Advanced AI assistant powered by GPT-5.2-Codex with 400K context, 128K output, and persistent memory.
git clone https://github.com/nschlaepfer/amarillo.git && cd amarillo
pip install -e .
cp .env.example .env # Add your OPENAI_API_KEY
amarilloOr run directly:
python -m amarillo.mainRequirements: Python 3.10+, OpenAI API key
| Feature | Description |
|---|---|
| GPT-5.2-Codex | 56.4% SWE-Bench Pro, 64.0% Terminal-Bench 2.0 |
| 400K Context | Handle massive codebases |
| ReasoningBank | Persistent memory across sessions |
| Multi-level Reasoning | none/low/medium/high/xhigh effort levels |
| Function Calling | OpenAI tools API integration |
| Scratch Pad | Organized note storage system |
pip install -e ".[dev]"pip install .Edit .env:
OPENAI_API_KEY_SANDBOX=sk-...
OPENAI_MODEL_CODE=gpt-5.2-codex
OPENAI_MODEL_REASONING_EFFORT=high
OPENAI_MODEL_COMPACTION=true| Command | Description |
|---|---|
exit / quit |
End session |
/save <cat> <file> <content> |
Save to scratch pad |
/view_scratch <cat> <file> |
View scratch pad file |
/list_scratch_pad [cat] |
List scratch pad files |
/edit_scratch <cat> <file> <content> |
Edit scratch pad file |
/delete_scratch <cat> <file> |
Delete scratch pad file |
/search_scratch <query> |
Search scratch pad |
/edit <idx> <content> |
Edit chat message |
/remove <idx> |
Remove chat message |
- deep_reasoning - Deep STEM reasoning with configurable effort
- retrieve_knowledge - Information retrieval
- coder - Code generation with GPT-5.2-Codex
- Scratch pad operations - save, edit, list, view, delete, search
Inspired by Poetiq's ARC-AGI solver and Emdash, these modules enable adaptive reasoning:
Iterative refinement with feedback accumulation:
- Generate hypothesis, evaluate, build feedback, iterate
- Solution pool with probabilistic selection
- Multi-expert voting for ensemble reasoning
- Early termination on success
from amarillo.reasoning import FluidReasoner, ExpertConfig
reasoner = FluidReasoner(client, evaluate_fn, config)
result = await reasoner.solve(task)
# Result includes: final_output, solutions, best_score, iteration_countSession state management with persistence:
- Registry pattern (getOrCreate)
- Automatic snapshot/restore
- Activity tracking and event listeners
- Workspace lifecycle (attach/detach/dispose)
from amarillo.memory import WorkspaceManager, get_workspace_manager
manager = get_workspace_manager()
task = manager.create_task("Implement sorting")
task.add_reasoning_step("analysis", "Breaking down the problem...")Partial credit evaluation system:
- Multi-dimensional scoring criteria
- Weighted aggregation
- Actionable feedback generation
- Pre-built scorers (syntax, keywords, execution)
from amarillo.reasoning import SoftScorer, code_syntax_scorer
scorer = SoftScorer(success_threshold=0.8)
scorer.add_criterion("syntax", code_syntax_scorer, weight=1.0)
result = scorer.score(solution, context)
# Result includes: score, success, dimension_scores, feedbackpython examples/fluid_intelligence_demo.pyamarillo/
├── src/amarillo/ # Main package
│ ├── __init__.py # Package exports
│ ├── main.py # CLI entry point
│ ├── core/ # Configuration & utilities
│ │ ├── config.py # Environment & settings
│ │ ├── constants.py # Magic numbers & defaults
│ │ └── utils.py # Helper functions
│ ├── memory/ # State management
│ │ ├── chat_history.py # Conversation tracking
│ │ ├── reasoning_bank.py # Persistent memory
│ │ ├── notes.py # Notes system
│ │ └── workspace_manager.py # Session state
│ ├── reasoning/ # Intelligence modules
│ │ ├── fluid_reasoning.py # Iterative refinement
│ │ ├── ensemble_reasoning.py # Multi-expert voting
│ │ └── scoring.py # Soft scoring system
│ └── tools/ # Tool definitions
│ ├── tools.py # Function schemas
│ └── file_tools.py # File operations
├── tests/ # Test suite
├── examples/ # Demo scripts
├── docs/ # Documentation & data
├── scripts/ # Utility scripts
├── pyproject.toml # Package configuration
├── requirements.txt # Dependencies
└── .env.example # Environment template
pytestblack src/ tests/
ruff check src/ tests/MIT - @nschlaepfer