What problem does it solve? When an AI coding session feels slow, teams guess at the cause and often optimize the wrong thing. This Skill answers "why was that session slow?" by measuring the transcript's own timestamps instead of guessing, attributing every gap to the activity that caused it and naming the exact commands behind the biggest delays. ## Core Features & Use Cases - Wall-clock attribution: Reads Claude Code session JSONL and Codex rollout JSONL, charging each time gap to categories like tests/build, subagent dispatch, adversarial review, model thinking, stalls, and user-away time. - Ranked gap analysis: Reports the top 20 individual gaps with timestamps and the specific command or prompt that preceded each, so findings are actionable rather than aggregate. - Token accounting: Emits a reproducible tokens block that correctly handles the opposite cache semantics of Claude and Codex usage records, plus sub-agent spend reported separately. - Use Case: After a pipeline felt sluggish, run --compare on two transcripts from before and after a fix to verify the change actually moved the category it was supposed to move. ## Quick Start Ask the agent to profile the most recent session transcript and explain where the wall-clock time went, using the profile-session skill with --last.
