What problem does it solve? Debugging multi-turn chat conversations is hard when failures originate in earlier turns and propagate. This Skill reconstructs an entire MLflow session from its traces, pinpoints the turn where things went wrong, and surfaces assessment signals without reading every trace in full. ## Core Features & Use Cases - Session Reconstruction: Discovers the input/output schema from the first trace, then extracts inputs, outputs, and assessments across all session traces using the MLflow CLI. - Assessment Interpretation: Distinguishes session-level from per-turn assessments, filters out scorer errors, and reads rationale fields to interpret values correctly. - Root-Cause Analysis: Correlates trace behavior with codebase patterns like context window management and cross-turn state. - Use Case: A user reports a wrong answer on turn 5 of a chatbot conversation. The Skill reconstructs the session, finds the factual error actually originated in turn 3, and traces it to a retriever returning an outdated document. ## Quick Start Ask the assistant to analyze the MLflow session with a given session ID and experiment ID to find where the conversation went wrong.
