Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Headroom startup hooks for Claude Code and GitHub Copilot CLI. Ensures the local Headroom runtime (token compression proxy) is available at session start and before shell/PowerShell tool use. Runs headroom init hook ensure with a 15-second timeout.
.claude/skills/kunanonj-cursor-plugin-headroom/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -62% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -24% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -47% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -80% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -75% | 0% |
Lightweight startup hooks that ensure the Headroom runtime is available for Claude Code and GitHub Copilot CLI sessions.
| Hook | Matcher | Command | Timeout | |------|---------|---------|---------| | SessionStart | startup, resume | headroom init hook ensure | 15s | | PreToolUse | Bash, PowerShell | headroom init hook ensure | 15s |
The headroom init hook ensure command checks for a matching durable headroom init deployment and starts it if needed. This ensures the Headroom compression proxy is running before any API calls are made.
headroom-ai package installed (uv tool install "headroom-ai[all]" or pip install "headroom-ai[all]")headroom init)| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-04 | pass→pass | 8,942 | 1,686 | -81% | 1 | 1 | 0% | 1,785 | 528 | -70% | 0 | 0 | — |
case-01 | fail→pass | 8,041 | 1,899 | -76% | 1 | 1 | 0% | 1,695 | 646 | -62% | 0 | 0 | — |
case-02 | fail→pass | 18,181 | 11,563 | -36% | 1 | 1 | 0% | 3,604 | 2,722 | -24% | 0 | 0 | — |
case-03 | fail→pass | 10,516 | 4,406 | -58% | 1 | 1 | 0% | 2,284 | 1,214 | -47% | 0 | 0 | — |
case-05 | pass→pass | 7,496 | 1,457 | -81% | 1 | 1 | 0% | 1,385 | 456 | -67% | 0 | 0 | — |
case-06 | fail→pass | 10,027 | 1,210 | -88% | 1 | 1 | 0% | 1,950 | 386 | -80% | 0 | 0 | — |
case-07 | fail→pass | 9,232 | 1,333 | -86% | 1 | 1 | 0% | 1,764 | 444 | -75% | 0 | 0 | — |
case-08 | fail→pass | 6,511 | 1,127 | -83% | 1 | 1 | 0% | 1,274 | 345 | -73% | 0 | 0 | — |
case-09 | fail→pass | 4,007 | 1,404 | -65% | 1 | 1 | 0% | 755 | 425 | -44% | 0 | 0 | — |
case-10 | fail→pass | 15,410 | 1,591 | -90% | 1 | 1 | 0% | 3,003 | 512 | -83% | 0 | 0 | — |
case-11 | fail→fail | 10,789 | 3,173 | -71% | 1 | 1 | 0% | 1,918 | 882 | -54% | 0 | 0 | — |
case-12 | fail→pass | 13,042 | 2,732 | -79% | 1 | 1 | 0% | 2,307 | 784 | -66% | 0 | 0 | — |
case-13 | fail→pass | 12,695 | 1,388 | -89% | 1 | 1 | 0% | 2,484 | 432 | -83% | 0 | 0 | — |
case-14 | pass→pass | 4,962 | 1,828 | -63% | 1 | 1 | 0% | 846 | 496 | -41% | 0 | 0 | — |
case-15 | fail→pass | 13,683 | 1,520 | -89% | 1 | 1 | 0% | 2,037 | 379 | -81% | 0 | 0 | — |
case-16 | fail→pass | 13,470 | 1,645 | -88% | 1 | 1 | 0% | 1,955 | 428 | -78% | 0 | 0 | — |
case-17 | fail→pass | 17,667 | 1,528 | -91% | 1 | 1 | 0% | 2,836 | 399 | -86% | 0 | 0 | — |
case-18 | pass→pass | 8,478 | 2,009 | -76% | 1 | 1 | 0% | 1,540 | 603 | -61% | 0 | 0 | — |
case-19 | fail→pass | 7,114 | 1,749 | -75% | 1 | 1 | 0% | 1,421 | 485 | -66% | 0 | 0 | — |
case-20 | fail→pass | 7,630 | 1,444 | -81% | 1 | 1 | 0% | 1,543 | 447 | -71% | 0 | 0 | — |
case-21 | fail→fail | 8,467 | 7,263 | -14% | 1 | 1 | 0% | 1,697 | 1,808 | +7% | 0 | 0 | — |
case-22 | pass→pass | 11,676 | 11,680 | +0% | 1 | 1 | 0% | 2,471 | 2,640 | +7% | 0 | 0 | — |
case-23 | pass→pass | 12,826 | 8,113 | -37% | 1 | 1 | 0% | 2,159 | 1,664 | -23% | 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. 23 cases were attempted. The headline lift of +65 percentage points is the difference between those two pass rates over the 23 comparable cases.
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.