v2.1.88: Massive context window consumption increase (~2% per prompt)

Description

Since updating to Claude Code v2.1.88, context window usage has increased dramatically compared to previous versions.

Observed behavior

• Every single prompt consumes approximately 2% of the context window
• In a 15-minute session, context usage reached 33% — this never happened before
• Previous versions (v2.1.87 and earlier) had significantly lower per-prompt consumption
• This is on a Max plan with 1M context — so 2% per prompt = ~20K tokens per turn just in overhead

Expected behavior

Context window consumption should be comparable to v2.1.87 where a 15-minute session would use a fraction of what v2.1.88 consumes.

Environment

• Claude Code version: v2.1.88
• Model: Opus 4.6 (1M context) with high effort
• Plan: Claude Max (unlimited)
• OS: macOS 15 (Darwin 25.3.0)
• Node: v22.17.0

Possible cause

Something may have changed in the prompt building or system prompt size between v2.1.87 and v2.1.88 that significantly increases token consumption per turn.

Workaround

Downgrading to v2.1.87 via npm install -g @anthropic-ai/claude-code@2.1.87 appears to reduce the issue.