// Classify oracle inputs/outputs into risk levels and return allow, rewrite, or refuse policy with concrete constraints. Use before and after specialist-agent generation, especially for finance, medical, legal, violence, self-harm, or fear-marketing risks.
Transform oracle guidance into concrete, calendar-ready tasks with suggested time windows, effort level, and practical notes. Use after ZiWei/MeiHua/Tarot outputs when users need execution plans.
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.
Audit consistency of repeated long-line oracle outputs and diagnose drift causes across routing, retrieval, and generation settings. Use when validating long-term ZiWei responses, regression testing, or before release quality gates.
Compose daily oracle cards from profile, date context, and recent conversation themes using a fixed stable format. Use for daily home feed generation, ritual cards, and lightweight recurring guidance.
Produce short-term MeiHua YiShu analysis for concrete near-term events, including reframed question, tendency, key variables, do/donot guidance, and contingency actions. Use for today/this-week event decisions and tactical consultation.
Convert retrieved philosophy or classics snippets into concise modern mindset guidance and practical reflection methods. Use when users need emotional reframing, growth-oriented interpretation, or RAG-backed meaning extraction.
| name | oracle-safety-guardian |
| description | Classify oracle inputs/outputs into risk levels and return allow, rewrite, or refuse policy with concrete constraints. Use before and after specialist-agent generation, especially for finance, medical, legal, violence, self-harm, or fear-marketing risks. |
Perform two-stage safety governance for oracle content: pre-check user input and post-check generated output.
mode: pre or postcontent: user query or generated answercontext: optional (profile summary, intent, tool trace)references/risk-grading.md.allowrewriterefuserewrite, provide strict rewrite constraints.refuse, provide safe alternative guidance.Return structured policy:
risk_level: S0/S1/S2/S3/S4decision: allow/rewrite/refusereasons: short listconstraints: list of mandatory constraintsdisclaimer_level: none/light/strongreferences/risk-grading.md before final decision.