Vibe-Skills

Routes natural language user intents to specific skill packs by analyzing intent keywords and context rather than allowing models to hallucinate tool selection. The router enforces priority and exclusivity rules, mapping requests through a deterministic decision tree that bridges user intent to governed execution paths. This prevents 'skill sleep' (where models forget available tools) by maintaining explicit routing authority separate from runtime execution.

Enforces a fixed, multi-stage execution pipeline (6 stages) that transforms requests through requirement clarification, planning, execution, verification, and governance gates. Each stage has defined entry/exit criteria and governance checkpoints, preventing 'black-box sprinting' where execution happens without requirement validation. The runtime maintains traceability and enforces stability through the VCO (Vibe Core Orchestrator) engine.

Provides a formal process for onboarding custom skills into the Vibe-Skills library, including skill contract definition, governance verification, testing infrastructure, and contribution review. Custom skills must define JSON schemas, implement skill contracts, pass verification gates, and undergo governance review before being added to the library. This ensures all skills meet quality and governance standards. The onboarding process is documented and reproducible.

Defines explicit skill contracts using JSON schemas that specify input types, output types, required parameters, and execution constraints. Contracts are validated at skill composition time (preventing incompatible combinations) and at execution time (ensuring inputs/outputs match schema). Schema validation is strict — skills that produce outputs not matching their contract will fail verification gates. This enables type-safe skill composition and prevents runtime type errors.

Provides testing infrastructure that validates skill execution independently of the runtime environment. Tests include unit tests for individual skills, integration tests for skill compositions, and replay tests that re-execute recorded execution traces to ensure reproducibility. Replay tests capture execution history and can re-run them to verify behavior hasn't changed. This enables regression testing and ensures skills behave consistently across versions.

Maintains a tool registry that maps skill identifiers to implementations and supports fallback chains where if a primary skill fails, alternative skills can be invoked automatically. Fallback chains are defined in skill pack manifests and can be nested (fallback to fallback). The registry tracks skill availability, version compatibility, and execution history. Failed skills are logged and can trigger alerts or manual intervention.

Generates proof bundles that contain execution traces, verification results, and governance validation reports for skills. Proof bundles serve as evidence that skills have been tested and validated. Platform promotion uses proof bundles to validate skills before promoting them to production. This creates an audit trail of skill validation and enables compliance verification.

Automatically scales agent execution between three modes: M (single-agent, lightweight), L (multi-stage, coordinated), and XL (multi-agent, distributed). The system analyzes task complexity and available resources to select the appropriate execution grade, then configures the runtime accordingly. This prevents over-provisioning simple tasks while ensuring complex workflows have sufficient coordination infrastructure.

Provides a curated library of 340+ specialized skills organized into domain-specific packs (planning, engineering, life sciences, etc.) with explicit pack manifests defining dependencies, contracts, and execution requirements. Skills are not loose functions but governed entities with JSON schemas, version constraints, and verification gates. The pack manifest system enables composition of complex workflows from verified skill combinations.

Integrates with the Model Context Protocol (MCP) to expose Vibe-Skills capabilities as standardized MCP entry points, enabling seamless integration with MCP-compatible clients (Claude, other LLMs). The integration maps skill packs to MCP resources and tools, allowing external models to discover and invoke Vibe-Skills without custom API wrappers. This enables Vibe-Skills to function as a backend service for multiple LLM frontends.

Enforces 129 configuration-based governance policies that prevent workspace drift, protect against skill misuse, and ensure execution compliance. Policies are declaratively defined (not hard-coded) and can be composed to create complex governance rules. The governance engine validates requests at routing, planning, and execution stages, blocking non-compliant operations before they execute. Policies cover skill access control, resource limits, output validation, and workspace isolation.

Implements multi-stage verification gates that validate skill execution, output compliance, and governance adherence before results are returned. Verification gates use JSON schemas to validate skill contracts, check output against content policies, and ensure resource usage stays within limits. Gates can be chained to create complex validation pipelines. Failed verifications block execution and provide detailed remediation guidance.

Manages skill pack versions, dependencies, and release lifecycles through explicit version governance rules. The system tracks version compatibility, enforces semantic versioning constraints, and prevents incompatible skill combinations from being deployed. Release lifecycle includes staging, testing, and promotion gates. This ensures reproducible agent behavior across deployments and prevents version drift.

Provides router overlay mechanisms that accelerate routing decisions by caching intent-to-skill mappings and using LLM-based pattern matching to predict likely skill selections. Overlays can be stacked to create multi-level routing hierarchies. The system learns from execution history to optimize routing performance, reducing latency for common intent patterns. Overlays are transparent to the underlying canonical router.

Provides host adapters that enable Vibe-Skills to run on different deployment targets (local, cloud, edge) without code changes. Distribution lanes define deployment configurations for different environments (dev, staging, production). Adapters handle platform-specific concerns (containerization, scaling, monitoring) while maintaining consistent skill execution semantics. This enables the same skill pack to run on Kubernetes, serverless, or local environments.