Regression: increased per-turn token consumption vs previous versions

Description

Per-turn token usage has increased compared to previous Claude Code versions, causing the context window to exhaust faster. No changes to plugins, skills, hooks, or CLAUDE.md configuration.

Context usage at session start (fresh session, zero messages)

System prompt:   4.3k tokens (2.2%)
System tools:   20.3k tokens (10.2%)
MCP tools:        905 tokens (0.5%)
Custom agents:   1.6k tokens (0.8%)
Memory files:    1.6k tokens (0.8%)
Skills:          2.3k tokens (1.2%)
Messages:          260 tokens (0.1%)
─────────────────────────────────
Total:            31k / 200k (16%)

31k tokens consumed before any work begins. Each subsequent turn adds ~2k tokens of overhead beyond the actual conversation content, which is higher than observed in previous versions with the same configuration.

Environment

• Platform: Linux (Fedora 43)
• Model: claude-opus-4-6
• 21 plugins installed (unchanged across versions)
• CLAUDE.md: ~1.3k tokens (global + project combined under "Memory files")

Expected behavior

Per-turn token overhead should remain consistent across Claude Code versions when the user's configuration has not changed. Previous versions exhibited lower per-turn consumption with the identical plugin and configuration setup.

🤖 Generated with Claude Code