Unexpected 2x token burn rate on MAX plan - possibly related to recent prompt caching changes

Summary

Heavy MAX plan user experiencing 59% weekly budget burn in just 2 days. This is a recent regression - I've been using Claude Code heavily for months and this burn rate only started occurring recently (possibly after prompt caching release).

Environment

• Plan: MAX (20x standard, ~$200/month)
• Model: claude-opus-4-5
• Usage: 59% of weekly budget in 2 days
• Expected: ~28% for 2 days of work
• Setting: ENABLE_TOOL_SEARCH=true

What Changed Recently

This burn rate is new. I've been a heavy Claude Code user running multi-agent GSD workflows for weeks. The same workload pattern previously fit within weekly budgets. Something changed recently.

Possible correlation: This started after prompt caching was released. Unknown if related, but timing is suspicious.

Identified Issues (Partial Picture)

1. Skill Listing Overhead (~11,500 tokens/message)

All ~230 skills are dumped into every system prompt:

The following skills are available for use with the Skill tool:
- keybindings-help: Use when the user wants to customize...
- simplify: Review changed code for reuse...
... (227 more entries)

This loads even with ENABLE_TOOL_SEARCH=true set. Expected behavior: skills should lazy-load via ToolSearch only.

2. Subagent Context Multiplication

Each spawned subagent appears to inherit the full system prompt (~120K tokens), including:

• All skills listing
• All CLAUDE.md content
• All rules
• Full memory

Spawning 3 parallel agents = 3× overhead before any actual work.

3. No Token Visibility

/usage doesn't show WHERE tokens went:

• No per-turn breakdown
• No subagent costs
• No skill/context overhead visibility

Without this, users can't diagnose issues.

What I Don't Know

This is not the full picture. The math I can do locally doesn't fully explain 59% in 2 days on a 20x plan. There may be:

1. Prompt caching not working as expected - Are cache misses causing full context re-sends?
2. Hidden token costs - Something in the pipeline consuming tokens I can't see
3. Billing calculation changes - Did MAX plan token accounting change?
4. Subagent overhead higher than documented - Each spawn may cost more than visible

Work Done in 2 Days

• 10 git commits
• ~5 GSD plan executions
• BDD test fixes
• Accessibility remediations
• Normal development workflow

This workload previously fit fine within weekly budgets.

Request

1. Investigate recent changes - What changed in the last 1-2 weeks that could cause this?
2. Token breakdown in /usage - Show context vs response vs subagent costs
3. Lazy skill loading - When ENABLE_TOOL_SEARCH=true, don't embed full skill list
4. Subagent context optimization - Don't duplicate full system prompt per spawn
5. Prompt caching visibility - Show cache hit/miss rates if caching is involved

Impact

At current burn rate:

• Weekly budget exhausted by day 3-4
• Paying $200/month for 3-4 days of usable access
• 85%+ of tokens appear to go to overhead, not work

Happy to provide more diagnostics if needed.