What problem does it solve? Debugging agent behavior requires visibility into what the LLM actually saw and did at each step, but production and local agent runs leave no inspectable record of messages, tool calls, or context assembly. This Skill records every agent execution step as a snapshot and provides a CLI to inspect, diff, and replay those traces. ## Core Features & Use Cases - Trace Inspection: Browse execution snapshots with commands for messages, tool calls, events, context engine payloads, available tools, and user memory at any step. - Remote Trace Pulling: Download production or staging snapshots by operation id via lh trace op or a standalone CLI with a configured object-storage base URL. - Context Window Mapping: Render per-call context composition with prefix-cache reuse analysis using ctx-map, including standalone HTML reports. - Model Replay: Re-issue a frozen recorded LLM payload against other models with an llm-rubric judge to test whether a different model would have completed the same run. - Use Case: An agent called the wrong tool in production. Pull the trace by operation id, inspect the step's available tools versus actual tool calls, then replay that call against another model to determine whether the fault lies in context assembly or model selection. ## Quick Start Ask the agent to inspect the latest agent trace and show the messages and tool calls for a specific step.
