AgentSkillsCN

oracle-agent-team-orchestrator

通过路由、检索与生成设置，协调DeepSeek Oracle的多智能体工作流，将用户意图导向专业Oracle技能，强化安全防线，并以后续问题与行动事项整合成统一的答案。当请求涉及长期ZiWei、短期MeiHua、塔罗牌、每日牌阵、哲学RAG，或混合多领域查询时，可使用此功能。

SKILL.md
--- frontmatter
name: oracle-agent-team-orchestrator
description: Orchestrate DeepSeek Oracle multi-agent workflows by routing user intents to specialist oracle skills, enforcing safety gates, and composing one unified answer with follow-up questions and action items. Use when requests involve long-term ZiWei, short-term MeiHua, tarot, daily card, philosophy RAG, or mixed multi-domain queries.

Oracle Agent Team Orchestrator

Overview

Route each user request to the right specialist skill, run mandatory safety checks, and return one unified response that is calming, actionable, and compliant.

Input Contract

Require these inputs (fill missing values with null):

  • •user_query
  • •conversation_history_summary
  • •user_profile_summary
  • •selected_school (east/west/mixed)
  • •enabled_schools (e.g. ziwei,meihua for MVP)
  • •safety_policy

Workflow

  1. •Run pre-check with $oracle-safety-guardian.
  2. •Classify intent:
  • •Long-term life trend -> $oracle-ziwei-agent
  • •Short-term event -> $oracle-meihua-agent
  • •Symbolic reflection -> $oracle-tarot-agent only if west mode is enabled
  • •Daily briefing -> $oracle-daily-card-composer
  • •Practical execution -> $oracle-actionizer
  • •Mindset/philosophy uplift -> $oracle-philosophy-rag-agent
  1. •Invoke 1-3 specialist skills only; avoid unnecessary fan-out.
  2. •Merge outputs using references/response-contract.md.
  3. •Run post-check with $oracle-safety-guardian.
  4. •If post-check is rewrite or refuse, follow that policy before final output.

Output Contract

Return this JSON-shaped structure in natural language sections:

  • •answer_text
  • •follow_up_questions (exactly 3)
  • •action_items (0-5)
  • •safety_disclaimer_level (none/light/strong)
  • •trace (internal; called skills + reason)

Quality Bar

  • •Always include: 1 calming sentence + 2-4 actionable suggestions + 1 risk reminder.
  • •Avoid fatalism and fear-based language.
  • •Keep long-line conclusions stable for the same user/profile input.
  • •Use Chinese by default unless user asks for another language.

References

  • •Read references/routing-matrix.md before routing.
  • •Read references/response-contract.md before composing final output.