qiti-yuanliu

What problem does it solve?

炁体源流 helps an LLM agent detect context drift, semantic contradictions, and degraded reasoning quality in multi-round conversations, then produce stable repair handoffs to protect alignment.

Core Features & Use Cases

• Semantic-level entropy & health scoring: computes entropy (conflict/gap/redundancy/topic drift/intent shift) and derives a health score to classify context quality.
• Contradiction and drift detection: flags semantic reversals, high negation density, info gaps, and topic/intention deviations, distinguishing user clarifications from high-risk resets.
• Repair handoff & “稳态契约” outputs: generates structured repair plans, stability contracts, checkpoints, risk hotspots, and recommendations that guide downstream skills (e.g., cross-segment trace) without directly rewriting the user’s goal.

Quick Start

Ask the agent to run qiti-yuanliu on the current conversation context stored in context.json and return context.health_report.json with metrics, alerts, and repair handoff.