Repository navigation
Description
hopin runs on the platform from its attested private GHCR image, with its secrets and its pull credential from platform-secrets, routed by Traefik as hopin.hallucinai.de, proven in the job cluster on pull requests and on main and then on the PC beside hal.
Caller
- Platform on k3s and Flux: from an empty PC to every app live, with hal archived #1, outline row P6: "hopin, speech-runtime and presentator's web run in k3d and on the PC beside hal." Target picture, Images: the PC pulls each app image privately from GHCR with a read-only token.
- P4: Flux decrypts SOPS-age secrets from platform-secrets #13, slice P4d: "the first private image pulls on the PC path … In k3d CI pulls private images with the job's
GITHUB_TOKEN(packages: read, the package grant), swapped in like P4a's source. Pulled with P6's hopin." - Traefik runs in the cluster from manifests rendered from its pinned chart and answers 404 for every host without a route #44, decision 9: this item's Ingress is Traefik's first real route, and its route cell is the proof that Traefik routes, beyond its own 404.
- P1: every app release yields an attested private image in GHCR #4 and its guard B1: this item carries the first app ref, which lands only after B1b (live-go plan, section 5).
- The live-go plan's slice S5 (the head's proposal of 06.11.2026, not ruled); hopin's facts from hopin's head (06.11.2026) and hopin's origin/main 49bd061.
Ruled
- Platform on k3s and Flux: from an empty PC to every app live, with hal archived #1, invariant 1, line A (the operator, 06.11.2026): „A => einfach sagen wenn soweit ist! …“. Card R1 below is ruled other (06.11.2026), and Platform on k3s and Flux: from an empty PC to every app live, with hal archived #1's invariant 1 now says that apps go live one by one, as a test, hopin first, without waiting for the cutover; that replaces "live stays on hal until the cutover". P3: the cluster skeleton reconciles in k3d on every pull request #11's invariants carry the live build: host cells run on the PC beside running hal.
- Q6 on P4: Flux decrypts SOPS-age secrets from platform-secrets #13 (the operator, 06.11.2026, „ja passt alles!“): his own classic PAT with only
read:packagesand an expiry is the pull credential on the PC. - Q8 and Q9 (same comment): the image pin changes only through a pull request he merges, and
apps/is the layout. - The bump path (the operator, 06.11.2026, on P1: every app release yields an attested private image in GHCR #4): the release's last job opens the bump pull request through the organisation's GitHub App, and Dependabot's docker ecosystem is not used for app images. This item's Deployment
image:field is what that pull request edits. - Only a
v*tag goes live (hal ruling of 16.10.2026, as P1: every app release yields an attested private image in GHCR #4 records it).
Decided by the head
Revised 06.11.2026 after the Codex plan review (REVISE): trust and isolation land together (decisions 3, 9 and 11), the cards R1 and R3a are corrected, the WebSocket cell H12 is two connections with a result table, and the rollback cell H15 drives #45's prune rule. The delta re-check the same day was APPROVE (#43 records it), and the operator ruled the cards R1, R3a and R3b below on 06.11.2026.
-
Layout.
apps/hopin/hopin.yamlholds the Deployment, the Service, the Ingress and the NetworkPolicy.- The Namespace
hopin, enforcing Pod Securityrestrictedand carryingkustomize.toolkit.fluxcd.io/prune: disabled, joinsinfrastructure/namespaces/namespaces.yaml, the layer The platform's own Cloudflare tunnel runs cloudflared in k3s in front of Traefik #45 creates for every namespace that receives a Secret fromplatform-secrets(its decision 5). clusters/pc/apps.yamlis the Kustomizationapps:path: ./apps,dependsOn: [platform-secrets],prune: true,wait: trueand atimeout.appsdoes not depend ontraefik: hopin is Ready without it, and the Ingress is data that Traefik reads when it runs.apps/is declared in[[root.allow]](Q9).
-
The pod.
- Deployment: one replica, strategy
Recreate(call state lives in memory, so two pods would split a call). A deploy ends every running call, as on hal. - Security:
runAsNonRoot, which the kubelet verifies against the image's numericUSER 1001(hopin'sDockerfileat 49bd061; hopin#537 closed with hopin's The acestep-worker header names songmaker's files that own its settings and image #542), so norunAsUseris set; a read-only root; anemptyDirat/tmp; no privilege escalation; all capabilities dropped; RuntimeDefault seccomp;automountServiceAccountToken: false. - Probe: readiness on
GET /healthat 8080. There is no liveness probe: a restart ends every call and session, and hal does not restart an unhealthy hopin either.
- Deployment: one replica, strategy
-
Settings, as
compose.yamlsets them, with these changes:HOPIN_TURN_PROVIDER: cloudflarewith the two TURN files, as card R3b's ruling A says;- under card R3a's gap,
HOPIN_AGENT__GATEWAY__ORIGIN: http://gateway.invalid. RFC 6761 reserves.invalidso it never resolves, and no key file exists. The game's four settings are required at start (README, Configuration); hopin reads the key at every turn, and a missing key file reads as unavailable (README). A comment in the manifest gives that why. Whether the setting is accepted at start is H1's; that a person sees the quiz as unavailable is observed in the live go's drive. Rejected: hal'shttp://litellm-front:4000, which names a host the platform does not have. HOPIN_TRUSTED_PROXIESonly in S5b, with the NetworkPolicy (decision 9). In S5a it stays unset: hopin's default""trusts no proxy (src/hopin/config/settings.py), so hopin keys every request on its direct peer.
-
Secrets, as files under
/run/secrets, from the Secrethopin(namespaces/hopin/hopin.yamlinplatform-secrets).- Keys, named as compose names its files:
oidc_client_secret,session_secret, andturn_key_idandturn_api_tokenwhen TURN is on. litellm_keyis not copied while the quiz gap holds, because nothing on the platform could use it; it joins when LiteLLM moves (P7).- The values are the operator's (O5). The client secret is the value hal's hopin uses, so a rollback signs in with the same client.
- The session key is new. hopin keeps its sessions in process memory (
InMemorySessions,src/hopin/app.py), and the cookie only names a session there (webauth v0.30.0, HMAC-signed cookies). A copy of hal's key would carry no session across the switch and would only add a second copy. Everyone signs in again at the switch and at a rollback: card R1 says so.
- Keys, named as compose names its files:
-
The pull credential.
- On the PC: the Secret
ghcr-pull, of typekubernetes.io/dockerconfigjson, in namespacehopin, one filenamespaces/hopin/ghcr-pull.yamlinplatform-secretsholding the operator's PAT (O3, Q6). The Deployment names it inimagePullSecrets. - For the first app it is one file, not P4d's "kept once and included in each app namespace". The include, a namespace transform in
platform-secrets, arrives with the second app that needs it. Until then one file has one rotation. The head edits P4: Flux decrypts SOPS-age secrets from platform-secrets #13's P4d row to say so. - In k3d: the job writes a dockerconfigjson from its own
GITHUB_TOKENinto the same Secretghcr-pullin its SOPS-encrypted artifact, beside dummy values for the Secrethopin, exactly as P4a serves its dummy. Flux decrypts it, and the token is masked and never printed. - The grant:
clustergainspackages: read, and hopin's package grants platform Read (O8). The head edits P3: the cluster skeleton reconciles in k3d on every pull request #11's D4 grant (pair in the P3d pairs file). - Weighed:
kubectl create secretin the job (outside Flux, not the PC's path); the PAT in/etc/rancher/k3s/registries.yaml(rejected in P4d); a PAT as a CI secret (a secret where the job token serves).
- On the PC: the Secret
-
The first app ref is one string,
ghcr.io/overnightworks/hopin:<YYYY.MM.DD.N>@sha256:<digest>, in the Deployment'simage:field, from hopin's first release through quality'simage.yml(hopin's The acestep-worker header names songmaker's files that own its settings and image #542).- The ref: hopin's first attested release,
ghcr.io/overnightworks/hopin:2026.11.06.2@sha256:aaa309151f36d714fe49d8a285f85c4942b4e145e0befd32472167978c496ae6, tagv2026.11.06.2at hopin's 127f26c (this item's journal, 06.11.2026). - When: it lands in S5a's landing commit. B1b has landed (f284a22), and its E1 judged this ref green in
images(PR The job images checks each app ref's tag, attestation and signer commit with the locked gh #67, run 42326095155). - Rejected: a ref by tag alone, a ref before the guard, and an image built on the PC and imported into containerd (live-go plan, section 5).
- The ref: hopin's first attested release,
-
What hopin needs at start (read from source, not driven):
- the four required settings and the game's four;
- the page build inside the image;
- a readable client secret, and a session secret of at least 32 characters;
- the two TURN files when TURN is on (
src/hopin/config/settings.py,auth/sign_in.py,main.py, README).
It calls no service at start. The OIDC endpoints are built from
HOPIN_OIDC_ISSUERwithout discovery (_build_oidc_client), and webauth fetches the key set only when it verifies a token (jwt_verification.py,_refresh_key_set). So k3d's dummy secrets start it, and its issuer is first reached at a sign-in. -
The route.
- The Ingress: class
traefik, one rule for hosthopin.hallucinai.de, path/(Prefix) to the Servicehopinon 8080. No TLS (Cloudflare terminates it), no default backend, no rule without a host, so P3d's 404 for unknown hosts holds. - WebSocket upgrades on
/wspass Traefik without annotations. - The back-channel path stays off the internet by the zone's rule at Cloudflare (hal's
deny_public), not by the route (R4, Neighbours).
- The Ingress: class
-
Trust and isolation land together, in S5b (the review's finding 2).
- The NetworkPolicy in namespace
hopinadmits ingress on 8080 only from namespacetraefik; traffic from the node, and with it the kubelet's probe, is always allowed. HOPIN_TRUSTED_PROXIES: 10.42.0.0/16, k3s's default cluster CIDR (P3b'shost/k3s/config.yamlat fa2f78e sets nocluster-cidr), the pod network where Traefik's pods live and change address on every restart. webauth readsX-Forwarded-Forfrom the right and stops at the first untrusted hop (webauth/proxies.py). The chain is Cloudflare's edge, then cloudflared (The platform's own Cloudflare tunnel runs cloudflared in k3s in front of Traefik #45), then Traefik trusting namespacecloudflared(The platform's own Cloudflare tunnel runs cloudflared in k3s in front of Traefik #45), then hopin trusting Traefik. So the hop it stops at is the visitor's address as Cloudflare's edge appended it.- Why together: trust without the policy would let every pod in the cluster name a client address to hopin, for its rate limits and logs; the policy without the trust has no sender to protect. P3d decision 7 pairs Traefik's trust with its policy for the same reason.
- The relationship check: the job checks that each Traefik pod's address lies in a CIDR of hopin's setting, both read from the cluster. Leaving the setting empty is refused: every visitor would share Traefik's rate-limit budget.
- Remaining risk as P3d decision 7 names it.
- The NetworkPolicy in namespace
-
The WebSocket timeout is decided by a driven cell, not by inference. P3d decision 8 infers from source that Traefik's default
readTimeoutof 60s cuts an upgraded connection. Yet hal's Traefik v3.7.16 serves hopin's/wstoday with no timeout override (hal'straefik.ymlat e15da11).-
hopin's side (hopin's head, from hopin's code at 49bd061, which this item re-read): uvicorn sends a WebSocket ping every 20s (its default
ws_ping_interval;uvicorn_optionsinmain.pysets none), and the browser answers each one. So an idle timeout never fires on/ws; only a deadline that traffic does not reset would cut it. -
What a cut costs: a person sees nothing, since the client reconnects after 1s and resumes its seat. hopin's log shows each cut as
websocket_errorwith code 1006 at error level (signaling/handler.py), one per participant per minute, which buries real errors. -
The cell H12 holds two connections through the platform's Traefik for 120s at the default
readTimeout, both from namespacecloudflaredto a scratch echo server: one idle, and one sending a ping every 20s, as hopin's traffic does. -
What each result decides:
Idle connection Pinging connection Means Decides held 120s held 120s no deadline at the default readTimeoutstays; P3d's inference is corrected on Traefik runs in the cluster from manifests rendered from its pinned chart and answers 404 for every host without a route #44cut near 60s held 120s a deadline that traffic resets readTimeoutstays, since hopin's pings reset it; Traefik runs in the cluster from manifests rendered from its pinned chart and answers 404 for every host without a route #44 records that an app without pings is cut after 60s idlecut near 60s cut near 60s a fixed deadline that traffic does not reset this slice sets ports.web.transport.respondingTimeouts.readTimeout: 0ininfrastructure/traefik/values.yamland regeneratesrendered.yamlwithmise run charts:render, as P3d decision 8 hands over, and H12 runs again at 0, where both must holdany other result not foreseen reported to the head, who decides before S5b lands -
Either way the result goes to hopin's and hal's heads.
-
-
Cut into two pull requests, as the head chose (07.11.2026).
- S5a, hopin runs: decisions 1, 2, 4 to 7, and 3 without
HOPIN_TRUSTED_PROXIES; about 150 lines in five files. - S5b, hopin is routed and isolated: decisions 8 to 10, the Ingress, the NetworkPolicy and
HOPIN_TRUSTED_PROXIESin one change; about 80 lines in three files.
S5a lands after The platform's own Cloudflare tunnel runs cloudflared in k3s in front of Traefik #45's S4a, whose namespaces layer and annotation guard it joins (decision 1, H15); its other preconditions, B1b and hopin's first release, are done. S5b lands after S5a and The platform's own Cloudflare tunnel runs cloudflared in k3s in front of Traefik #45's S4b. Weighed: one pull request (about 230 lines in six files); trust in S5a without the policy (every pod could forge a client address, the review's finding 2); trust and policy both in S5a (no sender before the route, and the rate-limit and relationship cells need the route).
- S5a, hopin runs: decisions 1, 2, 4 to 7, and 3 without
-
A revert of S5a keeps the Namespace and its Secrets (the review's finding 1). Its Secrets live in
platform-secrets, so the Namespace carries The platform's own Cloudflare tunnel runs cloudflared in k3s in front of Traefik #45's prune annotation, and the rollback cell H15 drives the revert while the Secrets stay, with a control without the annotation.
What is known
- hopin, origin/main 49bd061:
- the image runs as
hopin, uid 1001, withUSER 1001(Dockerfile), listens on 8080, and servesGET /healthwithout a session; every other path,/wsincluded, needs one; - state lives in process memory, sessions included (
InMemorySessions); sessions last 12 hours (SESSION_LIFETIME_SECONDS); trusted_proxiesdefaults to"", no proxy trusted (config/settings.py);- the sign-in budget is 40 requests in ten minutes per client address, and
/healthis exempt (config/constants.py); - TURN key files must match
[A-Za-z0-9_-]+and[A-Za-z0-9._~+/-]+=*(web/turn_relay.py), so CI's dummies are hex.
- the image runs as
- hal, origin/main e15da11:
- compose's
HOPIN_TRUSTED_PROXIESnames hal's network gateway and127.0.0.1; - Keycloak reaches hopin's back-channel logout only on
hal_backchannel_hopin, and the zone blocks that path publicly (docs/external.md,deny_public). The sign-in itself reaches Keycloak over the public internet.
- compose's
- The image: hopin's first attested release
v2026.11.06.2exists (decision 6): hopin's Release run 42283687917, built and attested through quality'simage.ymlat quality 7715afe, with the attestation's subject digest equal to buildx's (this item's journal, 06.11.2026). - The fixtures, read 06.11.2026 (the builder re-reads each; named with repository and digest so the admission item can list them):
- the client Pods' image, P3d's:
docker.io/curlimages/curl:8.23.0@sha256:3fb7e13e5d179a845c97d78204f3f1064d1907037f26a3ebf91710571151fd93(Docker Hub API); - H12's echo server:
docker.io/traefik/whoami:v1.12.0@sha256:c4717a8d1f0134a7444e24f881160e033991f23027c6c5a9a3f8fd22e70d1d44(Docker Hub API); - H12's WebSocket client:
ghcr.io/vi/websocat:v1.14.1@sha256:830f664fd592cd7deab9d6af6a3391aa0220f29bacf2b5a9d58f6e51c06a5344(GHCR's registry API, the index digest), whose--ping-intervalsends WebSocket pings (src/main.rsat v1.14.1). Its image sets noUSER, so its Pod sets a numeric user for namespacecloudflared'srestricted.
- the client Pods' image, P3d's:
- Invariants kept.
- No secret in the repository, a log or a test: the job's token and dummies exist only in the runner and the k3d cluster, masked.
clusterkeepspersist-credentials: falseand gains onlypackages: read.- No new mandatory job.
appssetswait: trueand atimeout.- hal's hopin keeps running until the live go.
- Rollback.
- Revert the commit. The root's prune deletes the Deployment, the Service and the Kustomization
apps. The Namespacehopinstays with its Secrets (prune disabled, The platform's own Cloudflare tunnel runs cloudflared in k3s in front of Traefik #45 decision 5), andplatform-secretsstays Ready (H15). The Namespace leaves for good only after its Secrets have leftplatform-secrets, The platform's own Cloudflare tunnel runs cloudflared in k3s in front of Traefik #45's named exit. - After a live go, in this order (hal#1151, proven by hal's Codex gate on 07.11.2026):
hal start hopinfirst, while hopin's host still names the tunnelplatform; then the catalog line flipped back on hal's main,hal updateandhal apply, which moves the record back only once hal's own hopin answers its ready path. The reverse order fails: a stopped hopin answers the stop barrier's 404, andhal applyrefuses. In minutes; everyone signs in again.
- Revert the commit. The root's prune deletes the Deployment, the Service and the Kustomization
- Deletes. Nothing in this repository. hopin's compose file and runbook leave in P11 through hopin's board.
- Size.
- S5a, about 150 lines in five files:
apps/hopin/hopin.yamlabout 72,clusters/pc/apps.yamlabout 18,infrastructure/namespaces/namespaces.yaml9,quality/config.toml3, theclusterjob about 45. - S5b, about 80 lines in three files:
apps/hopin/hopin.yamlabout 42, theclusterjob about 36,infrastructure/traefik/values.yaml2 only if H12 cuts the pinging connection, withrendered.yamlregenerated. AGENTS.mdchanges in both.
- S5a, about 150 lines in five files:
Operator actions
Each named once:
- Cards R1, R3a and R3b below: ruled 06.11.2026.
- O3: a classic PAT with only
read:packagesand an expiry of his choice, encrypted withmise run secrets:put-registry hopin ghcr-pull ghcr.io <login>(hidden prompt) asnamespaces/hopin/ghcr-pull.yamlinplatform-secrets, in the platform-secrets value item opened at P4b step 8, after P4b (A and B) and The PC creates its age key once and reaches platform-secrets with its own read-only deploy key #61's SB6. - O5: hopin's values with
mise run secrets:put hopin hopin <key>per key (stdin or hidden prompt) asnamespaces/hopin/hopin.yaml, in the same value item: the existing client secret, a new session key of at least 32 characters, and the two TURN values if R3b says on. - O7: hopin's package private.
- O8: on hopin's package, Manage Actions access →
platform: Read. Done: the operator confirmed it in chat on 07.11.2026 (this item's journal). - O9: dispatch hopin's first release, done 06.11.2026 (
v2026.11.06.2, this item's journal); then merge S5a's pull request.
#45's C12 (O11) and the DNS switch and live-go word (O10) belong to #45 and the live go, not to this item.
Files
apps/hopin/hopin.yaml(new):- S5a: the Deployment and the Service;
- S5b: the Ingress, the NetworkPolicy and
HOPIN_TRUSTED_PROXIES.
clusters/pc/apps.yaml(new): the Kustomizationapps.infrastructure/namespaces/namespaces.yaml: the Namespacehopin, with the restricted label and the prune annotation.quality/config.toml:[[root.allow]]forapps/, reason: Flux's documented layout, ruled Q9 on platform#1.infrastructure/traefik/values.yamland its generatedrendered.yaml(S5b, only if H12 cuts the pinging connection):readTimeout: 0onweb, regenerated withmise run charts:render..github/workflows/ci.yml, jobcluster:- S5a:
packages: read, the Secretshopinandghcr-pullin the job's artifact, and the assertions; - S5b: the route, NetworkPolicy, rate-limit and relationship cells, whose client Pods run in namespace
cloudflared.
- S5a:
AGENTS.md: the Layout line forapps/hopin/; theclusterline on the private pull and the route.
Verification: CI on the pull requests and the landing pushes; B1's images job green on the first ref; quality hygiene . printing hygiene: clean; the driven cells; the PC cell H13 after O3, O5 and O8.
Done when
- S1 Flux reconciles the Kustomization
appsafterplatform-secrets, withwait: trueand atimeout. Aplatform-secretsthat is not Ready holdsappsback with Flux's dependency reason. - S2 In the job, hopin's Deployment pulls its attested ref through the Secret
ghcr-pull, which the job made from its own token, and becomes Available with one Ready pod. The namespace enforcesrestrictedand carries the prune annotation. The container'simageIDrelates to the pinned digest as P3d's E4 established. Withoutghcr-pullthe pull fails. - S3 (S5a) hopin starts from the Secret
hopin's files under/run/secretswith no LiteLLM key and no trusted proxy. A missing session secret stops the start, naming the setting and no value, andclusterturns red. - S4 (S5b) From namespace
cloudflared,GET /healthwith Hosthopin.hallucinai.dethrough Traefik'swebgets 200, and an unknown host still gets 404. A Pod in another namespace gets no answer from hopin's Service. - S5 (S5b)
HOPIN_TRUSTED_PROXIESlands with the NetworkPolicy of S4, in one change. With it, 40 sign-in starts from one forwarded address make the 41st get 429, while the next start from another forwarded address does not. Every Traefik pod's address lies in the setting. - S6 (S5b) H12 is reported with both connections' close times, and
readTimeoutstays at the default or is 0 as its result row in decision 10 says. - S7 On the PC, after O3, O5 and O8,
appsis Ready at main's revision with hopin's pod pulled through the PAT Secret, and/healthanswers through Traefik from namespacecloudflared. - S8
AGENTS.mdnamesapps/and the route,quality hygiene .printshygiene: clean, and the cards R1, R3a and R3b are ruled. - S9 Reverting S5a's workload while its Secrets stay keeps the Namespace
hopin, andplatform-secretsstays Ready (H15).
Scenario matrix
Material risk (secrets in the pod, a private pull credential, the client-address trust): the first review drives the matrix, then the cross-provider gate. Driven cells run on scratch pull requests of the lane, closed after with their branches deleted. Host cells run on the PC beside running hal, under #1's invariant 1.
| Cell | What is driven | Expected observable | Proven by |
|---|---|---|---|
| H1 | S1 and S2, pull request | apps Ready after platform-secrets; hopin Available with one Ready pod, strategy Recreate |
CI |
| H2 | S2, pull request | the Deployment's ref is one string; images (B1) is green on it |
CI |
| H3 | S2, pull request | the container's imageID against the pinned digest, by the relationship P3d's E4 found |
CI |
| H4 | S3, scratch: the job's artifact without session_secret |
the pod is not Ready; its log names session_secret_file and no value; apps and cluster red |
driven |
| H5 | S2, scratch: the artifact without ghcr-pull |
the pull fails with the registry's refusal (the package is private); apps red |
driven |
| H6 | S1, scratch: the artifact's encryption broken | platform-secrets not Ready; apps reports its dependency not ready |
driven |
| H7 | S2 and S3, the diff | the security context without runAsUser, /tmp, the Secret mount, the probe, the settings (the gateway origin and its comment, no trusted proxy in S5a), no litellm_key, imagePullSecrets, the Namespace's two markers |
diff review |
| H8 | S4, pull request | /health through Traefik with hopin's host gets 200; an unknown host gets 404 |
CI |
| H9 | S4, pull request | a Pod in another namespace gets no answer on hopin:8080 within its timeout |
CI |
| H10 | S5, pull request | 40 /auth/start with X-Forwarded-For: 192.0.2.1, then the 41st gets 429; then one with 192.0.2.2 does not |
CI |
| H11 | S5, pull request | each Traefik pod's IP inside a CIDR of hopin's HOPIN_TRUSTED_PROXIES, both read from the cluster; the setting and the NetworkPolicy in the same diff |
CI and diff review |
| H12 | S6, scratch: whoami's /echo behind a scratch Ingress; from namespace cloudflared, two websocat connections held 120s at the default readTimeout, one idle and one with --ping-interval 20; again at 0 if decision 10's table says so |
each connection's close time or "held"; the table row they match and the value it decides | driven |
| H13 | S7, the PC | apps Ready, hopin pulled with the PAT, /health through Traefik |
the head, a host step |
| H14 | S1 to S6, push to main | cluster green with H1 to H3 and H8 to H11's observables |
CI, read at landing |
| H15 | S9, scratch, in one job run: the tree with S5a reconciled, then the GitRepository moved to a commit that reverts S5a's manifests (the Deployment, the Service, apps and the Namespace hopin) while the job's artifact keeps the Secrets hopin and ghcr-pull; then the same once more with the Namespace's prune annotation removed and #45's annotation guard skipped |
first: hopin's Deployment gone, the Namespace hopin and its Secrets still there, namespaces and platform-secrets Ready at the new revision; then the control: the Namespace deleted, and platform-secrets not Ready naming the missing namespace |
driven |
Driven cells: H4, H5, H6, H12, H15 (five), plus the host step H13.
Unknowns
| Unknown | Settled in |
|---|---|
Whether k3d's containerd pulls hopin's private package with cluster's job token through ghcr-pull (not verified; B1b's E1, PR #67, showed platform's job token with packages: read reading the package's digest from the registry in images) |
H1 |
Whether a pull request from a fork gets a job token that can pull it (platform is public; not verified); if not, its cluster turns red |
the first such pull request after S5a |
Whether a Dependabot-triggered run of cluster can pull it (not verified); if not, those pull requests turn red |
the first such pull request after S5a |
Whether hopin's agent settings accept http://gateway.invalid at start (not verified) |
H1 |
| Whether a person then sees the quiz as unavailable, as the README promises | the live go's drive |
Whether hopin writes outside /tmp at runtime (start is proven, sign-in and calls are not) |
the live go's drive |
Which deadline, if any, Traefik's default readTimeout puts on an upgraded WebSocket, and whether hal's hopin is cut each minute today |
H12 |
Whether the PC's pods reach auth.hallucinai.de over the public path |
H13 and the live go's sign-in |
| Whether k3s enforces the NetworkPolicy | #45's C8 before H9, and its C10 on the PC |
Neighbours
- Traefik runs in the cluster from manifests rendered from its pinned chart and answers 404 for every host without a route #44: this item is its first real route, and H12's result revises its decision 8.
- The platform's own Cloudflare tunnel runs cloudflared in k3s in front of Traefik #45: the namespaces layer and its prune rule, the tunnel, Traefik's trust; S4b lands before S5b, and its C12 runs after H13, before the live go.
- B1 under P1: every app release yields an attested private image in GHCR #4: B1b lands before S5a and judges its ref. P1: every app release yields an attested private image in GHCR #4's bump slice (route A) later edits the same
image:field. - The admission item: it lists the fixtures above.
- hopin's head hears H12's result.
- P4: Flux decrypts SOPS-age secrets from platform-secrets #13: P4d's remaining lines (the include for a second app, the PAT's expiry and revocation runbook) stay there. P4b (A and B) and The PC creates its age key once and reaches platform-secrets with its own read-only deploy key #61's SB6 come before O3 and O5, which the platform-secrets value item for hopin's values, opened at P4b step 8, carries.
- hal's head:
- R4, hal's Cloudflare catalog after the switch: it must keep the zone's block on hopin's
/auth/backchannel-logout; - H12's result for hal's own Traefik.
- R4, hal's Cloudflare catalog after the switch: it must keep the zone's block on hopin's
- The live go (S8 of the live-go plan): first the head's report that the platform's hopin answers
/healththrough the tunnelplatform(hal applychecks no host on another tunnel, so nothing in hal does), thenhal stop hopin, O10, the drive at mobile and desktop width (the quiz reading unavailable included), and the rollback drill in the order above. - S9 of the live-go plan: the 30-day bound on pod logs must hold before hopin's first log line on the PC is 29 days old.
version = 1
now = "S5a landed as 226d67a (#82, PR #96, 07.11.2026), and #45's S4b as 5607fe5. On the PC, platform-secrets PR #9 (#8) is merged, and hopin runs Ready at its attested digest. S5b is #117 (PR #118). H12 held, so readTimeout stays at the default. Its delta lane judges the merge of main 72260f4."
next = "Land #117 (S5b), then H13 on the PC (apps Ready, /health through Traefik), then #45's C12, then the live go (O10)."
done_when = "hopin runs on the platform from its attested private GHCR ref with Recreate, one replica, the image's uid 1001 and a read-only root, its secrets as files from platform-secrets and its pull Secret from the job token in k3d and from the PAT on the PC; the apps Kustomization waits on platform-secrets; Traefik routes hopin.hallucinai.de to it as its first real route, and in one change only Traefik reaches it and its trusted proxies key each visitor apart; the WebSocket timeout is set by its driven cell; a revert keeps the namespace and platform-secrets Ready; it is Ready on the PC; the cards R1, R3a and R3b are ruled; and the matrix H1 to H15 is reported."
scope = [".github/workflows/ci.yml", "AGENTS.md", "apps/hopin/hopin.yaml", "clusters/pc/apps.yaml", "infrastructure/namespaces/namespaces.yaml", "infrastructure/traefik/rendered.yaml", "infrastructure/traefik/values.yaml", "quality/config.toml"]
size = "L"
whole = "S5a does not split further: the Deployment, its Kustomization, its root declaration and the private pull that lets it start have no caller apart; S5b is the cut that serves the route, and its trust and isolation cannot part."
[[expectation]]
text = "HopIn zieht vor der großen Umstellung allein auf die neue Plattform um; nur seine Adresse wechselt, alle anderen Apps bleiben beim alten Stand."
asked_on = 2026-11-06
ruling = "other"
ruled_on = 2026-11-06
card_hash = "sha256:8960adee1aae9c2db3bcff298eaf0f4c182df2425071e5cf52bfd5d1bd4aacbc"
via = "chat"
other = "„3b) all zusammen oder von mir aus erst hopin für den test“; „warum nich alle? es tut nicht weh die apps haben keine user! also was ist los? sind die apps vorbereite? einfach fucking test go und wir sehen ob alle apss gehen oder nicht?“ (06.11.2026)"
question = "Soll HopIn als erste App schon vor der großen Umstellung von der neuen Plattform laufen, während alles andere beim alten Stand bleibt?"
problem = "Bis zur großen Umstellung zeigt kein echter Besucher, ob die neue Plattform trägt."
decision_kind = "release"
area = "Plattform"
topic = "Erste App live"
recommended = "A"
reason = "HopIn speichert nichts dauerhaft, ist öffentlich und braucht nur die Anmeldung, die es schon heute über das Internet erreicht (gelesen in HopIns Code und Startkonfiguration). Der Rückweg ist ein Adresseintrag und ein Startbefehl."
[[expectation.option]]
name = "HopIn zieht vorher um"
effect = "Nach deinem Wort zeigt die Adresse von HopIn auf die neue Plattform; alle anderen Apps laufen weiter wie heute."
notice = "HopIn sieht aus wie heute; nach dem Umschalten meldest du dich einmal neu an. Bevor die Adresse wechselt, gibst du das Wort."
cost = "Ein laufender Anruf bricht beim Umschalten ab, und jeder meldet sich danach einmal neu an: HopIn auf der neuen Plattform hat einen eigenen Sitzungsschlüssel und kennt keine laufende Sitzung. Beim Zurückschalten gilt dasselbe. Vorher übernimmt die neue Plattform HopIns Adresseintrag und die Regel, die seinen internen Abmeldeweg vor dem Internet sperrt, vom alten Stand; das ist eine eigene Aufgabe vor dem Umschalten, und dafür legst du voraussichtlich einmal einen Cloudflare-Schlüssel an. Bis zur großen Umstellung laufen zwei Plattformen nebeneinander, und die Lücken der nächsten Frage gelten für HopIn."
undo = "reversible"
undo_how = "Die Adresse zurück auf den alten Stand stellen und HopIn dort wieder starten, in wenigen Minuten."
contract = "Amends #1's invariant 1 line A (ruled 06.11.2026), whose contract keeps live on hal until the cutover, for one host: at hopin's live go, on the operator's word (O10), only `hopin.hallucinai.de` moves to the platform tunnel; every other host stays on hal until P10. The head edits #1's invariant 1 in the same change. S5 builds hopin's manifests for that live go; before the live go, #48 has platform's OpenTofu adopt `hopin.hallucinai.de` and the zone's `deny_public` rule for `/auth/backchannel-logout` from hal's state, and hal's head then drops hopin from hal's catalog in a hal lane (R4, hal's call), so the path is never public; the live go itself runs `hal stop hopin`, the record's move to the platform tunnel through #48's root, and the rollback drill (record back through the same root, `hal start hopin`). The platform's hopin has its own session key (O5) and keeps sessions in memory, so every session ends at the switch and at a rollback, and people sign in again. hal's data, `~/srv`, 8081 and 8090 stay untouched."
[[expectation.option]]
name = "Alle ziehen zusammen um"
effect = "HopIn läuft auf der neuen Plattform nur zum Testen; seine Adresse bleibt bis zur großen Umstellung beim alten Stand."
same_as_today = true
notice = "Bis zur großen Umstellung ändert sich für dich nichts."
cost = "Der erste echte Betrieb kommt erst mit der großen Umstellung; was nur echte Besucher zeigen, fällt erst dann auf."
undo = "reversible"
undo_how = "Später kann HopIn trotzdem vorher umziehen."
contract = "#1's invariant 1 line A stays as ruled: live stays on hal until the cutover (P10). S5 still builds and proves hopin in k3d and on the PC; no DNS record moves before P10, and the live go becomes part of P10's cutover."
[[expectation]]
text = "HopIn geht auf der neuen Plattform mit zwei benannten Lücken live, bis Anmeldedienst und Quiz-Dienst umgezogen sind."
asked_on = 2026-11-06
ruling = "A"
ruled_on = 2026-11-06
card_hash = "sha256:e93e9404b19136bc98374686d76b21ca469c556e9f9f56b74944ac152b62c680"
via = "chat"
question = "Darf HopIn auf der neuen Plattform starten, obwohl eine Abmeldung anderswo erst nach bis zu zwölf Stunden wirkt und das Quiz so lange ausfällt?"
problem = "Auf der neuen Plattform wirkt eine Abmeldung woanders erst nach bis zu zwölf Stunden, und das Quiz fällt aus."
decision_kind = "data_risk"
area = "HopIn"
topic = "Lücken bis zum Umzug"
recommended = "A"
reason = "Die Anmeldung erreicht HopIn auch auf der neuen Plattform über das Internet. Nur zwei Wege laufen beim alten Stand über dessen internes Netz: die Nachricht, mit der der Anmeldedienst eine Abmeldung an HopIn meldet, und der Quiz-Dienst. Ein Umweg dorthin würde das interne Netz des alten Stands öffnen. Eine Sitzung in HopIn endet nach höchstens zwölf Stunden (gelesen im Code), und der Rückweg ist ein Adresseintrag."
[[expectation.option]]
name = "Mit beiden Lücken starten"
effect = "HopIn läuft auf der neuen Plattform, und die Anmeldung geht wie heute. Eine Abmeldung oder Entfernung woanders beendet die HopIn-Sitzung erst, wenn sie nach höchstens zwölf Stunden abläuft. Das Quiz meldet nicht verfügbar."
notice = "Anrufe und Anmeldung laufen wie heute; das Quiz zeigt, dass es nicht verfügbar ist."
cost = "Wen du aus HopIn entfernst, kann bis zu zwölf Stunden weiter Anrufe starten. Beide Lücken schließen sich erst, wenn Anmeldedienst und Quiz-Dienst umgezogen sind."
undo = "reversible"
undo_how = "Die Adresse zurück auf den alten Stand stellen; dort gelten beide Lücken nicht."
contract = "hopin goes live on the platform with two named gaps until P7. The public sign-in works, since hopin reaches Keycloak over the internet. Keycloak's back-channel logout cannot reach the platform's hopin (Keycloak calls it only on hal_backchannel_hopin, and the zone blocks the path publicly), so a Keycloak sign-out or a removal from /apps/hopin ends a platform session only at its 12-hour lifetime. The quiz game master reads unavailable (`HOPIN_AGENT__GATEWAY__ORIGIN: http://gateway.invalid`, no `litellm_key` file), which the live go's drive observes. No host port, NodePort, hostNetwork or bridge to hal's Keycloak or LiteLLM is built. P7 closes both gaps; the gaps are named on #1's P6 and P7 rows."
[[expectation.option]]
name = "Auf den Umzug warten"
effect = "HopIn zieht erst um, wenn Anmeldedienst und Quiz-Dienst auf der neuen Plattform laufen; bis dahin bleibt es beim alten Stand."
same_as_today = true
notice = "Für dich ändert sich bis dahin nichts."
cost = "Der erste echte Betrieb rückt um die Arbeit an Anmeldedienst und Quiz-Dienst nach hinten, voraussichtlich um Tage."
undo = "reversible"
undo_how = "Später trotzdem mit den Lücken starten."
contract = "hopin's live go waits for P7 (Keycloak and LiteLLM on the platform). S5 still builds and proves hopin in k3d and on the PC, with the gateway origin and key set by P7's names."
[[expectation]]
text = "HopIn vermittelt Anrufe auf der neuen Plattform wie heute auch über den Umweg-Server von Cloudflare."
asked_on = 2026-11-06
ruling = "A"
ruled_on = 2026-11-06
card_hash = "sha256:26a1093ff7a42bb61d9431e23af34e4557664258e6acd8fc5e41c343e21bea86"
via = "chat"
question = "Sollen Anrufe auf der neuen Plattform auch dann zustande kommen, wenn ein Netz direkte Verbindungen sperrt, wie heute über einen Umweg bei Cloudflare?"
problem = "Ohne den Umweg scheitern Anrufe aus strengen Netzen, etwa manchen Mobilfunk- oder Firmennetzen."
decision_kind = "product"
area = "HopIn"
topic = "Umweg für Anrufe"
recommended = "A"
reason = "Beim alten Stand ist der Umweg heute an (gelesen in HopIns Startkonfiguration). Ohne ihn würde HopIn für manche schlechter als heute."
[[expectation.option]]
name = "Umweg bleibt an"
effect = "Der Schlüssel für den Umweg wird für die neue Plattform kopiert; Anrufe verhalten sich wie heute."
same_as_today = true
notice = "Anrufe kommen zustande wie heute. Den Schlüssel legst du einmal selbst verschlüsselt ab."
cost = "Bis der alte Stand abgeschaltet ist, liegt der Schlüssel zweimal vor."
undo = "reversible"
undo_how = "Mit einem Neustart ausschalten; schon ausgegebene Zugänge gelten noch bis zu zwölf Stunden."
contract = "`HOPIN_TURN_PROVIDER: cloudflare` with `HOPIN_TURN_KEY_ID_FILE` and `HOPIN_TURN_API_TOKEN_FILE` under /run/secrets; the operator copies hal's two TURN values into the Secret hopin in platform-secrets (O5). hal's copy leaves with hal (P11)."
[[expectation.option]]
name = "Umweg wird abgeschaltet"
effect = "HopIn bietet nur direkte Verbindungen an; der Schlüssel wird nicht kopiert."
notice = "Aus manchen Netzen kommt ein Anruf nicht zustande."
cost = "Anrufe aus strengen Netzen scheitern, bis der Umweg wieder an ist."
undo = "reversible"
undo_how = "Später den Schlüssel ablegen und einschalten."
contract = "`HOPIN_TURN_PROVIDER: off`, no TURN files in the Secret hopin, and no TURN value is copied."
Activity
FlexOr2 commentedon Nov 6, 2026
Journal, 06.11.2026: hopin's first attested release (O9's first half)
ghcr.io/overnightworks/hopin:2026.11.06.2@sha256:aaa309151f36d714fe49d8a285f85c4942b4e145e0befd32472167978c496ae6, tagv2026.11.06.2on hopin main 127f26c.image.ymlandrelease-tag.ymlare all green at quality 7715afe. The operator dispatched it with his own word in hopin's session.hal finds its own tunnel by id, so a second tunnel in the account bre…
A catalog host can name another tunnel that serves it (#1102)
A temporary tunnel-check host on the platform's tunnel (#1104)
FlexOr2 commentedon Nov 6, 2026
Journal, 07.11.2026: O8. Asked to grant
platformRead under the hopin package's Actions access, the operator answered in chat: „bzw ist doch schon rolde read“, meaning the role there is already Read. The head has not seen the page. The grant is proven by the first CI pull of hopin's image with platform's job token (E1 in #49), and a missing grant fails that pull loud.FlexOr2 commentedon Nov 6, 2026
Journal, 07.11.2026: O8 confirmed by the operator in chat: „hopin ist admin und platform ist role read“. Under the hopin package's Actions access,
hopinholds Admin andplatformholds Read. O8 is done; E1 still proves the pull.FlexOr2 commentedon Nov 7, 2026
The operator reported on 07.11.2026, in aco's head session, that he has done O8: on hopin's GHCR package, Manage Actions access → platform: Read. aco's head could not verify it, because its token lacks read:packages. H13 and S7 wait for it.
FlexOr2 commentedon Nov 7, 2026
Follow-ups from #82's lane (PR #96 at 7cdb037), routed here before its landing:
hopin(oidc_client_secret,session_secret,turn_key_id,turn_api_token) a second time instead of reading them from one owner. A new key fails CI at hopin's start, so drift is visible. Revisit when the second app's Secrets arrive.kubectl wait kustomizations --all --timeout=5mand the Kustomizationapps' owntimeout: 5m0send together, so the job turns red with only "timed out waiting for the condition on kustomizations/apps"; the cause (a refused setting in H4, a 401 in H5) shows only in the failure-only step "Print Flux's state", in H4 only because it fits its 20-line tail. Polish: the job is red and names the setting. Owner: S5b, which changes the same step.admission's objects, the root's" and a ragged paragraph. Polish; S5b re-flows it when it changes that paragraph.FlexOr2 commentedon Nov 7, 2026
Follow-ups from #82's delta run (PR #96 at 0266abc), routed here before its landing:
/run/secretsare proven only through the Ready pod and the Secret volume in the inventory; no step lists them. hopin's start checks the client secret, the session secret and both TURN files (decision 7), and H4 drives a missing session secret, so a Ready pod implies all four were read. H13 on the PC lists the four file names, never their content.infrastructure/traefik/: "k3s's default cluster CIDR, whichhost/k3s/config.yamlkeeps" is true only because the config sets nocluster-cidr; the job checks the relation itself. It is Traefik's web port admits only the tunnel, trusts its forwarded headers and carries hal's hardening #103's wording. S5b: hopin is routed through Traefik and isolated, trusting only Traefik's forwarded addresses #117 tightens it when it changes that paragraph.hopin runs in k3s from its attested private ref, pulled through ghcr-…
FlexOr2 commentedon Nov 7, 2026
S5a landed as 226d67a (#82, PR #96; the operator's word in platform's chat). platform-secrets PR #9 landed as 89f74c5 (#8, merged by the operator by hand).
H13 on the PC, first part: held (07.11.2026), read by the head read-only:
platform-secretsapplied 89f74c5, and the Secretshopinandghcr-pullexist in namespacehopin(names only).imagePullSecrets: ghcr-pull(the operator's PAT, Q6). ItsimageIDisghcr.io/overnightworks/hopin@sha256:aaa30915…, the pinned digest. The Deployment is 1/1,Recreate.appsturned Ready atmain@sha1:226d67aat 14:00:30Z. It had failed once, before the Secrets arrived, and passed on its next interval with no command./healthand the start page throughkubectl port-forward svc/hopin; both answered.What H13 still needs:
/healththrough Traefik from namespacecloudflared, which needs #117's route.FlexOr2 commentedon Nov 7, 2026
Follow-ups from S5b (#117, PR #118), none blocking:
clusterred on its bump PR, with a message that does not name the cause. The cell is ruled as 40. A constraint form needs the operator's ruling: starts until the first 429 within a cap, at least one admitted, then another address not refused with 429.trustedIPsand hopin'sHOPIN_TRUSTED_PROXIES. Each has a check that reads the cluster, so drift is caught. The second app that trusts proxies decides whether one owner holds the value.hopin is routed through Traefik as hopin.hallucinai.de, admits only T…
FlexOr2 commentedon Nov 7, 2026
H13 holds on the PC, as read by the head on 07.11.2026:
appsis Ready at 21046c9.HOPIN_TRUSTED_PROXIES, the Ingresshopinand the NetworkPolicyhttp-from-traefik-only.ghcr-pull(platform-secrets#14)./healththrough Traefik'sweb(ClusterIPtraefik, port 80) with Hosthopin.hallucinai.degets 200, and with an unrouted host it gets 404. The request was a GET from the node, which the policy admits like the kubelet's probes.Next is #45's C12 (#128, PR #130), on the operator's merge.
FlexOr2 commentedon Nov 7, 2026
The live go's operator blocks, from hal's head (derived from hal's code at b5676a56, 07.11.2026; nothing run). Order agreed in #45's journal: step 0 the head's probe of
tunnel-check.hallucinai.de/health, step 2 #132 merged and the check host's 404 seen, step 3 hal#1153 (hal#1151) landed, then:Step 4, the switch (each command takes one exact
yes; a non-zero exit stops the chain):hal stop hopinprints✓ hopin switched off barrier up · stopped · volumes kept · …;hal updateasksType yes to update the checkout to <sha>:and ends✓ checkout updated to <sha> · dependencies synced;hal applyshows four Cloudflare changes: the tunnel-check record destroyed, hopin's record target changed to the tunnelplatform, the deny rule updated with hopin kept, hopin's line dropped from hal's tunnel. It expectsPlan: 0 to add, 3 to change, 1 to destroy., asksType yes to apply exactly this plan:and ends withhal apply: done.✗orerror:line athal stop(exit 3);hal updaterefusing (local changes, no terminal, noyes);cloudflare: refused,cloudflare: abortedorcloudflare: tofu apply failed, orhal apply: the publish did not end clean.yes.hal start hopinrestores the old state.Rollback (only after hal's head has flipped the line back on hal's main; this order, per this item's Rollback):
hal start hopinasksType yes to start hopin:and endshal start: hopin is on, or printshal start: no change, hopin is not switched off;hal applychecks hopin's/healthlocally first, then plans hopin's record back to hal's tunnel and the line back into it, and endshal apply: done.hal start: aborted,partial,refusedorcannot plan(exit 3);cloudflare: hopin.hallucinai.de did not answer /health …followed byrefused, a ready host does not answer. That means hal's hopin is unhealthy: stop and report.hopin's public host moves to the platform tunnel (#1153)
FlexOr2 commentedon Nov 7, 2026
Live go, step 4 done (07.11.2026). The operator ran
hal stop hopin && hal update && hal apply(hal#1153 at 6ffa64e0) and reported done. Read at 19:43:06Z:https://hopin.hallucinai.de/health→ 200,/→ 401 (sign-in),POST /auth/backchannel-logout→ 403 at the edge (R4 holds),tunnel-check.hallucinai.deresolves no more. Five public/healthprobes raised Traefik's hopin service request counter on the PC from 25 to 34, so public requests reach the platform's Traefik; hal's hopin is stopped. Next: the drive at mobile and desktop width, the operator's own sign-in (the quiz reading unavailable), and the rollback drill.FlexOr2 commentedon Nov 7, 2026
The live go's drive (delegated tester, headless Chromium 1.49.1 through Playwright, 07.11.2026, after step 4), at 390x844 and 1280x800. Opening
/gets 200 after the redirect toauth.hallucinai.dewithclient_id=hopinandredirect_uri=https://hopin.hallucinai.de/auth/callback, with 0 console errors, 0 failed requests and no horizontal scroll at either width. The login page renders./healthreturns 200{"status":"ok"}. The login footer's Impressum, Datenschutz and Nutzungsbedingungen on hallucinai.de each return 200. Not driven: a real sign-in, the callback, and the quiz reading unavailable, which are the operator's own sign-in. The tester's follow-ups describe hopin as it is today (no public page of its own; the legal links live on the login page), not the move.FlexOr2 commentedon Nov 7, 2026
Live go done: hopin serves
hopin.hallucinai.defrom the platform since 19:43Z on 07.11.2026 (journal above), and the operator signs in through it. He declined the rollback drill: „3 nein sinnlos? was soll an platform nicht gehen? wir haben kein durck“ (07.11.2026, platform chat). If one is ever needed, the way back stays as recorded above:hal start hopinfirst, then the catalog line flipped back,hal update,hal apply. Done when is met.