correct
/correct
Zamienia powtarzający się błąd w regułę egzekwowaną przez repozytorium: analizuje historię gita, wybiera najwyższy poziom blokady (architektura, typy, lint, test, dokumentacja) i buduje ją przez lint.py lub .agents/rules.json.
- Autor
- r3dbars ↗
- Repozytorium
- transcripted ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 26.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- 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.
Co robi i kiedy
Zamienia powtarzający się błąd w regułę egzekwowaną przez repozytorium: analizuje historię gita, wybiera najwyższy poziom blokady (architektura, typy, lint, test, dokumentacja) i buduje ją przez lint.py lub .agents/rules.json.
Wyzwalacze z opisu
Oryginalny opis (EN)
Turn a repeated mistake into a rule the repo enforces. Use when the owner corrects an agent for a mistake that has happened before ("you did this again", "stop doing X"), when a review or red CI run shows the same kind of mistake a second time, or when the owner says /correct. Mines git history for the class, picks the highest level that can stop it (architecture, types, lint, test, docs last), builds it with lint.py or .agents/rules.json, proves it on a real past commit, and keeps docs/rules.md current.
Jak działa
- 1. Fix the instance first
- 2. Is it a class?
- 3. Pick the highest level that works
- 4. Build it
- 5. Prove it on the past mistake
- 6. Run it on today's tree
- 7. Keep the table current and run the gates
- 8. Report
Podgląd SKILL.md
---
name: correct
description: Turn a repeated mistake into a rule the repo enforces. Use when the owner corrects an agent for a mistake that has happened before ("you did this again", "stop doing X"), when a review or red CI run shows the same kind of mistake a second time, or when the owner says /correct. Mines git history for the class, picks the highest level that can stop it (architecture, types, lint, test, docs last), builds it with lint.py or .agents/rules.json, proves it on a real past commit, and keeps docs/rules.md current.
---
# Correct: fix the class, not the instance
Adapted from pstack's `/correct` skill for Transcripted's tools. The full process doc is `docs/lint-rules.md`; this is the loop.
Why it exists: every agent here sees only the files it opened, copies the nearest example, and takes the shortest path that compiles. Another sentence in `AGENTS.md` doesn't change that. A check that fails with a message naming the fix does.
## 1. Fix the instance first
Do what the owner asked on the code in front of you. Then decide whether it's a class.
## 2. Is it a class?
A class counts once it happened twice. Look for an earlier instance of the same wrong move (not the same file):
- `git log -S '<the bad call or pattern>' --oneline` and `git log -G '<regex>'`
- `git log --grep '<symptom words>' --oneline`, reverts (`git log --grep '^Revert'`), and commits titled fix/repair after a review
- `python3 scripts/dev/agent-context.py --symptom "<short description>"` to find the owner doc and existing checks
- `docs/rules.md` to see if a rule already covers it (then the rule has a hole: tighten it and add the new case as proof)
If it's truly the first time, fix it, say so, and stop. Don't write a rule from one data point.
For each instance, note the fixing commit and its parent (the bad version). You'll need at least one as proof.
## 3. Pick the highest level that works
Stop at the first one that can catch it:
1. **Architecture.** One owner, one way; delete the old way. Moving a type down or making one function the only writer beats any check.
2. **Types.** Make the bad state fail to compile. Module edges are enforced by `scripts/dev/check-module-boundaries.py`, so a type only the owner module can name is a real wall.
3. **Lint whose error names the fix.** A rule in `scripts/dev/lint_rules/`, or a `scripts/dev/check-*.py` script registered in `.agents/rules.json`. If the pattern is already common, ratchet it.
4. **Behavior test.** Per "Test rules" in `Tests/README.md`. Never a test that reads `Sources/` as text.
5. **Docs, last.** One line in the nearest folder `AGENTS.md`, only for a judgment call no check can make. Not the root `AGENTS.md` unless it applies everywhere.
Say which level you picked and why the higher ones don't work. Often it's two: fix the architecture and add a lint so the old way can't come back.
## 4. Build it
**A lint.py rule** (most corrections land here):
1. `scripts/dev/lint_rules/<rule_id_snake>.py` with `RULE` (`id`, `summary`, `why` citing the commits, `fix` naming the file/type/function to use, `owner_doc`, `level: "lint"`, `paths`, `ratchet`, `proof`) and `check(path, text)`. Use the helpers in `scripts/dev/lint_rules/_base.py` so comments and strings don't trip it.
2. `scripts/dev/lint_rules/fixtures/<rule-id>/` with `bad*` files that must flag and `good*` files that must not. Put the near-misses in `good*`.
3. Read the error as a fresh agent would. It should say exactly what to write instead. If it says only what's wrong, rewrite it.
**Anything else** (architecture, type, test, hook, doc, or a new `check-*.py`): build it, then add an entry to `.agents/rules.json` with `id`, `enforces`, `enforcer`, `level`, `owner_doc`, `why`, `runs_in` and `proof`.
## 5. Prove it on the past mistake
- lint.py: put `"<bad sha>:<path>"` in `proof`, then `python3 scripts/dev/lint.py --self-test`. It runs every fixture and every proof; your proof must flag.
- Other checks: run the check against the bad commit (`git show <sha>:<path>` into a temp tree, or the check's own `--self-test` case built from it) and show it fails. Put the sha in `proof`.
- A test: show it fails with the fix reverted, then passes.
No proof, no rule. If you can't make it fail on the real thing, the rule is guessing.
## 6. Run it on today's tree
`python3 scripts/dev/lint.py --all`. Every finding is either real or a false positive.
- False positives: tighten the rule until they're gone.
- Real and few: fix them in this change.
- Real and many: set `ratchet: True` and record counts with `python3 scripts/dev/lint.py --write-baseline --rule <rule-id>`. It touches only that rule and refuses to raise a rule that already has a baseline. Check the `.agents/lint-baseline.json` diff anyway.
If one line is genuinely fine, it gets `// lint-allow(<rule-id>): <reason> — expires YYYY-MM-DD, approved <handle>`. The approver is a human. Never approve your own.
## 7. Keep the table current and run the gates
1. `python3 scripts/dev/check-rule-registry.py --write` regenerates `docs/rules.md`.
2. `python3 scripts/dev/check-rule-registry.py` and `python3 scripts/dev/lint.py --self-test` pass.
3. `bash check.sh`.
4. If you added a doc line in step 3, keep it to one line and point at the check.
## 8. Report
Tell the owner, short:
- the class, and the two (or more) commits that show it happened before
- the level you picked and why not higher
- what enforces it now (rule id or script) and where it runs: edit hook, Stop hook, pre-push, CI
- proof: which past commit it flags
- what `--all` found on today's tree, how many were false positives, and whether you ratcheted
- anything that needs his call (a baseline bump, a suppression approval, a new `mayDependOn` grant)
Keep competitor names and real people's names out of rule text, fixtures, `why` lines and commit messages.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · r3dbars/transcripted | 100 | MIT |
| Calvin-LLC/skills | 100 | MIT |
| smarzban/agent-skills | 100 | MIT |
| SammyLin/skills | 100 | MIT |
| workerstack/repo | 100 | MIT |
| arjitj2/open-pstack | 100 | MIT |
| haowjy/creative-writing-skills | 100 | Apache-2.0 |
| workerstack/clients | 100 | MIT |
| CMMD-Center/cmmdstack | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 09869a9e060c (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
- 100honor-code
Zamienia powtarzające się błędy agentów w repozytorium w klasy problemów i naprawia każdą na możliwie najwyższym poziomie: architektura, typy, lint, test, a na końcu dokumentacja. - 90mistake-to-constraint
Zamienia powtarzający się błąd w najsilniejszą barierę: typ, test, lint lub regułę zapisaną w CONSTRAINTS.md. - 100agents-correct
Znajduje błędy, które agenci powtarzają w repozytorium, i uniemożliwia każdy z nich kolejno przez architekturę, typy, lint, testy, a na końcu dokumentację. - 100imp-guardrail
Zamienia błąd agenta lub człowieka w zabezpieczenie: wybiera najsilniejsze wymuszenie, na przykład regułę lintera lub test architektury, i dowodzi jego skuteczności. - 100fix-with-checks
Naprawia powtarzający się błąd z przeglądu lub CI wraz z regułą, testem albo typem, które wychwyciłyby go wcześniej. - 86lesson
Zamienia poprawkę lub powtarzający się błąd w zabezpieczenie: regułę, skrypt, test albo wpis w zasadach, jako ostatnią opcję.