11.27.4

Stabilize test tooling under parallel load: scope the idle watchdog to test steps, label signal-killed steps, and add an opt-in CHECK_LOW_RESOURCE e2e throttle.

11.27.3

Prevent API process crash by routing fire-and-forget auth verification email sends through sendAuthEmailSafely, so a failed send (SMTP down/delivery error) is logged instead of surfacing as an unhandled rejection.

11.27.2

Fix check.mjs hiding devDep vulns for library packages by running the chain's own audit command verbatim.

11.27.1

Introduce scripts/check.mjs: discovers each workspace project's check chain, runs steps with live status, and reports metrics per step.

11.27.0

Expose build identity (commit SHA + version) via health check so drifted containers are detectable after partial rollouts

11.26.3

Prevent broken unnamed enum $ref in OpenAPI output by skipping swaggerOpts.type when an enum is set.

11.26.2

Add COOKIE_PREFIX env to isolate session cookies on shared hosts, with cross-layer lockstep resolver and e2e regression guard

11.26.1

Expose roles in REST auth response so clients can gate admin areas without a second round-trip

11.26.0

AI Assistant Module: a fully opt-in, multi-tenant-capable, OWASP-aligned assistance system with DB-backed LLM connections, role-filtered tool registry, governed safety mechanisms, and MCP server integration — introduced on branch feature/ai-module.

Core Features:

• DB-backed LLM connections with AES-256-GCM encrypted API keys (admin CRUD, hasApiKey output, production-secret guard)
• Provider abstraction (OpenAiCompatibleProvider, ClaudeCliProvider) with auto-detection of supportsNativeTools / supportsJsonResponse / contextWindow
• Prioritized connection-resolution chain (global default → tenant default → user default → client selection → tenant/admin enforced)
• Role-filtered tool registry with mutating / destructive flags and pre-flight authorize() hooks
• Plan mode with all-or-nothing pre-flight authorization of every planned step
• Confirmation policy (mutating.default/enforced, destructive always-confirm, persistent tool grants "remember my choice")
• Scoped tool policies (deny / ask / allow against tool arguments via regex)
• Lifecycle hooks (PreToolUse / PostToolUse / SessionStart / Stop)
• Token budgets per user AND per tenant (day / month / none reset windows, cumulative usage report, HTTP 429 with i18n translation)
• Multi-turn conversations with $push-based message persistence and capped retention
• SSE streaming endpoint (POST /ai/stream) plus REST and GraphQL single-shot
• Multi-modal attachments (image URLs / dataUrls in prompts)
• Admin-defined named agent modes with allowedTools filter and prompt addendum
• Self-optimizing prompts (admin-editable tenant-scoped slots with override/reset, fragment-based builder, soft-delete via enabled: false)
• Governed learning loop (prompt hints from tool failures, admin-approval-gated or autoApply)
• User-facing prompt templates (scope: user|tenant, owner-only mutations)
• Runtime placeholder registry ({{userId}}, {{roles}}, {{tools}}, …, project-specific placeholders dynamically registrable)
• LLM-driven context compaction on context-window overflow with hard-trim fallback
• Deferred tool schemas plus built-in search_tools meta-tool for large tool catalogs
• Built-in ask_user_question tool for interactive clarification
• Audit logging into aiInteractions (admin-readable, prerequisite for budgets)
• MCP server (/ai/mcp Streamable HTTP) with lazy-loaded @modelcontextprotocol/sdk and 503 fallback when the SDK is missing
• OAuth 2.1 for MCP clients (HMAC-SHA256 access tokens with timingSafeEqual, PKCE S256-only, refresh-token rotation bound to client ID, dynamic client registration with persisted client_secret)
• SSRF allowlist for connection base URLs (ai.allowedBaseUrlHosts)
• Per-user rate limiting (max / windowSeconds)
• Three-layer security model (@Restricted / @Roles / securityCheck) with global stripping of apiKeyEncrypted and all token fields
• Multi-tenancy isolation on slots, prompts, hints, budgets, and conversations
• Full override hooks for every collaborator via ICoreModuleOverrides.ai

11.25.6

Bump dependencies and security overrides; disable BetterAuth user-cache in Playwright/lt-dev E2E for fresh roles