fluent-controlled-run
/fluent-controlled-run
Uruchamia, sprawdza i zatrzymuje pojedyncze przebiegi Fluent po zgodzie użytkownika i walidacji skryptu journal.
- Autor
- AstDen ↗
- Repozytorium
- kady-fluent-skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 12.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Uruchamia, sprawdza i zatrzymuje pojedyncze przebiegi Fluent po zgodzie użytkownika i walidacji skryptu journal.
Oryginalny opis (EN)
Starts, inspects and stops one Fluent journal run only after an explicit consent gate, a case-root confinement check and an offline validation gate. Use when a journal already passed fluent-journal-validation and the user explicitly asks to launch, monitor or stop a run. Not for judging convergence or correctness, and never for authoring a journal.
Jak działa
- When to use / when not to
- Inputs
- Procedure
- Outputs
- Refusals
- Hand-off
Podgląd SKILL.md
---
name: fluent-controlled-run
description: Starts, inspects and stops one Fluent journal run only after an explicit consent gate, a case-root confinement check and an offline validation gate. Use when a journal already passed fluent-journal-validation and the user explicitly asks to launch, monitor or stop a run. Not for judging convergence or correctness, and never for authoring a journal.
license: MIT
compatibility: Node ≥ 22.19 (bundled with Kady). No network. Needs a licensed Fluent executable and allowRun:true in config.json.
metadata:
version: "0.1.0"
pack: fluent-skill-pack
level: L3
---
# Controlled Fluent run
You are the only skill in this pack allowed to start a process, and you start exactly one
thing: the configured Fluent executable with one already-validated journal. You do not author
the journal, you do not judge it, and you do not decide whether the result is converged or
physically correct — a launched process proves only that a process was launched.
**On this machine Fluent is not installed.** On any run request, first read
`<home>/config.json` and check `fluentBin`. If it is absent, missing or not executable, the
honest answer is `fluent-not-found` (exit 4): report that and stop. Never guess a vendor path,
never offer to "just try it", and never imply that a run happened.
## When to use / when not to
- Use: `fluent-journal-validation` returned `clean` or `warn` for `<case-dir>/setup_case.jou`,
and the user explicitly asks "запусти", "launch it", "start the run", "останови расчёт",
"что с расчётом?".
- Use: a run started by this skill must be inspected (`status`) or stopped (`stop`).
- Do not use: the journal is missing, was hand-edited after validation, or its verdict is
`blocked` → `fluent-journal-authoring` (fix) then `fluent-journal-validation` (judge) first.
- Do not use: the user asks "сошлось?", "правильный результат?", "можно публиковать?" →
`fluent-result-qc` reads the transcript; physical correctness is the engineer's call. You
must not answer either question from process state.
- Do not use: there is no licence or no Fluent executable on this machine (see above).
## Inputs
- a case directory inside one of the configured `caseRoots`, containing `setup_case.jou` and,
ideally, a fresh `validation.json`;
- `<home>/config.json` with `fluentBin`, `caseRoots`, `allowRun`, `defaults` and `timeoutSec`;
- explicit consent: `allowRun: true` in that file plus a fresh "yes" from the user for this
specific launch (the `--confirm` flag).
### Config ritual (never skip)
1. Ask the user for two absolute paths: the Fluent executable (`fluentBin`) and the directories
cases may live under (`caseRoots`). Never guess either one from `PATH` or a sibling case.
2. Write a candidate JSON and load it with `node <home>/bin/run.mjs config --set <file.json>`;
the CLI validates types and requires absolute `caseRoots`, then writes `<home>/config.json`.
Read it back with `node <home>/bin/run.mjs config --show` and show the user what was stored.
3. `allowRun` defaults to `false`. Only the user turns it on, as its own explicit decision —
never silently, never bundled into "let me just start it", never because they seem rushed.
4. Keep `defaults` (`dimension`, `threads`, `mpi`, `affinity`) and `timeoutSec` visible in your
summary. The timeout is recorded metadata that `status` can flag; it is not a kill switch.
5. If `fluentBin` fails the executable check, the run is refused with `fluent-not-found`. On
this machine that is the expected outcome; say so instead of trying alternatives.
## Procedure
1. Confirm the gate: `validation.json` says `clean` or `warn`; `blocked` is never launched. A
missing or stale verdict means validate first; `run.mjs` re-validates on every start anyway.
2. Ask for this launch once, plainly, naming the case, the executable and the argv. Then run
`node <home>/bin/run.mjs start --case <case-dir> --name <name> --confirm` and read the JSON:
`run.pid`, `run.argv`, `run.validation{verdict,score,journalSha256}`, `run.backup`.
3. The executable receives exactly this argv (documented once, never guessed):
```
<fluentBin> 3ddp -t4 -g -i <case-dir>/setup_case.jou
```
`dimension` and `threads` come from `config.defaults`; `mpi` and `affinity` are recorded in
the run record but contribute no flag in v0.1. The process is launched detached, with its
own process group, `shell: false`, cwd = case dir, and stdout+stderr appended to
`<case-dir>/run.log`; the previous `run.log` and record are moved to
`<case-dir>/archive/pre-<name>-<timestamp>/` before a replacement starts.
4. Poll `node <home>/bin/run.mjs status --case <case-dir> --name <name>` every 30–60 seconds —
never in a tight loop. Report `alive`, `run.status`, `logTail.lastIteration`, the
convergence banner only as "the log says …", and any `logTail.errors` verbatim.
5. Stop with `node <home>/bin/run.mjs stop --case <case-dir> --name <name>`: SIGTERM to the
group, up to 5 s, then SIGKILL; the record becomes `stopped`. Never delete `<case-dir>/archive/`.
6. Report the record and log paths, what the transcript says and what is still unknown — never "converged".
## Outputs
- `start` → `{ok, run{name,caseDir,journal,argv,pid,startedAt,status,logPath,recordPath,
validation{verdict,score,journalSha256},backup}}`, and the record at `<case-dir>/.runs/<name>.json`.
- `status` → `{ok, run, alive, logTail{lines,lastIteration,converged,errors}}`; if the pid is
gone and the record says `running`, a derived `status` (done banner → `finished`, error →
`failed`, else `unknown`) and `finishedAt` are persisted; `converged` is log-quoted only.
- `stop` → `{ok, stopped, signal}`; `stopped: false` means there was nothing left to kill.
- `validation.json` is written by the gate on every start (`validate.mjs`); `run.log` only grows.
- Exit codes: 0 ok, 1 runtime, 2 usage, 3 data/contract, 4 refused; refusals carry `refused` and `ok:false`.
A run does not prove convergence (that is `fluent-result-qc` reading the transcript), does not
prove physical correctness (that is the engineer), and does not prove the journal was right.
It proves only: the binary started, it wrote this log, and here is the process we can still stop.
## Refusals
- No config → «Сначала нужен `config.json`: назовите абсолютные пути к Fluent и к каталогу
кейсов, я ничего не угадываю.»
- `allowRun` off → «Запуск выключен в конфиге. Включить `allowRun` можете только вы; я не
включаю его молча.»
- No `--confirm` → «Без явного подтверждения на этот запуск я Fluent не запускаю.»
- Blocked journal → «Вердикт `blocked`: журнал нельзя запускать. Сначала исправьте его в
авторском навыке.»
- Outside roots / symlink escape → «Кейс лежит вне разрешённых `caseRoots`; запуск отклонён.»
- No licence or binary → «Fluent не найден или не исполняемый: на этой машине он не
установлен. Прогон невозможен — это `fluent-not-found`, а не повод пробовать наугад.»
- Already running → «Под этим именем уже работает процесс; сначала `stop`, потом новый старт.»
- Request to judge convergence → «Сходимость оценивает `fluent-result-qc` по логу, не запуск.»
Never answer "запустилось?" by starting something; never edit the journal to make a launch pass.
## Hand-off
Hand the transcript to `fluent-result-qc` (`<case-dir>/run.log`, plus `qc --write` only with the
user's permission) and the record to `fluent-case-report`. Return to `fluent-case-builder` (the
router) with the run record path and one sentence of what the log shows. If the gate failed,
hand back to `fluent-journal-authoring` with the exact blockers instead of retrying the launch.
Wersje
W bazie nie ma innych skilli o nazwie fluent-controlled-run.
Migawka bazy z 13.11.2026 · wersja paczki 0904b0b5a373 (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
- 90fluent-journal-validation
Ocenia dziennik Fluent offline względem indeksu poleceń, sprawdza kolejność i zakończenia wierszy i wydaje werdykt: czysty, ostrzeżenie albo zablokowany. - 90fluent-case-report
Składa istniejące materiały przypadku CFD w raport z sumami sha256 i listą otwartych braków, gdy potrzebne jest czytelne podsumowanie dla człowieka. - 96fluent-cli
Pomaga budować aplikacje i pobierać dane w ServiceNow przez Fluent (now-sdk). - 95ansys-fluent-mcp
Steruje sesją solvera ANSYS Fluent przez serwer MCP: wczytuje przypadek i siatkę, zmienia ustawienia, uruchamia iteracje i odczytuje wyniki. - 100fluent-ci-cd
Uruchamia ServiceNow SDK (now-sdk) bez interfejsu w CI/CD, na przykład w GitHub Actions, i pomaga przy błędzie keys.ts out of date. - 90run-fluent-autoclave
Prowadzi, diagnozuje i waliduje symulacje CFD w ANSYS Fluent dla autoklawów z konwekcją wymuszoną, geometrią STEP i kontrolą bilansu masy.