Skip to main content
v2.6.0

Changelog

v2.6.0 moves the enterprise gateway onto OSS transports v2.6.0. Access profiles gain a Restrict mode, where the profiles of a user’s teams, business units and customers cap what the user can reach instead of adding to it. Licenses can now be bound to a cloud VPC, a Kubernetes or OpenShift cluster, or a Linux machine. Business units hold several budgets, like teams and customers. The OSS release adds named model routers, the Bifrost MCP server on /v1/mcp, Claude Code gateway sign-in, and realtime speech sessions.

✨ Features

  • Restrict Mode for Entity Access Profiles - With client.entity_access_profile_mode: "intersect" (shown as Restrict), the access profiles of a user’s teams and business units, and of their customers, cap the user instead of adding to their access. A level with no profile is skipped, and any one profile on a level is enough. The cap covers providers, models, provider keys, MCP tools and relays, for user requests, profile-managed keys, legacy keys a member holds, and keys owned by a team, business unit or customer. Budgets and rate limits are charged as before, and the setting applies from the next request. Docs
  • Several Budgets on Business Units - A business unit holds one budget per reset window plus calendar_aligned, like a team or customer. POST and PUT /api/governance/business-units[/{id}] take budgets[], rate_limit, calendar_aligned and reset_budget_usage, and config.json accepts inline budgets[]. The deprecated single budget / budget_id is still accepted. Teams, business units and customers also get the budget and rate-limit editor on their Usage tab (“Add limits”).
  • License Deployment Binding - A license can bind the gateway to where it runs, and a mismatch raises an alert in the logs, on the license page and in the top bar, then closes the license gate after an 8-day grace period. Licenses without a binding skip the check.
    • Cloud VPCs (vpcids) - The cloud account and VPC are read from AWS, GCP or Azure instance metadata. When IMDS is unreachable (for example EKS Auto Mode), AWS falls back to the EC2 DescribeNetworkInterfaces API with the pod’s credentials. GKE pods with Workload Identity read the network through the Compute API. AKS scale-set NICs are supported, and BIFROST_LICENSE_AZURE_CLIENT_ID picks the identity on nodes with several user-assigned identities.
    • Kubernetes and OpenShift clusters (k8scluster) - Compares the UID of the kube-system namespace, read with the pod’s service account.
    • Linux machines (machineids) - Compares a salted fingerprint of /etc/machine-id (or the DMI product UUID). The raw ID never leaves the host. The check is fully offline, so it works on bare metal and air-gapped installs.
    • bifrost-http --print-deployment-id prints the cluster and machine identifiers needed to request a license.
    • A missing cloud permission (aws_permission_missing, gcp_permission_missing, azure_identity_ambiguous) or an unreachable Kubernetes API reports an undetermined status and never starts the grace period.
  • Bifrost MCP Server on Enterprise - /v1/mcp serves the gateway’s API, including enterprise operations, as MCP tools. Each tool call passes through API keys, SCIM, RBAC and audit like a direct /api request, and tools/list shows each caller only the tools RBAC lets them call. When openapi_url is unset, the spec comes from the published https://getbifrost.ai/openapi.json. /api/session/me and /api/governance/users/me/limits are open to any signed-in caller, and /api/governance/users/{id}/limits is gated with access profile permissions. Docs
  • Routers Permission - Model router routes (/api/routing/model-routers) move from Governance to a new Routers RBAC resource. Admin and Developer get full access and Viewer gets View. Testing candidate phrases and pricing a suggestion’s impact need only View; classify keeps Create because it can make paid model calls. Docs
  • Model Routers in Clusters - Model router and routing rule changes are broadcast to every node, and each node rebuilds and warms its router from the shared store. Open dashboards on any node refresh.
  • Datasheet Permission - The model datasheet editor (/api/models/datasheet*) has its own Datasheet RBAC resource with View and Update, so editing model limits and prices no longer comes with ModelProvider access. Admin edits, Developer and Viewer read. Docs
  • Claude Code Gateway Sign-In with Enterprise Identity - Users signed in through the Claude Code gateway pass enforce_auth_on_inference, get the same access profile permits as their IdP token, and are logged with their team, business unit and customer. The consent page refuses users who hold no access, and an expired SSO session is refreshed so the page shows “Continue as <user>”. Docs
  • Guardrails on Embeddings - Text parts of embedding input are screened by input rules, can be redacted before they are forwarded, and are redacted in stored logs. Token-only inputs skip evaluation. Docs
  • Service Tier in Guardrail Rules - LLM rules can match on service_tier (auto, default, flex, priority, fast, ultrafast), for example to keep fast-mode traffic out of a rule. Requests with no tier read as "", so they are never silently exempted. The rule builder has a Service tier field.
  • Secrets Guardrail Rule Overrides - Secrets policies take disabled_rules and a per-rule rule_entropy threshold (0 to 6), edited in a new Rule overrides section. GET /api/guardrails/secrets/rules lists every detection rule with its default entropy. The stock generic-api-key rule now uses an entropy threshold of 3.8 and ignores UPPER_SNAKE environment variable names.
  • Bedrock Guardrail Usage - Bedrock ApplyGuardrail calls are recorded in guardrail metadata with their usage units and an estimated list-price cost. The cost is for display only and never counts toward request cost or budgets.
  • Noma Gateway Version - Noma scans send the running Bifrost version as gatewayVersion, and the setup hint points to Settings → Client Credentials with the Runtime Integrator role.
  • Fewer Workload Identity Limits - SSO providers can trust any number of extra issuers and audiences, and there is no cap on workload identity mappings, token size or subject length. Expired verified-token cache entries are swept on a timer.

🌎 Open Source Features

  • Model Routers - Named model routers classify each request into SIMPLE, MEDIUM or COMPLEX with a semantic, decision-model or LLM classifier, and send each tier to its own weighted provider/model/key targets with fallbacks. A router takes the requests its routing matches (global, team, customer, virtual key or user scope, an optional CEL condition and a priority) after the routing rules run, and is managed from the Routers section, /api/routing/model-routers or config.json. Request logs record which router routed each request. Docs
  • Complexity Router Phrase Suggestions - A model router can sample its traffic, cluster it on a schedule within a spend budget and suggest phrases for each tier. The suggestions view shows their impact on past traffic, adds accepted phrases, backfills from existing logs and can classify a test statement. The dashboard view is behind a feature flag.
  • Bifrost MCP Server - POST /v1/mcp serves Bifrost’s own HTTP API as MCP tools, so MCP clients such as Claude Code can manage a gateway and call its inference routes. Configured with client.bifrost_url, client.bifrost_mcp_allow_writes (reads only by default) and framework.pricing.openapi_url, or BIFROST_MCP_SPEC for air-gapped installs. Docs
  • Warp Management Actions - Warp answers questions about the gateway’s configuration and can create resources through the Bifrost MCP. Every write ends with a confirmation card and runs only after the user confirms, and writes are off unless enabled in Warp settings. New endpoints: GET /api/session/me, user limit lookups, GET /api/logs/search and GET /api/logs/{id}/trace. Docs
  • Claude Code Gateway Sign-In - Claude Code can sign in with /login (Cloud gateway) at <issuer_url>/gateway/claude-code: device-code sign-in through the consent page, silent token refresh, inference resolved to the approved virtual key, managed settings, and optional forwarding of Claude Code’s OTLP telemetry. Configured under oauth2_server_config.gateways.claude_code or Settings → Gateways. Docs
    The gateway reserves the /gateway route prefix. A server.base_path whose first segment is gateway is refused at startup, so move such a deployment to another prefix before upgrading.
  • Realtime Speech and Live Sessions - /v1/realtime serves Gemini Live on the Gemini API and Vertex AI, plus ElevenLabs Scribe and Soniox real-time transcription (intent=transcription). GenAI SDK clients in Vertex mode can open Live sessions on /genai. /v1/audio/speech/stream streams text in and audio back, billed once per session.
  • ElevenLabs Drop-In Routes - ElevenLabs SDKs can point at {gateway}/elevenlabs with a Bifrost virtual key in xi-api-key: text-to-speech (with and without timestamps), the model list filtered to the key, and the native Scribe and text-to-dialogue WebSockets. /elevenlabs_passthrough forwards other routes. Docs
  • Vertex Text-to-Speech - Vertex serves /v1/audio/speech through Cloud Text-to-Speech (Chirp 3 HD and other Google voices) and Gemini TTS models, with streaming for Gemini TTS.
  • Soniox Provider - New soniox provider for speech-to-text (async jobs behind /v1/audio/transcriptions), text-to-speech with streaming, and real-time transcription. Docs
  • Baseten Provider - New baseten provider for Model APIs and dedicated deployments: chat, Responses, embeddings, transcription, /v1/decisions, image generation on dedicated deployments, and rerank on dedicated BEI deployments. Docs
  • Kimchi Provider - New kimchi provider for chat, Responses, embeddings and list models. Docs
  • MachGen Provider - New machgen provider for text-to-image, image edit and async video generation with retrieve and download. /v1/models lists each key’s models with their task types. Docs
  • Sarvam Token Factory v2 - Sarvam chat and model listing move to Token Factory v2, and Sarvam Prism serves /v1/decisions.
    sarvam-30b is no longer served. Use sarvam-105b.
  • Gemini Decisions - Gemini decision models (gemini-decision-*) serve /v1/decisions natively, and a new input_file part carries PDFs, audio and video.
  • Regex Key Aliases - A key alias can be a regex: pattern over the requested model with a model_id template that reuses its capture groups (for example regex:(claude-.*) → global.anthropic.$1), so one row aliases a whole model family.
  • Datasheet Overrides - Model pricing and capabilities can be edited from a datasheet editor under Models, through override endpoints or datasheet_overrides in config.json. Overrides survive every upstream datasheet sync. Docs
  • Usage Estimates for Cancelled Streams - A cancelled or timed-out chat, text or Responses stream has its missing usage estimated with a tokenizer and priced like reported usage. Estimated fields are marked in usage.estimation.
  • Bedrock Native Structured Output - Bedrock Converse sends json_schema response formats through native outputConfig where supported and falls back to the synthetic tool otherwise. supports_native_structured_output overrides the default per model.
  • Custom Tools on Every Provider - Custom (freeform) tools such as Codex’s apply_patch are sent as function tools to models without native support, and come back as custom tool calls on Chat Completions and Responses, streaming included.
  • Codex WebSocket Through ChatGPT Sign-In - /chatgpt_passthrough/backend-api/codex/responses accepts Codex’s WebSocket and relays it with the caller’s ChatGPT token. Each turn is checked against the virtual key and logged and billed like an HTTP turn.
  • MCP Gateway Settings - slug_code_mode toggles Code Mode for one client’s /mcp/<slug> endpoint, mcp_client_metadata_url presents a self-hosted Client ID Metadata Document, and mcp_server_max_protocol_version caps the protocol version /mcp serves.
  • Configurable OpenAPI MCP Spec Size - The OpenAPI MCP spec limit rises from 5 MiB to 32 MiB by default and is set with mcp_openapi_max_spec_mb (up to 256).
  • MCP Logs Dashboard - MCP logs show a metric strip with sparklines and trends, a latency histogram, cancelled counts, readable error cards, and user, team and customer filters.
  • Stop Reason Column - The logs table has an optional stop reason column.
  • Trace Dimensions on Child Spans - An OTEL profile with apply_trace_dimensions_to_child_spans merges the request’s x-bf-dim-* dimensions onto every exported span, not only the root HTTP span.
  • Provider Queue Wait Metric - New bifrost_provider_queue_wait_seconds histogram records how long an attempt waits for a provider worker.
  • opencode Sessions - Requests from opencode carry a stable x-opencode-session header per virtual key.
  • Agent Tracing and Log Previews - Agent traces are easier to follow, and agent-loop log rows preview the latest tool call or result.
  • Anthropic Thinking Block Binding - thinking.block_binding is forwarded to the Claude API with its beta header.
  • Fable 5.2 Fast Mode on Vertex - "speed": "fast" is sent to Vertex for claude-fable-5-2.
  • Databricks Scoped OAuth Secrets - Service principals whose OAuth secret is limited to specific scopes work without configuration.
  • Virtual MCPs on Virtual Keys - The virtual key details sheet lists the Virtual MCPs assigned to the key.
  • FIPS-Approved Key Derivation - In FIPS 140-3 mode, the encryption data key and password hashes use PBKDF2-HMAC-SHA-256. Both formats stay readable, so a rolling deploy is safe.

⚙️ Upgrade Notes

  • Model router routes need Routers permissions - /api/routing/model-routers no longer reads Governance permissions. The default roles are granted on upgrade; grant Routers to custom roles yourself.
  • Datasheet editor needs Datasheet permissions - /api/models/datasheet* no longer reads ModelProvider permissions. The default roles are granted on upgrade; grant Datasheet to custom roles yourself.
  • Customer headers are ignored in Restrict mode - With entity_access_profile_mode: "intersect", x-bf-customer-id and x-bf-customer-name are not read, so they cannot lift the customer cap. Union mode is unchanged.
  • Stricter generic-api-key defaults - The secrets guardrail reports fewer short identifier-like values. Set rule_entropy for generic-api-key to 3.5 to restore the previous threshold.
  • Custom Go plugins must be rebuilt with Go 1.27.4 - The gateway now builds with Go 1.27.4, and plugins built with an older toolchain will not load.
  • Plugin URL allowlist - URL-based plugins download only from hosts in BIFROST_PLUGIN_URL_ALLOWLIST.
    The allowlist is empty by default, so no plugin URL is fetched until its host is listed. Set BIFROST_PLUGIN_URL_ALLOWLIST (comma-separated hosts, *.example.com or *) before upgrading if you load plugins from a URL.
  • Upstream latency metric - bifrost_upstream_latency_seconds no longer includes provider queue wait, retry backoff or other plugins’ hooks.
    Dashboards built on bifrost_upstream_latency_seconds will show lower values after upgrading. Queue wait is now in bifrost_provider_queue_wait_seconds.
  • ElevenLabs timestamp fields renamed - Speech alignment fields move to seconds.
    Breaking for Go SDK users: SpeechAlignment.CharStartTimesMs / CharEndTimesMs are renamed to CharacterStartTimesSeconds / CharacterEndTimesSeconds, and the JSON fields char_*_times_ms become character_*_times_seconds, in seconds.

🐞 Fixed

  • Go 1.27.4 - Security release with fixes to net/http (CVE-2026-97054), cgo, the go command, the compiler and the linker. Docker images move to golang:1.27.4-alpine3.24.
  • Dependency Security Updates - golang.org/x/net 0.62.0 (CVE-2026-78661, CVE-2026-97045), github.com/klauspost/compress 1.20.3 (CVE-2026-116591) and github.com/andybalholm/brotli 1.2.8 (CVE-2026-111497).
  • Progressive Guardrail Streams - A late finding in text the client already received no longer ends the stream; the delivered text stays as sent and later text is still masked.
  • Deleted Service Accounts - Deleting a service account revokes its workload identity connections in memory on every node, even when a node’s store read fails.
  • JWKS Refresh Data Race - The background signing-key refresh no longer reads a recycled request context.
  • Workload Tokens for Agents - An accepted workload token is forwarded to an agent whose registration enables credential forwarding, so agent handoffs no longer fail with 401.
  • ElevenLabs Virtual Key Header - xi-api-key is recognized as a virtual key header under enterprise identity.
  • SCIM PATCH and Entra Emails - An SSO login keeps the SCIM profile keys, and a userName-only PATCH no longer replaces an Entra user’s email with their object ID.
  • SCIM-Deleted Users - Clearing a deleted user’s IdP data only touches users that stay deleted, so a user who signs in again keeps their profile.
  • DAC on User Write Routes - Detaching a user’s access profile and adding or deleting a user’s virtual key answer 404 for a user outside the caller’s scope.
  • DAC on Agent Log Deletes - DELETE /api/agents/history refuses rows outside the caller’s AgentLogs scope.
  • DAC on Model Configs - GET, PUT and DELETE /api/governance/model-configs/{id} follow the caller’s model config scope.
  • DAC on Prompts - Folder listing counts, folder deletes, prompt updates and the latest version read follow the caller’s prompt scope; deleting a folder that holds a hidden prompt answers 409.
  • DAC on MCP Logs - The MCP latency histogram and filter options follow the caller’s MCP scope.
  • DAC on Router Suggestion Impact - A router’s suggestion impact report prices only logs the caller can see.
  • Customer Counts - Team and business unit lists count only customers the caller can see.
  • Teams Linked by customer_id - Customer reads, filters, DAC and the org chart include teams linked through their own customer_id, and detaching a team from a customer also clears that column.
  • Semantic Classifier Warmup on Peers - A failed router warmup restarts on every node after a provider or key change, not only on the node that served the request.
  • Budget Burn Chart - Users with several access profiles get burn lines for budgets whose spend can be attributed.
  • Model Armor Credentials Field - The service account JSON field is a multi-line textarea.
  • Datadog Tag Warnings - Governance tag warnings are no longer logged when the tags are not referenced.
  • Identity Input Limits - Size caps on ID-JAG assertions and client IDs, discovery and JWKS documents, and Entra search cursors are removed.
  • Mid-Conversation System Reminders - Claude Code’s trailing system reminder keeps one form across turns, so Bedrock InvokeModel no longer invalidates the prompt cache every turn.
  • Vertex Prompt Cache Misses - Vertex Claude request bodies always end with tools, system, messages, so turn-to-turn cache reads no longer miss.
  • Cached Tokens Counted Twice - Anthropic streaming usage counts each cached prompt token once, including upstreams that report the cache split on message_delta.
  • Prompt Cache on Aliases - Prompt cache auto-injection checks an alias’s canonical model.
  • prompt_cache_options - prompt_cache_options reaches only the models that accept it.
  • Anthropic Deferred Web Search - A deferred web search is replayed without a made-up web_search_tool_result, so continuations no longer 400.
  • Anthropic MCP History on Chat Providers - History with mcp_tool_use / mcp_tool_result or server-tool blocks no longer sends a roleless message to OpenAI-compatible providers.
  • Bedrock Structured Output Streams - output_config.format on Bedrock Claude streams one well-formed text block.
  • Bedrock base_url - Bedrock honours network_config.base_url for every Bedrock API.
  • Bedrock Beta Headers - A caller’s anthropic_beta on Converse is merged with the betas Bifrost derives.
  • Bedrock tool_reference Blocks - InvokeModel keeps tool_reference blocks inside a client tool_result.
  • Bedrock Computer Tools - Claude Responses requests with a computer tool are served through InvokeModel.
  • Vertex Anthropic Betas - Vertex receives Anthropic betas in the anthropic-beta header, including on CountTokens.
  • Bedrock Mantle OpenAI Models - Unsupported effort-dependent sampling parameters are stripped.
  • Logprobs Across Formats - Logprobs survive conversion between Chat Completions and Responses.
  • Gemini Function Call IDs - Gemini function calls without an ID get unique IDs.
  • Nested Custom Tool Names - Chat custom tools keep name and description nested under custom.
  • Gemini Image Output Format - output_format and output_compression map to Gemini image options on Vertex.
  • Mistral Reasoning Effort - reasoning_effort follows each Mistral model’s ladder instead of being rewritten to high.
  • ElevenLabs Timestamps - Speech with_timestamps returns character times instead of null.
  • ElevenLabs WAV and Scribe Access - Non-streaming wav speech returns a real WAV file, and Scribe works with wildcard virtual keys.
  • Empty Transcriptions - Transcribing silence with text, srt or vtt output returns empty text instead of a 500.
  • Unsupported Streaming Speech Models - Streaming speech on OpenAI tts-1 and tts-1-hd returns a 400 instead of a 500.
  • GenAI Audio by URL - A /genai request with an audio fileData URL is answered as a multimodal prompt.
  • Vertex Model Listing - us and eu keys list models from the same regional host as inference.
  • allowed_models and Key Aliases - A virtual key that allows a model also allows an alias that resolves to it.
  • Unprefixed Models on Keyless Providers - Bare Claude Code model names work on the Anthropic passthrough.
  • Model Parameters After URL Sync - Stored model parameters stay in effect after the background sync.
  • Deterministic Config Hashes - Provider config hashes no longer change with map order, so API-added keys are not dropped on restart.
  • Seedance Video Tokens - Seedance video settlements record their output tokens.
  • Active Requests Gauge Leak - A stream cancelled before the first byte no longer leaves bifrost_active_requests raised.
  • Upstream Latency - bifrost_upstream_latency_seconds no longer includes queue wait, retry backoff or other plugins’ hooks.
  • Large File Uploads - Multipart uploads are streamed to temp files instead of held in memory.
  • Virtual Key Budgets - Updating a config.json virtual key’s budget, moving a quarterly budget, and updating a team’s or customer’s rate-limit maximum alone no longer fail.
  • Team and Customer Saves - Saving a customer or team no longer rewrites its teams, virtual keys or customer.
  • Duplicate Provider Configs - A virtual key naming the same provider twice is refused with a 400.
  • MCP OAuth Consent - The consent page no longer gets stuck when an expired SSO session is refreshed mid-flow.
  • OAuth Grant Expiry - The OAuth Grants page uses the configured access token TTL.
  • OpenAPI MCP Tool Arguments - Every argument gets a name clients accept, so none are dropped.
  • A2A Agents - Valid A2A v1 agent cards register, and Agent Logs show Task artifacts and history.
  • Log Retention on Postgres - Retention deletes shrink their batch on statement timeout, and retention tables get tighter autovacuum settings.
  • ClickHouse Hybrid Storage - Object uploads update has_object in batches, which stops memory growth and dropped uploads.
  • Log Filters - Relay log filter values come from the last 30 days, and MCP log team and customer filters match multi-entity calls.
  • Encoded Request Paths - Transport plugins see the raw encoded path.
  • Dashboard - Log headers and the date picker fit on mobile, and the datasheet editor hides entries without a provider.

🗄️ Database Migrations

Enterprise

  • ent_add_datasheet_rbac_resource - Adds the Datasheet resource with View and Update, and grants Admin both and Developer and Viewer View. Rollback is a no-op, so the resource and grants stay after a downgrade.
  • ent_add_routers_rbac_resource - Adds the Routers resource with Create, View, Update and Delete, and grants Admin and Developer all four and Viewer View. Rollback is a no-op.
  • ent_add_license_deployments_table - Creates enterprise_license_deployments, where every instance records the account and VPC, cluster or machine it runs in. Reversible: rollback drops the table, which loses only the grace clock of an ongoing mismatch.
  • ent_migrate_business_unit_budgets_to_budgets_table - Adds governance_business_units.calendar_aligned and governance_budgets.business_unit_id, and backfills each unit’s budget from its old budget_id link. A budget several units share goes to the unit created first, and a budget another entity already owns stays with that owner; both log a warning. The old budget_id values are kept so a replica on the previous release keeps enforcing them. Rollback drops calendar_aligned only.

Open Source

This release adds columns to the logs table (logs_add_model_routing_columns) and the agent_logs table (agent_logs_add_caller_correlation_id_column). Bifrost handles concurrent access on the application side, but on large log tables the change still takes locks and adds load. Deploy during a low-traffic window.
  • add_usage_estimation_tokenizers_json_column - Adds the nullable tokenizer settings column to config_client for usage estimation. Reversible: rolls back by dropping the column.
  • add_datasheet_overrides_table - Creates governance_datasheet_overrides. Non-reversible: dropping it would delete user datasheet edits. The table is additive and older binaries ignore it.
  • add_databricks_scopes_column - Adds databricks_scopes to config_keys. Reversible: rolls back by dropping the column.
  • add_model_routing_suggestion_schedules_table - Creates the phrase suggestion schedules table. Reversible: rolls back by dropping the table.
  • add_model_routers_table - Creates model_routers. Non-reversible: dropping it would delete every model router.
  • add_model_router_targets_and_routing - Creates model_router_targets and adds routing columns to model_routers. Non-reversible: dropping it would delete every router’s targets.
  • add_model_router_suggestions_tier_columns - Adds suggestion tier columns to model_routers. Reversible: rolls back by dropping the columns.
  • add_mcp_client_server_type_on_slug_column - Adds the nullable slug_code_mode setting to config_mcp_clients. Reversible: rolls back by dropping the column.
  • add_mcp_client_metadata_url_columns - Adds config_client.mcp_client_metadata_url and oauth_configs.client_metadata_document. Non-reversible: dropping them would discard the hosted document URL and which OAuth configs use Bifrost’s document.
  • add_client_config_mcp_openapi_max_spec_mb_column - Adds mcp_openapi_max_spec_mb to config_client. Reversible: rolls back by dropping the column.
  • add_client_config_bifrost_url_and_bifrost_mcp_columns - Adds the Bifrost URL and Bifrost MCP settings to config_client. Reversible: rolls back by dropping the columns.
  • add_framework_config_openapi_url_column - Adds the OpenAPI URL to the framework config. Reversible: rolls back by dropping the column.
  • add_warp_management_writes_column - Adds management_writes to the Warp config. Reversible: without the column Warp offers reads only.
  • add_client_config_entity_access_profile_mode_column - Adds entity_access_profile_mode (default union) to config_client. Reversible: rolls back by dropping the column.
  • add_budget_business_unit_id_column - Adds a nullable business_unit_id owner column to governance_budgets. Reversible: rolls back by dropping the column.
  • add_budget_business_unit_id_index - Builds the business_unit_id index concurrently on PostgreSQL. Reversible: rolls back by dropping the index.
  • add_oauth2_authorize_requests_device_code_columns - Adds the nullable device_code_hash and user_code_hash columns to oauth2_authorize_requests, with a bounded lock wait on PostgreSQL. Reversible: only short-lived sign-in state is dropped.
  • add_oauth2_authorize_requests_device_code_indexes - Builds their unique indexes concurrently on PostgreSQL, so writes are never blocked. Reversible: rolls back by dropping the indexes.
  • add_client_config_mcp_server_max_protocol_version_column - Adds mcp_server_max_protocol_version to config_client. Reversible: without the column /mcp serves every version, as before.
  • agent_logs_add_caller_correlation_id_column - Adds caller_correlation_id to agent_logs. Non-reversible once any row records a value: the rollback refuses while one does.
  • logs_add_model_routing_suggestion_tables - Creates the log-store tables for phrase suggestions. Reversible: rolls back by dropping the tables.
  • logs_add_model_routing_columns - Adds model router columns to logs. Non-reversible once any log records a router: the router a request was routed by cannot be recaptured, so the rollback refuses while one does.
  • logs_tune_retention_autovacuum - Sets autovacuum_vacuum_scale_factor=0.01 and autovacuum_vacuum_threshold=10000 on the Postgres retention tables. Reversible: rolls back by resetting the options.
All new migrations add tables, nullable columns or indexes and never rewrite existing rows, so they are safe for a rolling deploy.

🐙 Closed OSS Issues

  • #123 - Files API Support
  • #3770 - Propagate x-bf-dim-* headers to all spans in a trace, not just the root HTTP span
  • #4965 - Bedrock never activates fine-grained-tool-streaming - tool.eager_input_streaming is silently dropped
  • #5272 - Bedrock provider ignores Helm network_config.base_url (and AWS_ENDPOINT_URL_*), falls back to default public runtime URL
  • #5354 - Anthropic streaming usage double-counts cache-read tokens (~2× prompt_tokens inflation)
  • #5510 - Anthropic passthrough double-counts cache-read tokens for Kimi-compatible streams
  • #5792 - Anthropic passthrough rejects bare (unprefixed) Claude Code model names with “could not auto resolve a provider”
  • #6406 - Vertex: prompt cache mostly misses because messages is serialized before system/tools
  • #6815 - Forward x-opencode-session on opencode requests
  • #7963 - Claude Code gateway sign-in (/login → Cloud gateway) for Bifrost
  • #8073 - Custom model-parameters rows stop applying after the background model-parameters URL sync
  • #8094 - Provider config hash is non-deterministic when maps have 2+ entries; API-added keys dropped on restart
  • #8467 - ElevenLabs Scribe is blocked by wildcard virtual-key permissions despite appearing in Models
  • #8976 - /bedrock/model/{id}/invoke drops tool_reference blocks inside tool_result, so deferred tools are never expanded
  • #9276 - Add Kimchi as a first-class provider
  • #9579 - bifrost_upstream_latency_seconds includes provider queue wait
  • #9591 - /anthropic/v1/messages injects web_search_tool_result after a deferred server_tool_use, so the mixed-tool continuation 400s
  • #9721 - Anthropic→OpenAI conversion drops role on messages containing mcp_tool_use/mcp_tool_result blocks
  • #9731 - Mistral: every reasoning_effort is rewritten to “high”
  • #9788 - ElevenLabs speech with_timestamps returns null char_start_times_ms / char_end_times_ms
  • #9846 - Add option to show stop reason as a column
  • #10026 - prompt_cache auto_inject checks r.Model, not the canonical model
  • #10132 - Updating the budget of a config.json virtual key fails with “invalid reset duration format”

📀 Base OSS version

transports/v2.6.0

🔌 If you are compiling plugin against this release - use following deps