v1.5.3

OpenClaw Gateway integration, Demo Mode, Design Overlay, comprehensive i18n audit, and dependency updates.

Backend.AI GO v1.5.3

104 commits since v1.5.2.

New Features

• OpenClaw Gateway integration with setup wizard, environment detection, container execution mode, messaging channels, skill browser, supervisor approval workflow, and Squad channel bridge (–,,)
• Demo Mode for automated prompt playback with UX polish — navigation, per-step agents, follow-up questions, step reordering, and context cleanup
• Design Overlay mode with Component Inspector, viewport presets, theme/locale switcher, state simulator, and page navigator
• Sticky interactive toolbar for models directory
• Partial model file cleanup and orphan management UI
• CoworkPage AnalyticsView and HistoryView components with favicon assets
• JSON syntax highlighting and readable formatting in tool call blocks
• Base Docker images and build workflow for OpenClaw Gateway and Agent Runner

Improvements

• Comprehensive i18n audit: replaced ~95 hardcoded strings across 47 components with translations in all 6 locales
• Upgrade Flatpak runtime from GNOME 45 to 47
• Rename Safetensors format identifier to Mlx across codebase
• Redesign Claw and Cowork sidebar icons
• Use default cursor for interactive elements in Tauri app mode
• Hide Claw menu in production mode
• 256 new tests for demoStore, lifecycleStore, supervisorStore, and inspectorUtils

Bug Fixes

• Replace hardcoded strings with i18n translations across 47 UI components
• Resolve 102 test failures from store migration to getApiAdapter
• Fix auth handler returning 500 instead of 401 for invalid API keys
• Tool toggle checkbox in agent profile editor does not persist state change
• Add timeout and safety net to update shutdown to prevent hang
• Guard against undefined fields from backend partial responses
• Complete Tauri mock implementation for reliable UI-only development
• Improve diagnostic information for model load failures
• Categorize unsupported model architectures for better error messages
• Normalize safetensors format to mlx in engine display and resolver
• Clear "starting" state after HuggingFace download completes
• Pin wizard navigation buttons to consistent bottom position
• Resolve Windows build errors from windows crate 0.62.2 HANDLE type change
• Prevent double base64 encoding of Windows updater signature
• Resolve mkdocs build errors and webpage high CPU usage

CI/CD Improvements

• Consolidate macOS build and signing into single self-hosted job
• Replace Apple codesign with rcodesign for macOS signing
• Multiple macOS code signing fixes for macOS 26 and CI runners
• Multiple Windows signing and packaging fixes
• Limit Rust build parallelism on Linux x64 to prevent OOM kill
• Use persistent CARGO_TARGET_DIR on Windows self-hosted runner for incremental builds
• Add 7-day auto-pruning for persistent cargo target cache
• Bump GitHub Actions to Node.js 24-compatible versions
• Add source_ref input to packaging workflow for independent code/version selection
• Opt into Node.js 24 for project board automation workflow

Technical Details

• Bump cargo dependencies and unify reqwest to 0.13
• Bump CLI cargo dependencies and replace deprecated serde_yaml
• Fix mkdocs deps for pygments 2.20 crash and edit_uri warnings
• Use inline Jinja2 escaping for Go template syntax in docs

Dependencies

• mlxcel updated to v0.0.13
• Continuum Router updated to v1.4.5
• reqwest unified to 0.13
• Cargo dependencies updated across main app and CLI

Breaking Changes

None

Known Issues

None

Full Changelog

v1.5.2

Apple Container runtime for macOS 26, chat message edit redesign, GGUF quantization grouping, Cowork streaming thinking steps, and enhanced folder management UX.

New Features

• Apple Container runtime support with bgo-container-cli for running agents in isolated containers on macOS 26
• Redesigned chat message edit mode with improved usability and inline editing
• Directory-based GGUF quantization grouping and batch download
• Live status indicators on fallback flow diagram showing real-time node health
• Real-time streaming of Cowork agent thinking steps via SSE delta events
• Save to Creation bookmark button on agent final answer steps
• Rich content rendering in run detail view with Markdown and structured tool calls
• Error detail display on failed Squad agents, tasks, and executions
• Smooth transition animation and analyzing state for Cowork task submission
• Contextual placeholder and tool badges in Cowork task input
• Enhanced Cowork empty state with quick actions, folder summary, and run history

Improvements

• Cowork analytics moved to Statistics page as a unified third tab
• Folder management consolidated with permission controls directly on chips
• Runtime selector relocated to settings drawer
• Default profile label changed from "General Assistant" to "General Purpose"
• Smart folder chip paths with permission indicators
• Staggered fade-in animation and hover lift effect for run history sidebar items
• Run detail metadata redesign with summary stat pills and human-readable model names

Bug Fixes

• Align bgo-container-cli with actual Containerization framework API
• Pass working directory to agent tools and add i18n for section title
• Align Cowork empty state two-column tops and heights
• Resolve elapsed timer stuck at 0:00 and compact Cowork empty state layout
• Add retry logic for transient API errors in agent loop
• Scope agent tool default paths to permitted folders in Cowork mode
• Improve quantization group card display with label, bits info, and column alignment
• Eliminate redundant Rust recompilation in Windows packaging workflow

CI/CD Improvements

• Add bgo-container-cli stub binary for Tauri packaging on macOS 15 CI runners

Technical Details

• Continuum Router updated to v1.4.4
• mlxcel updated to v0.0.11
• Swift 6.2 / macOS 26 Containerization framework integration
• What's New modal content updated for all 6 locales

Dependencies

• continuum-router: v1.4.1 → v1.4.4
• mlxcel: v0.0.10 → v0.0.11

Breaking Changes

None

Known Issues

• bgo-container-cli requires macOS 26 (Tahoe) with Xcode 26+; a stub binary is used on older macOS versions

What's Changed

• feat: add live status indicators to fallback flow diagram by @inureyes
• feat: add directory-based GGUF quantization grouping and batch download by @inureyes
• fix: improve quantization group card display with label, bits info, and column alignment by @inureyes
• feat: scope agent tool default paths to permitted folders in Cowork mode by @inureyes
• feat: replace No Agent label with General Purpose default in profile selector by @inureyes
• feat: enhance Cowork empty state with quick actions, folder summary, and history by @inureyes
• feat: scope agent tool default paths to permitted folders in Cowork mode by @inureyes
• feat: relocate runtime selector to settings drawer by @inureyes
• feat: improve permitted folder chip readability with smart paths and permission indicators by @inureyes
• feat: add contextual placeholder and tool badges to Cowork task input by @inureyes
• feat: consolidate folder management UX with permission controls on chips by @inureyes
• fix: add retry logic for transient API errors in agent loop by @inureyes
• feat: add smooth transition animation and analyzing state for cowork task submission by @inureyes
• fix: resolve elapsed timer stuck at 0:00 and compact cowork empty state layout by @inureyes
• fix: align cowork empty state two-column tops and heights by @inureyes
• fix: pass working directory to agent filesystem tools by @inureyes
• feat: show error details on failed Squad agents, tasks, and executions by @inureyes
• feat: add streaming thinking steps, redesign run detail view, and save-to-creation by @inureyes
• feat: add bgo-container-cli for Apple Container runtime by @inureyes
• feat: redesign chat message edit mode for better usability by @inureyes
• fix: align bgo-container-cli with actual Containerization framework API by @inureyes

v1.5.1

Distributed model pool, fallback chain UI, engine variants, and headless mode stability improvements.

Backend.AI GO v1.5.1

46 commits since v1.5.0.

New Features

• Distributed model pool with failover, circuit breaker, and routing statistics
• Build distributed model index from health check data
• Wire inference proxy to route requests to remote nodes
• Enable remote model load/unload on connected nodes
• Frontend UI for distributed model pool and routing management
• Fallback chain configuration UI in API Settings
• Interactive fallback chain flow diagram with glass morphism visual overhaul
• Sankey diagram for real-time fallback traffic monitoring
• Mid-stream fallback configuration for router config generation and UI
• Pop-out window support for fallback chain diagram and network topology
• Engine variant field with CUDA target support
• Variant-aware build and download scripts for mlxcel
• Variant field in publish-engine registry packages.json
• Auto-generate default admin key on first headless mode startup
• Replace direct tauriInvoke calls with API adapter for headless mode
• Migrate modelStore to API adapter for headless mode support

Improvements

• Migrate all custom modal/dialog implementations to common Dialog component
• Convert all Tauri adapter imports to dynamic for headless mode
• Make API Settings save button prominent when unsaved changes exist
• Add query parameter handling for Settings page tab navigation

Bug Fixes

• Relax overly restrictive Python sandbox import blocking
• Add memory tool handlers for non-squad single agents
• Align FallbackConfig schema with continuum-router expected YAML structure
• Prevent API key loss from encrypted file storage corruption
• Prevent conversation hover actions from overlapping with metadata text
• Tauri adapter returning undefined for available engines list
• Resolve headless mode OOBE and engine management issues
• Eliminate remaining TAURI_INTERNALS crashes in headless mode
• Remove macOS traffic light padding in headless/web mode sidebar
• Engine card variants null safety and model store error recovery
• Resolve storage path mismatch and align banner layout in headless mode
• Resolve headless mode errors across multiple pages
• Decrypt values when loading encrypted key file and handle corrupt files
• Use encrypted file storage and registry-based auth in headless mode

CI/CD Improvements

• Update build-engine workflow to support mlxcel variant builds

Technical Details

• Update build_mlxcel.sh to support variant and Linux CUDA targets
• Update download-mlxcel.sh to support variant-aware CUDA asset names

Dependencies

• mlxcel updated to 0.0.10
• continuum-router updated to 1.4.1
• llama.cpp updated to b8373

Breaking Changes

None

Known Issues

None

Full Changelog

v1.5.0

Container execution, multi-channel messaging, unified model selector, and credential proxy server with 50+ improvements and fixes.

Backend.AI GO v1.5.0

53 commits since v1.4.1.

New Features

• Container agent runner image and build system
• Sentinel marker protocol and container output parser
• Mount security validator with allowlist
• Squad container execution mode integration
• IPC file system for host-container bidirectional communication
• Container lifecycle management with concurrency control
• Group and session namespace management
• Container agent task scheduler with cron/interval/once
• Channel registry and message router infrastructure
• Telegram bot channel integration
• Slack bot channel integration
• Discord bot channel integration
• WhatsApp channel integration
• Squad multi-channel messaging tab UI
• Sender allowlist and message access control
• Cowork sub-agent container execution mode
• IPC permission enforcement and security hardening
• Unified model selector component across all model-consuming screens
• Container resource monitoring and audit logging
• Credential proxy server for API key isolation
• Container runtime detection and abstraction layer
• Relocate and rename folder permissions to working folders in Cowork page

Improvements

• Migrate ChatInterface to useUnifiedModelSelector hook
• Extract shared SettingsSection, Accordion components and redesign Squad channels tab
• Refresh What's New modal content for v1.5.0

Bug Fixes

• Hide search bar and format filters on model detail view
• Initialize Access Key Registry in headless mode
• MLX config files downloaded with wrong format causing tokenizer missing
• Render delete confirmation dialog in top-layer to appear above image viewer modal
• Resolve model ID mismatch when unloading from unified selector
• Surface image save failures and add retry mechanism for Creations page
• Resolve infinite loading spinner when Draw session image rehydration fails
• Add z-index to creation card actions to fix unclickable delete button
• Tighten enterprise headline layout
• Polish webpage landing copy and layout
• GTC banner and header polish
• Refine enterprise landing section, Korean hero tagline, hero copy hierarchy, landing English copy, webpage landing theme

CI/CD Improvements

None

Technical Details

• Container execution mode provides isolated sandboxed environments for Squad and Cowork sub-agents
• Multi-channel messaging supports Telegram, Slack, Discord, and WhatsApp via a unified channel registry and message router
• Unified model selector replaces per-page model dropdowns with a consistent component across Chat, Agent, Squad, Draw, and Translation
• Credential proxy server isolates API keys from agent containers via a localhost proxy
• IPC file system enables bidirectional host-container communication through a virtual filesystem

Dependencies

• Bump continuum-router to v1.4.0
• Bump mlxcel to v0.0.6

Breaking Changes

None

Known Issues

None

Full Changelog

v1.4.1

Squad agent memory tools, DGX Spark GPU support, and MLX model improvements.

Backend.AI GO v1.4.1

24 commits since v1.4.0.

New Features

• Expose read_memory, write_memory, and search_memory as squad agent tools for persistent context across tasks
• Add download all button for batch downloading MLX safetensors model files
• Add GPU variant field for GPU-specific engine packages with variant identification
• Display GPU variant in engine UI and fix composite ID collision
• Add Linux ARM64 CUDA 13 (DGX Spark / GB10) engine build target

Improvements

• Add project board automation workflow
• Update and add tab screenshot tests for new pages and correct tab labels
• Add mock handlers for squad, audio, diffusion, MCP, ACP, memory, scheduler, cowork, and translation
• Add GPU variant guide comment in build-engine workflow

Bug Fixes

• Sync searchFormat with selectedFormats in model selection
• Restore agent tool toggle functionality in squad configuration
• Replace raw HTML cursor injection with React component in streaming chat
• Persist MLX model metadata correctly across app restarts
• Use cmd shell for Windows Tauri build to preserve JSON config quotes
• Include variant suffix in package filename references
• Include variant in llama-cpp build job name
• Add variant field to manifest JSON schema
• Resolve nvcc discovery and CUDA arch for DGX Spark build
• Use correct runner labels for DGX Spark build job

CI/CD Improvements

• Eliminate redundant Rust recompilation in Windows packaging
• Add Linux ARM64 CUDA 13 (DGX Spark / GB10) engine build target

Technical Details

• GPU variant support enables multiple GPU-specific builds of the same engine (e.g., CUDA 12 vs CUDA 13)
• Squad agents now have full memory access for context persistence across multi-agent workflows

Dependencies

None

Breaking Changes

None

Known Issues

None

Full Changelog

v1.4.0

Agent Squad, ACP, Automations, CLI, and 6 new provider types with 178 commits since v1.3.1.

Backend.AI GO v1.4.0

178 commits since v1.3.1.

New Features

• Agent Squad: Multi-agent collaboration system with planner agent, task decomposition, TaskBoard Kanban UI, workspace file explorer, real-time execution events, memory search across agent banks, squad templates, execution budget controls, execution logs and report generation, inline agent editing, agent duplication, persistent chat sessions per agent, live work summary ticker, and comprehensive documentation (–,,,,,,,)
• ACP (Agent Communication Protocol): Settings UI page, agent profile selection and registry integration, session monitor UI, permission policy bridge with existing tool permission system, and server process lifecycle management (–)
• Automations (Scheduler): Cron-based task scheduling with data model, storage layer, ScheduleManager, TaskExecutor, template engine, input sources, output actions, i18n keys, schedule form with tool selection, execution history tab, desktop notifications, and Management API REST endpoints (–,,–,)
• Comprehensive CLI: 50+ new commands covering engine, pool, provider, HuggingFace, model, statistics, logging, conversations, memory, plugins, MCP, scheduling, benchmarks, lifecycle, access keys, media, translation, and agent management; Unix domain socket communication; CLI auto-discovery via discovery file (–)
• Architecture-based server parameter profiles for llama-server
• 6 new provider types to match continuum-router backend types
• Format filter chips for HuggingFace model browser and local models
• LLM-powered one-line summarization utility
• Thinking summary in collapsed ThinkingBlock accordion
• Tag label support (Beta/Experimental) for sidebar navigation items
• Inference engine name display in loaded model status panel
• Runtime context size display in model config drawer
• Save to Creations button in text tab result footer
• Fetch_url tool localhost/private IP access via user-configurable allowlist
• Localized built-in agent profile names and descriptions
• Engine-specific icons for sd-server and whisper-server
• Mlxcel engine support with build and publish workflows
• Model structure modal for safetensor (MLX) models
• Graceful shutdown for sidecar processes on Windows

Improvements

• Full-bleed chat interface with centered content
• Agent chat full-bleed like main chat page
• Replace left border accent with ambient glow + edge bleed on cards
• Unify SquadCard and ScheduleCard into shared InfoCard pattern
• Reuse TextInput common component in Squad and Agent chat inputs
• Consolidate SettingsSlider into SliderInput across all usages
• Unify Tabs and TabGroup into a single common component
• Convert BudgetConfig from collapsible details to Drawer
• Rename Schedules to Automations across all locale files
• Split AGENTS.md into modular .claude/rules/ with path scoping
• Enforce mandatory component reuse and no-duplicate-module policy
• Optimize chunk splitting and fix build warnings

Bug Fixes

• Streaming chat stuttering from O(n²) artifact extraction and excessive markdown re-parses
• Auto-scroll stopping during streaming
• Hardcoded llama-server references replaced with dynamic engine names
• Safetensor model alias and ID handling with parent directory names (–)
• Linux CUDA runtime detection and GPU logging
• ModelConfig parameters integration into inference server startup
• useModelMaxContext setting connected to model loading logic
• Update checks routed through channel-aware command with CA cert support
• Model display name showing raw IDs in chat selector
• Scheduler model ID mismatch causing unload failures
• Engine update state preserved across auto-refresh
• Squad task status, tool-call limit, and log errors
• Squad write_file tool silently creating 0-byte files
• Squad JSON parse error returned to LLM instead of silent empty fallback
• Persisted execution history and activity feed loading after restart
• Non-functional engine update button; added Update All
• Workspace tab height chain for proper file tree and viewer scrolling
• Hardcoded page title switch replaced with i18n fallback chain
• Document translation result panel expanded to full height
• Artifact auto-open throttled to once per streaming session
• Deduplicate sharded safetensors entries on Tier 3 scan
• Mlxcel version resolution and metallib packaging fixes

CI/CD Improvements

• Upgrade macOS runner from macos-14 to macos-15
• Add min_app_version option to engine build and publish workflows
• Increase Windows build timeout to 90 minutes
• Rename GH_PAT to GH_PAT_DOWNLOAD_ARTIFACT in packaging workflow

Technical Details

• Squad uses file-based agent memory banks with workspace directory lifecycle management
• Squad planner agent handles task decomposition with dependency graphs
• ACP uses stdio-based transport with permission policy bridging
• Scheduler uses cron engine with catch-up execution and template rendering
• CLI communicates via Unix domain socket to Management API
• Architecture-based server profiles auto-tune llama-server parameters per CPU architecture

Dependencies

None

Breaking Changes

None

Known Issues

None

What's Changed

• perf: fix build warnings and optimize chunk splitting by @inureyes
• fix: move logging initialization after AppPaths resolution by @inureyes
• fix: route all update checks through channel-aware command with CA cert support by @inureyes
• fix: implement Linux CUDA runtime detection and improve GPU logging by @inureyes
• fix: filter llama-server-specific CLI arguments when launching non-llama engines by @inureyes
• feat: support model structure modal for safetensor (MLX) models by @inureyes
• fix: pass parent directory as --model for safetensor models on non-llama engines by @inureyes
• fix: use parent directory name for safetensor model aliases and IDs by @inureyes
• fix: replace hardcoded llama-server references with dynamic engine names by @inureyes
• fix: use model alias for backend display name and auto-shorten long socket paths by @inureyes
• fix: replace hardcoded llama-server strings in log messages with dynamic engine name by @inureyes
• fix: connect useModelMaxContext setting to model loading logic by @inureyes
• feat(scheduler): define data model, storage layer, and cron engine by @inureyes
• fix: integrate ModelConfig parameters into inference server startup by @inureyes
• docs: enforce mandatory component reuse and no-duplicate-module policy by @inureyes
• feat(scheduler): implement ScheduleManager and TaskExecutor by @inureyes
• feat(scheduler): add Tauri IPC commands for schedule management by @inureyes
• feat(scheduler): add SchedulesPage and sidebar navigation under Agent section by @inureyes
• fix: resolve streaming chat stuttering from O(n²) artifact extraction and excessive markdown re-parses by @inureyes
• fix(chat): resolve auto-scroll stopping during streaming by @inureyes
• feat(scheduler): implement schedule creation/edit form and execution history drawer by @inureyes
• feat(scheduler): add i18n translation keys for scheduled tasks by @inureyes
• feat(scheduler): add Management API REST endpoints for API parity by @inureyes
• feat(scheduler): implement template engine, input sources, and output actions by @inureyes
• feat: add mlxcel engine to build and publish workflows by @inureyes
• feat(cli): add extended model management commands by @inureyes
• feat(cli): add HuggingFace integration commands by @inureyes
• feat(cli): add pool management commands by @inureyes
• feat(cli): add engine management commands by @inureyes
• fix: trigger model loading in scheduler executor when auto_load_model is enabled by @inureyes
• fix: preserve engine update state across auto-refresh and reduce lock scope by @inureyes
• feat(cli): add provider management commands by @inureyes
• fix: implement desktop notifications and use common ModelSelect in schedule form by @inureyes
• feat(cli): add settings, storage and monitoring commands by @inureyes
• feat(cli): add statistics and logging commands by @inureyes
• feat(cli): extend loaded model and router CLI operations by @inureyes
• feat: add localization support for built-in agent profile names and descriptions by @inureyes
• feat(cli): add conversation and memory management commands by @inureyes
• feat(cli): add plugin, MCP, and scheduling commands by @inureyes
• feat: add CLI auto-discovery via Management API discovery file by @inureyes
• feat(cli): add benchmark, lifecycle, and access key commands by @inureyes
• feat(cli): add media, translation, and agent management commands by @inureyes
• feat: add Unix domain socket support to Management API for CLI communication by @inureyes
• fix: add notification permission check for desktop notification output action by @inureyes
• fix: use folder picker and auto-generate filenames for Save to File action by @inureyes
• feat: add Execution History tab to Schedules page by @inureyes
• refactor: promote SliderInput to common component and unify all slider usages by @inureyes
• refactor: unify Tabs and TabGroup into a single common component by @inureyes
• feat: add tool selection to schedule creation/edit form by @inureyes
• fix: resolve scheduler model ID mismatch causing unload failures by @inureyes
• feat: add engine-specific icons for sd-server and whisper-server by @inureyes
• feat: add ACP server process lifecycle management by @inureyes
• feat: show schedule execution progress and completion in notification drawer by @inureyes
• refactor: rename Schedules to Automations across all locale files by @inureyes
• feat: add ACP settings, Tauri commands, and Management API endpoints by @inureyes
• feat: add ACP permission policy bridge with existing tool permission system by @inureyes
• feat: add ACP settings UI page by @inureyes...

v1.3.1

Statistics page source tab, multi-GPU UI, custom CA certificates, sidebar Agent section, and 16 bug fixes.

Backend.AI GO v1.3.1

41 commits since v1.3.0.

New Features

• Add source tab to Statistics page for In-App vs Router API stats
• Multi-GPU UI with dynamic GPU detection and main GPU passthrough
• Custom CA certificate support for update channel
• Restructure sidebar navigation with separate Agent section
• Persistent model selection guide tab on the Models page
• ModelSelect dropdown in agent profile editor replacing text input
• Alphabetical sorting for target language list in glossary editor
• CSS animation for BenchmarkIllustration empty state

Improvements

• Replace Qwen3 featured models with Qwen3.5-4B

Bug Fixes

• Resolve router stats fetch failure by adding Unix socket support and fixing response format
• Add spacing between dashboard toolbar row and widget grid
• Create versioned .so symlinks after engine extraction on Linux
• Load conversation from URL query parameter on chat page
• Load models on agent marketplace page mount
• Resolve race condition in loaded model display
• Show queued status for models in recommended list
• Use tauri async runtime for shutdown safety net to prevent panic
• Use instance-specific socket paths for Continuum Router
• Prevent context window usage resetting to 0 after response completion
• Resolve app shutdown hanging indefinitely at model pool stop step
• Resolve inability to delete duplicated built-in agent profiles
• Narrow webpage deploy paths to mkdocs-rendered directories only
• Clean up leftover comma artifacts in sanitized release notes
• Add actions:write permission for Homebrew cask workflow trigger
• Add missing --repo flag to gh workflow run in sync-release

CI/CD Improvements

• Narrow webpage deploy paths to mkdocs-rendered directories only
• Add actions:write permission for Homebrew cask workflow trigger
• Add missing --repo flag to gh workflow run in sync-release

Technical Details

None

Dependencies

• Update Continuum Router to v1.2.0
• Update all-smi to 0.17.3
• Bump llama.cpp default version from b7717 to b8189

Breaking Changes

None

Known Issues

None

What's Changed

• feat: add CSS animation to BenchmarkIllustration empty state by @inureyes
• fix: resolve inability to delete duplicated built-in agent profiles by @inureyes
• feat: replace text input with ModelSelect in agent profile editor by @inureyes
• feat: sort target language list alphabetically in glossary editor by @inureyes
• feat: add persistent model selection guide tab to models page by @inureyes
• fix: resolve app shutdown hanging indefinitely at model pool stop step by @inureyes
• update: replace Qwen3 featured models with Qwen3.5-4B by @inureyes
• update: bump llama.cpp default version to b8189 by @inureyes
• fix: prevent context window usage resetting to 0 after response completion by @inureyes
• chore: apply code formatting and add technical reports by @inureyes
• fix: use instance-specific socket paths for Continuum Router by @inureyes
• fix: use tauri async runtime for shutdown safety net to prevent panic by @inureyes
• fix: show queued status for models in recommended list by @inureyes
• fix: resolve race condition in loaded model display by @inureyes
• feat: restructure sidebar navigation with separate Agent section by @inureyes
• fix: load models on agent marketplace page mount by @inureyes
• fix: load conversation from URL query parameter on chat page by @inureyes
• fix: create versioned .so symlinks after engine extraction on Linux by @inureyes
• feat: support custom CA certificates for update channel by @inureyes
• feat: improve multi-GPU UI with dynamic GPU detection and main GPU passthrough by @inureyes
• fix: add spacing between dashboard toolbar row and widget grid by @inureyes
• feat: add source tab to Statistics page for In-App vs Router API stats by @inureyes
• fix: resolve router stats fetch failure by adding Unix socket support and fixing response format by @inureyes
• chore: update continuum-router to v1.2.0 and all-smi to 0.17.3 by @inureyes

v1.3.0

Backend.AI GO v1.3.0 — Plugin system, Companion AI, mDNS auto-discovery, new languages, and 30+ bug fixes.

Backend.AI GO v1.3.0

95 commits since v1.2.2.

New Features

• Plugin system with SDK, permission-gated API surface, JS module loader with error isolation, UI slot system, management REST API, IPC commands, filesystem manager, manifest schema, settings persistence, zip installation, plugin store connection, and auto-registration
• Companion AI with floating widget UI, chat bubble with streaming responses, persona presets, and Memory Bank integration (,,,)
• Serper (Google Search) web search provider with abstraction layer
• mDNS auto-discovery with stale node detection, direct node registration, and capability advertisement
• Audio transcription built-in tool for speech-to-text
• Built-in performance monitoring panel
• Spanish (es) language support
• Traditional Chinese (zh-TW) language support
• Simplified Chinese (zh-CN) language support
• Animated SVG splash screen during initial app loading
• Windows AF_UNIX socket support via socket2
• Keyboard shortcuts displayed with separate styled key badges
• Expose Add Widget and Reset buttons directly without Customize mode
• API key warning in ToolSelector when web_search is enabled
• Dependency version update script and Claude command
• Triage and resolve commands for user feedback workflow

Improvements

• Consolidate custom dropdown menus to shared ContextMenu component
• Hide plugins menu in production, show only in dev mode
• Reorganize sidebar menu layout
• Migrate plugin components to common UI primitives
• Rename Chinese locale codes from script subtags to region-based codes
• Improve UX feedback during slow prefill with memory context
• Update Continuum Router version to v1.1.0
• Refresh What's New content for changes since v1.2.2
• Translate Cowork menu name to Korean
• Translate Bliss theme name to Korean

Bug Fixes

• Resolve TypeScript type errors and ESLint violations across test files
• Repair 238 failing tests across 23 frontend files and 16 Rust doc tests
• Show user-defined provider name instead of OPENAI for custom providers
• Resolve composite model ID before title generation API call
• Send clear flags when selecting "None" in default model settings
• Auto-fetch plugins on app startup and re-render after async load
• Refresh relative timestamps periodically in useFormattedTime hook
• Replace AI-generated benchmark icons with provided SVG design
• Rename status messages to reflect response generation phases
• Add retry logic and path validation for engine update on Windows
• Remove hardcoded FPS sparkline maxValue to support high-refresh-rate monitors
• Escape Jinja2 comment syntax in troubleshooting heading anchors
• Replace bash build-plugins script with cross-platform Node.js version
• Add HTTP fallback when socket streaming fails on Windows
• Exclude llama.cpp build artifacts from Vite dependency scan
• Use hover overlay for Draw sidebar action buttons
• Correct benchmark page header and title/subtitle text
• Resolve chat title generation failures and add recovery
• Convert custom provider env var names to uppercase with underscores
• Add plugin build step to dev mode beforeDevCommand
• Preserve other engines' update badges during single engine update
• Complete Windows AF_UNIX socket support across all socket stubs
• Add missing enable_performance_monitor field to Rust backend
• Limit supervisor decisions timeline and add missing node i18n keys
• Restart router when toggling external access to apply bind host change
• Correct ServiceRemoved event matching in mDNS discovery
• Filter out self-node from mDNS discovery results
• Use Button icon prop instead of SVG children for correct flex layout
• Correct API server toggle showing OFF on Windows when TCP is active

CI/CD Improvements

• Auto-trigger publish workflows after successful builds

Technical Details

• Plugin architecture with isolated JS module loading, permission-gated SDK, and filesystem storage backend
• Companion AI system with persona management, memory persistence, and streaming chat interface
• mDNS service discovery with automatic capability advertisement and stale node cleanup
• Windows AF_UNIX socket support using socket2 crate with TCP fallback

Dependencies

• Continuum Router updated to v1.1.0

Breaking Changes

None

Known Issues

None

What's Changed

• feat: define plugin manifest schema and shared type definitions by @inureyes
• feat: implement plugin filesystem manager and storage backend by @inureyes
• feat: add audio_transcribe built-in tool for speech-to-text by @inureyes
• feat(ci): auto-trigger publish workflows after successful builds by @inureyes
• feat: implement plugin Tauri IPC commands by @inureyes
• fix: correct API server toggle showing OFF on Windows when TCP is active by @inureyes
• feat: add plugin management REST API endpoints by @inureyes
• fix: use Button icon prop instead of SVG children for correct flex layout by @inureyes
• feat: connect plugin store to real Tauri backend by @inureyes
• fix: filter out self-node from mDNS discovery results by @inureyes
• fix: correct ServiceRemoved event matching in mDNS discovery by @inureyes
• feat: add stale node detection and periodic cleanup for discovered nodes by @inureyes
• fix: restart router when toggling external access to apply bind host change by @inureyes
• feat: advertise real engine and model capabilities via mDNS by @inureyes
• feat: implement plugin JS module loader with error isolation by @inureyes
• feat: implement Plugin SDK with permission-gated API surface by @inureyes
• feat: implement direct node registration from discovered nodes by @inureyes
• feat: implement UI slot system for plugin rendering by @inureyes
• feat: display keyboard shortcuts with separate styled key badges by @inureyes
• test: add comprehensive tests for mDNS auto-discovery system by @inureyes
• docs: update auto-discovery documentation with mDNS details by @inureyes
• i18n: translate Bliss theme name to Korean (푸르름) by @inureyes
• feat: implement plugin installation from zip and improve uninstall flow by @inureyes
• i18n: translate Cowork menu name to Korean (협업) by @inureyes
• docs: add Supervisor Agent manual by @inureyes
• feat: add built-in performance monitoring panel by @inureyes
• feat: implement plugin settings persistence and runtime configuration by @inureyes
• fix: limit supervisor decisions timeline and add missing node i18n keys by @inureyes
• feat: add Windows AF_UNIX socket support via socket2 by @inureyes
• fix: add missing enable_performance_monitor field to Rust backend by @inureyes
• feat: add Serper (Google Search) web search provider with abstraction layer by @inureyes
• feat: add companion plugin floating widget UI by @inureyes
• feat: add companion chat bubble interface with streaming responses by @inureyes
• feat: add companion persona system with presets by @inureyes
• feat: add companion memory integration via Memory Bank by @inureyes
• fix: preserve other engines' update badges during single engine update by @inureyes
• refactor: apply code formatting and fix lint errors by @inureyes
• feat: add built-in plugin build pipeline and auto-registration by @inureyes
• feat: improve UX feedback during slow prefill with memory context by @inureyes
• feat: add Traditional Chinese (zh-Hant) language support by @inureyes
• fix: add plugin build step to dev mode beforeDevCommand by @inureyes
• feat: add Simplified Chinese (zh-Hans) language support by @inureyes
• feat: add Spanish (es) language support by @inureyes
• fix: convert custom provider env var names to uppercase with underscores by @inureyes
• fix: resolve chat title generation failures and add recovery by @inureyes
• refactor: rename Chinese locale codes from script subtags to region-based codes by @inureyes
• docs: enhance FAQ and Troubleshooting with macOS Gatekeeper guidance by @inureyes
• feat: add animated SVG splash screen during initial app loading by @inureyes
• fix: correct benchmark page header and title/subtitle text by @inureyes
• fix: exclude llama.cpp build artifacts from Vite dependency scan by @inureyes
• fix: use hover overlay for Draw sidebar action buttons by @inureyes
• feat: add triage and resolve commands for user feedback workflow by @inureyes
• chore: update Continuum Router version to v1.1.0 by @inureyes
• fix: add HTTP fallback when socket streaming fails on Windows by @inureyes
• feat: add dependency version update script and Claude command by @inureyes
• refactor: migrate plugin components to common UI primitives by @inureyes
• update: reorganize sidebar menu layout by @inureyes
• fix: replace bash build-plugins script with cross-platform Node.js version by @inureyes
• fix: escape Jinja2 comment syntax in troubleshooting heading anchors by @inureyes
• fix: remove hardcoded FPS sparkline maxValue for high-refresh-rate monitors by @inureyes
• fix: add retry logic and path validation for engine update on Windows by @inureyes
• fix: rename status messages to reflect response generation phases by @inureyes
• update: refresh What's New content for changes since v1.2.2 by @inureyes
• docs: add OS-specific installation troubleshooting FAQ entries by @inureyes
• feat: show API key warning in ToolSelector when web_search is enabled by @inureyes
• update: hide plugins menu in production, show only in dev mode by @inureyes
• fix: replace AI-generated benchmark icons with provided SVG design by @inureyes
• feat: expose Add Widget and Reset buttons directly without Customize mode by @inureyes
• fix: refresh relative timestamps periodically in useFormattedTime hook by @inureyes
• fix: auto-fetch plugins on app startu...

v1.2.2

Whisper-cpp audio transcription, Japanese language support, supervisor predictive analytics, and Docker/CI improvements.

Backend.AI GO v1.2.2

28 commits since v1.2.1.

New Features

• Integrate whisper-cpp as audio transcription engine in Rust backend
• Add whisper-cpp engine build script and CI workflow
• Add Japanese (ja) language support (,)
• Harden Python sandboxing with explicit module import blocking
• Support downloading and loading HuggingFace model files in subdirectories
• Implement supervisor phase 3 — predictive analytics, adaptive tuning, and external integration
• Auto-promote stable releases to latest in finalize step

Improvements

• Extract hardcoded language types into SUPPORTED_LANGUAGES constant
• Differentiate hero headline colors with per-line glow
• Refresh What's New content with all changes since v1.2.0

Bug Fixes

• Use native fetch for SSE streaming on Windows
• Register 5 missing tools in ToolRegistry::with_builtin_tools
• Merge agent profile and memory into single system message
• Replace invalid SwaggerUI route parameter name for axum 0.8 compatibility
• Remove unsupported Vulkan targets from whisper-cpp build matrix
• Fetch release data at build time instead of client-side
• Only trigger webpage rebuild for stable releases

CI/CD Improvements

• Trigger webpage rebuild after release finalization
• Add protobuf-compiler to Docker build dependencies
• Bump Rust image to 1.93 to satisfy dependency MSRV requirements
• Allow docs/legal through .dockerignore for frontend build
• Limit Docker workflow push trigger to Dockerfile changes only
• Copy docs/legal into Docker frontend-builder stage

Technical Details

• Supervisor phase 3 adds predictive analytics, adaptive tuning, and external integration capabilities
• Whisper-cpp integration provides native audio transcription without external dependencies
• Python sandbox hardening blocks explicit module imports for improved security
• SUPPORTED_LANGUAGES constant replaces hardcoded language types for i18n extensibility

Dependencies

None

Breaking Changes

None

Known Issues

None

Full Changelog

v1.2.1

Artifact Viewer, Supervisor Agent, and stability improvements since v1.2.0.

Backend.AI GO v1.2.1

14 commits since v1.2.0.

New Features

• Artifact Viewer Phase 1 — Side panel and multi-window renderer for code, SVG, HTML, and Markdown content generated during LLM conversations. Includes streaming-aware artifact detection, Code/Preview tabs, pop-out windows, copy/download toolbar, and keyboard shortcut (Ctrl/Cmd+Shift+A)
• Artifact Viewer Phase 2 — Mermaid diagram renderer with theme-aware dark/light mode and SVG export, CSV interactive table viewer with sortable columns, search/filter, and virtual scrolling for large datasets, and LaTeX equation renderer with KaTeX. Includes Creations persistence with file-based storage and deduplication
• Supervisor Agent — Centralized resource arbitration with priority-based policy engine. Features 5 built-in policies (Safety/OOM Prevention, Availability/Auto-Restart, Pinned Model Protection, Idle Eviction, Resource Optimization), multi-tier tick loop, bounded audit log, and Decision Debugger settings panel. Full Tauri IPC (9 commands) and REST API (8 endpoints) parity

Improvements

• Unified provider addition flow to use ProviderSetupWizard exclusively, with Health Check Endpoint field and optional API key hints for vLLM/Custom providers
• Updated What's New modal content for v1.2.1 release highlights

Bug Fixes

• Wire ArtifactPanel into layout and connect streaming pipeline
• Make links visible in user chat message bubbles
• Raise portal dropdown z-index above dialog overlay
• Remove hardcoded max-height from translation text area
• Replace blanket process cleanup with registry-based orphan detection for safer multi-instance support, with improved port/socket conflict messaging and 5 new crash recovery tests
• Restore translation textarea height to fill available viewport space

CI/CD Improvements

None

Technical Details

• Artifact system: ArtifactExtractor streaming code fence parser with state machine, Zustand store for artifact state management, React.lazy + Suspense for code-splitting heavy renderers (Mermaid, CSV, LaTeX)
• Supervisor system: 8 Rust modules (config, actions, snapshot, policy, resource_arbiter, fallback, audit, agent), weighted eviction scoring, FallbackManager for Continuum Router model failover, SnapshotProvider trait for testability
• Security: Sandboxed iframes for HTML artifacts, SVG event handler stripping, blob URL origin isolation
• Process management: ProcessRegistry-based orphan detection with PID reuse binary verification, stale socket auto-cleanup

Dependencies

None

Breaking Changes

None

Known Issues

None

Full Changelog