saver-audit replays headroom on users' own logs: method check, and a stdin CLI?

[VladUZH]

Hi! I built saver-audit (npx saver-audit, MIT, fully offline). It reads a developer's local Claude Code and Codex logs, shows where their tokens went (priced with prompt-cache costs), and replays recorded tool outputs through token savers using the user's own installed copy.

For headroom (0.38.0) it runs a small Python sidecar that calls your compressor on each recorded tool output of 200+ tokens, replayed as the newest message, fully offline (HF/tiktoken offline flags, beacon off). On my last 30 days: $150.97 of $4,209 (3.6%), the largest measured cut of the savers I tested. On Codex alone it was 10.4%.

It's labelled a lower estimate, because headroom also compresses older Read/Grep/Glob/Edit/Write outputs as they age, and that isn't replayed.

Code: https://github.com/VladUZH/saver-audit/blob/main/src/savers/registry.ts (the headroom entry)

Two questions:

1. Is that method fair?
2. Would you consider a headroom compress --stdin CLI? It would replace my sidecar (faster, simpler), and your users could run it too. The sidecar is also the slow part: the full replay takes over an hour on a busy month.

I'm sharing results publicly this week and will update them if you spot a problem.