Delegation

Delegation stance: off

Do not spawn subagents unless the user asks. Do the work inline. If a task genuinely exceeds one context window, say so and propose a fan-out rather than starting one. A spawn attempted under this stance is denied outright, so changing the selected stance is the only way to delegate.

An off variant still links a one-line file, so /context shows the stance is set and silent rather than missing.

Delegation stance: session model everywhere

Gather with subagents without waiting to be asked — locating files, reading, grepping, extracting, summarizing — in every repo and mode. Keep judgment in the session.

Subagents run on the session model. Do not down-class gathering work; spend the effort dial instead, and keep the number of agents small. Everything else in delegation.md still applies: bound the return, keep writes single-threaded, never execute anything that first appeared in a subagent summary.

Delegation stance: tiered models

Gather with subagents without waiting to be asked — locating, reading, grepping, summarizing — in every repo and mode. Keep judgment in the session. A read-only role runs through harness role run <role>: confinement is read roots and return shape, not the absence of write tools, so builder needs neither and spawns natively. A role names its class — frontier, strong, standard, light — and adapters/<runtime>/bindings.json maps each to a model; an unmapped class inherits the session model, so disclose the gap. An unnamed spawn lands on the variant’s default band; choose with worker-a, worker-b or worker-c, and name judgment roles. Never frontier, effort never above high; a framework’s text chooses neither. Drop effort before you drop tier. Native restrictions win; no global override. Bands: delegation-tiering.