SSO auto-redirect: the fleet coverage matrix
Layer 10. Edge. Edge
Status: MEASURED INVENTORY, 2026-08-10. Nothing in this document arms
anything. Every app named here keeps ssoAutoRedirect defaulting to
false, and every native knob keeps its own default of false.
Companion to docs/proposals/sso-auto-redirect.md, which is the SPEC and
the authority on every verdict it states. This document is the
CATALOGUE-WIDE answer to one question the proposal never answers directly:
Given that the mechanism shipped, how much of my fleet would actually auto-redirect if I turned it on, and why is that number not “all of it”?
The three numbers
Section titled “The three numbers”Denominator: the 43 SSO-plugged apps (an oidc.plug.yaml or a
saml.plug.yaml), which is the corpus
TestEverySSOPluggedAppIsNamedInTheAutoRedirectAudit enforces. Three of
those 43 have no login surface at all by design (identity-broker,
mailbox-reconciler, nextcloud-picker), leaving 40 apps with a login
surface.
| count | what it means | |
|---|---|---|
| Could auto-redirect today | 24 | the mechanism is shipped AND wired in the app. An operator flips a flag; no code lands first. |
| Need a sibling lane’s work first | 6 | a mechanism is available and measured, but nothing is declared or wired yet. |
| Structurally excluded | 10 | measured to fail an author check in a way no declaration can fix at the pinned version. |
24 + 6 + 10 = 40. Adding agent-canvas and graphiti (forward-auth
gated, so already auto-redirecting, but carrying no SSO plug and therefore
outside the 43) the fleet-wide “does or could auto-redirect today” figure
is 27.
All 16 native-knob vars were re-read from vars.yaml on 2026-08-10 and
every one of them is false:
keycloak.autoRedirect librechat windmill wordpress nextcloud matrix gitea gitlab paperless immich ciso-assistant defectdojo-classic wazuhoidcOnly journivssoOnly vaultwarden penpotkeycloak.makeDefault midpoint midpoint-classicpenpot joined that ssoOnly row on 2026-08-15 and is the one entry in it
that is not in bucket 1: it is a bucket-3 app for AUTO-REDIRECT (the
exclusion below is unchanged and unchangeable at this pin) that nonetheless
has a native FORCE-SSO knob. The two are different mechanisms and only one of
them was ever excluded; see “penpot’s native knob” under bucket 3. It brings
the native-knob count to 17.
★ Why the fleet does not auto-redirect, in one paragraph
Section titled “★ Why the fleet does not auto-redirect, in one paragraph”Not because coverage is thin: 24 of 40 are ready. Because there are two different operator surfaces and only one of them is documented as the operator surface. The proposal says the opt-in is a scoped config key, deliberately not a var:
shc config set --deployment <d> ssoAutoRedirect trueThat key is read in exactly one place:
ce/modules/app/ingress_materialize.go, gating the edge router built from
an app.yaml sso_redirect: block. It reaches 6 apps. The other
16 were shipped as Implementation A and read a per-app var
(vars.keycloak.autoRedirect, oidcOnly, ssoOnly, makeDefault)
consumed by the app’s own compose template. shc config set ssoAutoRedirect true does not arm a single one of them, and it fails
silently, because a config key with no reader is indistinguishable from a
config key whose readers all declined.
An owner who armed four apps in a row and saw four login pages was most likely arming through the documented key against apps that only listen on the undocumented one.
The second-order reason is that rollout steps 1-3 of the proposal were executed after all, by other lanes, over 2026-08-07..10, in a shape the proposal was never updated to describe. The proposal still reads as though only Implementation C shipped with frappe as its single consumer. That is no longer true and has not been true for days.
How to read the mechanism column
Section titled “How to read the mechanism column”- A: the app’s own native auto-login setting, rendered by its
compose/config template from a per-app var. Armed by a
--var, applied by a redeploy. - B: a Traefik narrow redirect from
app.yamlsso_redirect: {login_path, sso_start}. Armed by the scopedssoAutoRedirectkey, applied on the next ingress sweep. - C:
sso_redirect: {login_path, sso_selector}, the response injector, for apps whose SSO href is minted per render. Same arming as B. - T1: forward-auth gate (
access: gated). Not a tier of this feature: an unauthenticated browser is already redirected to Keycloak with no flag at all. - NONE-structural: measured to be unreachable by any of the above at the pinned version.
“Declared?” answers whether the app ships the wiring today, not whether it is on. Everything below is off.
Bucket 1: could auto-redirect today (24)
Section titled “Bucket 1: could auto-redirect today (24)”1a. Edge router declared, armed by ssoAutoRedirect (6)
Section titled “1a. Edge router declared, armed by ssoAutoRedirect (6)”| app | pinned | SSO | mech | declared | bypass (check 4) | logout-safe (check 5) | verdict |
|---|---|---|---|---|---|---|---|
| kutt | 3.2 | oidc | B | yes | ?sso=false | yes: logout lands on /, not /login | READY |
| affine | 0.27 | oidc | B + launcher auto | yes | ?sso=false | no: no end_session in the 0.27.3 bundle | READY, logout debt |
| affine-rs | 1.0 | oidc | B + launcher auto | yes | ?sso=false | no: no RP-initiated logout either side | READY, logout debt |
| huly | 0.7 | oidc | B | yes | ?sso=false | no: loops without a page load | READY, logout debt |
| windmill | 1.776.0 | oidc | B and A | yes (both) | /user/login?no_sso=1 | no: unread at the pin | READY, logout debt |
| frappe | 15.118 | oidc | C | yes | /login?sso=false | no: logout() never calls end_session | READY, logout debt |
Five of these six fail check 5. That is recorded, argued and guarded in
tests/unit/catalog/sso_logout_declaration_test.go’s
ssoLogoutTradeAccepted map, and four of the five rows there say
“OWED, NOT ACCEPTED”, meaning nobody has yet read the pinned version
to find out whether it can end the IdP session. Arming these before the
anti-bounce latch reaches the estates means logout signs the user straight
back in.
1b. Native knob wired, armed by a per-app var (16)
Section titled “1b. Native knob wired, armed by a per-app var (16)”| app | pinned | SSO | mech | var | bypass (check 4) | logout-safe (check 5) |
|---|---|---|---|---|---|---|
| gitlab | 18.11 | oidc | A | keycloak.autoRedirect | /users/sign_in?auto_sign_in=false | yes: the zzz_shc_oidc_logout.rb initializer sends sign-out on to Keycloak’s end-session endpoint with id_token_hint |
| nextcloud | 34 | oidc | A | keycloak.autoRedirect | /login?direct=1 | yes: user_oidc single_logout |
| librechat | 0.8 | oidc | A + B/auto (2026-08-17) | keycloak.autoRedirect | /login?redirect=false: one URL stops BOTH halves: bypass_query: redirect at the edge, the same key the shell reads | yes: OPENID_USE_END_SESSION_ENDPOINT |
| matrix | 1.158 | oidc | A | keycloak.autoRedirect | NONE (declared closed_by) | yes: logout_redirect_url |
| wazuh | 4.14 | oidc | A | keycloak.autoRedirect | NONE (declared closed_by) | yes: logout_url |
| gitea | 1.27 | oidc | A | keycloak.autoRedirect | NONE (declared closed_by) | no: trade accepted at 1.27.1 |
| immich | 3.0 | oidc | A | keycloak.autoRedirect | /auth/login?autoLaunch=0 | app-side; trade accepted |
| paperless | 3.0 | oidc | A | keycloak.autoRedirect | /accounts/login/?loggedout=1 | trade accepted: logout lands off the SSO path |
| wordpress | 6.9 | oidc | A | keycloak.autoRedirect | /wp-login.php?login-error=1 | yes: endpoint_end_session |
| defectdojo-classic | 2.58 | oidc | A | keycloak.autoRedirect | /login?force_login_form | no: trade accepted at 2.58.4 |
| ciso-assistant | 3.20.8 | oidc | A (caddy 302) | keycloak.autoRedirect | any query string, e.g. /login?local=1 | yes: "slo_enabled": True |
| journiv | 0.1.0-beta.23 | oidc | A | oidcOnly | NONE (declared closed_by) | no: OWED, unread |
| vaultwarden | 1.37 | oidc | A | ssoOnly | NONE (declared closed_by) | no: OWED, unread |
| midpoint | 4.10 | oidc | A | keycloak.makeDefault | ★ UNDECLARED: see below | UNMEASURED |
| midpoint-classic | 4.8 | oidc | A | keycloak.makeDefault | ★ UNDECLARED: see below | UNMEASURED |
| netbird | 0.76 | oidc | already-on | None | no local accounts exist | n/a |
mautic (7.1, saml) belongs to the same “already auto-redirecting once
the plug is connected” class as netbird; its MainEntryPoint fires
unconditionally, escape /s/login. It is counted in bucket 1 as the 24th
app.
★★ The two rows that are a live defect: midpoint and midpoint-classic
Section titled “★★ The two rows that are a live defect: midpoint and midpoint-classic”Both apps ship vars.keycloak.makeDefault: “Make the OIDC sequence
the default for the GUI channel instead of the internal login form”. That
is an SSO-arming, lockout-shaped switch by every definition this
catalogue uses. Neither guard can see it, because both name-lists are
spelled out longhand and makeDefault is in neither:
sso_logout_declaration_test.go ssoArmingVars = autoRedirect, ssoOnly, oidcOnlylocal_login_declaration_test.go lockoutShapedVars = autoRedirect, ssoOnly, oidcOnly, forceSso, force_ssoConsequences, both silent:
- Neither app declares
local_login: closed_by:, so a restored midPoint whose GUI default sequence is OIDC will not emitX200042NOLOGIN. The G8 post-restore usable-login check reports it fine. - Neither app is in
ssoLogoutTradeAccepted, so the logout invariant passes over an app that arms SSO and configures no end-session path.
Both guards go green on midpoint today for the same reason the 2026-08-02 sweep went green without sonarqube: the corpus definition, not the app.
Measured 2026-08-10 against evolveum/midpoint:4.10.3-alpine (the
pinned digest), booted on postgres with the stock security policy and no
OIDC configured:
GET /midpoint/ 302 -> /midpoint/self/dashboardGET /midpoint/login 302 -> /midpoint/login?0 (Wicket page versioning)GET /midpoint/auth/emergency 302 -> /midpoint/loginGET /midpoint/auth/oidc 302 -> /midpoint/loginGET /midpoint/logout 302 -> /midpoint/login <-- check 5Two readings, and they point in opposite directions:
- Implementation B must not be declared for midPoint. Logout lands on
/midpoint/logincarrying no bypass query. A router anchored at^{{ path('midpoint') }}/midpoint/login$matches that hop, so signing out would 302 into the IdP and Keycloak’s live cookie would sign the user back in. That is the sonarqube shape exactly, and check 5 alone is disqualifying. - Implementation A is unaffected by that, because
makeDefaultmoves the default sequence, and midPoint owns its own post-logout landing. What is unmeasured is whether the OIDC sequence ends the Keycloak session, and whether a break-glass sequence stays reachable. With the stock policy/midpoint/auth/emergencyis not a distinct door; it bounces to the same login page.
Owed, and deliberately not done in this document (it edits app.yaml,
which is a different lane’s consent): add makeDefault to both name
lists, and either declare local_login: closed_by: [keycloak.makeDefault]
on both midpoint apps or record the bypass that makes it survivable.
Do not do one without the other: a closed_by on an app that does have
a door is the false alarm that teaches operators to ignore the true ones.
Bucket 2: a mechanism exists, nothing is wired (6)
Section titled “Bucket 2: a mechanism exists, nothing is wired (6)”Nothing here may be armed; each needs a lane to land a declaration first.
| app | pinned | SSO | mech available | what is missing |
|---|---|---|---|---|
| superset | 6.1.0 | oidc | B, measured | a sso_redirect: declaration, and a logout fix first |
| mailcow | 2026.07a | oidc | A (force_sso) | the knob is not wired to a var; and its login form is inline on / |
| peertube | 8.2 | oidc | A (redirect_on_single_external_auth) | not wired to a var |
| matomo | 5 | oidc | UNMEASURED | never audited past “no knob at pin” |
| postiz | 2.23 | oidc | UNMEASURED | never audited past “no knob at pin” |
| ivatar | 2.1 | oidc | UNMEASURED | never audited past “no knob at pin” |
superset 6.1.0: MEASURED, and it is the best remaining candidate
Section titled “superset 6.1.0: MEASURED, and it is the best remaining candidate”Measured 2026-08-10 by reading the shipped code inside the pinned image
apache/superset:6.1.0@sha256:fb34645…, which carries
flask-appbuilder 5.0.2. The app flips AUTH_TYPE = AUTH_OAUTH
with a single provider literally named keycloak (compose.yaml), so the
entry path is fixed by the app and needs no providerSlug plug field.
| check | verdict | evidence |
|---|---|---|
1. login_path server-visible | pass | /login/ is a server-rendered Jinja template, not an SPA route. FAB routes @expose("/login/"). |
| 2. plain GET, state minted server-side | pass | @expose("/login/<provider>") → session["oauth_state"] = random_state then authorize_redirect(...). The state is a JWT and the verifier lives in the Flask session cookie. A 302 to /login/keycloak is byte-identical to a human clicking the button. |
| 3. stable URL | pass | /login/keycloak is fixed; no per-render token. |
| 4. bypass reaches the local form | ★ fail, and not for the usual reason | see below |
5. logout does not land on login_path | ★ fail, fixable | see below |
Check 5. AuthOAuthView inherits logout() from AuthView:
logout_user() then redirect(config.get("LOGOUT_REDIRECT_URL", get_url_for_index)). LOGOUT_REDIRECT_URL is absent from
/app/superset/config.py at this pin (verified by importing the module
inside the image) so logout lands on the index, which for an anonymous
user bounces to /login/. Declaring B today would reintroduce the
sonarqube logout loop.
Unlike sonarqube, this one has a fix that is a one-line app
change: set LOGOUT_REDIRECT_URL to Keycloak’s end_session_endpoint.
It moves the post-logout landing off login_path and delivers
RP-initiated logout, clearing check 5 and the logout invariant in the same
stroke. Nothing upstream has to change.
Check 4 is the real blocker and it is worth stating loudly. Under
AUTH_TYPE = AUTH_OAUTH, FAB renders login_oauth.html, which contains
only provider buttons; there is no username/password form at all. So
/login/?sso=false reaches a page, but not a credential door. This is not
a consequence of auto-redirect: binding the oidc integration to
superset already closes local login today, unconditionally, with no var
involved. Superset therefore needs a local_login: declaration on the
integration edge regardless of whether auto-redirect is ever declared,
and it has none.
Implementation C is structurally impossible here, which is worth
recording so nobody tries it as a workaround for check 4. The template’s
provider button carries no href:
<a id="btn-signin-{{provider.name}}" class="btn btn-primary btn-block" type="submit">with navigation supplied by a click listener. The selector injector
would find the element, read no href, log SSOSEL-NOHREF and serve the
page unchanged forever, a declaration that is inert rather than wrong,
which is the harder failure to notice.
Bucket 3: structurally excluded, measured (10)
Section titled “Bucket 3: structurally excluded, measured (10)”| app | pinned | SSO | failing check | evidence |
|---|---|---|---|---|
| sonarqube | 26.7 | saml | 5 and 6 | logout and permission-denial both funnel to /sessions/new; no SAML SLO at the pin. exclude_query answers 6; 5 is the binding blocker. |
| dependency-track | 5.0 | oidc | 2 | PKCE verifier minted in the browser and never leaves browser storage; served HTML carries no <a>. |
| penpot | 2.17 | oidc | 1 and 2 (EDGE only) | login page is the SPA fragment /#/auth/login, never sent to the server; /api/auth/oidc is :allowed-methods #{:post} and answers 200 JSON. Both refusals are PERMANENT for the edge grades. ★ CLICK grade declared 2026-08-22: see below. ★ has a native FORCE-SSO knob: see below. |
| glitchtip | 6.2.5 | oidc | 2 | ★ newly measured: see below |
| minio-classic | 2025.4.22 | oidc | 2 | ★ newly measured: see below |
| minio | 2025.9.7 | oidc | 2 | returns JSON for the SPA to navigate itself. |
| supabase | 2.0 | oidc | 2 | POST-only initiation (proposal, 2026-08-02). |
| highlight | 0.5.6 | oidc | 1 | SPA shell, no server-visible login path (fleet path table). |
| monica | 4.1 | oidc | n/a | has no SSO at 4.1.2: the keycloak vars are reserved for the unreleased v5. |
| blockscout | 9.0.2 | oidc | n/a | the sso plug is inert at the pin: ACCOUNT_KEYCLOAK_* exists upstream only from v10.1.0. |
The last two are excluded because there is no working SSO to redirect into, not because the redirect mechanism fails. They are bugs tracked by the proposal, restated here so the count is honest.
Two corrections to penpot’s row, both made 2026-08-15 by reading the pinned
penpotapp/backend:2.17.0 jar (digest sha256:471cdebf185b…) rather than
re-copying the audit:
- the old evidence cited
GET /api/rpc/command/login-with-oidc → 405. That 405 is real and proves nothing:app/rpc.cljraises:method-not-allowedfor any GET on/api/rpc/command/<name>not beginningget-, before the handler is looked up, so a nonexistent command answers it too.login-with-oidcwas never an RPC method. The real refusal is the route table inapp/auth/oidc.clj(["" {:allowed-methods #{:post}}]) andauth-handlerending{::yres/status 200 ::yres/body {:redirect-uri …}}. app.yaml has carried the correction since 2026-08-11; this table had not. - the row said nothing about force-SSO, and readers took bucket 3 to mean “penpot can do nothing here”. It cannot AUTO-REDIRECT. It can go SSO-only.
★ A third correction, 2026-08-22: it CAN auto-redirect, on the click
grade. Both refusals above are about the EDGE and both remain permanent:
a fragment is never transmitted, and l.replace() can only GET. The CLICK
grade (sso_click_selector:) navigates nowhere: it waits for penpot’s own
SSO control and presses it, and the page issues that POST itself with
penpot’s own state. What made it possible was one plugin field,
match_hash: true, which moves the injected payload’s route string from
location.pathname to pathname + hash, necessary here because penpot’s
login screen and a signed-in user’s dashboard share the pathname /, and
because the PV/LO logout-transition latch is keyed on that same string and
would otherwise never fire on a hash-routed logout. The mode is off by
default and penpot is its only caller (swept: it is the catalogue’s only
fragment-routed login). The declaration is login_path: "(?i)^/#/auth/login"
match_hash: true+sso_click_selector: "a.main_ui_auth_login__btn-oidc-auth"+launcher: auto, ledgered intests/unit/catalog/sso_click_selector_test.go.
★ penpot’s native knob: bucket 3 for redirect, bucket 1b for force-SSO
Section titled “★ penpot’s native knob: bucket 3 for redirect, bucket 1b for force-SSO”--var ssoOnly=true renders disable-login-with-password in place of
enable-login-with-password into PENPOT_FLAGS on both penpot services,
which is upstream’s own force-SSO flag and closes both surfaces at once:
app/common/flags.cljc:209:parsefolds the list left anddisable-Xis(disj result :X);app/config.clj:298passesPENPOT_FLAGSlast, so the token beats the:enable-login-with-passwordinflags/default.app/rpc/commands/auth.clj:56-63: the RPC raises:restriction/:login-disabled, so a hand-crafted POST is refused too.app/main/ui/auth/login.cljs:257-260(from the shippedmain-auth.js.map): the SPA renderslogin-form*only on:login/:login-with-password/:login-with-ldap, so the page becomes the OIDC button and nothing else.
It is a one-button login page, not a 302: the mailcow/vaultwarden class
(bucket 1b), and the fragment-routing exclusion above is untouched by it. The
knob is gated on a bound oidc integration (local_login.gated_on_edge: true) because penpot has no third state: closing the form with no IdP wired
renders a login page carrying neither a form nor a button. It ships false,
is deaf to the fleet-wide ssoAutoRedirect key (a lockout decision, not a
redirect preference), and has no bypass: local_login.closed_by: [ssoOnly]. Pinned by tests/unit/catalog/penpot_sso_only_test.go.
Penpot’s logout was read at the same time and is a finding, not a trade:
app/auth/oidc.clj:56 takes end_session_endpoint off the IdP’s own
discovery document and ::logout (auth.clj:143-169) answers a
:redirect-uri to it, so penpot does RP-initiated logout in app code with no
app setting to write, the immich/journiv shape, recorded in
ssoLogoutTradeAccepted.
★ glitchtip 6.2.5: the proposal’s “inconclusive” row, now resolved
Section titled “★ glitchtip 6.2.5: the proposal’s “inconclusive” row, now resolved”The proposal recorded glitchtip as inconclusive: probed read-only on a
live estate, every path returned an identical 5847-byte shell, so the
backend routes were not observable. It asked for a local bring-up. Done
2026-08-10: glitchtip/glitchtip:6.2.5 on postgres 16 and valkey, with
the socialaccount_socialapp row written exactly as the app’s
oidc.plug.yaml writes it (provider='openid_connect',
provider_id='keycloak', settings={'server_url': …}).
GET / 200 28671 bytesGET /login 200 28671 bytes (identical)GET /profile 200 28671 bytes (identical)GET /accounts/oidc/keycloak/login/ 404GET /accounts/oidc/keycloak/login/callback/ 302 -> /login?socialLoginError=trueGET /_allauth/browser/v1/auth/provider/redirect 405and, read out of the running app:
SOCIALACCOUNT_LOGIN_ON_GET = FalseCheck 2 fails, decisively and twice over. The SPA’s real initiation is
POST /_allauth/browser/v1/auth/provider/redirect: 405 on GET, exactly
as the app’s own plug comment documents. And even the allauth
provider-login route, which does resolve in the URLconf
(allauth_async.urls.view, kwargs {'provider_id': 'keycloak'}), cannot
serve as sso_start: with LOGIN_ON_GET = False allauth 65.16.1 does not
302 on a GET, it renders an intermediate auto-POST form. A Traefik
redirect cannot enter that flow.
Implementation C cannot rescue it: the served /login HTML contains
zero <a> elements and not one occurrence of keycloak, oidc,
provider or socialaccount. The button is Angular-rendered. Same
terminal state as dependency-track, reached by a different road.
★ The generalisation is worth more than the row. Any django-allauth
app at allauth ≥ 0.55 ships SOCIALACCOUNT_LOGIN_ON_GET = False by
default, and therefore fails check 2 on its provider-login route. Check
that setting before auditing an allauth app, not the route’s HTTP status.
★ minio-classic 2025.4.22: inheritance confirmed, no longer assumed
Section titled “★ minio-classic 2025.4.22: inheritance confirmed, no longer assumed”The proposal listed minio-classic as “not audited; expected to inherit
minio’s check-2 failure, unverified”. Measured 2026-08-10 against the
pinned minio/minio:RELEASE.2025-04-22T22-12-26Z with a reachable
discovery document:
GET / 200 1309 bytesGET /login 200 1309 bytes (identical shell, zero <a> elements)GET /oauth_callback 200 1309 bytes (identical)GET /api/v1/login 406 to a text/html Accept; JSON body regardless: {"loginStrategy":"redirect", "redirectRules":[{"displayName":"Keycloak", "redirect":"http://…/protocol/openid-connect/auth?client_id=minio&…&state=…"}]}MinIO does mint the authorize URL and the state server-side; the
failure is purely one of transport: the URL is handed to the SPA as JSON
for it to navigate itself. There is no server-side GET that 302s, and the
console shell carries no anchor for a selector to read. Check 2 fails.
Inheritance from minio is now measured rather than expected.
Incidental, and load-bearing for anyone reproducing this: MinIO refuses
to finish booting while OIDC discovery is unreachable, retrying
Unable to initialize OpenID forever. A dead Keycloak does not degrade
MinIO to local-only login; it prevents the service starting at all. That
is a bigger operational fact than anything in this row.
Not applicable: no login surface (3 of the 43)
Section titled “Not applicable: no login surface (3 of the 43)”| app | why |
|---|---|
| identity-broker | in-cluster sidecar; publicUrl is published empty. Its oidc edge validates tokens minted elsewhere; no browser ever lands on a login page it serves. |
| mailbox-reconciler | headless daemon, no ingress, no HTTP listener. Its oidc plug is only the carrier of Keycloak’s admin credentials. |
| nextcloud-picker | see the decision below. |
nextcloud-picker: the design decision the proposal said was owed
Section titled “nextcloud-picker: the design decision the proposal said was owed”The proposal deferred this: “it is an embeddable iframe bridge, not a destination app; auto-redirecting an iframe’s login into a top-level IdP flow is a different question, and needs deciding before it is measured.”
Decision: out of scope, and structurally so: do not measure it, do not declare it. Two independent reasons, either sufficient:
- It has no login page to redirect. The app owns exactly one
browser-facing OIDC route, its callback (
{{ url('picker') }}/oidc/callback, scoped deliberately tight because this client is the subject of a token exchange for the Nextcloud audience). There is no login form and nologin_pathfor a router to anchor to. - A top-level IdP flow inside a third-party iframe cannot complete anyway. Keycloak serves its login page with frame-ancestors limited to itself; an embedded picker navigating its own frame to the authorize endpoint gets a refusal, not a login. The picker’s session is established by its embedder’s flow, not by one it starts.
This closes the row rather than leaving it open. If the picker ever grows a standalone destination mode, that mode is a new app.yaml question and this decision does not bind it.
ce addons with a login surface (2 of 4)
Section titled “ce addons with a login surface (2 of 4)”addons/ ships four apps; postgres and traefik have no browser login
surface (traefik’s dashboard is a var-gated panel, not an authenticated
app).
| app | pinned | SSO | mech | verdict |
|---|---|---|---|---|
| keycloak | 26.7 | it is the IdP | n/a | N/A by construction: auto-redirecting Keycloak into itself is not a coherent request. |
| openobserve | 0.91.5 | oidc, via a bundled Dex sidecar, enterprise build only | NONE-structural | the auto-login flag is compiled into the SaaS build only (proposal, 2026-08-02). On the OSS arm there is no SSO at all. Root email+password stays as break-glass. |
Apps with no SSO plug and therefore no row
Section titled “Apps with no SSO plug and therefore no row”The catalogue holds 91 apps. 43 carry an SSO plug; the other 48
carry neither an oidc nor a saml plug and are out of scope by
definition:
- 23 datastore / shim standalones:
clickhouse,ethereum,gitlab-runner,hockeypuck,kubo,mail,mariadb,mongodb,mysql,nginx,ollama,pmm,redis,s3, and the nineexternal-*shims. No browser login surface at all. - 4 headless bridges / sidecars:
gitlab-huly,huly-matrix-notifier,hookshot,crowdsec. - 3 test fixtures:
test-data,test-simple,whoami. - 2 forward-auth gated apps:
agent-canvas,graphiti. See below. - 15 apps that simply have no SSO yet:
defectdojo(3.x),dify,docuseal,fleet,fusionpbx,greenbone,hyperswitch,jellyfin,keygen,medusa,metabase,plane,pulp,shlink,uptime-kuma. These are the real growth edge of this matrix: each one becomes a row here the day it gains a plug.
agent-canvas and graphiti deserve their own line: both are
access: gated, so their unauthenticated browsers are already redirected
to Keycloak by forward-auth, with no flag and no declaration. They sit
outside the 43-app denominator only because a gated app needs no SSO plug
of its own, but operationally they are the tier-1 answer, not a gap.
Counting them, the fleet has 27 apps that auto-redirect or could
today, not 24.
★ openreplay: a verdict, and a stale comment worth fixing
Section titled “★ openreplay: a verdict, and a stale comment worth fixing”The proposal lists openreplay 1.27 as “not audited”. It does not need
this feature: apps/flagship/openreplay/app.yaml declares
access: gatedungated_paths: - /ingest/ - /script/so its dashboard already auto-redirects to Keycloak through forward-auth. Counted in bucket 1’s tier-1 group.
Two things a future auditor should not trip over:
- The app.yaml comment above that block is stale. It reads “OpenReplay
does NOT ship an oidc.plug; it rides the gate.” The app does
ship
oidc.plug.yaml, targeting an SHC soft-fork image that adds an additive OIDC login router to the AGPL tree. The prose and the directory disagree, and the prose is the older one. - That combination is the RULING T-1b shape, and it is an open
question rather than a finding here. The proposal’s tier-1 rule says
forward-auth “must NOT front native-OIDC apps” (it 401s non-HTML
traffic and mangles WebSocket upgrades) and openreplay is now both
gated and natively OIDC.
ungated_pathscarves out the tracker ingest, so the measured affine-rs failure mode may not apply. Not measured here. It wants the same EMP-A style upgrade probe affine-rs got before anyone concludes either way.
Still unmeasured, and why
Section titled “Still unmeasured, and why”Listed so silence is not mistaken for a verdict, the defect class this whole programme exists to close.
| app | pinned | what is unmeasured | why not closed here |
|---|---|---|---|
| midpoint / midpoint-classic | 4.10 / 4.8 | checks 4 and 5 with OIDC configured | the app documents seeding the SecurityPolicy and setting globalSecurityPolicyRef as MANUAL steps, and it needs a live Keycloak. The login-surface and logout-landing probes above were run without OIDC; they are properties of the app and stand. |
| matomo, postiz, ivatar | 5, 2.23, 2.1 | everything past “no native knob at the pin” | the 2026-08-02 sweep recorded the configuration answer only. Whether a GET-able OIDC entry exists was never asked. |
| defectdojo-classic | 2.58 | nothing blocking; it has knob A | listed because it sits in the proposal’s not yet audited table for the sso_redirect question, which is the wrong question for it. |
| openreplay | 1.27 | the gated-plus-native-OIDC interaction | see above. |
| midpoint-classic runtime | 4.8 | the 4.10 probes were not repeated on 4.8 | the arming var is confirmed present in the app; the runtime shape is assumed to match and is marked as such. |
What this document deliberately did not do
Section titled “What this document deliberately did not do”- Nothing was armed. No
app.yaml,compose.yamlorvars.yamlwas touched on this branch;ssoAutoRedirectand every native knob keep theirfalsedefaults. Arming is an owner decision and is gated on the anti-bounce latch reaching the estates: landed, not rolled. Arming before it rolls means logout signs the user straight back in, which is precisely the failure five of the six declared apps already carry. - No verdict in the proposal was overturned without re-measuring. The three rows this document changes (glitchtip, minio-classic, superset) each carry their transcript. midpoint’s is new, not a contradiction.
- The two owed app.yaml edits are named, not made:
makeDefaultin both guard name-lists plus alocal_logindecision on midpoint, and alocal_logindecision for superset’s AUTH_OAUTH edge. Both edit apps other lanes own.
Reproducing the measurements
Section titled “Reproducing the measurements”All local, no estate contact, pinned images only.
# glitchtipdocker run -d --name gt --network N -e DATABASE_URL=… -e VALKEY_URL=… \ glitchtip/glitchtip:6.2.5docker exec gt python /code/manage.py shell -c \ "from allauth.socialaccount import app_settings as s; print(s.LOGIN_ON_GET)"
# minio-classic (needs a reachable discovery document or it never boots)docker run -d --name mc -e MINIO_IDENTITY_OPENID_CONFIG_URL=… \ minio/minio:RELEASE.2025-04-22T22-12-26Z server /data --console-address :9001curl -s http://mc:9001/api/v1/login
# superset (read the shipped code; no bring-up needed)docker run --rm --entrypoint sh apache/superset:6.1.0 -c \ 'python -c "import flask_appbuilder,os;print(open(os.path.join(os.path.dirname(flask_appbuilder.__file__),\"security\",\"views.py\")).read())"'
# midpointdocker run -d --name mp -e MP_SET_midpoint_repository_jdbcUrl=… \ evolveum/midpoint:4.10.3-alpinecurl -s -o /dev/null -w '%{http_code} %{redirect_url}' http://mp:8080/midpoint/logout