Token Efficient Task Router

Problem

General-purpose agents often treat every request as a full-scale job: reading all context, drafting full plans, touching many files, or producing long reports. For asks like “diagnose only”, “3 bullets first”, or “do it directly”, this wastes token spend and increases the chance of unwanted side effects. This skill adds a pre-execution routing layer that classifies each request by clarity, complexity, risk, and expected token cost before choosing an Ask, Plan, Craft, or Expert mode.

How It Works

• Classify before acting: simple, clear, low-risk tasks proceed in the smallest useful mode without extra confirmation.
• Load references on demand: start with focus_activation_rules.md, then pick one primary focus and at most one assist focus instead of loading all mode files.
• Budget first: for uncertain work, lock cognition, context, and delivery budgets, such as Direct, Light, or Saver, before broad execution.
• Safety gates: file changes, batch refactors, or destructive operations default to Plan or explicit confirmation; Rush and Unmetered policies do not remove safety checks.
• Cost observability: when spend tracking is needed, report real platform usage as Observed or give honest estimates, without presenting estimates as exact metered values.

Boundaries

Use it for ambiguous requests, knowledge-base or project-folder tasks, multi-step work, and cross-platform agent workflows. It is not a license to over-process simple tasks, bypass platform permissions, read private data, or perform destructive changes without confirmation. If the platform exposes real usage signals, prefer them; otherwise estimate using budgets, file count, material count, turn count, and output size.