The certificate's Intune managedDevice id is looked up in the tenant's Intune. With the tenant
switch ValidateIntuneDeviceBinding on, that lookup is the last accepting link of the validator
chain: an enrolled device that is not being retired, wiped or deleted is admitted as
ValidatorType.IntuneEnrollment. A 404 for a certificate younger than 60 minutes is
NotFoundRecentEnrollment (503, never cached); a missing grant is 503 with 120 s.
For devices another validator admitted, the lookup only observes: in the background, never
admitting or rejecting, and only when the app token carries
DeviceManagementManagedDevices.Read.All, so a tenant without the add-on costs no Graph call.
Outcomes ride on the request row (CertDeviceBinding, role, serial match, owner and enrollment
type).
TenantConfiguration.HasAnyDeviceValidation() is the single gate precondition; the feature flags
expose it as deviceValidationEnabled (replaces validateAutopilotDevice).
The agent parses ValidatorType strictly: release the agent before this backend.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The registrations (name, callback URL, client id) live in their own table,
not in TenantConfiguration, so the report only showed the limit. Fetch
GET tenants/{tenantId}/mcp-client-registrations fail-soft alongside the
other side views and render them in the Tenant Status section, plus the
platform switch. The limit row is relabelled to say it is a limit.
Claude-Session: https://claude.ai/code/session_01DbRs7NCppizQ7wWFjN7n3M
Co-authored-by: Claude <noreply@anthropic.com>
The query recognizes only npm limiter packages, never oauthRateLimit /
oauthTokenRateLimit, and flags a handler as soon as it calls a function
named like *auth*, *login* or *verify*. All six alerts it ever raised were
on throttled OAuth routes (the last two came from tenantAuthority()).
What it stood in for is now a test: every route of the OAuth router must
run a per-IP limiter first unless it is one of the three named discovery
documents, /oauth/token must carry the strict budget, and a router.use()
middleware fails the check. Verified by mutation: removing the limiter
from /oauth/authorize fails the test with the route name.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The registration lookup put a client_id-derived value into the backend URL
path unencoded. The amc_ regex already limits it to 32 hex characters, so
nothing changes at runtime, but every other request-derived path segment in
the MCP server goes through encodeURIComponent, and CodeQL counts only that
(or a ?/# prefix) as a request-forgery barrier - a regex capture group is
not one. Closes code-scanning alert js/request-forgery on oauth.ts.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A ?serial= parameter (serial number or exact device name) runs one lookup
as soon as sign-in has resolved the tenant, so organisations can send users
a link to follow their own device. Every hit writes the found session's
serial back into the URL for reload, bookmark and copying. Authorization is
unchanged: sign-in stays mandatory and roleless callers still need the exact
serial server-side. The query survives the login redirect because
ProtectedRoute keeps pathname + search. progress_serial_submitted carries
source form | link.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Settings > Tenant > AI Integration (between SLA Targets and Bootstrap
Sessions) explains the normal connection first - server URL for Claude,
ChatGPT, VS Code and other hosted assistants, nothing to register - and
shows the self-hosted client registration below it (Tenant Admin, platform
switch on). The page is visible to admins, platform scopes and anyone with
MCP access. The self-hosted example text is neutral.
- A tenant may hold one self-hosted client registration. The GA-only
TenantConfiguration.McpClientRegistrationLimit (1-10) raises it on request:
reverted for non-GA callers on PUT, refused on PATCH, edited in the Global
Admin tenant dialog. MCP_SERVER_URL joins the portal URL registry.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A self-hosted AI client redirects to its own domain, which the vendor
redirect allowlist never admits, so it could not use the browser sign-in.
Tenant Admins can now register the client's exact OAuth callback in the
portal and configure the client with client id amc_<registrationId>.
- MCP OAuth proxy: an amc_ id is resolved through the backend; its
registration is the redirect allowlist entry for that exact callback.
Authorize and token run at the registering tenant's Entra authority, and
the token endpoint discards a token of any other tenant (code and
refresh). Lookups are cached for 60 s, a failed lookup fails closed. Every
other client keeps the allowlist check before its resolution.
- Backend: table McpClientRegistrations, TenantAdminOrGA routes
tenants/{tenantId}/mcp-client-registrations (list, create, delete; at
most three, https or loopback http, exact URI, audited) and the anonymous,
per-IP rate-limited lookup auth/mcp/client-registrations/{id} (under
/api/auth for the client-certificate exclusion list).
- AdminConfiguration.McpClientRegistrationEnabled (default off) is the
platform switch: off hides the portal section and makes every lookup 404.
auth/me carries it for the sidebar.
- Portal: Settings > Tenant > Self-hosted AI clients (Tenant Admin / Global
Admin), and the switch in Admin > Settings > Global.
- Tests: MCP tenant-client flow (authority, tid check, revocation, cache,
allowlist), backend round trip, wire parity, validation, cap, switch,
audit; nav visibility. Manifests and wire types regenerated.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The proxy issues no client secrets, but its authorization server metadata
named no token_endpoint_auth_methods_supported (RFC 8414 then implies
client_secret_basic) and /oauth/register echoed a requested secret-based
method. A client following that default registered with client_secret_basic
and then failed the token exchange ("client_secret_basic authentication
requires a client_secret"), reproduced with the SDK calls a self-hosted chat
client makes. The metadata now lists ["none"] and registration always answers
none (RFC 7591 lets the server replace a requested value).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A user token obtained by an application other than the platform's own
registrations (for example a self-hosted AI client using on-behalf-of) was
accepted with the user's full role; only audience, issuer, signature and
lifetime were checked. Such tokens are now classified on every request.
- ClientAppBindingMiddleware (after authentication, before policy enforcement)
records the request dimensions ClientAppBinding (registered | unregistered |
unidentified) and ClientAppId, plus a throttled Warning.
- AdminConfiguration.EnforceClientAppBinding (default off; Admin > Settings >
Global): when on, a foreign client is refused with 403 ClientAppNotRegistered
unless the application is an enabled member of the caller's tenant (the
app:<client-id> row a service principal uses). An admitted caller keeps their
identity but gets the application caps: at most Viewer, no platform role,
DelegatedReader on managed tenants, roleless tiers closed except auth/mcp.
- AdminIdentity.IsCapped (application or admitted foreign client) is what the
global, delegated and member-role resolvers switch on; RequestContext carries it.
- Own clients = the audience trust set plus EntraId:TrustedClientAppIds, a
client-only trust for the local development registration that never widens
the accepted audiences.
- Tests: classification, unforgeable cap marker, middleware in both modes, caps
through the real policy middleware, pipeline order; manifests and wire types
regenerated.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A backend 401 on `initialize` logged only the UPN, and the backend's own trace
("IDX10214 audience validation failed") never names the token; a registration
rejection logged only a count of redirect_uris. Diagnosing a self-hosted client
(LibreChat) that first failed DCR on its own callback host and then presented
tokens minted for another resource took a customer round trip.
- access-guard: `invalid-token-claims` and `backend-rejected-token` carry
`aud=... azp=... tid=...` (unverified claims; control characters stripped, capped)
- oauth/register: the rejection names origin + path of up to three rejected
redirect_uris (never the query); the error_description names the actual
defaults (VS Code, not Google)
- tests for both lines; skill catalog in .claude/commands (mcp-telemetry) updated
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Route payloads (__next.*.txt, index.txt), static JSON and the router's HEAD
self-requests were 20k of 62k dependency rows per week and nobody reads them.
lib/webTelemetryExclusions.ts builds the patterns (own origins minus /api/,
plus relative non-API URLs) for two layers: excludeRequestFromAutoTracking-
Patterns for string fetches, and a dependency initializer that drops the
finished item by its absolute target - the SDK's own check sees an empty URL
for the URL objects Next's router passes to fetch (verified against the ingest
batch: 11 such fetches, 0 shipped; API, SignalR and token rows still shipped).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
P1 - full loads
- staticwebapp.config.json: /_next/static/* immutable, /landing/* max-age=86400
(SWA default was max-age=30 for hashed assets); guard test pins it
- components/NavLink.tsx: prefetch on hover/focus/touch for the portal chrome
(prefetch={false} also disabled the hover prefetch); landing links stay off
- app/layout.tsx: preconnect to the API origin
- lib/dashboardSeed.ts + lib/prefetchSeeds.ts: the dashboard's first list and
stats requests start with auth/me and are handed to the hooks as one-shot
seeds (the own-tenant list URL carries no tenantId; the backend reads the JWT)
- lib/msalConfig.ts: the login request carries the API scope, so one token
redemption serves the first API call; Graph consent via extraScopesToConsent
P2 - chatter per navigation
- lib/signalrGroupRegistry.ts + contexts/SignalRContext.tsx: group membership
is reference-counted with a 2 s grace leave; reconnect rejoins run in
parallel; on/off/invoke and the context value are memoised; the dashboard's
unconditional leave no longer drops the notification bell's global-admins
membership; consumers pair every join with a cleanup leave
- app/sessions/hooks: both group joins in parallel, the join catch-up is
dropped while the initial walk is in flight, identical session deltas skip
the state update
- lib/cachedAuthFetch.ts: TTL cache for feature flags, tenant list and latest
versions with invalidation on every write path; FeedbackBubble defers its
status call
P3 - pure derivations only (D-274): memoised filter arrays, one-pass phase
summary, status counts, useDeferredValue on both search inputs
P4 - landing
- lib/authHint.ts + AuthGate + globals.css: the loading overlay is applied by
an inline head script only when the site root is opened by a browser with an
auth hint; the prerendered HTML carries no overlay; CSS lifts it after 8 s
- hero as WebP srcset (81 KB instead of a 180 KB PNG), dead landing-v1 CSS
removed, terminal demo pauses off-screen, duplicate twitter image dropped
Telemetry (D-276)
- lib/appInsightsRuntime.ts: the SDK is its own chunk, deferred until idle on
the public host, remote config fetch off (blkCdnCfg), initial trackPageView
on both hosts, Web Vitals as web_vital_* custom metrics
Verified: tsc, eslint --max-warnings 0, vitest 127 files / 1442 tests, next
build, Playwright click-through of every portal route, settings section,
session views and public pages on the production bundle against the live API.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A platform script that ran 10 s was shown as "1h 00m" (session 4377911b): its
start line was self-anchored (+120 min) while its result line, re-read after an
overwrite rewind and therefore never fresh, fell back to the agent's cached zone
(+60 min), was rejected as ahead of the clock, and the duration was taken from
the raw pair. The same shape fired script_timeout_suspected daily in one tenant
on 2-9 s scripts read as 17 h. Fleet 2026-09-20..23: about 1 % of script
durations, 28 of 28 inflated ones carrying derivedTimestamp.
Agent: ScriptExecutionState binds the provenance of its start, result and exit
lines (CmTraceLineProvenance, persisted); EmitScriptCompleted judges the pair -
verified (both measured, or both assumed with the same offset from the same
file), unverified, or suppressed with durationSuppressedReason
(rejected-endpoint | offset-grid, CmTraceOffsetCalibrator.IsOnOffsetGrid) - and
always emits durationProvenance plus startSourceOffset*. A held completion's
sourceOffset* tags now describe its own result line; script_timeout_suspected
only judges a measurable duration; the emitted-run marker is capped at the
clock so a future-shifted result no longer swallows the next run's start.
Web: the reducer applies the same rule to stored events from older agents,
pairing a final with the latest script_started of its policy, and the panel
shows "n/a" with the explanation where the run time would be.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ImeRegistryAppStateObserver reads a fourth IME registry surface,
Win32Apps\OperationalState\<userGuid>\<appGuid>: the install deadline a
downloaded app waits for (ExecutionDeadlineTime, IME 1.106) and the
app-in-use deferral counters (DeferralsUsed, DeferUntilTime,
DeferralMaxDeferrals, DeferralAutoDeferred, IME 1.105). Both travel on the
existing registry_app_state event as data (installDeadlineUtc, deferralsUsed,
deferralMaxDeferrals, deferUntilUtc, deferralAutoDeferred) with the change
triggers installDeadline / deferralsUsed / deferUntil, and the message says
the state in words ("downloaded, install deadline ... UTC", "deferred by user
1/3 until ... UTC", "auto-deferred ..."). No new event type.
ClassifyStatusServiceStatus maps IME's legacy AppInstallStatus explicitly:
1000/1002/1004 installed, 1001/1003 installing, 2000 band notApplicable,
3000 failed, anything else unjudged. The 1000 band was treated as installed,
but IME maps every InProgress enforcement state to 1001, so a running app
counted as terminal success on the registry rail (observability only).
Tests: explicit status mapping, both registry timestamp shapes, the new
change triggers, the wait-state text, and the tick flow for a deadline and a
deferral event. Mutation proof: classifier and trigger broken on purpose
turn exactly the five expected tests red.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ajv-cli 5.0.0 is the latest release and still declares fast-json-patch ^2.0.0,
so neither Dependabot nor npm audit fix can move past CVE-2021-4279. An npm
override forces 3.1.1 into the lockfile; the 3.x API is backward compatible and
the validate command used in combine-rules.yml never touches the package.
Verified: npm ci from the lockfile plus all four ajv validate steps pass locally.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The offboard endpoint tombstones the tenant (Disabled=true, "Offboarding in
progress") before it answers 202, so the departing admin submits the farewell
feedback from an already-suspended tenant. Since the suspension gate moved onto
the shared authorization path (43ddc601, 2026-08-29) every feedback POST was
403 TenantSuspended — invisible, because nobody clicked and the two stored
feedbacks predate the gate (D-270).
- Endpoint catalog: `AllowedDuringOffboarding` on exactly
POST tenants/{tenantId}/offboard/feedback; the middleware ignores the
offboarding tombstone for that route only. An operator suspension still
denies; policy tier, cross-tenant check and the function's active-marker
check are untouched. Middleware tests for all four cases plus a catalog
ratchet pinning the flag to the one route.
- Farewell mail: IOffboardFarewellEmailSender.SendAsync returns whether the
provider accepted; the offboarding worker records FarewellEmailSent /
FarewellEmailSkipped / FarewellEmailFailed like the welcome mail — the
send's own success log is Information and never reaches Application
Insights. Missing Email:ApiKey logs Warning. Handler tests per outcome;
OpsAlertRules catalog and the generated manifests (Web + MCP vocabulary)
carry the three types.
- Web: auth/me 403 TenantSuspended no longer shows alert() and forces a
logout. AuthContext keeps the user signed in with claims-only info and
ProtectedRoute renders a suspended page: for the offboarding tombstone
"Your organization is being offboarded" with the farewell feedback form
(a reload during the 6-minute preparation window keeps the form), for any
other suspension the reason text plus contact links; both with sign-out.
The feedback widget moved from OffboardingSection into
components/OffboardingFeedbackForm.tsx.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Every tool description and both role instructions now fit the 2048-char host cap
(Claude Code truncated the Global Admin instructions and two descriptions silently).
Pagination, projection and scope mechanics are said once in the instructions and
the shared pageSize/continuation/tenantId/days texts; role markers left tools only
that role can see; get_resource lists its sections on the argument. The SDK $schema
URL is stripped from tools/list.
Measured (compact tools/list): Global Admin 111.4 KB -> 89.1 KB, Reader 94.7 -> 74.3,
tenant user 64.5 -> 51.3, delegated 75.1 -> 55.6 KB. catalog-budget.test.ts pins the
new ceilings, the variant counts and the empty over-cap list.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
list_tenants named a real customer domain as its example; the filterTenantList test
carried three real tenants. Both now use synthetic names and GUIDs.
catalog-budget.test.ts pins the per-role wire catalog as a ratchet (size, repeated
argument texts, no foreign domains or GUIDs, the 2048-char host truncation cap with
the four known over-cap texts listed). scripts/audit-catalog.ts dumps and sizes the
catalog per role for before/after measurements.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
zizmor (static analysis for GitHub Actions) now runs twice: locally before a
workflow commit (hard rule in .claude/CLAUDE.md) and as the `zizmor` job next to
the pin gate in action-pins.yml (zizmorcore/zizmor-action pinned by SHA, zizmor
1.30.1 pinned, SARIF into Code scanning, severity low and above). Like the pin
gate it reports and does not block. zizmor.yml holds the single reasoned ignore.
Baseline fixes so the gate starts green:
- persist-credentials: false on every checkout except combine-rules.yml, the one
job that pushes with the checkout credentials (ignored in zizmor.yml, with reason)
- publish-scripts.yml: id-token: write moves from the workflow to the publish job;
indexnow-notify.yml: permissions: {}
- deploy-mcp.yml: env values referenced as shell variables inside bash steps
instead of template expansion
- pester.yml / publish-scripts.yml: reusable workflow via the self-repository form
$/ (GitHub changelog 2026-07-30); the pin gate treats $/ like ./
- combine-rules.yml: ajv-cli installed with npm ci from rules/scripts/package-lock.json
(no ts-node peer, no lifecycle scripts) instead of an ad-hoc global install;
Dependabot watches /rules/scripts
- azure/login pin comment corrected to the tag the SHA actually is (v3.0.2; v3
moved on to v3.1.0), no runtime change
- Dependabot: 7-day cooldown on every ecosystem (security updates are exempt)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The XML summary added on TenantManagerItem.Name in ef280b7b rides into
shared-manifests.json as the field's doc; CI run 35701895623 caught the drift.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
GET delegations/managers no longer exposes the name of an operator-created
Tenant Group to the customer; both operator shapes (groups and direct grants)
carry the neutral label. Assignees, reader count and source stay visible.
Web badge reads "Provisioned by operators". D-267.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
network_state_change and network_connectivity_check join the non-activity
list next to power_state_change. A flapping dock or access point kept the
idle-stopped periodic collectors running for the whole episode and reset
the stall clock of the enrollment it was stalling. The events are emitted
unchanged; the device-info network snapshots stay activity.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
IME re-runs remediation policies on its own schedule for as long as the
device is up, so an enrollment that waits for its user reported every run
of every policy. The tracker now emits a policy's first run and every run
whose result differs from the last reported one (compliance, exit code,
stderr, remediation status, error code); repeats are counted and surface
as one script_recurrence_summary per policy when the tracker stops.
The start signal of a known policy is held until its result is known,
because the running indicators pair a start with a later result; a start
without any result is still emitted after a grace. The counters live in
the tracker state file next to the file positions, so a hard kill rolls
both back together. Platform scripts are untouched. The summary does not
count as enrollment activity.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Oliver Kieselbach