Changelog

Aleph follows a milestone-driven development process. Since the project is pre-1.0, changes are organized by month and milestone rather than semantic version tags.

---

February 2026

Desktop Bridge Overhaul

A complete redesign of the desktop integration layer, migrating from a native Swift macOS app to a cross-platform Tauri-based bridge architecture.

Server-Centric Build Architecture

• Introduced BridgeSupervisor — the Rust server now spawns and manages the Tauri bridge as a subprocess
• Implemented DesktopBridgeManager for server-side bridge lifecycle management
• Added JSON-RPC protocol over Unix Domain Sockets (UDS) for server-to-bridge communication
• Bridge handshake with capability negotiation (aleph.handshake, system.ping)

Browser Automation

• New BrowserTool with Chrome DevTools Protocol (CDP) transport via chromiumoxide
• Cross-platform Chromium discovery (macOS, Linux, Windows)
• ARIA snapshot extraction with ref_id resolution for semantic element targeting
• Full action set: click, type, fill, scroll, hover with ARIA ref targeting
• Integration tests for complete browser workflows

Vision & Approval System

• VisionTool with pluggable providers: Claude vision and platform-native OCR
• ApprovalPolicy trait and configuration for gating sensitive tool actions
• Approval checks integrated into BrowserTool and DesktopTool

Desktop Capabilities (Phase 3)

• 12 bridge methods implemented: click, type_text, key_combo, launch_app, window_list, focus_window, canvas_overlay, tray.update_status, ocr, ax_tree, scroll, drag
• Windows support: input simulation via WinRT, OCR, UI Automation accessibility tree
• macOS: Vision framework OCR, Accessibility API inspection

Breaking Changes

• UniFFI Swift bindings removed — replaced by UDS IPC protocol
• macOS Swift app deprecated and removed — Tauri bridge is the sole desktop frontend
• macOS-specific CI workflows and build scripts removed

Cron System Redesign

• Replaced window-based scheduling with a state-machine scheduling engine
• Added delivery pipeline with DeliveryTarget trait (webhook support)
• Template system, resource management, and job chaining
• Schema migration with state-machine, resilience, delivery, and chain fields

Memory System Enhancements

• Storage-time deduplication via vector similarity
• JSONL backup service with rolling retention
• ScoringPipeline wired into HybridRetrieval search flow
• NoiseFilter integrated into the ingestion pipeline
• Adaptive retrieval and backup settings added to MemoryConfig

---

Early February 2026

Swarm Intelligence Architecture

• Event bus foundation with AgentLoopEvent enum and publishing (shadow mode)
• Semantic aggregator for cross-agent context
• Context injector with ContextProvider trait integrated into MessageBuilder
• SwarmContextProvider implementation
• Collective memory for multi-agent shared state
• Phase 1 Shadow Monitor Mode completed

System State Bus (SSB)

• Core infrastructure: event routing, subscriptions, and state tracking
• Robustness and privacy layer with PII filtering
• Action dispatcher for automated responses to system events
• Platform Abstraction Layer (PAL) with macOS-specific sensors
• Integration with SystemStateBus for perception data

Control Plane UI

• Sidebar narrow/wide mode switching
• Tooltip and StatusBadge components for alert indicators
• Real-time alert system via WebSocket subscriptions
• Settings with nested routing and SettingsLayout
• Real Gateway RPC integration (replaced mock data)

Collaborative Skill Evolution

• SuccessManifest data structure for tracking successful tool patterns
• ConstraintValidator for skill safety checks
• CollaborativeSolidificationPipeline for multi-agent skill sharing
• Integration with tool execution for automatic manifest creation

---

Late January 2026

POE (Principle-Operation-Evaluation) Architecture

A new execution framework for structured, validated agent task execution.

• Core types: principles, operations, evaluation criteria
• Entropy-based budget manager for resource allocation
• Hard validation rules and semantic validation via LLM judge
• Composite validator combining rule-based and semantic checks
• POEManager for orchestrating POE lifecycles
• Worker abstraction with AgentLoopWorker integration
• Gateway RPC handlers: poe.run, poe.status, poe.cancel, poe.list
• Experience crystallization feeding into the skill evolution system

Agent-Action Interaction

• QuestionKind types for structured user interaction (confirm, select, input)
• UserAnswer type for structured responses
• AskUserRich decision variant in the agent loop
• on_user_question callback in LoopCallback trait
• CLI implementation via inquire crate

Skill Evolution System

• EvolutionTracker for logging tool execution patterns
• SolidificationDetector for detecting repeated successful patterns
• SkillGenerator for automatic SKILL.md creation
• GitCommitter for auto-committing evolved skills
• SkillCompiler (Phase 10) for compiling skills into executable form

Spec-Driven Development

• SpecWriter for automated requirement analysis
• TestWriter for test generation from specifications
• LlmJudge for evaluation of generated code against specs
• SpecDrivenWorkflow orchestrator tying it all together

Resilient Task Execution

• ResilientTask trait for defining fault-tolerant tasks
• ResilientExecutor with retry logic, exponential backoff, and fallback strategies
• Cron integration with PodcastTask example

---

Mid-January 2026

rig-core Migration

A major internal migration from the custom AI provider layer to the rig-core framework.

• Upgraded rig-core from 0.6 to 0.28
• Complete V1-to-V2 core migration (AetherCore removed, AetherV2Core is now the sole implementation)
• New UniFFI v2 bindings with async callback pattern
• RigAgentManager with process methods for tool-calling agents
• MemoryStore with SQLite and fastembed for vector search
• SearchTool (Tavily API) and WebFetchTool (HTML extraction) as built-in tools
• Legacy routing, semantic, and tools modules removed

Aether Cowork Mode

• Task orchestration core (Phase 1)
• Swift UI progress panel for task tracking
• UniFFI bindings for task orchestration
• Configuration system and settings UI

macOS UI Polish

• Liquid Glass implementation for macOS 26+ with consistent appearance
• Multi-modal attachment architecture with encoding field
• Memory storage integration into conversation processing
• Clipboard image attachments passed to AI via rig-core

---

Late December 2025

Internationalization

• Full English and Simplified Chinese localization support
• Localized UI strings throughout the macOS app

Permission System

• Mandatory permission gating at startup (Accessibility, Input Monitoring)
• Auto-restart after permission grant
• Debounced permission status monitoring
• In-app permission prompts (replaced system dialogs)

Provider Management

• Provider activation/deactivation toggle
• Inline connection test results
• Custom provider authentication fixes
• Settings window layout improvements

macOS 26 Window Design

• Phase 1 & 2 implementation of new window chrome
• Minimum window size constraints
• Window reopening via URL scheme

---

Mid-December 2025

Memory Module (Phase 4)

• RAG-based contextual memory with vector search
• Memory retention policies
• macOS context capture (Swift)
• Performance benchmarks and comprehensive test suite
• Core integration with conversation pipeline

AI Provider Integration

• Foundation AI provider interface
• Multi-provider router implementation
• AetherCore and memory integration
• Config management backend with hot-reload
• Routing rules editor
• Behavior settings and hotkey customization
• PII scrubbing for privacy
• Typewriter effect output mode

Halo Overlay

• Enhanced Halo overlay UI
• Three input modes: cut, copy, and halo
• Input mode selection UI before AI processing
• Clipboard context mechanism (10-second window)
• Double-shift default hotkey (avoiding input method conflicts)

---

Pre-December 2025

Foundation

The initial development period established the core Rust server, Gateway WebSocket protocol, agent loop, session management, and the original macOS Swift app. Key milestones included:

• Rust core with async runtime (Tokio)
• Gateway JSON-RPC 2.0 protocol over WebSocket
• Multi-provider AI routing (Claude, GPT, Gemini, Ollama)
• Telegram and Discord channel adapters
• Three-tier command execution security model
• MCP (Model Context Protocol) tool integration
• SQLite-backed conversation persistence
• Configuration via TOML with environment variable overrides

---

Tags