What problem does it solve?
When a session is flagged as a cost outlier, operators need to know which specific messages drove the spend. This Skill breaks down per-message costs inside one session, exposing whether the expense came from output tokens, ephemeral cache writes, or model escalations.
Core Features & Use Cases
- Per-message cost ranking: Parses a session's jsonl transcript, prices every assistant message with usage data, and lists the top-N most expensive messages with input, output, cache write, and cache read token counts.
- Percentile context: Computes p50, p90, and p99 message costs and flags the top message when it exceeds 2x the p99, identifying in-session outliers.
- Flexible targeting: Select a session by ID or use the most recent one, and filter messages with a --since timestamp to drill into a specific time range.
- Use Case: After cost-anomaly flags a session as a >3.5 sigma outlier, run this Skill on that session ID to discover that a single message wrote 881K tokens to ephemeral cache, explaining a $16.58 charge.
Quick Start
Ask the AI to show the top 10 most expensive messages in the latest Claude session and explain where the cost came from.