●Change8

v1.83.7.rc.1

Breaking Changes
📦 litellmView on GitHub →
⚠ 1 breaking✨ 7 features🐛 14 fixes🔧 5 symbols

Summary

This release introduces Docker image signing verification instructions and includes numerous bug fixes across proxy, S3, logging, and guardrails. It also features a breaking change reducing the default Prometheus latency histogram buckets.

⚠️ Breaking Changes

  • The default Prometheus LATENCY_BUCKETS have been reduced from 35 to 18 boundaries. Users relying on specific 'le' values in existing Prometheus dashboards or PromQL SLO queries (e.g., le="1.5") must update those queries as those series will no longer exist.

Migration Steps

  1. Review and update any Prometheus dashboards or PromQL SLO queries that reference specific latency histogram bucket values ('le' values) due to the reduction in default LATENCY_BUCKETS.

✨ New Features

  • Added support for Anthropic's advisor_20260301 tool type.
  • Implemented optional skip system message functionality in unified guardrail inputs.
  • Added support for skipping dummy user continue for assistant prefix prefill in Bedrock.
  • Introduced Azure routing, managed container IDs, and delete response parsing features for containers.
  • Added file content streaming support for OpenAI and related utilities.
  • Added applyguardrail to inline IAM.
  • Implemented Team member permission for /spend/logs for team-wide spend logs (UI + RBAC).

🐛 Bug Fixes

  • Improved input validation on management endpoints in the proxy.
  • Used parameterized query for combined_view token lookup in the proxy.
  • Harden file path resolution in skill archive extraction.
  • Aligned v1 guardrail and agent list responses with v2 field handling.
  • Flushed Tremor Tooltip timers in user_edit_view tests.
  • Fixed a2a client default 60 second timeout.
  • Ensured spend/cost logging runs when stream=True for websearch_interception.
  • Mapped refusal stop_reason to incomplete status in streaming responses.
  • Preserved proxy key-auth metadata on /v1/messages Langfuse traces.
  • Added retry with exponential backoff for transient S3 503/500 errors.
  • Preserved dict guardrail HTTPException.detail and bedrock context in the proxy.
  • Implemented session-TZ-independent date filtering for spend/error log queries.
  • Aligned Org and Team Endpoint Permission Checks.
  • Fixed proxy to pass-through multipart uploads and Bedrock JSON body.

Affected Symbols

LATENCY_BUCKETS/v1/messagesAnthropicBedrockS3