
  • Stopped pinning large request payloads past request end.
  • Discovered missing OAuth scopes and token_url when authorization_url is set manually for MCP.
  • Returned empty dict from _classifier_call_metadata when metadata is absent for complexity_router.
  • Resolved chat routes at render time so navigation works under server_root_path in the UI.
  • Enforced minimum custom key length and masked short keys in key_name for key management.

Affected Symbols

generic_guardrail_apiInvokeGuardrailChecksauto-routerBedrock model cost mapCachingCompresr guardrailPOST /guardrails/apply_guardrailBetaBadgeProjects sidebar itemUsage cardse2e/claude_code testsproxyBedrockwebsearchResponses APIanthropic-adapterMCPCLI SSOget_combined_callback_listspend logsanthropicllm_guardcomplexity_routerUIkey management