Stop AI agents from breaking your architecture without you noticing. A coordination layer for repos with multiple LLM coding agents.
SKL is worth trying if you are running two or more AI coding agents on the same repo and have been surprised by what one of them changed, or want to know when agents make architectural decisions without asking.
SKL is probably not worth it if you use one agent at a time, review all agent output before it reaches your repo, or your codebase is small enough to hold in one context window.
- Install the VS Code extension
- Run SKL: Init (Phase 0) from the command palette
- Set
export SKL_AGENT_ID=Agent-1in your agent's terminal - Push a branch
You will see what your agent changed, what risk signals fired, and whether it touched files outside its expected scope — without configuring anything else. When you are ready for scope enforcement, RFCs, and the full governance layer: run SKL: Upgrade to Full SKL.
CLI alternative:
npm install -g @skl/cli
skl init
skl statusSKL is a Git hook plus a shared state file. When an agent pushes code, the hook records what changed, runs static analysis to flag risk signals, and checks whether the change crosses scope boundaries. You see the activity in plain language. If a change is architectural, an RFC is generated and the merge is blocked until you resolve it. If two agents make conflicting assumptions, SKL surfaces the conflict before either change lands.
| Feature | Phase 0 | Full SKL |
|---|---|---|
| Activity logging | ✓ | ✓ |
| Risk signals on push | ✓ | ✓ |
| Scope definitions required | ✗ | ✓ |
| Scope enforcement on push | ✗ | ✓ |
| RFC gate for architectural changes | ✗ | ✓ |
| Orchestrator review session | ✗ | ✓ |
| Human digest review | ✗ | ✓ |
| Queue max | 50 | Configurable |
- Import scanning is static only — dynamic imports are not detected
- Scope definition quality is load-bearing — a poorly defined scope produces false confidence in enforcement
- The Orchestrator requires an LLM (uses your existing Copilot model) and degrades to template fallbacks if none is available
- Full SKL adds review overhead — this is intentional, not a bug
- Not suitable for deep refactors, framework migrations, or cross-cutting changes where the correct scope is unknowable until implementation is underway
SKL is in active development. See PLAN.md for the full build history
and SPEC.md for the v1.4 reference specification that this extension
implements.
Issues and pull requests welcome. Please read PLAN.md before
contributing to understand the architecture and constraints.
A VS Code extension implementing the SKL v1.4 coordination protocol for small concurrent LLM agent teams.
SKL makes parallelism debuggable. It is not a safety system. It is a fail-stop coordination layer with bounded damage and explicit uncertainty.
| Stage | Description | Status |
|---|---|---|
| 1 | Data layer, schemas, initialization | ✅ Complete |
| 2 | Enforcement hook, Queue panel | ✅ Complete |
| 3 | Orchestrator | 🔄 In progress |
| 4 | CI integration, acceptance criteria | ⏳ Planned |
The full SKL v1.4 reference specification is in SPEC.md.
Built using spec-driven development with a single Copilot agent.
See PLAN.md for the staged build plan.
This project is in early development. Contributions are welcome once Stage 3 is complete. Check back soon.