What problem does it solve? Answering where time actually went during Claude Code sessions — inside subagents versus the main agent, by model, by tool, by bash command — normally requires re-reading and re-aggregating raw telemetry on every question, which is slow, token-expensive, and non-reproducible. This skill provides a predefined, deterministic analysis engine over the recorded ptp telemetry store. ## Core Features & Use Cases - Predefined analysis engine: Runs scripts/ptp-telemetry-analyze.js (Node standard library only, no install step) over the raw NDJSON span store, with a CSV fallback when no raw files exist. - Subagent vs main-agent split: Classifies leaf spans by parent links or per-session timestamp containment against wrapper windows, naming the method used in a mandatory data-quality footer. - Honest degradation: Reports unavailable rather than 0 for figures the CSV fallback cannot measure, and states exactly when retention has narrowed the analyzed scope. - Use Case: After several weeks of ptp pipeline runs, ask which models and bash commands consumed the most time and how much work happened inside subagents, and receive a reproducible report with per-key counts and a data-quality footer. ## Quick Start Ask the assistant to run the ptp telemetry analyze report for this repository to show where time went by model, tool, and subagent.
