Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Context window manager. Snapshots session state, compacts conversation, and resumes via hook. Use when context exceeds 70% or a large new task begins.
.claude/skills/hashgraph-online-context/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -8% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -50% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 24% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 8% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -35% | 0% |
Before compacting, ensure critical state is saved:
$HARNESS_DIR/sessions/Suggest or trigger compaction with a clear summary:
Summary for compaction:
- Working on: [task description]
- Files modified: [list]
- Current status: [what's done, what remains]
- Key decisions: [important choices made]
- Next step: [what to do after compaction]After compaction, the resume hook will reload:
$HARNESS_DIR/sessions/$HARNESS_DIR/memory/$HARNESS_DIR/evolved/| Excuse | Rebuttal | What to do instead | |--------|----------|-------------------| | "I still have context left" | Quality degrades well before the hard limit. 70% is the trigger. | Compact proactively. Better to lose 5 min than lose coherence. | | "Compacting will lose important context" | That's why you preserve first. The resume hook restores it. | Save state → compact → resume loads snapshot + memory + skills. | | "I'll just re-read the files" | Re-reading 5 large files wastes 30%+ of your fresh context. | Summarize key facts before compacting. Read only what you need after. |
Before compacting, confirm ALL of these:
$HARNESS_DIR/sessions/ (show file name)Compacting without a summary = guaranteed context loss.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-06 | pass→pass | 16,562 | 6,943 | -58% | 1 | 1 | 0% | 1,665 | 1,485 | -11% | 0 | 0 | — |
case-01 | fail→fail | 17,116 | 16,968 | -1% | 1 | 1 | 0% | 2,196 | 965 | -56% | 0 | 0 | — |
case-02 | fail→fail | 14,035 | 19,656 | +40% | 1 | 1 | 0% | 1,313 | 2,069 | +58% | 0 | 0 | — |
case-03 | fail→fail | 14,667 | 10,815 | -26% | 1 | 1 | 0% | 1,384 | 1,211 | -13% | 0 | 0 | — |
case-04 | fail→pass | 16,385 | 10,427 | -36% | 1 | 1 | 0% | 1,367 | 1,251 | -8% | 0 | 0 | — |
case-05 | fail→pass | 19,848 | 8,345 | -58% | 1 | 1 | 0% | 1,941 | 972 | -50% | 0 | 0 | — |
case-07 | fail→pass | 7,782 | 4,899 | -37% | 1 | 1 | 0% | 1,064 | 1,316 | +24% | 0 | 0 | — |
case-08 | pass→pass | 9,574 | 11,765 | +23% | 1 | 1 | 0% | 1,481 | 1,758 | +19% | 0 | 0 | — |
case-09 | fail→pass | 17,081 | 21,839 | +28% | 1 | 1 | 0% | 2,508 | 2,701 | +8% | 0 | 0 | — |
case-10 | fail→pass | 21,209 | 11,873 | -44% | 1 | 1 | 0% | 2,190 | 1,424 | -35% | 0 | 0 | — |
case-11 | fail→pass | 10,622 | 8,729 | -18% | 1 | 1 | 0% | 1,251 | 1,166 | -7% | 0 | 0 | — |
case-12 | fail→pass | 10,093 | 11,410 | +13% | 1 | 1 | 0% | 1,375 | 1,448 | +5% | 0 | 0 | — |
case-13 | fail→pass | 16,501 | 8,239 | -50% | 1 | 1 | 0% | 2,195 | 1,005 | -54% | 0 | 0 | — |
case-14 | pass→pass | 14,208 | 4,472 | -69% | 1 | 1 | 0% | 1,369 | 1,062 | -22% | 0 | 0 | — |
case-15 | pass→pass | 13,652 | 5,305 | -61% | 1 | 1 | 0% | 1,300 | 1,428 | +10% | 0 | 0 | — |
case-16 | pass→pass | 9,293 | 10,783 | +16% | 1 | 1 | 0% | 1,220 | 1,545 | +27% | 0 | 0 | — |
case-17 | fail→pass | 18,482 | 14,805 | -20% | 1 | 1 | 0% | 1,964 | 1,912 | -3% | 0 | 0 | — |
case-18 | fail→pass | 15,560 | 8,207 | -47% | 1 | 1 | 0% | 1,590 | 983 | -38% | 0 | 0 | — |
case-19 | pass→pass | 11,502 | 3,673 | -68% | 1 | 1 | 0% | 1,056 | 1,191 | +13% | 0 | 0 | — |
case-20 | pass→pass | 13,225 | 8,349 | -37% | 1 | 1 | 0% | 1,257 | 1,607 | +28% | 0 | 0 | — |
case-21 | pass→pass | 9,215 | 3,369 | -63% | 1 | 1 | 0% | 649 | 1,074 | +65% | 0 | 0 | — |
case-22 | fail→fail | 22,378 | 10,112 | -55% | 1 | 1 | 0% | 3,131 | 2,054 | -34% | 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. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 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 +45 percentage points is the difference between those two pass rates over the 20 comparable cases. 2 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.