Session Token Ledger

Analyze local OpenClaw session token usage from a generated SQLite ledger and markdown summaries. Use when the user asks for a token audit, token体检报告, context-bloat diagnosis, session efficiency review, largest sessions, daily token trends, input/output ratio analysis, cache-read heavy sessions, or wants SQL/SQLite-based inspection of local OpenClaw session usage.

Analyze local OpenClaw session-token usage from a generated ledger stored under this skill.

What lives here

• scripts/rebuild_sqlite.py — rebuild the local ledger from OpenClaw session transcripts
• scripts/report.py — generate markdown reports for all sessions or one session
• references/overview.md — schema, views, and design notes
• references/queries.sql — canned SQLite queries for common audits
• references/hook-setup.md — optional OpenClaw hook setup for automatic rebuilds
• references/ANOMALIES.md — populated after rebuild with suspicious or bad transcript lines
• assets/session_tokens.db — populated after rebuild; SQLite ledger
• assets/index.json — populated after rebuild; machine-readable session index
• assets/TOTAL_TOKENS.txt — populated after rebuild; quick aggregate summary
• assets/YYYY-MM-DD_N.md — populated after rebuild; one markdown file per completed session

Default workflow

1. Rebuild the ledger with python3 scripts/rebuild_sqlite.py if data looks stale.
2. Start with assets/index.json for session list and top-level totals.
3. If the user wants trends or rankings, inspect assets/session_tokens.db.
4. If the user wants a narrative for one session, read the matching assets/YYYY-MM-DD_N.md file.
5. If the user wants root-cause analysis, inspect these views first:
   • largest_sessions
   • usage_efficiency
   • bloated_sessions
   • top_context_hogs
   • daily_efficiency
6. If something looks wrong, read references/ANOMALIES.md.

Read these references only when needed

• Read references/overview.md when you need schema or view meaning.
• Read references/queries.sql when you need canned SQL.
• Read references/hook-setup.md when you want automatic rebuilds after /new, /reset, /stop, or gateway startup.

Quick path

Generate reports:

python3 scripts/report.py
python3 scripts/report.py --session SESSION_ID
python3 scripts/report.py --save
python3 scripts/report.py --session SESSION_ID --save

Use read-only SQLite queries when available:

sqlite3 -readonly assets/session_tokens.db "SELECT * FROM overall_summary;"
sqlite3 -readonly assets/session_tokens.db "SELECT * FROM largest_sessions LIMIT 10;"
sqlite3 -readonly assets/session_tokens.db "SELECT * FROM bloated_sessions LIMIT 10;"
sqlite3 -readonly assets/session_tokens.db "SELECT * FROM top_context_hogs LIMIT 10;"
sqlite3 -readonly assets/session_tokens.db "SELECT * FROM daily_efficiency ORDER BY date DESC;"

For one session:

sqlite3 -readonly assets/session_tokens.db "SELECT * FROM usage_efficiency WHERE session_id='SESSION_ID';"

If sqlite3 CLI is unavailable, use the bundled Python scripts instead.

Reporting rules

• Lead with the plain-English conclusion.
• Separate total tokens, input tokens, output tokens, and cache read.
• Call out whether waste came from long context, too many topic switches, long outputs, or repeated tool/doc loading.
• For subscription-style billing, emphasize token totals and efficiency, not fake precision on dollar cost.
• When giving recommendations, prefer a short ranked list over a long essay.

Boundaries

• Treat this ledger as local analysis data, not ground truth for provider billing.
• Do not modify the database unless the user explicitly asks to rebuild or update the ledger.
• Rebuilds skip the currently active live .jsonl session when a matching .lock file exists, so the ledger defaults to completed sessions only.
• Prefer querying the DB over manually re-deriving totals from raw session logs unless the ledger appears stale or broken.