Agents are probabilistic. Infrastructure cannot be.
WatchLLM builds deterministic, local-first infrastructure that governs how autonomous coding agents read, write, and execute code — before unsafe changes reach disk.
| Layer | Component | Description |
|---|---|---|
| Governance | kernel |
Deterministic local runtime enforcement engine (Python, Tree-sitter AST) |
| Governance | runtime |
High-performance Rust + Wasm implementation with JSON schema parity |
| Orchestration | orchestrator |
Policy graph, model routing, and sub-agent spawning |
| Contracts | schemas |
Central event, rule, policy, and violation schema contracts (v1) |
| Memory | klyd |
Persistent architectural memory — extract, store, enforce design decisions |
| Observability | replay |
Execution graph replay, DAG rendering, and runtime forensics |
| Reliability | reliability |
Adversarial stress-testing, prompt injection, and sandboxed evaluation |
| Quality | benchmarks |
Ecosystem-wide performance and regression benchmarking |
| Editor | vscode |
VS Code extension — save interception, diagnostics, and inline governance |
| Docs | docs |
Architecture, RFCs, threat model, roadmap, and gap analysis |
| Demos | examples |
Curated scenarios: secrets, boundary enforcement, auth flow |