What problem does it solve? Agent coding sessions accumulate hidden workflow friction—repeated failed commands, rate-limit pressure, auto-review loops, and context drift—that is hard to spot by reading raw rollout or session transcripts by hand. This Skill scans local trace files deterministically, groups friction hits into costed episodes and clusters, and produces redacted reports so recurring problems can be routed to the right durable fix. ## Core Features & Use Cases - Deterministic trace scanning: analyze_rollouts.py scans bounded sets of rollout/session files for concrete friction signals such as repeated command failures, GitHub rate limits, auth loops, and stale validation results, with redacted output and explicit scan budgets. - Episode segmentation and clustering: segment_rollout_episodes.py converts line-level hits into costed episodes with outcome metadata, and cluster_rollout_episodes.py collapses them into root-cause clusters with compact redacted trajectory skeletons for review. - Auto-review ledger classification: classify_auto_review_ledger.py distinguishes current-target findings from stale detached auto-review worktree proposals so outdated findings do not block active work. - Optional local LLM review: bounded LM Studio scout and memory-extraction pipelines review redacted findings locally, with strict JSON validation and trusted-endpoint rules. - Use Case: After noticing an agent repeatedly retrying GitHub commands, point the analyzer at the recent Codex sessions directory, cluster the resulting episodes, and receive a proposal classifying the top cluster as rate-limit pressure with a recommended helper fix. ## Quick Start Ask the agent to audit recent rollout session files for workflow friction using the rollout-friction skill, providing the path to your sessions directory.
