Expected size and the strength of the native-path bias
Read this file the first time a request's expected size is in doubt. SKILL.md states the bias itself (prefer the
orchestrator's family when models are roughly equally suitable, and prefer a native path); this file is the size anchor
for how strongly it applies.
Apply the bias according to expected task size. These are judgment anchors, not hard thresholds:
| expected size | practical meaning | native-path bias |
|---|---|---|
| tiny | Setup and review may cost as much as doing the task directly. | Very strong. Usually do it directly or delegate natively. |
| short | One bounded operation with little context or expected iteration. | Strong. Cross harnesses only for a meaningful capability or total-cost advantage. |
| medium | A substantive task where model work dominates fixed startup overhead. | Moderate. Consider likely retries and review burden alongside startup overhead. |
| large | Extended work where cross-harness startup is a minor part of the run. | Weak. Choose the route most likely to finish cleanly at lower expected total cost. |