mirzahusadzic/cogx: v2.5.0: Multi-Provider LLM Architecture

🚀 v2.5.0: Multi-Provider LLM Architecture

Transformative Release: First-time integration of Google Gemini ADK as a powerful new LLM provider, enabled by a flexible provider abstraction architecture. This release introduces full Gemini agent capabilities with tool execution, memory recall, and multi-turn conversations. Claude integration also receives significant enhancements including thinking blocks, API key support, and optional SDK. Both providers benefit from comprehensive TUI improvements and production-ready code quality.

🎯 Major Features

Full Gemini ADK Agent Integration

This release officially brings the Gemini ADK agent to Cognition CLI - a monumental addition offering a new, highly capable LLM provider for agentic workflows.

Key Capabilities:

• Multi-Turn Conversations (BIDI Streaming): Continuous multi-turn interactions, allowing the agent to persist through tool calls (e.g., git diff) and complete complex workflows without prematurely stopping
• Tool Execution with Guardrails: Gemini can leverage Cognition's tools (write_file, glob, bash, etc.) with the new tool permission system
• Memory Recall: Gemini agents can utilize the recall tool for semantic memory within conversations
• Robustness Improvements:
  • Unique session ID generation with entropy to prevent collisions
  • Sophisticated error handling for benign JSON parsing errors from experimental ADK SDK
  • BIDI mode made opt-in with automatic enablement in TUI for agent workflows
  • Graceful handling of SDK parsing errors when assistant messages exist

Enhanced Claude Integration

Claude also received significant improvements in this release, working better than ever:

• Thinking Blocks: Full support for extended thinking mode with visual "thinking blocks" in TUI, providing transparency into Claude's reasoning process
• API Key Authentication: Added support for ANTHROPIC_API_KEY environment variable while preserving existing OAuth functionality
• Flexible SDK: Made Claude Agent SDK optional, allowing basic completions without SDK dependency for licensing flexibility
• Unified Experience: Benefits from all TUI improvements (ESC interrupt, tool confirmations, multiline input) equally with Gemini

LLM Provider Abstraction Layer

To support multiple LLMs and enable future extensibility, a new pluggable LLM Provider Abstraction Layer has been introduced:

• Standardized interface for interacting with different LLMs (Claude, Gemini, and future providers)
• Dynamic provider loading and registration with health checks
• Persistent default provider configuration
• Seamless TUI integration for consistent UX across all providers

Architecture Benefits:

• Future-proof design for easy LLM provider additions
• Standardized AgentProvider and LLMProvider interfaces
• Provider registry with health checks and capability detection
• Extensible tool system with permission callbacks

🎨 TUI Improvements

Comprehensive InputBox Overhaul

The InputBox component received significant attention for enhanced user experience:

• Multiline Support: Users can now input multiline prompts for more expressive commands
• Improved Paste Handling: Robust paste operations with better large content handling
• Cursor & Backspace: Refined cursor rendering and backspace behavior, preventing UI collapses

General Stability & User Experience

• ESC Interrupt: Critical UX improvement - users can press ESC to interrupt ongoing agent responses from any LLM provider (Gemini and Claude)
• Thinking Block Reliability: Fixed bug where TUI could get stuck with thinking blocks through accurate abort tracking
• Dynamic Provider Display: Session tip messages now dynamically show current LLM provider name
• Turn Counting: Improved turn counting and display in completion status for better multi-turn interaction context

Tool System Enhancements

• Tool Confirmation Dialog: More robust and properly rendered confirmation dialogs with friendly tool names
• Permission System: Integrated permission callbacks for safe tool execution across providers

🐛 Bug Fixes & Improvements

Backward Compatibility & Deprecation Management

• Added backward compatibility for deprecated --path flag in init command
• Supports both --path and --project-root with clear deprecation warnings
• Ensures existing scripts and workflows continue working seamlessly during transition

Core Framework & Tooling

• Enhanced code quality and maintainability
• Standardized environment variable naming (e.g., GEMINI_API_KEY)
• Improved modularity of LLM provider components
• Graceful SDK error handling (no process.exit(1) crashes, OAuth error flexibility)

📝 Documentation

• Extensive TSDoc added to TUI components and hooks for better developer experience
• Updated COMPREHENSIVE_ANALYSIS.md with LLM provider abstraction details
• Comprehensive documentation for Gemini integration architecture and capabilities
• Added developer documentation for implementing custom providers
• Supported Providers: Officially supports Claude (Anthropic) and Gemini (Google) as LLM providers

⚠️ Breaking Changes (Mitigated)

CLI Flag Standardization

The --path flag has been deprecated in favor of --project-root for consistency across commands. However, backward compatibility is maintained - the old flag still works with a deprecation warning.

Migration Guide:

# Old (still works with warning)
cognition-cli init --path /my/project

# New (recommended)
cognition-cli init --project-root /my/project

🔧 Release Statistics

• 60 commits since v2.4.2
• 628 tests passing (38 skipped)
• All format, lint, and build checks passing
• Comprehensive integration tests for Gemini provider

📚 Links

• Full Changelog
• Documentation
• The Lattice Book - Complete reference manual
• Installation Guide

📋 Previous Releases

• v2.4.2 - Documentation & Stability
• v2.4.1 - Cross-Overlay Workflows
• v2.4.0 - Production Excellence
• v2.0.0 - Σ (Sigma) Infinite Context

Built with ❤️ for human-AI symbiosis