open-mind-capability-discovery 0.1.0

Capability discovery engine for open-mind — reads CAPABILITY.toml files and builds integration graphs
Documentation
  • Coverage
  • 88.37%
    38 out of 43 items documented0 out of 20 items with examples
  • Size
  • Source code size: 24.1 MB This is the summed size of all the files inside the crates.io package for this release.
  • Documentation size: 737.39 kB This is the summed size of all files generated by rustdoc for all configured targets
  • Ø build duration
  • this release: 6s Average build duration of successful builds.
  • all releases: 6s Average build duration of successful builds in releases after 2024-10-23.
  • Links
  • crates.io
  • Dependencies
  • Versions
  • Owners
  • SuperInstance

What Makes This Different

Open Interpreter lets LLMs run code. open-mind lets LLMs understand code — and then make decisions about how to execute it.

Three capabilities that open-interpreter doesn't have:

1. 🧠 Induction Engine

Ingest any GitHub repo (or local codebase) and the induction engine builds a living model of it:

  • Parses every function and class using AST (Python) or tree-sitter (Rust, C, C++, JavaScript, TypeScript)
  • Extracts signatures, docstrings, type hints, call graphs
  • Identifies test coverage — which functions are tested, which aren't
  • Builds dual-side vectors — input vectors (what triggers this function?) and output vectors (what does it produce?)
  • Stores everything in SQLite for fast retrieval
from interpreter.induction import ingest

result = ingest("https://github.com/your-org/your-repo")
print(f"Parsed {len(result.functions)} functions from {len(result.test_files)} test files")

2. 🎸 Tripartite Synchronizer

For every function in your codebase, the synchronizer decides: how much thinking does this need?

Decision When Cost Example
HARDCODE Hot path, tested, deterministic 0 tokens, ~1ms add(a, b) — just run it
CACHED Pre-computed, stable output 0 tokens, ~5ms Config lookups — replay the result
HYBRID Mostly stable, some edge cases ~50 tokens API calls — cache + fallback
MODEL Novel, creative, untested ~500 tokens, ~2s New feature generation — ask the LLM

Three inputs drive the decision:

  • Hardware profile — GPU? RAM? Battery? Edge device?
  • Application profile — Latency? Safety? Scale?
  • User profile — Manual control? Creative? Consistent?
from interpreter.induction import TripartiteSynchronizer, TriHardwareProfile, TriApplicationProfile

sync = TripartiteSynchronizer()
hw = TriHardwareProfile(compute_power=0.8, gpu_available=True)
app = TriApplicationProfile(latency_requirement_ms=10, safety_critical=True)
decision = sync.decide(hw, app, user_profile)

print(f"Decision: {decision.value}")  # "hardcode"
print(f"Reasoning: {decision.reasoning}")

3. 🔌 Export Bridges

Connect the induction engine to the real world:

  • lever-runner export → skill packs for the lever-runner command system
  • pincherOS .nail export → pre-computed results for edge devices (ESP32, Jetson)
  • Vector search → semantic function lookup across ingested repos
from interpreter.induction import export_lever_pack, export_nail

# Export HARDCODE functions as lever-runner commands
export_lever_pack_batch(hardcoded_functions, db_path="commands.db")

# Export CACHED functions as .nail files for pincherOS
export_nail_batch(cached_functions, export_dir="./nail-files/")

Quick Start

Install

git clone https://github.com/SuperInstance/open-mind.git
cd open-mind
pip install -e .

Use as Open Interpreter (everything still works)

interpreter

All original open-interpreter functionality is preserved. Chat with an LLM, run code, iterate.

Use the Induction Engine

from interpreter.induction import ingest, VectorBuilder, TripartiteSynchronizer

# Ingest a repo
result = ingest("https://github.com/SuperInstance/ternary-core")

# Build vectors
builder = VectorBuilder()
for func in result.functions:
    builder.build_function_vectors(func, "https://github.com/SuperInstance/ternary-core")

# Search for functions
matches = builder.search_input("multiply ternary vectors")
for m in matches:
    print(f"  {m.function_name}: {m.input_text[:80]}")

Run the Full Pipeline Demo

python demo_full_pipeline.py https://github.com/SuperInstance/lever-runner

This runs the complete loop: ingest → vectorize → hardware probe → tripartite decisions → export lever-runner pack → export .nail files.

Run the Tripartite Demo

python demo_tripartite.py

Shows the synchronizer making decisions for different hardware/application/user scenarios.

Architecture

open-mind/
├── interpreter/
│   ├── core/               # Open Interpreter core (unchanged)
│   │   ├── core.py         # Interpreter class
│   │   └── respond.py      # Response handling
│   ├── induction/           # ✨ THE NEW STUFF
│   │   ├── ingester.py     # Repo → functions/classes pipeline
│   │   ├── multi_lang_parser.py  # tree-sitter multi-language parser
│   │   ├── vector_builder.py     # Dual-side vector builder
│   │   ├── synchronizer.py       # Tripartite decision engine
│   │   ├── hardware_probe.py     # GPU/RAM/CPU detection
│   │   ├── spreader.py           # Continuous iteration engine
│   │   ├── profiles.py           # Pre-built scenarios
│   │   ├── export_lever.py       # lever-runner bridge
│   │   └── export_nail.py        # pincherOS bridge
│   ├── llm/                 # LLM setup (OpenAI, local, etc.)
│   ├── cli/                 # CLI interface
│   ├── code_interpreters/   # Python, JS, R, Shell execution
│   ├── terminal_interface/  # Terminal UI
│   ├── rag/                 # RAG pipeline
│   └── utils/               # Utilities
├── tests/
│   ├── test_induction.py    # Induction engine tests
│   ├── test_synchronizer.py # Synchronizer tests
│   └── test_interpreter.py  # Original interpreter tests
├── demo_full_pipeline.py    # Full pipeline demo
├── demo_tripartite.py       # Tripartite demo
└── induction-results/       # Stored ingestion results

The Induction Engine — Deep Dive

How Ingestion Works

  1. Clone the repo (or use a local path)
  2. Walk the file tree, skipping .git/, target/, node_modules/
  3. Detect language by file extension
  4. Parse each file:
    • Python → built-in ast module (always works)
    • Rust, C, C++, JS, TS → tree-sitter (optional, requires pip install tree-sitter-*)
  5. Extract: functions (name, signature, docstring, args, types, return type, calls), classes (name, bases, methods, docstring)
  6. Build call graph: who calls whom
  7. Detect tests: test_*.py, *_test.py, files with assert/#[test]
  8. Mark tested functions: search for function names in test content
  9. Return IngestResult: everything structured and queryable

The IngestResult

@dataclass
class IngestResult:
    repo_url: str
    local_path: str
    functions: list[FunctionInfo]   # Every parsed function
    classes: list[ClassInfo]        # Every parsed class
    test_files: list[str]           # Test file paths
    file_structure: dict            # Directory tree
    call_graph: dict[str, list[str]] # caller → [callees]
    stats: dict                     # Summary statistics

FunctionInfo

@dataclass
class FunctionInfo:
    name: str                    # "tdot"
    module: str                  # "src.lib"
    file_path: str               # "src/lib.rs"
    line_start: int              # 42
    line_end: int                # 48
    signature: str               # "def tdot(a: &[Trit], b: &[Trit]) -> Trit"
    docstring: Optional[str]     # "Inner product of two ternary vectors mod 3"
    source_code: str             # Full function source
    decorators: list[str]        # ["#[inline]"]
    arg_names: list[str]         # ["a", "b"]
    arg_types: dict              # {"a": "&[Trit]", "b": "&[Trit]"}
    return_type: Optional[str]   # "Trit"
    calls: list[str]             # Functions this calls
    called_by: list[str]         # Functions that call this (populated later)
    has_tests: bool              # Is this function tested?

The Tripartite Synchronizer — Deep Dive

Three Inputs

Hardware Profile (the body):

TriHardwareProfile(
    compute_power=0.8,      # 0-1, how much compute is available
    gpu_available=True,     # Is there a GPU?
    memory_gb=32.0,         # Available RAM
    battery_level=None,     # Battery % (None = plugged in)
    device_type="workstation",  # workstation, laptop, edge, mobile
)

Auto-detected via probe_hardware():

from interpreter.induction import probe_hardware
hw = probe_hardware()
print(f"GPU: {hw.gpu_name}, RAM: {hw.ram_gb}GB, Type: {hw.device_type}")

Application Profile (the task):

TriApplicationProfile(
    latency_requirement_ms=10,    # How fast must this be?
    accuracy_requirement=0.95,    # How correct must this be?
    safety_critical=True,         # Can errors hurt people?
    scale=1000,                   # How many times will this run?
    deterministic=True,           # Must this be reproducible?
)

User Profile (the human):

TriUserProfile(
    wants_manual_control=False,   # User wants to approve?
    wants_creativity=0.2,         # 0=deterministic, 1=creative
    wants_consistency=0.9,        # 0=variety, 1=same every time
    tolerance_for_error=0.1,      # 0=perfect, 1=yolo
)

Decision Matrix

Hardware Application User Decision
GPU, fast Safety-critical Consistent HARDCODE
Edge, low power Any Any CACHED
Any Novel, creative Creative MODEL
Any Mostly stable Manual control HYBRID
Battery low Flexible Consistent CACHED
Workstation Untested Explorer MODEL

Multi-Language Support

The induction engine parses code in multiple languages via tree-sitter:

Language Extensions Parser
Python .py Built-in ast (always works)
Rust .rs tree-sitter-rust
C .c, .h tree-sitter-c
C++ .cpp, .hpp, .cc tree-sitter-cpp
JavaScript .js tree-sitter-javascript
TypeScript .ts tree-sitter-typescript

Install tree-sitter support:

pip install tree-sitter tree-sitter-python tree-sitter-rust tree-sitter-c tree-sitter-javascript

Without tree-sitter, only Python files are parsed. With it, the engine handles all six languages.

The Spreader — Continuous Iteration

The Spreader runs multiple passes over an ingested repo, refining the model each time:

from interpreter.induction import Spreader

spreader = Spreader()
report = spreader.run(result, continuous=True, max_passes=5)
print(f"After {report.passes} passes: {report.functions_analyzed} functions analyzed")

Each pass:

  1. Re-evaluates tripartite decisions with updated call graphs
  2. Propagates confidence scores through the call chain
  3. Identifies new test coverage from changed files
  4. Updates vector embeddings with new context

Export Formats

lever-runner Skill Pack

Export HARDCODE functions as commands for the lever-runner system:

from interpreter.induction import export_lever_pack_batch

functions = [
    {"function_name": "authenticate", "module_path": "auth", "description": "Verify user credentials"},
    {"function_name": "hash_password", "module_path": "auth", "description": "Hash a password with bcrypt"},
]
records = export_lever_pack_batch(functions, db_path="commands.db")

pincherOS .nail Files

Export CACHED functions as pre-computed results for pincherOS edge devices:

from interpreter.induction import export_nail_batch

cached = [
    {"function_name": "get_config", "cached_output": "{...}", "description": "Load device configuration"},
]
manifests = export_nail_batch(cached, export_dir="./nail-cache/")

Hardware Probe

Auto-detect your system's capabilities:

from interpreter.induction import probe_hardware

hw = probe_hardware()
print(f"""
Device:  {hw.device_type}
Arch:    {hw.arch}
CPU:     {hw.cpu_cores} cores
RAM:     {hw.ram_gb:.1f} GB
GPU:     {'✅ ' + hw.gpu_name if hw.gpu else '❌ none'}
Battery: {f'{hw.battery_pct:.0f}%' if hw.battery_pct else 'plugged in'}
""")

Pre-Built Profiles

Common scenarios are pre-configured:

from interpreter.induction.profiles import PROFILES

# Autonomous driving — safety first, low latency
driving = PROFILES["autonomous_driving"]

# Creative writing — high creativity, flexible latency
creative = PROFILES["creative_writing"]

# Edge device — low power, cached results
edge = PROFILES["edge_sensor"]

# Developer tool — balanced, hybrid defaults
dev = PROFILES["developer_tool"]

Relation to the SuperInstance Ecosystem

open-mind is part of the SuperInstance ecosystem:

Project Role Link
openmind Standalone package (Python + CLI + Jupyter) SuperInstance/openmind
open-mind This repo — open-interpreter + induction engine You are here
agent-knowledge Agent-to-agent documentation SuperInstance/agent-knowledge
openmind-esp32-bridge Rust: motor neurons for ESP32 SuperInstance/openmind-esp32-bridge
openmind-conductor Rust: multi-agent orchestration SuperInstance/openmind-conductor
openmind-mirror Rust: self-reflection + coherence SuperInstance/openmind-mirror
openmind-cellular Resource-adaptive computation SuperInstance/openmind-cellular
ai-writings Creative writing + manifestos SuperInstance/ai-writings

The Standalone Package

If you don't need the open-interpreter chat interface and just want the induction engine + muscle memory + cellular computation, use the standalone package:

pip install openmind

See SuperInstance/openmind for the standalone package with 95 tests, CLI, Jupyter magic, and resource-adaptive computation.

The Agent Knowledge Base

For AI agents discovering this ecosystem, read the a2a documentation at agent-knowledge:

Development

git clone https://github.com/SuperInstance/open-mind.git
cd open-mind
pip install -e ".[dev]"
pytest

Running Tests

# All tests
pytest tests/

# Just induction tests
pytest tests/test_induction.py -v

# Just synchronizer tests
pytest tests/test_synchronizer.py -v

Project Structure

interpreter/
├── induction/           # ← Your playground
│   ├── ingester.py     # Try: add a new language parser
│   ├── synchronizer.py # Try: add a new decision factor
│   ├── vector_builder.py # Try: add a real embedding model
│   └── profiles.py     # Try: add a new scenario profile
├── core/               # Open Interpreter core
├── llm/                # LLM integrations
└── code_interpreters/  # Code execution environments

License

MIT — same as Open Interpreter.

Credits

Based on Open Interpreter by Killian Lucas.

Induction engine, tripartite synchronizer, and hardware probe by SuperInstance.