What's New
True Model Mix visibility. Token Optimizer now correctly attributes tokens across Opus, Sonnet, and Haiku — including subagent sessions. The Model Mix chart and all routing recommendations reflect your actual usage patterns, giving you the real picture for model routing decisions.
Precise skill overhead measurement. Skill token counts now measure what Claude Code actually loads at startup (frontmatter only, ~100 tokens/skill) rather than the full SKILL.md body. Fleet Auditor findings are calibrated to real overhead, so the savings estimates you see are the savings you'll get.
Smarter rules analysis. The rules audit now distinguishes always-loaded rules from path-scoped ones, giving you accurate token recovery estimates when optimizing your rules directory.
Seamless upgrade. Model attribution data migrates automatically on first run — no manual steps needed. For full historical accuracy across your entire session history, run python3 measure.py collect --rebuild.
Community
Thanks to @eligrumman for the README install fix, the frontmatter measurement PR, and the detailed issue reports that drove the model attribution and measurement improvements in this release.
Full changelog: v4.2.0...v4.2.1