v1.5.0-prerelease4

​

✨ Features

• Claude Opus 4.7 — Added compatibility for Anthropic’s Claude Opus 4.7 model, including adaptive thinking, task-budgets beta header, display parameter handling, and “xhigh” effort mapping
• Anthropic Structured Outputs — Added response_format and structured output support for Anthropic models across chat completions and Responses API, with order-preserving merge of additional model request fields (thanks @emirhanmutlu-natuvion!)
• MCP Tool Annotations — Preserve MCP tool annotations (title, readOnly, destructive, idempotent, openWorld) in bidirectional conversion so agents can reason about tool behavior
• Anthropic Server Tools — Expanded Anthropic chat schema and Responses converters to surface server-side tools (web search, code execution, computer use containers) end-to-end
• OCR Request Support — Added OCR request type with stream terminal detection, full body accumulation for passthrough streams, input logging with detail view, and per-request pricing support
• Team Budgets — New team budget system with per-team spending tracking, atomic ratelimit updates, and database structure support
• Single Log Export — Export individual log entries from the logs view and MCP logs sheet
• Deny-by-Default Virtual Keys — Virtual key provider and MCP configs now block all access when empty; automatic migration backfills existing keys to preserve behavior
• User Agent Detection — Improved multi-user-agent detection with tool call reduplication fix for mixed-client environments
• Per-User OAuth Codemode — OAuth server selection and validation per-user in codemode

​

🐞 Fixed

• Provider Queue Shutdown Panic — Eliminated send on closed channel panics in provider queue shutdown by leaving channels open and exiting workers via the done signal; stale producers transparently re-route to new queues during UpdateProvider
• OpenAI Tool Result Output — Flatten array-form tool_result output into a newline-joined string for the Responses API so strict upstreams (Ollama Cloud, openai-go typed models) no longer reject with HTTP 400 (thanks @martingiguere!)
• vLLM Token Usage — Treat delta.content="" the same as nil in streaming so the synthesis chunk retains its finish_reason, restoring token usage attribution in logs and UI
• Gemini Tool Outputs — Handle content block tool outputs in Responses API path for function_call_output messages (thanks @tom-diacono!)
• Bedrock Streaming — Emit message_stop event for Anthropic invoke stream and case-insensitive anthropic-beta header merging (thanks @tefimov!)
• Bedrock Tool Images — Preserve image content blocks in tool results when converting Anthropic Messages to Bedrock Converse API (thanks @Edward-Upton!)
• Gemini Thinking Level — Preserve thinkingLevel parameters across round-trip conversions and correct finish reason mapping
• Anthropic WebSearch — Removed the Claude Code user agent restriction so WebSearch tool arguments flow for all clients
• Responses Streaming Errors — Capture errors mid-stream in the Responses API so transport clients see failures instead of silent termination
• Anthropic Request Fallbacks — Dropped fallback fields from outgoing Anthropic requests to avoid schema validation errors
• Tool Execution Header — Remove redundant static header assignment in tool execution flow
• Virtual Key Configs — Virtual key configurations cleaned up correctly on provider changes; fix key creation and management edge cases
• Virtual Key Management — Fix virtual key creation validation and update handling
• vLLM Extra Params — Extra parameters now properly passed through to vLLM providers
• OAuth Query Params — Preserve existing query parameters when building OAuth upstream authorize URLs
• Streaming Timeouts — Separate streaming clients per provider to prevent read timeout collisions
• Plugin Timer Concurrency — Fix concurrent map access in plugin timer causing potential race conditions
• Async Context Propagation — Preserve context values in async requests so downstream handlers retain request-scoped data
• Custom Providers — Allow custom providers without a list-models endpoint to accept any model rather than restricting on virtual key registration
• OTel Insecure Default — OTel plugin now defaults insecure to true when omitted, enabling HTTP collectors without explicit config; OTel semconv updated to v1.40.0
• Helm mcpClientConfig — Fixed templating for mcpClientConfig (thanks @crust3780!)
• Helm Chart — Refreshed helm chart with validation fixes

• feat: claude-opus-4-7 compatibility with adaptive thinking, task-budgets beta header, display parameter handling, and xhigh effort mapping
• feat: add Anthropic structured output and response_format support across chat completions and Responses API (thanks @emirhanmutlu-natuvion!)
• feat: preserve MCP tool annotations in bidirectional conversion between MCP tools and Bifrost chat tools
• feat: expand Anthropic chat schema and Responses converters to surface server-side tools (web search, code execution, computer use containers)
• feat: add OCR request type support with stream terminal detection and full body accumulation for passthrough streams
• feat: add user agent detection for multiple user agents and fix tool call reduplication
• feat: virtual key provider and MCP configs are now deny-by-default; empty configs block all access
• fix: make provider config weight optional; null weight excludes provider from weighted routing
• fix: use separate streaming clients per provider to prevent read timeouts
• fix: concurrent map access in plugin timer
• fix: extra params passthrough for vllm providers
• fix: remove redundant static header assignment in tool execution
• fix: add OCR request pricing support
• fix: usage of per-user OAuth servers in codemode
• fix: adds validation on direct api keys
• fix: OpenAI provider - flatten array-form tool_result output for Responses API (thanks @martingiguere!)
• fix: Gemini provider - handle content block tool outputs in Responses API path (thanks @tom-diacono!)
• fix: case-insensitive anthropic-beta merge in MergeBetaHeaders
• fix: Bedrock provider - emit message_stop event for Anthropic invoke stream (thanks @tefimov!)
• fix: Bedrock provider - preserve image content in tool results for Converse API (thanks @Edward-Upton!)
• fix: gemini preserves thinkingLevel parameters during round-trip and corrects finish reason mapping
• fix: WebSearch tool argument handling for all clients by removing Claude Code user agent restriction
• fix: capture responses streaming API errors
• fix: delete fallbacks from outgoing Anthropic requests
• fix: token usage for vllm streaming (treat delta.content="" same as nil)
• fix: provider queue shutdown panic - eliminated send on closed channel by leaving channels open and exiting via done signal

• feat: add team budget system with per-team spending tracking, DB tables, and atomic ratelimit integration
• feat: add access profile filter to exclude access-profile-managed virtual keys
• feat: add OCR input logging and pricing metadata support
• feat: add tiered and priority/flex pricing support — 272k token tier fields, 200k priority variants, and tier selection based on service_tier in responses
• fix: use separate DB connection for migrations to prevent connection conflicts
• fix: preserve existing query params in OAuth upstream authorize URL
• fix: clean up virtual key configs when provider changes
• fix: calendar_aligned propagation in v1.5.0-prerelease4 migration
• fix: virtual key creation and management handling
• fix: preserve context values in async requests
• fix: capture responses streaming API errors
• fix: allow custom providers without a list models endpoint to register any model
• fix: don’t mark OAuth config expired on transient refresh failures
• fix: only treat invalid_grant and unauthorized_client as permanent OAuth errors

• feat: add chat-to-responses conversion for models that only support the Responses API
• refactor: integrate model catalog to determine per-model conversion requirements
• chore: upgraded core to v1.5.3 and framework to v1.3.3

• feat: add team budget support with per-team spending tracking and DB tables
• feat: atomic budget and ratelimit update operations for concurrent safety
• refactor: budget DB table restructure to support team budgets
• fix: allow custom providers without a list models endpoint to pass in any model

• chore: upgraded core to v1.5.3 and framework to v1.3.3

• feat: add OCR input logging with request type metadata and detail view
• fix: handle stream terminal detection in logging operations
• fix: capture responses streaming API errors

• chore: upgraded core to v1.5.3 and framework to v1.3.3

• chore: upgraded core to v1.5.3 and framework to v1.3.3

• fix: update OTel semconv from v1.39.0 to v1.40.0
• fix: default insecure to true when omitted so HTTP collectors work without explicit config
• fix: include fallbacks in emitted OTel metrics
• chore: upgraded core to v1.5.3 and framework to v1.3.3

• refactor: Updates prompt plugin flow to change the headers used, and better code quality
• chore: upgraded core to v1.5.3 and framework to v1.3.3

• chore: upgraded core to v1.5.3 and framework to v1.3.3

• chore: upgraded core to v1.5.3 and framework to v1.3.3