intent-routing

name	intent-routing
description	意图识别与路由 — 解析用户自然语言输入，识别意图类别，路由到合适的 Agent 或内置命令
version	1.0.0
license	MIT
metadata	{"author":"blackplume233","actant-tags":"intent,routing,nlp,steward","layer":"kernel","agent":"steward"}

Intent Routing

Overview

Parse user natural language input, classify intent categories, and route to the appropriate Agent or built-in command.

Intent Categories

Category	Description	Routing Target
task-delegation	User wants work done by an Agent	Steward → domain Agent (Researcher, Maintainer, etc.)
query	User asks a question or seeks information	Steward → Researcher or direct answer
system-command	User invokes system control (start/stop, config, status)	Built-in command handler
conversation	Greeting, clarification, follow-up	Steward handles directly, maintain context

Routing Rules

1. Keyword matching: Slash commands (/qa, /issue), explicit verbs ("create", "run", "check")
2. Semantic inference: Use context and prior turns to disambiguate ambiguous input
3. Fallback: When intent unclear, ask clarifying question before routing

Fallback Behavior

• If confidence < threshold: respond with "I'm not sure if you want to... [options]. Which applies?"
• If multiple valid routes: prefer the one with highest prior success rate for similar queries
• If no valid route: explain capabilities and suggest alternatives