Problem
skillkit context reports CLAUDE.md as ~1.3K tokens, but it doesn't resolve @path/file.md references. Those files get inlined by Claude Code at session start and count against every API call.
Example
My CLAUDE.md has 9 @ imports totaling ~1,386 lines (~12.5K tokens). Skillkit reports 3.9K total, real load is ~19K. That's a 5x undercount.
CLAUDE.md itself: ~2,500 tokens
9 @ imports: ~12,500 tokens <-- invisible to skillkit
Skills metadata: ~2,600 tokens
MEMORY.md (200 lines): ~1,500 tokens <-- also not counted
Expected behavior
skillkit context should:
- Parse
@path/file.md lines in CLAUDE.md
- Read and count tokens for each referenced file
- Show them in the CLAUDE.MD BREAKDOWN section
- Include MEMORY.md (~/.claude/projects/*/memory/MEMORY.md) in the total
Notes
Claude Code loads @ references eagerly at session start, not lazily. They're textual includes, like #include in C. Every referenced file is present in every API call for the entire session.
Problem
skillkit contextreports CLAUDE.md as ~1.3K tokens, but it doesn't resolve@path/file.mdreferences. Those files get inlined by Claude Code at session start and count against every API call.Example
My CLAUDE.md has 9
@imports totaling ~1,386 lines (~12.5K tokens). Skillkit reports 3.9K total, real load is ~19K. That's a 5x undercount.Expected behavior
skillkit contextshould:@path/file.mdlines in CLAUDE.mdNotes
Claude Code loads
@references eagerly at session start, not lazily. They're textual includes, like#includein C. Every referenced file is present in every API call for the entire session.