shc.yaml (design)
Status: built (SHC-1361). The author reference is shc.yaml. The loader (ce
modules/app/shcyaml.go,LoadShcYAML) reads a rendered shc.yaml, refuses every key this table does not name, and writes the compose document the compose path renders; an install reads shc.yaml whenever an app has one, andshc convertwrites one from a compose template. Clay, 2026-10-02: an app may be its own file, “just an shc.yaml file”, and “it wouldn’t respect the compose file if you did that.” The schema “doesn’t do all the backwards compatibility stuff”; “you can only do expose for ports, it has to be explicit the way our schema is, and it doesn’t do the x-shc stuff.” This note is that schema. It is written against cee8398c2e232c, the commit shc next pinned on 2026-10-04 (it has moved on since), and it edits neither the replica project contract nor groups.
A compose file stays the bring-your-own and import path, with its enrichment and its x-shc-* keys. This file is the other way to write an app, and the recommended one. The loader’s refusals are X400530SHYUNK (not a key), X400531SHYCMP (a compose key), X400532SHYXSH (an x-shc key), X400533SHYPRT (a host side on a port), X400534SHYHST (a host path), X400535SHYVAL (a value), X400536SHYPRV (the privileged class outside privileges:) and X400537SHYDOC (not one mapping). Each names the service and the key.
1. The file
Section titled “1. The file”An app directory is what it is now (app.yaml, vars.yaml, backup.yaml, plugs, sockets, jobs) with shc.yaml in place of compose.yaml. When shc.yaml is present the compose file is not read. Both files produce the same stored specification, so there is one renderer.
That specification is deploymentspec.Spec (modules/deploymentspec/spec.go), which SpecFromRender records (modules/app/spec_record.go). Each replica is a sealed compose document plus a pod.Record (core/pod/record.go). Spec.Services is the data mounts and the scale ceiling (modules/app/spec_services.go). Publishing is not a key of this file. It stays the deployment flag, --expose (the new line).
The file is a Jinja2 template, strict on undefined names, and a group is a {% members %} block. That block is the one way to write a group. There is no groups key. The tag, the stored member list (pod.Member, kept on the manifest as groups) and the scale grammar are the groups note. This note only uses them.
The schema is read on the rendered document, which is the document the spec is built from. A key that an untaken {% if %} never wrote is not there, same as today.
2. Where a row comes from
Section titled “2. Where a row comes from”Every row is one of these, at the pin:
- a field the stored spec keeps, or a key the replica split reads into
pod.Record - a directive in
xshcRegistry(modules/app/extension_lint.go) that an app may declare - a service key in the incus executor’s
serviceKeys(core/executor/incus/translate.go), which is the closed set a linked runtime runs, exceptportsanddeploy read_onlyandtty, which the render leaves on the service map it stores and which a flagship compose writes, and which incus then refuses on that replica (Replica.Refused)
ports is not a key. Listening ports are expose. deploy is not a key. Its contents are replicas, quota and restart on the service, and where at the top level.
Every service key a catalog app writes and SHC honours has a row, so every catalog app can be written as shc.yaml (main’s correction A, from the scan of 110 compose files on the ticket). A compose key no catalog app writes stays out until one needs it.
x-shc-quota and x-shc-restart are reserved: the render writes them and refuses an authored one. The author writes quota and restart instead. Three strings in the ticket’s 2026-10-02 list are not directives at this pin. x-shc-bak is the backup archive content type. x-shc-volume-prep is the comment on volume.PrepOp; the mount tail that builds that op is the volume fields below. x-shc-conn is the worked example of a bad near-miss in nearestExtension, not a key anyone decodes.
A key this table does not name is an error, including an x-shc-* key, a compose key the table does not list, and a host path. Nothing is dropped and nothing is rewritten into a different meaning. The render may still stamp platform labels, set restart: "no" on the stored document, and write the quota marker. That is the renderer, not a second reading of the author’s key.
service. means any service. volume. means any entry of that service’s volumes. secret. and config. mean any entry of the top-level mappings.
3. The schema
Section titled “3. The schema”| Key | Type | Meaning | Replaces |
|---|---|---|---|
services | mapping of name to service | The services the render groups into pods and seals as replica projects. | services |
secrets | mapping of name to secret | Secrets the document names. A service lists the names it mounts. | secrets |
secret.vault | string ref+vault://… | The secret’s only source. The value is not in the file. file and environment sources stay refused. | x-shc-vault |
configs | mapping of name to config | Inline configs. A service lists the names it consumes. | configs |
config.content | string | The config bytes, stored inline. | configs.<name>.content |
config.name | string, optional | Engine name of that config. | configs.<name>.name |
where | list of where term | The deployment’s where. Each entry is one term of the selector --where reads: node=<hostname>, <label>=<value>, <label>!=<value>. The terms are sorted, kept once each, joined by & and stored as package_rules, as a Swarm constraint is now. node.role has no where form. | deploy.placement.constraints |
service.image | string | The image the replica runs. | image |
service.command | string or list of string | command | |
service.entrypoint | string or list of string | entrypoint | |
service.working_dir | string | working_dir | |
service.user | string | user | |
service.hostname | string | hostname | |
service.domainname | string | domainname | |
service.environment | mapping of string to string | A list of KEY=value is an error. The author writes the mapping. | environment |
service.env_file | list of string | Paths inside the package. The replica project rewrites them, as it does now. | env_file |
service.expose | list of port | Listening ports only. An entry is an integer, or a string that is a port or port/protocol. A quoted port is a port, because a template has to quote a number. There is no host:container form. Members that declare a port are the replica’s ready list (core/pod/project.go). | expose and ports |
service.labels | mapping of string to string | Plain labels only. A key starting with shc. is an error (X400357LBLSHC). A list of k=v is an error. The platform stamps its own labels on the stored document and overwrites these names. | labels and deploy.labels |
service.networks | list of name, or mapping | Networks the service joins. A mapping may set aliases (a list of string); the render still appends <env>.<service>.<stack>. An authored address is not a key: on the fabric the render writes the pin. | networks |
service.depends_on | list of name, or mapping of name to dependency | Inside the pod it stays in the replica document. Onto another pod it becomes pod.Record.After. A list means condition started, required. | depends_on |
depends_on.condition | started, healthy or completed | The words the record stores. | service_started, service_healthy, service_completed_successfully |
depends_on.required | bool | Default true. Required false on a service the app does not have is dropped, as today. | depends_on.<service>.required |
depends_on.restart | bool | Restart this replica after the dependency is recreated. Default false. | depends_on.<service>.restart |
service.healthcheck | mapping | The check the runtime runs, stored on the replica document. | healthcheck |
healthcheck.test | string or list of string | The incus translator reads this field. | healthcheck.test |
healthcheck.interval | duration | healthcheck.interval | |
healthcheck.timeout | duration | healthcheck.timeout | |
healthcheck.retries | integer | healthcheck.retries | |
healthcheck.start_period | duration | healthcheck.start_period | |
healthcheck.disable | bool | healthcheck.disable | |
service.health.restart_after | duration, a number of seconds, or 0 | How long the container may stay unhealthy before its node restarts it. Stored as the label shc.health.restart_after. 0 disables. | x-shc-health.restart_after |
service.restart | no, always, on-failure or unless-stopped | The policy the node agent applies. Absent means unless-stopped. The stored document gets restart: "no" and the render writes x-shc-restart. | restart and deploy.restart_policy.condition |
service.replicas | integer | How many replica projects this pod gets when --scale does not name it. Absent means 1. Two members of one pod that disagree are refused. Removed from the replica document, as deploy.replicas and scale are now. | deploy.replicas and the integer scale |
service.scale.max | integer, 1 or more | Replica ceiling. deploymentspec.Service.MaxScale. A pod’s ceiling is the smallest its members state. | x-shc-scale.max |
service.scale.backup | integer | Replica count while a backup runs, usually 0 or 1. Absent leaves the service as it is; keep, min and max are refused (Clay, 2026-10-07). | x-shc-scale.backup |
service.quota.cpu | cpu quantity | The cpu quota. The render sets the container limit to quota times deploy.limit_ratio (1.5 unless configured) and writes x-shc-quota. The author does not write that marker, nor cpu_source. | deploy.resources.limits.cpus |
service.quota.memory | memory quantity | The memory quota, same rule. | deploy.resources.limits.memory |
service.quota.cpu_reservation | cpu quantity, optional | What this member reserves, when the author writes it. Above the quota it is lowered, as today. Absent means quota times the reservation ratio. Summed into pod.Record.Reserve. | deploy.resources.reservations.cpus |
service.quota.memory_reservation | memory quantity, optional | Same, for memory. | deploy.resources.reservations.memory |
service.disruption.max_unavailable | integer, 0 or more | How many replicas a drain may take down at once. | x-shc-disruption.max_unavailable |
service.primary | bool, or an integer port 1 to 65535 | The service a bare publish of a hostname routes to. A port pins which listening port. A quoted port is an error: it would elect nothing. | x-shc-primary |
service.path | absolute URL path, not / | The path this service serves under the primary’s hostname. No ?, #, !, backtick or control character. | x-shc-path |
service.path_strip | bool | The edge strips that path before dialling. True with no path is an error, as today. | x-shc-path-strip |
service.lifecycle | one-shot | A run-once service. The only value. The render stamps shc.lifecycle=one-shot and strips the key before the engine sees the document. | x-shc-lifecycle |
service.volumes | list of volume | Mounts. A named SHC volume, a package file, or a platform mount. No host path. | volumes and top-level x-shc-volumes |
volume.name | string | The volume’s name, deploymentspec.Mount.Name. Required when there is no source. The top-level volume declaration goes away; the render declares it. | the source of a named volume, and the key under x-shc-volumes |
volume.path | absolute path | Where it is mounted in the container. | the target of a volumes entry |
volume.source | ./… or /etc/shc/<name> | A package file, or a platform mount. Any other absolute path is a host path and an error. The four inert host files compose still allows are not keys here. | a bind source |
volume.read_only | bool | the ro token | |
volume.mode | octal string | Owner mode the prep step applies. | the perm field of ParseMount (core/volume/parse.go) |
volume.owner | string uid:gid | Owner the prep step applies, both numeric or both names. | the uid and gid fields of ParseMount |
volume.noinit | bool | Do not seed the mount from the image. | the noinit flag of ParseMount |
volume.size | size quantity | Block volume size. Only on a named volume. A size that resolves to none is an error. | x-shc-volumes.<name>.size |
volume.access | rwo or rox | Default rwo. Any other token is an error. Only on a named volume. | x-shc-volumes.<name>.access |
service.tmpfs | string or list of string | A tmpfs mount. It is not a volume: it has no host side. | tmpfs |
service.shm_size | string or integer | shm_size | |
service.cap_add | list of string | Capabilities in the runtime’s default set. An escalating one is privileges.cap_add. | cap_add |
service.security_opt | list of string | Options that keep confinement on. An unconfined one is privileges.security_opt. | security_opt |
service.privileges | mapping | The one block that asks for the privileged class core/privcap defines. It only asks: the install is refused unless the tenant grant (privilege.grants.<tenant>), the deployment grant or the package’s waiver admits it. | |
privileges.docker_socket | true, false or read_only | The host Docker socket, bound at /var/run/docker.sock. The socket written as a mount source is an error naming this key. | a docker.sock bind |
privileges.privileged | bool | privileged | |
privileges.cap_add | list of string | Capabilities outside the runtime’s default set. A default one here is an error naming cap_add. | cap_add |
privileges.security_opt | list of string | Unconfined options. A confining one here is an error naming security_opt. | security_opt |
privileges.devices | list of string | Host devices. A /dev path written as a mount source is an error naming this key. | devices |
service.extra_hosts | list of string, or mapping | extra_hosts | |
service.pull_policy | string | pull_policy | |
service.group_add | list of string | group_add | |
service.oci_runtime | string | The OCI runtime, such as nvidia. Not SHC’s runtime. | runtime |
service.live_labels | mapping of label key to check | Commands SHC runs inside each replica; the first stdout line is that replica’s label value. Each check reads test, interval and timeout. A group’s writer is no key: the top-level aliases picks it by these labels (SHC-2459). | x-shc-live-labels |
service.platform | string | platform | |
service.init | bool | init | |
service.read_only | bool | Root filesystem read-only. The service map keeps it. Incus refuses the replica. | read_only |
service.tty | bool | Same as read_only: kept on the document, refused by incus. | tty |
service.sysctls | mapping | A tenant package may not set a host sysctl, as today. | sysctls |
service.ulimits | mapping | ulimits | |
service.dns | list of string | Kept when the author writes it. A service that writes none and has no network_mode gets the node resolver, as today. | dns |
service.dns_search | list of string | The replica project appends the deployment zone. | dns_search |
service.stop_grace_period | duration | stop_grace_period | |
service.stop_signal | string | stop_signal | |
service.pre_start | list of hook | Each hook reads command (required), user, working_dir and environment. | pre_start |
service.post_start | list of hook | Same. | post_start |
service.secrets | list of secret name | secrets | |
service.configs | list of config name, or {name, path} | configs, short form or source and target | |
service.network_mode | none, host or service:<name> | service:<name> puts both services in one pod. A tenant package may not say host. Incus refuses the key and the replica records that. | network_mode |
service.pid | host or service:<name> | service:<name> is the same pod grouping. | pid |
service.ipc | host or service:<name> | Same. | ipc |
service.volumes_from | list of service name | A service source joins that service’s pod. container:<name> is an error in every package (X403293VOLFRM). | volumes_from |
service.links | list of service or service:alias | Inside the pod it stays. To another pod, under the service’s own name, it becomes an after with started. Under an alias the document does not split, as today. | links |
Durability is not a key here. backup.yaml still says which volumes are durable, and the spec records that on the mount.
4. Nextcloud
Section titled “4. Nextcloud”The package on next is apps/flagship/nextcloud. There is no apps/nextcloud. The jobs, plugs, sockets, app.yaml, vars.yaml and backup.yaml stay as they are. This is the compose file rewritten in the schema above. The environment is the same template.
services: nextcloud: image: "{{ vars.images.nextcloud }}" primary: true networks: [default] quota: cpu: "1.0" memory: 2G privileges: security_opt: - apparmor:unconfined restart: unless-stopped expose: [80] environment: NEXTCLOUD_ADMIN_USER: "{{ vars.adminUser }}" NEXTCLOUD_ADMIN_PASSWORD: "{{ vars.adminPassword }}" NEXTCLOUD_TRUSTED_DOMAINS: "{{ hosts_all('nextcloud') | join(' ') }}" OVERWRITEPROTOCOL: "{{ 'https' if tls else 'http' }}" OVERWRITEHOST: "{{ hosts.nextcloud }}" OVERWRITECLIURL: "{{ url('nextcloud') }}" OVERWRITEWEBROOT: "{{ vars.basePath | default(path('nextcloud')) }}" NEXTCLOUD_INIT_HTACCESS: "true" POSTGRES_HOST: "{{ integrations.postgres.socket.hostname | default(vars.postgres.hostname) }}" POSTGRES_DB: "{{ integrations.postgres.plug.data.database | default(vars.postgres.database) }}" POSTGRES_USER: "{{ integrations.postgres.plug.data.username | default(vars.postgres.username) }}" POSTGRES_PASSWORD: "{{ integrations.postgres.plug.data.password | default(vars.postgres.password) }}" REDIS_HOST: "{{ hosts.valkey }}" REDIS_HOST_PORT: "6379" {% if integrations.s3 is defined or vars.s3.endpoint %} OBJECTSTORE_S3_HOST: "{{ integrations.s3.socket.endpoint | default(vars.s3.endpoint) }}" OBJECTSTORE_S3_PORT: "{{ integrations.s3.socket.port | default(vars.s3.port) }}" OBJECTSTORE_S3_BUCKET: "{{ integrations.s3.plug.data.bucket | default(vars.s3.bucket) }}" OBJECTSTORE_S3_KEY: "{{ integrations.s3.plug.data.accessKey | default(integrations.s3.socket.accessKey) | default(vars.s3.accessKey) }}" OBJECTSTORE_S3_SECRET: "{{ integrations.s3.plug.data.secretKey | default(integrations.s3.socket.secretKey) | default(vars.s3.secretKey) }}" OBJECTSTORE_S3_SSL: "{{ integrations.s3.socket.useSsl | default(vars.s3.useSsl) }}" OBJECTSTORE_S3_REGION: "{{ integrations.s3.socket.region | default(vars.s3.region) | default('us-east-1') }}" OBJECTSTORE_S3_USEPATH_STYLE: "true" OBJECTSTORE_S3_AUTOCREATE: "true" {% endif %} SMTP_HOST: "{{ integrations.smtp.socket.hostname | default(vars.smtp.host) }}" SMTP_PORT: "{{ integrations.smtp.socket.smtpPort | default(vars.smtp.port) }}" SMTP_SECURE: "{{ integrations.smtp.socket.smtpEncryption | default(vars.smtp.secure) }}" SMTP_NAME: "{{ integrations.smtp.socket.username | default(vars.smtp.username) }}" SMTP_PASSWORD: "{{ integrations.smtp.socket.password | default(vars.smtp.password) }}" SMTP_AUTHTYPE: "{{ vars.smtp.authtype | default('LOGIN') }}" MAIL_FROM_ADDRESS: "{{ integrations.smtp.plug.data.fromAddress | default(vars.smtp.fromAddress) }}" MAIL_DOMAIN: "{{ vars.smtp.domain }}" {% if shc.ca_root_b64 %} NC_default_certificates_bundle_path: /etc/shc/tls.ca SSL_CERT_FILE: /etc/shc/tls.ca NODE_EXTRA_CA_CERTS: /etc/shc/tls.ca REQUESTS_CA_BUNDLE: /etc/shc/tls.ca CURL_CA_BUNDLE: /etc/shc/tls.ca {% endif %} volumes: - name: nextcloud-data path: /var/www/html mode: "0755" owner: "33:33" {% if shc.ca_root_b64 %} - source: /etc/shc/tls.ca path: /etc/shc/tls.ca read_only: true {% endif %} depends_on: valkey: condition: started healthcheck: test: ["CMD-SHELL", "curl -f http://localhost:80/status.php || exit 1"] interval: 30s timeout: 10s retries: 5 start_period: 120s
notify-push: image: "{{ vars.images.nextcloud }}" path: /push path_strip: true networks: [default] restart: unless-stopped user: "33:33" quota: cpu: "0.25" memory: 256M expose: [7867] environment: PORT: "7867" NEXTCLOUD_URL: "http://{{ environment }}.nextcloud.{{ stack }}" volumes: - name: nextcloud-data path: /var/www/html mode: "0755" owner: "33:33" depends_on: nextcloud: condition: started valkey: condition: started entrypoint: ["/bin/sh", "-c"] command: - | set -eu bin="/var/www/html/custom_apps/notify_push/bin/$$(uname -m)/notify_push" i=0 while [ ! -x "$${bin}" ]; do i=$$((i + 1)) if [ "$${i}" -gt 120 ]; then echo "notify_push binary never appeared at $${bin} after 10 minutes;" >&2 echo "the notify_push app is not installed on this instance" >&2 exit 1 fi sleep 5 done exec "$${bin}" /var/www/html/config/config.php healthcheck: test: ["CMD-SHELL", "curl -sS -o /dev/null --max-time 5 http://localhost:7867/ws || exit 1"] interval: 30s timeout: 10s retries: 3 start_period: 630s
valkey: image: "{{ vars.images.valkey }}" scale: max: 1 restart: unless-stopped command: ["valkey-server", "--save", "", "--appendonly", "no"] quota: cpu: "0.25" memory: 256M volumes: - name: valkey-data path: /data healthcheck: test: ["CMD", "valkey-cli", "ping"] interval: 10s timeout: 5s retries: 5 start_period: 5s
nextcloud-apppass: image: "{{ vars.images.broker }}" entrypoint: ["/usr/local/bin/nextcloud-apppass"] path: /apppass path_strip: true networks: [default] restart: unless-stopped quota: cpu: "0.25" memory: 64M expose: [8420] environment: NCBROKER_LISTEN: ":8420" NCBROKER_ISSUER: "{% if vars.apppass.issuer %}{{ vars.apppass.issuer }}{% elif integrations.oidc is defined and integrations.oidc.socket.publicUrl %}{{ integrations.oidc.socket.publicUrl ~ (integrations.oidc.socket.basePath | default('')) ~ '/realms/' ~ (integrations.oidc.plug.data.realm | default(tenant)) }}{% endif %}" NCBROKER_AUDIENCE: "{{ vars.apppass.audience | default(url('nextcloud') ~ '/apppass') }}" NCBROKER_AZP: "{{ vars.apppass.azp | default('') }}" NCBROKER_GROUP: "{{ vars.apppass.group | default('') }}" NCBROKER_NEXTCLOUD_URL: "http://{{ environment }}.nextcloud.{{ stack }}" NCBROKER_NEXTCLOUD_HOST: "{{ hosts.nextcloud }}" NCBROKER_PUBLIC_URL: "{{ url('nextcloud') }}" NCBROKER_KEY: "{{ vars.apppass.key }}" depends_on: nextcloud: condition: started healthcheck: test: ["CMD-SHELL", "wget -q -O /dev/null http://127.0.0.1:8420/healthz || exit 1"] interval: 15s timeout: 5s retries: 3 start_period: 10s
nextcloud-pgclient: image: "{{ vars.images.postgresClient }}" networks: [default] restart: unless-stopped quota: cpu: "0.5" memory: 512M entrypoint: ["sh", "-c"] command: ["sleep infinity"] environment: NC_DATABASE_URL: "postgresql://{{ integrations.postgres.plug.data.username | default(vars.postgres.username) }}:{{ integrations.postgres.plug.data.password | default(vars.postgres.password) }}@{{ integrations.postgres.socket.hostname | default(vars.postgres.hostname) }}:{{ integrations.postgres.socket.port | default(vars.postgres.port) }}/{{ integrations.postgres.plug.data.database | default(vars.postgres.database) }}"nextcloud-data is one named volume, mounted by two services, with the prep owner the compose file spelled :0755:33:33. /etc/shc/tls.ca is the platform mount, not a host path the app picked. Valkey’s ceiling is scale.max, not x-shc-scale. The push and apppass routes are path and path_strip. Nothing in the file publishes a host port.
5. A group
Section titled “5. A group”One app under apps/ has a {% members %} block (apps/standalone/mongodb-replicaset). The group written out here is pgha, the app the postgres tutorial installs: one primary and numbered streaming standbys. app.yaml, vars.yaml and the two scripts stay. The compose file becomes:
services:{% members "db", default=3, max=5 as m %} db-{{ m.n }}: image: postgres:16-alpine entrypoint: ["sh", "/opt/pgha/member.sh"] environment: MEMBER: "{{ m.n }}" MEMBER_LOCATION: "{{ m.location }}" POSTGRES_PASSWORD: "{{ vars.password }}" REPL_PASSWORD: "{{ vars.replPassword }}" PGDATA: /var/lib/postgresql/data/pgdata volumes: - name: data-{{ m.n }} path: /var/lib/postgresql/data - source: ./bin path: /opt/pgha read_only: true - source: ./initdb path: /docker-entrypoint-initdb.d read_only: true expose: [5432] healthcheck: test: ["CMD", "pg_isready", "-U", "postgres"] interval: 10s retries: 12 quota: cpu: "0.5" memory: 256M{% endmembers %}data-{{ m.n }} is a named volume per member. ./bin and ./initdb are package files, which is why a volume entry can carry a source instead of a name. The member number and the location are the member object the groups note already defines. Scaling stays --scale db=N and --scale db=<location>:N.
Picks (main 10-07, Clay may change)
Section titled “Picks (main 10-07, Clay may change)”Main picked these on 2026-10-07 and corrected four after Clay read them. They are spellings: each is written once in the loader, so a rename is one edit.
whereis one top-level key, a list of where terms. The terms land on the deployment.- The publish flag stays
--expose. It is not a key of this file. lifecycle: one-shotstays. There is no boolean.exposeis the listening-ports key. There is nohost:containerform.- Package files (
source: ./…) and the platform mount/etc/shc/<name>are mount entries. Any other absolute source is a host path and refused, the four inert host files compose allows included. scale.restoreis refused until a restore opens that window. Nothing is parsed and ignored.http_uiis refused from an author. It is an importer’s mark.- A dependency’s conditions are
started,healthyandcompleted. The list form stays. The converter renames compose’s words. - Replaced by correction A.
- A mount’s fields are
read_only,mode,ownerandnoinit, besidename,source,path,sizeandaccess. pre_startandpost_startare closed field by field.replicasreplacesdeploy.replicasand the integerscale.scaleis only the object.- There is no
healthcheck.start_interval.
Corrections, which win over the list:
- A. Every service key a catalog app writes and SHC honours has a row, so every catalog app can be written as shc.yaml. The scan of 110 compose files is on SHC-1361.
- B. The Docker socket and the rest of the privileged class are asked for in one block,
privileges:, and admitted only by the grants that admit them in a compose file. One taxonomy:core/privcapandmodules/privilege. - C.
scale.statefulis not a key. A service that needs its own volume per copy is a{% members %}group. No catalog app writesstatefultoday. - D.
replicasbesidescalestays spelt in one place, so main may fold them intoscale: {default, max}on his word.
Main’s rulings of 2026-10-07 03:46Z: the four inert host files are refused and the converter reports them; a catalog app’s x-shc-refused-* guard is not a key, the Jinja guard stays and the converter keeps it; oci_runtime is compose’s runtime, because runtime means SHC’s engine; ce’s four system addons stay compose. A live label (x-shc-live-labels) is not where: it writes labels and where reads them, so both are keys, live_labels on the service and where at the top.