agent-tracing

Debug agent execution by reading stored execution snapshots to inspect step-level LLM calls, tool invocations, and context engine inputs. The skill renders detailed JSON artifacts locally, providing complete visibility into what was sent to the LLM after processing.

An execution snapshot is a deterministic local JSON artifact recording multi-step LLM workflows, runtime events, and context engine payloads. It captures tool calls, messages, and low-level events like llm_start, enabling post-execution analysis for troubleshooting.

Run the command `agent-tracing inspect latest` from your repo-root working directory to immediately view an overview of the newest agent execution trace. The CLI reads stored JSON files to render step, message, tool, and context engine details.

Yes, you can list, inspect, and clean in-progress partial snapshots to troubleshoot interrupted agent runs. The CLI provides commands to manage these incomplete traces, ensuring interrupted context engine payloads and tool calls are accessible for diagnosis.

No, agent tracing has zero external dependencies. It records and stores deterministic snapshot artifacts locally as JSON files, then provides a standalone CLI that renders execution details directly from the repo-root working directory.

Analyze context engine behavior by viewing its input parameters and the final LLM payload produced after processing. The skill compares available tools against what the agent actually invoked, including arguments and outputs, for deeper diagnosis.