episode-complete

Pattern extraction captures tool sequences, decision points, and error patterns from task episodes. After completing a task, the episode-complete Skill scores outcomes (verdict, time, tokens, artifacts) and generates reflections that identify what worked and what to improve, then stores these patterns for future retrieval and heuristic updates.

Outcome scoring records verdict, execution time in milliseconds, token usage, generated artifacts, and errors encountered. This outcome data feeds into reward scoring and reflection generation, which extract actionable patterns to refine future task performance.

When you call complete_episode with a constructed TaskOutcome, the Skill derives patterns from tool sequences and decision points, computes a RewardScore, generates reflections on successes and gaps, then updates persistent storage and indices. This feeds learned heuristics back into the system.

Episode-complete applies during post-task closure across learning episodes regardless of task type. It requires a constructed TaskOutcome with outcome data; the Skill then handles pattern extraction, reflection, and heuristic updates automatically without dependencies on specific frameworks or components.

Reflection documents what worked and what to improve—high-level learnings from the episode. Pattern extraction identifies reusable structures: tool sequences (how tools were chained), decision points (where choices mattered), and error patterns (failure modes). Both update persistent storage for future tasks.