v1.95.0-rc.2
📦 litellmView on GitHub →
✨ 13 features🐛 21 fixes🔧 15 symbols
Summary
This release introduces enhanced Docker image signature verification using cosign and includes numerous bug fixes and new features across various components like MCP, UI, and proxy. It also adds new model providers and pricing information.
Migration Steps
- Verify Docker images using cosign with the pinned commit hash (recommended) or the release tag.
✨ New Features
- Add Meta Model API provider and muse-spark-1.1.
- Add Azure GPT-5.6 (sol/terra/luna) pricing and metadata.
- Add dcr_bridge column and plumbing for client-forwarded auth modes.
- Relay upstream 401 on client-forwarded pass-through tool calls.
- Add shared composable DataTable component.
- Dcr_bridge authorize and token relay redirect handling with mandatory S256.
- Dcr_bridge discovery facade and register relay.
- Add sealed envelope module for dcr_bridge client-held credentials.
- Emit the gen_ai.client.operation.exception event on failed LLM calls.
- Configure the coordination redis independently of the response cache.
- Make coordination redis a first-class chart and terraform surface.
- Add dcr_bridge envelope consumer helpers.
- Add LLM-based classifier option to complexity router.
🐛 Bug Fixes
- Lower realtime server-VAD threshold to 0.5.
- Mask credentials embedded in guardrail_response before persist.
- Honor store_prompts_in_spend_logs for guardrail_information.
- Sum multi-round session cost in logs UI.
- Stop persisting the DCR client onto true_passthrough and oauth_delegate server rows.
- Batch credentials wiring and compose harness for live proxy suite.
- Fail closed and surface semantic filter context window errors.
- Keep mid-conversation system messages in place for Claude Invoke.
- Add coordination_redis routes to component allowlist.
- Build redis usage cache from REDIS_* env when cache backend is not Redis.
- Match multi-mode guardrail_mode without false-COMPLIANT.
- Merge credentials within the client-forwarded class on an auth-type switch.
- Only use SSLConnection when ssl is truthy in connection pool kwargs.
- Strip trailing slash from --base-url in lite CLI.
- Add jp.anthropic.claude-opus-4-8 to model cost map.
- Preserve reasoning_tokens through chat->responses usage translation.
- Gate in-place system role messages on model support for Claude Invoke.
- Raise APIError on in-stream error events; widen ErrorEventError.param to accept dict.
- Skip budget metric DB/cache lookups when gauges are NoOpMetric.
- Strip @version suffix in _model_map_lookup_candidates so @default Vertex AI models resolve to adaptive thinking.
- Split log...