v1.91.0-rc.1
📦 litellmView on GitHub →
✨ 7 features🐛 19 fixes🔧 6 symbols
Summary
This release introduces Docker image signing verification via cosign and significant improvements to MCP OAuth handling and Prometheus metrics. Numerous bug fixes address issues across guardrails, Bedrock model listing, logging, and proxy performance.
Migration Steps
- Users running LiteLLM via Docker are encouraged to verify image signatures using the provided cosign commands to ensure image integrity.
✨ New Features
- Added support for verifying LiteLLM Docker image signatures using cosign.
- Added requested_model label to Prometheus spend and requests metrics.
- Implemented shared OAuth token foundation for MCP, including challenge, seam storage, expiry-aware cache, and single-flight refresh.
- Added headroom guardrail for message compression.
- Migrated authorization_code MCP to the v2 resolver (single-replica) [1/2].
- Implemented cross-replica single-flight refresh for the v2 per-user OAuth store [2/2].
- Added per-team litellm_team_members_metric gauge for Prometheus.
🐛 Bug Fixes
- Fixed guardrails to match policy-pipeline block response to direct guardrail attachment.
- Stopped listing bedrock_mantle models under the Bedrock provider in the UI.
- Stopped logging tool-call input in the MCP client.
- Scheduled spend logging via the durable logging worker for passthrough.
- Wrapped agentic loop response in a fake stream for streaming websearch requests.
- Synchronized tool_choice when converting web_search tools.
- Forwarded all multipart files with repeated field names in passthrough.
- Preserved Anthropic server_tool_use web search usage in cost tracking.
- Pointed AgentOps OTLP exporter at otlp.agentops.ai.
- Stopped the pass-through route registry from growing on every reload (PERF-13).
- Normalized Messages system role and adaptive-thinking for Claude Invoke on Bedrock.
- Bound event-loop blocking from oversized requests in the proxy.
- Cancelled upstream LLM stream when client disconnects during time-to-first-token in the proxy.
- Logged pre-call guardrail blocks at WARNING instead of ERROR with a traceback.
- Honored user_api_key_cache_ttl for management-object cache writes in proxy/auth.
- Fixed loop-scope async Lua script registration in Redis.
- Appended rawPredict suffix for custom api_base on /v1/models endpoint for vertex_ai.
- Showed an agent's attached virtual key in the UI.
- Persisted global retry_policy via /config/update.