What this audit covers. 16 real Claude Code coding-agent sessions over ~18 days — reading code, editing files, running commands & tests (845 steps; Bash, Edit and Read dominate), across Claude Sonnet, Opus & Haiku.
in the last 18 days
57%
total opportunity identified
$114.51 safe today+ $144.43 pending validation
where the safe savings came from
- Most of it $97.35 from compacting live turns the model confirmed safe
- Then $11.81 from trimming bloated tool output (whitespace, repeated logs)
- And $5.35 from dropping data it never looked at again
- And $0.00 from hiding tools the agent never uses