Repository navigation
Closed
Container
7 / 77 of 7 issues completed
Copy link
Description
skat runs on the platform from its attested private image, routed by Traefik
Caller: overnightworks/skat#6 (S2, "Skat is live"). That slice serves the operator's ruling of 07.11.2026 that Skat is its own fleet app. overnightworks/skat is a public app repository built the way hopin is.
Facts from skat's head (08.11.2026, from skat's code and its release)
- Port 8080;
GET/HEAD /healthanswers{"status":"ok"}, public, unlimited, checking nothing; it serves readiness and liveness for now. skat serveexits unless its database is exactly at the image's migration head;skat migrate(Alembic, under a Postgres advisory lock) needs only the database settings and runs before the server, as an init container of the same image. An older image refuses a newer schema.- Settings (prefix
SKAT_):OIDC_ISSUER,OIDC_CLIENT_SECRET_FILE,SESSION_SECRET_FILE,PUBLIC_ORIGIN(sets the allowed Host and the redirect<origin>/auth/callback), optionalOIDC_CLIENT_ID(skat) andTRUSTED_PROXIES. From skat's next release,DATABASE_URLwithout a password (on the PCpostgresql://skat@postgres:5432/skat) plusDATABASE_PASSWORD_FILE; skat refuses a URL that carries a password. - uid 1001, no runtime writes (read-only root, an emptyDir at
/tmp); egress to its Postgres and the issuer's token and certs endpoints; Keycloak's back-channel logout posts to it; sign-in 40 per 10 minutes per client address, read fromTRUSTED_PROXIES; cookies always Secure; server-sent events with a 15-second heartbeat.
Decided by the head (08.11.2026; the plan is in this item's journal)
- skat gets its own Postgres in namespace
skatfrominfrastructure/postgres, P7a's shape; there is no shared instance to join. This answers skat's Postgres database is backed up, with a restore drill #123's open question; skat's Postgres database is backed up, with a restore drill #123 keeps the backup and its restore drill. - skat's Postgres and its Deployment share one Flux Kustomization
skatoverapps/skat, becausedb:load-dumpsuspends the Kustomization named like the namespace and a Deployment under another one would be scaled back up mid-load. The plan review (Opus, 08.11.2026, journal) held this boundary and chose one Kustomization per app as the target. Every app reconciles through its own Flux Kustomization named like its namespace #145 builds that shape for hopin and speech-runtime and deletesappsbefore P2, instead of an interim list inapps/kustomization.yamlthat Every app reconciles through its own Flux Kustomization named like its namespace #145 would replace. - The password has one owner, the namespace's Secret
postgres: skat reads it as a file (skat's head chose this over a URL secret or a composing init container). - The release bump pull request is P1: every app release yields an attested private image in GHCR #4's bump workflow, not this item's.
Invariants every slice keeps
- One password owner; no URL holds a password.
- The Namespace exists before its Secrets, the Secrets before the Deployment that mounts them; every new Kustomization has
wait: true, atimeoutand depends onadmission,platform-secretsandstorage. - hopin's and speech-runtime's rendered objects stay byte-identical; hal is never touched.
- Pods pass Pod Security
restrictedand admission; skat's image is a full app ref that passesimages; no new vendor. - The host goes public only through Skat's catalog entry: its platform host and its legal block hal#1148, after the realm holds skat's client.
- No secret in this tree or a log.
Slices
- P1 Namespace
skat(restricted, prune disabled) in the namespaces layer. Landed as 2d9f0d2 (P1: Namespace skat #142); on the PCskatexists, enforcingrestricted, prune disabled. - P2 skat's Postgres in the Kustomization
skatoverapps/skat, after Every app reconciles through its own Flux Kustomization named like its namespace #145. Landed as 63f0884 (P2: skat's Postgres in the Kustomization skat #150, PR skat's Postgres runs in the Kustomization skat over apps/skat #153); on the PCskatis Ready, its Postgres runs as uid 70 on a Retain claim andpg_isreadypasses.apps/skat/kustomization.yamlincludes the Postgres base with namespaceskatand skat'skey-tablestables chairs receipts(skat's head, 08.11.2026: all of skat's schema today; a later history table joins on its word).clusters/pc/apps.yamlgainsskat, withwait: trueand atimeout, depending onadmission,platform-secretsandstorage.- The
clusterjob's Postgres-Secret loop (ci.yml, "no layer under infrastructure/ includes") covers every layer that includes the base, those underapps/included. - AGENTS.md (
infrastructure/postgres/, the CI Secret text,clusters/pc/apps.yaml),docs/app.mdwhere it names an app's database, and the comments naming "keycloak and litellm include" follow. - CI proves:
- skat's Postgres is Ready as uid 70 on a Retain claim;
skat's inventory holds the StatefulSet, Service, NetworkPolicy and ConfigMap of namespaceskat, and no other Kustomization holds an object of that namespace but the Secrets platform-secrets serves;skatdepends onadmission,platform-secretsandstorage.
- On the PC
pg_isreadypasses, which needs overnightworks/platform-secrets#18's Postgres password.
- P3 skat runs: Deployment and Service at its release that reads the password file, the init container
skat migrate, read-only root,/healthprobes; CI proves it Available at the full ref with its Secrets before it, and its Deployment carrieskustomize.toolkit.fluxcd.io/name=skat. Theclusterjob's app list (APP_MANIFESTS) and its dummy Secrets gain skat with skat's own keys, since the step writes hopin's keys today. Needs overnightworks/platform-secrets#18 and skat's release. - P4 skat's Ingress for
skat.hallucinai.de, its NetworkPolicy admitting onlytraefikon 8080,SKAT_TRUSTED_PROXIESat the pod CIDR, and its tunnel line; CI proves/health200 through Traefik from namespacecloudflaredand no answer from elsewhere; on the PC/health200 through Traefik, which frees hal#1148.
Named risks
- The first start's
skat migratemay run before Postgres is ready; the init container's backoff heals it within the timeout. - Until skat's Postgres database is backed up, with a restore drill #123 lands there is no backup: a release that migrates gets a
pg_dumpfirst, and a rollback after a migration needs it. - Keycloak's back-channel logout cannot reach skat while Keycloak runs on hal, the gap hopin has too; P7: Postgres, Keycloak, LiteLLM and backups run on the platform, tried on the PC with copies of hal's backups #116's cutover gives skat's port 8080 an ingress from namespace
keycloak. /healthstays green while the database is down after start; a check that fails then belongs to skat (skat#5).- Whether skat's package grants the platform repository's Actions read access is unknown; P3's CI shows it.
Done when
skat's pod runs on the PC from its attested image pinned by tag and digest, on its own Postgres, routed and isolated like hopin, and answers /health 200 through Traefik on the node; its tunnel line for skat.hallucinai.de is on main.
version = 1
now = "P4 landed on 09.11.2026 (#273, PR #276, 0bf0c46): skat's Ingress, NetworkPolicy, trusted proxies and tunnel line are on main, and on the PC /health answers 200 through Traefik for skat.hallucinai.de; skat's legal texts joined the pages the same day (#274, #275)."
next = "hal#1148 points skat.hallucinai.de's record at the platform's tunnel; then the public name answers."
done_when = "skat's pod runs on the PC from its attested image on its own Postgres, routed and isolated like hopin, answers /health 200 through Traefik on the node, and its tunnel line is on main."
size = "L"
slice = []
Metadata
Metadata
Assignees
Labels
No labels
Activity
FlexOr2 commentedon Nov 8, 2026
Pulled 08.11.2026. Plan (phase planner, Opus high): P1 Namespace, P2 Postgres in Kustomization skat over apps/skat, P3 Deployment with skat migrate as init container, P4 route, isolation and tunnel line; #123 keeps backup and drill after #116's P7d. URL: skat reads a password file from Secret postgres (skat's head chose it; options rejected: a URL secret with a second copy, an init container composing the URL). Risks as listed in the body.
FlexOr2 commentedon Nov 8, 2026
Plan review of the Kustomization boundary (Opus plan-reviewer, 08.11.2026): REVISE.
The boundary holds.
db:load-dumpsuspends the Kustomization named like the namespace, scales its Deployments to 0 and deletes its labelled Jobs (mise.toml465-479). A Deployment inappswould be scaled back up mid-load.Alternatives weighed:
infrastructure/skat: breaksdocs/app.md's rule that an app's manifests live underapps/<app>/.appsdropped: the target.db:load-dumpsuspending a named Kustomization: rejected, since it would also suspend hopin and speech-runtime.Blocking findings:
clusterjob's Postgres-Secret loop (ci.yml439-449) covers onlyinfrastructure/*/kustomization.yamlthat include../postgres, so in CI skat would get no Secretpostgres.appsis a predecessor without an exit.docs/app.mdand the comments "keycloak and litellm include" would turn false.skat's objects;skat's dependsOnplatform-secretsandstorage; from P3 on, skat's Deployment carryingkustomize.toolkit.fluxcd.io/name=skat.Also:
appshas nokustomization.yaml, so Flux's generated one would also applyapps/skat, and two Kustomizations would own the same objects.Follow-ups: skat's
key-tables; P2's PCpg_isreadyneeds platform-secrets#18's Postgres password.Head's answer. All four blocking findings go into the body.
apps' exit becomes #145 (one Kustomization per app,appsgone), ordered before P2. The review's interim list inapps/kustomization.yamlwould be a stopgap that #145 replaces, and skat's users can wait for it. skat'skey-tablesgo to skat's head. platform-secrets#18 now carries the Postgres password, put by an agent. A delta re-check of #145 and of this body follows.The Namespace skat exists, restricted and never pruned (#143)
FlexOr2 commentedon Nov 8, 2026
Delta re-check (same Opus reviewer, 08.11.2026): APPROVE. All four blocking findings are resolved, and #145's two-step hand-off is sound under Flux: deleting a Kustomization whose live spec has prune off skips garbage collection, the same field manager adopts the objects by server-side apply, and pod templates do not change. Every site that names
appsis covered. Follow-ups, now in the bodies: #145 S2's way back (prune off on the new Kustomizations before a revert) and P3'sAPP_MANIFESTSand dummy Secrets for skat.FlexOr2 commentedon Nov 8, 2026
skat's head, 08.11.2026: skat#35 has landed as 4bc0811 on skat's main.
SKAT_DATABASE_URLwithout a password, plusSKAT_DATABASE_PASSWORD_FILE: an absolute path to a file holding only the password, where a trailing newline is fine.postgresql://and names a host, a user and a database. It carries no password and no query parameter, so TLS options reach libpq only throughPG*variables.FlexOr2 commentedon Nov 8, 2026
Journal, 08.11.2026, from skat's head:
hand_cardsand a columntables.dealer, through migration 0003. When the release that carries it is named,hand_cardsjoins skat'skey-tables, by skat's word.2026.11.08.N, comes from main once the operator has started skat's release workflow. Starting it is refused in skat's head's permission mode. skat's head sends the tag and digest once the release is attested.FlexOr2 commentedon Nov 8, 2026
Journal, 08.11.2026, agreed with skat's head (skat#61, S10.1):
SKAT_PUBLIC_ORIGINis written as a browser writes an origin:https, a lowercase host, no trailing dot, no:443, no path, query or trailing slash.https://skat.hallucinai.de.SKAT_CALL_ORIGIN(hopin's origin, which may frame skat) stays unset, so nothing frames skat. skat's head names the value once hopin's stage ships.FlexOr2 commentedon Nov 8, 2026
P3 (#166, PR #168, tip 2d8336d): review Sonnet/medium APPROVE, no blocking finding. Two follow-ups for P4 and the head:
apps/skat/sentence "It has no route, NetworkPolicy or tunnel line yet" misleads, since namespaceskatholds the base's NetworkPolicypostgres, and "yet" is plan narrative. P4 rewrites that sentence anyway.kustomize.toolkit.fluxcd.io/name=skat. No CI cell asserts the label directly. The head reads it on the PC after the landing and journals it here.db:load-dumpdepends on that label to scale the Deployment.FlexOr2 commentedon Nov 8, 2026
Journal, 08.11.2026, PR #168 (P3, #166) at 2d8336d:
imagesandclusterare red for one cause.imagesreportsMANIFEST_UNKNOWNforghcr.io/overnightworks/skat@sha256:677fe5eb…, andclustershowsErrImagePullfor the same ref. The digest and the attestation are right: GitHub's attestation API holds one provenance for that digest, subjectghcr.io/overnightworks/skat, built from 43153b9, the commit of tag v2026.11.08.1. The job's token cannot read the package, becauseplatformhas no Read under skat's package's Manage Actions access, the stepO8was for hopin andO1for speech-runtime. It is a web-only setting and the operator's. Skat's public step stays held by skat's head (skat#5's line 10); P3 and P4 go on.skat runs from release 2026.11.08.1 with its migrations (#168)
FlexOr2 commentedon Nov 9, 2026
Journal, 09.11.2026: skat answers live through the platform. hal's head checked after the operator's
hal update,hal setupandhal apply(hal at e9cfd360;~/srv/skatat v2026.11.08.1, the checkout hal#1148 needs):GET https://skat.hallucinai.de/healthreturns 200 through the edge and platform's tunnel line (P4: skat's route, isolation, trusted proxies and tunnel line #273, PR skat is routed through Traefik, closed to every other Pod, and has its tunnel line #276).GETandPOST /auth/backchannel-logoutreturn 403 from Cloudflare, so the edge refuses them./datenschutzcarries#skat,#skat-konto,#skat-geraetand#skat-protokolle, and/nutzungsbedingungencarries#skatand#skat-leistung. hal serves these pages until P9a-5.hal#1148 closes once the operator's
hal check hostsexits 0.FlexOr2 commentedon Nov 9, 2026
A follow-up from PR #292's review: skat's release is set twice, at
apps/skat/skat.yaml:33(the init containermigrate) and:54(serve), andtests/legal/app_texts_equal_their_release.sh:31reads only the first one, withhead -n 1.That was already so before #292. A bump that moved only one ref would pass the legal check against the wrong release. The
imagesjob judges both refs, but no check holds the two equal. The owner is this item: a check that both refs are one, or a kustomizeimages:entry that sets both.FlexOr2 commentedon Nov 9, 2026
Done when is met: skat's pod runs on the PC from its attested image on its own Postgres, its Ingress, NetworkPolicy and tunnel line are on main (#273, PR #276, 0bf0c46), and after overnightworks/hal#1148 closed the public name answers:
GET https://skat.hallucinai.de/healthreturns 200 (read again on 09.11.2026). The follow-up from PR #292's review is closed by 7ba3a15 (#342):tests/legal/app_texts_equal_their_release.shreads every image ref of an app and fails when they name more than one version.