batuta 0.7.3

Sovereign AI orchestration: autonomous agents, ML serving, code analysis, and transpilation pipelines
Documentation

Installation | Quick Start | Features | Stack Components | Documentation

Table of Contents

What is Batuta?

Batuta is the orchestration CLI for the Sovereign AI Stack -- a pure-Rust ecosystem for privacy-preserving ML infrastructure. It coordinates 15+ crates spanning compute, training, inference, and serving -- with zero Python dependencies.

# Analyze any codebase
batuta analyze --tdg .

# Query the stack oracle
batuta oracle "How do I serve a Llama model locally?"

# Serve models (OpenAI-compatible API)
batuta serve ./model.gguf --port 8080

# Autonomous coding agent
batuta code --prompt "Summarize this codebase"

Installation

# From crates.io
cargo install batuta

# With autonomous agents
cargo install batuta --features agents

Quick Start

Analyze a project

batuta analyze --tdg .

Analysis Results
  Files: 440 total, 98,000 lines
  Languages: Rust (95%), TOML (3%), Markdown (2%)
  TDG Score: 98.4 (Grade: A+)

Hunt for bugs

batuta bug-hunter analyze .

Finds unwraps, panics, unsafe blocks, error swallowing, and 20+ fault patterns across your codebase.

Query the stack oracle

batuta oracle "How do I train a random forest?"

Returns component recommendations with working code examples and TDD test companions.

Serve a model

batuta serve ./model.gguf --port 8080

Starts an OpenAI-compatible server at http://localhost:8080/v1/chat/completions.

Features

  • Code Analysis -- TDG scoring, bug hunting, Popperian falsification testing
  • Oracle Queries -- Natural language queries with RAG-based documentation search
  • Model Serving -- OpenAI-compatible endpoints with privacy tiers (Sovereign/Private/Standard)
  • Autonomous Agents -- Perceive-reason-act loop with 9 tools and formal contract invariants
  • Stack Orchestration -- Version drift detection, publish-status, release pipelines for 15+ crates
  • Transpilation -- Python/Shell/C to Rust conversion via depyler/bashrs/decy
  • Playbooks -- Deterministic YAML pipelines with BLAKE3 content-addressed caching

Agent Runtime

The agent runtime provides an autonomous coding assistant with local LLM inference:

# Single-prompt mode (GPU-accelerated)
batuta code --prompt "Explain the error handling in this project"

# Interactive chat
batuta code

# With explicit model
batuta code --model ./Qwen3-8B-Q4_K_M.gguf --prompt "Add unit tests"

Agents are configured via TOML manifests with capability-gated tools (shell, filesystem, network, RAG, MCP), privacy enforcement, and circuit-breaker guards.

See the Agent Runtime Book Chapter for details.

Stack Components

+-------------------------------------------------------------+
|                    batuta (Orchestration)                    |
+-------------------------------------------------------------+
|  whisper-apr (ASR)  |  realizar (Inference)  | pacha (Reg)  |
+---------------------+------------------------+--------------+
|   aprender (ML)   |  entrenar (Training)  | jugar (Games)  |
+-------------------+-----------------------+----------------+
|   simular (Sim)   |   profesor (Edu)      |                |
+-------------------+-----------------------+----------------+
|                 repartir (Distributed Compute)              |
+-------------------------------------------------------------+
|  trueno-zram (Compression)  |  trueno-ublk (Block Device)  |
+-----------------------------+------------------------------+
|               trueno (SIMD/GPU Compute Primitives)          |
+-------------------------------------------------------------+
Component Version Description
trueno 0.16 SIMD/GPU compute (AVX2/AVX-512/NEON, wgpu, LZ4)
aprender 0.27 ML algorithms: regression, trees, clustering, NLP
entrenar 0.7 Training: autograd, LoRA/QLoRA, quantization
realizar 0.8 LLM inference for GGUF/SafeTensors/APR models
repartir 2.0 Distributed compute (CPU/GPU/Remote executors)
whisper-apr 0.2 Pure Rust Whisper ASR (WASM-first)
ttop 2.0 Sovereign system monitor (14 panels, GPU support)
presentar-terminal 0.3 Zero-alloc TUI rendering
pacha 0.2 Model registry with Ed25519 signatures
renacer 0.10 Syscall tracing with semantic validation
pmat 3.x Code quality analysis and TDG scoring

CLI Reference

batuta analyze        Analyze project structure, languages, TDG score
batuta bug-hunter     Proactive bug hunting (fault patterns, mutation targets)
batuta falsify        Popperian falsification checklist
batuta oracle         Natural language queries about the Sovereign AI Stack
batuta serve          ML model serving (OpenAI-compatible API)
batuta code           Autonomous coding agent (local LLM inference)
batuta stack          Stack version management, drift detection
batuta playbook       Deterministic YAML pipeline runner
batuta transpile      Code transpilation (Python/Shell/C -> Rust)
batuta hf             HuggingFace Hub integration

Privacy Tiers

Tier Behavior Use Case
Sovereign Blocks ALL external API calls Healthcare, Government
Private VPC/dedicated endpoints only Financial services
Standard Public APIs allowed General deployment

Quality

Metric Value
Tests 6,258 passing
Coverage 95%+ line coverage
TDG Score 94.7 (A)
Clippy Zero warnings
Contracts 13 provable contracts, 129 FALSIFY tests

Falsifiable Quality Commitments

Commitment Threshold Verification
Test coverage >= 95% line coverage cargo llvm-cov (CI enforced)
Clippy clean Zero warnings cargo clippy -- -D warnings
Contract enforcement 13 contracts, 129 tests pv lint + FALSIFY suite
TDG grade A or above pmat analyze tdg
Build time < 2 minutes incremental cargo build --timings

Development

git clone https://github.com/paiml/batuta.git
cd batuta

cargo build --release          # Build
cargo test --lib               # Unit tests
cargo clippy -- -D warnings    # Lint
make book                      # Build documentation

Documentation

Contributing

See CONTRIBUTING.md for guidelines.

License

MIT License -- see LICENSE for details.