| name | cc-optimizer:analyze |
| description | Analyze Claude Code transcripts: aggregate token usage, models, cache metrics per project (zero model tokens spent, pure arithmetic). |
| argument-hint | [path/to/sessions] [output.json] |
| allowed-tools | Bash, Read |
/cc-optimizer:analyze
Low-level analyzer: aggregates Claude Code session transcripts into per-project statistics without spending any model tokens.
Usage
/cc-optimizer:analyze # Uses ~/.claude/projects
/cc-optimizer:analyze /path/to/sessions optimizer/out/raw-stats.json
Output: raw-stats.json
{
"projects": {
"project-slug": {
"sessions": 150,
"total_tokens": 8048909762,
"input_tokens": 156819318,
"output_tokens": 257451667,
"cache_creation": 2150345336,
"cache_read": 30117123696,
"cache_hit_ratio": 0.9288,
"by_model": {
"claude-opus-4-8": { "msgs": 5000, "tokens": 18012088391 },
...
}
}
}
}
What's measured (from your transcripts)
- Total tokens by model
- Cache creation vs. read (1h vs 5m ephemeral)
- Cache hit ratio
- Output-size histogram (identifies trivial work on expensive models)
- Subagent vs. main task spending
What's NOT spent
- Zero model API calls — pure Python arithmetic on recorded
message.usage data
- All heavy lifting is deterministic, reusable, reproducible
