A deployment that stands up its own organisation-wide namespace — to use
instead of the built-in global one — finds it invisible to everybody. The
namespace listing only returns namespaces the caller belongs to, and the
only thing that ever added members automatically was hard-wired to the slug
"global".
Make that list a setting. namespace.default-membership holds the slugs every
newly activated account is enrolled in, defaulting to ["global"], which is
what every deployment did before. Publishing only requires membership of any
role, so being enrolled is enough to publish there; no extra grant needed.
GlobalNamespaceMembershipService becomes DefaultNamespaceMembershipService,
since it no longer means one specific namespace.
Where the strictness sits:
- Saving validates every slug resolves to an ACTIVE namespace, so a typo
fails at the moment an administrator makes it.
- Enrolling tolerates a slug that no longer resolves: it logs and skips.
A namespace that was deleted or renamed must not cost somebody their
login.
Adding a namespace to the list after people have signed up leaves them out,
the same trap the personal-namespace work hit, so this ships with the same
preview-then-apply backfill.
Verified against a real PostgreSQL, end to end: create a namespace, reject an
unknown slug, save with whitespace and duplicates and see them normalised,
preview, apply, re-preview showing nothing left, and a fresh registration
landing in the global namespace, the new shared one, and its own personal one
at once.
The backfill preview returned 500 on PostgreSQL:
SQLState 42883: function lower(bytea) does not exist
It reused UserAccountRepository.search(keyword, status, pageable) with a
null keyword. That query compares the keyword with lower(...), and a null
bind leaves PostgreSQL to infer the parameter type as bytea, so lower()
has no matching signature.
Nothing had exercised that branch before: the admin user list goes through
AdminUserSearchRepository, and the member-candidate lookup always passes a
real keyword. The backfill was the first caller to pass null.
Give callers that want every account in a status a query without a keyword
to bind, rather than papering over the null with a cast or an empty string.
Neither test layer would have caught this. The unit tests mock the
repository, and the integration tests run on H2 in PostgreSQL mode, which
accepts the null-typed bind that PostgreSQL rejects. Verified instead
against a real PostgreSQL: preview, apply, and a second preview showing
alreadyProvisioned with nothing left to do, with namespace_member rows
confirming each owner holds OWNER on a TEAM namespace.
Turning provisioning on only affects accounts activated afterwards, which
on a registry that has already been running means nobody. The first person
to hit this on our deployment was the operator who enabled it: they signed
in, got no namespace, and had no way to find out why.
Two fixes.
Backfill. POST /api/v1/admin/settings/personal-namespace/backfill walks the
active accounts and gives a namespace to those without one, skipping system
accounts and anyone who already owns a non-global namespace. Details worth
knowing:
- dryRun reports the plan — each account and the slug it would take —
without writing. The console requires a preview before it will enable the
apply button.
- The response lists only accounts that changed or could not be placed;
the rest are counted, so an operator reads the work rather than the whole
directory.
- A run stops at a per-run account cap and reports truncated rather than
looking like it covered everything.
- Slugs promised earlier in a run are reserved, so one batch cannot hand
the same slug to two accounts.
- Not @Transactional: each namespace is created in its own transaction, so
an account that cannot be placed does not discard the rest of the run.
Diagnosability. The skip paths — provisioning disabled, account already owns
a namespace, no slug available — were silent returns, which is why "nothing
happened and I cannot tell why" was the actual user experience. They now log
their reason; account activation is rare enough that the extra lines cost
nothing.
Self-hosted deployments want every new account to have somewhere of its
own to publish, without asking an administrator for a namespace first and
without pushing drafts into `global`.
Add an operator-controlled policy, off by default so upgrading changes no
behaviour. When enabled, an account that becomes usable gets a namespace
it owns. "Private" here means a team namespace whose only member is that
account: namespaces have no visibility flag, and skill visibility stays a
property of each skill.
Trigger points. UserActivatedEvent is published wherever an account first
becomes usable:
- LocalAuthService.register
- IdentityBindingService.bindOrCreate, for ACTIVE first logins
- AdminUserAppService.updateUserStatus, on a transition into ACTIVE
The third matters for deployments that gate access behind approval: those
accounts are created PENDING at the first OAuth attempt and only become
usable when an administrator approves them.
Why an AFTER_COMMIT listener rather than a call alongside
GlobalNamespaceMembershipService.ensureMember. Both namespace.created_by
and namespace_member.user_id reference user_account(id), which rules out
each obvious alternative:
- Joining the registration transaction lets a slug clash roll the
registration back, so a namespace failure costs the user their account
— or, on OAuth, their login.
- Suspending it with REQUIRES_NEW leaves the new transaction unable to see
the uncommitted user_account row, so the foreign key check blocks on the
outer transaction's row lock and the two wait on each other.
Provisioning after commit avoids both. The listener is deliberately not
@Async, so the namespace exists by the time the user's next request
arrives, and it swallows failures.
Naming. Two templates over ${username}, ${email_prefix} and ${user_id};
unknown placeholders are left in place so a typo is visible rather than
silently dropped. ${username} falls back to the email local part and then
to the user id. Slugs go through the existing slug rules, which is why the
console renders a live preview: underscores are not legal in a slug, so
`${username}_space` yields `alice-space`. A taken or reserved candidate
gets a numeric suffix, so `admin` becomes `admin-2`. Owning any non-global
namespace already skips provisioning, which keeps re-enabling an account
from handing out a second one.
The templates are not exposed in application.yml: they contain ${...},
which Spring would resolve as property references, and Boot 3.2 predates
placeholder escaping. Only the enable flag lives there; templates are set
in the console and default from PersonalNamespaceProvisioningProperties.
Updating the policy writes an audit entry with the before and after.
SkillHub has no mechanism for settings an operator can change without a
redeploy: the only per-deployment knobs live in application.yml, and the
only stored preferences are per-user notification preferences.
Add a generic store. One row holds one setting group serialized as JSON,
so a group can gain fields without a schema migration.
Reads take the caller's defaults:
<T> T get(String settingKey, Class<T> type, T defaults)
which gives two properties worth keeping:
- A group nobody has overridden has no row, and resolves to whatever the
deployment configured. Configuration-file-only deployments keep working
exactly as before, and an upgrade changes no behaviour.
- A stored document that can no longer be parsed also falls back to the
defaults, with a warning. One malformed row must not take down the flows
that read settings, such as login.
Groups are deserialized with unknown fields ignored so a rolling upgrade
can read documents written by a newer node.
No consumer yet; the following commit adds the first one.
Downloading a skill whose slug is non-ASCII (e.g. a Chinese name) through
the ClawHub CLI compatibility route failed: the 302 Location header was
built by string-concatenating the raw slug, and Tomcat encodes header
values as ISO-8859-1, so a character outside 0-255 makes it drop the
Location header entirely and the download breaks. The skillhub CLI path
was unaffected because it doesn't go through this redirect.
Build the Location with UriComponentsBuilder.pathSegment(...).encode(), so
each segment is percent-encoded while the '/' separators stay literal.
"需求" becomes %E9%9C%80%E6%B1%82 and the header is ISO-8859-1-writable.
Fixes#658
Signed-off-by: FenjuFu <92919259+FenjuFu@users.noreply.github.com>
Remove terminal review tasks before deleting an allowed skill version. Lock all versions of the aggregate in stable order so concurrent deletes preserve the last-version invariant and return business errors instead of 500 responses.
Signed-off-by: XiaoSeS <87064762+XiaoSeS@users.noreply.github.com>
* fix(api): tell callers why a request was forbidden
The scope filter already computes an exact reason ("Missing API token
scope: skill:delete", "API token cannot access endpoint: /x") and the
access-denied handler discarded it, returning a bare "Forbidden" for
every case: missing scope, endpoint closed to API tokens, and paths
that simply don't exist. Clients cannot tell those apart, so they
guess — the published CLI reports every 403 as "token may lack
required scope", which sent us debugging token scopes for an hour when
the real causes were a revoked token and a mistyped namespace path.
The reason now rides in the response via a new error.forbidden.detail
message (en + zh), and is logged alongside the exception type.
Signed-off-by: Gal Eyal <gal.e@popai.health>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(api): safely expose API token denial reasons
Signed-off-by: XiaoSeS <87064762+XiaoSeS@users.noreply.github.com>
---------
Signed-off-by: XiaoSeS <87064762+XiaoSeS@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: XiaoSeS <87064762+XiaoSeS@users.noreply.github.com>
The shared RedisTemplate uses GenericJackson2JsonRedisSerializer with
the application ObjectMapper, which embeds no type information, so
stored DeviceCodeData deserializes as a LinkedHashMap. The typed casts
in pollToken and authorizeDeviceCode then throw ClassCastException on
every call, making the whole device authorization flow unusable
(every poll returns 500).
Convert the raw value with ObjectMapper.convertValue instead of
casting; this reads both the current untyped map format and any typed
format, so no stored-data migration is needed. Adds bean setters to
DeviceCodeData for map conversion and regression tests that feed the
service exactly what Redis returns in production (untyped maps).
Fixes#604
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Gal Eyal <gal.e@popai.health>
* fix(publish): delete review tasks of any status when replacing a version
Re-uploading a rejected version under the same version number returned
HTTP 500. deleteReplaceableVersionArtifacts only removed a PENDING review
task, but a rejected version owns a REJECTED one; that row kept a foreign
key on the skill_version, so the subsequent delete hit a constraint
violation that surfaced as a 500.
Delete every review task attached to the version instead.
Signed-off-by: FenjuFu <92919259+FenjuFu@users.noreply.github.com>
* test(publish): drop the spring-test dependency from the new test
skillhub-domain has no spring-test on its test classpath, so
ReflectionTestUtils does not resolve there. Use plain JDK reflection for
setting the generated id and invoking the private method.
Signed-off-by: FenjuFu <92919259+FenjuFu@users.noreply.github.com>
* fix(publish): constrain rejected version replacement
Signed-off-by: XiaoSeS <87064762+XiaoSeS@users.noreply.github.com>
* test(publish): verify replaced review is deleted
Signed-off-by: XiaoSeS <87064762+XiaoSeS@users.noreply.github.com>
* test(e2e): use generated API response types
Signed-off-by: XiaoSeS <87064762+XiaoSeS@users.noreply.github.com>
---------
Signed-off-by: FenjuFu <92919259+FenjuFu@users.noreply.github.com>
Signed-off-by: XiaoSeS <87064762+XiaoSeS@users.noreply.github.com>
Co-authored-by: XiaoSeS <87064762+XiaoSeS@users.noreply.github.com>
The @Async rebuildSkill fix relied on a fresh thread giving a clean
transaction boundary. But skillhubEventExecutor uses CallerRunsPolicy:
under saturation the rejected task runs on the caller (request) thread,
back inside the afterCommit synchronization phase — the original failure
context where the @Transactional index write is silently dropped.
Mark SearchIndexService.index as REQUIRES_NEW so it always suspends any
lingering post-commit synchronization and commits in its own transaction,
independent of whether the async dispatch actually happened.
Add regression tests: detach removes the label keyword, and a synchronous
rebuild inside the afterCommit phase still persists the document (fails
without REQUIRES_NEW).
Signed-off-by: shychee <shychee96@gmail.com>
Attaching or detaching a skill label triggers a search index rebuild via
an afterCommit callback. Because LabelSearchSyncService.rebuildSkill ran
synchronously on the request thread, the @Transactional index write
executed inside the already-committed transaction-synchronization phase
and was silently dropped -- the search document was never written, so
label keywords never became searchable.
Move rebuildSkill onto the skillhubEventExecutor with @Async (matching the
existing rebuildSkills batch path) so the rebuild runs on a fresh thread
and transaction. Add an integration test that fails on the old synchronous
path and passes with the async fix.
Signed-off-by: shychee <shychee96@gmail.com>
