v1.100.0-dev.1
Breaking Changes📦 litellmView on GitHub →
⚠ 1 breaking✨ 4 features🐛 32 fixes🔧 14 symbols
Summary
This release introduces Docker image signing with cosign for enhanced security and includes numerous bug fixes across various integrations and the UI. Key features include per-team New Relic trace routing and improved Azure AI Foundry authentication.
⚠️ Breaking Changes
- The `prompt_token_calculator` utility has been deleted. If your code relied on this function, you will need to find an alternative way to calculate prompt tokens.
Migration Steps
- If you were using the `prompt_token_calculator` function, you will need to replace it with an alternative method for calculating prompt tokens.
✨ New Features
- Added support for per-team New Relic trace routing via team callbacks.
- Introduced security controls for vector-store uploads on the /v1/rag/ingest endpoint.
- Added a Gemini Family auto-router preset to the UI.
- Enabled Entra ID / OAuth authentication on all Azure AI Foundry routes.
🐛 Bug Fixes
- Resolved an issue where the conversation was not being kept when chaining `previous_response_id` on the bridge.
- Fixed an issue with New Relic trace routing.
- Ensured the migration DDL guard runs correctly and stops reading comments as SQL.
- Corrected the display of team and organization TPM/RPM limits of 0 to show as 0 instead of 'Unlimited'.
- Fixed the UI to theme the created-key box for dark mode.
- Restored the public model name tooltip layout in the add model flow.
- Repointed the key detail URL to the rotated hash after regenerating keys.
- Made playground chat bubbles theme-aware.
- Corrected the skill install command and marketplace setup UX.
- Skipped health checks for strategy routers.
- Fixed Databricks to bill cached tokens at cache rates and added missing Claude pricing.
- Ensured Anthropic models round-trip thinking blocks to OpenAI backends on /v1/messages.
- Normalized agent card protocolBinding casing before transport match for A2A.
- Tracked cost and spend for Google Interactions API requests.
- Stopped emitting an empty assistant delta after the finish_reason chunk for Bedrock.
- Reconciled Anthropic enum with declared type in output_format schema.
- Fixed the UI to boot the UI image as an arbitrary uid by anchoring nginx writes under /tmp.
- Correctly parsed form-encoded video edit/extension bodies after authentication for the proxy.
- Kept legacy thinking budget_tokens on Claude 4.6 models on /v1/messages for Anthropic.
- Ensured the prompt_token_calculator counts Claude models again.
- Kept every value of a repeated form key and gated the tests that hid it for the proxy.
- Applied model_info.health_check_params to health check probes.
- Routed all generation endpoints and fixed video cost tracking for RunwayML.
- Attributed spend and released budget reservation on router-model /v1/llm and /azure routes for passthrough.
- Matched OpenAI SDK wire format on image/video routes.
- Carried queued tasks across event-loop changes instead of dropping them for the logging_worker.
- Pinned require_managed_files enforcement behind a marker-gated stack phase for e2e tests.
- Ensured configured auto-routers are listed in the usage picker before they have traffic.
- Kept both ends of a clipped classifier context turn for the complexity router.
- Ignored recursive form-field flatteners in recursive_detector.
- Stopped leaking the caller's virtual key on credential-less Vertex passthrough.
- Stopped copying forwarded credentials into retry breadcrumbs for the router.