Route, don't guess

The machine's orchestrator is designed to route each task to the most optimal model based on the task's complexity, rather than simply choosing a higher-cost model for complex tasks. This routing layer is a contract file, a handful of dispatch templates, and an append-only ledger, all read by a Claude Code session before delegating any work.

There's no external scheduler or installed dependencies. The routing policy focuses on task shape, specifically the number of correction rounds required and whether a round can be objectively checked. Tasks that can be objectively checked are routed down a cost ladder, starting with scripts and local models, before moving to low-cost hosted tiers.

Tasks that require iterative and judgment-heavy processes, such as design, debugging, architecture, and review, stay with the top-tier model throughout. Low-cost tiers incur several additional costs, including cold-start tax, review tax, and retry tax. The policy eliminates the use of low-cost models for tasks requiring taste-sensitive output, terminal review, green light verification, and any other tasks.

Every delegated run ends with a status block in the chat and a row appended to a ledger file, which records what was routed where, whether it passed, and any lessons learned. The policy tightens based on field evidence from the ledger, not just design sessions. The project's near miss involved a decision to run git rm --cached against a sibling project's repository, which was stopped by the session before execution.

This exposed the default behavior of any session potentially editing neighboring projects' files, which now routes through a proposal note instead of asserting the edit. The whole project was built using native primitives of Claude Code over the course of eleven days, with each day closing with four working files that provided context for the next session.