# Public Runtime Watchdog

This runbook keeps the public Vignette runtime recoverable after Windows
reboot, update, or process crash. It covers:

- engine gateway: `http://127.0.0.1:9099`
- prod API: `http://127.0.0.1:8001`
- Cloudflare tunnel for `https://api-vignette.chanpaca.net`

## Secret Handling

Do not put secrets in scheduled task arguments.

The scripts use the existing runtime locations:

- API secrets stay in `apps/api/.env`.
- Cloudflared credentials stay under the current user's `.cloudflared` config.
- Claude CLI OAuth/config stays in the current Windows user profile.

The installer creates a per-user interactive scheduled task. It starts at user
logon and repeats as a watchdog. Fully unattended boot before any user logs in
requires an operator-managed service account or Task Scheduler credential; do
that in Windows, not by adding secrets to these scripts.

## Install Or Update

From the repo root:

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\install-public-runtime-task.ps1 -RunNow
```

The installer is idempotent. Re-running it updates the same task:

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\install-public-runtime-task.ps1 -IntervalMinutes 5
```

Do not add future domains to the default watchdog until DNS and Cloudflare
routing are live. For example, `api-vnet.18ka.net` is intentionally excluded
from the default checks while that domain has no DNS. After a public domain is
actually reachable, add it explicitly:

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\install-public-runtime-task.ps1 `
  -AdditionalPublicHealthUrls https://api-vnet.18ka.net/health
```

Task name:

```powershell
VignettePublicRuntimeWatchdog
```

## Manual Start

Use this when you want to force a runtime restore immediately:

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\start-public-runtime.ps1
```

`start-public-runtime.ps1`는 관리자·인증 제어면과 엔진을 분리한다. 이미
`environment=prod`, `db=true`인 API와 정상 웹·터널은 유지하고, 엔진만 실패한 경우
엔진만 복구한다. 엔진이 늦게 준비돼도 관리자·인증 API 기동을 막지 않는다.

API 코드 변경을 운영 프로세스에 반영할 때는 다른 표면을 유지한 채 API만 명시적으로 교체한다.

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\start-public-runtime.ps1 `
  -ForceApiRestart `
  -SkipEngineRestart `
  -SkipWebRestart `
  -SkipCloudflaredRestart
```

## Health Checks

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\watch-public-runtime.ps1 -CheckOnly
Invoke-RestMethod http://127.0.0.1:9099/health
Invoke-RestMethod http://127.0.0.1:8001/health
Invoke-RestMethod https://api-vignette.chanpaca.net/health
```

Optional extra public host check, only after the host resolves:

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\watch-public-runtime.ps1 `
  -CheckOnly `
  -AdditionalPublicHealthUrls https://api-vnet.18ka.net/health
```

Optional real Claude CLI readiness smoke:

```powershell
Invoke-RestMethod http://127.0.0.1:9099/ready
```

`/ready` can consume a small Claude budget because it performs a real generation.

## Logs And Task State

```powershell
Get-ScheduledTask -TaskName VignettePublicRuntimeWatchdog
Get-ScheduledTaskInfo -TaskName VignettePublicRuntimeWatchdog
Get-Content .\public-runtime-watchdog.log -Tail 50
Get-Content .\apps\api\engine.public.err.log -Tail 50
Get-Content .\apps\api\api.public.err.log -Tail 50
Get-Content .\cloudflared.public.err.log -Tail 50
```

## Remove

```powershell
Unregister-ScheduledTask -TaskName VignettePublicRuntimeWatchdog -Confirm:$false
```

## Recovery Notes

If local health is good but public health fails, inspect the cloudflared process
and `C:\Users\<user>\.cloudflared\vignette-config.yml`.

If engine health fails, verify that `claude` runs for the same Windows user that
owns the scheduled task and that the user has completed Claude CLI login.

엔진 장애 중에도 `http://127.0.0.1:8001/health`의 `environment=prod`, `db=true`가
유지되면 관리자·인증 제어면은 정상이다. 이때 watchdog은 API·웹·터널을 재시작하지
않는다. 예약 작업 확인 기준은 `VignettePublicRuntime`의 `LastTaskResult=0`과
`public-runtime-watchdog.failcount=0`이다.

If API health fails with `environment`, `db`, or auth configuration errors,
inspect `apps/api/.env`; do not copy secret values into scripts or task
arguments.