15 Commits
Author SHA1 Message Date
homelab-updaterandClaude Opus 5.5 765a1e0eee docs: TASK_STATE after deploying haiku
ci / check (push) Successful in 1m36s
ci / publish (push) Successful in 1m27s
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-08 00:19:51 +02:00
homelab-updaterandClaude Opus 5.5 b29aec3dee docs: TASK_STATE after adding haiku
ci / check (push) Successful in 1m39s
ci / publish (push) Successful in 1m43s
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-07 23:57:25 +02:00
homelab-updaterandClaude Opus 5.5 964f2ad815 feat(models): add haiku as a model choice and make it the default
Haiku 5.5 is cheaper and close enough to Sonnet for the release notes
search and assessment, the repair, the migration plan and the
post-upgrade check, so an app that names no model now gets haiku.

- ModelName (haiku|sonnet|opus) is the one vocabulary; modelOr picks the
  default where the old code fell back to sonnet or opus.
- A default "haiku" profile is seeded next to sonnet and opus.
- Migration 26 rebuilds app_config so its CHECKs accept haiku and its
  release notes default is haiku; existing rows are copied verbatim.
- claude CLI 2.1.293: the first release that knows claude-haiku-5-5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-07 23:57:25 +02:00
homelab-updaterandClaude Opus 5.5 4b9fa93ed0 docs: TASK_STATE after the post-upgrade check change
ci / check (push) Successful in 1m32s
ci / publish (push) Successful in 1m28s
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-07 11:59:38 +02:00
homelab-updaterandClaude Opus 5.5 c17b170739 feat(post-upgrade): run the model only when the update may ask for something
ci / check (push) Successful in 1m33s
ci / publish (push) Successful in 1m35s
The post-upgrade model ran after every update: 17 of 18 checks found nothing,
mostly LinuxServer rebuilds of the same app version. And "off" did not turn it
off: the workflow port read a null executor as sonnet.

- A null postUpgradeExecutor is off again: no session, no model.
- New per-app postUpgradeWhen (migration 25): when_needed (default) or always.
- when_needed runs the model for an update the owner approved (release notes
  read), a major one or a migration; otherwise it reads the service's last 200
  log lines in an observe session and calls the model only on errors there.
- ExecutorClient.sessionLogs reads /agent/logs with the session token, like
  sessionCompose.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-07 11:54:48 +02:00
homelab-updaterandClaude Opus 5.5 bcd144fbdc fix(runs): count a model run once when the terminal node carries its result on
ci / check (push) Successful in 1m35s
ci / publish (push) Successful in 1m33s
A terminal node's result is the record it received, so after post_upgrade the run
listed the same post-upgrade check twice and counted its cost twice (run view, run
list and model costs). Identical calls now count once.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-07 11:31:30 +02:00
homelab-updaterandClaude Opus 5.5 d05b6c27f3 docs: TASK_STATE after the registry rate limit fix
ci / check (push) Successful in 1m34s
ci / publish (push) Successful in 1m25s
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-07 00:34:18 +02:00
homelab-updaterandClaude Opus 5.5 e060a59b67 fix(registry): wait out a ghcr.io rate limit instead of reporting the version unknown
ci / check (push) Successful in 1m38s
ci / publish (push) Successful in 1m25s
ghcr.io answered 429 to the tag list for jackett, radarr and sonarr, and image_tags
turned that straight into version_unknown with only "skip" left. The registry client
now asks again after 429, honouring Retry-After up to five minutes; the driver waits
30 s, 2 min and 5 min, the server does not wait.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-07 00:28:08 +02:00
homelab-updaterandClaude Opus 5.5 8966fec9cf docs: TASK_STATE after the migrate and verify fixes
ci / check (push) Successful in 1m35s
ci / publish (push) Successful in 1m20s
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-03 23:55:43 +02:00
homelab-updaterandClaude Opus 5.5 1b8bfc4446 fix(verify): wait out the old container while the service is recreated
ci / check (push) Successful in 2m31s
ci / publish (push) Successful in 3m9s
For a Git project Arcane recreates the service while verify polls, and
for a moment compose has the old container next to the new one. The
executor refuses to inspect that (multiple_containers), and verify
failed at once, so a healthy update went to model.repair (deluge,
homeassistant, nextcloud, nextcloud-redis). That refusal now counts as
an ordinary failed attempt within the usual 24 x 5 s; any other refusal
still fails at once.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-03 23:31:33 +02:00
homelab-updaterandClaude Opus 5.5 d00c8ad4aa fix(migrate): run an approved migration that keeps the compose file
ci / check (push) Successful in 1m35s
ci / publish (push) Successful in 1m29s
A migration plan may leave composeYaml null ("only the image changes"),
but the docker.migrate request required a compose file, so building the
request failed (invalid_privileged_action_request) right after the
backup. Gitea 1.27.3 -> 28.0.0 stopped there. The request now carries
null and the executor keeps the current compose.yaml (read at apply
time, matching the backup) and only pins the new image. The plan's
earlier snapshot is not used: another update of the same project may
have changed the file since.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-02 21:51:59 +02:00
homelab-updaterandClaude Opus 5.5 2e3d61982f docs: TASK_STATE and phase H status after deploying the login
ci / check (push) Successful in 1m34s
ci / publish (push) Successful in 1m22s
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-30 14:06:22 +02:00
homelab-updaterandClaude Opus 5.5 6cf584894d fix(deploy): start the auth service, and roll back a start that fails part-way
ci / check (push) Successful in 1m36s
ci / publish (push) Successful in 1m22s
Compose writes an inline config only into a writable container, so the
auth service is no longer read_only; it trusts X-Forwarded-* only from
the Docker networks. The first deploy of the login stopped at compose up
and update.sh exited there, leaving the app without its auth backend
and the driver not started: a failed up now goes back to the previous
compose file and image like a start that does not come up healthy.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-30 14:02:01 +02:00
homelab-updaterandClaude Opus 5.5 7131ca9b12 fix(deploy): keep VM103's disk from filling with images and build cache
ci / check (push) Successful in 1m33s
ci / publish (push) Successful in 1m25s
Every CI build and deploy left a 1.2 GB image of this app on the VM and
the build cache kept growing, until the root disk was full and apt in
CI failed. After a good deploy update.sh keeps the running image, the
previous one and the three newest; the publish job prunes build cache
older than two days.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-30 13:44:30 +02:00
homelab-updaterandClaude Opus 5.5 4b6b179525 docs: Trilium connected to Arcane Git Sync
ci / check (push) Failing after 6s
ci / publish (push) Skipped
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-30 12:59:56 +02:00
57 changed files with 591 additions and 128 deletions

No files matched your search

+1 -1
View File
@@ -5,7 +5,7 @@ ARG NODE_IMAGE=node:24-slim@sha256:2fe369e969550cde8e867afc3fe370b260140cab4a23d
ARG DOCKER_CLI_IMAGE=docker:28.5.2-cli@sha256:625d9431a9f54c5a2bc90f24f0e1c3d55b1349fd857dd85035f98c2c9acbdd4d
# The claude CLI for the model sidecar, pinned and updated deliberately (docs/DESIGN.md, "Wykonawcy modeli").
ARG CLAUDE_CODE_VERSION=2.1.267
ARG CLAUDE_CODE_VERSION=2.1.293
# The codex CLI for the model sidecar, pinned and updated deliberately (docs/DESIGN.md, "Wykonawcy modeli").
ARG CODEX_VERSION=0.156.1
+17 -13
View File
@@ -1,32 +1,36 @@
# Stan zadania — nowoczesny frontend (Faza G)
# Stan zadania — faza H (wygoda na co dzień)
## Cel
Bardziej nowoczesny i graficzny frontend: okna/karty, wykresy z danych, które API już zwraca. Plan i status: `docs/IMPLEMENTATION.md`, „Faza G”.
Aplikacja wygodna na co dzień: stan „tu i teraz”, historia wersji, cisza nocna i poranny raport, decyzja z telefonu, szybka nawigacja, logowanie Keycloak z dostępem spoza LAN. Plan i status: `docs/IMPLEMENTATION.md`, „Faza H”.
## Decyzje
- Sidebar + karty, ciemny motyw domyślny z przełącznikiem (`src/web/theme.ts`).
- ECharts (`src/web/charts/echarts.ts`, osobny chunk), opcje wykresów jako czyste funkcje w `src/web/charts/options.ts`, agregacje w `src/web/stats.ts`; kolory z tokenów CSS, kategorie `--series-1..8`.
- Okna: `src/web/components/dialogs.tsx` (`useConfirm`, `useToast`, `Drawer`); `window.confirm` nie jest już używany.
- Bez zmian w backendzie; lista runów pobierana w całości — przy dużej liczbie runów rozważyć endpoint agregatów.
- Logowanie: oauth2-proxy (usługa `auth`) jako forwardAuth Traefika, Keycloak realm `home`, klient `homelab-upgrade` (założony przez właściciela), wpuszczany tylko `AUTH_ALLOWED_EMAILS`; sekrety w `.env` na VM103 (root 600). Publiczne tylko `manifest.webmanifest` i `icon.svg`.
- Cisza nocna 01:00–07:30, raport o 7:30 (strefa harmonogramów); ustawienia w UI (`#/settings`).
- Wdrożenia tylko przyciskiem `deploy` (Actions albo MCP `dispatch_workflow`), nigdy przez Arcane.
## Zrobione
- G0–G7 scalone do `master` (`aec96b2`, fast-forward z `frontend-refresh`) i wdrożone 2026-09-30 na VM103: `./update.sh aec96b2…`, wszystkie usługi na obrazie CI (build 174), bez obrazów hotfix; `compose.yaml` na VM = `deploy/compose.yaml`. Kopie przed wdrożeniem: `/srv/appdata/homelab-upgrade/data/queue.sqlite.before-phase-g-*`, `/srv/appdata/homelab-upgrade/compose.yaml.before-phase-g-*`.
- Wdrożenie przyciskiem: workflow `deploy` (Actions → deploy → Run workflow z `master`, albo `dispatch_workflow` przez MCP) uruchamia `update.sh` z wdrażanego obrazu (`--install`: kopia bazy, `compose.yaml` z obrazu, powrót do poprzedniego obrazu i compose). Tagi `v*` budują obrazy wydań. Sprawdzone na VM103: run 206 wdrożył `136f9d2` (build 177), run 207 odrzucił niepoprawną wersję przed jakąkolwiek zmianą. Opis: `docs/DEPLOY.md`, „Aktualizacja”.
- Gitea (`homelab-compose`, `gitea/compose.yaml`): router `gitea-web-lan` bez limitu zapytań dla LAN, VPN i sieci Dockera (`ClientIP`, priorytet 200); limit Traefika zostaje dla internetu. Powód: CI dostawało 429 przy `docker login`.
- Graf Cytoscape nie ponawia nieudanego startu canvas przy każdym renderze (wcześniej czyścił listę zastępczą).
- H1–H6 na `master` i na VM103: migawka hosta, historia wersji, raport i cisza nocna, decyzje z telefonu, PWA, `Ctrl+K`, logowanie Keycloak.
- Poprawki po pierwszych nocach: migracja z `composeYaml: null` (executor bierze bieżący compose i przypina obraz; `d00c8ad`, Gitea 28.0.0 zaktualizowana), `verify` czeka na zniknięcie starego kontenera przy przebudowie przez Arcane zamiast od razu wołać `model.repair` (`1b8bfc4`).
- `version_unknown` u jackett/radarr/sonarr (5–6.10) to był HTTP 429 z ghcr.io przy liście tagów, nie wersja: klient rejestru ponawia po 429 (driver 30 s / 2 min / 5 min, `Retry-After` do 5 min; serwer bez czekania), wdrożone `e060a59`.
- Podwójny wpis modelu w runie: węzeł końcowy powtarzał wynik `post_upgrade`; liczony raz (`bcd144f`).
- Krok modelu po aktualizacji (`c17b170`): `postUpgradeExecutor: null` znów wyłącza (port workflow czytał null jako sonnet), nowe `postUpgradeWhen` (`when_needed` domyślnie / `always`, migracja 25). `when_needed`: model tylko przy aktualizacji zatwierdzonej przez właściciela, major, migracji albo błędach w ostatnich 200 liniach logów (sesja observe, bez modelu). Na VM103 wyłączone dla 11 aplikacji (arr, jackett, bazarr, jellyfin, ombi, deluge, orcaslicer, homepage, trilium, nextcloud-redis), reszta sonnet + `when_needed`.
- Haiku 5.5 jako trzeci model (`haiku`): domyślny dla notatek wydania (szukanie + ocena), naprawy, planu migracji i sprawdzenia po aktualizacji, gdy aplikacja nie wybrała modelu; profil `haiku` w seedzie, migracja 26 (przebudowa `app_config` dla CHECK), CLI w obrazie 2.1.293 (pierwsza wersja znająca `claude-haiku-5-5`). Wdrożone `b29aec3` (8.10); na VM103 wszystkie 22 aplikacje przełączone z `sonnet` na `haiku` (wyłączone zostały wyłączone).
- Homelab-compose: whitelist deluge przepuszcza bramę sieci `proxy` (`172.16.0.1/32`), inaczej `verify` dostawał 403.
- Proxmox: pełny dysk 30.09 zawiesił pmxcfs (brak vzdump 1–2.10); po restarcie usług PVE kopie działają. VM103: przycięte obrazy (dysk 89% → 64%).
## Pozostałe
- Przegląd na żywo w przeglądarce (oba motywy, 375 px i desktop), poprawki wyglądu po przeglądzie.
- Właściciel loguje się raz w przeglądarce i na telefonie (potwierdzenie Keycloak → aplikacja).
- Ze strumienia Git/Arcane: migracja klientów Influx przed włączeniem aktualizacji.
- Do rozważenia: uproszczenie przebiegów (ustawienia per aplikacja zamiast edycji grafu, deklaratywny warunek zamiast funkcji TS), wyłączenie godzinnego sprawdzania aktualizacji obrazów w Arcane (dubluje nasz system), bramka „verify failed” bez opcji „accept”, sprzątanie `# saved as ./compose.yaml` z planów migracji, automatyczny `pg_dump` przed migracją bazy.
## Następny krok
Właściciel ogląda nowe UI na https://upgrade.mariuszpacek.pl i zgłasza poprawki.
W pierwszym runie z modelem sprawdzić, że alias `haiku` w CLI 2.1.293 to Haiku 5.5 (`modelUsage` w odpowiedzi `claude`; próba 8.10 trafiła na limit sesji konta) i czy jakość ocen wystarcza. Dalej obserwować 429 i `when_needed`.
## Walidacja
`npm run check` zielone: 107 plików, 1260+ testów. Testy agregacji i opcji wykresów: `npx vitest run src/web/stats.test.ts src/web/charts`.
`npm run check`: 1320 testów PASS. `deploy/update.sh` sprawdzany na atrapie `docker` (instalacja, kopia bazy, nieudany start i nieudane `up`, przycinanie obrazów).
+3 -1
View File
@@ -63,7 +63,7 @@ services:
image: quay.io/oauth2-proxy/oauth2-proxy:v7.15.4@sha256:b1b2021fe8f4004573e8d690dec6c7bb29cc44364572cf8510a05bf3a0ae2ded
restart: unless-stopped
user: "65532:65532"
read_only: true
# Not read_only: compose writes an inline config (auth-emails) only into a writable container filesystem.
cap_drop: [ALL]
security_opt: ["no-new-privileges:true"]
environment:
@@ -76,6 +76,8 @@ services:
OAUTH2_PROXY_REDIRECT_URL: https://upgrade.mariuszpacek.pl/oauth2/callback
OAUTH2_PROXY_HTTP_ADDRESS: 0.0.0.0:4180
OAUTH2_PROXY_REVERSE_PROXY: "true"
# Only Traefik (on the Docker proxy network) may say who the client is through X-Forwarded-*.
OAUTH2_PROXY_TRUSTED_PROXY_IPS: 172.16.0.0/12
OAUTH2_PROXY_UPSTREAMS: static://202
OAUTH2_PROXY_EMAIL_DOMAINS: "*"
OAUTH2_PROXY_AUTHENTICATED_EMAILS_FILE: /etc/oauth2-proxy/emails
+3 -3
View File
@@ -156,9 +156,9 @@ if ! TAG="$tag" docker compose pull; then
restore_compose
exit 1
fi
TAG="$tag" docker compose up -d --remove-orphans
if wait_until_up; then
# A start that fails part-way (compose refuses a service, a container does not start) goes back like one that does
# not come up healthy.
if TAG="$tag" docker compose up -d --remove-orphans && wait_until_up; then
# Only the TAG line changes: .env also holds HA_TOKEN and other settings.
if [[ -n "$tag_arg" ]]; then
touch .env
+4 -4
View File
@@ -98,7 +98,7 @@ Notatki wydań nie są ustawieniem: są w **skillu** projektu (Apps → aplikacj
Jedno ustawienie w Apps → aplikacja, domyślnie wyłączone:
- **Post-upgrade check executor**: `sonnet` / `opus` = po udanym `verify` model sprawdza, czy aktualizacja wymaga czynności poinstalacyjnych (czyta ocenę notatek, skill i stan usługi; co robić po aktualizacji, można dopisać w skillu). Model wykona tylko to, co dopuszcza `config/agent-exec.json` (dokładne komendy, bez wildcardów); reszta trafia do właściciela jako STOP `migration_required`. Dziś lista ma tylko komendy Nextclouda: kontrola odbiorcza ze skilla (tylko odczyt: `php occ status` z `--output=json` i bez, `php occ config:app:get core lastcron`, `php occ app:list`, `pidof crond` w usłudze `cron`) oraz to, co po upgrade brakuje w bazie (`php occ db:add-missing-indices`, `db:add-missing-columns`, `db:add-missing-primary-keys`; użytkownik `www-data`, usługa `app`). Model widzi listę przez `hl exec-rules` i nie zgaduje. Dla pozostałych aplikacji ten krok tylko wypisze, co zrobić ręcznie.
- **Post-upgrade check executor**: `haiku` (domyślny) / `sonnet` / `opus` = po udanym `verify` model sprawdza, czy aktualizacja wymaga czynności poinstalacyjnych (czyta ocenę notatek, skill i stan usługi; co robić po aktualizacji, można dopisać w skillu). Model wykona tylko to, co dopuszcza `config/agent-exec.json` (dokładne komendy, bez wildcardów); reszta trafia do właściciela jako STOP `migration_required`. Dziś lista ma tylko komendy Nextclouda: kontrola odbiorcza ze skilla (tylko odczyt: `php occ status` z `--output=json` i bez, `php occ config:app:get core lastcron`, `php occ app:list`, `pidof crond` w usłudze `cron`) oraz to, co po upgrade brakuje w bazie (`php occ db:add-missing-indices`, `db:add-missing-columns`, `db:add-missing-primary-keys`; użytkownik `www-data`, usługa `app`). Model widzi listę przez `hl exec-rules` i nie zgaduje. Dla pozostałych aplikacji ten krok tylko wypisze, co zrobić ręcznie.
Schemat bazy przechodzi na v10 (v8 dodał `post_upgrade_executor`, v9 zdejmuje niepotrzebną `release_notes_source`, v10 dodaje niepodłączony jeszcze katalog i instancje nowego silnika workflow); starszy obraz jej nie otworzy, więc przed wdrożeniem kopia `queue.sqlite*`.
@@ -191,11 +191,11 @@ Ręczny powrót do dowolnej kopii: STOP w kolejce → rollback, albo z kroku z `
## Arcane i repo compose
Repozytorium `ssh://git@git.mariuszpacek.pl:2244/mariusz.pacek/homelab-compose.git` jest prywatne, na `master`, zawiera 19 katalogów projektów z samym `compose.yaml`, archiwum `to_delete/dockge` oraz `.gitignore`; pliki `.env` i sekrety nie są śledzone. Główny checkout VM103: `/srv/appdata/homelab-upgrade/git/repo`. Dziesięć compose używających `env_file` w formacie `raw` skonwertowano przez `tools/compose-env-compat.py`; pełny wynik `docker compose config` porównano ściśle, wartości lokalne pozostały bez zmian.
Repozytorium `ssh://git@git.mariuszpacek.pl:2244/mariusz.pacek/homelab-compose.git` jest prywatne, na `master`, zawiera 20 katalogów projektów z samym `compose.yaml`, archiwum `to_delete/dockge` oraz `.gitignore`; pliki `.env` i sekrety nie są śledzone. Główny checkout VM103: `/srv/appdata/homelab-upgrade/git/repo`. Dziesięć compose używających `env_file` w formacie `raw` skonwertowano przez `tools/compose-env-compat.py`; pełny wynik `docker compose config` porównano ściśle, wartości lokalne pozostały bez zmian.
Arcane v2.13.1 działa pod `https://arcane.mariuszpacek.pl`. Projekty używają Git Sync w trybie **Deploy** (w UI: Pull), ścieżka `<projekt>/compose.yaml`, `syncDirectory=false`, interwał 5 minut. Autosync jest włączony dla wszystkich 18 projektów; stosy Arcane i `homelab-upgrade` są wyłączone, a `upgrade-lab` działa bezpośrednio. Wszystkie manifesty projektów zawierają etykietę `com.getarcaneapp.arcane.updater: "false"`. Arcane sam wykonuje redeploy projektu po zmianie compose, więc executor w trybie git nie uruchamia `pull/up`.
Arcane v2.13.1 działa pod `https://arcane.mariuszpacek.pl`. Projekty używają Git Sync w trybie **Deploy** (w UI: Pull), ścieżka `<projekt>/compose.yaml`, `syncDirectory=false`, interwał 5 minut. Autosync jest włączony dla wszystkich 19 projektów; stosy Arcane i `homelab-upgrade` są wyłączone, a `upgrade-lab` działa bezpośrednio. Wszystkie manifesty projektów zawierają etykietę `com.getarcaneapp.arcane.updater: "false"`. Arcane sam wykonuje redeploy projektu po zmianie compose, więc executor w trybie git nie uruchamia `pull/up`.
Początkowe synchronizacje 18 projektów zachowały SHA plików env i nie zmieniły startów kontenerów; Arcane może utworzyć pusty `.env`. Źródła nie wymagały restartu podczas importu. Test Jackett przez API executora potwierdził ścieżkę komentarz → commit `47c2b7f` → push → udany sync Arcane i HTTP 301; rollback executora wykonał `git revert`, przywrócił pierwotny SHA i zakończył się powodzeniem.
Początkowe synchronizacje 18 projektów (i Trilium, dołączonego 2026-09-30) zachowały SHA plików env i nie zmieniły startów kontenerów; Arcane może utworzyć pusty `.env`. Źródła nie wymagały restartu podczas importu. Test Jackett przez API executora potwierdził ścieżkę komentarz → commit `47c2b7f` → push → udany sync Arcane i HTTP 301; rollback executora wykonał `git revert`, przywrócił pierwotny SHA i zakończył się powodzeniem.
`config/workflow-schedules.json` rozkłada starty w strefie `Europe/Warsaw` (czas polski): poniedziałek Bazarr 19:00, Jackett 21:00; wtorek Radarr 19:00, Sonarr 21:00; środa Deluge 19:00, Ombi 21:00; czwartek Homepage 19:00, OrcaSlicer 21:00; piątek Gitea 18:00, baza 19:00, Keycloak 20:00, baza 22:00; sobota Home Assistant 18:00, Nextcloud 20:00, baza 22:00, Redis 23:00; niedziela Trilium 16:00, Grafana 18:00, Jellyfin 20:00. InfluxDB i Dockge są wyłączone.
+1 -1
View File
@@ -95,7 +95,7 @@ Projekt compose może mieć kilka usług z własnymi obrazami (grafana + influxd
## Model naprawia (decyzja właściciela, 2026-09-19 wieczór)
„Skrypt robi upgrade, dopóki wszystko przechodzi; kiedy STOP albo kontener nie wstał, wchodzi model, patrzy i naprawia.”
- Krok `repair`: gdy `verify` nie przeszedł, a aplikacja ma „Migration executor” (sonnet/opus), `verify` sam kolejkuje `repair` (jedna próba; drugi nieudany `verify` = STOP dla właściciela). Na dowolnym STOP właściciel może wybrać „Repair (model)”.
- Krok `repair`: gdy `verify` nie przeszedł, a aplikacja ma „Migration executor” (haiku/sonnet/opus; bez wyboru haiku), `verify` sam kolejkuje `repair` (jedna próba; drugi nieudany `verify` = STOP dla właściciela). Na dowolnym STOP właściciel może wybrać „Repair (model)”.
- Kontener `model`: niezmodyfikowany `claude` CLI (przypięta wersja w obrazie) zalogowany kontem właściciela (`CLAUDE_CONFIG_DIR=/claude`). Bez `docker.sock`, bez plików projektów i bazy.
- Model działa tylko przez narzędzie `hl` w sesji executora otwartej na jeden projekt: `ps`, `logs`, `compose` (sekrety jako `<secret:KLUCZ>`), `write`, `pull`, `up`, `restart`, `exec-rules` (lista komend, które `exec` w ogóle przyjmie: model patrzy, zamiast zgadywać), `exec` w kontenerach tego projektu. Egzekwowane w kodzie (`domain/agent-guard.ts`): zapis compose sprawdzany listą zakazów z „Granic” względem manifestu z otwarcia sesji; wyjścia maskowane wartościami sekretów projektu; hook `PreToolUse` przepuszcza w Bash tylko `hl`, a Read/Edit/Write tylko w katalogu sesji; trasy drivera w executorze i `/run` modelu odpowiadają tylko adresowi drivera. Każde wywołanie jest w logu sesji i w wyniku kroku.
- Wyjątek za zgodą właściciela (2026-09-23): na bramce po nieudanej naprawie przycisk „Repair with full access” uruchamia jedną ponowną naprawę w sesji `full`. W niej, tylko w tym projekcie, `exec` przyjmuje dowolną komendę i użytkownika, `run` uruchamia jednorazowy kontener usługi z jej wolumenami (np. `chown` katalogu danych przy crash-loopie), a `write` sprawdza tylko poprawność compose, bez listy zakazów. Maskowanie sekretów, log sesji, kopia i rollback działają dalej. To jedna próba na run: po niej zostaje rollback albo skip.
+2
View File
@@ -180,3 +180,5 @@ Decyzje: logowanie Keycloak dla wszystkich (oauth2-proxy jako forwardAuth Traefi
- **H4** decyzja z telefonu: strona runu z bramką na górze i przyciskami przyklejonymi do dołu ekranu; push otwiera ją bezpośrednio.
- **H5** czasy względne, szybkie wyszukiwanie `Ctrl+K`, dry run wszystkich aplikacji z dostępną aktualizacją.
- **H6** logowanie Keycloak i zdjęcie ograniczenia do LAN.
Status 2026-09-30: H1–H6 wdrożone na VM103 (ostatnio `6cf5848`). Przy pierwszym wdrożeniu H6 compose odrzucił inline config w usłudze `read_only`, a `update.sh` przerwał się bez wycofania (ok. 12 min awarii, wycofane przyciskiem do `fc68d30`); poprawione: `auth` bez `read_only`, nieudane `up -d` wraca do poprzedniego obrazu i compose. Tego samego dnia dysk VM103 zapełnił się obrazami aplikacji i cache buildów CI: `update.sh` zostawia teraz uruchomiony, poprzedni i 3 najnowsze obrazy, CI czyści cache starszy niż 2 dni.
+35 -3
View File
@@ -89,6 +89,38 @@ describe('component-operation-executor: verify', () => {
expect(calls).toBe(4);
});
it('retries while compose still has the old container next to the new one, then passes', async () => {
const clock = new FakeClock(NOW);
const executor = new FakeExecutorClient();
let calls = 0;
executor.on('inspect', () => {
calls += 1;
return calls <= 2
? { outcome: 'refused', reason: 'multiple_containers', detail: 'service "jackett" of "jackett" has 2 containers' }
: { outcome: 'done', result: inspectResult({ state: 'running', health: 'healthy', image: PINNED }) };
});
const app = createComponentOperationExecutor({ clock, executor, backupVmid: 103 });
const outcome = await settle(app.execute('verify', 1, verifyInput()), clock);
expect(outcome).toMatchObject({ event: 'healthy', result: { healthy: true, attempts: 3 } });
});
it('gives up on a service that keeps two containers after the usual attempts, and at once on any other refusal', async () => {
const clock = new FakeClock(NOW);
const executor = new FakeExecutorClient();
executor.on('inspect', { outcome: 'refused', reason: 'multiple_containers', detail: 'service "jackett" of "jackett" has 2 containers' });
const app = createComponentOperationExecutor({ clock, executor, backupVmid: 103 });
const outcome = await settle(app.execute('verify', 1, verifyInput()), clock);
expect(outcome).toMatchObject({ event: 'unhealthy', result: { healthy: false, reason: 'inspect_unavailable', attempts: 24 } });
const other = new FakeExecutorClient();
other.on('inspect', { outcome: 'refused', reason: 'unknown_service', detail: 'no service' });
const once = await settle(createComponentOperationExecutor({ clock, executor: other, backupVmid: 103 }).execute('verify', 1, verifyInput()), clock);
expect(once).toMatchObject({ event: 'unhealthy', result: { healthy: false, reason: 'inspect_unavailable', detail: 'no service' } });
expect(other.calls).toHaveLength(1);
});
it('gives up after 24 unhealthy attempts, 5s apart, once the container is no longer starting', async () => {
const clock = new FakeClock(NOW);
const executor = new FakeExecutorClient();
@@ -249,17 +281,17 @@ describe('component operation executor: migration_plan', () => {
.on('sessionCompose', { outcome: 'done', result: { text: 'services:\n influxdb:\n image: influxdb:1.8\n' } })
.on('closeSession', { outcome: 'done', result: { project: 'grafana', service: 'influxdb', manifest_changed: false, manifest_sha256: 'a'.repeat(64), log: [{ at: NOW, op: 'compose', ok: true, detail: 'read' }] } });
it('has opus look at the app in an observe session and hands over the plan with its digest', async () => {
it('has haiku look at the app in an observe session and hands over the plan with its digest', async () => {
const model = new FakeModelExecutor().reply({ output: plan, costUsd: 0.4 });
const executor = observer();
const done = await createComponentOperationExecutor({ backupVmid: 103, clock: new FakeClock(NOW), skills, model, executor }).execute('migration_plan', 1, context);
expect(done).toMatchObject({
event: 'migration_needed',
result: { plan: { needed: true, inputs: [{ name: 'ADMIN_PASSWORD' }] }, model: 'opus', costUsd: 0.4, actions: ['compose: read'], composeBefore: 'services:\n influxdb:\n image: influxdb:1.8\n' },
result: { plan: { needed: true, inputs: [{ name: 'ADMIN_PASSWORD' }] }, model: 'haiku', costUsd: 0.4, actions: ['compose: read'], composeBefore: 'services:\n influxdb:\n image: influxdb:1.8\n' },
});
expect((done.result as { planDigest: string }).planDigest).toMatch(/^[0-9a-f]{64}$/);
expect(model.requests[0]).toMatchObject({ executor: 'opus', profile: 'observe', sessionToken: 'observe-token' });
expect(model.requests[0]).toMatchObject({ executor: 'haiku', profile: 'observe', sessionToken: 'observe-token' });
expect(model.requests[0]?.prompt).toContain('Before update: create an admin user');
expect(executor.calls.map((call) => call.op)).toEqual(['openSession', 'sessionCompose', 'closeSession']);
expect(executor.calls[0]?.request).toEqual({ project: 'grafana', service: 'influxdb', mode: 'observe' });
+10 -5
View File
@@ -4,7 +4,7 @@ import { resolveImageRef } from './ops/image-resolve.ts';
import { backupVerdict, HostBackupRecord } from '../domain/backup-check.ts';
import { isWithinAuto } from '../domain/version.ts';
import { describeError } from '../domain/mask.ts';
import { nextVerifyAction, VERIFY_INTERVAL_MS, verifyChecks } from '../domain/verify.ts';
import { nextVerifyAction, VERIFY_ATTEMPTS, VERIFY_INTERVAL_MS, verifyChecks } from '../domain/verify.ts';
import type { VerifyObservation } from '../domain/verify.ts';
import type { Clock, ExecutorClient, ModelExecutor, PrivilegedActionRun, RegistryClient, SkillStore } from '../domain/ports.ts';
import type { SessionLogEntry } from '../domain/executor-api.ts';
@@ -15,7 +15,7 @@ import { buildMigrationPlanPrompt, MIGRATION_PLAN_SCHEMA, MigrationPlan, migrati
import { ModelError } from '../domain/ports.ts';
import type { GroupMember } from './ops/group.ts';
import type { ComponentOperationId } from '../domain/workflow-component-operation.ts';
import { toJson } from '../domain/steps.ts';
import { modelOr, toJson } from '../domain/steps.ts';
import type { JsonValue } from '../domain/steps.ts';
import { validateImage } from '../domain/compose-edit.ts';
import { splitImage } from '../domain/image-ref.ts';
@@ -229,7 +229,12 @@ export function createComponentOperationExecutor(deps: ComponentOperationExecuto
for (let attempt = 1; ; ) {
const inspected = await executor.inspect({ project, service }, signal);
if (inspected.outcome !== 'done') {
return { event: 'unhealthy', result: { healthy: false, reason: 'inspect_unavailable', detail: inspected.detail } };
// While compose (or Arcane, for a Git project) recreates the service, the old container still sits next to the new one.
const transient = inspected.outcome === 'refused' && inspected.reason === 'multiple_containers' && attempt < VERIFY_ATTEMPTS;
if (!transient) return { event: 'unhealthy', result: { healthy: false, reason: 'inspect_unavailable', detail: inspected.detail, attempts: attempt } };
attempt += 1;
await deps.clock.sleep(VERIFY_INTERVAL_MS, signal);
continue;
}
const { container } = inspected.result;
const http = httpUrl === null || expectStatus === null ? null : await probe(httpUrl, expectStatus);
@@ -280,7 +285,7 @@ async function releaseNotes(deps: ComponentOperationExecutorDeps, record: Record
versionChange: firstString(record['change'], selection['change'], 'unknown'),
auto: firstString(record['auto'], 'none'),
promptHint: typeof record['promptHint'] === 'string' ? record['promptHint'] : null,
model: record['releaseNotesExecutor'] === 'opus' ? 'opus' : 'sonnet',
model: modelOr(record['releaseNotesExecutor']),
group: async () => (groupMembers === undefined ? [] : [...await groupMembers(app, signal)]),
}, signal);
@@ -316,7 +321,7 @@ async function migrationPlan(deps: ComponentOperationExecutorDeps, record: Recor
const project = asString(record['project']);
const service = asString(record['service']);
const app = firstString(record['component'], record['app'], project);
const executorName = record['migrateExecutor'] === 'sonnet' ? 'sonnet' : 'opus';
const executorName = modelOr(record['migrateExecutor']);
const nextSteps = asRecord(notes['report'])['nextSteps'];
const prompt = buildMigrationPlanPrompt({
app, project, service,
+12
View File
@@ -716,6 +716,18 @@ services:
expect(JSON.stringify(result)).not.toContain('old-db-secret');
});
it('without a new compose file pins the image in the current one, secrets untouched', async () => {
const { context, request } = await prepared();
const result = await context.executor.migrate({ ...request, compose_yaml: null, env: {} });
const written = WITH_SECRET.replace(`image: ${OLD}`, `image: ${NEW}`);
expect(written).not.toBe(WITH_SECRET);
expect(result).toMatchObject({ outcome: 'done', result: { state: 'applied', image: NEW, applied_manifest_sha256: sha(written), env_keys: [] } });
expect(readFileSync(context.manifest, 'utf8')).toBe(written);
expect(readFileSync(join(context.dir, '.env'), 'utf8')).toBe(ENV);
});
it('refuses an unknown secret, an unfilled input, a changed manifest, and leaves everything as it was', async () => {
const { context, request, envCandidate } = await prepared();
const cases = [
+4 -1
View File
@@ -588,7 +588,10 @@ export function createComposeExecutor(deps: ComposeExecutorDeps): ComposeExecuto
return refused('manifest_changed', `compose.yaml of "${project}" is ${short(current)}, not ${short(request.expect_manifest_sha256)} as planned`);
}
// The plan was written against the masked file (hl compose): its <secret:KEY> are this file's values.
const unmasked = unmaskComposeSecrets(request.compose_yaml, maskComposeSecrets(manifest.value.toString('utf8')).originals);
// Without a new file the migration keeps the current one (read just now, so it matches the backup) and only pins the image.
const unmasked = request.compose_yaml === null
? { ok: true as const, text: manifest.value.toString('utf8') }
: unmaskComposeSecrets(request.compose_yaml, maskComposeSecrets(manifest.value.toString('utf8')).originals);
if (!unmasked.ok) return refused('invalid_manifest', `the migration's compose.yaml has an ${unmasked.reason}`);
if (/<input:[A-Za-z0-9_]+>/.test(unmasked.text)) return refused('invalid_manifest', "the migration's compose.yaml still has an <input:NAME> the owner did not fill");
const pinned = rewriteImage(unmasked.text, service, image);
+5 -5
View File
@@ -7,7 +7,7 @@ const PINNED = `lscr.io/linuxserver/jackett:v0.24.2644-ls36@sha256:${'e'.repeat(
const config = (over: Partial<AppConfig> = {}): AppConfig => ({
app: 'jackett', project: 'jackett', enabled: true, position: 1, service: 'jackett', imageRepo: 'lscr.io/linuxserver/jackett',
tagPattern: '^v.*$', auto: 'patch', httpUrl: 'http://jackett:9117', expectStatus: 200, releaseNotesExecutor: 'sonnet',
migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, notes: null, apply: true, ...over,
migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, postUpgradeWhen: 'when_needed', notes: null, apply: true, ...over,
});
function prompts(app: AppConfig) {
@@ -24,16 +24,16 @@ describe('model action prompts', () => {
it('asks for the post-upgrade check what the old chain asks, with the versions, the group and the skill', async () => {
const asked = await prompts(config()).postUpgrade(target);
expect(asked.executor).toBe('sonnet');
expect(asked.executor).toBe('haiku');
expect(asked.prompt).toContain('v0.24.2627-ls34');
expect(asked.prompt).toContain('v0.24.2644-ls36');
expect(asked.prompt).toContain('flaresolverr');
expect(asked.prompt).toContain('Jackett keeps its indexers in /config.');
});
it('uses the models the app chose, opus by default for a repair', async () => {
expect((await prompts(config()).repair({ ...target, reason: 'verify_failed' })).executor).toBe('opus');
expect((await prompts(config({ postUpgradeExecutor: 'opus', migrateExecutor: 'sonnet' })).postUpgrade(target)).executor).toBe('opus');
it('uses the models the app chose, haiku by default for a repair', async () => {
expect((await prompts(config()).repair({ ...target, reason: 'verify_failed' })).executor).toBe('haiku');
expect((await prompts(config({ postUpgradeExecutor: 'opus', postUpgradeWhen: 'when_needed', migrateExecutor: 'sonnet' })).postUpgrade(target)).executor).toBe('opus');
const repair = await prompts(config({ migrateExecutor: 'sonnet' })).repair({ ...target, previousImage: null, reason: 'verify_failed: unhealthy' });
expect(repair.executor).toBe('sonnet');
expect(repair.prompt).toContain('verify_failed: unhealthy');
+5 -4
View File
@@ -2,6 +2,7 @@ import { splitImage } from '../domain/image-ref.ts';
import { buildPostUpgradePrompt } from '../domain/post-upgrade.ts';
import { buildRepairPrompt } from '../domain/repair.ts';
import type { AppConfigStore, ModelActionPrompts, SkillStore } from '../domain/ports.ts';
import { modelOr } from '../domain/steps.ts';
import type { AppConfig } from '../domain/steps.ts';
import type { GroupMember } from './ops/group.ts';
@@ -15,8 +16,8 @@ const tagOf = (image: string | null): string => (image === null ? 'unknown' : sp
/**
* The workflow engine's model actions ask what the old chain asks (domain/post-upgrade.ts, domain/repair.ts), with the
* app's compose group and skill, and the model the app chose: postUpgradeExecutor (sonnet by default) for the check
* after an update, migrateExecutor (opus by default) for a repair.
* app's compose group and skill, and the model the app chose: postUpgradeExecutor for the check after an update,
* migrateExecutor for a repair (each haiku by default).
*/
export function createModelActionPrompts(deps: ModelActionPromptDeps): ModelActionPrompts {
const configOf = (component: string): AppConfig => {
@@ -41,7 +42,7 @@ export function createModelActionPrompts(deps: ModelActionPromptDeps): ModelActi
update: { current: tagOf(request.previousImage), candidate: tagOf(request.image), pinned: request.image },
group, skill, migration: request.migration,
});
return { executor: config.postUpgradeExecutor === 'opus' ? 'opus' : 'sonnet', prompt };
return { executor: modelOr(config.postUpgradeExecutor), prompt };
},
async repair(request, signal) {
const { config, group, skill } = await facts(request.component, signal);
@@ -50,7 +51,7 @@ export function createModelActionPrompts(deps: ModelActionPromptDeps): ModelActi
update: { current: tagOf(request.previousImage), candidate: tagOf(request.image), pinned: request.image },
httpUrl: config.httpUrl, expectStatus: config.expectStatus, group, skill, fullAccess: request.fullAccess === true,
});
return { executor: config.migrateExecutor === 'sonnet' ? 'sonnet' : 'opus', prompt };
return { executor: modelOr(config.migrateExecutor), prompt };
},
};
}
+2 -1
View File
@@ -6,6 +6,7 @@ import { STRUCTURED_OUTPUT } from '../../domain/agent-guard.ts';
import { describeError, truncate } from '../../domain/mask.ts';
import type { CommandRunner, ModelFailure } from '../../domain/ports.ts';
import type { ModelSidecarRunResult } from '../../domain/model-api.ts';
import type { ModelName } from '../../domain/steps.ts';
// Runs the unmodified `claude` CLI, logged in with the owner's own account (docs/DESIGN.md, "Wykonawcy modeli"): print mode,
// JSON answer checked against a schema, a fresh empty working directory (no CLAUDE.md, no settings of anyone else), and for
@@ -30,7 +31,7 @@ export interface ClaudeDeps {
}
export interface ClaudeRunRequest {
readonly model: 'sonnet' | 'opus';
readonly model: ModelName;
readonly profile: 'readonly' | 'observe' | 'tools';
readonly prompt: string;
readonly outputSchema: unknown;
+1 -1
View File
@@ -88,7 +88,7 @@ describe('privileged action dispatcher', () => {
const current = {
app: 'jackett', project: 'jackett', enabled: true, position: 1, service: 'jackett', imageRepo: 'linuxserver/jackett',
tagPattern: '.*', auto: 'patch' as const, httpUrl: null, expectStatus: 200, releaseNotesExecutor: 'sonnet' as const,
migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, notes: null, apply: false,
migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, postUpgradeWhen: 'when_needed' as const, notes: null, apply: false,
};
const liveConfig = { get: () => current, list: () => [current], upsert: () => undefined };
const backupPolicy = PrivilegedActionPolicy.parse({
+1 -1
View File
@@ -22,7 +22,7 @@ const config = (over: Partial<AppConfig> = {}): AppConfig => ({
releaseNotesExecutor: 'sonnet',
migrateExecutor: null,
promptHint: null,
postUpgradeExecutor: null,
postUpgradeExecutor: null, postUpgradeWhen: 'when_needed',
notes: null,
apply: true,
...over,
+4 -3
View File
@@ -4,6 +4,7 @@ import { ModelError } from '../domain/ports.ts';
import type { ModelExecutor, SkillStore } from '../domain/ports.ts';
import { buildReleaseNotesPrompt, RELEASE_NOTES_ANSWER_SCHEMA, ReleaseNotesAnswer } from '../domain/release-notes.ts';
import type { ReleaseNotesContext } from '../domain/release-notes.ts';
import type { ModelName } from '../domain/steps.ts';
// Before an update a model finds the release notes itself (web search and fetch, the owner's list of trusted sources as
// a preference) and lays its assessment in front of the owner (requirements, breaking changes, what to do before and
@@ -14,8 +15,8 @@ export interface ReleaseNotesRead {
}
export type ReleaseNotesAssessment =
| { readonly kind: 'failed'; readonly why: string; readonly notes: ReleaseNotesRead; readonly model: 'sonnet' | 'opus'; readonly costUsd: number }
| { readonly kind: 'assessed'; readonly report: AgentReport; readonly notes: ReleaseNotesRead; readonly model: 'sonnet' | 'opus'; readonly costUsd: number };
| { readonly kind: 'failed'; readonly why: string; readonly notes: ReleaseNotesRead; readonly model: ModelName; readonly costUsd: number }
| { readonly kind: 'assessed'; readonly report: AgentReport; readonly notes: ReleaseNotesRead; readonly model: ModelName; readonly costUsd: number };
export interface ReleaseNotesQuestion {
readonly app: string;
@@ -27,7 +28,7 @@ export interface ReleaseNotesQuestion {
readonly versionChange: string;
readonly auto: string;
readonly promptHint: string | null;
readonly model: 'sonnet' | 'opus';
readonly model: ModelName;
readonly group: () => Promise<ReleaseNotesContext['group']>;
}
+1 -1
View File
@@ -44,7 +44,7 @@ async function setup() {
appConfigStore.upsert({
app: 'jackett', project: 'jackett', enabled: true, position: 1, service: 'jackett', imageRepo: 'linuxserver/jackett',
tagPattern: 'version-*', auto: 'patch', httpUrl: null, expectStatus: 200, releaseNotesExecutor: 'script',
migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, notes: null, apply: true,
migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, postUpgradeWhen: 'when_needed', notes: null, apply: true,
});
skillStore.upsert({ project: 'jackett', content: 'Jackett torrent proxy', updatedAt: now });
await ensureAppWorkflows({ workflowCatalog, functionCatalog, functionRuntime: passingRuntime, extensions, newId: () => crypto.randomUUID(), now: () => now }, ['jackett']);
+8 -6
View File
@@ -8,7 +8,8 @@ import type {
WorkflowInstanceStore,
} from '../domain/ports.ts';
import type { WorkflowInstance } from '../domain/workflow.ts';
import type { AppConfig, JsonValue } from '../domain/steps.ts';
import { modelOr } from '../domain/steps.ts';
import type { AppConfig, JsonValue, ModelName } from '../domain/steps.ts';
import { buildComponentUpdateSnapshot, buildSystemSnapshot } from '../domain/workflow-snapshot-builder.ts';
import { appWorkflowId } from './workflow-seed.ts';
@@ -123,13 +124,13 @@ export function startSystemRun(deps: WorkflowRunDeps, workflowId: string): Start
* project and service), so only what the context cannot say is here: apply's dry-run flag and the model profiles.
*/
function realRunRequests(deps: WorkflowRunDeps, config: AppConfig): Record<string, JsonValue> {
const profile = (model: 'sonnet' | 'opus') => {
const profile = (model: ModelName) => {
const latest = deps.modelProfileStore.getLatestVersion(model);
return latest === null ? null : { profileId: latest.profileId, version: latest.version };
};
// The same models the old chain uses: the app's choice, else sonnet checks after an update and opus repairs.
const postUpgrade = config.postUpgradeExecutor === 'opus' ? 'opus' : 'sonnet';
const repair = config.migrateExecutor === 'sonnet' ? 'sonnet' : 'opus';
// The app's choice, else the default model.
const postUpgrade = modelOr(config.postUpgradeExecutor);
const repair = modelOr(config.migrateExecutor);
return {
backupRequest: { project: config.project },
applyDryRunRequest: { dryRun: true },
@@ -139,7 +140,8 @@ function realRunRequests(deps: WorkflowRunDeps, config: AppConfig): Record<strin
repairFullRequest: { reason: 'owner_requested', access: 'full' },
rollbackRequest: {},
migrateRequest: {},
postUpgradeRequest: {},
// Without a model chosen the check is off (domain/steps.ts); with one, the app says when it runs.
postUpgradeRequest: { when: config.postUpgradeExecutor === null ? 'off' : config.postUpgradeWhen },
postUpgradeProfile: profile(postUpgrade),
repairProfile: profile(repair),
};
+24 -6
View File
@@ -104,13 +104,14 @@ const MIGRATION_PLAN = {
stepsBefore: [], stepsAfter: ['hl ps'], rollback: 'Restore the backup.',
};
function fakeOperations(tags: 'candidate_within_auto' | 'candidate_above_auto' = 'candidate_within_auto', migration: 'none' | 'needed' = 'none'): ComponentOperationRunner {
function fakeOperations(tags: 'candidate_within_auto' | 'candidate_above_auto' = 'candidate_within_auto', migration: 'none' | 'needed' | 'image_only' = 'none'): ComponentOperationRunner {
const plan = migration === 'image_only' ? { ...MIGRATION_PLAN, composeYaml: null, inputs: [] } : MIGRATION_PLAN;
const results: Record<string, { event: string; result: JsonValue }> = {
image_version: { event: 'version_detected', result: { event: 'version_detected', detectedVersion: '0.22.0', imageVersionState: 'detected' } },
image_tags: { event: tags, result: { event: tags, current: '0.22.0', candidate: 'version-v0.22.1', change: tags === 'candidate_within_auto' ? 'patch' : 'major' } },
release_notes: { event: 'no_source', result: { sources: [] } },
migration_plan: migration === 'needed'
? { event: 'migration_needed', result: { plan: MIGRATION_PLAN, planDigest: 'd'.repeat(64), summary: MIGRATION_PLAN.summary, model: 'opus', costUsd: 0.3, actions: [] } }
migration_plan: migration !== 'none'
? { event: 'migration_needed', result: { plan, planDigest: 'd'.repeat(64), summary: plan.summary, model: 'opus', costUsd: 0.3, actions: [] } }
: { event: 'no_migration', result: { plan: null, summary: 'Nothing to migrate.' } },
image_resolve: {
event: 'resolved',
@@ -146,7 +147,7 @@ async function setup(
appConfigStore.upsert({
app: 'jackett', project: 'jackett', enabled: true, position: 1, service: 'jackett', imageRepo: 'linuxserver/jackett',
tagPattern: 'version-*', auto: 'patch', httpUrl: null, expectStatus: 200, releaseNotesExecutor: 'script',
migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, notes: null, apply: true,
migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, postUpgradeWhen: 'when_needed', notes: null, apply: true,
});
skillStore.upsert({ project: 'jackett', content: 'Jackett torrent proxy', updatedAt: now });
// The real sandbox: the template functions run in bwrap exactly as on VM103.
@@ -253,7 +254,7 @@ describe('workflow runner', () => {
{ image: PINNED, backupId: BACKUP_ID, expectedManifestSha256: MANIFEST, dryRun: true },
{ image: PINNED, backupId: BACKUP_ID, expectedManifestSha256: MANIFEST, dryRun: false },
]);
expect(docker.requests.find((request) => request.action === 'model.postUpgrade')).toMatchObject({ image: PINNED, backupId: BACKUP_ID, profile: { profileId: 'sonnet' } });
expect(docker.requests.find((request) => request.action === 'model.postUpgrade')).toMatchObject({ image: PINNED, backupId: BACKUP_ID, profile: { profileId: 'haiku' }, when: 'off', reasons: [] });
});
it('rolls the app back when the owner chooses rollback after the model asked for action', async () => {
@@ -321,10 +322,27 @@ describe('workflow runner', () => {
action: 'docker.migrate', version: 1, component: 'jackett', project: 'jackett', service: 'jackett', image: PINNED, backupId: BACKUP_ID,
expectedManifestSha256: MANIFEST, composeYaml: MIGRATION_PLAN.composeYaml, planDigest: 'd'.repeat(64), inputs: [{ name: 'ADMIN_PASSWORD', target: 'env' }],
});
expect(docker.requests.find((request) => request.action === 'model.postUpgrade')).toMatchObject({ migration: { summary: MIGRATION_PLAN.summary, stepsAfter: ['hl ps'] } });
expect(docker.requests.find((request) => request.action === 'model.postUpgrade')).toMatchObject({
migration: { summary: MIGRATION_PLAN.summary, stepsAfter: ['hl ps'] }, reasons: ['approved_update', 'major', 'migration'],
});
expect(instanceStore.getReplayStatus('run-1').status).toBe('ok');
});
it('runs an approved migration that keeps the compose file: the executor only pins the new image', async () => {
const { runner, instanceStore, docker, runDeps } = await setup(fakeOperations('candidate_above_auto', 'image_only'));
startWorkflowRun(runDeps, 'jackett', 'real');
await runUntilIdle(runner);
const instance = instanceStore.get('run-1')!;
expect(instanceStore.submitOwnerDecision({
decision: { schemaVersion: 1, decisionId: crypto.randomUUID(), signal: 'owner-decision.v1', instanceId: 'run-1', waitNodeId: 'gate_migration', action: 'approve', note: null },
expectedLifecycleSeq: instance.projection.seq, expectedExecutionSeq: instanceStore.getReplayStatus('run-1').replaySeq, now,
}).ok).toBe(true);
expect(await runUntilIdle(runner)).toContain('apply_migration');
expect(docker.requests.find((request) => request.action === 'docker.migrate')).toMatchObject({ composeYaml: null, inputs: [] });
expect(instanceStore.get('run-1')?.projection.status).toBe('succeeded');
});
it('goes the usual way above the auto threshold when the model finds no migration', async () => {
const { runner, runDeps } = await setup(fakeOperations('candidate_above_auto'));
startShadowRun(runDeps, 'jackett');
+3 -3
View File
@@ -60,10 +60,10 @@ describe('ensureUpdatePolicy', () => {
});
describe('ensureDefaultModelProfiles', () => {
it('adds read-only Claude profiles for the sonnet and opus executors once', () => {
it('adds read-only Claude profiles for the haiku, sonnet and opus executors once', () => {
const store = createSqliteModelProfileStore(openDatabase(':memory:'));
expect(ensureDefaultModelProfiles(store, now)).toEqual(['sonnet', 'opus']);
expect(ensureDefaultModelProfiles(store, now)).toEqual(['haiku', 'sonnet', 'opus']);
expect(ensureDefaultModelProfiles(store, now)).toEqual([]);
expect(store.getLatestVersion('opus')?.document).toMatchObject({ provider: 'claude', model: 'opus', mode: 'readonly', timeoutMs: 1_800_000 });
});
@@ -74,7 +74,7 @@ describe('ensureDefaultModelProfiles', () => {
store.publishVersion({ profileId: 'sonnet', expectedVersion: 1, document: first('sonnet'), now });
store.publishVersion({ profileId: 'opus', expectedVersion: 1, document: first('opus', { provider: 'codex' }), now });
expect(ensureDefaultModelProfiles(store, now)).toEqual(['sonnet']);
expect(ensureDefaultModelProfiles(store, now)).toEqual(['haiku', 'sonnet']);
expect(store.getLatestVersion('sonnet')).toMatchObject({ version: 2, document: { timeoutMs: 1_800_000 } });
expect(store.getLatestVersion('opus')).toMatchObject({ version: 1, document: { provider: 'codex', timeoutMs: 300_000 } });
});
+5 -4
View File
@@ -12,6 +12,7 @@ import {
} from '../domain/workflow-update.ts';
import { createWorkflowCatalogApp, createWorkflowGraphValidator } from './workflow-catalog.ts';
import { createWorkflowFunctionCatalogApp } from './workflow-function-catalog.ts';
import { ModelName } from '../domain/steps.ts';
/** Each app gets its own update workflow, first published as a copy of the component-update template. */
export { appWorkflowId };
@@ -28,7 +29,7 @@ export function ensureUpdatePolicy(store: PrivilegedActionStore, now: string): b
}
/** A default profile's first document; one still exactly like it was never changed by the owner. */
const FIRST_DEFAULT_PROFILE = (model: 'sonnet' | 'opus'): ModelProfileDocumentV1 => ({
const FIRST_DEFAULT_PROFILE = (model: ModelName): ModelProfileDocumentV1 => ({
schemaVersion: 1, provider: 'claude', model, mode: 'readonly', timeoutMs: 300_000, maxCostUsd: null, tools: [], systemPrompt: null, context: null,
});
@@ -36,17 +37,17 @@ const FIRST_DEFAULT_PROFILE = (model: 'sonnet' | 'opus'): ModelProfileDocumentV1
* The default profile: the model sidecar's own limit (30 min, MODEL_TIMEOUT_MS), since a repair session takes long; the
* dispatcher stops a run at its profile's timeout.
*/
const DEFAULT_PROFILE = (model: 'sonnet' | 'opus'): ModelProfileDocumentV1 => ({ ...FIRST_DEFAULT_PROFILE(model), timeoutMs: 30 * 60_000 });
const DEFAULT_PROFILE = (model: ModelName): ModelProfileDocumentV1 => ({ ...FIRST_DEFAULT_PROFILE(model), timeoutMs: 30 * 60_000 });
/**
* App configs name 'sonnet' and 'opus' as executors; a model request without a pinned profile uses the profile of that
* App configs name 'haiku', 'sonnet' and 'opus' as executors; a model request without a pinned profile uses the profile of that
* name, and a workflow snapshot pins one for each. Adds the missing ones as Claude profiles, and moves a profile the owner
* never changed from its first document to the current default; the owner edits them on the Models page (the provider
* picks the Claude or Codex sidecar).
*/
export function ensureDefaultModelProfiles(store: ModelProfileStore, now: string): readonly string[] {
const added: string[] = [];
for (const model of ['sonnet', 'opus'] as const) {
for (const model of ModelName.options) {
const latest = store.getLatestVersion(model);
if (latest !== null && !isDeepStrictEqual(latest.document, FIRST_DEFAULT_PROFILE(model))) continue;
const result = store.publishVersion({ profileId: model, expectedVersion: (latest?.version ?? 0) + 1, document: DEFAULT_PROFILE(model), now });
+1 -1
View File
@@ -16,7 +16,7 @@ const bazarr: AppConfig = {
releaseNotesExecutor: 'sonnet',
migrateExecutor: null,
promptHint: null,
postUpgradeExecutor: null,
postUpgradeExecutor: null, postUpgradeWhen: 'when_needed',
notes: null,
apply: false,
};
+5 -2
View File
@@ -1,5 +1,5 @@
import { z } from 'zod';
import { AutoLevel, Executor } from './steps.ts';
import { AutoLevel, Executor, ModelName, PostUpgradeWhen } from './steps.ts';
import type { AppConfig } from './steps.ts';
import { compileTagPattern } from './tags.ts';
@@ -15,7 +15,7 @@ export function countWords(text: string): number {
}
/** Model steps run on a model; `script` is not a choice for them. */
const ModelChoice = z.enum(['sonnet', 'opus']);
const ModelChoice = ModelName;
const cleared = (value: unknown) => (typeof value === 'string' && value.trim() === '' ? null : value);
@@ -41,6 +41,7 @@ export const AppConfigPatch = z
migrateExecutor: ModelChoice.nullable(),
promptHint: z.preprocess(cleared, z.string().trim().max(500).nullable()),
postUpgradeExecutor: ModelChoice.nullable(),
postUpgradeWhen: PostUpgradeWhen,
notes: z.preprocess(
cleared,
z
@@ -61,6 +62,8 @@ export const AppConfigValue = z.strictObject({
service: z.string().min(1), imageRepo: z.string().min(1), tagPattern: z.string(), auto: AutoLevel,
httpUrl: z.string().nullable(), expectStatus: z.number().int(), releaseNotesExecutor: Executor,
migrateExecutor: Executor.nullable(), promptHint: z.string().nullable(), postUpgradeExecutor: Executor.nullable(),
// Revisions from before 2026-10-07 have none: the same default as the app_config column (migration 25).
postUpgradeWhen: PostUpgradeWhen.default('when_needed'),
notes: z.string().nullable(), apply: z.boolean(),
});
+2 -1
View File
@@ -88,7 +88,8 @@ export const MigrateRequest = z.strictObject({
image: PinnedImage,
expect_manifest_sha256: Sha256,
backup_id: BackupId,
compose_yaml: z.string().min(1).max(65_536),
/** null: the plan keeps the current compose.yaml; only the service's image changes. */
compose_yaml: z.string().min(1).max(65_536).nullable(),
env: z.record(z.string().regex(/^[A-Z][A-Z0-9_]{0,63}$/), z.string().max(4096).regex(/^[^\n\r]*$/, 'one line')),
});
export type MigrateRequest = z.infer<typeof MigrateRequest>;
+2 -1
View File
@@ -1,5 +1,6 @@
import { z } from 'zod';
import type { ModelFailure } from './ports.ts';
import { ModelName } from './steps.ts';
// The model sidecar's API: the driver asks it to run one prompt (readonly, or with the hl tool in a repair session).
@@ -16,7 +17,7 @@ export type ModelSidecarRunResult =
| { readonly ok: false; readonly kind: ModelFailure; readonly detail: string; readonly costUsd?: number | null };
export const ModelRunRequest = z.strictObject({
executor: z.enum(['sonnet', 'opus']),
executor: ModelName,
profile: z.enum(['readonly', 'observe', 'tools']),
prompt: z.string().min(1).max(400_000),
output_schema: z.record(z.string(), z.unknown()),
+1 -1
View File
@@ -16,7 +16,7 @@ const config = (app: string, position: number, enabled = true): AppConfig => ({
releaseNotesExecutor: 'sonnet',
migrateExecutor: null,
promptHint: null,
postUpgradeExecutor: null,
postUpgradeExecutor: null, postUpgradeWhen: 'when_needed',
notes: null,
apply: false,
});
+6 -3
View File
@@ -18,6 +18,7 @@ import type {
SessionOpenRequest,
SessionOpenResult,
SessionLogResult,
AgentLogsRequest,
AgentTextResult,
SessionLogEntry,
SessionSummary,
@@ -25,7 +26,7 @@ import type {
import type { PreflightReport } from './preflight.ts';
import type { Skill, SkillVersion } from './skill.ts';
import type { AppConfigRevision } from './app-config.ts';
import type { AppConfig, Executor, JsonValue } from './steps.ts';
import type { AppConfig, Executor, JsonValue, ModelName } from './steps.ts';
import type { HostStatus } from './host-status.ts';
import type { NotificationSettings } from './notification-settings.ts';
import type { AppendableWorkflowInstanceEventInput, PublishedWorkflowVersion, WorkflowDraft, WorkflowDocument, WorkflowInstance, WorkflowInstanceCreateInput, WorkflowInstanceEvent, WorkflowInstanceStatus, WorkflowValidation } from './workflow.ts';
@@ -403,6 +404,8 @@ export interface ExecutorClient {
sessionLog(request: SessionCloseRequest, signal?: AbortSignal): Promise<ExecutorCall<SessionLogResult>>;
/** The project's compose.yaml as the session sees it (secrets masked as `<secret:NAME>`). */
sessionCompose(request: SessionCloseRequest, signal?: AbortSignal): Promise<ExecutorCall<AgentTextResult>>;
/** A service's last log lines as the session reads them (secrets masked), like the model's `hl logs`. */
sessionLogs(request: SessionCloseRequest & AgentLogsRequest, signal?: AbortSignal): Promise<ExecutorCall<AgentTextResult>>;
}
export interface ModelRequest {
@@ -476,14 +479,14 @@ export interface ModelActionPrompts {
readonly migration?: { readonly summary: string; readonly stepsAfter: readonly string[] };
},
signal?: AbortSignal,
): Promise<{ readonly executor: 'sonnet' | 'opus'; readonly prompt: string }>;
): Promise<{ readonly executor: ModelName; readonly prompt: string }>;
repair(
request: {
readonly component: string; readonly project: string; readonly service: string; readonly image: string; readonly reason: string; readonly previousImage: string | null;
readonly fullAccess?: boolean;
},
signal?: AbortSignal,
): Promise<{ readonly executor: 'sonnet' | 'opus'; readonly prompt: string }>;
): Promise<{ readonly executor: ModelName; readonly prompt: string }>;
}
/** Runs a model (the `claude` CLI logged in with the owner's account, in the model sidecar). Throws ModelError. */
+33 -1
View File
@@ -1,5 +1,5 @@
import { describe, expect, it } from 'vitest';
import { buildPostUpgradePrompt, POST_UPGRADE_ANSWER_SCHEMA, PostUpgradeAnswer } from './post-upgrade.ts';
import { buildPostUpgradePrompt, POST_UPGRADE_ANSWER_SCHEMA, PostUpgradeAnswer, suspiciousLogLines } from './post-upgrade.ts';
const context = {
app: 'nextcloud',
@@ -58,3 +58,35 @@ describe('PostUpgradeAnswer', () => {
expect(POST_UPGRADE_ANSWER_SCHEMA).toMatchObject({ type: 'object', required: ['outcome', 'summary', 'actions'] });
});
});
describe('suspiciousLogLines', () => {
it('finds errors, fatal ones, panics, exceptions and failed migrations', () => {
const logs = [
'[migrations] started',
'[migrations] no migrations found',
'2026-10-07 09:27:01 INFO Starting Sonarr',
'2026-10-07 09:27:02 ERROR Unable to connect to indexer',
'level=fatal msg="database is locked"',
'panic: runtime error: index out of range',
'System.NullReferenceException: Object reference not set',
'Traceback (most recent call last):',
'Migration failed: column already exists',
].join('\n');
expect(suspiciousLogLines(logs)).toEqual([
'2026-10-07 09:27:02 ERROR Unable to connect to indexer',
'level=fatal msg="database is locked"',
'panic: runtime error: index out of range',
'System.NullReferenceException: Object reference not set',
'Traceback (most recent call last):',
'Migration failed: column already exists',
]);
});
it('a clean start, and lines that only say there are no errors, find nothing', () => {
expect(suspiciousLogLines('[migrations] no migrations found\nStarted in 2.1 s, 0 errors\nno errors found\n')).toEqual([]);
});
it('keeps at most the first lines it is asked for', () => {
expect(suspiciousLogLines(Array.from({ length: 30 }, (_, index) => `ERROR ${String(index)}`).join('\n'), 3)).toEqual(['ERROR 0', 'ERROR 1', 'ERROR 2']);
});
});
+14 -1
View File
@@ -2,7 +2,20 @@ import { z } from 'zod';
// The step after a successful update (owner, 2026-09-20): some apps ask for actions once the new version runs (Nextcloud: missing
// database indices, occ repair). A model looks at what the release notes analysis and the app's skill say, does what the tool
// allows and lists the rest for the owner. Per app, off by default: then the chain stays scripts only.
// allows and lists the rest for the owner. Per app, off by default: then the chain stays scripts only. When on, by default only
// when the update may ask for something (owner, 2026-10-07): one the owner approved, a major one, a migration, or errors in the logs.
/** Why a post-upgrade check runs the model. */
export const PostUpgradeReason = z.enum(['approved_update', 'major', 'migration', 'log_errors']);
export type PostUpgradeReason = z.infer<typeof PostUpgradeReason>;
const SUSPICIOUS = /\b(error|fatal|panic|traceback|critical)\b|exception\b|migrations? failed/i;
const NO_ERRORS = /\b(no|0) errors?\b/i;
/** The log lines after a start that look like trouble, the first `max` of them. */
export function suspiciousLogLines(logs: string, max = 10): string[] {
return logs.split('\n').map((line) => line.trim()).filter((line) => SUSPICIOUS.test(line) && !NO_ERRORS.test(line)).slice(0, max);
}
export const PostUpgradeAnswer = z.object({
outcome: z
+13 -2
View File
@@ -1,4 +1,6 @@
import { z } from 'zod';
import { PostUpgradeReason } from './post-upgrade.ts';
import { ModelName } from './steps.ts';
/**
* The executor is intentionally a closed vocabulary. Adding an action is a
@@ -128,7 +130,8 @@ export const DockerMigrateRequest = z.strictObject({
image: PinnedImage,
backupId: BackupId,
expectedManifestSha256: Sha256,
composeYaml: z.string().min(1).max(65_536),
/** null when the plan keeps the compose file as it is apart from the image. */
composeYaml: z.string().min(1).max(65_536).nullable(),
planDigest: Sha256,
inputs: z.array(z.strictObject({ name: z.string().regex(/^[A-Z][A-Z0-9_]{0,63}$/), target: z.enum(['env', 'compose']) })).max(20),
});
@@ -182,6 +185,10 @@ export const ModelPostUpgradeRequest = z.strictObject({
profile: ProfileReference,
/** After an approved migration (docker.migrate): its summary and the hl steps its plan runs once the app is up. */
migration: z.strictObject({ summary: z.string().max(2000), stepsAfter: z.array(z.string().min(1).max(500)).max(30) }).optional(),
/** The app's setting: off = no model, when_needed = only with a reason (or errors in the logs). Older requests have none: always. */
when: z.enum(['off', 'when_needed', 'always']).optional(),
/** What in this run asks for a look: an update the owner approved, a major one, a migration. */
reasons: z.array(PostUpgradeReason).max(4).optional(),
});
export type ModelPostUpgradeRequest = z.infer<typeof ModelPostUpgradeRequest>;
@@ -322,7 +329,7 @@ const ModelActionReport = z.strictObject({
outcome: z.string().min(1).max(64),
summary: z.string().max(4000),
actions: z.array(z.string().max(500)).max(50),
model: z.enum(['sonnet', 'opus']),
model: ModelName,
costUsd: z.number().nonnegative(),
});
@@ -354,6 +361,10 @@ export const ModelPostUpgradeResult = z.strictObject({
manifestChanged: z.boolean(),
manifestSha256: Sha256,
report: ModelActionReport.optional(),
/** Why the model ran; absent on results from before 2026-10-07. */
reasons: z.array(PostUpgradeReason).max(4).optional(),
/** The model did not run: the app has it off, or nothing in the run or the logs asked for it. */
skipped: z.enum(['off', 'not_needed']).optional(),
}).strict();
export type ModelPostUpgradeResult = z.infer<typeof ModelPostUpgradeResult>;
+16 -1
View File
@@ -3,10 +3,24 @@ import { z } from 'zod';
// Shared vocabulary: model executors, auto levels, reason codes, JSON and an app's configuration. The zod schema is the
// single source of the type. (The file is named after the retired step queue it started with.)
export const Executor = z.enum(['script', 'sonnet', 'opus']);
/** The Claude models a model step runs on; the profile of the same name picks the sidecar (Models page). */
export const ModelName = z.enum(['haiku', 'sonnet', 'opus']);
export type ModelName = z.infer<typeof ModelName>;
/** The model of a model step whose app chose none (owner, 2026-10-07: haiku is cheap and good enough). */
export const DEFAULT_MODEL: ModelName = 'haiku';
/** The app's model choice, or the default when it names none (null, `script`, anything else). */
export const modelOr = (choice: unknown): ModelName => {
const parsed = ModelName.safeParse(choice);
return parsed.success ? parsed.data : DEFAULT_MODEL;
};
export const Executor = z.enum(['script', ...ModelName.options]);
export type Executor = z.infer<typeof Executor>;
export const AutoLevel = z.enum(['none', 'patch', 'minor', 'major']);
/** When an app's post-upgrade check runs its model: only when the update may ask for something, or after every update. */
export const PostUpgradeWhen = z.enum(['when_needed', 'always']);
export type PostUpgradeWhen = z.infer<typeof PostUpgradeWhen>;
export type AutoLevel = z.infer<typeof AutoLevel>;
/** Why an operation stopped or failed (app/ops); unknown cause = `unexpected`. */
@@ -59,6 +73,7 @@ export interface AppConfig {
readonly promptHint: string | null;
/** null = after an update only scripts check it; a model here checks what the update asks to be done afterwards. */
readonly postUpgradeExecutor: Executor | null;
readonly postUpgradeWhen: PostUpgradeWhen;
readonly notes: string | null;
/** Real updates through the executor (backup, apply, verify). Off = only dry runs. */
readonly apply: boolean;
+1 -1
View File
@@ -21,7 +21,7 @@ export const PRIVILEGED_ACTION_DESCRIPTIONS: Readonly<Record<PrivilegedActionId,
'docker.diskUsage': { title: 'Check the disk', kind: 'read', summary: 'Executor reads the free space of the Docker filesystem and what images and build cache could give back.' },
'docker.imagePrune': { title: 'Remove unused images', kind: 'change', summary: 'Executor removes Docker images no container uses (and the build cache when asked); volumes are never touched.' },
'model.repair': { title: 'Model repairs the app', kind: 'model', summary: 'A model reads logs and the compose file and may change them to make the app healthy again.' },
'model.postUpgrade': { title: 'Model checks the update', kind: 'model', summary: 'A model checks the updated app (hl ps, hl logs), may restart it, and asks the owner if something needs a decision.' },
'model.postUpgrade': { title: 'Model checks the update', kind: 'model', summary: 'A model checks the updated app (hl ps, hl logs), may restart it, and asks the owner if something needs a decision. Off per app; by default only for an approved or major update, a migration, or errors in the logs.' },
};
export const COMPONENT_OPERATION_DESCRIPTIONS: Readonly<Record<ComponentOperationId, BlockDescription>> = {
+14
View File
@@ -27,4 +27,18 @@ describe('workflowModelCalls', () => {
expect(calls[2]).toMatchObject({ instanceId: 'run-1', workflowId: 'update-radarr', summary: 'Clean start.', actions: ['hl ps'], at: '2026-09-22T21:35:00.000Z' });
expect(totalCostUsd(calls)).toBeCloseTo(0.82);
});
it('counts a model run once when a later node carries its result on, as the terminal node does', () => {
const postUpgrade = { action: 'model.postUpgrade', state: 'completed', report: { outcome: 'nothing_to_do', summary: 'Clean start.', actions: ['hl ps'], model: 'sonnet', costUsd: 0.2 } };
const notes = { verdict: 'attention', summary: 'A migration runs.', model: 'opus', costUsd: 0.12 };
const calls = workflowModelCalls({ id: 'run-1', workflowId: 'update-radarr' }, [
attempt(1, 'release_notes', notes),
attempt(2, 'gate', notes),
attempt(3, 'post_upgrade', postUpgrade),
attempt(4, 'terminal_updated', postUpgrade),
]);
expect(calls.map((call) => call.nodeId)).toEqual(['release_notes', 'post_upgrade']);
expect(totalCostUsd(calls)).toBeCloseTo(0.32);
});
});
+32 -23
View File
@@ -22,34 +22,43 @@ const cost = (value: unknown): number => (typeof value === 'number' && Number.is
/**
* The model runs a workflow run made, read from its node attempts: a model action's result carries `report`
* (app/privileged-action-executor-adapter), the release notes operation's result carries `verdict` and `model`.
* A node that passes the record on (a terminal node's result is its input) repeats that result; the same call counts once.
*/
export function workflowModelCalls(instance: { readonly id: string; readonly workflowId: string }, attempts: readonly NodeAttempt[]): WorkflowModelCall[] {
return attempts.flatMap((attempt): WorkflowModelCall[] => {
const result = record(attempt.result);
const at = attempt.finishedAt ?? attempt.startedAt;
const base = { instanceId: instance.id, workflowId: instance.workflowId, nodeId: attempt.nodeId, at };
const action = text(result['action']);
if (action === 'model.postUpgrade' || action === 'model.repair') {
const report = record(result['report']);
if (Object.keys(report).length === 0) return [];
return [{
...base, kind: action === 'model.repair' ? 'repair' : 'postUpgrade', model: text(report['model']),
outcome: text(report['outcome']) ?? 'unknown', summary: text(report['summary']) ?? '',
actions: Array.isArray(report['actions']) ? report['actions'].filter((item): item is string => typeof item === 'string') : [],
costUsd: cost(report['costUsd']),
}];
}
// Only a release notes result that reached a model names one; "no source" and unreadable notes cost nothing.
if ('verdict' in result && text(result['model']) !== null) {
return [{
...base, kind: 'releaseNotes', model: text(result['model']), outcome: text(result['verdict']) ?? 'unknown',
summary: text(result['summary']) ?? '', actions: [], costUsd: cost(result['costUsd']),
}];
}
return [];
const seen = new Set<string>();
return attempts.flatMap((attempt) => modelCallOf(instance, attempt)).filter(({ nodeId: _nodeId, at: _at, ...call }) => {
const key = JSON.stringify(call);
if (seen.has(key)) return false;
seen.add(key);
return true;
});
}
function modelCallOf(instance: { readonly id: string; readonly workflowId: string }, attempt: NodeAttempt): WorkflowModelCall[] {
const result = record(attempt.result);
const at = attempt.finishedAt ?? attempt.startedAt;
const base = { instanceId: instance.id, workflowId: instance.workflowId, nodeId: attempt.nodeId, at };
const action = text(result['action']);
if (action === 'model.postUpgrade' || action === 'model.repair') {
const report = record(result['report']);
if (Object.keys(report).length === 0) return [];
return [{
...base, kind: action === 'model.repair' ? 'repair' : 'postUpgrade', model: text(report['model']),
outcome: text(report['outcome']) ?? 'unknown', summary: text(report['summary']) ?? '',
actions: Array.isArray(report['actions']) ? report['actions'].filter((item): item is string => typeof item === 'string') : [],
costUsd: cost(report['costUsd']),
}];
}
// Only a release notes result that reached a model names one; "no source" and unreadable notes cost nothing.
if ('verdict' in result && text(result['model']) !== null) {
return [{
...base, kind: 'releaseNotes', model: text(result['model']), outcome: text(result['verdict']) ?? 'unknown',
summary: text(result['summary']) ?? '', actions: [], costUsd: cost(result['costUsd']),
}];
}
return [];
}
export function totalCostUsd(calls: readonly WorkflowModelCall[]): number {
return calls.reduce((sum, call) => sum + call.costUsd, 0);
}
@@ -7,6 +7,7 @@ import {
PrivilegedActionVersion,
} from './privileged-action.ts';
import type { PrivilegedActionPolicy as PrivilegedActionPolicyType, PrivilegedActionResult as ActionResult } from './privileged-action.ts';
import type { PostUpgradeReason } from './post-upgrade.ts';
import { privilegedActionIdentityMatches } from './privileged-action-invocation.ts';
import type { JsonValue } from './steps.ts';
import { bindWorkflowResult, bindingInput, DataBindingV1 } from './workflow-data-binding.ts';
@@ -83,6 +84,15 @@ export function resultEvent(result: ActionResult): string {
export function createPrivilegedActionExtensionRegistry(policy?: unknown): WorkflowExtensionRegistry {
const parsedPolicy = policy === undefined ? null : PrivilegedActionPolicy.safeParse(policy);
const policyValue = parsedPolicy === null ? undefined : parsedPolicy.success ? parsedPolicy.data : null;
/** What in the run asks for a post-upgrade look: release notes are read only for an update above the app's auto level. */
function postUpgradeReasons(context: JsonValue, migrated: boolean): PostUpgradeReason[] {
const reasons: PostUpgradeReason[] = [];
if (isRecord(jsonPointer(context, '/releaseNotes'))) reasons.push('approved_update');
if (jsonPointer(context, '/imageSelection/change') === 'major') reasons.push('major');
if (migrated) reasons.push('migration');
return reasons;
}
function deriveMergedSelected(
parsedConfig: PrivilegedActionNodeConfig,
context: JsonValue,
@@ -135,7 +145,7 @@ function deriveMergedSelected(
const plan = jsonPointer(context, '/migrationPlan/plan');
const digest = jsonPointer(context, '/migrationPlan/planDigest');
if (isRecord(plan)) {
if (!('composeYaml' in result) && typeof plan.composeYaml === 'string') result['composeYaml'] = plan.composeYaml;
if (!('composeYaml' in result) && (typeof plan.composeYaml === 'string' || plan.composeYaml === null)) result['composeYaml'] = plan.composeYaml;
if (!('inputs' in result) && Array.isArray(plan.inputs)) {
result['inputs'] = plan.inputs.filter(isRecord).map((input) => ({ name: input.name ?? null, target: input.target ?? null }));
}
@@ -151,6 +161,7 @@ function deriveMergedSelected(
result['migration'] = { summary: plan.summary, stepsAfter: plan.stepsAfter };
}
}
if (action === 'model.postUpgrade' && !('reasons' in result)) result['reasons'] = postUpgradeReasons(context, 'migration' in result);
// Profile & Reason for model actions
if (action === 'model.repair' || action === 'model.postUpgrade') {
+2 -1
View File
@@ -2,7 +2,7 @@ import { HostStatus } from '../domain/host-status.ts';
import { NotificationSettings } from '../domain/notification-settings.ts';
import { z } from 'zod';
import { SkillContent } from '../domain/skill.ts';
import { AutoLevel, Executor, Json } from '../domain/steps.ts';
import { AutoLevel, Executor, Json, PostUpgradeWhen } from '../domain/steps.ts';
import {
WorkflowDraft,
PublishedWorkflowVersion,
@@ -65,6 +65,7 @@ export const AppConfigResponse = z.object({
migrateExecutor: Executor.nullable(),
promptHint: z.string().nullable(),
postUpgradeExecutor: Executor.nullable(),
postUpgradeWhen: PostUpgradeWhen,
notes: z.string().nullable(),
apply: z.boolean(),
});
+13
View File
@@ -95,6 +95,19 @@ describe('HTTP executor client failures', () => {
expect(seen).toEqual([{ url: 'http://executor:8081/agent/compose', authorization: `Bearer ${'a'.repeat(32)}` }]);
});
it('reads a service’s logs with the session token, like the model’s hl logs', async () => {
const seen: { url: string; authorization: string | null; body: unknown }[] = [];
const client = createHttpExecutorClient({
baseUrl: 'http://executor:8081',
fetch: (input, init) => {
seen.push({ url: input instanceof Request ? input.url : input.toString(), authorization: new Headers(init?.headers).get('authorization'), body: JSON.parse(init?.body as string) });
return answering({ outcome: 'done', result: { text: 'started\n' } })();
},
});
expect(await client.sessionLogs({ token: 'a'.repeat(32), service: 'sonarr', tail: 200 })).toEqual({ outcome: 'done', result: { text: 'started\n' } });
expect(seen).toEqual([{ url: 'http://executor:8081/agent/logs', authorization: `Bearer ${'a'.repeat(32)}`, body: { service: 'sonarr', tail: 200 } }]);
});
it('an unreachable executor is unavailable, naming where it looked', async () => {
const client = createHttpExecutorClient({ baseUrl: 'http://executor:8081', fetch: () => Promise.reject(new TypeError('fetch failed', { cause: { code: 'ECONNREFUSED' } })) });
expect(await client.inspect({ project: 'jackett', service: 'jackett' })).toEqual({ outcome: 'unavailable', detail: 'executor unreachable at http://executor:8081/inspect: fetch failed (ECONNREFUSED)' });
+1
View File
@@ -74,5 +74,6 @@ export function createHttpExecutorClient(options: HttpExecutorClientOptions): Ex
closeSession: (request, signal) => post('/session/close', request, SessionCloseResponse, signal),
sessionLog: (request, signal) => post('/session/log', request, SessionLogResponse, signal),
sessionCompose: ({ token }, signal) => post('/agent/compose', {}, AgentTextResponse, signal, token),
sessionLogs: ({ token, ...request }, signal) => post('/agent/logs', request, AgentTextResponse, signal, token),
};
}
+41
View File
@@ -137,6 +137,47 @@ describe('createHttpRegistryClient', () => {
});
});
describe('HTTP 429', () => {
const tooMany = (headers: Record<string, string> = {}) => new Response('', { status: 429, headers });
it('waits and asks again, so a short rate limit does not fail the lookup', async () => {
const waits: number[] = [];
const { fetchFn, calls } = fakeRegistry({ [HUB]: [tooMany(), tooMany(), json({ tags: ['1'] })] });
const client = createHttpRegistryClient({ fetch: fetchFn, rateLimitWaitsMs: [5_000, 30_000], sleep: (ms) => (waits.push(ms), Promise.resolve()) });
expect(await client.listTags('nextcloud')).toEqual(['1']);
expect(waits).toEqual([5_000, 30_000]);
expect(calls).toHaveLength(3);
});
it('waits as long as Retry-After asks when that is longer, up to five minutes', async () => {
const waits: number[] = [];
const { fetchFn } = fakeRegistry({ [HUB]: [tooMany({ 'retry-after': '20' }), tooMany({ 'retry-after': '3600' }), json({ tags: [] })] });
await createHttpRegistryClient({ fetch: fetchFn, rateLimitWaitsMs: [5_000, 30_000], sleep: (ms) => (waits.push(ms), Promise.resolve()) }).listTags('nextcloud');
expect(waits).toEqual([20_000, 300_000]);
});
it('does not ask again unless pauses are given', async () => {
const { fetchFn, calls } = fakeRegistry({ [HUB]: tooMany() });
await createHttpRegistryClient({ fetch: fetchFn }).listTags('nextcloud').catch(() => undefined);
expect(calls).toHaveLength(1);
});
it('gives up as rate_limited once the pauses are used', async () => {
const { fetchFn, calls } = fakeRegistry({ [HUB]: tooMany() });
const error = await createHttpRegistryClient({ fetch: fetchFn, rateLimitWaitsMs: [1] }).listTags('nextcloud').catch((cause: unknown) => cause);
expect((error as RegistryError).kind).toBe('rate_limited');
expect(calls).toHaveLength(2);
});
it('an abort during the pause is rethrown as it is', async () => {
const controller = new AbortController();
const { fetchFn } = fakeRegistry({ [HUB]: tooMany() });
const running = createHttpRegistryClient({ fetch: fetchFn, rateLimitWaitsMs: [5_000] }).listTags('nextcloud', controller.signal);
setTimeout(() => controller.abort(new Error('step time limit')), 20);
await expect(running).rejects.toThrow('step time limit');
});
});
describe('time', () => {
const hanging = (_input: string | URL | Request, init?: RequestInit) =>
new Promise<Response>((_resolve, reject) => init?.signal?.addEventListener('abort', () => reject(init.signal?.reason as Error)));
+26 -1
View File
@@ -1,4 +1,5 @@
import { createHash } from 'node:crypto';
import { setTimeout as delay } from 'node:timers/promises';
import { z } from 'zod';
import { findPlatformEntry, parseImageRef, platformNames } from '../domain/image-ref.ts';
import type { IndexEntry } from '../domain/image-ref.ts';
@@ -14,11 +15,16 @@ export interface HttpRegistryOptions {
readonly maxPages?: number;
/** Largest response body we read (tag pages, manifests, image configs). */
readonly maxBodyBytes?: number;
/** Pauses before asking again after HTTP 429, one retry per entry; none by default. ghcr.io answers 429 now and then and is fine later. */
readonly rateLimitWaitsMs?: readonly number[];
readonly sleep?: (ms: number, signal: AbortSignal | undefined) => Promise<void>;
}
const DEFAULT_TIMEOUT_MS = 30_000;
const DEFAULT_MAX_PAGES = 50;
const DEFAULT_MAX_BODY_BYTES = 4 * 1024 * 1024;
/** A longer `Retry-After` than this is not waited out. */
const MAX_RATE_LIMIT_WAIT_MS = 300_000;
const MANIFEST_ACCEPT = [
'application/vnd.oci.image.index.v1+json',
@@ -58,6 +64,8 @@ export function createHttpRegistryClient(options: HttpRegistryOptions = {}): Reg
const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
const maxPages = options.maxPages ?? DEFAULT_MAX_PAGES;
const maxBodyBytes = options.maxBodyBytes ?? DEFAULT_MAX_BODY_BYTES;
const rateLimitWaitsMs = options.rateLimitWaitsMs ?? [];
const sleep = options.sleep ?? ((ms: number, signal: AbortSignal | undefined) => delay(ms, undefined, signal === undefined ? {} : { signal }));
/** One repository on one registry: the token is fetched once and reused by every request. */
function openSession(location: RepoLocation, signal: AbortSignal | undefined) {
@@ -65,8 +73,25 @@ export function createHttpRegistryClient(options: HttpRegistryOptions = {}): Reg
const what = `${location.registry}/${location.path}`;
let token: string | null = null;
/** One request with the caller's abort and our own time limit. A caller abort is rethrown untouched. */
/** `request`, asked again after HTTP 429 as long as `rateLimitWaitsMs` has pauses left; the last 429 is returned. */
async function request(url: string, authorization: string | null, accept?: string): Promise<Response> {
for (let attempt = 0; ; attempt++) {
const response = await requestOnce(url, authorization, accept);
const wait = rateLimitWaitsMs[attempt];
if (response.status !== 429 || wait === undefined) return response;
const retryAfterMs = Number(response.headers.get('retry-after')) * 1000;
const pause = Number.isFinite(retryAfterMs) && retryAfterMs > wait ? Math.min(retryAfterMs, MAX_RATE_LIMIT_WAIT_MS) : wait;
void response.body?.cancel().catch(() => undefined);
try {
await sleep(pause, signal);
} catch (cause) {
throw signal?.aborted === true ? (signal.reason as Error) : cause;
}
}
}
/** One request with the caller's abort and our own time limit. A caller abort is rethrown untouched. */
async function requestOnce(url: string, authorization: string | null, accept?: string): Promise<Response> {
const timeout = AbortSignal.timeout(timeoutMs);
const combined = signal === undefined ? timeout : AbortSignal.any([signal, timeout]);
const headers: Record<string, string> = {};
+1 -1
View File
@@ -16,7 +16,7 @@ export interface ModelDispatcherOptions {
/**
* Sends each model request to the sidecar of its profile's provider (Claude, Codex). The profile is the pinned
* `profileRef` of a run's snapshot, or else the latest profile named like the request's executor (`sonnet`, `opus`: the
* `profileRef` of a run's snapshot, or else the latest profile named like the request's executor (`haiku`, `sonnet`, `opus`: the
* app's configuration names the model by profile id). There is no default provider and no fallback: an unknown
* profile or a provider without a sidecar is an error, and one provider's failure never moves the request to another.
* The profile's timeout and cost cap apply to the run.
@@ -82,6 +82,58 @@ describe('executor-backed privileged action adapter', () => {
}
});
describe('when the post-upgrade model runs', () => {
const nothingToDo = () => new FakeModelExecutor().reply({ output: { outcome: 'nothing_to_do', summary: 'Nothing to do.', actions: [] }, costUsd: 0.2 });
const withLogs = (text: string) => session().on('sessionLogs', { outcome: 'done', result: { text } });
it('off: no session and no model, the hash comes from the service', async () => {
const client = new FakeExecutorClient()
.on('inspect', { outcome: 'done', result: { project: 'jackett', service: 'jackett', manifest_sha256: 'd'.repeat(64), env_sha256: null, manifest_image: PINNED, container: null } });
const model = nothingToDo();
const result = await createExecutorBackedPrivilegedActionExecutor({ executor: client, modelExecutor: model, prompts: prompts() }).execute({ ...postUpgrade, when: 'off', reasons: ['major'] });
expect(result).toEqual({ ...target, action: 'model.postUpgrade', version: 1, state: 'completed', manifestChanged: false, manifestSha256: 'd'.repeat(64), reasons: ['major'], skipped: 'off' });
expect(model.requests).toEqual([]);
expect(client.calls.map((call) => call.op)).toEqual(['inspect']);
});
it('when needed, with nothing in the run asking and clean logs: reads the logs and stops there', async () => {
const client = withLogs('[migrations] no migrations found\nStarted.\n');
const model = nothingToDo();
const result = await createExecutorBackedPrivilegedActionExecutor({ executor: client, modelExecutor: model, prompts: prompts() }).execute({ ...postUpgrade, when: 'when_needed', reasons: [] });
expect(result).toMatchObject({ state: 'completed', manifestChanged: false, manifestSha256: 'c'.repeat(64), reasons: [], skipped: 'not_needed' });
expect(result).not.toHaveProperty('report');
expect(model.requests).toEqual([]);
expect(client.calls.map((call) => [call.op, call.request])).toEqual([
['openSession', { project: 'jackett', service: 'jackett', mode: 'observe' }],
['sessionLogs', { token: 'tok', service: 'jackett', tail: 200 }],
['closeSession', { token: 'tok' }],
]);
});
it('when needed, errors in the logs (or logs it cannot read) bring the model in', async () => {
for (const client of [withLogs('Started.\nERROR Unable to open database\n'), session().on('sessionLogs', { outcome: 'failed', detail: 'no container', manifest_replaced: false })]) {
const model = nothingToDo();
const result = await createExecutorBackedPrivilegedActionExecutor({ executor: client, modelExecutor: model, prompts: prompts() }).execute({ ...postUpgrade, when: 'when_needed', reasons: [] });
expect(result).toMatchObject({ state: 'completed', reasons: ['log_errors'], report: { outcome: 'nothing_to_do', costUsd: 0.2 } });
expect(result).not.toHaveProperty('skipped');
expect(model.requests).toHaveLength(1);
}
});
it('when needed, a reason from the run brings the model in without reading the logs first', async () => {
const client = session();
const model = nothingToDo();
const result = await createExecutorBackedPrivilegedActionExecutor({ executor: client, modelExecutor: model, prompts: prompts() }).execute({ ...postUpgrade, when: 'when_needed', reasons: ['approved_update', 'major'] });
expect(result).toMatchObject({ state: 'completed', reasons: ['approved_update', 'major'], report: { outcome: 'nothing_to_do' } });
expect(model.requests).toHaveLength(1);
expect(client.calls.map((call) => call.op)).not.toContain('sessionLogs');
});
});
it('takes the manifest hash from the service when the session gave no summary', async () => {
const client = session()
.on('closeSession', { outcome: 'failed', detail: 'gone', manifest_replaced: false })
@@ -169,6 +221,18 @@ describe('executor-backed privileged action adapter', () => {
expect(store.stored).toEqual({});
});
it('sends no compose file when the plan keeps the current one', async () => {
const client = new FakeExecutorClient().on('migrate', {
outcome: 'done',
result: { project: 'grafana', service: 'influxdb', state: 'applied', image: PINNED, manifest_sha256: 'a'.repeat(64), applied_manifest_sha256: 'b'.repeat(64), env_keys: [] },
});
const adapter = createExecutorBackedPrivilegedActionExecutor({ executor: client, runInputs: inputs({}) });
await adapter.execute({ ...request, composeYaml: null, inputs: [] }, undefined, { instanceId: 'run-1', nodeAttemptOrdinal: 4 });
expect(client.calls[0]?.request).toMatchObject({ compose_yaml: null, env: {} });
});
it('refuses to run with a value missing, and keeps the values when the executor refused', async () => {
const client = new FakeExecutorClient().on('migrate', { outcome: 'refused', reason: 'invalid_manifest', detail: 'bad' });
await expect(createExecutorBackedPrivilegedActionExecutor({ executor: client, runInputs: inputs({ ADMIN_USER: 'admin' }) }).execute(request, undefined, { instanceId: 'run-1', nodeAttemptOrdinal: 4 }))
@@ -6,7 +6,8 @@ import type { SessionSummary } from '../domain/executor-api.ts';
import { watchModelSession } from '../app/model-activity.ts';
import type { ModelActivitySink } from '../app/model-activity.ts';
import { REPAIR_ANSWER_SCHEMA, RepairAnswer } from '../domain/repair.ts';
import { POST_UPGRADE_ANSWER_SCHEMA, PostUpgradeAnswer } from '../domain/post-upgrade.ts';
import { POST_UPGRADE_ANSWER_SCHEMA, PostUpgradeAnswer, suspiciousLogLines } from '../domain/post-upgrade.ts';
import type { PostUpgradeReason } from '../domain/post-upgrade.ts';
/** Drops closing tags the model sometimes leaves at the end of its summary (`…</summary>\n</invoke>`). */
function cleanSummary(summary: string): string {
@@ -15,6 +16,8 @@ function cleanSummary(summary: string): string {
/** The session log operations a repair result may list (ModelRepairResult.actions). */
const SESSION_OPERATIONS: ReadonlySet<string> = new Set(['ps', 'logs', 'compose', 'write', 'pull', 'up', 'restart', 'exec-rules', 'exec', 'run']);
/** How many of the service's last log lines a post-upgrade check scans before it decides whether a model looks. */
const POST_UPGRADE_LOG_LINES = 200;
/** The closed driver-side mapping. It owns no command construction and only the executor client can reach Docker. */
export function createExecutorBackedPrivilegedActionExecutor(deps: {
@@ -84,6 +87,31 @@ export function createExecutorBackedPrivilegedActionExecutor(deps: {
return { output, costUsd, failure, model, manifestChanged: summary?.manifest_changed ?? false, manifestSha256, actions };
}
/**
* Whether a post-upgrade check needs its model, and why. When needed only: a reason from the run, else the service's
* last log lines read in an observe session (no model); errors there, or logs it cannot read, bring the model in.
*/
async function postUpgradeNeed(
request: Extract<PrivilegedActionRequest, { readonly action: 'model.postUpgrade' }>,
signal?: AbortSignal,
): Promise<{ readonly run: true; readonly reasons: readonly PostUpgradeReason[] } | { readonly run: false; readonly skipped: 'off' | 'not_needed'; readonly manifestSha256: string | null }> {
const reasons = request.reasons ?? [];
if (request.when === 'off') return { run: false, skipped: 'off', manifestSha256: null };
if (request.when !== 'when_needed' || reasons.length > 0) return { run: true, reasons };
const session = completed(request.action, await deps.executor.openSession({ project: request.project, service: request.service, mode: 'observe' }, signal));
const close = () => deps.executor.closeSession({ token: session.token }, signal);
let logs: ExecutorCall<{ readonly text: string }>;
try {
logs = await deps.executor.sessionLogs({ token: session.token, service: request.service, tail: POST_UPGRADE_LOG_LINES }, signal);
} catch (cause) {
await close().catch(() => undefined);
throw cause;
}
const closed = await close();
if (logs.outcome !== 'done' || suspiciousLogLines(logs.result.text).length > 0) return { run: true, reasons: ['log_errors'] };
return { run: false, skipped: 'not_needed', manifestSha256: closed.outcome === 'done' ? closed.result.manifest_sha256 : null };
}
return {
async execute(request, signal, run) {
const parsed = PrivilegedActionRequest.parse(request);
@@ -141,9 +169,10 @@ export function createExecutorBackedPrivilegedActionExecutor(deps: {
const values = deps.runInputs.values(run.instanceId);
const missing = parsed.inputs.filter((input) => values[input.name] === undefined).map((input) => input.name);
if (missing.length > 0) throw new Error(`docker.migrate is missing the owner's values for ${missing.join(', ')}`);
const composeYaml = parsed.inputs
const plannedCompose = parsed.composeYaml;
const composeYaml = plannedCompose === null ? null : parsed.inputs
.filter((input) => input.target === 'compose')
.reduce((text, input) => text.replaceAll(`<input:${input.name}>`, values[input.name] ?? ''), parsed.composeYaml);
.reduce((text, input) => text.replaceAll(`<input:${input.name}>`, values[input.name] ?? ''), plannedCompose);
const env = Object.fromEntries(parsed.inputs.filter((input) => input.target === 'env').map((input) => [input.name, values[input.name] ?? '']));
const answer = await deps.executor.migrate({
project: parsed.project, service: parsed.service, image: parsed.image, expect_manifest_sha256: parsed.expectedManifestSha256,
@@ -195,6 +224,17 @@ export function createExecutorBackedPrivilegedActionExecutor(deps: {
break;
}
case 'model.postUpgrade': {
const need = await postUpgradeNeed(parsed, signal);
if (!need.run) {
const manifestSha256 = need.manifestSha256
?? completed(parsed.action, await deps.executor.inspect({ project: parsed.project, service: parsed.service }, signal)).manifest_sha256;
result = {
action: parsed.action, version: 1, component: parsed.component, project: parsed.project, service: parsed.service,
image: parsed.image, backupId: parsed.backupId, state: 'completed', manifestChanged: false, manifestSha256,
reasons: parsed.reasons ?? [], skipped: need.skipped,
};
break;
}
const asked = await ask(parsed, 'postUpgrade', signal, run);
const answer = PostUpgradeAnswer.safeParse(asked.output);
const report = answer.success
@@ -208,6 +248,7 @@ export function createExecutorBackedPrivilegedActionExecutor(deps: {
state: done ? 'completed' : 'action_required',
manifestChanged: asked.manifestChanged, manifestSha256: asked.manifestSha256,
report: { ...report, model: asked.model, costUsd: asked.costUsd },
reasons: need.reasons,
};
break;
}
+4 -3
View File
@@ -21,16 +21,16 @@ export function createSqliteAppConfigStore(db: DatabaseSync): VersionedAppConfig
const previousConfig = previous === undefined ? null : parseAppConfig(previous);
db.prepare(
`INSERT INTO app_config (app, enabled, position, service, image_repo, tag_pattern, auto, http_url, expect_status,
release_notes_executor, migrate_executor, prompt_hint, post_upgrade_executor, notes, apply, project)
release_notes_executor, migrate_executor, prompt_hint, post_upgrade_executor, post_upgrade_when, notes, apply, project)
VALUES (:app, :enabled, :position, :service, :imageRepo, :tagPattern, :auto, :httpUrl, :expectStatus,
:releaseNotesExecutor, :migrateExecutor, :promptHint, :postUpgradeExecutor, :notes, :apply, :project)
:releaseNotesExecutor, :migrateExecutor, :promptHint, :postUpgradeExecutor, :postUpgradeWhen, :notes, :apply, :project)
ON CONFLICT (app) DO UPDATE SET
enabled = excluded.enabled, position = excluded.position, service = excluded.service,
image_repo = excluded.image_repo, tag_pattern = excluded.tag_pattern, auto = excluded.auto,
http_url = excluded.http_url, expect_status = excluded.expect_status,
release_notes_executor = excluded.release_notes_executor, migrate_executor = excluded.migrate_executor,
prompt_hint = excluded.prompt_hint,
post_upgrade_executor = excluded.post_upgrade_executor, notes = excluded.notes, apply = excluded.apply, project = excluded.project`,
post_upgrade_executor = excluded.post_upgrade_executor, post_upgrade_when = excluded.post_upgrade_when, notes = excluded.notes, apply = excluded.apply, project = excluded.project`,
).run({
app: config.app,
enabled: config.enabled ? 1 : 0,
@@ -45,6 +45,7 @@ export function createSqliteAppConfigStore(db: DatabaseSync): VersionedAppConfig
migrateExecutor: config.migrateExecutor,
promptHint: config.promptHint,
postUpgradeExecutor: config.postUpgradeExecutor,
postUpgradeWhen: config.postUpgradeWhen,
notes: config.notes,
apply: config.apply ? 1 : 0,
project: config.project,
+14 -1
View File
@@ -20,7 +20,7 @@ describe('migrate', () => {
]);
expect(columns(db, 'app_config')).toEqual([
'app', 'enabled', 'position', 'service', 'image_repo', 'tag_pattern', 'auto', 'http_url', 'expect_status',
'release_notes_executor', 'migrate_executor', 'prompt_hint', 'notes', 'apply', 'project', 'post_upgrade_executor',
'release_notes_executor', 'migrate_executor', 'prompt_hint', 'notes', 'apply', 'project', 'post_upgrade_executor', 'post_upgrade_when',
]);
expect(columns(db, 'workflows')).toEqual(['id', 'name', 'created_at']);
expect(columns(db, 'workflow_drafts')).toEqual(['workflow_id', 'revision', 'document', 'validation_revision', 'validation_document', 'diagnostics']);
@@ -120,6 +120,19 @@ describe('migrate', () => {
expect(db.prepare('PRAGMA user_version').get()).toEqual({ user_version: 19 });
});
it('v26 accepts haiku as an app model and keeps every app config verbatim', () => {
const db = new DatabaseSync(':memory:');
migrate(db, MIGRATIONS.slice(0, 25));
db.exec(`INSERT INTO app_config (app, service, image_repo, release_notes_executor, migrate_executor, post_upgrade_executor, post_upgrade_when, notes, apply, project)
VALUES ('bazarr', 'bazarr', 'lscr.io/bazarr', 'opus', 'sonnet', NULL, 'always', 'n', 1, 'arr')`);
const before = db.prepare('SELECT * FROM app_config').all();
migrate(db, MIGRATIONS.slice(0, 26));
expect(db.prepare('SELECT * FROM app_config').all()).toEqual(before);
db.exec("INSERT INTO app_config (app, service, image_repo, migrate_executor, post_upgrade_executor) VALUES ('new', 'new', 'r', 'haiku', 'haiku')");
expect(db.prepare("SELECT release_notes_executor FROM app_config WHERE app = 'new'").get()).toEqual({ release_notes_executor: 'haiku' });
expect(() => db.exec("UPDATE app_config SET release_notes_executor = 'gpt' WHERE app = 'new'")).toThrow(/CHECK/);
});
it('v7 adds a nullable progress to steps already there', () => {
const db = new DatabaseSync(':memory:');
migrate(db, MIGRATIONS.slice(0, 6));
+40
View File
@@ -802,6 +802,46 @@ export const MIGRATIONS: readonly Migration[] = [
) STRICT;
`,
},
{
// When an app's post-upgrade check runs its model (owner, 2026-10-07): by default only when the update may ask for something.
version: 25,
sql: `
ALTER TABLE app_config ADD COLUMN post_upgrade_when TEXT NOT NULL DEFAULT 'when_needed' CHECK (post_upgrade_when IN ('when_needed', 'always'));
`,
},
{
// Haiku as a model choice and the default release notes model (owner, 2026-10-07). SQLite cannot change a CHECK in
// place, so the table is rebuilt; the retired steps table keeps its old one, nothing writes it any more.
version: 26,
sql: `
ALTER TABLE app_config RENAME TO app_config_v25;
CREATE TABLE app_config (
app TEXT PRIMARY KEY,
enabled INTEGER NOT NULL DEFAULT 1 CHECK (enabled IN (0, 1)),
position INTEGER NOT NULL DEFAULT 0,
service TEXT NOT NULL,
image_repo TEXT NOT NULL,
tag_pattern TEXT NOT NULL DEFAULT '.',
auto TEXT NOT NULL DEFAULT 'patch' CHECK (auto IN ('none', 'patch', 'minor', 'major')),
http_url TEXT,
expect_status INTEGER NOT NULL DEFAULT 200,
release_notes_executor TEXT NOT NULL DEFAULT 'haiku' CHECK (release_notes_executor IN ('script', 'haiku', 'sonnet', 'opus')),
migrate_executor TEXT CHECK (migrate_executor IN ('script', 'haiku', 'sonnet', 'opus')),
prompt_hint TEXT,
notes TEXT,
apply INTEGER NOT NULL DEFAULT 0 CHECK (apply IN (0, 1)),
project TEXT NOT NULL DEFAULT '',
post_upgrade_executor TEXT CHECK (post_upgrade_executor IN ('script', 'haiku', 'sonnet', 'opus')),
post_upgrade_when TEXT NOT NULL DEFAULT 'when_needed' CHECK (post_upgrade_when IN ('when_needed', 'always'))
) STRICT;
INSERT INTO app_config (app, enabled, position, service, image_repo, tag_pattern, auto, http_url, expect_status,
release_notes_executor, migrate_executor, prompt_hint, notes, apply, project, post_upgrade_executor, post_upgrade_when)
SELECT app, enabled, position, service, image_repo, tag_pattern, auto, http_url, expect_status,
release_notes_executor, migrate_executor, prompt_hint, notes, apply, project, post_upgrade_executor, post_upgrade_when
FROM app_config_v25;
DROP TABLE app_config_v25;
`,
},
];
function schemaVersion(db: DatabaseSync): number {
+3 -1
View File
@@ -1,5 +1,5 @@
import { z } from 'zod';
import { AutoLevel, Executor } from '../../domain/steps.ts';
import { AutoLevel, Executor, PostUpgradeWhen } from '../../domain/steps.ts';
import type { AppConfig, JsonValue } from '../../domain/steps.ts';
// Rows are validated where they leave the database, so a corrupted or foreign row fails loudly here
@@ -20,6 +20,7 @@ const AppConfigRow = z
migrate_executor: Executor.nullable(),
prompt_hint: z.string().nullable(),
post_upgrade_executor: Executor.nullable(),
post_upgrade_when: PostUpgradeWhen,
notes: z.string().nullable(),
apply: z.number(),
project: z.string().min(1),
@@ -40,6 +41,7 @@ const AppConfigRow = z
migrateExecutor: row.migrate_executor,
promptHint: row.prompt_hint,
postUpgradeExecutor: row.post_upgrade_executor,
postUpgradeWhen: row.post_upgrade_when,
notes: row.notes,
apply: row.apply === 1,
}),
@@ -119,7 +119,7 @@ async function setupEnvironment() {
releaseNotesExecutor: 'sonnet',
migrateExecutor: null,
promptHint: null,
postUpgradeExecutor: null,
postUpgradeExecutor: null, postUpgradeWhen: 'when_needed',
notes: null,
apply: true,
});
+1 -1
View File
@@ -8,7 +8,7 @@ const one = { app: 'bazarr', service: 'bazarr', imageRepo: 'linuxserver/bazarr',
describe('parseAppsFile', () => {
it('fills the defaults: enabled, models off, real updates off, positions from the file order', () => {
const [first, second] = parseAppsFile([one, { ...one, app: 'sonarr', service: 'sonarr', enabled: false, notes: 'x' }]);
expect(first).toEqual({ ...one, enabled: true, position: 1, releaseNotesExecutor: 'sonnet', migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, notes: null, apply: false, project: 'bazarr' });
expect(first).toEqual({ ...one, enabled: true, position: 1, releaseNotesExecutor: 'haiku', migrateExecutor: null, promptHint: null, postUpgradeExecutor: null, postUpgradeWhen: 'when_needed', notes: null, apply: false, project: 'bazarr' });
expect(second).toMatchObject({ app: 'sonarr', enabled: false, position: 2, notes: 'x' });
});
+3 -2
View File
@@ -1,6 +1,6 @@
import { z } from 'zod';
import { parseProjectName } from '../domain/compose.ts';
import { AutoLevel, Executor } from '../domain/steps.ts';
import { AutoLevel, DEFAULT_MODEL, Executor, PostUpgradeWhen } from '../domain/steps.ts';
import type { AppConfig } from '../domain/steps.ts';
import { compileTagPattern } from '../domain/tags.ts';
@@ -19,10 +19,11 @@ const AppEntry = z.strictObject({
auto: AutoLevel,
httpUrl: z.string().nullable().default(null),
expectStatus: z.number().int().min(100).max(599).default(200),
releaseNotesExecutor: Executor.default('sonnet'),
releaseNotesExecutor: Executor.default(DEFAULT_MODEL),
migrateExecutor: Executor.nullable().default(null),
promptHint: z.string().nullable().default(null),
postUpgradeExecutor: Executor.nullable().default(null),
postUpgradeWhen: PostUpgradeWhen.default('when_needed'),
notes: z.string().nullable().default(null),
/** Real updates stay off until the owner turns them on in the app. */
apply: z.boolean().default(false),
+3 -1
View File
@@ -42,6 +42,8 @@ import { missingSkills, readSkillFiles } from './skills-file.ts';
import { createWorkflowExtensions } from '../domain/workflow-extensions.ts';
const logger = createConsoleLogger();
/** ghcr.io answers 429 now and then; the schedule is spread out, so a run can wait minutes for the tag list. */
const REGISTRY_RATE_LIMIT_WAITS_MS = [30_000, 120_000, 300_000];
let config;
try {
@@ -118,7 +120,7 @@ const runner = createWorkflowRunner({
functionCatalog,
functionRuntime: createFunctionRuntimeClient({ baseUrl: config.functionRuntimeUrl }),
componentOperations: createComponentOperationExecutor({
registry: createHttpRegistryClient(),
registry: createHttpRegistryClient({ rateLimitWaitsMs: REGISTRY_RATE_LIMIT_WAITS_MS }),
hostBackup,
backupVmid: config.backupVmid,
clock: systemClock,
+6
View File
@@ -17,6 +17,7 @@ import type {
SessionOpenRequest,
SessionOpenResult,
SessionLogResult,
AgentLogsRequest,
AgentTextResult,
SessionSummary,
} from '../domain/executor-api.ts';
@@ -135,6 +136,7 @@ interface ExecutorOps {
closeSession: [SessionCloseRequest, SessionSummary];
sessionLog: [SessionCloseRequest, SessionLogResult];
sessionCompose: [SessionCloseRequest, AgentTextResult];
sessionLogs: [SessionCloseRequest & AgentLogsRequest, AgentTextResult];
}
type ExecutorOp = keyof ExecutorOps;
type ExecutorScript<K extends ExecutorOp> = (request: ExecutorOps[K][0]) => ExecutorCall<ExecutorOps[K][1]>;
@@ -204,4 +206,8 @@ export class FakeExecutorClient implements ExecutorClient {
sessionCompose(request: SessionCloseRequest) {
return this.#call('sessionCompose', request);
}
sessionLogs(request: SessionCloseRequest & AgentLogsRequest) {
return this.#call('sessionLogs', request);
}
}
+1 -1
View File
@@ -25,7 +25,7 @@ export function seedTestApps(appConfig: AppConfigStore): void {
releaseNotesExecutor: 'sonnet',
migrateExecutor: null,
promptHint: null,
postUpgradeExecutor: null,
postUpgradeExecutor: null, postUpgradeWhen: 'when_needed',
notes: null,
apply: false,
});
+12
View File
@@ -206,6 +206,18 @@ describe('app configuration', () => {
await eventually(() => expect((requests.filter((r) => r.method === 'POST').at(-1)!.body as Record<string, unknown>)['postUpgradeExecutor']).toBeNull());
});
it('sets when the post-upgrade model runs, once a model is chosen', async () => {
await open();
expect(form()!.querySelector('[name="postUpgradeWhen"]')).toBeNull();
await choose(field('postUpgradeExecutor'), 'sonnet');
expect(field<HTMLSelectElement>('postUpgradeWhen').value).toBe('when_needed');
await choose(field('postUpgradeWhen'), 'always');
button(form()!, 'Save').click();
await eventually(() => expect(form()!.textContent).toContain('Saved'));
expect(requests.find((r) => r.method === 'POST')!.body).toEqual({ postUpgradeExecutor: 'sonnet', postUpgradeWhen: 'always' });
expect((await server.inject('/api/apps/bazarr/config')).json()).toMatchObject({ postUpgradeWhen: 'always' });
});
it('checks the change in the browser before sending it', async () => {
await open();
await type(field('tagPattern'), '(?P<v>x)');
+7 -4
View File
@@ -17,6 +17,7 @@ interface Draft {
releaseNotesExecutor: string;
migrateExecutor: string;
postUpgradeExecutor: string;
postUpgradeWhen: string;
promptHint: string;
notes: string;
}
@@ -31,6 +32,7 @@ const toDraft = (config: AppConfigView): Draft => ({
releaseNotesExecutor: config.releaseNotesExecutor,
migrateExecutor: config.migrateExecutor ?? 'off',
postUpgradeExecutor: config.postUpgradeExecutor ?? 'off',
postUpgradeWhen: config.postUpgradeWhen,
promptHint: config.promptHint ?? '',
notes: config.notes ?? '',
});
@@ -102,7 +104,7 @@ function ConfigForm({ initial }: { initial: AppConfigView }) {
<input type="text" name={key} value={draft[key]} disabled={pending} onInput={(event) => set(key, event.currentTarget.value)} />
</label>
);
const select = (key: 'auto' | 'releaseNotesExecutor' | 'migrateExecutor' | 'postUpgradeExecutor', label: string, options: readonly string[]) => (
const select = (key: 'auto' | 'releaseNotesExecutor' | 'migrateExecutor' | 'postUpgradeExecutor' | 'postUpgradeWhen', label: string, options: readonly string[]) => (
<label>
{label}
<select name={key} value={draft[key]} disabled={pending} onChange={(event) => set(key, event.currentTarget.value)}>
@@ -153,9 +155,10 @@ function ConfigForm({ initial }: { initial: AppConfigView }) {
</div>
{text('httpUrl', 'Check URL', 'empty = none')}
{text('expectStatus', 'Expected HTTP status')}
{select('releaseNotesExecutor', 'Release notes executor', ['sonnet', 'opus'])}
{select('migrateExecutor', 'Repair executor', ['off', 'sonnet', 'opus'])}
{select('postUpgradeExecutor', 'Post-upgrade check executor', ['off', 'sonnet', 'opus'])}
{select('releaseNotesExecutor', 'Release notes executor', ['haiku', 'sonnet', 'opus'])}
{select('migrateExecutor', 'Repair executor', ['off', 'haiku', 'sonnet', 'opus'])}
{select('postUpgradeExecutor', 'Post-upgrade check executor', ['off', 'haiku', 'sonnet', 'opus'])}
{draft.postUpgradeExecutor === 'off' ? null : select('postUpgradeWhen', 'Post-upgrade check runs', ['when_needed', 'always'])}
{text('promptHint', 'Prompt hint')}
{text('notes', 'Notes', 'up to 10 words')}
<div class="actions">
+1 -1
View File
@@ -176,7 +176,7 @@ export function ModelsView() {
</button>
</div>
<p class="muted">
Apps pick a model as <code>sonnet</code> or <code>opus</code>: the profile of that name decides which sidecar runs it
Apps pick a model as <code>haiku</code>, <code>sonnet</code> or <code>opus</code>: the profile of that name decides which sidecar runs it
(Provider: Claude or Codex), how long it may take and what it may cost. A run keeps the profile version it started with.
</p>