v0.1.10

What's New

AiChat Agent

New agent integrating aichat CLI for multi-provider LLM access (20+ backends). Includes:

• start_aichat_session / resume_aichat_session MCP tools
• Full council participation (Round 1 + Round 2 deliberation)
• Role system and team preset support
• Configurable model via AICHAT_MODEL env var

Improvements

• Skip timed-out agents in council Round 2 instead of failing
• Animated council data flow diagram in README

Changelog

• 1af9f3f Add aichat agent support for multi-provider LLM access
• fc3dd62 Add aichat agent to README documentation
• 50e8938 Skip timed-out agents in council R2
• 73c6782 Add animated council data flow diagram to README

Install / Upgrade

uv tool install --force git+https://github.com/agentic-mcp-tools/owlex.git

v0.1.8

What's New

ClaudeOR Agent

• New agent: Claude Code via OpenRouter for alternative models (DeepSeek, GPT-4o, Gemini, etc.)
• Tools: start_claudeor_session, resume_claudeor_session
• Env vars: OPENROUTER_API_KEY, CLAUDEOR_MODEL

Bug Fixes

• Fix stdin hang for Claude CLI in non-streaming mode (use DEVNULL instead of PIPE)
• Check env vars at runtime for ClaudeOR (fixes .mcp.json config issues)
• Fix OpenRouter env vars per official docs

Documentation

• Add ClaudeOR to team presets table
• Add optimal team preset (roles matched to model strengths)
• Add specialist roles documentation

v0.1.7 - OpenCode Integration & Claude Code Skills

Highlights

• OpenCode as standalone agent - New start_opencode_session and resume_opencode_session tools for direct OpenCode
  invocation
• Claude Code skills - Non-blocking /codex, /gemini, /council, and /critique slash commands
• Fixed JSON double-encoding - MCP responses now return proper objects instead of escaped JSON strings

What's New

Features

• OpenCode standalone MCP tools for independent agent invocation
• Claude Code plugin with delegation subagents
• MCP resources and notification support
• Skills for non-blocking agent calls: codex, gemini, council, critique

Fixes

• Fixed double JSON encoding in all MCP tool responses
• Fixed Codex model name (gpt5.2-codex)

Improvements

• Improved README with value proposition and inline demo GIF
• Optimized demo GIF size (2.5MB)

Council

All three agents now supported:

• Codex (OpenAI)
• Gemini (Google)
• OpenCode (open-source)

Full Changelog: v0.1.6...v0.1.7

v0.1.6

• Async task creation for council deliberation
• Proper task status tracking (pending → running → completed/failed)
• Error handling with task failure states

Full Changelog

v0.1.5...v0.1.6

v0.1.5

Full Changelog: v0.1.1...v0.1.5

v0.1.2 - Polymorphic Runners & Hardening

What's New

Architecture Refactoring

• Unified run_agent method - Single polymorphic entry point in TaskEngine replacing agent-specific methods
• Agent enum - Type-safe agent identification
• AGENT_RUNNERS registry - Clean mapping of agents to their runners
• Extracted prompts module - Council deliberation templates in dedicated module

Security & Robustness

• CLI flag injection prevention - Validates session_ref parameters to block malicious flags
• Config crash prevention - Graceful fallbacks for invalid environment variables
• Structured error codes - ErrorCode enum for consistent MCP responses
• Zombie task cleanup - Automatic cleanup of stale tasks on startup
• Graceful termination - SIGTERM handling for subprocess cleanup
• Gemini CLI fix - Resolved stdin wait issue causing timeouts

Council Improvements

• Removed synthesis phase (MCP client now synthesizes responses)
• Simplified round 1/2 structure with deliberation and critique modes

Testing

• 86 tests total, all passing
• New test_config.py - 8 tests for config parsing edge cases
• New test_server.py - 19 tests for MCP validation paths
• New test_agents.py - Flag injection prevention tests
• Updated council tests for new architecture

v0.1.1 - Gemini CLI Integration

• add adviser tool use using gemini. ## What's New

  Gemini CLI Support

  Owlex now supports Gemini CLI alongside Codex CLI, enabling multi-AI delegation from Claude Code:

  • start_gemini_session - Fresh Gemini session with 1M token context
  • resume_gemini_session - Continue with preserved conversation history
  • Async task execution with wait_for_task / get_task_result

  When to Use Each

  Task	Provider	Why
  Large codebase analysis	Gemini	1M token context
  Code review & bug finding	Codex	Purpose-built for flaws
  Multimodal (images, video)	Gemini	State-of-the-art
  PRD & requirements	Codex	Socratic questioning

  Other Changes

  • MIT License added
  • Improved notification messages
  • Better error handling for encoding issues
  • README badges and project metadata

  ---

  Full Changelog: v0.1.0...v0.1.1

v0.1.0 - Simplified Package Structure

What's New

• New tool: start_codex_session - Start fresh Codex sessions without prior context
• Simplified API - Reduced to 4 essential tools:
  • start_codex_session - new session
  • resume_codex_session - continue from last session (uses --last by default)
  • wait_for_task - wait for completion
  • get_task_result - check task status
• New options: --cd (working_directory) and --search (enable_search)
• Package restructure - Now installable via uv tool install

Breaking Changes

• Removed legacy tools (review_plan, execute_codex_command, etc.)
• Changed from single script to Python package structure