episode-log-steps

Logging execution steps captures detailed records of tool invocations, actions, latency, token usage, and outcomes. Create an ExecutionStep with tool, action, latency_ms, tokens, success, and observation fields, then attach it to the episode using log_step to build a searchable execution history for learning and debugging.

Execution step logs should record the tool used, the action performed, latency in milliseconds, token count, success status, and observations about the result. This structured schema enables pattern extraction, performance analysis, and error tracking across builds, tests, code generation, and other tasks.

Yes, batching guidance in the Skill allows you to log related steps together to organize your episode logs meaningfully. Group steps by logical phase—such as decision points, tool chains, or error sequences—to make your logs easier to search and analyze.

Episode logs stored in JSON-encoded memory provide a complete, queryable record of execution steps. Use latency, token, and success metrics alongside observations to identify bottlenecks, error patterns, and inefficiencies across your recorded episodes.

Execution step logs conform to a structured ExecutionStep schema stored in batch-friendly JSON format within episode memory. This standardized structure supports parsing, filtering, and analysis of tool usage, decision points, errors, and milestones across your tasks.