Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Inspect disk pressure with SBH and run one
.claude/skills/boshu2-sbh/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | 54% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -4% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 106% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 25% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -37% | 0% |
SBH exposes disk-pressure status, ballast, scanning, and recovery commands. This skill gathers evidence and performs at most the explicit action authorized by the caller. Its authorized actions can delete files and change host storage configuration, so no mutation runs without explicit caller authority. The command surface below is anchored to sbh 0.4.27; re-verify against sbh --version on another host before relying on an exact flag.
Evidence-first recovery works because disk pressure has cheap reversible remedies and expensive irreversible ones; a factual baseline is what tells them apart before anything is deleted. Order remediation by irreversibility: status and dry runs first, ballast release next, cleanup of unprotected files after that, and emergency deletion last — never skip forward while a more reversible step remains untried.
Named failure mode — wrong-mount relief: reclaiming gigabytes on a filesystem that is not the constrained mount and declaring the pressure resolved.
Anti-pattern: escalating straight to emergency --yes because pressure is critical. Corrective: urgency raises the stakes of an irreversible mistake; it never lowers the authorization bar or the ordering above.
sbh --json status, sbh check, or a dry run on the exact mountbecause recovery decisions need a factual baseline.
clean --yes, emergency --yes, ballast release, tune --apply,unprotect, or service/configuration changes without explicit authority because those operations mutate host state.
.git/, open-file, young-file, non-writable-parent, and.sbh-protect vetoes because they prevent unsafe or ineffective cleanup.
free space on another filesystem does not resolve the pressure.
caller owns any further mutation. A negative result is returned to the caller; this skill does not retry or escalate it.
Return mount, free bytes, pressure state, dry-run candidates, authorization used, exact command and exit code, bytes reclaimed, protection vetoes, and checked/not checked surfaces.
Full command documentation: <https://github.com/Dicklesworthstone/storage_ballast_helper>
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 15,676 | 6,450 | -59% | 1 | 1 | 0% | 2,782 | 907 | -67% | 0 | 0 | — |
case-02 | fail→fail | 6,209 | 7,073 | +14% | 1 | 1 | 0% | 162 | 931 | +475% | 0 | 0 | — |
case-03 | fail→fail | 6,523 | 6,032 | -8% | 1 | 1 | 0% | 414 | 890 | +115% | 0 | 0 | — |
case-04 | pass→pass | 8,378 | 6,879 | -18% | 1 | 1 | 0% | 1,229 | 1,649 | +34% | 0 | 0 | — |
case-05 | fail→fail | 9,985 | 5,693 | -43% | 1 | 1 | 0% | 1,640 | 848 | -48% | 0 | 0 | — |
case-06 | fail→pass | 10,756 | 10,903 | +1% | 1 | 1 | 0% | 1,617 | 2,496 | +54% | 0 | 0 | — |
case-07 | fail→pass | 10,445 | 5,580 | -47% | 1 | 1 | 0% | 1,456 | 1,395 | -4% | 0 | 0 | — |
case-08 | fail→pass | 7,460 | 9,847 | +32% | 1 | 1 | 0% | 1,046 | 2,150 | +106% | 0 | 0 | — |
case-09 | pass→pass | 9,032 | 7,505 | -17% | 1 | 1 | 0% | 1,399 | 1,738 | +24% | 0 | 0 | — |
case-10 | fail→pass | 9,199 | 13,643 | +48% | 1 | 1 | 0% | 1,524 | 1,898 | +25% | 0 | 0 | — |
case-11 | fail→pass | 10,426 | 3,645 | -65% | 1 | 1 | 0% | 1,700 | 1,070 | -37% | 0 | 0 | — |
case-12 | fail→fail | 9,690 | 6,999 | -28% | 1 | 1 | 0% | 1,587 | 1,586 | -0% | 0 | 0 | — |
case-13 | fail→pass | 9,683 | 9,988 | +3% | 1 | 1 | 0% | 1,469 | 2,192 | +49% | 0 | 0 | — |
case-14 | pass→fail | 8,257 | 12,774 | +55% | 1 | 1 | 0% | 1,119 | 2,686 | +140% | 0 | 0 | — |
case-15 | fail→pass | 5,436 | 9,876 | +82% | 1 | 1 | 0% | 886 | 2,269 | +156% | 0 | 0 | — |
case-16 | fail→pass | 14,009 | 11,805 | -16% | 1 | 1 | 0% | 2,342 | 2,337 | -0% | 0 | 0 | — |
case-17 | pass→fail | 7,897 | 6,861 | -13% | 1 | 1 | 0% | 1,218 | 851 | -30% | 0 | 0 | — |
case-18 | fail→pass | 7,664 | 7,283 | -5% | 1 | 1 | 0% | 1,349 | 1,783 | +32% | 0 | 0 | — |
case-19 | fail→fail | 13,902 | 7,380 | -47% | 1 | 1 | 0% | 2,411 | 925 | -62% | 0 | 0 | — |
case-20 | fail→pass | 8,426 | 10,059 | +19% | 1 | 1 | 0% | 1,156 | 1,766 | +53% | 0 | 0 | — |
case-21 | fail→pass | 7,392 | 14,201 | +92% | 1 | 1 | 0% | 1,124 | 2,073 | +84% | 0 | 0 | — |
case-22 | pass→pass | 7,427 | 12,203 | +64% | 1 | 1 | 0% | 1,307 | 1,900 | +45% | 0 | 0 | — |
case-23 | pass→pass | 7,339 | 8,779 | +20% | 1 | 1 | 0% | 1,244 | 1,888 | +52% | 0 | 0 | — |
case-24 | pass→pass | 5,594 | 14,008 | +150% | 1 | 1 | 0% | 1,028 | 2,271 | +121% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 24 cases were attempted, and 18 counted toward the lift figure. The other 6 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +38 percentage points is the difference between those two pass rates over the 18 comparable cases. 5 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.