Temporal Reasoning Sleuth is a skill for AI agents that addresses a specific class of failure: incorrect or incomplete answers to temporal and causal queries over long event histories. Rather than a tool with callable endpoints, it provides architecture patterns and retrieval strategies that developers can apply when building agents over organizational knowledge bases, incident histories, architecture decision records, or similar chronological data stores.
The skill addresses three distinct query types. Sequence queries ask what happened between two points in time. Causal queries ask what caused a given outcome or what sequence of decisions led to a particular state. Counterfactual queries ask what would have happened if a specific decision had differed. Each of these query types breaks down differently when agents receive the wrong context.
Three failure modes are explicitly covered: “lost in the middle” failures where agents overlook events buried deep in a chronological dump; context poisoning where events retrieved without surrounding causal context produce wrong conclusions; and unanswerable history questions that arise when causal relationships are not explicitly modeled.
The concrete patterns provided include event graphs with timestamped causal edges, pre-computed causal chain indexes, and windowed context synthesis that compresses distant history to preserve critical signal within a context window. This skill is most relevant when understanding why something happened is as important as knowing what happened.