What's New
Scoring that rewards good setups instead of just punishing. The coach health score now starts at 70 (not 100) and earns points for active skills, lean CLAUDE.md, configured hooks, and intentional effort settings. An optimized setup scores higher than a default one.
Proportional thresholds. Detectors now require patterns to affect >5% of your messages before flagging. 7 loops out of 972 messages is noise, not a signal. Same for monolithic prompts, retry chains, and other session-level patterns.
Accurate cost breakdown. The Most Expensive Prompts section now shows the actual token split (cache write vs cache read vs fresh input) so you understand why a prompt costs what it costs.
No more false positives:
- Web search overhead won't flag if you average <1 search per session
- Verbose skill descriptions threshold raised to 200 chars (was 120)
- MCP servers aren't flagged unless you have 20+ AND they exceed 3% of context
- Skills aren't penalized by count, only by actual unused percentage
- Opus usage threshold raised to 85% (was 70%), no double-penalty with per-session detector
- CLAUDE.md guidance is constructive ('Could Be Leaner' with restructuring help) not blaming ('Novel')
Detector improvements: All 9 detectors now emit occurrence counts for proportional filtering. Thresholds raised across the board to reduce noise at power-user scale.
Transcript preservation. Auto-sets cleanupPeriodDays on first run so Claude Code doesn't silently delete your session history after 30 days.
Full changelog: v4.3.0...v4.3.1