Note that Cloudflare requires per_page in 5..50 so nobody optimizes
it back to per_page=1, which returned 400 and, before terminal
classification, produced a 400 storm. Also correct a stale team_id
reference in the Vercel user-fetch comment: the OAuth callback
surfaces the team as teamId.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
The Vercel connector reads the team from the camelCase teamId callback
parameter, unlike the snake_case params most providers use. Nothing
guarded the name, so a regression back to team_id would silently drop
the team on every Vercel connect and leave the source resolving no
users.
Extract the read into vercelCallbackTeamID and pin the exact parameter
name with a test, since the surrounding handler needs a live connector
registry and database to exercise directly.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
Org-defaulting for picker providers only ran in the GraphQL resolver,
so a picker-provider source created or updated through the MCP API
connected fine but resolved no users until the org was picked. Move
the defaulting into the accessreview service as
AutoSelectDefaultOrganization and call it from both surfaces, moving
the providerOrgConfigs picker dispatch alongside it (the three console
picker resolvers now dispatch through service accessors, behavior
unchanged).
Also harden the moved path: resolve the provider from cheap connector
metadata before building the authenticated HTTP client, so the ~50
non-picker providers no longer pay a decrypt/refresh/DB-write on every
create/update; bound the outbound ListOrgs call with a 10s timeout so
a hung provider cannot stall the mutation; and re-check inside the
ConfigureAccessReviewSource tx (OnlyIfUnset) so an org the user picks
while ListOrgs is in flight is not overwritten by the first listed
org.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
Once the source-name worker sets name_synced_at, nothing cleared it,
so a source that hit a terminal failure (Brex 403, wrong Sentry org
slug) kept its generic name forever — even after the user reconnected
with the right scope or picked the correct org, contradicting the
Brex guidance to reconnect.
Clear name_synced_at whenever the connection changes: on UpdateSource
when a connector is (re)set, inside ConfigureAccessReviewSource when
the org is (re)selected, and on the OAuth reconnect path via the new
ResetSourceNameSyncForConnector service method. The worker then
re-claims the row and re-resolves the display name.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
The source-name worker drains claims until ErrNoTask with no
inter-task delay. When connector load or HTTP-client setup failed the
worker logged an error and returned nil without marking the source
synced, leaving name_synced_at NULL. The row stayed claimable, so a
dead connector (notably a revoked OAuth refresh token failing the
eager token refresh) was re-claimed immediately every cycle and hot-
looped the vendor token endpoint.
Treat a setup failure like a terminal resolution failure: keep the
generic name and mark the source synced with a warning so it stops
re-claiming. A later reconnect or reconfigure clears name_synced_at
to retry.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
The source-name worker keeps a generic name and marks the source
synced only when a resolver reports ErrTerminalNameResolution. The
Grafana and Metabase name resolvers, plus Tailscale via its shared
fetchUsers, returned a plain error on any non-2xx, so a revoked
credential (401/403) never reached the terminal path and the worker
re-claimed the row every drain cycle with no delay.
Route their non-2xx through nameStatusError so permanent 4xx wrap
ErrTerminalNameResolution while 5xx stay retryable. The sentinel is
inert on the ListAccounts sync path, which does not inspect it.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
Picker providers (GitHub, Sentry, GitLab, Bitbucket, Heroku, Asana,
Netlify, ClickUp, DocuSign) require the user to pick an org/workspace in
a follow-up step after connecting. When that step is skipped, the source
stays connected but unconfigured, and its first campaign silently
resolves no users because the driver needs an org — the same "connects
fine, campaign fetches nobody" symptom seen on Sentry.
When a connector is linked to a source, auto-select the first workspace
the connection can list, so the source is usable immediately. An org the
user already chose is never overridden, and the picker stays visible
(the frontend shows the selector whenever an org is selected) so they
can switch when several are listed.
Best-effort: a provider that is unreachable or lists nothing leaves the
source in its existing needs-configuration state rather than failing the
create/update mutation.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
The Brex name resolver calls GET /v2/company to build the source
display name, but the connector only requested openid, offline_access,
and users.readonly -- and users.readonly covers /v2/users (the account
fetch), not /v2/company. So the company endpoint 403'd for every Brex
source, which the source-name worker retried forever (2.9M errors in 7
days) until the terminal-error handling stopped the loop.
Add companies.readonly so the endpoint resolves. Verified as the exact
scope string against Brex's OAuth authorize flow (the "Companies: Read
only" developer scope maps to companies.readonly). Existing Brex
connectors must reconnect to re-consent to the added scope; until they
do, /v2/company still 403s but is now handled cleanly (terminal ->
generic name, no loop) rather than silently retried.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
The access-review source-name worker never marked a source synced when
name resolution errored, so it re-claimed the source on every poll and
retried at vendor-latency cadence. Two permanently-failing sources
generated millions of error logs (Brex /v2/company 403 and Cloudflare
/accounts 400) and hammered vendor APIs (8.6M 403s to Brex in 30 days) --
a ban risk, all for best-effort display metadata.
Generalize the Google-403 special case: name resolvers now classify a
non-2xx response through nameStatusError, which wraps
ErrTerminalNameResolution for permanent client errors (400, 401, 403,
404) and returns a plain, retryable error for everything else (5xx,
network). The worker treats a terminal error as done -- it keeps the
generic name and marks the source synced -- while transient failures
keep retrying as before.
Also fix the Cloudflare name resolver's own bug: it requested
per_page=1, but Cloudflare's List Accounts endpoint requires per_page in
5..50 and 400s otherwise (the driver already uses 50). That 400 was the
sole cause of the Cloudflare retry storm; bump it to 50.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
Clerk's Backend API secret key only exposes the customer's application
end-users (GET /v1/users) -- the people who sign into the customer's
product -- not the Clerk workspace/dashboard team who administer the
platform. Those admins have no Backend API and are managed only in the
Clerk Dashboard UI. So an access-review campaign against Clerk reviews
the wrong population entirely (verified in prod: one source paginating
16K+ end-users), which is a compliance defect, not just the pagination
timeouts it also caused.
Unregister Clerk from the builtin provider registry and drop its driver,
provider registration, tests, and cassette so it can no longer be added
or fetched. Keep the ConnectorProviderClerk enum value, its IsValid
membership, the GraphQL enum value, and the frontend logo so existing
stored CLERK connectors still validate, serialize, and display (users
can still see and remove them). Drop it from ConnectorProviders so the
registry stays in sync with the registered set.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
Vercel appends the customer's team to the OAuth completion callback as
teamId (camelCase), but the handler read query.Get("team_id"). The
lookup always missed, so every team install fell through to the
/v2/user personal-account fallback -- which returns 404 for a
team-scoped integration token -- leaving TeamID empty. The Vercel
driver then refused to build ("team_id is required") and every
access-review campaign targeting a Vercel team fetched zero accounts.
Personal-account installs were unaffected because they never send a
teamId and legitimately use the /v2/user path, which is why the
breakage only showed up for team installs.
Read the parameter under its real name so the team is captured and
persisted in the connector settings.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
Commit f5703d390 replaced the fmt.Sprintf URL construction with
url.JoinPath, which calls path.Join and therefore strips a trailing
slash unless the final element carries one. Sentry's API only routes
slashed paths and answers 404 without redirecting, so every
ListAccounts call failed on its first request and no access-review
campaign targeting Sentry could fetch a single account.
The failure was invisible for two reasons. queryMembers maps 404 to
errSentryOrgNotAccessible, so a routing bug surfaced to users as
"reconnect the connector with the correct organization" -- advice that
could never help, because the slug was never wrong. And commit
74ce2bc5d edited the recorded request URL in testdata/sentry.yaml to
match the new construction instead of re-recording the cassette, which
kept CI green; that cassette still carries Sentry's own Link header
with the trailing slash, contradicting its own request line.
Pass the slash on the final JoinPath element in both the members
endpoint and the organization name resolver, revert the cassette to the
URL Sentry actually served, and add a regression test that drives the
driver against a server which 404s unslashed paths, so the URL shape is
pinned independently of the cassette matcher.
Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>
Operators could see that a rate-limit cooldown was active, but not
when it ends, and failure logs omitted most of the CA problem
document. Add a until-timestamp gauge and log the full acme.Error
surface so cooldowns and ACME responses are diagnosable.
Signed-off-by: Ludovic Vielle <ludovic@probo.com>
Drop removed Connect Organization profile fields that break Get Many,
and return those fields from Compliance Portal Get.
Signed-off-by: Sacha Al Himdani <sacha@probo.com>
Compliance portal already embeds and serves the public SPA, so the
old @probo/trust app was unused. Drop its Makefile/CI stubs, Relay
project, CORS port, and the helper that only it imported.
Signed-off-by: Émile Ré <emile@probo.com>
Several race and validity gaps could leave certificate provisioning
stuck, unusable, or noisy:
- Accept the HTTP-01 challenge only after the key authorization is
committed, so the CA cannot hit the token before this instance can
serve it and invalidate the order.
- Persist challenge metadata under a blocking write-back lock; a row
merely locked by a competing transaction no longer silently drops the
accepted order.
- Abandon a recovered VALID order and restart instead of issuing it
with a freshly generated key that cannot match the existing cert.
- Exclude rate-limited rows from the ten-minute stale reset so the
resumable order survives the ACME cooldown.
- Size the provisioning poll lease to exceed the max processing window
so a released claim lock cannot let another worker process the same
row concurrently.
- Parse Retry-After as unsigned seconds and clamp overflow so malformed
values fall back to the default cooldown instead of disabling it.
- Normalize the acme_errors problem_type label to the RFC 8555 set to
bound Prometheus cardinality.
Signed-off-by: Bryan Frimin <bryan@probo.com>
Address several provisioning defects that either stalled the retry
budget or crashed the process:
- Classify CAA resolver/transport failures apart from a real CAA policy
denial. Both shared the "caa records" wording, so a transient resolver
error was persisted as customer misconfiguration and retried forever
without consuming the retry budget. A new ErrCAANotPermitted sentinel
now marks the genuine misconfiguration; other CAA errors are treated
as ordinary transient failures.
- Honor an explicit Retry-After: 0 (or a past date) as permission for an
immediate retry instead of promoting it to the one-hour default
cooldown. acme.RateLimit collapses zero, invalid, and absent headers
to a zero duration, so the header is now parsed directly to tell an
explicit zero apart from a missing one.
- Reuse already-registered Prometheus collectors when a second
ACMEService shares a registerer. The fixed-name collectors were
MustRegistered, so a duplicate registration panicked the process.
- Persist provisioning failures on a context detached from the process
tick deadline. A timed-out attempt reached persistFailure with an
expired context, so the write-back failed and the retry budget never
advanced, leaving the certificate indefinitely retriable.
- Use pgx.StrictNamedArgs in the certificate FOR UPDATE loaders to match
the coredata SQL contract.
Signed-off-by: Bryan Frimin <bryan@probo.com>
Rate limits no longer inflate ssl_retry_count into an instant FAILED
path. Centralize outcomes in decideProvisioningOutcome, keep ACME
order state on transient and rate-limit errors, bound each Process
tick with a timeout, and block on FOR UPDATE when persisting a
freshly issued certificate.
Signed-off-by: Ludovic Vielle <ludovic@probo.com>
Signed-off-by: Bryan Frimin <bryan@probo.com>
A hardcoded 'de breaks vowel harmony for many
entity names. Use "için" instead so the hero
title stays grammatical for any {{name}}.
Signed-off-by: Émile Ré <emile@probo.com>
Non-English heroTitle strings still used "trust"
equivalents. Switch them to each locale's IT
compliance term so they match English and the
portal tagline.
Signed-off-by: Émile Ré <emile@probo.com>
Store a short entity name instead of the full home
heading so orgs can brand portals for sub-entities.
Restore hero i18n composition and keep the English
document title composed from the entity name.
Signed-off-by: Émile Ré <emile@probo.com>
Select menus portal at z-3 while drawers sit at z-5, so
the popup opened behind the drawer. Portal into the drawer
panel instead. Point .env.example VITE_API_URL at the slug
subdomain with a local Vite note.
Signed-off-by: Émile Ré <emile@probo.com>
SNI middleware already stores scheme://host; SEOFromRequest
no longer takes a base URL or re-strips the path. Expose
context helpers and drop the obsolete strip tests.
Signed-off-by: Émile Ré <emile@probo.com>
Restore VITE_API_URL to :8080 in .env.example with notes
for local Vite (:5174 + proxy target). Log compliance
portal lookup errors in the SNI middleware with context.
Signed-off-by: Émile Ré <emile@probo.com>
portalOrigin fell back to the raw input when scheme or host
was missing, so paths like /fr/documents still produced
canonical and hreflang tags. Return an empty origin instead.
Signed-off-by: Émile Ré <emile@probo.com>
Custom-domain base URLs included the request path, so
canonical/hreflang doubled routes; emit SEO links only with
an absolute origin. Rewrite Korean mismatch strings to avoid
literal particle parentheses, rename HtmlLang to HTMLLang, and
ship public favicons so Vite no longer 404s the fallback icon.
Signed-off-by: Émile Ré <emile@probo.com>
Swallow locale mutation rejections after the toast, close the
mobile drawer on locale change, escape SEO paths, share the
IAM locale list with SEO, and finish dropping /trust leftovers.
Signed-off-by: Émile Ré <emile@probo.com>
Portals are host-routed on slug subdomains, so SEO path
stripping and the frontend basename helper were dead code.
Signed-off-by: Émile Ré <emile@probo.com>
Hardcoding the trust-center origin broke local portal
development. Read COMPLIANCE_PORTAL_PROXY_TARGET so each
slug can point Vite's /graphql proxy at the right HTTPS host.
Signed-off-by: Émile Ré <emile@probo.com>
Schedule persist and navigate together, and defer the
mismatch banner so a one-frame URL/identity desync does
not flash during language changes from the menu.
Signed-off-by: Émile Ré <emile@probo.com>
Dismiss the user menu on language pick, and update the
URL together with an optimistic identity write so the
mismatch callout does not flash mid-switch.
Signed-off-by: Émile Ré <emile@probo.com>
Render the switch-to-saved button in the identity locale,
translate remaining locale stubs, and use a checkmark for
selected dropdown radio items.
Signed-off-by: Émile Ré <emile@probo.com>
