Skip to content

Releases: kriely/CLMA

v1.1.0 - AAN Chain Iteration, Node.js Sandbox, CI and Major Fixes

04 May 11:46
@kriely kriely
v1.1.0
e51c147

Choose a tag to compare

View all tags
v1.1.0 - AAN Chain Iteration, Node.js Sandbox, CI and Major Fixes Latest
Latest

v1.1.0 — AAN Chain Iteration, Node.js Sandbox, CI & Major Fixes

This release brings major improvements to the Adaptive Agent Network (AAN), adds JavaScript/Node.js execution support, and overhauls the project's test infrastructure.

✨ New Features

AAN Chain Closed-Loop Iteration

  • _aan_execute_chain now supports up to 3 rounds of closed-loop iteration
  • When the Evaluator score falls below threshold (default 0.7), the Refiner automatically receives the previous round's Verifier feedback + score guidance
  • Each round produces progressively higher-quality output
  • 6 new test scenarios covering: iteration events, threshold stop, round numbering, max 3 rounds, mid-stream cancellation

JavaScript / Node.js Sandbox Support

  • New execution environment for JavaScript code via Node.js
  • Fully integrated with the sandbox tiering system
  • Configurable timeout and enable/disable in Tools Configuration panel

CI Pipeline

  • GitHub Actions CI workflow (ci.yml) for automated testing
  • Runs on push and pull request to main branch

🐛 Bug Fixes

  • AAN Router classification fix: Most queries now correctly route to chain topology instead of incorrectly falling through to direct mode
  • Removed stale test files: Cleaned up 6 root-level test scripts migrated to organized tests/ directory
  • Configuration panels: API/Rules/Tools configuration panels in Web UI now properly save and restore all settings

🛠 Technical Improvements

  • Test infrastructure overhaul: Consolidated all tests into tests/test_aan.py (506 lines) with comprehensive coverage
  • Session store refactored: Improved state isolation between queries in session_store.py
  • Web UI v2: Major frontend overhaul in app.js (+1659 lines) and style.css (+1059 lines) with SSE event-driven architecture
  • Removed brand names from documentation: Cleaned up references in README and docs

📁 Included in Source Bundle

The attached clma-v1.1.0-source.tar.gz (9.9MB) contains the complete source tree:

  • C++17 Core Engine → src/, include/
  • Python Interface → python_interface/ (Flask Web UI + agent orchestration)
  • Pybind11 Bindings → CMakeLists.txt
  • Tests → tests/test_aan.py
  • Web UI → templates/index.html, static/app.js, static/style.css
  • Documentation → README.md, docs/, blog/
  • GitHub Config → .github/workflows/ci.yml

Excludes: venv/, pycache/, build/, compiled objects, and runtime data.

📋 Full Changelog (since v1.0.0)

5267ab2 AAN Chain closed-loop iteration
46b1729 feat(sandbox): add JavaScript/Node.js execution support
41466af fix(aan): repair router topology classification
080e5c2 Create ci.yml
ff6ae4a docs: add Single-Loop and Multi-Loop UI screenshots
458c8fb Update README.md
951f834 docs: remove brand names from Motivation section
a09af7b chore: clean up git-tracked artifacts

GitHub: https://github.com/kriely/CLMA

Assets 3
Loading

v1.0.0

30 Apr 12:37
@kriely kriely
v1.0.0
e92d911

Choose a tag to compare

View all tags
v1.0.0

v1.0.0 — Initial Release

CLMA: Closed-Loop Multi-Agent Framework

A self-verifying code generation system that automates the validation loop — no human-in-the-middle required.

Core Features

  • Self-verifying pipeline: Refiner → Reasoner → Solver → Verifier → Evaluator → auto-iterate until score ≥ threshold
  • 4 execution modes: Fast Path (~2s), Single Loop (~5s), DAG (~10s), Nested Multi-Loop (~40s), auto-selected by query classifier
  • C++17 core engine: Orchestrator, DAG Processor, RuleEngine, TokenMonitor, LoopController, PluginManager, Sandbox
  • 5 agent roles: fully pluggable via C++ .so plugins
  • Flask Web UI: dark theme, real-time SVG flow graph, SSE streaming, score gauge, execution timeline
  • 5 LLM providers: OpenAI, Anthropic, DeepSeek, Gemini, Local (Ollama/vLLM) — unified adapter pattern
  • 114 passing tests: 62 C++ unit tests + 46 Python tests + 6 integration tests

Benchmark (DeepSeek API)

Task Mode Time Score
Hello World Fast Path 2.3s 0.97
Fibonacci Single Loop 4.7s 0.99
Quicksort Single Loop 5.1s 0.98
Batch file rename Single Loop 8.1s 0.98
Microservice architecture Nested Loop 39s 1.0

Installation

git clone https://github.com/kriely/CLMA.git
cd CLMA
python3 -m venv venv
source venv/bin/activate
pip install flask flask-cors pybind11
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)
cd ..
./run_webui.sh

Contributors

Loading