The flagship app tier: one best-in-class pick per league
Status: MECHANISM SHIPPED; FOUR LABELS ARE ASSIGNED AND THE ROSTER IS
ENFORCED. This change shipped the flagship: key, its registry plumbing, and
the flagship-contract fitness check, with zero apps labeled and the check
green. §7’s candidate table is a proposal with gap notes, not a ruling.
AMENDED 2026-08-11: the first labels land, and errors is no longer
vacant. Three owner-ruled assignments were executed together (the errors
replacement, ruled 2026-08-06, carried out on 2026-08-11):
| League | App | Ruling record |
|---|---|---|
errors | glitchtip | 2026-08-06-errors-and-uptime-league.md §0 |
replay | openreplay | 2026-08-06-session-replay-league.md §0 |
uptime | uptime-kuma | errors-and-uptime §0, awarded on a re-measurement that FAILED GlitchTip’s status-page condition |
and highlight was RETIRED out of the catalogue in the same commit
(apps/RETIRED), which is what emptied the errors vacancy in §7. §7’s rows
for those four apps are superseded by the two decision docs above; the rest of
the table is untouched and still a proposal.
AMENDED 2026-08-14: analytics is filled, and the roster became
machine-held. Two things landed together:
| League | App | Ruling record |
|---|---|---|
analytics | superset | 2026-08-06-bi-league-pick.md §0, the commissioned re-evaluation §7’s analytics row was waiting on |
and the flagship-league-roster fitness check (§4) now pins the ruled
leagues, including storage’s VACANCY, as a set equality. That second half
is the point: until it landed, flagship-contract validated only the claims
that existed, so deleting every label in the catalogue would have kept it
green. That is half-done-sweep F9,
and it is closed here, not merely worked around by adding labels.
Date: 2026-08-06. Measured against shc origin/main @ f2e70a2ec (the
oidc/saml socket split is LANDED at this pin: 31 apps carry oidc.plug.yaml,
mautic and sonarqube carry saml.plug.yaml) and ce
v0.0.23-0.20260806060121-65e94fca6bab (the go.mod pin).
AMENDED 2026-08-07 by owner ruling:
2026-08-07-flagship-payment-axis.md:
“if you are forced to pay to use the app it’s not a flagship.”
Requirement 4 is rewritten around user-facing payment; the no-SaaS concern it
used to carry moves to the managed-cloud eligibility axis (§9), which is
explicitly NOT a flagship criterion. §7 verdicts that turned on the old axis
are corrected in place and marked [2026-08-07]. Two factual errors in §7
measured at the f2e70a2ec pin, since overtaken by commits that landed
before this doc was committed, are corrected in place and marked
[corrected].
1. What a flagship is
Section titled “1. What a flagship is”A flagship is SHC’s single best-in-class pick for a function (its league): the app the platform stands behind when an operator asks “what do I install for git / mail / knowledge / …”. The bar is owner-defined and is not to be watered down:
- SSO ladder. IdP-initiated or auto-redirect preferred (a native app knob or SHC’s traefik auto-redirect), else userless-but-gated (no user model upstream; the forward-auth gate IS the auth layer), else at minimum a visible SSO button, and the claimed tier must be reachable WITHOUT a paid license. An app whose SSO is paywalled (see highlight, §7) cannot claim any rung by paying.
- Zero-touch install: the hard bar.
shc install <app>with its integrations attached must produce a fully functional app INCLUDING working SSO with no special setup: no copy-pasting keys between apps, no click-ops in the app’s admin console for baseline functionality. Every credential is minted by plugs / vault autogen; every baseline config write happens in package jobs. If baseline function needs a human in the app’s UI or a hand-shuttled key, the app is RULED OUT until the package automates it. - Exactly one app per league. The label is a superlative; two flagships in one league is a contradiction the fitness check refuses.
- Free to use, and the bar is what the USER must pay (owner ruling
2026-08-07: “if you are forced to pay to use the app it’s not a
flagship”). An app is DISQUALIFIED when a user must pay to use the app at
all, to reach its core functionality, or to reach its SSO
(requirement 1’s no-paid-tier rung rule is this bar’s SSO-specific case;
the two say the same thing). The test is what the user is forced to
pay, not which license family the code carries, and not what SHC
would owe a vendor as a managed-service provider.
- Source-available / no-SaaS terms do NOT disqualify. SSPL, BSL/BUSL 1.1, ELv2, RSALv2 and FCL restrict operating the software as a service for third parties. That binds SHC’s unshipped managed cloud, never the self-hoster the label is written for, and it is tracked on its own axis (§9). It is not a flagship criterion.
- A redistribution grant is still a packaging precondition. SHC ships
images, soft-forks and custom builds, so the license must permit the
modification and redistribution the package performs; copyleft is
honored via public fork repos
(
docs/guide/apps/custom-builds.md). This is a precondition on the packaging, not the disqualifier above. - The apps this correctly disqualifies today are already segregated in
apps/enterprise/: highlight (SSO gated behind a license key;apps/enterprise/highlight/compose.yaml:57) and metabase (“SAML/JWT/OIDC SSO are paid”;apps/enterprise/metabase/app.yaml:75). The bucket’s own definition, apps where “SSO/core needs a paid license with no free workaround” (commitbcc4ca425), is this same test stated app-side, so the two now read compatibly and no membership moves. [2026-08-11] highlight is no longer in that bucket or in the tree: it was RETIRED (apps/RETIRED). The paid-SSO reading above is why it could never holderrors; the reason it was removed rather than kept as a disqualified package is requirement 8: LaunchDarkly’s acquisition left no maintained upstream. metabase is now the sole member ofapps/enterprise/, and the bucket’s definition is unchanged.
- Lifecycle honesty. From-empty install converges with zero hand steps,
complete
backup.yaml, staged-safe upgrades, digest-pinned images. - Estate citizenship. Consumes estate services via plugs (postgres/s3/smtp/oidc/llm), never bundled copies of them; protocol-named sockets; no default credentials post-install; vault-autogen secrets; measured memory limits; no runtime internet fetches.
- Honest health. Real healthchecks on every service; “12/12 healthy”
must be a liveness claim, not a running-state claim (the huly G26 lesson,
docs/handoff/2026-08-05-sso-certs-and-the-gap-inventory.md). - Maintained upstream: NO exception process (owner ruling 2026-08-06).
EOL / frozen lines are ineligible, full stop. In particular no
-classicapp can EVER carry the flagship label: a classic line is by definition no longer officially maintained upstream. A league whose only candidate is a classic line stays VACANT (§5); it does not inherit the label by default. The-classicname rule is enforced mechanically by the fitness check (§4e); the broader maintained-upstream judgement stays with the owner.
Requirements 1 and 2 compose into the sharpest test: a flagship’s SSO must
come up from shc install <app> -i oidc=<keycloak> alone. The 2026-08-05
field sessions proved how far the estate is from that: SELECT COUNT(*) FROM integration was 0 on both estates and every live login was hand-wired with
kcadm/occ/ORM writes, so every candidate whose live SSO was hand-driven
carries a gap note in §7 naming exactly what the package must automate before
the label is earned.
2. Label grammar
Section titled “2. Label grammar”# app.yaml, top levelflagship: git-
The value is the league: an unquoted lowercase token matching
^[a-z][a-z0-9-]*$, declared exactly once. Quoted, capitalised, empty or duplicated declarations are refused by the fitness check; the index and the check must always read the same bytes. -
League names name the FUNCTION, never the product (the DNS doctrine:
knowledge.notaffine.; the product gets swapped, the function does not):knowledge,projects,git,mail,cloud,design,sbom,siem,grc,workflows,errors,devices,agents,license,quality,storage,commerce,chat,memory,analytics. The set is open (a new league is a new token, no schema change), and a league may be VACANT (§5). -
The label is owner-assigned. An app PR must not add or move a
flagship:key without an owner ruling; the fitness check enforces the contract’s mechanical half but cannot referee “best in class”. -
Userless marker grammar. An app with no upstream user model claims the userless-but-gated rung by declaring BOTH a top-level
access: openoraccess: gatedAND a rationale comment in app.yaml:# flagship-userless: no user model upstream; every request rides the# forward-auth gate, so the estate SSO IS the auth layeraccess: gatedThe
# flagship-userless:marker is what separates a documented decision from a routing accident;access:alone does not satisfy the check.
3. Where the label surfaces
Section titled “3. Where the label surfaces”The key rides the ordinary app-metadata path, additively at every hop:
| Hop | File | Change |
|---|---|---|
| package | apps/<name>/app.yaml | top-level flagship: <league> |
| build | scripts/build-apps-archives.go | appManifest.Flagship (yaml) → buildMeta.Flagship (json, always emitted, "" when unset) |
| publish | scripts/publish-apps-registry.sh | the manifest jq projection carries flagship: (.flagship // "") |
| index | scripts/gen-apps-registry-index.go | httpIndexApp.Flagship (flagship: beside category: on the app HEAD; omitempty, so the unlabeled majority renders no key) |
Head-only, like category:; a league claim is app-level metadata, not
per-revision. Absent and empty are kept distinct the same way images keeps
them: an old manifest with no key and an unlabeled app both render NO
flagship: line (tests/unit/ci/flagship_index_test.go).
Compatibility is evidence, not intent (same argument as images and
channels:): ce’s repo.ParseHTTPIndex decodes with a plain yaml.Unmarshal
and the whole ce module contains no KnownFields(true), so an unmodelled key
is ignored, not rejected, verified against the pinned ce module and asserted
by flagship_index_test.go, which parses a flagship-carrying index with the
SHIPPED parser.
The ce seam: documented, not touched
Section titled “The ce seam: documented, not touched”shc carries the label end-to-end (app.yaml → meta → manifest → index). For ce to surface it, three ce-side structs would need the field, none of which this change touches (ce never gets speculative edits; a ce change is its own consented step, verified against shc’s tests before pushing):
ce modules/repo/httpindex.goHTTPIndexApp: to readflagship:off the synced index into the catalog row (note ce DELIBERATELY droppedparent_app/vappfrom this struct; flagship would be a considered addition, not a reflex).ce modules/app/detail.goappDetailManifest+AppDetailResponse: forshc app info/ the HUD to display the league.- Any catalog listing surface that wants a flagship badge.
Until then the label is fully functional as a PUBLISHING and ENFORCEMENT contract: the index carries it, operators and tooling can read it, and the fitness gate polices it; ce simply does not render it yet.
4. Enforcement
Section titled “4. Enforcement”scripts/arch-fitness.sh’s flagship-contract check, hard-fail, runs in
make lint. Identify it by NAME: the name is what hard and detector_die
print, what FITNESS FAIL [flagship-contract] says, and the only handle that
survives a reorder. (An earlier revision of this section called it “check 22”;
that number was wrong: check 22 is the socket/plug stem naming policy and the
flagship block’s comment header reads Check 23, and the numbering is a
comment-header convention that has already drifted, with two Check 13
headers and three retired numbers. Do not cite these checks by number.) For
every apps/*/app.yaml carrying a top-level flagship::
- (a) SSO surface: the package ships
oidc.plug.yaml,saml.plug.yamlorldap.plug.yaml, OR declares the documented userless posture (§2’saccess:+# flagship-userless:marker pair). - (b)
backup.yaml: a flagship whose data cannot be restored is not a flagship. - (c)
tests/tests.bats: a flagship without tests is a claim, not a tier. - (d) League exclusivity: no two apps claim one league; the token obeys §2’s grammar and is declared exactly once.
- (e) The
-classicban: an app whose name ends in-classicis refused the label unconditionally, however complete its package: classic lines are unmaintained upstream by definition and requirement 8 has no exception process (owner ruling 2026-08-06).
The check refuses to run (exit 2, detector_die) when its corpus resolves
empty: the fitness-guards-scanned-NOTHING lesson: a check over nothing must
die by name, never pass. FLAGSHIP_APPS_DIR overrides the corpus root for the
unit test only (the GUARD_*_DIR seam from coded-registry-guard.sh).
The second check: flagship-league-roster, added 2026-08-14
Section titled “The second check: flagship-league-roster, added 2026-08-14”★ flagship-contract is a CLAIM VALIDATOR, and that was a hole. It walks
the manifests carrying flagship: and says nothing about the ones that do not,
so a tree with the labels deleted reports clean over zero claims. That is
half-done-sweep F9 stated exactly:
“Every artifact of a shipped feature exists and the check passes, because it
validates claims and there are none”, with F10 noting the pair “hide each
other”.
The flagship-league-roster check (hard-fail, same script, same make lint) is the floor. It pins every owner-RULED league with its ruling record and
demands a set equality with the catalogue, not a count floor, which is the
shape that let the bucket layout look right while being wrong
(docs/app-development/app-buckets.md §1):
- (f1) a ruled league with no claimant is red, the arm that makes a deleted or never-landed declaration impossible to hide;
- (f2) a ruled league claimed by the wrong app is red; the label cannot move without a ruling;
- (f3) a league ruled VACANT with any claimant is red. A vacancy is a
ruling (§5), and
storageis the case (e) does not cover:apps/flagship/miniocarries no-classicsuffix, so the contract check would accept the label on it; - (f4) a label naming a league on no roster row is red; §2 says the label is owner-assigned, and the roster row is where that ruling is recorded.
The roster has its own corpus seam, FLAGSHIP_ROSTER_DIR, deliberately separate
from FLAGSHIP_APPS_DIR: a roster floor over the contract check’s three-package
synthetic trees would fail every one of that check’s own cases. It carries the
same empty-corpus refusal, plus one of its own: an all-VACANT roster is a
detector_die, because a roster of nothing reproduces the very vacuity it
exists to close.
tests/unit/ci/flagship_fitness_test.go drives the REAL script (never a
reimplementation) against synthetic trees: green on the real tree, green on a
valid claim and on a documented userless posture, red on a missing SSO
surface / missing backup.yaml / missing tests.bats / duplicate league /
malformed token / undocumented access:, and exit-2 on an empty corpus. For
the roster it builds a corpus from the real tree’s app.yaml files, the only
bytes that check reads, and mutates one manifest per case: a deleted
declaration, a moved label, a claim on the vacant league, and an
unruled league token, each required to fail naming the check, the league
and the app. The deleted-declaration case also asserts flagship-contract
stays SILENT, so the roster is provably the thing that caught it.
What the check deliberately does NOT measure, the judgement half: zero-touch proven live, license posture, maintained upstream, measured memory limits, honest healthchecks per service. Those are owner-ruled at label time against §1, with §7’s gap notes as the worksheet. A grep cannot referee them, so the check does not pretend to.
5. Exceptions and vacancies
Section titled “5. Exceptions and vacancies”- There is NO maintained-upstream exception (owner ruling 2026-08-06;
this supersedes an earlier draft of this tier that treated minio-classic as
a documented exception case). An EOL/frozen line, and categorically any
-classicapp, cannot hold the label. The honest answer for such a league is a VACANCY plus a commissioned fix: forstorage, an SHC-maintained console fork is separately commissioned (§7). - Userless posture (requirement 1). The
# flagship-userless:grammar (§2), machine-checked, so it cannot silently rot. This is a documented RUNG of the SSO ladder, not an exception to it. - Vacant league. A league with no qualifying app stays VACANT rather than
taking the least-bad candidate,
errorsandstoragetoday (§7): a label is a claim operators act on, and a flagship that fails the no-paid-SSO bar or the maintained-upstream bar is a worse answer than none. [2026-08-14]errorswas filled by glitchtip on 2026-08-11;storageis still vacant, and the vacancy is now MACHINE-HELD: it is aVACANTrow on §4’sflagship-league-roster, so labellingapps/flagship/minio(which carries no-classicsuffix, and so slips past the (e) ban) fails by name. A vacancy that only lived in prose was a vacancy nothing defended. - Revocation. The label moves by owner ruling in an ordinary MR-less commit (landing doctrine): delete the key here, add it there; the fitness check enforces the single-claimant invariant through the transition. [2026-08-14] And the roster row moves with it: deleting a key without changing §4’s roster now fails, which is the difference between a revocation and a disappearance.
6. What ships in this change
Section titled “6. What ships in this change”flagship:plumbed app.yaml → meta.json → manifest.json → index.yaml (§3), withflagship_index_test.gopinning emit/omit/parse behavior.- the arch-fitness
flagship-contractcheck +flagship_fitness_test.go(§4). docs/app-development/app-definition.mddocuments the key.- Zero apps labeled; the check is green on the real tree.
★ That last line is the shape a reader should distrust, and it is why F9
happened. “Zero labeled and the check is green” reads as a finished mechanism
awaiting an owner pass. It was also a guard that could not fail. Both halves
were true; only one was said out loud. The state as of 2026-08-14 is four
labels, errors, replay, uptime, analytics, plus the
flagship-league-roster floor (§4) that makes green mean the ruled labels are
present and correct instead of nothing was checked.
7. CANDIDATE TABLE: a PROPOSAL, not a ruling
Section titled “7. CANDIDATE TABLE: a PROPOSAL, not a ruling”The owner assigns labels. Facts below were measured against the tree at
f2e70a2ec (plug files, backup.yaml, tests.bats, healthchecks, in-package
knobs) and the 2026-08-05 field record
(docs/handoff/2026-08-05-sso-certs-and-the-gap-inventory.md); license and
upstream-posture notes marked verify were NOT re-verified this session.
Nothing in this table is labeled yet, and per §1.2 any candidate whose SSO
or baseline config was hand-wired on the live estates keeps its gap note until
the package automates it.
[2026-08-07] Two kinds of edit have since been made in place. Rows marked
[2026-08-07] carry a verdict re-decided under the rewritten requirement 4
(huly, keygen, metabase, and a no-change confirmation on highlight). Rows
marked [corrected] carry a FACTUAL error, not a judgement change:
ciso-assistant and graphiti. Both errors have the same cause and it is worth
naming, because it is a repeatable failure mode: this table was measured at
f2e70a2ec (15:51:39) but committed at 16:52:05, and two commits landed in
the 61-minute gap that falsified two of its cells. A measured pin is only
honest while the commit follows it closely; re-measure before committing, or
say the pin is stale.
| League | Proposed | Mechanical floor (plug/backup/bats) | SSO tier reachable today | Gap notes against the bars: what the package must automate / prove before the label is earned |
|---|---|---|---|---|
knowledge | affine | oidc+llm+postgres+s3+smtp / yes / yes | visible button (structurally cannot auto-redirect; 2026-08-03 audit) | Live edge exists via integrate.sso on both estates. Owed: a zero-touch from-empty install+login proof through the oidc plug alone; note in-app search is off by design (AFFINE_INDEXER_ENABLED: false). License posture verify. |
projects | huly | oidc+postgres+s3+smtp / yes / yes | native OIDC flow: the ONLY candidate with a real user’s full login proven live (2026-08-05 10:59:15) | G26 closed at this pin: 13/13 services now declare healthchecks. Still open: G28 (stats hairpins out the public edge), G30 (four advertised endpoints with no route). Must stay UNGATED (forward-auth breaks native-OIDC apps; owner posture). [2026-08-07] Requirement 4 is MET, and this is the case that forced the ruling. Huly is EPL-2.0 (verified in-repo: docs/guide/apps/custom-builds.md) and nobody pays to use it. It bundles two no-SaaS components (apps/huly/vars.yaml:147-148: elasticsearch:7.14.2, ELv2/SSPL; redpanda:v24.3.6, BSL 1.1) which the OLD requirement 4 would have read as disqualifying; they are now a §9 managed-cloud item, not a flagship one. Both are also tag-pinned, not digest-pinned, so requirement 5 still has a gap here. |
git | gitea | oidc+postgres+smtp / yes / yes | native auto-redirect knob in the package (keycloak.autoRedirect, honest lockout warning authored in-file) | Strongest mechanical candidate in the catalogue. Owed: zero-touch live proof. Upstream-governance note (Gitea Ltd; the forgejo fork is not packaged); owner may prefer to revisit the pick, not the league. gitlab is the runner-up: heavier by an order of magnitude (measured-limits bar) and carries a restore-failure history (3 pre-existing app-internal bugs, giant-app memory). |
mail | mailcow | oidc+smtp / yes / yes | SOGo SSO shipped in 2026.07a+13 | The default-admin P0 is now AUTOMATED at this pin (mailcow-init.job.yaml rotates “moohoo”, deactivates other superadmins, verifies the invariant); needs a live re-proof. Still owed: the SMTP-relay auth blocker (dovecot delegates to mailcowauth.php, never packaged) and the documented list of mail steps SHC does not do (DNS/DKIM/rDNS are estate steps, but every in-app one must move into jobs). League has no packaged alternative, but mailcow is actively maintained upstream so requirement 8 is met; the zero-touch bar is what is furthest from met here. |
cloud | nextcloud | oidc+llm+mysql+postgres+s3+smtp / yes / yes | plug targets user_oidc | G4 blocks the label: the shipped plug drives user_oidc while the live estate runs sociallogin, and the socket’s reconcile branch would OVERWRITE the working redirect URI on attach, the exact reason SSO attach was REFUSED on nc (owner decision). The package must pick one intake, make attach non-destructive, and prove zero-touch from empty. |
design | penpot | oidc+postgres+s3+smtp / yes / yes | button (no auto-redirect knob found in-package) | No live estate history (clean slate). Owed: zero-touch proof; check penpot’s native OIDC registration flags for an auto-redirect rung. MPL-2.0. |
sbom | dependency-track | oidc+postgres / yes / yes | button (frontend OIDC, free) | Apache-2.0; OIDC is unpaid upstream. Owed: zero-touch proof. |
siem | wazuh | oidc+smtp / yes / yes | dashboard SSO via the oidc plug | GPL-2.0. SSO free upstream. Owed: zero-touch proof (the dashboard/indexer security config is the fiddly part, exactly what §1.2 exists to force into jobs). |
grc | ciso-assistant | oidc+llm+postgres / yes / yes | OIDC login, written through the app’s own settings API | [corrected 2026-08-07] This row previously read “no SSO plug … the package ships no oidc/saml plug at all”. That was FALSE when this doc was committed. apps/ciso-assistant/oidc.plug.yaml exists and lands allauth’s openid_connect provider into the singleton sso settings row through PUT /api/settings/sso/, with read-back verification and a documented force_sso lockout refusal. It landed in 7e04fbe14 at 16:29 on 2026-08-06, after the f2e70a2ec pin this table was measured against (15:51) but before this doc was committed (16:52). The package clears bar (a). AGPL. Owed: a zero-touch live proof; note the plug’s own header records that its context: service job is invisible to a swarm-runtime provider, so SSO stays unconfigured on a swarm-deployed install. |
workflows | windmill | oidc+llm+postgres+smtp / yes / yes | OIDC login via plug | Verify the no-paid bar: windmill’s SAML/SCIM are EE; the plug-driven custom-OIDC login path must be confirmed CE-reachable, else windmill joins highlight below. AGPL core. |
errors | VACANT: highlight ruled out | (highlight: oidc+clickhouse+postgres+s3 / yes / yes) | n/a | highlight FAILS the no-paid-SSO bar: its SSO is license-gated upstream (apps/enterprise/highlight/compose.yaml:57 passes LICENSE_KEY), and it has its own user model, so the userless-but-gated rung is not available to it either (gating an app with identity intake double-credentials it). The league stays vacant until a qualifying candidate is packaged (no sentry/glitchtip in the catalogue today). [2026-08-07] Verdict UNCHANGED and now doubly held: requirements 1 and 4 disqualify it on the same axis. Note this is a payment verdict, not a legality one; docs/reviews/2026-08-06-highlight-license-review.md finds an SSO-enabling soft-fork legal (the enterprise/ carve-out directory holds no code), and it is the maintained-upstream bar (requirement 8) that rules the fork out, not requirement 4. |
replay | openreplay: LABEL ASSIGNED [2026-08-11] | oidc+clickhouse+postgres+s3+smtp / yes / yes | native OIDC via SHC’s additive AGPL soft-fork of chalice, free to everyone | Row added 2026-08-11; the league had no row here because the package post-dates this table. 2026-08-06-session-replay-league.md §0 resolves the §5 tension: upstream’s SAML2 SSO is EE-paywalled, but the fork’s OIDC router is AGPL and unpaid, so the no-paid-SSO bar is met on the merits. Stays access: gated with /ingest/ + /script/ exempt; the machine audience is carved out, so the gate costs a second credential and breaks nothing. ★ Open honesty gap, and it is a REAL requirement-7 shortfall rather than a note: exactly 2 of the package’s 22 compose services declare a healthcheck: (measured 2026-08-11; grep -c 'healthcheck:' against the shipped compose). Three of the 22 are one-shot init/migration containers that correctly have none, so the honest denominator is 19 long-running services with 2 covered. “N/N healthy” on this stack is therefore a RUNNING-state claim, not a liveness one, the exact huly G26 failure requirement 7 exists to forbid. The label is the owner’s ruling and this row does not relitigate it, but the gap is open and owed, not waived. Memory limits are likewise sized, not measured. |
uptime | uptime-kuma: LABEL ASSIGNED [2026-08-11] | no SSO plug (matrix+smtp) / yes / yes | userless-but-gated: access: gated + # flagship-userless: now both authored | Row added 2026-08-11. Awarded on a MEASUREMENT, not a preference: GlitchTip was re-checked at v6.2.5/v6.2.6 for a public status page and has none (issue #36 open since 2022), so the uptime/status league does not collapse into errors. The gap this table would have predicted was closed in the same commit; the label forced ungated_paths for /status/, /api/status-page/, /api/badge/, /assets/ and /upload/, because ce’s disposition default is gated and the “public status page” the league is named for was not reachable by the public. MIT, no paid tier, no user model to pay for. |
devices | fleet | saml+mysql+sentry+smtp / yes / yes | button (no auto-redirect; declining it is a MEASURED refusal, see below) | [2026-08-11] This row previously read “no SSO plug (mysql+smtp) … Fleet Free includes SSO (SAML), verify, but the package ships no saml.plug.yaml”. The *verify* is CONFIRMED (§8: one licence check in the whole SSO path, on enable_jit_provisioning alone) and the gap is closed: apps/integrated/fleet/saml.plug.yaml exists, binding the saml socket because fleet’s web login is SAML 2.0 ONLY (no OIDC path to /login at any 4.x pin). The package clears bar (a). MIT. Owed: a real browser SAML round-trip; the realm descriptor advertises WantAuthnRequestsSigned="true" while fleet structurally cannot sign an AuthnRequest, and that is the same green-provisioning/dead-first-login shape the mautic bug had; recorded as an unproven consumer in plug_socket_agreement_test.go. Also owed: a zero-touch proof, and closing the estate skew (live v4.89.2 vs the v4.90.0 pin, recorded 2026-08-11). No sso_redirect, and that is measured, not skipped: the login form is client-rendered inside an SPA shell (check 1), fleet registers no SAML SLO endpoint (check 5), and, uniquely here, SSO-vs-password is a PER-USER flag, so auto-redirecting the login page would strand every password user including the bootstrap admin, which is this deployment’s only break-glass. |
agents | openhands | llm only / yes / yes | userless-but-gated (candidate) | No upstream SSO; single-user tool. Path to the label: declare access: gated + the # flagship-userless: marker (§2); neither is authored yet. MIT. |
license | keygen | postgres+smtp / yes / yes | none today | [2026-08-07] The license blocker is WITHDRAWN; SSO is now the whole question. This row previously read “requirement 4 disqualifies no-SaaS licenses, so this needs an explicit owner ruling before any SSO work is even worth doing”, a verdict decided entirely on the retired axis. Keygen CE is recorded in-repo as “fair-source (self-hostable CE)” whose restriction “targets reselling Keygen itself as a service” (docs/internal/ee-licensing-design.md:286-292), a no-compete/no-SaaS shape, which §9 now handles and requirement 4 no longer refuses. SSO work is worth doing. Two things stay owed: the package ships no SSO plug at all (bar (a) fails mechanically), and whether self-hosting Keygen CE is free OF CHARGE is unverified in-repo; the exact Fair Core License terms have been marked verify since this doc landed and still are. Note the estate’s own interest: SHC’s paid-tier licensing runs on self-hosted Keygen (docs/internal/ee-licensing-design.md), so SHC is a user of this app, not only a packager of it. |
quality | sonarqube | saml+postgres+smtp / yes / yes | SAML (Community Build supports SAML unpaid) | One of only two saml.plug.yaml consumers; the plug is new and unproven live; the mautic SAML lesson (a green job against a client the protocol rejects) is the exact failure mode to test for. Owed: a real browser SAML round-trip. |
storage | VACANT: minio-classic BANNED by the -classic rule | (minio-classic: oidc / yes / yes) | (console OIDC) | No qualifying console-bearing pick exists (owner ruling 2026-08-06). minio-classic is the only full-featured candidate (upstream removed the admin console from the current line (packaged as apps/minio)), but a -classic app can never carry the label and requirement 8 has no exception process. The league stays vacant until the separately commissioned SHC-maintained console fork exists and is packaged; assess that fork against all eight bars when it lands. |
commerce | medusa | not in the tree at this pin | n/a | A medusa packaging lane is in flight (untracked in the main checkout, not at f2e70a2ec). Assess against all eight bars when it lands; until then the league is unclaimable; the fitness check would fail a label naming a nonexistent package’s league only when the package exists, so simply do not label. |
chat | librechat | oidc+llm+mongodb+s3+smtp / yes / yes | native auto-redirect knob in the package (keycloak.autoRedirect → OPENID_AUTO_REDIRECT, opt-in, constraint documented in-file) | MIT; OIDC free. Owed: zero-touch proof. (matrix serves messaging, a different function; its SSO+socket story is live-proven but it is not this league’s candidate.) |
memory | graphiti | llm only / yes / yes | userless-but-gated (candidate) | [corrected 2026-08-07] This row previously read “declares no access: key and no marker”. The access: half is FALSE: apps/graphiti/app.yaml:40-41 declares access: gated and exposure: lan, with a long in-file argument for both axes, landed in 212d7a966 at 16:29 on 2026-08-06, again after the f2e70a2ec pin, before this doc was committed. What is genuinely still missing is ONLY the # flagship-userless: marker, which the check requires in addition to access:, verified against the check’s own conditional, which ANDs the two greps. Authoring one comment line is the whole remaining gap. Apache-2.0. |
analytics | superset: LABEL ASSIGNED [2026-08-14] | oidc+postgres / yes / yes | button (flask-appbuilder AUTH_OAUTH, OSS, no paid tier) | Row rewritten 2026-08-14. It previously read “no proposal: metabase deliberately not proposed … the league is being re-evaluated for an alternative (separately commissioned); owner directive 2026-08-06 … assess the re-evaluation’s pick against all eight bars when it lands.” The re-evaluation landed: 2026-08-06-bi-league-pick.md, status ACCEPTED, Superset over Lightdash/Redash/Grafana on five bars. The eight-bar assessment that row asked for is transcribed into apps/flagship/superset/app.yaml beside the key, pinned to the 6.1.0 image. Bars 4, 5, 6, 7 and 8 are MET on the merits and 7 is unusually strong for this catalogue: 3 of 3 long-running services declare a healthcheck (the two others are one-shot init containers), so “healthy” is a liveness claim here, not the huly-G26 running-state claim; all four images are digest-pinned. ★ Open gap, and it is a real one: requirement 1 is met only at the VISIBLE-BUTTON rung; FAB’s AuthOAuthView.login renders the provider chooser even with exactly one provider, so the package carries no auto-redirect knob. Under the 2026-08-14 auto-redirect ruling §2 that is a package gap to fix, not an accepted state; an edge-flavour or superset_config.py arming path is owed. Also owed, like every row here: a from-empty zero-touch install+login proof (requirement 2). [2026-08-07] on metabase stands and is unchanged: requirement 4 disqualifies it outright: “SAML/JWT/OIDC SSO are paid” in metabase OSS (apps/enterprise/metabase/app.yaml:75), a permanent bar, not an interim one, and why it sits in apps/enterprise/. Whether metabase is retired from the catalogue remains a separate owner decision (BI pick doc, header). |
Reading the table against §1.2: not one candidate has a zero-touch install+SSO proof today. The nearest are huly (login proven live, but through an estate-managed edge, not a from-empty install) and gitea/librechat (auto-redirect authored in-package, never proven from empty). That is the honest state of the estate, and it is why this change ships the tier with zero labels: the mechanism is ready; the proofs are owed.
8. Verified vs assumed
Section titled “8. Verified vs assumed”Verified this session (measured in the tree / test-run): the oidc/saml
split is landed at the pin (31+2 plug files); per-candidate plug/backup/bats
presence; huly 13/13 healthchecks; gitea + librechat auto-redirect knobs;
mailcow-init’s default-admin rotation; ce’s parser leniency (no
KnownFields in the pinned module); the fitness check’s green/red/refuse
behavior (9 unit cases + a live run); index emit/omit behavior (unit case);
validate-templates 358/358 clean.
Assumed / owed verification (marked in §7): license identifications
(affine, keygen’s FCL terms, windmill’s CE-OIDC reach, fleet-free SSO;
see below); every zero-touch claim (none exist; all owed); sonarqube’s SAML
round-trip; mailcow’s rotation re-proof on a live install.
[2026-08-11] fleet-free SSO: CONFIRMED, and the package gap it exposed is
now closed. Measured twice and independently: (a) by reading the pinned
image’s own Go type metadata and running its config_dump, (b) against
upstream source at tag v4.90.0. There is exactly one licence check in
fleet’s entire SSO configuration path (server/service/appconfig.go,
validateSSOSettings), and it fires only on enable_jit_provisioning:
if !lic.IsPremium() { if p.SSOSettings.EnableJITProvisioning { invalid.Append("enable_jit_provisioning", ErrMissingLicense.Error()) }}Nothing guards enable_sso, entity_id, metadata, metadata_url,
idp_name, idp_image_url or enable_sso_idp_login, and the runtime login
path (InitiateSSO / InitSSOCallback) checks only EnableSSO. Upstream’s
own pricing-features-table.yml at that tag agrees: Single sign-on (SSO) …
tier: Free, with Automatic user creation (JIT, SCIM) … tier: Premium listed
separately. The package supplies no FLEET_LICENSE_KEY, so a shc deployment is
Free, and Free is enough.
Two things follow, both of which the §7 devices row now records:
- Fleet is genuinely a SAML consumer, not an OIDC one. Its web login is
SAML 2.0 only (
server/sso/is entirely crewjam/saml); the OIDC that does appear in the image belongs to Apple Platform SSO and Okta/Entra conditional access, MDM subsystems that never touch/login. It is the third consumer of thesamlsocket. - The gap was packaging, not tier.
apps/integrated/fleet/saml.plug.yamlnow exists. It is not live-proven; see the §7 row andtests/unit/catalog/plug_socket_agreement_test.go’ssamlConsumersAwaitingLoginProoffor the browser round-trip still owed.
[2026-08-07] Resolved out of that list: huly is EPL-2.0, verified in-repo
(docs/guide/apps/custom-builds.md); ciso-assistant’s SSO is not gated; it
ships a working oidc plug (apps/ciso-assistant/oidc.plug.yaml), so the
question the row asked is moot. Still unverified and now stated as such
rather than merely marked: whether Keygen CE self-hosting is free of charge,
and the Redpanda BSL Additional Use Grant on the exact tag huly pins
(v24.3.6); the licensing-postures page’s own rule is never to generalize a
BSL grant across versions.
9. The managed-cloud eligibility axis: NOT a flagship criterion
Section titled “9. The managed-cloud eligibility axis: NOT a flagship criterion”Owner ruling 2026-08-07:
2026-08-07-flagship-payment-axis.md.
The concern the old requirement 4 carried (“SSPL/BSL-style no-SaaS licenses disqualify”) is real, but it is not a question about the app. It is a question about SHC as a service provider, and it belongs on its own axis:
Managed-cloud eligibility. May SHC operate this component on behalf of paying third parties? SSPL, BSL/BUSL 1.1, ELv2, RSALv2 and FCL all bear on that question. None of them bear on whether an operator can install and use the app on their own estate, which is the only thing a flagship label claims.
Three consequences, stated so they are not re-litigated:
- A no-SaaS component never costs an app the flagship label. Huly’s
bundled elasticsearch and redpanda are the worked case (§7,
projects). - The constraint is recorded, not live. SHC’s managed cloud is Phase 2
roadmap:
docs/internal/product/business-plan.md:855(§14.3, Months 6–12), with Managed tiers launching progressively from month 7. Nothing in the catalogue is sold as a managed service today, so no clause in this family currently binds SHC. The axis exists so the answer is already written when that changes. - The substance lives in one place. Per-family detail, the
backing-service-only rule, and the shapes that rule cannot express are in
docs/guide/apps/licensing-postures.md. This section is a pointer and a boundary marker, not a second copy.