$ how-boost-is-different
Command compression that proves it helped.
The short answer
Compression ratio is not the outcome
Across the complete suite, Boost made the same agent loop faster and cheaper.
- Full-suite duration
- −28.9%
- Cost per task
- −13.5%
- Evaluated workload
- 89 tasks
Terminal-Bench 2.0, Claude Code with Haiku 4.5, complete 89-task suite. Values compare Boost with the uncompacted baseline. Full benchmark report.
Task success was validated separately: both cohorts passed 30.9% in an earlier 81-task run. Prior capability check.
The Boost advantage
Why Boost is the better command-output choice
All four reduce tokens, but at different layers. RTK compresses commands, Headroom compresses the wider context stream, and Caveman shortens assistant replies. Boost stays at the command boundary and adds a measurable safety loop: preserve requested detail, recover hidden output, learn from every retrieval, and verify that agents still complete the task.
| Capability | Boost | RTK | HRHeadroom | CaveCaveman |
|---|---|---|---|---|
| Command output compression | ✓ | ✓ | ✓ | × |
| Full-context and RAG compression | ✓ | × | ✓ | × |
| Assistant reply compression | × | × | × | ✓ |
| Command output recovery | ✓ | ✓ | ✓ | × |
| Explicit verbose flags preserve command detail | ✓ | ✓ | × | × |
| Native approval sees original executable | ✓ | × | — | — |
| Boost learning loop | ||||
| Retrieval records filter, version, and reason | ✓ | × | × | × |
| Disables filters that repeatedly hide useful detail | ✓ | × | × | × |
| End-to-end agent task + cost A/B | ✓ | × | × | × |
✓ documented capability · × not a documented primary capability · — not applicable at that layer. Public project docs reviewed July 19, 2026.
Approval safety compares whether the host sees the original executable. RTK separately evaluates host permission rules before rewriting.
The honest summary
- Does better
- Task-level proof, native approval visibility, versioned rollback learning, and OTLP evidence.
- Also does
- Flag-aware command compression, full-context/RAG compression, and raw-output recovery.
- Does not do
- Shorter assistant replies.
When you do not need Boost
Skip it if your agent rarely runs shell commands or your main problem is verbose replies.
Data, telemetry, and preview status
Redacted originals stay local. Boost sends product telemetry to JFrog under the Preview Agreement. External OTLP export is optional.
Try it on your workload
Let your numbers decide
Run one session. Open boost report.
curl -fsSL https://boost.jfrog.com/install.sh | bash