●Change8

v1.96.0-rc.1

Breaking Changes
📦 litellmView on GitHub →
⚠ 1 breaking✨ 12 features🐛 30 fixes🔧 38 symbols

Summary

This release introduces Docker image signature verification using cosign and includes numerous bug fixes and feature enhancements across the proxy, UI, and MCP components. Key updates involve improved security, better management endpoints, and enhanced routing capabilities.

⚠️ Breaking Changes

  • The `feat(proxy)!: gate all mock testing request params behind a single config flag` indicates a breaking change. Users may need to update their configuration to enable mock testing requests, as this functionality is now gated by a single flag.

Migration Steps

  1. Update `aiohttp` dependency to at least `3.14.2` to resolve pooled-connection timeouts.
  2. Users enabling mock testing requests may need to update their configuration to use the new single config flag.

✨ New Features

  • Verify Docker image signatures using cosign with pinned commit hash or release tag.
  • Expose classifier context window fields on Auto-Router screens.
  • Add generic list handler for /management/v1.
  • Source the ID-JAG subject from the user's stored SSO assertion.
  • Add GET /management/v1/budgets.
  • Show which log rows are the auto-router's own classifier calls.
  • Add sorting, filtering and search to the budgets page.
  • Add global exclude_metrics and exclude_labels options for Prometheus.
  • Allow AI API keys to read /model/info.
  • Let the classifier see assistant turns and rate what a short reply approves.
  • Push config sync to pods via redis pub/sub.
  • Add pod-hardening and migration-Job knobs to the componentized chart.

🐛 Bug Fixes

  • Annotate connected-app reachability on the gateway connect page.
  • Recover the tool-name prefix boundary from registered prefixes.
  • Exclude skipped tests from coverage-registry numerator.
  • Deny MCP access when a named entitlement cannot be read.
  • Dispose recycled client sessions deterministically.
  • Run bedrock guardrail on MCP tool calls in during_mcp_call mode.
  • Drop unsupported prisma select kwarg from team lookup.
  • Do not send duplicate thoughtSignature copies to Gemini.
  • Nest source object in Claude Code marketplace settings snippet.
  • Exempt values frozen in place by tuple/frozenset/MappingProxyType from LIT002.
  • Keep the session view open when selecting a log inside it.
  • Enforce tool entitlements on every MCP tool dispatch path.
  • Raise aiohttp floor to 3.14.2 to clear pooled-connection timeouts.
  • Drop unsupported prisma select kwarg from key lookup.
  • Stop model writes 500ing on another pod's delete.
  • Show pass through route selections and match team id substrings in team search.
  • Align team member add with existing user provisioning rules.
  • Keep config-defined agents registered and accept the documented `agents:` key.
  • Run unit tests on pushes to main and litellm_internal_staging.
  • Enforce token limits when the pre-call increment is zero.
  • Render pod-level securityContext on the migration Job.
  • Honor request-level num_retries over a deployment's litellm_params value.
  • Bound each spend-log write statement by payload bytes.
  • Bound DB statement and lock time via general_settings.
  • Cache AssumeRole credentials per attributed identity.
  • Derive the throughput SLO per replica and surface locust's error breakdown.
  • Bake prisma offline in the componentized migrations image.
  • Encode public model group on background-created output file ids.
  • Honor USE_DDTRACE in the gateway and backend deployments.
  • Give gateway and backend pro...

Affected Symbols

cosignghcr.io/berriai/litellmlitellm/litellm:v1.96.0-rc.1aiohttpprismaGeminiClaude Code marketplacetuplefrozensetMappingProxyTypeLIT002Datadog MCP tool-call teststhroughput SLO load testBedrock guardrailTeam lookupKey lookupModel writesTeam endpointsConfig-defined agentsPrisma offlineGateway deploymentsBackend deploymentsMigration JobPod-hardeningPrometheusRedis pub/subAuto-RouterGateway connect pageRegistered prefixesCoverage-registrySSO assertionClient sessionsMCP tool dispatch pathToken limitsSecurityContextSpend-log write statementAssumeRole credentialsLocust