{
	"route": {
		"instructions": "The state is a request to a coding agent. Decide where the strong model's reasoning has to be spent: before the first edit, or after it. Prewalk lets a strong model investigate and plan, then hands the work to a faster, cheaper model; it pays off only when the hard reasoning is front-loaded and the work left after planning is bounded.\n\n`request` is the task to classify and is authoritative. `previous_request` and `previous_response`, when present, are the user's earlier request and the agent's last reply. Use them only to resolve what a short or referential request points at (\"do that\", \"continue\", \"fix what you found\"). A request that stands on its own is classified on its own words, whatever came before it.\n\nJudge where the reasoning is needed, not how large or hard the task is. Volume of work never makes a task a prewalk candidate, and neither does difficulty alone. Needing a plan is not enough: the hard part must be settled by the time the first edit is made. A wrong prewalk costs more than a missed one. If torn between prewalk and primary, choose primary. If torn on whether a small task is worth any hand-off, choose direct.",
		"criteria": {
			"prewalk": "Both hold: starting correctly needs real up-front investigation, repository understanding, diagnosis, design, or planning; and once that fixes the approach, the implementation and verification left are bounded enough for a faster model. Tracing an unfamiliar subsystem before a localized change; reading several interacting files before implementing a clear design; diagnosing a bug down to a concrete fix whose implementation is straightforward; a migration or refactor whose hard part is finding the right transformation and everything it touches.",
			"primary": "Strong reasoning stays needed while implementing, debugging, or validating, so handing off at the first edit would be premature. Unresolved races, concurrency, or invariant problems; subtle state-machine or lifecycle changes; a bug whose cause is likely still uncertain after a first look; high-risk changes that need iterative reasoning while editing and testing; difficult algorithms; work where the first edit is only the start of the hard thinking.",
			"direct": "No up-front phase worth handing off from: explanations and read-only questions; trivial or mechanical edits; fully specified changes with an obvious implementation; tiny localized fixes; requests that involve no implementation."
		}
	}
}