Runs FIRST for every message.
Complexity
complexity[4]{level,criteria,approach}:
Quick,"Single file / simple fix / clear scope","Direct implementation"
Standard,"2-5 files / feature / some unknowns","Scout then implement"
Deep,"6+ files / architecture / vague scope","run-orchestrator"
Project,"Multi-feature / multi-session / weight ≥ 3 on bridge heuristic AND no active plan","/aura-frog:plan bootstrap then per-task /run anchored"
Model Selection
Inherit the session model for all substantive work. Only override: down-shift to haiku for trivial mechanical work (classification, detection, bookkeeping). Never force-upgrade to a named model. See rules/core/small-to-large-routing.md.
Detection (priority order)
Task content > explicit tech (+60) > intent keywords (+50) > project context (+40) > file patterns (+20). Scoring: Primary ≥80 (leads), Secondary 50-79, Optional 30-49, Skip <30. tester: always secondary unless explicit test request.
Cache: .claude/cache/agent-detection-cache.json (reuse within workflow; invalidate on new workflow / phase 1 / user override).
Detail (Project-level heuristic, signal weights, model rationale): skills/agent-detector/reference.md — load on demand only.