Skip to main content
​
v2.1.0

​
Changelog

v2.1.0 moves the enterprise gateway onto OSS transports v2.1.0 and ships four governance changes at once. Projects give a named, budgeted scope that a request opts into per call, with per-member budget and rate-limit shares. Virtual MCPs replace MCP Tool Groups and are wired into access profiles and projects. Business Units become a property of the user instead of the team, and Data Access Control can now be set per RBAC resource on a role. Users can hold several access profiles at the same time, with scheduled automatic key rotation and a toggle that governs keys members create themselves. The OSS base adds the Databricks and GitHub Copilot providers, embedding-based complexity routing, prompt-cache auto-injection, and a large set of streaming and reasoning fixes.

​
⚠️ Breaking Changes

  • Business Units Move from Teams to Users - A business unit is now assigned to users directly through a user-to-BU table, not through the user’s team. SCIM and OIDC sync write BU membership onto the user, DAC derives the caller’s BU from the user, and the Teams page renames “Members” to “Users”. Existing team-level assignments are copied onto every current team member by migration. The legacy business_unit_id column on teams is no longer read or written but is not dropped in this release.
  • MCP Tool Groups Replaced by Virtual MCPs - The tool-group routes are deprecated and RBAC grants on MCPToolGroups are copied onto a new VirtualMCPs resource. mcp.tool_groups in config.json is deprecated in favour of mcp.virtual_mcps. Existing tool groups keep working through the deprecated routes until they are migrated.
  • Access-Profile-Managed Keys No Longer Carry a Governance Mirror - Provider configs, budgets, rate limits, and MCP configs that were copied onto access-profile-managed virtual keys are stripped by migration, and the SyncVKFromAccessProfile propagation path is removed. Governance for those keys is resolved from the profile at request time, so edits to a profile take effect without a propagation job.
  • User-Scoped Model Configs Retired - Creating model configs with scope: user is no longer possible. Existing rows that uniquely match an access profile’s declared per-model budget or rate limit are re-stamped to scope: access_profile; unmatched rows are left in place and still enforced.
  • Routing Metadata and Governance Error Codes Renamed (OSS) - The routing field on responses and log rows is renamed from routing_debug to routing_metadata with no alias. Governance error codes virtual_key_not_found and virtual_key_blocked are renamed to access_not_found and access_blocked. Update any consumer matching on those names.
  • allow_on_all_virtual_keys Renamed (OSS) - The MCP client field is now allow_by_default (the old key is still accepted) and the list filter all_virtual_keys becomes allowed_by_default. Default-allowed MCP clients are granted to access profiles and keyless users unconditionally.
  • Plugin version Field Removed (OSS) - Plugin sync is now driven by a SHA-256 hash of the config entry. A leftover version key in config.json, the API, or Helm values is ignored.
  • Project Delete force Flag Removed - Deleting a project always drops its members with it.

​
✨ Features

  • Projects - A named, budgeted governance scope that a request opts into with the x-bf-project-id or x-bf-project-name header. A project carries an access rule, members, budgets, rate limits, provider configs, MCP configs, virtual MCPs, and allow_all_providers. Members can receive equal-split or explicit shares of each project cap, and shares are redivided by a background job when membership changes or a user is deleted. Projects are managed through /api/governance/projects, reconciled from governance.projects in config.json with hash-based sync, gossiped across the cluster, scoped by DAC, and audited. A request naming a project it cannot use is refused with access_blocked. project_id and project_name land on logs, MCP tool logs, span attributes, metric labels, and the BigQuery, Datadog, and Splunk connectors, with a project_ids filter on user rankings. The UI adds a Governance → Projects page with overview, budgets and limits, provider, MCP, and Members tabs, per-member cap meters, and warnings when an equal split leaves a member without a whole unit.
  • Virtual MCPs Across the Enterprise Surface - Virtual MCPs get a dedicated RBAC resource (with grants migrated from tool groups), DAC scoping, cluster gossip, and an Access Profiles section. They can be attached to access profiles (replacing the per-server tool overrides, which are folded into per-client allowlists) and to projects, and are resolved from the live governance cache at request time. Direct MCP servers gain an endpoint_slug and are served at /mcp/<slug>. config.json and Helm accept mcp.virtual_mcps, and virtual MCP e2e coverage is added.
  • Per-Entity Data Access Control - A role’s data visibility can now be set independently for each scopable RBAC resource (logs, MCP logs, virtual keys, RBAC, and the rest) instead of one global default. Overrides are stored in a new role DAC table, fail closed to own-data on invalid values, and are editable from a Data Access tab on the role sheet with an effective-scope badge on the roles table. The role API, me/permissions, and config.json sync carry the overrides. Docs
  • Optimistic Concurrency on Role Updates - Roles carry a revision counter that every mutation increments inside one transaction. Sending the loaded revision turns an update into a compare-and-swap; a stale write is refused with HTTP 412 and the UI reloads the latest role and asks the editor to reapply. Omitting the revision keeps last-writer-wins for older clients. Docs
  • Business Units per User - Users can belong to several business units, each edge carrying its source (manual or attribute_mapping). SCIM group push and OIDC claims sync are authoritative for their own edges, with delta removes, stale-edge pruning, and a live per-user BU roster endpoint. Group push now runs the same side effects as team sync (claims mirror, role recompute, governance and access-profile reload), and BU changes are broadcast so peers refresh DAC. The Business Units UI is rebuilt as a per-user property.
  • Multiple Access Profiles per User - A user can hold several active access profiles at once. Each holder is stamped with its source (manual, role_default, or attribute_mapping), sources are ranked so an attribute-mapping row is adopted as the role default when the role grants the same template, and role-detach only revokes role-default holders. A role can grant more than one profile and a SCIM attribute mapping rule can target several profiles. At request time exactly one profile pays per attempt (a payable profile wins over an exhausted one), rate limits are unioned across all live profiles, and provider-candidate exclusion mirrors the same logic. The user detail sheet is restructured into tabs with expandable profile cards, a Customers tab, team and BU deep links, and confirm dialogs for membership removal. Docs
  • Scheduled Virtual Key Rotation - Access profiles gain auto_rotation_interval (1h to 365d, off by default) with next_rotation_at and read-only last_rotation_at. A cluster-deduplicated background job rotates managed keys in batches, honours the rotation cooldown, posts dashboard notifications, and advances the schedule; a manual rotation inside the window is respected. Docs
  • Govern Member-Created Virtual Keys - The “Govern virtual keys created by members” behaviour is now an RBAC operation, VirtualKeys:CreateStandalone, granted by migration to every role that already holds VirtualKeys:Create. Members without it have keys they create adopted into their active profile (budgets, rate limits, provider and model scope, MCP grants applied and the key locked); if adoption fails the key is rolled back. GET /api/governance/users/me/vk-creation-policy reports the outcome ahead of time so the create form can lock fields, and auto-provisioned keys are named after the user’s display name or email with a short unique suffix.
  • Allow All Providers on Access Profiles and Projects - allow_all_providers on an access profile or project grants every configured provider, including ones added later, without listing them in provider_configs. Explicit provider entries still apply their model lists. Exposed in the UI, the API, config.json, and Helm, with e2e coverage.
  • Access-Profile-Scoped Model Limits - Per-model budgets and rate limits materialized from an access profile are stored under scope: access_profile keyed by the user’s profile copy, tagged with a structured SourceRef naming the owning profile, and rendered read-only in the UI with a notice naming the profile. User access profiles can serve live usage with from_memory=true, per-model budget lines cascade on delete, and legacy user-scoped limits are still gathered for enforcement.
  • Virtual Key Reveal and Copy Audit - Revealing or copying a key value in the dashboard fires GET /api/governance/virtual-keys/{id}/reveal or /copy, which records an access audit event with the target key and returns 204. The two paths let operators tell a reveal from a clipboard copy when filtering audit logs. Audit logs also gain server-side sorting with URL-persisted sort state.
  • Guardrails: Conversation Window and Native Redaction - Guardrail rules gain an explicit send_all_conversation_turns toggle; when off, “Previous conversation turns to include” counts only preceding turns, fixing an off-by-one in the window. Redaction now rewrites native raw passthrough requests (Claude Code style traffic), synchronizes native SSE stream rewriting with the normalized redaction output, and supports multi-target Responses stream redaction with per-field segment state. The CrowdStrike guardrail receives the user id and name. Docs
  • Cluster: Dead-Node Archive and Config Gossip - Usage counters of dead nodes are archived to the config store instead of being held in memory forever, bounding governance memory on long-lived clusters; a node’s own unflushed usage is preserved when it archives a peer. Complexity-analyzer config changes are gossiped so every node refetches the authoritative config. MCP client updates are broadcast even when the enable or edit dial fails, and peers acknowledge committed-but-dial-failed messages instead of retrying forever. Docs
  • Kubernetes Discovery Hardening - Peer discovery only returns pods that are ready and not terminating, so an address about to be recycled is never handed out. At startup the node checks the configured k8s_namespace against the namespace projected by the service account and verifies that its own pod matches k8s_label_selector, refusing to join another deployment’s pods when a config was copied between environments. A missing namespace falls back to default with a warning, a missing selector warns that every pod in the namespace will be treated as a peer, and the local pod lookup is bounded by a timeout. Docs
  • Per-Instance MCP Failure Reasons - The node-state heartbeat carries each instance’s last MCP failure alongside its connection state, so the aggregate breakdown in the server sheet shows why each pod failed even when every pod agrees on Unstable. The KV key prefix is bumped so mixed-version rolling upgrades degrade cleanly.
  • Tracing for Async Jobs and Overhead Span Toggle - Batch and video settlement emit spans from the async sweeper, and the Datadog, Kafka, and Pub/Sub plugins gain an export_overhead_spans toggle so internal overhead spans can be kept out of the export.
  • SCIM Wizard Refinements - Group role mappings accept a wildcard * value, the trusted-networks trigger moves into the provider-config step for self-hostable providers, the breadcrumb shows the selected provider name, and role profiles are reconciled when a token carries no access-profile mapping attribute.
  • MCP Self-Service Routes on View Permission - Session re-auth and per-user header flow routes are gated on MCPGateway:View instead of the create and update permissions, so view-only roles can re-authenticate their own sessions without MCP client management rights.
  • Edge License Enforcement on Agent Token Endpoints - The agent token and refresh endpoints now check license expiry on every call and return HTTP 402 with the message “Edge trial expired” when the Edge trial has lapsed, and the license state re-checks expiry on every read instead of relying on the async watcher.

​
🌎 Open Source Features

Enterprise-labelled items in the OSS release notes (Projects, scheduled rotation, allow-all-providers, per-instance MCP failure detail, tracing controls) are described above and not repeated here.
  • Databricks Provider - First-class databricks/<model> provider covering Model Serving and Unity AI Gateway with PAT or OAuth M2M auth via databricks_key_config (workspace_url, api_format, client_id/client_secret, forward_gateway_tags). Requests are sanitized per model against datasheet capabilities, reasoning_effort is translated to Anthropic thinking on Claude endpoints, remote images are inlined, native Responses calls fall back to chat emulation, and upstream error messages are surfaced. The UI adds the key form and a guided migration from a custom provider named databricks. Docs
  • GitHub Copilot Provider - github_copilot provider that mints installation tokens server-to-server from GitHub App credentials (github_copilot_key_config: app_id, installation_id, repository_id, private_key, optional github_domain), supporting chat completions, Responses, and list models with cost tracked in GitHub AI Credits. Configurable through config.json and the API only; it is hidden from the Add Provider picker pending release testing.
  • Semantic Complexity Routing - The keyword scorer in the complexity router is replaced by an embedding-based classifier over three tiers with curated exemplar phrases (backfilled by migration, 750 combined phrase cap), a pluggable vector store including an embedded chromem backend with cross-node warm coordination, an optional LLM classifier fallback (semantic.fallback: llm), and session-aware routing that keeps a session at its highest observed tier. New status, generations, and retry endpoints under /api/routing/complexity-analyzer-*, complexity_* log columns and filters, and routing embedding and LLM request and cost counters. tier_boundaries is deprecated and ignored. Docs
  • Per-Request Grants - Every request settles its identity (virtual key, MCP JWT, WebSocket key, ephemeral secret, GenAI session) onto one resolved access grant that governance checks, charges, and filters with, so checked and billed limits cannot diverge and async jobs, WebRTC relays, and WebSocket upgrades keep their identity. MCP runs a single shared server with tool visibility decided per request by governance admission, and every models listing (including integration routes) is narrowed by resolved access.
  • Virtual Key Rotation Cooldown - New client.vk_rotation_cooldown setting (duration string, e.g. “5m”): after a rotation the previous key value keeps authenticating until the grace window expires. config.json VK sync treats a changed value as an explicit rotation (with console warning) and recognizes the previously rotated-out value as “no change”.
  • Prompt Cache Auto-Injection - New provider prompt_cache block (auto_inject, ttl, cache_control_injection_points) synthesizes cache breakpoints for clients that send none, so agentic clients like Codex stop paying the cache-write rate every turn. Off by default, capped at four markers, never touches caller-supplied markers, overridable per request with x-bf-prompt-cache-auto-inject. Edited from a Prompt Caching tab in the provider sheet and extended to the gpt-5.6 family.
  • Azure DeepSeek Chat Completions Routing - Responses requests to Azure DeepSeek models from coding harnesses are routed to Chat Completions because the DeepSeek Responses endpoint rejects reasoning.effort; models without a Responses endpoint fall back the same way, including on Bedrock Mantle. Controlled by compat.azure_deepseek (default true). The compat plugin also logs every dropped parameter and request-type conversion as structured per-request log entries.
  • Native Passthrough Redaction - Guardrail PII redaction applies to Anthropic Messages and Gemini GenAI passthrough traffic, rewriting only content-bearing fields, and to native SSE streams through a paused-buffer codec that rewrites content_block_delta text before release.
  • Video Job Accounting - Async video generation is billed at settlement: a settler polls jobs to a terminal state, prices from captured params or provider-reported dimensions with resolution-banded per-second rates (480p, 720p, 1024p, 1080p, 4k), records failures at zero, and parks unpriceable jobs for backfill. The batch_jobs table is generalised into a provider job table with kind and params columns, and the log detail sheet gains a Video Details block.
  • Webhook Deliveries Page - GET /api/webhooks/deliveries searches delivery history across all endpoints by endpoint, event, outcome, status class, request or delivery ID, and time window, paginated by delivery group. A Webhooks → Deliveries page adds filters, live polling, manual redelivery, and deep links from each endpoint. Docs
  • Request ID Lookup and Period Comparison in Logs - Logs, stats, and histogram endpoints accept an exact request_id that bypasses the time window; the search box auto-detects a UUID or id: prefix. GET /api/logs/stats?compare_to_previous=true returns the previous period, powering a segmented metric strip with sparklines and change badges.
  • Hidden Request Types - logs_store.hidden_request_types (Helm storage.logsStore.hiddenRequestTypes) hides whole request types such as count_tokens from every log read path without affecting writes, cost recalculation, or access control; shown read-only under Config → Logging.
  • Tool Call Names Filter - Logs gain a tool_call_names column, recorded even when content logging is off, with a matching filter on the logs and histogram endpoints and in the logs sidebar.
  • Served and Canonical Model in Logs - The model the provider actually served is persisted as served_model and shown when it differs from the request, and the logs model column displays the canonical name with the requested name as fallback.
  • MCP Connection Failure Details - GET /api/mcp/clients returns last_failure (stage, message, timestamps) and per-node node_states, OAuth tokens record a status_reason, and the server sheet shows the failure in the state badge popover plus a credential block with scopes, refresh-token presence, and expiry.
  • Scoped Model Limits and Quota Sources - Model configs and quota budgets carry a structured SourceRef naming what governs them, GET /api/governance/model-configs accepts a comma-separated scope, quota responses tag each budget and rate limit with its source and list every contributing rate_limits entry, and read-only scopes render as view-only in the UI. Provider-scoped budgets now participate in load-balancing candidate exclusion.
  • Plugin Config Hash Reconciliation - Plugin sync is driven by a SHA-256 hash of the config entry, so a changed entry syncs automatically. Custom Go plugins can use SecretVar in their config, and a plugin’s created_at survives updates.
  • Tracing Controls - New export_overhead_spans toggle (Helm and the Configure Tracing sheet) controls whether internal overhead spans are exported, converter work is split into finer span buckets, and a Latency and Overhead Breakdown docs page explains the log detail view.
  • Log Level Tabs for Plugin and Routing Logs - Routing decision and plugin logs carry a level and can be filtered by it in the log detail view.
  • UI Improvements - Custom providers whose name collides with a first-party integration prompt a switch, access-profile-managed keys get a fallback creation view, sheets get a refreshed design with sticky headers, the logs page handles small screens, cached and uncached input tokens are broken down in a tooltip, and the Raw JSON tab explains when raw storage is disabled.
  • Helm Chart Updates - Values and schema for projects, Databricks keys, access-profile mappings, VK rotation cooldown, virtual MCPs, allow_all_providers, hidden request types, and guardrail send_all_conversation_turns; the SCIM block renders as-is when enabled: false.
  • Baseten on Hugging Face - Baseten is discoverable as a Hugging Face inference provider.
  • Magic Hour in MCP Library - Magic Hour is added to the MCP library.

​
🐞 Fixed

  • Dead-Node Archive Loses Local Usage - Archiving a dead node moved the local baselines and current usage by the same amount through the store’s CAS loops, so this node’s own unflushed delta is preserved and concurrent request accounting is retained.
  • Access Profile Budget and Rate-Limit Persistence - Budget IDs survive team and profile edits, zero-limit budgets are dropped and invalid rate limits rejected on save, presence flags cover every nullable field, per-model rate-limit IDs are collected before a wipe and provider-budget join rows deleted explicitly to avoid FK violations, and provider rate limits that cap nothing are dropped instead of stored.
  • Access Profile Lifecycle - Cascade delete removes every user copy of a profile, not just the first; orphaned user profiles are purged and the parent FK cascades; scope=access_profile model configs are deleted on force delete; legacy user-scoped model configs are deleted when the user is deleted; virtual keys are reclaimed only when a user’s last profile is revoked; non-manual detach is blocked while a user holds several profiles; the highest-precedence profile is returned in user listings; higher-ranked source rows are skipped during source sync to avoid spurious governance reloads; reconcile errors are logged instead of discarded; and config.json user-copy propagation is deferred to a queued job so it cannot fail before initialization.
  • Access-Profile-Backed Keys Pruned by config.json - A prune guard keeps config.json reconciliation from deleting access-profile-owned virtual keys, and those keys are excluded from direct Virtual MCP assignments.
  • Project Member Shares - Equal-split member adds and cap edits that leave no whole unit per member are refused, member shares are zeroed beside their source on budget reset, equal split now divides provider and model caps by head count, member share IDs are included in calendar-alignment adoption, roster reads and member removal are scoped to the caller’s DAC principal, and redivision after user deletion is queued instead of run inline.
  • Virtual MCP Resolution - Access-profile and project virtual MCPs are resolved from the live cache, project-attached virtual MCPs are deleted on cascade, and the MCP and Virtual MCP loaders are consistent in the UI.
  • Complexity Routing Dependencies - The complexity router’s dependencies are wired at enterprise bootstrap, so the semantic classifier and vector store are available on the enterprise build.
  • Inspect Handler Identity - The inspect handler settles identity onto the grant so pre-hooks are not refused as unsettled requests.
  • SCIM and Sessions - Logout no longer leaves a stale session, BU group push now mirrors claims, recomputes roles, and reloads governance like team sync does, and user BU changes are broadcast so peers refresh DAC.
  • Guardrail Reasoning Blocks - Reasoning content stays immutable across every redaction path.
  • Governance Usage Snapshot - A focused usage snapshot replaces the full-state snapshot on hot paths.
  • UI Fixes - Sheet header padding, access-profile badge and truncated-badge tooltips, “+N” chips for budget and rate-limit cells, container-query mapping editors and a chip-based profile multi-select, attribute-mapping button overflow, a read-only notice naming the owning profile with model budgets re-synced on calendar-aligned flips, Projects styling and breadcrumbs, the per-user profile list invalidated on detach, the redundant “Managed by” prefix removed, humanized entity names in the DAC override list, and a free-text BigQuery location input.
  • Streaming Hangs and Connection Leaks (OSS) - A patched fasthttp fixes a race when closing streams, abandoned streams are drained in the background so the upstream connection returns to the pool, streams that send only heartbeats after finish_reason now terminate, and a does_not_send_done_marker toggle on custom providers ends the stream at finish_reason for upstreams that never send [DONE].
  • DeepSeek Reasoning Lost on Multi-Turn (OSS) - Assistant reasoning_content is aliased instead of stripped for Groq and Cerebras, so thinking survives multi-turn requests on the OpenAI-compatible inbound.
  • Reasoning Summary Stream Events (OSS) - summary_index, summary text, and signatures are populated on reasoning_summary_* events for Anthropic, Bedrock, and Gemini.
  • Responses-to-Chat finish_reason (OSS) - Chat finish_reason is derived from the Responses terminal state and incomplete details instead of being dropped.
  • Anthropic Stream Truncation (OSS) - response.incomplete and response.failed emit proper message_delta, message_stop, and error events instead of truncating the stream, and a missing terminal text suffix is synthesized from output_text.done.
  • Anthropic-on-Vertex Passthrough (OSS) - Usage and stream terminal detection for Anthropic models in Vertex GenAI passthrough mode go through the Anthropic parsers.
  • Anthropic Passthrough for Non-Claude Models (OSS) - Raw-body passthrough is cleared based on the resolved provider and model pair, so non-Claude models on Vertex, Azure, and Bedrock Mantle are converted instead of forwarded as Anthropic payloads, and also when the provider does not support the output config format.
  • Unsupported Reasoning Signature (OSS) - The encrypted reasoning signature is stripped when the upstream reports the field as unsupported, such as Bedrock Converse replaying a Claude signature onto a non-Anthropic model.
  • Bedrock Reasoning Blocks (OSS) - Unsigned reasoning blocks are dropped from Converse replays to Claude, while Nova and MiniMax keep receiving them, and native Grok or OpenAI reasoning summaries on Converse responses are rendered instead of dropped.
  • Bedrock Null Content on Empty Assistant Messages (OSS) - An assistant message with no text and no tool calls no longer serializes as content:null, which Converse rejected.
  • Bedrock Model Routing to Converse (OSS) - Bedrock models route to the Converse API as intended.
  • GenAI Signature Drop (OSS) - Standalone thoughtSignature parts with empty text are no longer dropped on native GenAI.
  • Ollama max_tokens (OSS) - Ollama receives max_tokens instead of the unsupported max_completion_tokens.
  • Cohere Rerank Documents (OSS) - Rerank documents are sent as Cohere v2 strings rather than objects.
  • Nullable Response Fields (OSS) - Spec-required nullable response fields are marshalled as null instead of omitted.
  • Model Arrays from OpenAI-Compatible APIs (OSS) - Top-level arrays returned by OpenAI-compatible model listings are accepted.
  • GPT-6 Astra Reasoning Effort (OSS) - Max reasoning effort is preserved instead of being downgraded to high.
  • Forced Tool Choice (OSS) - Anthropic tool_choice: any maps to required on OpenAI egress, gated on the provider capability flag.
  • Azure Reasoning Efforts (OSS) - Reasoning effort handling for Azure-hosted models.
  • Thinking Block Modification Error (OSS) - Replayed thinking blocks no longer trigger a modification error.
  • Custom Provider in List Models (OSS) - Custom providers are skipped in list models when the request is not allowed to use them.
  • Vertex GenAI Model Names (OSS) - Vertex GenAI resource model names are normalised to bare IDs for governance and key selection.
  • Allowed Models Wildcard (OSS) - allowed_models: ["*"] no longer returns model_blocked when the live list-models store is empty for a provider.
  • OpenRouter Prompt Caching on Responses (OSS) - cache_control breakpoints are translated correctly for OpenRouter Claude models on the Responses API.
  • Plugin Config Reverted on Restart (OSS) - Plugin config edited via UI or API is no longer reverted from config.json on restart under source_of_truth: split.
  • Realtime Observability and Auth (OSS) - WebSocket Responses turns emit llm.call spans, realtime auth survives KV replication, and realtime WebSocket and WebRTC routes honour enforce_auth_on_inference.
  • Budget State Preserved Across Edits (OSS) - Changing a budget’s reset frequency or fiscal-quarter setting no longer resets accumulated usage or drops quarter_start_month, new model budgets start empty, and budget IDs survive edits.
  • Routing Rule Persistence (OSS) - Stale routing rules are deleted inside the merge transaction to avoid priority collisions, and rule reads honour the row-visibility query scope.
  • created_at Preserved on Sync (OSS) - created_at survives config sync and updates for budgets, rate limits, teams, customers, model configs, pricing overrides, routing rules, and plugins.
  • Vault Encryption Deadlocks at Boot (OSS) - Rows are encrypted one per transaction with cursor pagination and concurrent vault writes, preventing deadlocks and boot hangs.
  • MCP Client Deletion (OSS) - Legacy FK constraints on oauth_user_tokens and oauth_user_sessions are dropped so deleting an MCP client no longer fails, and the client ID is resolved before vault hooks run.
  • SSRF Hardening for MCP (OSS) - Unauthenticated callers cannot register stdio MCP clients or private addresses, all MCP HTTP clients dial through the SSRF guard, and the Teredo prefix is blocked.
  • Rate Limits on Model-less Passthrough (OSS) - Rate limits apply to passthrough requests that carry no model.
  • Redis Semantic Cache (OSS) - Hex value fields are handled and the score filter is removed from the Redis store.
  • Prompt Child Scoping (OSS) - Prompt child reads and writes are scoped to their parent prompt.
  • Billed Usage on Failed Requests (OSS) - Tracing emits billed token and cost attributes on failed requests.
  • File Response MIME Type (OSS) - File responses carry the MIME type.
  • Logs Filter Search Case (OSS) - Filter data search is case-insensitive on SQLite, Postgres, and ClickHouse.
  • UI Fixes (OSS) - Logout no longer cascades into 401s, virtual key loading state is consistent, managed VK state uses the server flag, and background polling pauses while an edit sheet is open.
  • Dependency and Security Updates (OSS) - Dependabot and CodeQL fixes across modules.

​
🗄️ Database Migrations

Enterprise config store:
  • ent_add_projects_tables - Creates enterprise_projects and its child tables (provider configs, MCP configs, members, member budget shares, member rate-limit shares) plus the budget join tables, with cascading FKs. Rollback drops every project table and discards all projects.
  • ent_add_projects_rbac_resource - Seeds the Projects RBAC resource with Create, View, Update, and Delete; Admin gets all four, Developer and Viewer get View. Rollback is a no-op.
  • ent_add_project_virtual_mcps_table - Creates the project-to-Virtual-MCP link table. Rollback is a no-op so assignments are never dropped.
  • ent_add_project_allow_all_providers_column - Adds allow_all_providers (default false) to enterprise_projects. Reversible.
  • ent_add_access_profile_allow_all_providers_column - Adds allow_all_providers (default false) to enterprise_access_profiles and enterprise_user_access_profiles. Reversible.
  • ent_add_access_profile_auto_rotation_columns - Adds auto_rotation_interval_ns, next_rotation_at, and last_rotation_at to enterprise_access_profiles with an index on next_rotation_at. Reversible.
  • ent_add_role_entity_dac_table - Creates enterprise_governance_role_dac keyed by (role_id, resource_id) for per-entity DAC overrides. Rollback drops the table and every override.
  • ent_add_role_revision_column - Adds revision BIGINT NOT NULL DEFAULT 0 to enterprise_governance_roles. Reversible.
  • ent_add_user_business_units_table - Creates enterprise_user_business_units for per-user BU edges. Reversible.
  • ent_backfill_user_business_units_from_teams - Copies each team’s manual BU onto every current team member, then resolves attribute BU mappings against each user’s stored SCIM profile and claims. One-way; rollback is a no-op.
  • ent_relabel_legacy_scim_business_unit_edges - Rewrites edges whose source is the legacy scim literal to attribute_mapping, in batches of 500. One-way; rollback is a no-op.
  • ent_add_user_access_profile_source_column - Adds an indexed source column to enterprise_user_access_profiles. Reversible.
  • ent_backfill_user_access_profile_source - Stamps role_default on holders whose parent is the user’s current role default and manual on every remaining blank row. Rollback is a no-op.
  • ent_add_virtual_keys_create_standalone_rbac_operation - Adds the VirtualKeys:CreateStandalone operation and grants it to every role that holds VirtualKeys:Create. Rollback is a no-op.
  • ent_drop_access_profile_virtual_key_mirror - Deletes mirrored provider configs, VK-scoped model configs, budgets, rate limits, and MCP configs from access-profile-managed virtual keys. Non-reversible: the mirror was derived state.
  • ent_adopt_legacy_user_model_configs_to_access_profile_scope - Re-stamps scope=user model configs that uniquely match a profile’s declared per-model limits to scope=access_profile; ambiguous rows are left as-is. Non-reversible.
  • ent_ap_per_model_budget_lines_fks_cascade - Recreates four per-model budget line FKs with ON DELETE CASCADE. Rollback is a no-op.
  • ent_delete_orphan_user_access_profiles_cascade_parent_fk - Deletes user profile copies with no parent (and their budget and rate-limit rows), then makes the parent FK cascade. Non-reversible.
  • ent_fold_access_profile_mcp_servers_overrides_into_configs - Folds the legacy per-server MCP override tables into per-client allowlist config tables for profiles and user copies, then drops the four legacy tables. Rollback recreates the legacy tables empty; folded data is not reconstructed.
  • ent_migrate_mcp_tool_groups_rbac_to_virtual_mcps - Creates the VirtualMCPs RBAC resource and copies every grant on MCPToolGroups onto it. Rollback is a no-op.
  • ent_add_governance_node_archives_table - Creates the dead-node governance archive table. Rollback drops the table and any archived usage.
  • ent_add_guardrail_rule_send_all_conversation_turns_column - Adds send_all_conversation_turns to enterprise_guardrail_rules and sets it false for rules with max_turns_to_send > 0. Reversible.
OSS config store (from transports v2.1.0):
  • backfill_default_complexity_exemplars_v2 - Rewrites the complexity_semantic_config governance row, appending curated default exemplar phrases and seeding the semantic row on pre-split installs. Non-reversible.
  • add_vk_rotation_cooldown_columns, add_vk_rotation_cooldown_client_column - Adds previous_value, previous_value_hash, previous_value_expires_at, rotated_at, and an index to governance_virtual_keys, and vk_rotation_cooldown_ns to config_client. Reversible.
  • drop_legacy_oauth_user_fk_constraints - Drops the MCP client and virtual key FK constraints on oauth_user_tokens and oauth_user_sessions. Reversible unless orphan rows accumulated.
  • add_virtual_mcp_tables, add_mcp_client_endpoint_slug - Creates the virtual MCP and VK-to-virtual-MCP tables, adds endpoint_slug to config_mcp_clients, and backfills unique slugs. Non-reversible.
  • add_video_resolution_pricing_columns - Adds resolution-banded video per-second rate columns to model_pricing. Non-reversible; additive.
  • add_provider_job_kind_columns, swap_provider_job_indexes - Adds kind (default batch) and params to batch_jobs and swaps the identity and sweeper indexes to include kind, concurrently on Postgres. Reversible only while no non-batch jobs exist.
  • add_compat_azure_deepseek_column - Adds compat_azure_deepseek to config_client, true on existing rows. Reversible.
  • clear_plugin_config_hashes - Blanks config_hash on every plugin row so hash-based reconciliation starts clean. Rollback is a no-op.
  • add_mcp_oauth_token_status_reason_column - Adds status_reason to mcp_oauth_tokens. Reversible.
  • add_databricks_key_config_columns - Adds the five databricks_* key columns. Reversible.
  • add_github_copilot_config_columns - Adds the five github_copilot_* key columns. Non-reversible; additive.
  • add_allow_all_providers_to_virtual_key, backfill_vk_allow_all_providers_hash - Adds allow_all_providers (default false) to governance_virtual_keys and recomputes every VK config_hash. Column add is reversible; hash rollback is a no-op.
  • add_prompt_cache_json_column - Adds prompt_cache_json to the provider table. Reversible.
OSS log store (from transports v2.1.0): nine additive migrations (logs_add_complexity_routing_columns, logs_add_session_id_column, logs_add_routing_metadata_column, webhook_deliveries_add_filter_indexes_v1, logs_add_video_debug_column, logs_add_project_columns, mcp_tool_logs_add_project_columns, logs_add_served_model_column, logs_add_tool_call_names_column) add nullable columns to logs and mcp_tool_logs and filter indexes on webhook_deliveries. No backfill, no data rewrite. All reversible.
This release adds new columns to the log store. Each ADD COLUMN takes an ACCESS EXCLUSIVE lock on logs, the highest-volume table, and on Postgres the migration waits at most 5 seconds for that lock before failing the boot and retrying on the next one. Upgrade during a low-activity window so the lock is acquired immediately and no queries queue behind it.
To apply the log store schema ahead of the upgrade, run the statements below against the log store database. They match what the migrator executes, and every statement is idempotent. After the DDL you must also record the nine migration IDs in the migrations table (shown after the SQLite block) so the next boot treats them as applied.Postgres:
SQLite (no IF NOT EXISTS on ADD COLUMN; skip any column that already exists):
Then record the migrations in the log store’s migrations table (Postgres and SQLite). Run in this order. On SQLite use CURRENT_TIMESTAMP instead of NOW():
ClickHouse has no migration ledger; Bifrost reconciles missing columns on boot with ADD COLUMN IF NOT EXISTS (add ON CLUSTER when configured):

​
🐙 Closed OSS Issues

  • #2765 - Bedrock provider does not sanitize empty content blocks (regression from #1189 fix)
  • #5887 - DeepSeek thinking silently lost on ALL multi-turn requests via OpenAI-compat inbound (v1.6.7; regression from v1.6.3)
  • #6073 - GenAI passthrough in Vertex mode breaks Anthropic models
  • #6132 - Ollama provider: max_tokens / max_completion_tokens silently dropped from forwarded request
  • #6143 - data race - fasthttp requestStream released to pool while SSE reader is still inside Read (stream cancellation)
  • #6180 - explicit prompt cache for Bedrock Mantle GPT-5.6 Responses
  • #6265 - Realtime/WebSocket Responses turns produce no llm.call span, so span-based observability connectors export them unattributed
  • #6290 - OpenRouter Claude prompt caching remains broken on Responses API
  • #6434 - Plugin config edited via UI/API is reverted from config.json on every restart under source_of_truth: split
  • #6624 - Bedrock reasoning signature field is dropped for Anthropic models, which require it present
  • #6631 - add Baseten to Hugging Face inference providers
  • #6640 - v2.0.0 rerank sends documents as objects to Cohere-based custom providers, breaking servers that expect Cohere v2 strings
  • #6657 - Fireworks virtual key with allowed_models: [”*”] blocks every model (empty synced catalog; explicit list works)
  • #6689 - Responses omit spec-required nullable fields
  • #6690 - [MCP Library] Add: Magic Hour
  • #6711 - Support array responses from OpenAI-compatible model APIs
  • #6730 - Native GenAI drops empty text from standalone thoughtSignature parts
  • #6775 - OSS ui typecheck fails since #6618 (VKCreationPolicyResponse missing from fallback types)
  • #6784 - Chat completion stream hangs forever after finish_reason when upstream omits [DONE] but keeps sending heartbeats
  • #6831 - Responses-to-Chat mux drops non-streaming finish_reason
  • #6880 - GPT-6 Astra max reasoning effort is silently downgraded to high
  • #6887 - Anthropic tool_choice {type: any} forwarded to OpenAI as “any” instead of “required”
  • #6914 - Team current spend is reset after adjusting budget limit even when choosing Preserve Usage

​
📀 Base OSS version

transports/v2.1.0 (pinned as github.com/maximhq/bifrost/transports v1.6.12-0.20260908151718-096d5c8b3736), with core v1.8.5, framework v1.6.1, governance v1.7.1, and logging v1.7.1.

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