Route each Claude Code task to the cheapest Claude model tier that can do it well — via model-pinned subagents.
Agent
Highest-intelligence read-only advisor (Fable 5). Consult at commitment boundaries — before an architecture / migration / API-contract / large-refactor decision, or when a problem has resisted two attempts. Returns a verdict, not a survey. Never writes code.
Agent
Fast read-only search/exploration lane (Haiku 4.5). Use for broad fan-out — locate code, map naming conventions, gather files across a large tree, summarize or classify at volume. Returns conclusions, not file dumps. Never writes.
Agent
Hard-build lane (Opus 5) — the write lane for escalated, uncertain, or reasoning-heavy implementation and root-cause debugging. Use when a change is hard as code (gnarly logic, a non-trivial algorithm, a stubborn bug) or when the implementer (Sonnet) lane has failed the same task twice. Writes code and proves it.…
Agent
Standard implementation worker (Sonnet 5) — the build lane. Use for normal feature/bugfix implementation and moderate-reasoning multi-file changes where the approach is already decided. Writes code and proves it with tests.
Agent
Strong read-only reviewer / synthesizer (Opus 5). Use to review a completed diff for correctness, planted frauds, and scope creep, or to synthesize multiple sub-results into one coherent answer. Cites file:line. Returns a VERIFIED / VERIFIED WITH CAVEATS / REFUTED verdict. Never writes code.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: