2 months ago
- rtk (Rust Token Killer) claims 60–90% token savings by compressing shell command output, but a paired A/B benchmark shows no actual savings.
- Most agent bytes never reach the hook because Claude Code uses built-in tools (Read, Grep) and many commands are not covered; only about 20% of tool output is eligible.
- Measured token cost increased by a median +7.6% at low reasoning effort and was flat at high effort, with no quality degradation.
- rtk's own analytics overestimate savings by counting raw output that Claude Code would truncate and ignoring cached re-reads.
- The tool's design is honest and functional, but its counterfactual assumptions are incorrect, leading to a net cost increase rather than savings.