migrate
/migrate
Przenosi Bench lub Experiment napisany dla madbench v0.41.0 lub nowszego na zainstalowaną wersję madbench, stosując zmiany z pliku changes.md, gdy po aktualizacji bench się nie ładuje.
- Autor
- madappgang ↗
- Repozytorium
- magus ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 26.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 3
- Skrypty
- 0
- Komentarze HTML
- 1
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 26.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- SKILL.md zawiera komentarze HTML (1). Nie widać ich w podglądzie Markdown, a model je czyta.
Co robi i kiedy
Przenosi Bench lub Experiment napisany dla madbench v0.41.0 lub nowszego na zainstalowaną wersję madbench, stosując zmiany z pliku changes.md, gdy po aktualizacji bench się nie ładuje.
Argumenty
/migrate <bench or Experiment path>
Oryginalny opis (EN)
Brings a madbench Bench or Experiment written for madbench v0.41.0 or later to the installed madbench, applying the changes `changes.md` records between the two versions. Use when a bench no longer loads after an upgrade.
Jak działa
- Step 0 — the Glossary and versions, before anything else
- Step 1 — the two versions
- Step 2 — apply changes.md
- Step 3 — load it
- Step 4 — the free gates, in order
- Step 5 — the words
- Step 6 — stored reports are re-run, never edited
- Report
Podgląd SKILL.md
---
name: migrate
description: Brings a madbench Bench or Experiment written for madbench v0.41.0 or later to the installed madbench, applying the changes `changes.md` records between the two versions. Use when a bench no longer loads after an upgrade.
argument-hint: "<bench or Experiment path>"
disable-model-invocation: true
---
<!-- Owned by the madbench repository and shipped by `madbench skills`; a copy anywhere else
is a copy, fixed upstream. `changes.md` beside this file is the only record of key, flag
and file changes since v0.41.0; this file carries no key map. -->
# Migrate a Bench or Experiment to the installed madbench
Target: $ARGUMENTS
With no path, ask for one. A directory means every `madbench.yaml`, `*.madbench.yaml` and
`*.experiment.yaml` in it, one file at a time.
## Step 0 — the Glossary and versions, before anything else
1. Read `../madbench-evals/glossary.md`, relative to this skill's directory. Write only its
words for madbench concepts; translate the user's words, never echo them.
2. Run `madbench skills --check` and quote its output (Step 0 of `../madbench/SKILL.md`
says how to read it). You migrate to the **installed** binary, and `changes.md` must
cover it: an unpinned set older than the binary is updated first (`madbench skills`),
with the user's yes.
## Step 1 — the two versions
- **Installed:** `madbench version`, quoted.
- **Written for:** the version the user says the file was written for. With no answer, take
v0.41.0 **provisionally**: Step 3 can still find that the file is older.
This skill knows v0.41.0 and every release after it, and nothing earlier. A file written for
a madbench older than v0.41.0 is outside what it knows: say so, and rewrite the file against
`../madbench-evals/schema.md` as new authoring, showing the user every change. This is the
**rewrite path**. Do not reconstruct an older format from memory, and do not report such a
file upstream: an older format is not a madbench gap.
## Step 2 — apply `changes.md`
Read `changes.md` beside this file. A `## vX.Y.Z` section is what release X.Y.Z changed, so a
file written for v0.41.0 takes the `## v0.42.0` section. Take every section newer than the
file's version and not newer than the installed one, oldest first. Apply each old → new line in
those sections exactly as written, and nothing else. A section with no lines changed nothing
that a file spells.
## Step 3 — load it
```bash
madbench list <file>
```
Quote the output verbatim. Exit 0 means it parses; go to Step 4.
An unknown key fails like any typo: `field X not found in type …`, naming the key, the line
and the type it was read into, and nothing else. A key the applied sections did not cover is
not guessed from memory. Look it up in `../madbench-evals/schema.md` and
`../madbench-evals/glossary.md`, make the change, and run `madbench list` again.
A key that neither `schema.md`, `glossary.md` nor an applied `changes.md` section covers goes
one of two ways, decided by Step 1's version:
- **The version is provisional** (the user did not state one) **or older than v0.41.0:** the
file predates v0.41.0. Tell the user, and take Step 1's rewrite
path. No bug report.
- **The user stated v0.41.0 or later:** it is a gap, because `changes.md` should have
recorded it. Follow the gap rule in `../madbench-evals/SKILL.md` (a bug report upstream,
drafted and shown before it is sent).
**Loop until it loads.** The strict decoder names every unknown key it reads in one pass, but
it never reads beneath an unknown key, so fixing a parent can reveal unknown keys under it on
the next pass. Each pass: quote, fix the keys named, re-run. Never batch-rewrite keys the
loader has not named.
## Step 4 — the free gates, in order
`list` proves the file parses, not that it can run. Check types are built at preflight, so
an unknown check type passes `list` and fails here as `unknown check type`, and a key a Check
type does not take fails here as `unknown key "<key>"; the keys are …`. Treat both the way
Step 3 treats an unknown key.
```bash
madbench preflight <file> # exit 3 = blocked, each finding naming its fix; 1 = a Check could not be built
madbench check <file> # the negative control: every graded Check must fail
```
Quote each verbatim, with its exit code. Fix a preflight finding exactly as it is named, then
re-run. A `madbench check` that reports `WRONGLY PASSED` is not a migration error: it is a
Bench that never measured anything, and it goes back to the user as a finding (the review
checklist in `../madbench-evals/SKILL.md`), not a silent fix.
An Experiment: run the gates on the Experiment file, which loads the Bench it names once per
variant.
## Step 5 — the words
With the file loading, rewrite what the loader cannot see: `description:` values, Scenario
descriptions, YAML comments, and any README beside the Bench. Every deprecated word or other
tool's word from `glossary.md` used for a madbench concept becomes the Glossary's word —
"test case" → Scenario, "fixture" → testdata, "arm" → variant. A word naming another tool's
own feature stays: that is a quotation.
`madbench glossary check <the Bench's directory>` lists each one as `path:line word →
replacement`, and exits 0 when none is left.
## Step 6 — stored reports are re-run, never edited
A report or store whose `schema_version` is not the installed madbench's is refused by
`madbench report` and `madbench grade`. Do not hand-edit one into the current shape. Re-run
the Bench and keep the new report.
## Report
```
madbench <version> · skills for <MADBENCH_VERSION>
File <path>
Written for <version, and whether the user said it or it is the provisional v0.41.0 — or "older than v0.41.0: rewritten as new authoring">
Changes applied <each changes.md line applied, with its section — or "none">
Load <each `madbench list` line quoted, and the change made for it — or "loaded first time">
Free gates list / preflight / check — exit and one line each, verbatim
Words <each description or comment rewritten, old → new — or "none">
Reports <reports of another schema_version found, and "re-run" — or "none">
Gaps <none | in a file the user stated is v0.41.0 or later, an unknown key neither changes.md nor schema.md covers, and the bug report, sent or awaiting approval>
```
Pliki w paczce (3)
- MADBENCH_VERSION 1 KB
- SKILL.md 6 KB
- changes.md 7 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · madappgang/magus | 100 | MIT |
| marcel-bich/marcel-bich-claude-marketplace | 100 | MIT |
| abatyuk/okf | 100 | MIT |
| markusleben/ha-nova | 100 | MIT |
| toss/es-toolkit | 100 | MIT |
| rvdbreemen/adr-kit | 100 | MIT |
| asgeirtj/system_prompts_leaks | 100 | CC0-1.0 |
| bennypowers/cem | 100 | GPL-3.0 |
| hainrixz/claude-db | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 6a5655f5fde8 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100madbench-evals
Tworzy, uruchamia i debuguje benche madbench natywnie: YAML benchu, checki, metryki, dane testowe w stanie czerwonym, dwie kontrole i kody wyjścia, a zamiast wrappera przygotowuje szkic zgłoszenia upstream. - 86esp32-bench
Pomaga przy pracy z obwodem lub płytką ESP32 w projektach elektronicznych. - 100bench
Prowadzi przy dodawaniu i edycji benchmarków, mierzeniu czasu działania i kosztu typów dla kompilatora oraz zmianach w plikach bench.md i typebench.md w projekcie Esposter. - 100sanity-bench
Uruchamia i odczytuje wyniki hermetycznego zestawu testów wydajności Studio w perf/bench, lokalnie i przez workflow GitHub Actions, wraz z porównaniami commitów i werdyktami. - 90bench-compare
Porównuje wydajność API między wersjami: wykrywa regresje i szuka przyczyny, gdy ktoś pyta, czy wydajność spadła. - 81bench-harness
Prawdopodobnie ocenia oprogramowanie sterujące stroną w harnessie testowym.