v1.4.12

​

✨ Features

• WebSocket based responses API — Added WebSocket transport for responses API (OpenAI)
• Anthropic Passthrough — Added native Anthropic passthrough endpoint
• Prompt Repository — Added HTTP handlers for prompt management with RBAC (folders, prompts, versions, sessions)
• Streaming Request Decompression — Threshold-gated streaming decompression with pooled readers, replacing BodyUncompressed()
• Model Parameters API — Added model parameters table and API endpoint with in-memory caching
• Virtual Key Limit Resets — Added virtual key limit reset functionality
• Session Stickiness — Added session stickiness in key selection for consistent routing
• Pricing Engine Refactor — Unified cost calculation with quality-based image and video pricing
• Image Configuration — Added size/aspect ratio config for Gemini and size-to-resolution conversion for Replicate
• Large Payload Support — Added large payload awareness across transport hooks, plugins, and response streaming
• Raw Request/Response Storage — Allow storing raw request/response without returning them to clients (thanks @Vaibhav701161!)
• ChatReasoning Enabled Field — Added Enabled field to ChatReasoning struct (thanks @mango766!)

​

🐞 Fixed

• Deterministic Tool Schema — Fixed deterministic tool schema serialization for Anthropic prompt caching (thanks @Edward-Upton!)
• CORS Wildcard — Fixed CORS issue with allowing * origin
• TLS Termination — Allow TLS termination inside Bifrost server through config
• Bedrock toolChoice — Fixed toolChoice silently dropped on Bedrock /converse and /converse-stream endpoints
• Count Tokens Passthrough — Fixed request body passthrough for count tokens endpoint for Anthropic and Vertex
• Chat Finish Reason — Map chat finish_reason to responses status and preserve terminal stream semantics
• Tool Call Indexes — Fixed streaming tool call indices for parallel tool calls in chat completions stream
• Video Pricing — Fixed video pricing calculation
• SQLite Migration — Prevented CASCADE deletion during routing targets migration
• Log Serialization — Reduced logstore serialization overhead and batch cost updates
• Log List Queries — Avoid loading raw_request/raw_response in log list queries (thanks @Vaibhav701161!)
• MCP Reconnection — Improved MCP client reconnection with exponential backoff and connection timeout
• Responses Input Messages — Set responses input messages in gen_ai.input.messages
• Helm Fixes — Fixed Helm chart and test issues

• feat: WebSocket and Realtime API support
• feat: Anthropic passthrough support
• feat: threshold-gated streaming request decompression with pooled readers
• feat: refactored model catalog pricing engine with unified cost calculation
• feat: quality-based image pricing and image size/aspect ratio for Gemini
• feat: size-to-resolution conversion for Replicate image models
• feat: session stickiness in key selection
• feat: add Enabled field to ChatReasoning struct (thanks @mango766!)
• feat: allow storing raw request/response without returning to clients (thanks @Vaibhav701161!)
• feat: RBAC for prompt repository
• fix: deterministic tool schema serialization for prompt caching (thanks @Edward-Upton!)
• fix: skip body building for large payload flow
• fix: TLS termination inside Bifrost server through config
• fix: map chat finish_reason to responses status and preserve terminal stream semantics
• fix: set responses input messages in gen_ai.input.messages
• fix: video pricing fixes
• fix: remove resolution parameter from image generation
• fix: MCP client reconnection with exponential backoff and connection timeout

• fix: record ttft in nanoseconds instead of milliseconds to avoid truncation to 0
• feat: add routing_targets table with 1:many relationship to routing_rules; migrates existing single-target rules to the new table with weight=1; drops legacy provider and model columns from routing_rules
• feat: add per-target key_id pinning support in routing_targets
• fix: avoid postgres cached-plan failures during provider hash backfill @dannyball710
• feat: prompt repository with folder, prompt, version, and session schemas and backend
• feat: model parameters table and API endpoint with in-memory caching
• feat: large payload awareness for plugins and logstore
• feat: large payload transport hooks and response streaming
• feat: chat token detail OTEL span attributes
• feat: hide deleted virtual keys from filter options
• feat: virtual key search/filtering and pagination
• feat: allow storing raw request/response without returning to clients (thanks @Vaibhav701161!)
• fix: MCP client reconnection with exponential backoff and connection timeout
• fix: prevent SQLite CASCADE deletion during routing targets migration
• fix: reduce logstore serialization overhead and batch cost updates
• fix: video pricing fixes
• fix: avoid loading raw_request/raw_response in log list queries (thanks @Vaibhav701161!)
• chore: upgraded core to v1.4.8

• feat: pricing engine integration with unified cost calculation
• feat: large payload awareness
• chore: upgraded core to v1.4.8 and framework to v1.2.27

• chore: upgraded core to v1.4.8 and framework to v1.2.27

• chore: upgraded core to v1.4.8 and framework to v1.2.27

• feat: passthrough support for log capture
• feat: large payload awareness for logstore
• feat: async log write improvements
• fix: reduce logstore serialization overhead and batch cost updates
• fix: avoid loading raw_request/raw_response in log list queries (thanks @Vaibhav701161!)
• chore: upgraded core to v1.4.8 and framework to v1.2.27

• feat: WebSocket and Realtime API support
• chore: upgraded core to v1.4.8 and framework to v1.2.27

• chore: upgraded core to v1.4.8 and framework to v1.2.27

• fix: set responses input messages in gen_ai.input.messages
• chore: upgraded core to v1.4.8 and framework to v1.2.27

• feat: add Enabled field to ChatReasoning struct (thanks @mango766!)
• feat: large payload awareness
• chore: upgraded core to v1.4.8 and framework to v1.2.27

• feat: pricing engine integration
• chore: upgraded core to v1.4.8 and framework to v1.2.27