Change8

v1.99.0-dev.2

Breaking Changes
📦 litellmView on GitHub →
1 breaking7 features🐛 24 fixes🔧 21 symbols

Summary

This release introduces security enhancements with signed Docker images and includes numerous bug fixes across proxy, router, and UI components. It also features improvements to spend logging, end-to-end testing, and adds a new CLI command for configuring Claude Code.

⚠️ Breaking Changes

  • feat(auto-router)!: scope shadow eval jobs to multiple keys. This change might affect existing auto-router configurations that relied on a single key scope for shadow evaluations. Users may need to update their configurations to specify multiple keys if applicable.

Migration Steps

  1. Users may need to update their auto-router configurations to specify multiple keys if they relied on a single key scope for shadow evaluations.
  2. Update Docker image verification to use cosign with either the pinned commit hash or release tag as per the provided instructions.

✨ New Features

  • Docker images are now signed with cosign for enhanced security. Verification instructions are provided using both pinned commit hash and release tags.
  • Added lifecycle timestamps to spend logs.
  • Introduced record/replay transport seam and fixture bundle format for end-to-end testing.
  • Added canonical content-based match keys for record-and-replay.
  • Added `lite login --config-claude` CLI command to wire Claude Code at login.
  • Configurable reasoning override floor for the complexity router.
  • Auto-suppress the no-Redis banner for confirmed single-worker deployments.

🐛 Bug Fixes

  • Log spend for OpenAI passthrough embeddings with unmapped models.
  • Keep acreate_file fallbacks inside the requested model group.
  • Record estimated input tokens in spend logs for failed dispatched requests.
  • Accept bool thinking param instead of crashing with AttributeError.
  • Restore the cache control Role and Index field hints.
  • Rebuild nested and list paths in the mounted-field projection.
  • Gate the pass-through guardrail field inputs when the section is disabled.
  • Keep a host PROXY_BASE_URL out of request-derived URL tests.
  • Read through to the DB on registry misses so just-created models, guardrails, and agents resolve on sibling replicas.
  • Forward the per-server auth header on OpenAPI tool calls.
  • Restore tab strip styling and panel persistence lost in the shadcn migration.
  • Migrate the antd Button call sites onto the shadcn Button.
  • Accept provider cost objects when propagating usage cost.
  • Gate the reasoning override on a non-SIMPLE score.
  • Stop reporting failed OpenAPI tool calls as successes.
  • Accept inherited model sentinels in project key limits.
  • Add provider-announced deprecation_date to 205 registry entries.
  • Correct gemini 3.1 flash image and deepseek v4 pricing, add openai deprecation dates.
  • Set prompt_cache_min_tokens=4096 for Gemini 3.5/3.6/3.7 Flash and 3.1 Pro Preview.
  • Report provider thinking tokens instead of classifying them as text for Anthropic and Bedrock.
  • Stop one bad output line from zeroing an entire batch's spend.
  • Resolve bare model names against wildcard deployments in model access groups.
  • Drop stale user search answers so Enter commits the current match.
  • Retry spen

Affected Symbols