trajectory-audit

You can audit token usage and cost distribution across multiple Claude Code sessions by parsing JSONL transcript logs. This process identifies token-weighted cost inefficiencies, tool-use error rates, and file mutation churn to optimize agent performance and billing.

Your session may be experiencing stuck loops, file mutation churn, or shell timeout kills. Behavioral auditing detects these patterns in JSONL logs to pinpoint exactly why a specific Claude Code session is looping and burning through tokens.

Verify your prompt-caching strategy by evaluating cache-read shares and cache-CREATE bursts within your session transcripts. The audit calculates whether your caching implementation is genuinely reducing token billing or causing inefficiencies.

Yes, Python is required to execute the session_audit.py script. This script parses the JSONL transcript logs and generates the markdown summary reports needed for token accounting and behavioral pattern analysis.

The best way is to audit cross-session token usage by running the session parser. It provides exact token accounting and cost breakdowns by billing bucket and model provider across multiple Claude Code runs to identify your most expensive projects.

The required file format is JSONL. The skill parses JSONL transcript logs to extract behavioral patterns, enabling the evaluation of tool-use error rates, file mutation churn, and shell timeout kills across your Claude Code sessions.