Repository navigation
2.2.3 changelog - #7528
2.2.3 changelog#7528
Conversation
|
Navigate logical layers of code changes, visualize relationships, and explore their blast radius. ℹ️ Recent review info⚙️ Run configurationConfiguration used: Repository: maximhq/bifrost/.coderabbit.yaml Review profile: CHILL Plan: Team Run ID: 📒 Files selected for processing (30)
Included review availability: Your plan provides up to 10 included reviews per hour; 7 remain after this review. 📝 SummarySummary by CodeRabbit
WalkthroughThis release updates changelogs and version values for core, framework, plugins, and transports. The transport changelog lists routing, provider, Responses, and telemetry changes, and states that no database migrations were added. ChangesRelease updates
Priority: ⬇️ Low Estimated code review effort: 2 (Simple) | ~8 minutes Suggested reviewers: 🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |

Summary
Releases core v1.10.2, framework v1.7.4, transports v2.2.3, and corresponding plugin patch versions. This release focuses on routing reliability, OpenAI Responses API correctness, and several provider-specific bug fixes.
Changes
key_idorprovider_key_nameto pin a particular provider key. The framework'sRoutingFallbacktype accepts either the legacyprovider/modelstring or the new object form. Unpinned fallbacks are unchanged, preserving existing config hashes.asyncflag,output_schema, andtunnel_idare forwarded on Responses tool calls.asyncis stripped for models that do not support it via thesupports_async_toolsdatasheet override. Structured MCP tool-call errors, object-formconversation, array-formallowed_tools,approval_request_id, andin/ninfile search filters now round-trip correctly.supports_prompt_cache_breakpoint.reasoning.effort: "none"is forwarded forgpt-6-solandgpt-6-luna.temperature,top_logprobs, andlogprobsare stripped alongsidetop_pwhen the model and effort do not support them. An omittedreasoning.effortcounts asnoneonly for models that default to no reasoning.cache_controlbreakpoints are preserved for Anthropic models routed through OpenRouter.tool_choice: "auto"for gpt-oss models and recovers leaked parameter tags with surrounding whitespace.enablednow preserves the stored value instead of writing NULL.user_labels_enabledis now persisted with the telemetry config.conversationare read for cache metadata.Type of change
Affected areas
How to test
go test ./...key_idpin route to the correct provider key.async,output_schema, andtunnel_idare forwarded on Responses tool calls to OpenAI and stripped for unsupported models.cache_controlbreakpoints are preserved for Anthropic models via OpenRouter.temperature,top_logprobs, andlogprobsare stripped for OpenAI reasoning models when effort does not support them.enableddoes not overwrite the stored value.Breaking changes
Go SDK callers using
ResponsesMCPApprovalResponse.ApprovalResponseIDmust rename it toApprovalRequestID. The message type is nowmcp_approval_response.ResponsesToolMessage.ErrorandResponsesParameters.Conversationare now union types instead of*string.Related issues
#7470, #7242, #7240, #7492, #7239, #7241, #7521, #7468, #7473, #7461, #7490, #7379, #7380, #7381
Security considerations
None beyond the existing provider credential handling. Pinned key IDs are resolved through the existing key-access control path.
Checklist
docs/contributing/README.mdand followed the guidelines