Skip to content

Latest commit

 

History

History
43 lines (33 loc) · 2.13 KB

File metadata and controls

43 lines (33 loc) · 2.13 KB

ARCHITECTURE.md

# Architecture Overview

Tinkerflow-AI is built as a modular system with three main components:

┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐ │ GitHub App │────▶│ Governance │────▶│ LLM Gateway │ │ (webhooks) │ │ Engine │ │ (router) │ └─────────────────┘ └─────────────────┘ └─────────────────┘ │ │ ▼ ▼ ┌─────────────────┐ ┌─────────────────┐ │ Policy │ │ Providers │ │ Checker │ │ (OpenAI, etc.)│ └─────────────────┘ └─────────────────┘


### Components

1. **GitHub App** – Listens for pull_request events, authenticates with GitHub, and triggers governance.
2. **Governance Engine** – Orchestrates policy evaluation, proof generation, and feedback posting.
3. **LLM Gateway** – Routes prompts to appropriate LLM providers with fallback logic.
4. **Policy Checker** – Evaluates policies against PR diffs using regex, AST, or LLM calls.
5. **Providers** – Adapters for Ollama, OpenAI, Anthropic, etc.

### Data Flow

1. Webhook received → authentication middleware verifies signature.
2. Payload parsed, PR diff fetched via GitHub API.
3. Governance engine loads policies from repo config or global config.
4. Each policy evaluated; results are hashed to create cryptographic proofs.
5. Proofs optionally verified (external service or local).
6. Check run created; if any policy fails, PR merge blocked.
7. Results logged and audit trail stored.

### Deterministic Governance

All decisions are based solely on the PR content and fixed policies. No external factors alter the outcome, ensuring reproducibility.