Skip to main content
​
v2.1.44

​
Changelog

  • Added bifrost.auditLogs.omitIpAddresses (default false) to stop recording client IP addresses in audit logs while keeping the rest of the audit trail. IPs stored before the flag was turned on are hidden from the API, filter data, search and exports, and archive copies written after the change omit them, but they stay in the database and in archive objects already written. Renders into audit_logs.omit_ip_addresses.
  • Added excluded_attributes to the OTEL, Datadog, Kafka, Pub/Sub, Splunk and BigQuery plugin configs (bifrost.plugins.<connector>.config.excluded_attributes, and profiles[*].excluded_attributes for OTEL). Names span attributes to leave out of exported traces, matched exactly (e.g. gen_ai.request.tools), for dropping large attributes that are not needed downstream. Unmatched names are ignored. On BigQuery these are column names (tools) rather than attribute names, because BigQuery writes typed columns; the column stays in the table schema and is left empty, and trace_id / timestamp are rejected. Renders into excluded_attributes.
  • Added bifrost.plugins.bigquery.config.export_raw_payloads and bifrost.plugins.otel.config.export_raw_payloads (also profiles[*].export_raw_payloads), default false. Stores the raw provider request and response bodies. Requires the provider’s store_raw_request_response and is suppressed by disable_content_logging. Previously the field existed in Bifrost but had no Helm path, so it could not be enabled from a declarative install. Renders into export_raw_payloads.
  • Added bifrost.agents[] — declarative A2A agent registrations rendered into the config.json agents section. Each entry takes name, agent_card_url, and optionally tenant, enabled, allow_by_default, forward_accepted_credential, forward_accepted_credential_overrides_auth, discovery_auth, runtime_auth, extension_uris, and virtual_key_ids. Reconciled by name on startup: created if absent, updated when the declaration changes, and left alone otherwise (merge mode). When bifrost.sourceOfTruth is config.json and agents is present, stored registrations absent from the rendered list are removed; set agents: [] to remove all stored registrations.
  • Added bifrost.client.a2aExternalClientUrl (a2a_external_client_url) — Bifrost’s public base URL for served agent cards and A2A push-notification callback URLs. Push notifications stay disabled while unset. Supports env. syntax.
  • Added bifrost.server.a2aGrpcBaseDomain (a2a_grpc_base_domain) and bifrost.server.a2aGrpcPort (a2a_grpc_port) — enable the shared A2A gRPC listener; per-agent hostnames are advertised as <agent-name>.<base-domain>:<port>. gRPC stays disabled while either is unset.
  • Added bifrost.server.a2aAllowPrivatePushCallbacks (a2a_allow_private_push_callbacks) — lets A2A push delivery reach loopback and private network callback targets. Off by default; intended only for controlled development and test environments.
  • Added storage.logsStore.clickhouse.maxQuerySize (renders into logs_store.config.max_query_size), the ClickHouse max_query_size setting in bytes. Defaults to 16 MiB so team-scoped log reads with thousands of member and virtual key ids no longer fail against the 256 KiB server default.
  • Added bifrost.governance.complexityAnalyzerConfig.decision (renders into complexity_analyzer_config.decision): a decision-model classifier for the complexity router. provider and model select the decision model (default typesafe/jev-latest). Supported providers are Typesafe (Jev), OpenAI (gpt-6-luna, also on a custom provider with base_provider_type: openai), OpenRouter (the decision models it serves, such as ~typesafe/jev-latest), Cloudflare (Clef: @cf/cloudflare/clef, @cf/cloudflare/clef-flash), Perplexity (pplx-decider-*), and Ollama (locally pulled System One models: nimble, tev1, clef, clef-flash, laya). Laya and Nimble (by Bespoke) can also be served by a custom provider with base_provider_type: typesafe. Requests go through Bifrost’s /v1/decisions to the selected provider’s native decisions endpoint, using that provider’s own credentials. criteria holds per-tier overrides of the shipped definition, signals, and examples, keyed by SIMPLE, MEDIUM, or COMPLEX (exact case). Any tier or field left out sends the shipped default; a definition is at most 500 characters, and each list at most 12 items of 300 characters.
  • Added the cloudflare provider (Cloudflare Workers AI) to bifrost.providers. Every key needs cloudflare_key_config.account_id (supports env.); the base URL defaults to https://api.cloudflare.com/client/v4. It serves chat completions, Responses, embeddings, model listing, and Clef decisions (@cf/cloudflare/clef, @cf/cloudflare/clef-flash).
  • Added bifrost.mcp.toolManagerConfig.maxInstructionsPerClient and .maxInstructionsTotal (max_instructions_per_client / max_instructions_total) to bound forwarded MCP server instructions in bytes; 0 keeps the built-in defaults.
  • Added access_profile to bifrost.governance.customers[], .teams[], and .businessUnits[] (renders into governance.{customers,teams,business_units}[].access_profile), the enterprise access profile the entity holds in place of its own budgets and rate limit.
  • bifrost.governance.roles[].dac and .entity_dac now accept business-unit-data and customer-data (governance.roles[].dac / entity_dac).
  • Documented ttft_timeout_ms on bifrost.governance.routingRules[].targets[] (governance.routing_rules[].targets[].ttft_timeout_ms) in the values schema.
  • Added bifrost.proxyConfig (renders into the top-level proxy_config): the global outbound proxy from the dashboard’s proxy settings page, including enableForScim (enable_for_scim), enableForInference and enableForApi; reconciled with the same config hash as the other sections.
  • Added bifrost.scim.config.attributeProjectMappings ({ attribute, value, project }, every SSO provider) — renders into scim_config.config.attributeProjectMappings. Every matching rule adds the user to that project by name (projects are never auto-created); memberships a rule added are removed when it stops matching, while members added from the dashboard are kept.
  • Added bifrost.scim.config.bulkSyncInterval (default 24h, every SSO provider) — renders into scim_config.config.bulkSyncInterval. How often the directory reconcile deprovisions users the IdP no longer returns: whole days or h/m/s pairs (12h, 1h30m, 7d), between 1h and 30d. Inert when SCIM is enabled or the provider has no directory API access.
  • Added bifrost.mcp.toolManagerConfig.codeModeLimits (maxSourceBytes, maxSteps, maxMemoryBytes, maxLogBytes, maxToolCalls, maxValueBytes, maxNestingDepth) to tune the limits on each code mode execution; an omitted or 0 field keeps the built-in default. Renders into mcp.tool_manager_config.code_mode_limits. Code mode no longer limits concurrent executions.
  • Added a template-time guard: the chart now fails to render when bifrost.client.mcpServerAuthMode is oauth or both and bifrost.client.oauth2ServerConfig.issuerUrl (client.oauth2_server_config.issuer_url) is empty or unset, instead of letting the pod crash-loop on Bifrost’s startup check. env.VAR_NAME references pass.
  • Added storage.logsStore.objectStorageExcludeRequestTypes (renders into logs_store.object_storage_exclude_request_types): request types (e.g. list_models) whose logs stay in the logs database in full instead of being offloaded to object storage. Hidden logs are still offloaded.
  • Added ignore_provider_cost to bifrost.providers.<name> and bifrost.governance.providers[] (providers.<name>.ignore_provider_cost / governance.providers[].ignore_provider_cost): ignore the provider-reported usage.cost and price the request from Bifrost’s model pricing and pricing overrides.
  • Added storage.configStore.statementTimeout / .idleInTransactionSessionTimeout and the same pair on storage.logsStore (config_store.config.statement_timeout, idle_in_transaction_session_timeout, and the logs_store.config equivalents). Server-side Postgres timeouts on every runtime pool connection, default 120s; "0" keeps the server default.
  • Added bifrost.client.compat.forceReasoningOnlyModelsToResponses (client.compat.force_reasoning_only_models_to_responses, default true): routes chat requests to the Responses API for models that cannot reason with tool calls, so reasoning is kept.
  • Added bifrost.client.deleteExpiredVirtualKeys (client.delete_expired_virtual_keys, default false) and per-key delete_after_expire on bifrost.governance.virtualKeys[]: a daily job deletes virtual keys past expires_at; the per-key value overrides the client setting.
  • Added disable_content_logging and business_unit_id to bifrost.governance.virtualKeys[] (governance.virtual_keys[]): a per-key content-logging override and the owning business unit (mutually exclusive with team_id / customer_id). Grant agents to a key from the agent side with bifrost.agents[].virtual_key_ids.
  • Added bifrost.client.oauth2ServerConfig.allowedRedirectUris (client.oauth2_server_config.allowed_redirect_uris): exact remote OAuth client callback URIs to allow. Loopback and Cursor callbacks need no entry; other remote callbacks are denied by default.
  • Added bifrost.mcp.clientConfigs[].requirePublicTarget (mcp.client_configs[].require_public_target): every connection to the MCP server must resolve to a public address. Once stored as true it is never cleared.
  • Added cache_tool_call_responses to bifrost.plugins.semanticCache.config (semantic_cache plugin config): cache responses that carry tool calls. Off by default, since a replayed tool call runs in another caller’s agent loop.
  • bifrost.governance.routingRules[].fallbacks entries now accept an object { provider, model, provider_key_name } that pins a provider key, alongside the provider/model string form (governance.routing_rules[].fallbacks).
  • Added displayName to the generic OIDC bifrost.scim.config (scim_config.config.displayName): the login button label (Sign in with <displayName>), default SSO.
  • Added topologySpreadConstraints (top-level value): pod topology spread constraints rendered into the Bifrost Deployment or StatefulSet pod spec, to spread replicas across nodes and zones. A constraint without a labelSelector defaults to the release’s server pod labels, so only Bifrost’s own replicas are counted.
  • Added extraContainers (top-level value): regular sidecar containers appended to the Bifrost pod after the Bifrost container. Native sidecars (Kubernetes 1.29+) follow the Kubernetes convention, an initContainers entry with restartPolicy: Always, and the chart now refuses to render one on clusters older than 1.29 instead of leaving the pod blocked on an init container that never exits.