v1.6.0

​

✨ Features

• Runware Provider — Added Runware provider support, including image and video generation operations
• Runway Image Operations — Added Runway image generation operations
• Customer Attribution — Added x-bf-customer-id and x-bf-customer-name header support for per-customer attribution
• Enriched Model Listing — list models now returns ContextLength, MaxInputTokens, MaxOutputTokens, Architecture, and WebSearch pricing sourced from pricing entries
• Streaming Pause/Resume — Added pause/resume flows for streaming calls
• Session Trace Grouping — Added group_traces_by_session support to the OTEL/Datadog plugins
• Root Span Content Toggle — Added a toggle to disable root-span content logging
• Password Policy — Added password-policy validation with inline error and sticky save button to the security view
• Error Sanitization — Internal error details (stack traces, SQL) are now sanitized before being sent to clients
• Cluster Discovery Env Refs — Added env.VAR_NAME support to dns_names in cluster discovery config
• Server Logs Config — Added configurable server logging
• Bedrock Streaming Errors — Added __type return for Bedrock errors in streaming paths
• Secret References — Added typed SecretVar env/vault reference support (env.*, vault.*) across config and UI, replacing EnvVar

​

🐞 Fixed

• Streaming Memory — Reduced memory usage on streaming request/response paths
• Provider Config Sync — allow_all_keys and blacklisted_models now sync from the config.json source of truth (thanks @acarpe!) (closes #4640)
• Provider Keys Payload — PUT /api/providers/{provider} no longer silently discards keys/blocked-model edits (thanks @aeciolevy!) (closes #4648)
• Custom Header Base URL — Fixed base-URL protocol handling when a custom header is set
• VK Quota Usage — Fixed the start time for virtual-key quota model usage
• Model Budgets — Fixed model budget attachment from virtual keys
• Structured Streaming Errors — Preserved structured errors for streaming plugin blocks
• Log Hygiene — Removed leaking request bodies from console logs

​

🐙 Closed GitHub Issues

• #2347 — MCP tool ordering is non-deterministic, breaking prefix-based prompt caching
• #3443 — Anthropic→OpenAI streaming tool_call deltas violate OpenAI spec on continuation chunks
• #4068 — Bedrock: mid-conversation system messages hoisted into top-level system block break prompt caching
• #4413 — OpenAI Responses streaming returns empty error.message on context_length_exceeded
• #4460 — GLM-5.2 reasoning_effort “max” silently downgraded to “high”
• #4496 — Frequent intermittent broken pipe / closed connection errors with vllm provider
• #4544 — Cerebras + /anthropic endpoint fails after first turn with 400 provider API error
• #4606 — Realtime socket request observability logs not recorded since v1.5.2
• #4608 — ResponsesMessage drops author/recipient/encrypted_content, breaking Codex multi_agent_v2 subagent spawning
• #4617 — idle-timeout timer goroutine can panic in closeBodyStream and crash the process
• #4622 — Bedrock Converse document blocks with format xlsx/xls/doc/docx silently rewritten to pdf
• #4627 — Gemini video reference fields sent under parameters
• #4640 — provider config key_ids:["*"] not synced to allow_all_keys for existing virtual keys
• #4648 — PUT /api/providers/{provider} silently discards payload.Keys

• feat: add Runware provider with image and video generation operations
• feat: add Runway image generation operations
• feat: add x-bf-customer-id and x-bf-customer-name header support
• feat: enrich list models response with ContextLength, MaxInputTokens, MaxOutputTokens, Architecture, and WebSearch pricing from pricing entries
• feat: add pause/resume flows for streaming calls
• feat: add OpenCode gateway providers (Go, Zen) (thanks @neta79!)
• feat: add advisor tool compatibility for Claude
• feat: add Mantle list models endpoint and OpenAI-compatible endpoint support
• feat: add typed SecretVar env/vault reference support (env.*, vault.*) with store/resolve/remove hooks, replacing EnvVar
• fix: streaming memory improvements
• fix: enrich Fireworks list models from config
• fix(bedrock): correct GLM-5 stop sequences and remove erroneous web search error
• fix(bedrock): broaden document file extension support (xlsx/xls/doc/docx) (closes #4622)
• fix(bedrock): preserve tool_result order to match parallel tool_use blocks (thanks @alexef!)
• fix(bedrock): inline mid-conversation system reminders to preserve prompt cache (thanks @mickgvirtu!) (closes #4068)
• fix(bedrock): set TTL in cache points
• fix: add GLM-5.2 to models supporting max reasoning effort (thanks @is911!) (closes #4460)
• fix: retry on stale connections (closes #4496)
• fix: drop assistant reasoning for Cerebras (closes #4544)
• fix(gemini): map video reference fields to instances (thanks @vojthor!) (closes #4627)
• fix: preserve request conversion for Gemini/Vertex batch requests
• fix: populate error message in OpenAI responses streaming error events (thanks @Purvi09!) (closes #4413)
• fix: preserve structured errors for streaming plugin blocks
• fix: accept object-valued tool-call arguments (e.g. tool_search_call) on the Responses API streaming path (thanks @devonpmack!)
• fix: add key pool filter
• fix: ctx panic fix for fasthttp
• fix: recover from closeBodyStream panic in idle-timeout timer goroutine (thanks @KamilDziemba!) (closes #4617)
• fix(responses): preserve author, recipient, encrypted_content in multi-agent items and message copies (thanks @Shaik-Sirajuddin!) (closes #4608)
• fix: anthropic tool call handling (thanks @rmarku!) (closes #3443)
• fix: commit routing-pinned key ID to reserved BifrostContextKeyAPIKeyID after PreRequestHook unblock
• fix: preserve extra_content on ChatAssistantMessageToolCall so provider-specific tool-call metadata (e.g. Gemini thought_signature) round-trips losslessly (thanks @nghodkicisco!)
• fix: surface Anthropic prompt-cache diagnostics (cache-diagnosis-2026-04-07 beta) via new CacheDiagnostics/CacheMissReason schema
• fix: deterministic MCP tool ordering for prompt cache stability (thanks @ahoblitz!) (closes #2347)

• feat: add global GORM vault callbacks with VaultPathKeyer interface and map[string]EnvVar support, replacing per-model BeforeSave/AfterDelete vault hooks
• feat: add typed SecretVar env/vault reference support (env.*, vault.*), replacing EnvVar
• feat: add configurable server logs
• fix: streaming memory improvements
• fix: include allow_all_keys and blacklisted_models in virtual key hash (closes #4640)
• fix: sync provider config allow_all_keys from config.json source of truth (thanks @acarpe!)
• fix: populate error message in OpenAI responses streaming error events (thanks @Purvi09!)
• fix: shift Add-column migrations to raw SQL
• fix: add responses_input_history fallback to hybrid logstore check for logs UI
• fix(responses): deep-copy author, recipient, encrypted_content in message copies (thanks @Shaik-Sirajuddin!)
• refactor: replace gorm DropColumn with dropColumnIfExists helper
• revert: TableVirtualKey.Value from SecretVar back to string, dropping env/vault reference support for virtual key values

• chore: upgraded core to v1.6.0 and framework to v1.4.0

• feat: add x-bf-customer-id and x-bf-customer-name support
• chore: upgraded core to v1.6.0 and framework to v1.4.0

• refactor: rename EnvVar to SecretVar
• chore: upgraded core to v1.6.0 and framework to v1.4.0

• chore: upgraded core to v1.6.0 and framework to v1.4.0

• refactor: rename EnvVar to SecretVar
• chore: upgraded core to v1.6.0 and framework to v1.4.0

• refactor: rename EnvVar to SecretVar
• chore: upgraded core to v1.6.0 and framework to v1.4.0

• chore: upgraded core to v1.6.0 and framework to v1.4.0

• feat: add group_traces_by_session support
• feat: add root span content logging disable toggle
• feat: add typed SecretVar env/vault reference support
• chore: upgraded core to v1.6.0 and framework to v1.4.0

• chore: upgraded core to v1.6.0 and framework to v1.4.0

• refactor: rename EnvVar to SecretVar
• chore: upgraded core to v1.6.0 and framework to v1.4.0

• refactor: rename EnvVar to SecretVar
• feat: add typed SecretVar env/vault reference support
• chore: upgraded core to v1.6.0 and framework to v1.4.0