fix: use lora_alpha instead of alpha for PEFT compatibility

fix: set train_split to 0.999 to avoid train_test_split integer edge case

fix: remove bitsandbytes dependency to fix Kaggle CUDA compatibility

feat: standalone training and merge scripts for Kaggle

fix: add synthetic data fallback when training-data not available

fix: use correct input_path key for train_lora and ensure data path is correct

fix: update Kaggle notebook with proper data handling and training configuration

fix: change to /kaggle/working before cloning to avoid directory error

fix: support input_path in train_lora for JSONL files

fix: Kaggle notebook typo - os.chdir

fix: add train_dir and eval_dir to config for Kaggle

fix: fix git clone command in Kaggle notebook

fix: update Kaggle notebook with self-contained cells

docs: update README with training options and RTMP extraction

feat: add Kaggle training notebook

feat: combine RTMP tools with main training data

feat: add RTMP tool extraction scripts

feat: add local training script for Mac MPS

fix: skip model download to avoid crashes

fix: skip model download if it exists, avoid crashes

fix: add retry logic for git clone

fix: define all paths in first cell for kernel restart safety

fix: add imports to each cell for robustness

fix: create sample data if real data not found in repo

refactor: simplify notebook with single ROOT_DIR variable

fix: use different base path to avoid nested directory issue

fix: force fresh clone to get latest script fixes

fix: handle nested directory and data path issues

fix: add nested directory detection and auto-fix

fix: correct data path from training-data/final to data/final

fix: add error checking and absolute paths for config and data

fix: use absolute paths and fix bitsandbytes version

fix: syntax errors in notebook and create_mini_dataset script

fix: update notebook for Colab T4 GPU training

Fix colab training notebook for local Mac MPS training

docs: Add official launch plan

feat: Add free deployment support for Stack 2.9

refactor: Squeeze folders further - cleaner structure

refactor: Clean up project structure - fewer root folders

feat: add TypeScript modules and reorganize project structure

feat: zero-cost Colab training pipeline and proper evaluation

feat: Together AI integration complete

Address critical gaps: 1) Create comprehensive TOOLS.md with all 38 tools including parameters, examples, and JSON schema. 2) Fix get_tool_schemas() to auto-generate accurate schemas using inspect. 3) Fix architecture diagram alignment in ARCHITECTURE.md. 4) Remove vague 'autonomous sub-agents' claim from README; clarify enhancement sourcing. Verified cloud deployment scripts (RunPod/Vast) exist and are documented in stack-2.9-deploy/README.md.

Update README: add requested badges (Apache 2.0, OpenRouter, Hugging Face, HumanEval, MBPP) and acknowledge sub-agent enhancements

Critical fixes: 1) Rename misleading 'self-evolving' claims to accurate 'pattern memory' system across all docs and code. 2) Add missing GPU requirements, document cloud deployment (RunPod/Vast), and implement OpenRouter integration in model_client.py with factory support. 3) Document 37 built-in tools with full schemas in docs/tools.md. 4) Expose fraudulent evaluation scores (76.8% HumanEval, 82.3% MBPP, 94.1% Tool Use), remove them from README/BENCHMARKS/website, add EVALUATION.md audit report, and deprecation warnings to flawed eval scripts. Also updated HuggingFace Space demo with correct terminology.

docs: Add GitHub topics manifest

docs: Complete restructure README like premium AI products

feat: Complete Stack 2.9 TUI with benchmarks and pattern management

feat: Add evaluation and pattern management to CLI

feat: Add GSM8K math benchmark with model integration