Derived the language set from actual enrollment geography
(get_geographic_metrics, 365d, 10k sessions across ~50 countries) and
documented the country->language mapping on the key arrays. Newly
covered: Finnish, Hungarian, Lithuanian, Latvian, Romanian, Croatian/
Bosnian, Slovak, Indonesian/Malay, Vietnamese, Thai, Arabic, Bulgarian
radio-type variant and Swedish "Radiotyp". New fi/id parser tests; the
unknown-language tests now use Greek (outside the geography set).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Extend the localized label sets with the remaining languages seen in
enrollments: Russian, Japanese, Chinese (Simplified + Traditional,
incl. the 信道/通道 channel variants) and Korean (incl. the two radio-
type spellings). SSID stays untranslated everywhere. New zh-CN and ko
parser tests; suite green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The netsh fallback previously matched English labels only, silently
dropping fields on localized OS images. The parser now matches the
label variants of the languages commonly seen in enrollments
(EN/DE/FR/ES/PT/IT/NL/PL/CS/TR; Scandinavian locales reuse the EN/DE
spellings) and is split into a pure Parse(output) function covered by
unit tests (EN/DE/ES samples, "94 %" spacing, AP-BSSID collision,
unknown-language degradation - SSID itself is untranslated in all
locales). Unknown labels still fall through harmlessly; the native
WLAN API remains the primary path.
V2.Core suite 2095/2095 green; fallback live-verified against netsh
on dev hardware (identical values to the API path).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Field regression on 2.0.1383 (session f7c56065, HP ProBook 450 G7,
build 26200.8037): the native WLAN API path returned nothing in the
OOBE/SYSTEM context and the null path was silent, so wifi_signal_info
disappeared entirely - while netsh-based 2.0.1380/1382 sessions on the
same build kept reporting SSID/signal the same morning. Suspected (not
yet proven) cause: Wi-Fi location gating of WlanQueryInterface on
24H2+; inbox netsh is unaffected.
- WifiInfoReader: new TryGetCurrentConnection overload reporting WHY
the read yielded nothing (native rc per step, interface state);
callers log it at Debug so the agent log explains missing WiFi info.
- NetshWifiFallback (new): parses netsh wlan show interfaces ONLY when
the WLAN API yields nothing on a WiFi-connected device; SystemPaths.
Netsh restored for it.
- WifiConnectionInfo.SignalPercent is now int? (netsh fallback cannot
always parse localized "Signal"); event payload emits the field only
when present - same shape as before.
V2.Core suite 2085/2085 green; interactive positive path re-verified
(SSID/94%/802.11ax/ch116, empty diagnostics).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Narrows the column so more columns fit on one page; matches the
two-line style of Device/Model/Status cells.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
netsh wlan show interfaces output is localized: "Radio type"/"Channel"
(and on some languages even "Signal") never matched on non-English
images, silently dropping fields from wifi_signal_info and network
change snapshots. It also spawned a child process per collection.
Replace both netsh call sites with wlanapi.dll P/Invoke
(WlanQueryInterface: current_connection + channel_number):
- New WlanNativeMethods (raw interop) + WifiInfoReader (fail-soft,
prefers the active NIC's interface GUID, falls back to the first
connected WLAN interface; returns null on machines without WLAN).
- DeviceInfoCollector.CollectWiFiSignalInfo and
NetworkChangeDetector.CollectWiFiInfo now use the reader; RadioType
strings stay netsh-compatible ("802.11ax" ...), event fields
(wifiSsid/wifiSignalPercent/wifiRadioType/wifiChannel) unchanged.
- SystemPaths.Netsh removed (no remaining references).
V2.Core suite 2085/2085 green incl. new WifiInfoReader tests; struct
marshaling verified against native sizeof (604/68/36/532).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Off by default; sortable via avgApiLatencyMs. Data comes from the
SessionsIndex mirror the list already reads — no extra query or table.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Fleet data (2026-07-28, agent 2.0.1380) showed the late-start detector was
outcome-blind: tenant 659c3a90 collected 12 Warnings with the post-mortem/
hung-script note on enrollments that succeeded — the bootstrap merely ran
late in the ESP queue and the agent observed the healthy final minutes live.
New pure helper ObservationCoverage.DescribeLateStart:
- Succeeded -> Info with a tail-only-coverage note (IME-replay hint,
bootstrap-ordering review if recurring); no alarm framing.
- Failed/TimedOut -> Warning with the existing post-mortem/hung-script note.
- oobeStateAtAgentStart refines the message only: "completed" = pure
post-mortem, "in_progress" = live tail observed; not_started/unavailable
add no claim (value flips to not_started after mid-OOBE reboots and
"completed" is routine on post-desktop restarts).
The coverage gate (10 min boot-to-start / 5 min uptime) is unchanged, and
the OOBE state is deliberately NOT a gate: the canonical hung-script
failure (f148976f) terminates on the still-shown ESP screen, i.e. with
OOBE reading in_progress.
8 new tests (outcome x oobe-state matrix); V2.Core suite 2067 green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- SearchSessionsFunction fields whitelist += avgApiLatencyMs,
apiRequestCount so search_sessions lean sweeps (e.g.
fields=geoCountry,avgApiLatencyMs,apiRequestCount) can aggregate
latency without full SessionSummary payloads
- search_sessions tool doc lists the new fields incl. the weighting
hint; get_geographic_metrics doc names avgApiLatencyMs /
apiLatencyVsGlobalPct as the backend-region distance metric
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The outbound_ip trace event stays available for analysis (timeline,
MCP); only the UI row is removed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The extra grid cell in the Session Info card forced a fourth row once
Device Validation was present. The latency reading is a "when needed"
detail, so it moves into the collapsed Device Details card as a new
Connectivity section instead — together with values that were collected
but never displayed anywhere: the estimated internet bandwidth
(network_bandwidth_estimate incl. DO-throttle hint), total telemetry
upload volume (agent_metrics_snapshot), and the outbound IP.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Surfaces the persisted per-session AvgApiLatencyMs in the geographic
views so operators can compare regions when deciding on new backend
locations (Asia/US):
- ComputeGeographicMetrics: per-location and global request-weighted
latency (sum(avg*requests)/sum(requests)) + ApiLatencyVsGlobalPct;
serves both the tenant and the GA cross-tenant endpoint
- GeoMetricsSessionProjection extended (column-projected scan would
otherwise drop the new fields); equivalence test pins it
- Drilldown: LocationSessionRow carries the per-session values; sessions
table + summary tile show them
- geographic-performance page: sortable "API Latency" column with
absolute color buckets (<250 green .. >=800 red), global benchmark
- PlatformMetricsService: session AvgLatency now exact from cumulative
counters when present (fallback: mean of interval averages)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The V2 agent already measures HTTP round-trip latency per request
(NetworkMetrics, cumulative _totalLatencyMs) but only emitted the
per-interval delta average. Operators deciding on new backend regions
(Asia/US) need the session-wide average, aggregatable by GeoCountry.
Measured baseline: Jakarta sessions average ~600-1050 ms client-side vs
94 ms server-side processing — the gap is pure network distance to GWC.
- Agent: agent_metrics_snapshot now carries cumulative net_total_latency_ms
(total/requests = whole-session average; idempotent by construction)
- Ingest: last matching snapshot of a batch projects AvgApiLatencyMs +
ApiRequestCount onto the Session row (fire-and-forget, fail-soft,
last-write-wins; agents predating the field simply never match)
- SessionsIndex: both fields mirrored (merge + BuildSessionIndexEntity
superset) so list/GA views can aggregate latency per geo
- Web: session detail shows "API Latency" when present
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
GET metrics/ime-versions is MemberRead and the archive is one global
partition, so every tenant member reads the same platform-wide rows. The only
thing between an ordinary tenant user and another tenant's identifiers was an
inline projection with no test behind it.
There is no HTTP-level harness in this suite - nothing constructs an
HttpRequestData with a RequestContext - so the projection was unreachable from
a test while it lived inside Run. Extracted it to a static
BuildResponsePayload(versions, hasGlobalScope), matching the pattern
GetTpmPssUnsupportedFunction already uses for the same reason. Behaviour is
unchanged: the payload is returned as `object`, and System.Text.Json resolves
an `object` declared type via the runtime type, so both branches keep the wire
format they had inline. A test asserts that rather than trusting it.
10 tests, asserting on serialized JSON in camelCase (Program.cs's naming
policy) because a property dropped from an anonymous type is only a real
redaction once it is absent from the wire:
- non-global: firstSeenTenantId / firstSeenSessionId absent, and their GUID
VALUES absent anywhere in the payload (a model rename must not let the
identifier survive under another key)
- non-global: exactly {version, firstSeenAt, lastSeenAt, sessionCount}
- non-global: sessionCount survives - the counts are platform-wide by
decision, pinned so it cannot drift either way unnoticed
- global (Admin or Reader): full entry including both identifiers
- both scopes: array shape, empty archive, row order preserved
Mutation-checked both directions: forcing the global branch fails 3 tests,
dropping sessionCount from the projection fails 2. Full suite 3521 passed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The worker serializer omits null fields (WhenWritingNull), so the
aggregated global/metrics/device-journeys response carries no
repeatDevices at all. The strict `=== null` check missed undefined and
fell through to repeatDevices.length, crashing the page. Verified
against the deployed chunk: crash column 11464 is the `0===o.length`
read behind the `null===o` guard.
Loose `== null` check + optional DTO field; absent and null both render
the "select a tenant" hint.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Defender for Cloud reports CVEs against the mcp container image that the
repo scanners cannot see. The subassessment evidence paths split them in two:
/usr/local/lib/node_modules/npm/** - the npm 10.9.8 that node:22-slim
bundles, vendoring @sigstore/core 2.0.0, brace-expansion 2.0.2, tar 7.5.11
(critical), sigstore, picomatch, ip-address, plus npm itself (high, no fix
available). Present in all 20 images including the current one. Dependabot
reads package-lock.json, so none of this is visible to it.
/app/node_modules/** - body-parser 2.2.2, fast-uri 3.1.2, hono, protobufjs,
tar 7.5.16, all confined to images from 2026-07-21 and earlier and already
fixed by 7df208fa/c3e9621a. Those findings age out with the old ACR tags.
npm is build-time only here: CMD is `node dist/index.js`, the HEALTHCHECK is
`node -e`, the container app spec overrides no command and probes over HTTP,
and src/ contains no child_process call. So remove npm rather than upgrade it
- npm 11.x still vendors @sigstore/core 3.1.0, itself inside the vulnerable
<= 3.2.0 range, and would clear nothing.
Separately, tar 7.5.20 -> 7.5.22 in our own tree (GHSA-r292-9mhp-454m, via
@huggingface/transformers -> onnxruntime-node). 7.5.22 satisfies the declared
^7.0.1, so no override is needed. fast-uri deliberately stays on 3.1.4: it is
patched against all four fast-uri advisories, and ajv 8.18.0 declares ^3.0.1,
so forcing 4.x would be an override against a major range.
npm audit --omit=dev: 0 vulnerabilities. tsc clean, 455 tests pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
With ASP.NET Core integration the HttpRequestData body is a
forward-only stream whose Length property throws NotSupportedException
(verified in prod: 500 + System.NotSupportedException on the first
POST). The size cap is now enforced by reading at most cap+1 chars via
ReadBlockAsync instead of asking the stream for its length.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
New anonymous POST ops/alert-webhook receives Common Alert Schema
notifications from an Action Group webhook receiver and re-emits them
as ops events (category Platform, type AzureMonitorAlert), so Azure
Monitor alerts flow through the existing OpsEventService ->
OpsAlertDispatchService fan-out (Telegram/Teams/Slack + portal
ops-events list) like every internally raised operational event. The
delivery path is independent of the portal: a SWA outage cannot
suppress the notification; the action group's e-mail receiver stays as
the backend-independent fallback.
Security: action-group webhooks cannot send custom headers, so the
shared secret travels as ?secret= and is compared in constant time
(FixedTimeEquals) against the OpsAlertWebhookSecret app setting -
fail-closed while unset. 64 KB payload cap, defensive essentials-only
parsing (only alertRule is mandatory), severity mapped Sev0->Critical,
Sev1->Error, Sev2->Warning, Sev3/4->Info with unknown values falling
back to Info so schema drift can never escalate.
Dual-registered per convention: EndpointAccessPolicyCatalog entry
(PublicAnonymous), OPS_EVENT_TYPES in OpsAlertRulesSection (rule
wiring) and Platform category in OpsEventsSection (badge + filter).
Backend suite 3511 green (incl. new helper tests + catalog
completeness), web tsc + vitest 651 green.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Navbar + GlobalSidebar render the full navigation on every page, so the
App Router's default viewport prefetch fired one ?_rsc= request per
visible link against the SWA runtime (~10.4k requests / 14 days). That
origin serves document, chunks and RSC payloads over HTTP/1.1 (6
connections per origin); when the runtime queued the prefetches (p95
896ms, tail up to 234s), they held every socket and the portal froze
until the queue drained. Measured on the 2026-07-27 17:25 incident:
/admin/presence?_rsc= at 151s while the backend API answered in <2.2s.
All 16 nav links now set prefetch={false}. Every target route is a
client component behind ProtectedRoute that fetches its own data after
mount, so the prefetched RSC payload carried no benefit.
New OKF doc docs/web/portal-navigation-prefetch.md records the full
evidence chain and where browser telemetry lives (separate App Insights
resource, not reachable via MCP query_backend_logs).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A GA with a persisted tenant override opening /usage-metrics (or /sla) saw
their OWN tenant's stats until a manual refresh. useGlobalAdminScope seeded
the persisted selection in a post-mount effect, so the page effect first saw
effectiveTenantId = own tenant and fired an own-tenant fetch immediately
followed by the override fetch; with raw authenticatedFetch (no latest-wins
guard, unlike useAuthenticatedFetch since 82934a37) the last-RESOLVED
response won.
- useGlobalAdminScope: seed the GA selection render-time (same pattern as
the delegated seed) so consumers never observe a transient own-tenant
effectiveTenantId - the wrong fetch no longer starts at all.
- usage-metrics/sla pages: latest-wins sequence guard in fetchMetrics so a
superseded request can neither write data/notifications nor clear the
newer request's loading state (covers rapid selector switches).
Verified: tsc 0 errors, vitest 651/651.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three defects found by measuring production after the 1.5.1012 deploy:
- AppMetricsProjection dropped TerminalState, so the PR0 skip
classification never reached metrics/app (totalSkipped stuck at 0).
Column added; equivalence fixture now populates TerminalState/AppId/
EspBlocking/AppIdCollision and asserts them (proven: test fails
without the projection fix).
- BuildTimeAttributionAggregates excluded breakdowns on ANY quality
flag; 903/1000 carry BlockingSetUnknown, starving fleet stats
(Clean=12 vs 331 excluded on the largest tenant). New
TimeAttributionFlagQuality.DurationCriticalFlags mask - only
ClockSkew/PartialObservation/WGAnchors exclude; blocking-set flags
keep their measured spans (positive-evidence semantics unchanged).
- HasMeasuredDuration gets an upper plausibility bound (6h = default
AgentMaxLifetimeMinutes): 490 rows carried back-stamped completions
(up to 117 days; 60175s app rows in a 2697s session) that dominated
slowestApps with six-figure averages. Read-side, heals old + new rows.
Backend 3484/3484 green; solution build clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A tab still running the previous deployment's bundle requests lazily-loaded
chunks (next/dynamic charts, route chunks) whose content-hashed filenames no
longer exist after the SWA swap - the user is stuck on "Loading chart..." or a
dead navigation until a manual refresh. Standard remedy: reload once so the
fresh document references the new bundle.
- utils/chunkReloadRecovery.ts: chunk-error detection (webpack ChunkLoadError +
Chromium/Firefox/Safari dynamic-import variants) and a sessionStorage guard -
at most one auto-reload per tab per 5 minutes; without working storage no
auto-reload at all, so a reload loop is structurally impossible. Emits a
ChunkReloadRecovery App Insights event before reloading.
- components/ChunkReloadRecovery.tsx (root layout): catches rejected dynamic
imports via unhandledrejection and failed /_next/ script/link tags via a
capture-phase error listener (third-party scripts never trigger a reload).
- app/error.tsx: a chunk error reaching the boundary reloads once instead of
showing the error card; inside the guard window it falls through to the card.
Verified: tsc clean, vitest 651/651 (8 new tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
GA override on gather-rules called GET /api/global/config/{tenantId}, a route
that does not exist in the backend; the swallowed 404 left the unrestricted
badge showing the GA's HOME tenant's mode for every foreign tenant. Both paths
now read the member-readable config/{tenantId}/feature-flags (MemberRead has a
global-scope bypass and the route is RouteParam-scoped), with a reset-on-switch
so a failed fetch never carries the previous tenant's mode over. The dead
globalConfig.tenant entry is removed from the URL register.
useAuthenticatedFetch now enforces latest-wins: the scope hook seeds a
persisted foreign tenant only after mount, so an own-tenant fetch races the
override fetch and the last-RESOLVED response used to win, leaving the rules
list on the home tenant while the selector showed the override. Stale requests
no longer write data/error/loading; the spinner is owned by the last-started
non-silent request.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Customers who additionally deploy Microsoft 365 Apps as their own Win32 app saw two
rows in Install Progress (the IME view and the Office C2R lifecycle) with nothing
telling them apart, which read as a duplicate.
- New lib/installProgress.ts: buildInstallItems() extracted from the component so the
event-folding rules are unit-testable; component is now render-only.
- InstallItem gains `source` (ime | office-c2r | realmjoin) and a source-scoped `key`.
Rows were keyed by app name alone, so a Win32 app named exactly "Microsoft 365 Apps"
(= OfficeInstallDetector.AppName) collided with the C2R row and overwrote its state.
- Pill only for non-default sources: "Click-to-Run" / "RealmJoin", neutral grey so it
does not compete with the state badge, with a tooltip explaining what the row measures.
Plain Intune app rows stay unlabelled.
- Drops the "RJ: " name prefix, which broke sorting and truncation.
Verified: tsc clean, vitest 643/643 (16 new).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Codex review (P2): the ESP-config refresh triggers each run on their own
Task.Run (as do CollectAll's startup sweep and CollectAtEnrollmentStart),
so two concurrent collects could interleave read-before-emit - the LATER
emission then carried the OLDER (smaller) registry snapshot. The backend
uses the latest list-bearing esp_config_detected emission as positive
blocking evidence, so an out-of-order emission would silently shrink it.
CollectEspConfiguration now runs its read->emit sequence inside a
dedicated lock covering ALL callers; a later emission always carries an
equal-or-newer snapshot. The race predates the Q2 refresh triggers (they
only raised its frequency).
Deterministic concurrency test: 12 parallel refreshes against a probe
whose list grows per read - the emission sequence must be exactly 1..12.
Verified: Agent V2 Core 2059/2059.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
All six review findings verified against code and real; five fixed here,
the agent-side race lands in the companion agent commit.
[P1] WhiteGlove pause counted as time: ClampToWindows clipped intervals
against the HULL of both WG windows, but the excluded pause lies inside
the hull - a reboot gap or app interval straddling the pause counted the
entire pause (a ~2-day "reboot" in a 30-minute enrollment was possible).
Seconds are now the per-window intersection sum (chronological endpoints
kept); an interval wholly inside the pause makes no claim and no longer
raises ClockSkewDropped. What-if savings are additionally capped at the
session's WallClockSeconds - endpoint deltas can still straddle the
pause, and no removal can save more than the enrollment took.
[P1] Late-arriving events never updated breakdowns: the terminal-seam
compute is one-shot and the sweep only healed missing/version-stale rows.
SessionTimeBreakdown now stores EventCountAtCompute (0 = pre-fix row,
recomputed once); the sweep recomputes on EventCount drift and then also
re-runs ResolveEspBlockingForSessionAsync (newly exposed on
IMetricsRepository) so late esp_config_detected re-emits and app batches
reach both the breakdown and the blocking evidence.
[P2] Stale aggregate buckets: the sweeps now reconcile every partition
they wrote - buckets that existed but were not regenerated (sessions
deleted, class left the window) are deleted via new targeted deletes;
rolling30 rows additionally get a 48h ComputedAt freshness filter at
read time, so a fully quiet tenant stops serving year-old numbers as
"last 30 days".
[P2] Date windows included an extra calendar day: both inclusive range
ends meant days=30 returned 31 keys and days=1 summed yesterday AND
today. InclusiveWindowStart(today, days) = today-(days-1) in both
response builders; the sweep's rolling30 source is filtered to the same
30 calendar days (its query window stays one day wider for backfill).
[P3] Device-history banner counted later sessions as "previous
enrollments" when viewing an older session: selectPriorSessions
(deviceHistoryPrior.ts) anchors on the session's own chain entry (live
sessions: new sessionStartedAt prop), sessionId tie-break mirroring the
server chain sort.
Verified: solution build clean, Backend 3480/3480 (+6: three WG-pause
golden fixtures, window-start facts, RK builder/delete consistency,
EventCountAtCompute roundtrip), Web tsc + vitest 633/633 (+6).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Q1 (AppInstallTiming): a terminal->active transition (IME retry path,
e.g. Error->Downloading or Skipped->Downloading) re-arms CompletedAtUtc
so the final terminal payload covers first-start -> last-terminal instead
of freezing at the first terminal state. Plus a negative-duration guard
in AppInstallTiming.DurationSeconds (inverted endpoints -> null, field
omitted) - observed in production as durationSeconds=-2342 on a
Skipped-then-retried user-scope app (session 66f1c5e1).
Q2 (esp_config_detected): the ESP blocking lists are read too early -
device lists grow progressively, user-scope S-<SID> lists appear only
after sign-in (88 of 89 prod sessions had an empty user list at last
emission). DeviceInfoHost now re-collects the ESP configuration as
phase-driven one-shots: apps sub-phase declaration (AppsDevice/AppsUser
phase_transition = first app activity) and AccountSetup detection.
The StartupEventGate keeps unchanged payloads silent (fingerprint covers
list contents), so only grown lists actually re-emit. F1 EspBlocking
resolution (latest list-bearing emission) benefits without backend change.
Q5 (HardwareInfo): WMI values (serial, manufacturer, model) trimmed at
the source via NormalizeWmiValue; whitespace-only collapses to the
"Unknown" sentinel the retry logic already treats as unresolved.
Attach-window investigation (11a3b67a): NO agent fix - hypothesis
disproved against two real agent logs (diag bundles db5ce007/2.0.1337,
4a9e9273/2.0.1379). "completed without started" first apps were already
installed (WinGet detection-only, zero install lines in AppWorkload.log);
pre-attach replay provably works (pattern-match log). PR0's
HasMeasuredDuration already handles the class server-side.
Duplicate app_install_completed (66f1c5e1): mechanism narrowed to an
in-process re-read of exactly one AppWorkload rotation window (<24h-old
lines, historic suppression not applicable) - documented as follow-up,
no fix without a bundle-bearing repro per lessons.md.
Verified: solution build clean, Agent V2 Core 2058/2058 (14 new tests),
DecisionCore 493/493. No ConfigVersion bump, no new event types,
decision rail untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Analyze rules only (gather stats stay per-batch-deduped until Q6 is
fixed - never silently included). Per (tenant, rule) the radar compares
the trailing 7-day hit rate (FireCount/SessionsEvaluated from the daily
RuleStats rows) against the prior 28-day baseline and fires only when
the window has >=5 hit sessions and >=20 evaluated sessions, the rate
is >=2x baseline AND the Wilson 95% intervals are disjoint
(MetricsMath.WilsonInterval/RateIncreaseSeparated, hand-computed test
vectors). Suppressed for deleted rules, rules inside the CreatedAt
grace period (<35d), rules edited inside the window, empty baselines,
and the RuleRegressionRadarDisabled app-setting kill switch.
Alert episodes live as ruleregression|{ruleId} rows in the notification
tracker (third keyspace): AddEntity-409 dedup = one bell + one
RuleFrequencyRegression ops event (category Tenant, Warning, full
numbers, dual-registered in the web OPS_EVENT_TYPES) per episode;
numbers refresh while active without moving FirstNotifiedAt; episodes
re-arm under 1.5x baseline, on fire-stop, or via the tracker's 30d
retention prune (spec-intended reminder). On-fire dimension correlation
(osBuild/model/agentVersion/imeVersion; hit set from a RuleResults
range scan intersected with the window sessions) reports both shares
with the "correlated, not necessarily causal" wording contract or
honestly "no clear dimension concentration". Radar runs in the daily
maintenance pass after the rule-stats aggregation, anchored on
yesterday (whole days only).
Surfaces: regressions[] in the rule-stats response (MCP get_rule_stats
inherits it), tenant bell deep-linking to the rule card, and the
analyze-rules page gains a densified 30-day fire sparkline plus a
regression badge while an episode is active. OKF doc
backend/rule-regression-radar.md + index/log.
Backend 3474/3474 (27 new: Wilson vectors, every suppression branch,
window boundary dates, re-arm matrix, dimension gates, tracker
round-trips incl. tri-states, keyspace dedup/prune, wording pins),
web tsc + vitest 627/627, MCP tsc + 455 passed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three read routes over the PR4 data: metrics/device-history
(MemberRead + tenant query param - one route serves the web banner AND
MCP, GA cross-tenant via ?tenantId= with a sessionId-based owner
fallback; history:null is a normal outcome), metrics/device-journeys
and global/metrics/device-journeys (days selector honored because FTR
counts are additive - window rate = ratio of summed daily rows; the
tenant-less global variant serves the "global" partition and discloses
repeatDevices:null). Attempt numbers are server-computed
(DeviceJourneyCalculator.ComputeAttemptNumber: chain position within
the journey; live sessions via a virtual non-successful attempt so the
redeploy and 30d-gap rules apply) - no consumer re-derives journey
semantics. Repeat-devices violator list: last journey >=2 attempts with
newest terminal ref in-window, attempts desc then recency, cap 10,
failure reason from the newest failed attempt (bounded point-reads).
Web: DeviceHistoryBanner on session detail ("Attempt N for this device
- View history", expandable terminal chain, durations strictly the
authoritative durationSeconds, WG/DevPrep/admin-marked badges) and the
Fleet Health "First-time-right" section (rate gated at >=20 completed
journeys with disclosed n, junk-serial exclusion note, Monday-UTC
weekly trend, attempt histogram, repeat devices) following the 7/30/90
selector. MCP: get_device_history (serial mode / fleet FTR mode,
role-aware routing, all roles - GA_FULL privilege snapshot extended
deliberately). Docs: OKF backend/device-journeys.md + index/log.
Backend 3447/3447 (14 new: attempt-number vectors incl. gap boundaries
and live sessions, additive sums, days clamping, repeat-device
gate/sort/cap), web tsc + vitest 627/627, MCP tsc + 455 passed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Device-key normalization (trim+casefold, junk-serial exclusion incl. the
agent's "Unknown" WMI sentinel), pure journey grouping (terminal set
Succeeded/Failed/Incomplete; journey ends at first success; >30d gap from
the previous attempt's CompletedAt starts a new journey; WG Pending stays
an open session, never an attempt), and the DeviceHistories table
(PK tenant, RK encoded normalized serial): terminal-only session-ref
chains capped at 20 with derived journey counts, written inline at the
session-terminal seam (all three terminal statuses, DeletionState-guarded)
and healed by a rolling 30d maintenance sweep. The sweep prunes chain refs
of deleted sessions tombstone-driven (SessionTombstones cover cascade AND
retention deletions), deletes emptied rows, and recomputes daily FTR
aggregates (DeviceJourneyAggregates, PK tenant/"global", RK date):
completed journeys, first-time-right count, attempt histogram and
junk-serial disclosure — additive counts, bucketed on the completing
success session's StartedAt date. Chain refs carry the authoritative
DurationSeconds verbatim (never CompletedAt-StartedAt) plus an
adminMarked flag. Both tables registered in Constants.All + offboarding
TenantPartitionTables; aggregates under the 180d retention block;
deliberately NO per-session deletion-manifest step for DeviceHistories.
Backend 3433/3433 (51 new tests: junk/normalization vectors, terminal
set, exact 30d gap boundaries, WG one-row-one-attempt, redeploy split,
deletion pruning, entity round-trips incl. row-key escaping, FTR
aggregation core), Agent 2044+493 green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Insights expansion PR3 (tasks/insights-expansion-spec.md §F1 "Surfaces").
Completes F1: schema (PR1) -> persistence/aggregation (PR2) -> surfaces.
- Rolling range rows: the sweep now also writes "rolling30|{class}"
aggregate rows — range statistics over the window's actual sessions. A
median of per-day medians is not the range median, so the fleet panel
reads these and is deliberately fixed to "last 30 days"; the daily rows
stay in the payload for trends/MCP.
- Routes (all registered in EndpointAccessPolicyCatalog):
GET sessions/{sessionId}/time-attribution (MemberRead + tenant query
param, GA fallback via FindSessionTenantId; breakdown=null is SUCCESS —
a missing lane is a normal state, not an error), GET
metrics/time-attribution (MemberRead) and GET
global/metrics/time-attribution (GlobalReadOrAdmin + tenantId filter),
both returning { windowDays, classes (rolling), daily }.
- Web: TimeAttributionLane under the session-detail phase timeline —
proportional stacked bar of segment totals incl. the explicit
unattributed remainder (bar always sums to the session's authoritative
duration), ESP-blocking apps with critical-path occupancy, reboot badge
(cross-cutting, never a bar slice), quality-flag chips; fetched
fail-soft for terminal sessions only. Fleet Health gains the "Time
attribution" section: per-class stacked medians gated at n>=20
("insufficient data (n=...)" below), flagged/missing exclusions
disclosed, and a top-blocking-apps table with class-scoped rows and
"up to X (p75 Y)" what-if wording. Deliberate deviation: the spec's
per-app trend column is omitted — per-day app rows are too sparse under
the daily n>=5 gate; revisit once rolling history exists.
- MCP: get_time_attribution (one tool, session + fleet mode, role-aware
routing like get_app_install_metrics, registered for all roles) — the
privilege-leak-guard catalog snapshot extended deliberately.
- Docs: OKF concept doc backend/time-attribution.md (invariant, windows,
segment mapping, positive-evidence blocking, Q7 reboot spans,
persistence + surfaces) registered in index.md + log.md.
Backend 3382/3382 (+1 rolling-aggregate test), web tsc clean + vitest
627/627, MCP tsc clean + 455 passed. Post-deploy: web deploy +
MCP redeploy (new tool + docs corpus).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Insights expansion PR2 (tasks/insights-expansion-spec.md §F1 "Data & compute
changes"):
- SessionTimeBreakdowns table (PK=TenantId, RK=SessionId): span/reboot/
blocking-app JSON columns, QualityFlags, AttributionVersion,
EspAppsOccupancySeconds kept tri-state (absent column = unknown, never 0).
Written once at the terminal single-writer seam (TableSessionRepository,
next to the counter reconcile + EspBlocking resolution; Succeeded/Failed
only — Incomplete deliberately stores no duration). Deleted with its
session: new deletion-manifest step 16 (PkRkExact; tombstone moves to 19)
plus tenant-offboarding partition wipe — no separate age sweep.
- Self-healing 30d maintenance sweep instead of a one-off backfill (rides
the existing 2h tick + manual parity): backfills pre-deploy sessions,
heals inline-compute misses, folds late-terminating sessions into their
StartedAt-date aggregates, and enforces AttributionVersion purity — a
stale-version breakdown whose events aged out counts as missing rather
than mixing algorithm semantics (truthfulness rule 8). Skips sessions
with a deletion cascade in flight.
- TimeAttributionAggregates table (PK=TenantId or "global",
RK="{date}|{class}"): per enrollment class (whiteglove /
device_preparation / self_deploying / user_driven — never mixed),
median/p75/p90 per segment over the fixed six-segment stack (incl.
unattributed; absent segment = honest 0), clean/flagged/missing counts
disclosed (rule 7), rows written below the n>=20 UI gate on purpose
(the UI needs the n — rule 4). 180d retention alongside UsageMetrics.
- What-if bound: WhatIfSavingSeconds implements the spec formula
max(0, cpEnd - cpEndWithoutX) including the idle gap before X (the ESP
waits for X to start, too); sole-interval fallback = own start. Per-app
aggregate rows gate at >=5 sessions, top 20 by median interval, carry
median/p75 of interval and saving ("up to" wording is the consumer
contract).
Backend 3381/3381 (13 new tests: entity round-trips incl. tri-state,
aggregation bucketing/disclosure/gates, what-if vectors, class precedence,
manifest step-count pins 17/19), solution build clean, web untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Insights expansion PR1 (tasks/insights-expansion-spec.md §F1, audit §0.5):
- AppInstallSummary + AppId / EspBlocking (bool?, tri-state) / AppIdCollision:
sentinel-gated columns + mapper + both projections. appId adopted from app
events (first-seen wins); a second appId under the same display name flags
the name-keyed row as a collision (audit Q3) — in-batch and cross-batch
(sticky in the reconcile). Collision rows leave every per-app fleet
aggregate (app metrics payload, apps list, app analytics) with a disclosed
exclusion count; the fleet DO rollup and per-app session drill-down keep
them.
- EspBlocking resolution at the session-terminal single-writer seam
(TableSessionRepository.UpdateSessionStatusAsync, next to the counter
reconcile — covers ingest, admin marks, maintenance sweep, rule engine):
joins the session's app rows against the LATEST list-carrying
esp_config_detected emission as a positive-evidence set (audit Q2) —
listed => true, absent => stays unknown, never false. EspBlockingSets
parser handles truncation via the uncapped espTracked*Count fields.
- TimeAttributionCalculator (pure, AttributionVersion 1) + SessionTimeBreakdown
models: exact partition of the session's authoritative DurationSeconds into
device_prep / esp_apps / identity_hello / user_esp / desktop_handoff spans
plus an explicit unattributed remainder — invariant sum(spans)+unattributed
== wall clock enforced in every test. WhiteGlove = two observation windows
in ONE session (part1 anchored on whiteglove_part1_complete, pause never
attributed). Per-app blocking intervals from EVENT timestamps (first start
-> LAST terminal, audit Q1 retry freeze), occupancy = merged union clipped
to esp_apps. Reboot spans from event gaps around lastBootUtc (audit Q7),
deduped per boot. QualityFlags: ClockSkewDropped, PartialObservation,
BlockingSetUnknown, BlockingSetTruncated, WhiteGloveAnchorsIncomplete.
Persistence (SessionTimeBreakdowns) and surfaces follow in PR2/PR3.
Backend 3368/3368 (35 new tests: 19 calculator golden fixtures + 16
identity/blocking/aggregate-exclusion), solution build clean, web untouched
(new response fields are additive).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Source-data audit (tasks/insights-expansion-spec.md §0.5) verified against
production: 17.9% of Succeeded AppInstallSummaries rows carry no duration —
78% are genuine skips (app_install_completed with payload state=Skipped,
"Update for X" WinGet policies not applicable) whose state the ingest
discarded, the rest are real installs whose start the agent never observed
(attach window). Averaging them as 0s understated avg install duration by
18% and distorted the slowest-apps ranking.
- AppInstallSummary.TerminalState (Installed/Skipped/Postponed/Error) from
the closing event's payload state; sentinel-gated column + mapper +
AppsDashboardProjection. Status stays "Succeeded" for compatibility.
- MetricsMath.IsSkipTerminalState/HasMeasuredDuration: durations computed
over measured rows only; skips leave the failure/success rate (decision
2026-07-26 — a skip is not an install attempt); per-app skipped/unmeasured
counts + totals exposed; slowest ranking gates on measured installs.
Applied consistently: app metrics payload, apps dashboard (list/analytics/
time series/trend/versions/device models), SLA weekly/snapshot/breach.
- StoreAppInstallSummaryAsync: reconcile extracted to internal static
ReconcileAppInstallSummaryWithExisting (test seam). Fixes Q4 (terminal
batch before started batch left duration permanently unset) and hardens
terminal status against late/replayed InProgress batches.
- Web: measured-only hints on app detail cards, skip suffix in apps list,
fleet-health "measured installs" caption.
Backend 3333/3333 (20 new tests), web tsc clean, vitest 627/627.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The "Collect Logs" feature (b382e567) added a second opt-in path into hosted
diagnostics upload — the quick-config dialog on a session — and a new collection
trigger, neither of which the customer-facing pages described. The trust FAQ still
said hosted upload is enabled "in the settings UI", and the dialog itself named
hosted storage without stating that packages leave the tenant boundary, so the
"clearly marked disclosure" promise was not met on that path.
- CollectLogsButton: marked data-boundary disclosure in the quick-config dialog
(leaves your Azure tenant, per-tenant isolation, blob-scoped write-only token,
removal per Data Retention Days)
- privacy + about pages: both opt-in paths named; on-demand collection described
(Admin/Operator of the own tenant, same content as the automatic package,
delivered with the agent's next check-in, visible on the session timeline)
- docs repo counterpart in trust/security-faq.md (separate commit)
Distress-bell dedup: the TpmPssUnsupported one-shot bell had key-shape tests only.
New HardwareRejectionNotificationTrackerDedupTests pins the 409 dedup branch, the
fail-soft branch, the blank-input short-circuit, the disjoint hardware/tpmpss key
spaces, and the retention sweep — which also prunes tpmpss rows and thereby re-arms
that device's bell, now stated in the interface and implementation docs.
Backend tests 3313/3313, tsc clean, web vitest app/sessions 60/60.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Rule list cards (analyze rules, gather rules, IME log patterns) used a
single-line flex header where every badge was flex-shrink-0 and only the
title could shrink — on phone widths the title truncated to a few
letters. The header now wraps: on mobile the title drops onto its own
full-width row (order-last + w-full) under the badge row and may wrap;
on >=sm the layout is unchanged (title inline, truncated).
PhaseTimeline squeezed up to 8 flex-1 columns with 48px circles and
whitespace-nowrap labels into the viewport, so circles and label text
overlapped completely on mobile. The timeline now has a minimum width
of 80px per phase inside an overflow-x-auto container (swipe on phone,
no scrollbar on desktop) and labels wrap inside their own column
instead of overlaying neighbours.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dark mode: globals.css was missing the entire slate family (the
"Incomplete" status filter chip rendered white-on-white) and all light
hover tints outside gray/indigo/cyan (hover:bg-purple-50 in the tenant
suggestion dropdown flashed near-white under light text). Added slate
bg/text/border mappings mirroring gray, plus dark equivalents for every
hover:bg-*-50/100/200 class in use (blue, purple, red, amber, yellow,
green, orange, sky, teal, emerald, slate, gray-200/300).
Mobile: ScriptExecutions rows were single-line flex with flex-shrink-0
right clusters, so badges and status text overlapped/overflowed on
narrow screens. Header counters, cycle card headers and item rows now
flex-wrap with gap spacing instead of space-x.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A Global Admin editing a gather/analyze rule while viewing a FOREIGN
tenant (override) got a green "saved" banner, but the write landed in
the GA's OWN home tenant: the JWT-scoped rules/{type}/{ruleId} routes
resolve the tenant from the caller's token, and no cross-tenant write
route existed (reads had /api/global/rules/*, writes did not). The
reload re-read the foreign tenant and showed the unchanged rule.
Fix: symmetric cross-tenant WRITE routes mirroring the global reads.
Backend:
- GlobalRulesFunction: PUT/DELETE global/rules/gather/{ruleId} and
global/rules/analyze/{ruleId}; tenant from ?tenantId= (400 if
missing), same 1 MB guard + ValidateScopeAndEmitMode as the JWT
variants, same service calls (toggle-partial/built-in-state/full-edit
semantics are tenant-parameterized and identical).
- Catalog: 4 entries GlobalAdminOnly + TenantScoping.QueryParam — a
write tier, so the delegated-read rescue never fires; only a platform
admin may mutate another tenant's rules. No global CREATE route:
create stays JWT-only and the portal hides Create in override.
Web:
- api.rules.globalGatherRule/globalAnalyzeRule helpers; both rule pages
route save/toggle/mark-failed/notify/delete through a
*RuleWriteUrl(ruleId) switch on isGlobalOverride.
- Analyze template-enable (a create) is blocked in override with a
clear message instead of silently writing to the wrong tenant.
Tests: CrossTenantAccessTests guards the 4 routes are GA-only +
QueryParam and that a GA reaches the named target tenant. Backend
3295/3295, web vitest 627/627, tsc clean.
Note: the old bug planted orphan copies of foreign rules into the GA's
home tenant on each attempted cross-tenant edit; those are not
auto-cleaned.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
TpmPssUnsupported distress reports (TPM cannot RSA-PSS-sign, Schannel drops
the client cert on Win11 25H2+, agent can never authenticate) were operator-
only, yet the remediation is tenant-side: TPM firmware update or device
replacement.
- New GET audit/tpm-pss-unsupported (MemberRead), aggregated per serial,
mirroring audit/device-not-registered
- Bell notification "tpm_pss_unsupported" (Admin audience) once per
(tenant, serial); dedup via the existing rejection tracker table with a
disjoint "tpmpss|{serial}" row-key space; no webhook by design
- Web: TpmIncompatibleDevicesInsights panel in the Hardware Whitelist
section; renders only when a device is affected (no permanent card)
No agent change, no ConfigVersion bump. Backend tests 3287/3287, tsc clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Session 8e4cc4ae: two "Collect Logs" round trips completed on the agent (both
ZIPs landed in hosted storage) but the portal never flipped to Download and
the button hung at "Collecting..." — the ServerActionDispatcher confirms an
on-demand upload only via server_action_executed(actionType=request_diagnostics,
blobName=...) and emits NO diagnostics_uploaded (only the terminal upload path
does), so UpdateSessionDiagnosticsBlobAsync never ran.
The ingest classification now treats such an executed event as the upload
confirmation (a real diagnostics_uploaded in the same batch still wins — it
carries destination). Because the executed event has no destination field,
InferDiagnosticsDestination derives "Hosted" from the backend-built
"{tenantId}/{filename}" blob-name shape (CustomerSas blobs are bare
filenames); without it the download route would look in the customer's
container for a hosted blob.
Backend-only — works with the already-deployed agent fleet, no agent release.
Backend suite 3274/3274.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Operators previously saw editable settings forms bound to hardcoded defaults
whose Save always 403'd. Now they see the tenant's real configuration
read-only, with no save affordances at all.
Backend:
- GET config/{tenantId}: TenantAdminOrGlobalReader -> MemberRead. Safe because
the handler redacts by default (CanViewSecrets: only GA or own-tenant Admin
see secrets); Operators/Viewers receive RedactedCopyForReader().
- POST diagnostics/files: TenantAdminOrGA -> TenantAdminOrOperator (submitting
logs is a troubleshooting action; Viewer and read-only GlobalReader excluded).
- Tests: policy pins updated; new middleware tests (own-tenant Operator allowed
+ redacted, cross-tenant Operator 403); CanViewSecrets theory for
Operator/Viewer.
Web:
- TenantConfigContext loads the full config for all settings users (403
fallback to feature-flags minimal config for deploy-order safety) and
exposes canEditConfig; saveConfiguration guards on it.
- New ReadOnlyFieldset wraps section content (native fieldset disable, values
stay copyable); SaveResetBar is not rendered at all for read-only viewers.
Applied to Contact, SLA Targets, Hardware Whitelist (+ rejection-insights
toggle), Notifications, Agent Settings, Agent Analyzers, Diagnostics, Data
Management.
- Hidden entirely from Operators (in-page admin-only card + sidebar sub-item
gate): Autopilot Validation, Access Management, Offboarding, Unrestricted
Mode. GlobalReader sidebar parity kept via hasGlobalScope.
Tests: backend 3257/3257, web vitest 627/627, tsc clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Covers the queue-on-session/deliver-on-ingest channel, the
TenantAdminOrOperator tier with in-function per-type re-gating, the portal
Collect Logs flow (quick-config orders rotate_config before
request_diagnostics), sequence-based progress matching, and the hosted-prefix
sweep in the offboarding cascade. Registered in index.md, logged in log.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Queues a request_diagnostics ServerAction (delivered with the agent's next
ingest response) and tracks progress live: server_action_received flips the
button to "Collecting…", the DiagnosticsBlobName SignalR delta completes it
(Download Diagnostics appears), server_action_failed or a 3-min quiet timeout
surface as toasts.
- Button is ALWAYS rendered — disabled states carry tooltips (inactive
session, cross-tenant view, missing role, diagnostics not configured) as a
deliberate nudge towards enabling diagnostics upload.
- Tenant Admin on an unconfigured tenant gets a quick-config dialog: switches
the tenant to Hosted + OnFailure (GET config -> patch two fields -> PUT
verbatim), then queues rotate_config BEFORE request_diagnostics so the agent
collects against the fresh config. Operators/Viewers never see the dialog.
- Operator may collect when configured (new TenantAdminOrOperator backend tier).
- useSessionTenantConfig picks up diagnosticsUploadConfigured from the
member-readable feature-flags endpoint.
- Progress matching is sequence-based (backend-assigned) so agent clock skew
and at-least-once action redelivery cannot produce false matches.
Pure state logic in collectLogsLogic.ts with vitest matrix (21 tests).
tsc clean; web suite 627/627.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Oliver Kieselbach