agent-tracing

Debug unexpected LLM agent behavior by inspecting recorded execution snapshots and step-level traces. You can view completed traces to drill into individual steps and understand exactly what happened during the agent runtime.

Compare context-engine input against the final LLM payload by using agent tracing to deep-dive into context payloads. This reveals how DB messages, system role, model, tools, and memory are processed before reaching the LLM.

Inspect agent tool calls, messages, and raw runtime events by using CLI inspection flags on completed JSON traces. These flags allow you to view tool call results, event forensics, user memory, and payload diffs across steps.

View the context payload sent to the LLM by running the CLI inspection command on the latest trace. You can navigate to a specific step and apply flags to inspect the exact messages and context delivered to the model.

Yes, you need to generate local snapshots in development mode to use agent tracing. The diagnostic process requires reading completed JSON traces from the .agent-tracing directory to inspect recorded execution steps.

If your LLM agent behaves unexpectedly, inspect the payload diff between context-engine input and the final LLM payload. Analyzing execution snapshots and step-level traces helps identify where the context transformation diverged.