AI Execution Router — AI skill for Claude Code

ai-execution-router

A [Claude Code](https://claude.com/claude-code) skill that decides **how** an AI agent should work on each software-engineering issue: which kind of model, how much reasoning, which workflow, and how hard the result must be verified. It then resolves that decision to the best concrete model available **on the day the work runs**, and records both on the GitHub issue.

**Status: v0.** The policy has been audited against scripted scenarios and blind readings, but the skill has not yet run against a real repository. Expect changes after the first real runs.

Contents

• Why it exists
• The core idea: policy now, model later
• A quick example
• How a run works
• Classification
• Roles, effort and effort control
• The model registry
• Resolving a model
• Ties and preferences
• What gets written to GitHub
• Optional: evaluating models
• Installation
• Usage
• Configuration
• Repository layout
• Design principles
• Current status and limitations
• License

Why it exists

Handing issues to AI agents raises the same questions every time:

• Which model? A frontier model on a one-line rename wastes money. A fast model on concurrent message handling ships bugs.
• How much reasoning? Most providers now expose an effort or reasoning setting. The right level depends on the task, not on the model.
• Who checks the work? Code that touches money, security or concurrency needs a stronger, independent review than a text change.
• For how long is the answer valid? Model names change every few months. An issue that says "use