agent-tracing

What problem does it solve? Debugging agent runs requires visibility into what the LLM actually saw and did, but that context is scattered across logs, databases, and remote storage. This Skill records every agent execution step to disk as a snapshot and provides a CLI to inspect failed tool calls, LLM payloads, available tools, memory, and context engine input by operation ID. ## Core Features & Use Cases - Trace Inspection: List and inspect local, in-progress partial, and remote production snapshots, drilling into steps, messages, tool calls, events, and raw JSON. - Context Window Mapping: The ctx-map command renders each LLM call's message composition against the model's context window, including prefix-cache reuse and break points. - Model Replay: Re-issue a frozen recorded LLM call against other models with an llm-rubric judge to determine whether failures come from the model or the context. - Use Case: An agent operation failed in production. Run lh trace op inspect op_xxx to see the exact tool arguments and results, then use ctx-map to check whether an injected timestamp block broke prefix caching. ## Quick Start Ask the agent to inspect the latest agent trace and show the tool calls and LLM messages for the failing step.