System prompt token bloat: ~40-50% increase between v2.1.92 and v2.1.100

Summary

System prompt token consumption increased approximately 40-50% (~4,000 tokens) between v2.1.92 and v2.1.100 with zero changes to user configuration. This significantly reduces the usable context window for actual work.

Reproduction

1. Install v2.1.92: npm install -g @anthropic-ai/claude-code@2.1.92
2. Run claude -p "Estimate the token count of your system prompt" in a project with CLAUDE.md, memory, skills, and hooks configured
3. Upgrade to v2.1.100: npm install -g @anthropic-ai/claude-code@2.1.100
4. Run the same prompt again
5. Compare estimates

Observed results

| Version | Estimated system prompt tokens |
|---------|-------------------------------|
| v2.1.92 | ~8,000–10,000 |
| v2.1.100 | ~12,000–14,000 |

Identified causes

The following system prompt sections appear to have grown significantly:

• Agent tool definition: Extensive usage examples, agent type catalog, isolation mode docs — substantially larger than v2.1.92
• Bash tool definition: Now embeds full Git Safety Protocol, PR creation workflow, and commit workflow instructions inline
• Deferred tools mechanism: New ToolSearch tool + listing of 20+ deferred tool names adds overhead that didn't exist in v2.1.92
• system-reminder duplication: outputStyle reminder ("Explanatory output style is active...") is injected after every tool result, not once per turn. Task reminder also repeats frequently
• # Output efficiency section removed: Previously existed in v2.1.92, absent in v2.1.100 — ironic given the bloat

Additional context

• Output efficiency removal is concerning — the section that told Claude to be token-efficient was removed in the same version range that added thousands of tokens to the system prompt
• The system-reminder repetition is especially wasteful — the same 1-2 line reminder injected 5-10 times per conversation turn adds up fast
• Environment: WSL2 (Linux), project uses custom skills (17), hooks (9), memory files (13), agent definitions (4). None of these changed between versions.

Expected behavior

System prompt overhead should remain stable across minor versions, or at minimum not grow 40-50% without clear user-facing benefit. Repeated system-reminder injection should be deduplicated.

Environment

• OS: WSL2 (Linux 6.6.87.2-microsoft-standard-WSL2)
• Claude Code: tested on v2.1.92 and v2.1.100
• Currently downgraded to v2.1.92 as workaround