agent-routing

# Agent Routing — model + effort selection for subagents Before spawning any subagent, answer one question: **is the task mechanically checkable, or does it require judgment?** Checkable → Haiku at `effort: 'low'` behind a verifier. Judgment → up-tier. The routing table refines this split; the rest of the skill governs cascades and loops. **Do not up-tier checkable work "to be safe."** Measured Haiku 4.5 failure on every checkable task family tested is ≈0, at `effort: 'low'`, with chain-of-thought suppressed (evidence: references/calibration-2026-07-15.md). The burden of proof is on routing *up*: reflexive up-tiering costs 3–5× for safety that the data says is imaginary. Spend effort, not tier, and only where reasoning depth demonstrably falls short. To turn a judgment-shaped task INTO a Haiku-executable one (explicit procedures, n-shot examples), use the sibling `down-skilling` skill. This skill decides the routing; that one engineers the prompt. ## Context handoff — routing picks the tier; the prompt carries the context Subagents inherit nothing: not the conversation, not loaded skills, not the existence of artifacts already on disk. Every index, scan output, artifact path, or tool-invocation recipe the orchestrator relies on must be serialized into the subagent's prompt (or written to a file the prompt points at). Otherwise the agent falls back to blind rediscovery — and the tier premium is spent on crawling, not judgment. A Sonnet with no handoff wastes more than a H...