The self-distilling model router for AI coding CLIs — solve each problem class at the expensive tier once, execute it cheap forever. Just markdown, no proxy.
Agent Claude Code
Standard implementation agent for well-specified tasks tagged [T1] — features from a clear spec, unit tests, documentation, renames, config changes, mechanical multi-file edits, straightforward bugfixes with a known repro. MUST BE USED for [T1] tasks.
Agent Claude Code
Complex implementation agent for tasks tagged [T2] — multi-file refactors, concurrency/async, schema and data migrations, performance optimization, intricate state machines, bugs with unclear repro, subtle correctness work. MUST BE USED for [T2] tasks.
Agent Claude Code
Highest-reasoning tier, read-only. Use for (a) REVIEW of critical diffs — security, auth, payments, data migrations, public API contracts; (b) SOLVE — designing novel algorithms or root-causing bugs that survived two fix attempts. Expensive — invoke deliberately, with a complete dossier.
Agent Claude Code
Read-only codebase reconnaissance. Use PROACTIVELY before any planning or implementation to map relevant files, symbols, data flow, conventions, and risks. Never writes or modifies anything.