Repository navigation
v0.1.0 — Claude Code session forensics
Five skills for reading what Claude Code actually did.
Skills
| Skill | What it answers |
|---|---|
read-sessions |
Which model actually answered when you asked for another; where the tokens went; what compaction threw away |
audit-install |
Why a hook, skill, plugin or CLAUDE.md silently stopped working |
triage-failure |
What actually broke when a request failed with an empty error box |
audit-spend |
Where the spend went, per model and per prompt |
build-atlas |
One offline HTML page of the whole history |
Each ships as a zip with SKILL.md at the archive root. Install by unzipping into your skills directory, or copy the folder from skills/ in the repo.
Why the numbers can be trusted
Three defects were found by generating a real report and checking it against the raw logs, not by reading the code:
- Compaction loss was multiplied.
cumulativeDroppedTokensis cumulative per session; summing it over boundaries reported 437M dropped where the per-event deltas total 19.4M. - Resumed sessions were counted twice. A resumed session replays its history into a new file with the same cumulative cost. 398 files held 330 session ids; summing files turned ~$397 into $3,562.
- Spend was called an estimate when it was a placeholder. The client flags unknown-priced models itself, and 24 of 34 records carried that flag.
What is tested
- 169 unit tests (
npx vitest run --pool=forks) - 7 trigger evals (
claude plugin eval .) — recorded run 7/7 at--runs 1; the honest default is--runs 3and costs about three times as much
The limit worth knowing
Every figure rests on an undocumented file format. If Claude Code changes its shape, these tools report unmeasured and print no totals rather than printing zeroes. The report footer states the date the format was last verified, and the command to re-verify it.
Read-only. Node 20+. No runtime dependencies, no network, no telemetry.
MIT · acrot0