measuring-claude-session-start-tokenslisted
Install: claude install-skill sblattj/agent-chandlery
## More triggers
real-start-tokens (old name)
# measuring-claude-session-start-tokens
```bash
S=${SKILL_DIR:-~/.claude/skills/measuring-claude-session-start-tokens}/scripts/start_tokens.py
python3 $S # newest 5 fresh interactive sessions of THIS project (cwd)
python3 $S --all-projects # across every project
python3 $S --last 20 # more history
python3 $S X.jsonl # specific sessions
# --dir ~/.claude/projects/<proj> one project --project PATH --all-entrypoints include -p/sdk --json
# honors CLAUDE_CONFIG_DIR; helper lib/ccpaths.py ships beside scripts/
```
**Start** is exact: the first assistant `usage` (input + cache_creation + cache_read). Block columns are attachment chars / 4, so estimates. The system prompt and tool schemas aren't logged; they are the rest.
## Gotchas (all hit in real measuring)
- **Only `entrypoint: cli` is a real start.** A `claude -p` run reads ~21k low (no Artifact tool, a smaller system prompt), so the default filter drops it.
- **Rows are keyed by session start time.** File mtime moves whenever a session is resumed, so a morning session looked like it ran after the diet.
- **Compare like with like.** A resumed or compacted session starts with its summary, and a pasted image or long prompt adds to it. For a clean number, open a fresh session with a one-line prompt.
- **Agent view pre-warms sessions,** so the first session after a settings change can still carry the old settings. Measure twice.
- `/contex