chain-of-commandlisted

# Chain of command: Fable plans, Opus codes The main loop is the commanding officer: it sets strategy, gives the orders, and signs off on everything that comes back. Opus agents carry out the orders. Nothing that judges or decides gets delegated down. From this point forward in THIS session, follow this dispatch policy on every agent/workflow dispatch. It exists to concentrate the expensive top-tier model (Fable) where judgment lives and push volume work (recursion, code emission, bulk review passes) to Opus. ## The two lanes **Main loop (session model, Fable when available) - never delegated:** - Understanding the problem, reading the load-bearing code needed to plan - Strategy, architecture, and the written plan - Decomposing work into dispatches and writing the dispatch prompts - Judging and verifying everything that comes back before trusting it - Final review synthesis, gate decisions (ship / fix / re-run), and anything ambiguous or novel - Any decision a subagent escalates **Opus (`model: "opus"`, passed EXPLICITLY on every dispatch) - all volume work:** - Writing or editing code from an established plan or spec - Writing tests from a spec, mechanical refactors, migrations, fixture generation - Exploration that requires comprehension: tracing call paths to understand behavior, semantic hunts ("find where we handle X"), any search whose STRATEGY needs judgment - Recursive/fan-out review passes (dcr-style reviewers, fixers, verify swarms, adversarial refuters) - EXCE