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

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

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

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

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

Más de este repositorio

mismo repositorio

引导一个工程目录优化到目标 Agent-first 结构（审计差异、输出迁移建议、可选创建缺失目录与占位文件）。当用户请求“优化目录结构/对齐目标结构/仓库结构治理”时使用。

资产注册 — 管理所有托管资产（Docker/工作目录/进程/配置/数据）的注册、索引和健康巡检

容量监控 — 持续监控磁盘、内存、进程数等系统容量指标，提前预警

清理策略 — 基于规则的自动清理执行引擎，支持 dry-run 预览和安全确认

代码生成 — 根据 Issue 描述和代码上下文，自主编写高质量代码

上下文评估 — 评估用户的技术背景和需求，定制个性化的学习路径

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"}

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

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

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

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