Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Apply one rewrite-plan step at a time as small reviewable file edits, never rewriting whole files when a localised change suffices.
.claude/skills/nexu-io-patch-edit/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-11 | ✗→✓ | ▲ Improved | -40% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -14% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -58% | 0% |
| case-14 | ✗→✓ | ▲ Improved | -62% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -51% | 0% |
Spec §20.3 / §21.3.2: code-migration / tune-collab cannot ship a trustworthy diff when the agent rewrites whole files in one pass — the reviewer can't audit the change. This atom enforces "one step per turn, smallest possible localised edit".
plan/steps.json from rewrite-plan.iterations (the daemon's per-stage counter).code/index.json for the file's known imports + neighbours.The atom mutates files inside the project cwd via the file-edit tool. After every iteration, it writes a per-step receipt:
textproject-cwd/ └── plan/ ├── steps.json (status updated: 'pending' → 'completed' | 'skipped') └── receipts/ └── step-<id>.json # { files: ['...'], diffSummary, rationale, completedAt }
The atom completes when every step in plan/steps.json is in a terminal state (completed / skipped) OR iterations >= OD_MAX_DEVLOOP_ITERATIONS. Use:
jsonc{ "id": "patch-edit-loop", "atoms": ["patch-edit"], "repeat": true, "until": "plan.steps.terminal === plan.steps.total || iterations >= 8" }
files[].shell-tier file when the step's risk is not'high' or the user hasn't confirmed.
Implemented by the daemon runner in apps/daemon/src/plugins/atoms/patch-edit.ts. It validates step ownership and risk, applies unified diffs with atomic writes, and records receipts and progress.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-11 | fail→pass | 19,744 | 3,672 | -81% | 1 | 1 | 0% | 1,695 | 1,019 | -40% | 0 | 0 | — |
case-01 | fail→fail | 5,493 | 4,963 | -10% | 1 | 1 | 0% | 204 | 645 | +216% | 0 | 0 | — |
case-02 | fail→fail | 2,770 | 5,506 | +99% | 1 | 1 | 0% | 328 | 771 | +135% | 0 | 0 | — |
case-03 | fail→fail | 5,201 | 6,596 | +27% | 1 | 1 | 0% | 184 | 667 | +263% | 0 | 0 | — |
case-04 | pass→pass | 27,600 | 21,532 | -22% | 1 | 1 | 0% | 5,096 | 3,815 | -25% | 0 | 0 | — |
case-05 | fail→fail | 4,209 | 9,642 | +129% | 1 | 1 | 0% | 724 | 1,630 | +125% | 0 | 0 | — |
case-06 | fail→fail | 14,756 | 8,584 | -42% | 1 | 1 | 0% | 1,140 | 1,364 | +20% | 0 | 0 | — |
case-07 | fail→fail | 14,420 | 4,921 | -66% | 1 | 1 | 0% | 1,899 | 1,167 | -39% | 0 | 0 | — |
case-08 | pass→pass | 7,225 | 3,285 | -55% | 1 | 1 | 0% | 795 | 1,072 | +35% | 0 | 0 | — |
case-09 | pass→pass | 8,468 | 5,159 | -39% | 1 | 1 | 0% | 1,126 | 1,302 | +16% | 0 | 0 | — |
case-10 | fail→fail | 10,378 | 7,777 | -25% | 1 | 1 | 0% | 1,561 | 1,361 | -13% | 0 | 0 | — |
case-12 | fail→pass | 7,123 | 2,702 | -62% | 1 | 1 | 0% | 1,032 | 890 | -14% | 0 | 0 | — |
case-13 | fail→pass | 17,237 | 4,578 | -73% | 1 | 1 | 0% | 2,535 | 1,059 | -58% | 0 | 0 | — |
case-14 | fail→pass | 11,913 | 1,813 | -85% | 1 | 1 | 0% | 1,819 | 690 | -62% | 0 | 0 | — |
case-15 | fail→pass | 12,108 | 4,058 | -66% | 1 | 1 | 0% | 1,938 | 958 | -51% | 0 | 0 | — |
case-16 | pass→pass | 6,415 | 2,908 | -55% | 1 | 1 | 0% | 946 | 839 | -11% | 0 | 0 | — |
case-17 | fail→pass | 9,712 | 3,039 | -69% | 1 | 1 | 0% | 1,292 | 983 | -24% | 0 | 0 | — |
case-18 | pass→pass | 5,105 | 4,504 | -12% | 1 | 1 | 0% | 769 | 1,025 | +33% | 0 | 0 | — |
case-19 | fail→pass | 16,862 | 2,331 | -86% | 1 | 1 | 0% | 2,118 | 692 | -67% | 0 | 0 | — |
case-20 | fail→pass | 10,650 | 3,575 | -66% | 1 | 1 | 0% | 1,396 | 981 | -30% | 0 | 0 | — |
case-21 | pass→pass | 8,388 | 3,520 | -58% | 1 | 1 | 0% | 1,338 | 893 | -33% | 0 | 0 | — |
case-22 | fail→pass | 26,998 | 2,801 | -90% | 1 | 1 | 0% | 1,509 | 929 | -38% | 0 | 0 | — |
case-23 | fail→pass | 14,492 | 3,987 | -72% | 1 | 1 | 0% | 1,944 | 992 | -49% | 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, and 19 counted toward the lift figure. The other 4 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 +43 percentage points is the difference between those two pass rates over the 19 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.