Documentation Index
Fetch the complete documentation index at: https://docs.getbifrost.ai/llms.txt
Use this file to discover all available pages before exploring further.
Changelog
A focused patch on top of v1.4.1. Headline items: Entra app-role propagation in SCIM provisioning (App Registration roles are now fetched via Microsoft Graph and attached to each provisioned user, so role-based RBAC mappings finally work end-to-end),RequiredHeaders and RoutingChainMaxDepth plumbed into the governance plugin loader, and security dependency bumps for four open Dependabot alerts. Picks up an OSS base sitting just past transports/v1.5.2 (commit bef816a), which adds Chat ↔ Responses response_format round-trip conversion, separate push/pull telemetry toggles with a plugin hot-reload fix, MCP tool-manager config-hash sync, plus a batch of Gemini, Bedrock, Anthropic, and compat-plugin fixes.✨ Features
RBAC & Identity
- Entra App-Role Propagation in SCIM Sync - App Registration roles configured on the Entra app are now fetched via Microsoft Graph and attached to each provisioned user (both direct user assignments and group-member assignments). Role values populate the user’s
Rolesset so Entraroles-claim based role mappings can be matched against actual app-role assignments instead of being unverified. - Entra Role Filter Options in Sync UI -
FetchFilterOptionsnow returns the enabled App Registration roles (display name, value, description) as selectable role filters, removing the previous placeholder where Entra role options were always empty. Application.Read.AllProbed at Setup -CheckGraphPermissionsnow probes the App RegistrationappRolesendpoint and surfaces a warning if the configured client lacksApplication.Read.All, so admins find out at config time rather than during a silent role-sync failure.- Group-Member Fetch Memoization -
fetchUsersFromAppAssignmentscaches group-member lookups per group ID within a single sync pass, avoiding repeated Graph round-trips when multiple app-role assignments target the same group.
Governance
RequiredHeadersandRoutingChainMaxDepthWired to Governance Plugin - Both fields existed on the enterprise client config but were not reaching the governance plugin loader; they are now forwarded so header-based identity requirements and routing-chain depth limits take effect in governance evaluation.
OSS Base (post-transports/v1.5.2)
- Chat ↔ Responses
response_formatRound-Trip -response_format(json_object,text,json_schemawith fullname/description/strict/schema) is now bidirectionally mapped totext.formatbetween the Chat and Responses request schemas; previously it was silently dropped on conversion. A newSupportsResponseSchemacapability flag in the model catalog registers bothresponse_formatandtextas supported parameters when set. - Separate Push/Pull Telemetry Toggles + Plugin Hot-Reload Fix - Push-based and pull-based telemetry can now be toggled independently, and the plugin hot-reload path is fixed.
- MCP Tool Manager Config in Client Hash -
mcp.tool_manager_configis now included in the client config hash and synced on reload, so changes to tool-manager config are picked up by reload-driven flows. - Mistral Reasoning Effort in Model Catalog - Mistral reasoning effort entries added to the model catalog.
🐞 Fixed
Security & Dependencies
- CVE Bumps for Dependabot Alerts - Bumped
aws-sdk-go-v2/service/bedrockruntimev1.50.1 to v1.50.6 (eventstream DoS panic),ulikunitz/xzv0.5.12 to v0.5.15 (memory leak on corrupted LZMA),nwaples/rardecode/v2v2.1.0 to v2.2.2 (RAR dictionary DoS), andjackc/pgx/v5v5.9.1 to v5.9.2 (SQL injection via placeholder confusion).
OSS Base (post-transports/v1.5.2)
- Gemini
fallbacksKey Removed from Requests - Thefallbackskey is no longer included in outgoing Gemini requests (it is not a Gemini-native field and was causing schema rejections). - Gemini Fallback Propagation in GenAI - Fixed fallback propagation in the GenAI integration so configured fallbacks are honored.
- Gemini Raw Request Scoping - Raw-request handling is now applied only for the Gemini provider, not leaked across providers.
- Anthropic Advisor Model Passthrough - Fixed prefix stripping in the advisor tool for Anthropic so model passthrough resolves correctly.
- Empty Text / Signature Messages on Bedrock OpenAI - Drops messages with empty text or signature and converts thinking blocks for OpenAI Bedrock models.
- Anthropic Trailing Assistant Message Drop + Mistral Reasoning Effort Conversion - Drops the last assistant message for Anthropic models and converts unsupported reasoning effort values for Mistral.
- Compat Plugin Defaults Enabled - All compat plugin settings are now enabled by default.
- System-Only Message Role Conversion - Converts role
systemto roleuserwhen only a system message is present for non-OpenAI models. - Compat
cachePointDrop - Compat plugin now dropscachePointfor unsupported Bedrock models and non-Bedrock models. - OTEL Plugin Metrics - Resolved issues in the OTEL plugin metrics path.
- AWS SDK Patch Bumps (OSS) -
aws-sdk-go-v2v1.41.5 to v1.41.7 and related internal modules.
📀 Base OSS version
This release pins an OSS commit, not a tagged release. All five OSS modules (transports, core, framework, plugins/governance, plugins/logging) are pinned to commit bef816abe9c2 (2026-05-13), which is transports/v1.5.2 + 3 additional fixes:14df900aChat ↔ Responsesresponse_formatconversion +supports_response_schemacapability flag (#3454)725f2cb7AWS SDK patch bumps (#3461)bef816abRemovefallbackskey from Gemini requests (#3464)
go.mod:
