Enterprise-grade orchestration for AI-assisted development.
Projeto de estudo em Rust — Orquestração de nível empresarial para desenvolvimento assistido por IA. Todo o código-fonte está comentado em português brasileiro (PT-BR) para fins didáticos.
WOLRAM applies battle-tested enterprise automation patterns (inspired by UiPath's REFramework) to AI coding workflows — bringing state machine governance, retry logic, model routing, and full audit trails to LLM-powered development.
Brainstorm / Reasoning
|
v
TODO Generation (10-100 items)
|
v
┌───────────────────────────────────────────┐
│ JOB STATE MACHINE (per task) │
│ │
│ INIT ──> DEFINE AGENT ──> PROCESS ──> END│
│ │ │ │ │ │
│ │ Skill Router Execute Task Log │
│ │ Model Selector (w/ retries) Git │
│ │ │ │
│ │ ┌─────┴─────┐ │
│ │ Success Failure │
│ │ │ (Biz/System) │
│ │ Commit Retry <= Max? │
│ │ │ Y: Re-queue │
│ │ │ N: Log fail │
└───┴───────────────────┴─────────┴─────────┘
|
v
GIT INTEGRATION
- Job completed --> commit (with summary)
- Pool of jobs (e.g. hero page, login) --> branch
- Full build --> PR with audit trail
| Concept | Description |
|---|---|
| Job | A single task extracted from the TODO list |
| State Machine | Each job passes through INIT > DEFINE AGENT > PROCESS > END |
| Skill Router | Assigns the right agent/skill using weighted keyword scoring or LLM-based classification |
| Model Selector | Picks the most cost-effective model using keyword + heuristic scoring (Haiku for simple, Sonnet for medium, Opus for complex) |
| LLM Classification | Optional Haiku pre-classification call that routes jobs by skill and complexity, with automatic fallback to keyword scoring |
| Model Override | CLI --model flag to force a specific model tier, overriding both LLM and keyword-based selection |
| Business Failure | Task logic failed (wrong output, validation error) — retryable |
| System Failure | Infrastructure failed (API timeout, rate limit) — retryable |
| Audit Trail | Every job logs: timestamp, model, skill, status, retry count, cost |
cargo build # Compile
cargo run -- demo # Run the built-in state machine demo
cargo run -- run "implement X" # Run a single job (stub mode without API key)
cargo run -- run --model opus "fix the typo" # Force Opus model
cargo test # Run all tests (80 tests)Set ANTHROPIC_API_KEY to enable real API calls and LLM-based classification; otherwise jobs run in stub mode.
WOLRAM uses a layered routing strategy during the DEFINE_AGENT phase:
- LLM classification (if API key is set) — sends a Haiku call to classify the job into a skill and complexity level
- Weighted keyword scoring (fallback) — sums weights of matched keywords to pick the best skill and model tier
- CLI override (
--model) — always takes precedence for model selection
| Keyword | Skill | Weight |
|---|---|---|
| test, spec | testing | 10, 5 |
| refactor, clean up | refactoring | 10, 5 |
| doc, readme | documentation | 10, 5 |
| fix, bug, debug, error | bug_fix | 10, 10, 7, 5 |
| implement, add, create, build | code_generation | 5, 3, 5, 5 |
- Simple keywords (rename, format, typo, delete, remove, update) push toward Haiku
- Complex keywords (architect, refactor, redesign, migrate, multi-file, system, overhaul) push toward Opus
- Short descriptions (<20 chars) boost simple score; long descriptions (>100 chars) and high word count (>15 words) boost complex score
- Default is Sonnet when scores are inconclusive
- CLI interface (terminal-first)
- TODO generation from natural language prompts
- 4-stage state machine with configurable retry logic
- Intelligent skill/model routing per job
- Git integration (auto-commit, branching)
- Audit trail with timestamp, model, cost tracking
- LLM-based job classification with keyword fallback
- CLI model override (
--modelflag) - CI/CD pipeline via GitHub Actions
- Web interface (wolram.com.br)
- Rust (edition 2024) — core runtime
- Anthropic API — Claude models for task execution and job classification
- git2 — programmatic git operations
- clap — CLI argument parsing
- tokio — async runtime
- indicatif/console — terminal UI with spinners and colored output
Contributions are welcome! Please run cargo fmt, cargo clippy, and cargo test before submitting a PR.
Code comments are written in Brazilian Portuguese (PT-BR). Test function names follow Rust convention (English).
v0.1.0 — Core state machine, CLI, Anthropic HTTP client, intelligent skill/model routing, git integration, and terminal UI are implemented.
Marlow Sousa — @wolram | wolram.com.br