Changelog

All notable changes to this project are documented here. Format follows Keep a Changelog. Versions correspond to Roadmap milestones.

​

[Unreleased]

​

Added

• Docker Compose deployment: Single image (API + Frontend), named volumes for data persistence, standalone Next.js output mode
• Admin: login history recording: Auth endpoints now record every login attempt (IP address, User-Agent, success/failure with reason) to the LoginHistory table for security auditing
• Admin: file browser: New admin file browser endpoints for listing and downloading user-uploaded files with pagination support
• Admin: enriched agent/KB views: Admin agent list and knowledge base views now expose model_name, tools, kb_ids, and embedding_model fields for better visibility
• Chat: sensitive word blocking: Both ReAct and inject_message chat endpoints now enforce sensitive word filtering, blocking messages that contain configured sensitive words
• SSE: structured error handling: onError callback in useSSE hook upgraded from plain string to ApiError object for richer error context (error code, args)
• UI: formatTokens utility: New formatting helper for displaying token counts with locale-aware number formatting
• UI: Checkbox and FadeIn components: New shadcn-style checkbox and fade-in animation wrapper components

​

Changed

• Docs: ASCII → Mermaid diagrams: All box-drawing ASCII diagrams converted to native Mermaid charts (12 diagrams across 6 pages) for better rendering and interactivity
• Docs: dark mode theme: Refined Mintlify color config — warm amber accent in dark mode, gradient background decoration, proper background colors matching project theme

​

Fixed

• Docker i18n broken: Copy messages/ directory into standalone output and add outputFileTracingIncludes so next-intl can discover locale JSON files at runtime
• Docker startup race condition: Add API readiness wait loop (Python socket poll) in entrypoint before starting Next.js, preventing e.map is not a function SSR errors
• Artifacts: sort by conversation created_at: Sort artifacts by conversation creation time instead of file mtime (mtime is unreliable on Docker volumes)
• OAuth: auto-detect production API URL from window.location port: Browser auto-detection now works correctly for custom ports in production (not just 3000)

​

[v0.7.3] - 2026-02-28

​

Added

• Global MCP servers: Admin-provisioned MCP servers loaded in all chat sessions (not per-user/agent); query logic: or_(is_global==True, user_id==current_user.id)
• Backward-compat mode: Admin Platform reads legacy registration_enabled boolean if registration_mode is not set (auto-migrates to “open” mode)
• Structured audit logging: write_audit() helper centralizes audit log writes (admin_id, target_type, target_id, detail fields); replaces ad-hoc AuditLog() calls

​

Changed

• Audit log schema: Moved from raw meta dict to structured admin_id / target_type / target_id / detail columns for better compliance
• OAuth callback URL construction: {API_BASE_URL}/api/auth/oauth/{provider}/callback (auto-constructed by backend, no hardcoding)

​

Fixed

• Invite code backward-compat: Correctly interprets registration_mode when not explicitly set (reads from registration_enabled)
• Multi-user scenario: Global MCP servers no longer conflict with per-user tools

​

[v0.7.2] - 2026-02-21

​

Added

• Invite-only registration mode: New registration system with three modes (open/invite/disabled) and invite code CRUD
• Storage management: Per-user disk usage calculation, clear command, orphan file cleanup
• Per-user force logout: Admin ability to revoke all tokens and force re-authentication
• Conversation moderation: Admin ability to list and delete all user conversations

​

Changed

• Admin panels: Separated into Admin List (users, API keys, agents, etc.) with consistent dropdown-menu action patterns

​

Fixed

• Admin list pagination: All admin data lists now support pagination and search

​

[v0.7.1] - 2026-02-14

​

Added

• API health dashboard: System stats (users, conversations, tokens, model usage charts, tokens-by-agent breakdown), connector call metrics (success rate, latency)
• JWT auth: Token-based SSE auth, conversation ownership, per-user resource isolation
• Admin API: Agent management, per-user token quota enforcement (429 when exceeded)

​

Fixed

• Token quota enforcement: ReAct and DAG endpoints now check per-user quotas before execution

​

[v0.7] - 2026-02-07

​

Added

• Admin Platform v0.7: User management with search/pagination, role toggle, password reset, account enable/disable, invite-only registration, per-user quota, conversation moderation, storage management
• First-run setup wizard: On first launch, portal guides through admin account creation (username, password, email)
• Personal Center: Per-user global system instructions applied across all conversations
• Language preference: Per-user language setting (auto/en/zh) directs all LLM responses to chosen language

​

Changed

• Multi-tenant auth: JWT tokens enforce conversation ownership; agents inherit user context

​

Fixed

• OAuth account binding: Proper email matching for link/login flows across all providers

​

[v0.6.5] - 2026-01-31

​

Added

• Utility tools: email_send, json_transform, template_render, text_utils
• Connector response filtering: CONNECTOR_RESPONSE_MAX_CHARS and CONNECTOR_RESPONSE_MAX_ITEMS limits
• Embedding model options: Support for Jina, OpenAI, and custom embedding providers

​

Fixed

• RAG pipeline stability: Better error handling for failed chunk operations

​

[v0.6] - 2026-01-10

​

Added

• Connector Platform (core feature):
  • Connector CRUD (create, read, update, delete)
  • ConnectorToolAdapter converts Connector → BaseTool in ToolRegistry
  • Per-user credentials with AES-GCM encryption
  • Confirmation gate for write operations
  • Circuit breaker for connection failures
  • Operation classification (read/write/admin)
  • Audit logging for all tool calls
• MCP integration: Tool auto-discovery via MCP protocol, process isolation
• Connector data flow example: Query → Governance → MCP → Legacy System

​

Changed

• Tool architecture: Unified ToolRegistry supports built-in, MCP, and connector tools

​

Fixed

• Tool categorization: All built-in tools properly tagged with categories

​

[v0.5] - 2025-12-20

​

Added

• Full RAG pipeline: Jina embedding + LanceDB vector store + FTS + RRF hybrid retrieval + reranker
• Grounded Generation: Evidence-anchored RAG with [N] citations, conflict detection, explainable confidence scores
• KB document management: Chunk-level CRUD, text search, failed doc retry, auto-migrating vector store schema
• ContextGuard + Pinned Messages: Token budget manager; pinned messages protected from compaction

​

Changed

• Multi-turn memory: DbMemory in core/memory/db.py, CompactUtils in core/memory/compact.py
• LLM Compact: CompactUtils.llm_compact() summarizes old turns via fast LLM
• DAG Re-Planning: Loop up to MAX_REPLAN_ROUNDS=3; re-plans when analysis.achieved=False and confidence < 0.5

​

Fixed

• Token overflow handling: Automatic LLM-powered summarization respects token budgets

​

[v0.4] - 2025-11-15

​

Added

• Multi-turn conversations: DbMemory persists conversation history; smart truncation handles long sessions
• Tool step folding: UI collapses/expands tool call steps for readability
• HTTP request tool: Custom headers, auth, method selection
• Shell exec tool: Safe command execution with timeout
• Agent management: Create, configure, publish agents with bound models, tools, instructions
• Per-agent execution mode: Standard/Planner mode toggle, temperature control
• JWT authentication: Stateless token-based auth for API calls

​

Changed

• Memory architecture: ReAct fully wired to DbMemory; DAG injects history as enriched query prefix
• Admin health check: Monitors both ENV-configured and admin-configured models

​

Fixed

• Multi-user isolation: Conversations isolated by user; agents scoped to user

​

[v0.3] - 2025-10-20

​

Added

• Web tools: web_search (Jina/Tavily/Brave), web_fetch, configurable providers
• File operations tool: Read, write, list, delete with path validation
• MCP client: Protocol-first tool integration, process isolation
• Tool auto-discovery: discover_builtin_tools() in builtin/__init__.py
• Tool categories: computation, web, filesystem, knowledge, mcp, connector, general
• DAG visualization: Interactive flow graph with live status, dependency edges, click-to-scroll
• Python/Node.js execution: Code execution in Docker for isolation (CODE_EXEC_BACKEND=docker)
• Sandbox hardening: --network=none, --memory=256m, timeout enforcement

​

Changed

• Tool interface: All tools now categorized; registry supports filtering
• Execution model: DAG Executor runs steps concurrently via asyncio

​

Fixed

• Tool discovery robustness: Handles missing optional dependencies gracefully

​

[v0.2] - 2025-09-15

​

Added

• Retry & rate limiting: Automatic exponential backoff for transient failures
• Usage tracking: Per-request token/cost accounting
• Native function calling: Direct model support for tool selection (no JSON-only parsing)
• Multi-model support: FAST_LLM_MODEL for cheaper DAG step execution
• Memory system: WindowMemory (latest N turns), SummaryMemory (LLM-summarized), DbMemory (persistent)
• FastAPI backend: /api/execute (ReAct), /api/stream (SSE streaming)

​

Changed

• LLM abstraction: BaseLLM → OpenAICompatibleLLM with unified interface
• ModelRegistry: Centralized configuration, provider auto-selection

​

Fixed

• Error recovery: Automatic retry on transient API errors

​

[v0.1] - 2025-08-01

​

Added

• ReActAgent: Reason → Act → Observe loop with tools
• DAGPlanner: LLM-generated dependency graphs, DAGExecutor for concurrent execution, PlanAnalyzer for result verification
• BaseTool framework: Unified interface for all tools
• Calculator tool: Arithmetic expressions, formula evaluation
• Python exec tool: Code execution in isolated environment
• KaTeX rendering: Mathematical notation in streaming output
• JSON mode parsing: Robust structured output handling
• Portal UI: Next.js web interface with real-time streaming, dark/light/system theme
• Initial documentation: Wiki pages covering philosophy, architecture, roadmap

​

Fixed

• Streaming stability: Proper SSE error handling and connection cleanup