Token Autopsy — AI skill for Claude Code

token-autopsy

[](https://github.com/bugyal/token-autopsy/actions/workflows/ci.yml) [](https://github.com/bugyal/token-autopsy/releases/latest) [](https://pkg.go.dev/github.com/bugyal/token-autopsy)

**token-autopsy is a per-behavior cost autopsy for agent session transcripts.**

Why

Agent fleets (Claude Code, Codex, Cursor) burn tokens in ways the invoice never explains. Total spend per day is easy to get; *which behaviors* drove that spend is not. Teams end up tuning prompts blind while the actual waste — re-reading the same file nine times, retrying a failing command in a tight loop, pasting 40 MB of logs into context — goes unmeasured.

token-autopsy ingests a week of session transcripts and attributes token cost to the behaviors that caused it, so you can fix the workflow instead of just paying for it.

What it does

It walks a directory of `.jsonl` session transcripts (Claude Code format: one JSON object per line with `timestamp`, `message`, `tool_calls`; Codex/Cursor flat records tolerated), normalizes every message and tool call into a single event stream, detects cost-driving behaviors, and prints a table sorted by estimated token cost:

Behavior	What it detects	Attributed cost
re-reads	Same file_path in multiple Read/Glob calls within one session	Every read after the first
retry-storms	Identical tool call (same tool + args) repeated >2× in a row	Repeats beyond the threshold
tool-result-bloat	Tool result chars / call content chars ≥ 3.0 (flag with --bloat-ratio)	Result volume above the threshold
cache hit-rate	Sessions carrying cache_read_input_tokens / cache_creation_input_tokens metadata	Informational line under the table

Token counts come from tr