Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Persistence loop — keep working on a task until it's resolved or until a circuit breaker stops you. Different from /loop (which is time-based) — ralph is iteration-based with state.
.claude/skills/evolution-foundation-dev-ralph/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | -44% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 60% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -37% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -53% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -45% | 0% |
Derived from oh-my-claudecode (MIT, Yeachan Heo). Adapted for the EvoNexus Engineering Layer.
Persistence loop. Keep iterating on a task until it's resolved, with a circuit breaker for safety. Different from /loop (which runs a prompt every N minutes) — ralph runs the same agent repeatedly until a goal is met.
/loop (built-in)Iteration 1:
- Run agent (e.g., @bolt-executor) with task
- Check completion signal
- If met: STOP, success
- If not: record what was tried, what failed
- Increment iteration counter
Iteration 2:
- Run agent with task + accumulated state
- Check completion signal
- ...
Iteration 5 (max):
- Stop and report: "Reached max iterations without completion"
- Hand off to @apex-architect with full state dump
OR
3-identical-failure breaker:
- If iteration N produces the same error fingerprint as N-1 and N-2
- Stop immediately
- The issue is no longer "fix the bug" — it's "question the approach"Per iteration, emit:
[ralph iteration {N}/{max}]
Goal: {completion signal}
Tried: {what this iteration did}
Result: {success | failure: {reason}}
State: {what's been tried so far}Final report saved to workspace/development/debug/[C]ralph-{topic}-{date}.md.
@bolt-executor (most common driver)@hawk-debugger (for bug-fix loops)@oath-verifier (provides the completion signal)@apex-architect (escalation when circuit breaker trips)Interrupted by dev-cancel — preserves state for resume.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-12 | fail→pass | 14,986 | 5,693 | -62% | 1 | 1 | 0% | 2,532 | 1,418 | -44% | 0 | 0 | — |
case-13 | fail→fail | 11,405 | 10,251 | -10% | 1 | 1 | 0% | 2,404 | 2,523 | +5% | 0 | 0 | — |
case-01 | fail→fail | 18,652 | 4,474 | -76% | 1 | 1 | 0% | 3,269 | 1,168 | -64% | 0 | 0 | — |
case-02 | fail→fail | 4,544 | 5,525 | +22% | 1 | 1 | 0% | 255 | 1,039 | +307% | 0 | 0 | — |
case-03 | fail→fail | 6,935 | 5,451 | -21% | 1 | 1 | 0% | 180 | 1,372 | +662% | 0 | 0 | — |
case-04 | pass→pass | 12,084 | 4,748 | -61% | 1 | 1 | 0% | 1,756 | 1,365 | -22% | 0 | 0 | — |
case-05 | fail→pass | 14,661 | 1,856 | -87% | 1 | 1 | 0% | 567 | 908 | +60% | 0 | 0 | — |
case-06 | fail→pass | 9,031 | 1,634 | -82% | 1 | 1 | 0% | 1,302 | 814 | -37% | 0 | 0 | — |
case-07 | fail→pass | 13,861 | 4,271 | -69% | 1 | 1 | 0% | 2,171 | 1,031 | -53% | 0 | 0 | — |
case-08 | fail→pass | 10,197 | 2,157 | -79% | 1 | 1 | 0% | 1,662 | 917 | -45% | 0 | 0 | — |
case-09 | fail→pass | 12,223 | 3,487 | -71% | 1 | 1 | 0% | 1,822 | 1,145 | -37% | 0 | 0 | — |
case-10 | fail→pass | 10,600 | 1,757 | -83% | 1 | 1 | 0% | 1,733 | 847 | -51% | 0 | 0 | — |
case-11 | fail→pass | 18,545 | 3,413 | -82% | 1 | 1 | 0% | 1,305 | 1,184 | -9% | 0 | 0 | — |
case-14 | pass→fail | 3,206 | 9,927 | +210% | 1 | 1 | 0% | 445 | 1,713 | +285% | 0 | 0 | — |
case-15 | fail→fail | 18,504 | 5,742 | -69% | 1 | 1 | 0% | 3,608 | 851 | -76% | 0 | 0 | — |
case-16 | fail→pass | 11,498 | 3,568 | -69% | 1 | 1 | 0% | 1,729 | 1,210 | -30% | 0 | 0 | — |
case-17 | pass→pass | 12,603 | 4,175 | -67% | 1 | 1 | 0% | 2,027 | 1,282 | -37% | 0 | 0 | — |
case-18 | fail→pass | 10,673 | 4,246 | -60% | 1 | 1 | 0% | 1,586 | 1,212 | -24% | 0 | 0 | — |
case-19 | pass→pass | 7,065 | 2,535 | -64% | 1 | 1 | 0% | 1,038 | 976 | -6% | 0 | 0 | — |
case-20 | fail→pass | 24,347 | 4,062 | -83% | 1 | 1 | 0% | 1,013 | 1,168 | +15% | 0 | 0 | — |
case-21 | pass→pass | 10,014 | 6,053 | -40% | 1 | 1 | 0% | 1,604 | 1,557 | -3% | 0 | 0 | — |
case-22 | fail→pass | 12,048 | 6,632 | -45% | 1 | 1 | 0% | 1,893 | 1,620 | -14% | 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 17 counted toward the lift figure. The other 5 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 +50 percentage points is the difference between those two pass rates over the 17 comparable cases. 1 case got worse with the skill loaded, and it is 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.