v1.6.3

​

✨ Features

• DeepSeek Provider - Added DeepSeek as a first-class provider with dedicated request handling and thinking-mode gating
• AWS Bedrock Mantle Provider - Added bedrock_mantle as a first-class provider with SigV4 key config, native-Anthropic and OpenAI-compatible routing, DB migration, and UI support
• OAuth 2.1 Gateway Auth for MCP - Added a full OAuth2 authorization server for /mcp: discovery endpoints, dynamic client registration, authorize/token with PKCE and refresh token rotation, consent page, JWT Bearer authentication, session listing/revocation with sweep worker, OAuth Grants UI, and mcp_server_auth_mode config
• Virtual Key Expiry - Added an expiry field to virtual keys with governance enforcement
• ClickHouse Log Store (Beta) - Added ClickHouse support for the log store, including a hybrid store mode. This feature is in beta and may have some corner cases.
• IPv6 Support - Added IPv6 support to the HTTP transport
• Per-MCP-Server Tool Timeout - Added per-MCP-server tool execution timeout configuration (thanks @Purvi09!)
• OpenAI Responses Lifecycle APIs - Added missing OpenAI Responses lifecycle methods with explicit per-verb governance flags (thanks @17jmumford!)
• MCP Clients Filtering & Pagination - Added connection_type, auth_type, state, virtual_key, and server/client_id filters with pagination and a faceted filter sidebar on the MCP clients page
• Deprecated Model Marking - Models are now marked is_deprecated in pricing and catalog APIs instead of being filtered out of responses
• Log Attribution Columns - Added user, team, customer, and business-unit name columns to the logs list with multi-value attribution cells
• Latency on Errors - Error responses now carry latency information
• Env-Store Virtual Key Values - Virtual key values now use schemas.SecretVar, enabling env-store references
• Connector Multi-Attribution - Connectors can now attach multiple teams, customers, and business units
• Supplemental External Budgets - Added support for externally resolved supplemental budgets not tracked against a virtual key
• Cost Recalculation Progress - Cost recalculation now streams progress via SSE with batch processing
• Vendor-Prefix Pricing Fallback - Extended Bedrock vendor-prefix pricing fallback to OpenAI, Google, and xAI models
• Complexity Router Improvements - Added stemming alongside exact keyword match and a no-signal fallback to the complexity analyzer
• MCP VK Header - Added x-goog-api-key as a supported virtual-key header on the MCP auth path

​

🐞 Fixed

• Anthropic Redacted Thinking - Round-trip redacted_thinking blocks on chat completions so tool-use turns with extended thinking replay correctly (thanks @fus3r!)
• Bedrock Streaming Block Boundaries - Emit contentBlockStop events on the Bedrock ConverseStream egress (thanks @fus3r!)
• Cache Token Accounting - Report cached_tokens as reads only per the OpenAI spec so cache writes are not billed as reads (thanks @fus3r!)
• Streaming Retries & Fallbacks - Clear the per-attempt stream close claim so streaming retries and fallbacks work after SSE-embedded provider errors (thanks @fus3r!)
• Governance Team IDs - Decode URL-encoded team IDs in fetch, update, and delete endpoints (thanks @nnNyx!)
• Semantic Cache Keys - Resolve semantic cache internal embedding keys like external requests (thanks @nnNyx!)
• Gemini Batch Responses - Surface Gemini batch inline responses from the response field instead of dest (thanks @nnNyx!)
• Governance Rate-Limit CPU - Skip O(N) reference refresh on request-time rate-limit and budget reset
• Tier Cost Calculation - Evaluate tier costs via input tokens instead of total tokens
• Cancelled Requests - Fixed stats and log state for cancelled requests
• Billing on Failed Streams - Fixed billing on failed Responses stream requests for Anthropic and Bedrock, and cost for image generation and edit streaming
• Custom Provider Budgets - Custom providers with spaces in their names can now set budgets
• Model Parameters URL - Honor model_parameters_url changes in config.json like pricing_url (thanks @jeremym-tanium!)
• Bedrock Truncation Signal - Signal Bedrock max_output_tokens truncation on the Responses API (thanks @jeremym-tanium!)
• MCP Reconnect - Fixed MCP clients registering as connected with an empty tool set when ListTools fails during startup (thanks @HackToHell!)
• MCP Tool Ordering - Deterministic MCP tool ordering for prompt cache stability
• Vertex gs:// Images - Pass through gs:// image URLs on Vertex Gemini (thanks @G-XD!)
• Hybrid Log Token Usage - Rebuild token usage from denormalized columns in the hybrid log list (thanks @G-XD!)
• Anthropic Files - Preserve file ID document sources (thanks @mmacvicar!) and forward file IDs and content type on the Anthropic files integration
• Gemini Upload MIME Type - Preserve file upload MIME types (thanks @mmacvicar!)
• Content Logging Bypass - Sanitize ErrorDetailsParsed so raw payloads honor disable_content_logging (thanks @citrocat!), plus error-detail sanitization on the log update path
• Trace Store Memory Leak - Sweep orphaned deferred spans in trace store TTL cleanup (thanks @citrocat!) and complete deferred LLM spans on streaming goroutine exit
• Claude Code Passthrough Streaming - Consistent content_block indices for server tools (thanks @surki!)
• Codex Tool Search Round-Trip - Preserve codex tool_search_call and tool_search_output input items on the Responses API (thanks @raghu-nandan-bs!)
• Gemini Fixes - Guard tool call config, fix the 2.5-pro thinking budget value, OpenAI-through signature compatibility, and video reference field mapping (thanks @vojthor!)
• DeepSeek Thinking - Convert thinking to disabled when tool choice is required
• OpenAI Integration - Propagate max_tokens from the OpenAI integration and pass chunking_strategy through as an extra param
• Bedrock Error Types - Fixed error type setting in all integrations for Bedrock
• Perplexity Responses - Fixed Perplexity Responses API compatibility
• Secret Detection - Set SecretTypePlainText for plain-text JSON and non-prefixed secret values, and check whether virtual key values are secrets
• Empty Tool Results - Fixed empty tool call result insertion failures
• Error Redaction - Redact decoder details from invalid request payload errors
• Vertex Idle Timeout - Fixed idle timeout wiring in the Vertex path
• Web Fetch - Assorted web fetch fixes
• MCP Token Refresh - Skip background token refresh for disabled or unconfigured MCP clients and exclude terminal-status OAuth configs from the refresh query
• SSO Login Loop - Fixed an endless login loop on SSO
• UI Fixes - Governance form calendar-aligned toggle gating, dashboard array query params, MCP sessions table scrolling with sticky header, audit logs layout, and model catalog key aliases displayed as model names

​

🐙 Closed GitHub Issues

• #2347 - MCP tool ordering is non-deterministic, breaking prefix-based prompt caching
• #3106 - Governance team delete/fetch fails for SCIM-synced team IDs containing spaces or URL-sensitive characters
• #3121 - OpenAI responses.retrieve() not supported
• #3139 - Bifrost adds non-standard reasoning/reasoning_details fields to chat completions when using a custom provider for deepseek v4 models
• #3357 - Bifrost billing discrepancy for cancelled requests
• #3951 - Gemini batch: inline responses (dest.inlinedResponses) are silently dropped, leaving output_file_id null
• #4262 - Bedrock ConverseStream egress never emits contentBlockStop (breaks strands streaming)
• #4314 - MCP client registered as connected with empty tool set when ListTools fails during connect/reconnect
• #4402 - Vertex provider drops image blocks whose URL uses gs:// scheme
• #4446 - Add per MCP server level tool timeout configuration
• #4679 - Bedrock Responses API does not signal max_output_tokens truncation
• #4689 - Custom providers cannot set budget
• #4720 - chunking_strategy is dropped for OpenAI-compatible transcription requests
• #4721 - Logs table Tokens column shows N/A when hybrid object storage is enabled
• #4756 - semantic_cache internal embedding path bypasses plugin pipeline, causing “no keys found” while direct /v1/embeddings works
• #4777 - Image generation stream: completed chunk returns empty output_tokens_details, causing under-billing
• #4788 - DeepSeek Anthropic-compatible provider causes “stream closed” error in v1.6.0 (regression from v1.5.16)
• #4816 - /v1 chat completions folds cache-write tokens into prompt_tokens_details.cached_tokens
• #4851 - v1.6.2 governance rate-limit reset causes high CPU in BumpRateLimitUsage/updateRateLimitReferences
• #4863 - model_parameters_url in config.json is ignored after the DB value is set
• #4868 - Memory leak: orphaned deferred spans in TraceStore are never TTL-swept
• #4872 - Raw request/response payloads bypass disable_content_logging via ErrorDetailsParsed
• #4942 - redacted_thinking blocks are dropped on chat completions, breaking tool-use replay with extended thinking

• feat: added DeepSeek as a first-class provider (#4852)
• feat: added bedrock_mantle as a first-class provider with native-Anthropic and OpenAI-compatible routing and SigV4 key config (#4736, #4737)
• feat: added JWT Bearer authentication path for /mcp with session validation, a virtualKeysByID secondary index, and cached signing key and VK lookups (#4508, #4783)
• feat: added per-MCP-server tool execution timeout (#4472, closes #4446) (thanks @Purvi09!)
• feat: added missing OpenAI Responses lifecycle methods (#3125, closes #3121) (thanks @17jmumford!)
• feat: added IPv6 support (#4895)
• feat: added ClickHouse support for the log store (#4748)
• feat: extended Bedrock vendor-prefix pricing fallback to OpenAI, Google, and xAI models and folded bedrock_mantle onto bedrock lookups (#4924)
• feat: added is_deprecated to model pricing and catalog responses and mark deprecated models instead of filtering them (#4779, #4792, #4936)
• feat: added latency info on errors (#4867, #4876)
• feat: added multiple teams, customers, and business units to connectors (#4875)
• feat: virtual key values use schemas.SecretVar to support the env store (#4817)
• feat: chunking_strategy passes through as an extra param for OpenAI models (#4741, closes #4720)
• feat: simplified Responses lifecycle permissions to explicit per-verb flags (#4880)
• fix: round-trip Anthropic redacted_thinking blocks on chat completions so tool-use turns with redacted reasoning can be replayed (#4943, closes #4942) (thanks @fus3r!)
• fix: emit contentBlockStop events on the Bedrock ConverseStream egress so consumers that assemble messages on block boundaries get complete content (#4923, closes #4262) (thanks @fus3r!)
• fix: clear the per-attempt stream close claim so streaming retries and fallbacks are not dead on arrival after an SSE-embedded provider error (#4911, closes #4788) (thanks @fus3r!)
• fix: emit reads-only cached_tokens in usage per the OpenAI spec so cache writes are not priced as cache reads (#4906, closes #4816) (thanks @fus3r!)
• fix: forward file IDs and content type on the Anthropic files integration (#4956)
• fix: preserve Anthropic file ID document sources (#4832) (thanks @mmacvicar!)
• fix: preserve Gemini file upload MIME types for GenAI file URI completions (#4833) (thanks @mmacvicar!)
• fix: propagate max_tokens from the OpenAI integration (#4966)
• fix: error type setting in all integrations for Bedrock (#4958)
• fix: guard setting tool call config in Gemini (#4959)
• fix: Gemini 2.5-pro thinking budget value (#4947)
• fix: Gemini OpenAI-through signature compatibility (#4810)
• fix: surface Gemini batch inline responses from the response field, not dest (#4904, closes #3951) (thanks @nnNyx!)
• fix: Gemini video reference fields map to instances (thanks @vojthor!)
• fix: web fetch fixes (#4945)
• fix: set SecretTypePlainText for plain-text JSON and non-prefixed secret values (#4946) and check whether virtual key values are secrets (#4927)
• fix: idle timeout wiring in the Vertex path and recover from idle-timeout timer-goroutine panic (#4937)
• fix: set content type header consistently for Responses API requests (#4935)
• fix: deterministic MCP tool ordering for prompt cache stability (#4932, closes #2347)
• fix: consistent content_block indices for server tools on Claude Code passthrough streaming (#4890) (thanks @surki!)
• fix: empty tool call result insertion failures (#4925)
• fix: sanitize error details on the log update path and set the raw-storage log flag on standalone MCP tool executions (#4913)
• fix: complete deferred LLM span on streaming goroutine exit (#4885)
• fix: billing on failed Responses stream requests for Anthropic and Bedrock (#4842)
• fix: cost for image generation and image edit streaming (#4802, closes #4777)
• fix: Perplexity Responses API compatibility (#4813)
• fix: signal Bedrock max_output_tokens truncation on the Responses API (#4680, closes #4679) (thanks @jeremym-tanium!)
• fix: preserve codex tool_search_call and tool_search_output input items and accept object-valued tool-call arguments on the Responses API streaming path (#4121) (thanks @raghu-nandan-bs!)
• fix: MCP reconnect failure on startup (#4316, closes #4314) (thanks @HackToHell!)
• fix: pass through gs:// image URLs on Vertex Gemini (#4568, closes #4402) (thanks @G-XD!)
• fix: skip model check for Responses lifecycle APIs (#4920)
• chore: refactored Anthropic request building into BuildAnthropicChatRequestBody, shared completeRequest across Anthropic, Azure, and Bedrock, lazy BodySigner SigV4 signing, and BearerAuthHeader helper (#3309, #4394, #4425, #4735)

• feat: added ClickHouse support for the log store with a hybrid store mode (#4748, #4893)
• feat: added OAuth 2.1 gateway auth: AS discovery endpoints, signing key management, MCPServerAuthMode config, issuance endpoints (DCR, authorize, token) with PKCE and refresh token rotation, session listing, revocation, family-revocation, VK liveness checks, and sweep worker (#4505, #4506, #4509)
• feat: revoke VK-mode OAuth2 grants on VK deletion with user-liveness checks at refresh and request time (#4806)
• feat: push OAuth2 sessions filtering and pagination to SQL with total count (#4775)
• feat: added expiry field to virtual keys (#4887)
• feat: virtual key values use schemas.SecretVar to support the env store (#4817)
• feat: added BedrockMantleKeyConfig support to key hashing, schema/table mapping, and sensitive field clearing (#4737, #4886)
• feat: added per-MCP-server tool execution timeout (#4472, closes #4446) (thanks @Purvi09!)
• feat: added connection_type, auth_type, state, virtual_key, and server/client_id filters with pagination to the MCP clients list (#4839, #4767, #4841)
• feat: added user_name, team_ids, team_names, customer_ids, customer_names, business_unit_ids, and business_unit_names to log list select columns (#4866)
• feat: added UpsertModelParametersBatch for batched model parameter sync (#4800)
• feat: added is_deprecated to pricing and catalog responses and mark deprecated models instead of filtering them (#4779, #4792, #4936)
• feat: drop reasoning when tools are present but reasoning_with_tool_calls is unsupported (#4630)
• feat: extended vendor-prefix pricing fallback to OpenAI, Google, and xAI models (#4924)
• feat: lowered auth_code_ttl default to 300s and enforce a 900s maximum (#4822)
• fix: sweep orphaned deferred spans in trace store TTL cleanup (#4869, closes #4868) (thanks @citrocat!)
• fix: rebuild token usage from denormalized columns in hybrid log list (#4722, closes #4721) (thanks @G-XD!)
• fix: stats for cancelled requests (#4930)
• fix: tier costs evaluated via input tokens instead of total tokens (#4917)
• fix: append datasheet models for incomplete list-models calls (#4879)
• fix: skip background token refresh for disabled or unconfigured MCP clients and guarantee non-nil logger in sync workers (#4848)
• fix: exclude terminal-status OAuth configs from the expiring token refresh query (#4754)
• fix: check whether virtual key values are secrets (#4927)
• fix: web fetch fixes (#4945)
• fix: Perplexity Responses API compatibility (#4813)

• feat: drop reasoning when tools are present but reasoning_with_tool_calls is unsupported (#4630)
• fix: convert thinking to disabled when tool choice is required for DeepSeek (#4861)
• chore: upgraded core to v1.6.3 and framework to v1.4.3

• feat: added expiry enforcement for virtual keys (#4887)
• feat: complexity analyzer stemming support alongside exact keyword match and a no-signal fallback (#4708, #4791)
• feat: added virtualKeysByID secondary index with cached signing key and VK lookups on the /mcp JWT auth path (#4783)
• feat: virtual key values use schemas.SecretVar to support the env store (#4817)
• fix: skip O(N) reference refresh on request-time rate-limit and budget reset (#4883, closes #4851)
• fix: skip model check for Responses lifecycle APIs (#4920)
• fix: empty tool call result insertion failures (#4925)
• chore: upgraded core to v1.6.3 and framework to v1.4.3

• chore: upgraded core to v1.6.3 and framework to v1.4.3

• feat: latency info on errors (#4867)
• feat: stream cost recalculation progress via SSE with batch processing (#4778)
• fix: sanitize ErrorDetailsParsed so raw payloads honor disable_content_logging (#4873, closes #4872) (thanks @citrocat!)
• fix: sanitize error details on the log update path and remove redundant immediate error serialization (#4913)
• fix: cancelled state in logs (#4831, closes #3357)
• fix: empty tool call result insertion failures (#4925)
• chore: upgraded core to v1.6.3 and framework to v1.4.3

• chore: upgraded core to v1.6.3 and framework to v1.4.3

• chore: upgraded core to v1.6.3 and framework to v1.4.3

• chore: upgraded core to v1.6.3 and framework to v1.4.3

• chore: upgraded core to v1.6.3 and framework to v1.4.3

• chore: upgraded core to v1.6.3 and framework to v1.4.3

• fix: resolve internal embedding keys like external requests (#4903, closes #4756) (thanks @nnNyx!)
• fix: clear body-transport state for internal embedding requests via ClearContextForInternalRequest (#4918)
• chore: upgraded core to v1.6.3 and framework to v1.4.3

• chore: upgraded core to v1.6.3 and framework to v1.4.3