saver-audit replays rtk on users' own Claude Code/Codex logs: could you check the method?

[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 the recorded tool outputs through popular token savers using the user's own installed copy.

For rtk it runs rtk pipe --filter <name> on each recorded shell output whose command rtk's hook would rewrite to that filter. I checked the routing against rtk rewrite (0.50.0) on all 20,726 shell commands in my last 30 days: about 97% of the output I credit to rtk is output your hook actually rewrites. On those 30 days rtk cut $21.49 of $4,209 (0.5%) API-equivalent, labelled "replayed".

It's shown as a lower bound, because your hook also rewrites commands that have no pipe filter (curl, ls, wc, git status/log), which can't be replayed from logs. git-status and git-log are also left out, because live rtk changes their format.

Manifest: https://github.com/VladUZH/saver-audit/blob/main/src/savers/builtin/rtk.json

Is that fair to rtk? If a stdin mode for the curl/ls/git filters is possible, I'd happily use it. I'm sharing these results publicly this week and will update them if you spot a problem.