shopdb-flask/docs/notes/endpoint-telemetry-and-ui-phases.md

Endpoint telemetry and PC Details UI - phased working notes

Scratch/reference notes for a multi-phase effort. Not a contract document. When a phase lands, its details belong in AGENT-CONTRACT.md, the csproj changelog, or the panel declarations - not here. Delete this file when all phases are done.

Branch: wilm. Current versions at time of writing: agent 1.8.0, tray 1.1.0.

---

Ground truth established before scoping

Verified in code, not assumed. These constrain everything below.

Panel groups are a hardcoded frontend list. frontend/src/components/PluginAssetPanels.vue:423:

const RENDERED_GROUPS = ['details', 'monitors', 'software', 'activity']

Any panel whose group is not in that list silently falls back to details. Nothing is logged. The comment directly above it records that this has already caused two disappearing-card incidents (actions -> monitors, and the metrics -> monitors rename). Tab display labels are separately hardcoded in plugins/computers/frontend/views/PCDetail.vue:441-449.

Consequence: introducing a new tab is a THREE-file change that must land in one commit - panel group in plugins/endpoints/plugin.py, RENDERED_GROUPS, and the PCDetail label map. Split it and cards vanish with no error.

There is no pinned/locked card concept today. Layout is a CSS grid ordered purely by the integer position on each panel. There is no "top row", no span control except one special case: the Network interfaces table sets full-width span because at half width it wrapped every cell to three lines (PluginAssetPanels.vue:693).

Consequence: "lock this card top-left/top-right" does not exist yet and is itself a feature, not a config value. See Phase 5.

Two of the requested cards are not plugin panels at all. Both the Entra/Intune card (PCDetail.vue:202-243) and the Network/IP card (PCDetail.vue:249-265) are hand-written Vue inside PCDetail's details grid, not declarations in plugin.py. Moving them is a component refactor, not a position/group edit.

There is no LAPS retrieval feature in this codebase. Searched: no LAPS button, endpoint, or credential path exists. The only hit is a historical incident note in plugins/tools/frontend/views/SecureSendCreate.vue:254 about LAPS passwords once being sent over a chat link. The request assumes a click-to-retrieve LAPS button already exists on the Entra card; it does not. This needs its own scoping conversation - it is a credential-disclosure path with real audit/permission implications and must not be smuggled in as part of a layout change. See Open Questions.

Existing panel positions in monitors (so new cards do not collide): 10 status, 12 tracked services, 13 tracked processes, 14 hardware, 15 osinfo, 16 interfaces, 17 storage, 18 operator reports, 19 listeners, 40 user context. activity uses 20. software uses 30.

---

Phase 1 - Mapped drive reachability (agent + tray)

STATUS: code complete, NOT fully validated. See "Where we left off" at the end of this file before resuming.

Highest value, and the reason the user-context work exists. Extends the probe already added in agent 1.8.0 / tray 1.1.0.

No active reachability testing. An earlier draft proposed touching each share to see if it responded, with a per-mapping timeout. Rejected: Windows already reports the status, so probing the share is wasted work and reintroduces the hang risk the bounded probe exists to avoid. Read the status, do not test it.

The two status sources disagree, and that disagreement is the signal. Captured live on the same box at the same moment, same two mappings:

Get-SmbMapping        net use
OK   Q: ...specdata   Disconnected  Q: ...specdata
OK   S: ...shared      Reconnecting  S: ...shared

Get-SmbMapping reported OK for both while net use reported Disconnected and Reconnecting. So neither source alone is trustworthy: collecting only Get-SmbMapping (the nicer, more structured one) would have reported a healthy Q: on a box whose Q: was actually down - exactly the failure this card exists to catch.

Collect BOTH and keep them as separate fields per mapping. Do not silently reconcile them in the agent; the server/UI decides presentation, and preserving the disagreement is diagnostically useful in its own right. The card should surface the WORSE of the two rather than the prettier one.

Status vocabulary must include Reconnecting - it appears in the real capture above and an earlier draft's Ok/Disconnected/Unknown set would have forced it into the wrong bucket. Keep Unknown for probe failure so a failed check never renders as green Ok.

Elevation matters and confirms the tray approach. The capture was taken from an elevated shell (start powershell -verb runas). Because of UAC token separation, an elevated context sees a DIFFERENT mapped-drive table than the normal user - often empty. The tray runs unelevated as the operator, which is the correct context to enumerate from. Do not "improve" this later by elevating the probe; that would silently empty the card.

Mechanism: the probe already writes JSON to the spool file, so this rides the existing handoff - run both commands in-session, write the combined result, service reads it. No new transport.

Server: persist as part of the existing usercontext payload key (no new column, no migration needed - it rides usercontextjson).

Note in code, commented out, per request: a future one-click remediation command will re-establish a Disconnected mapping. Do not implement now, but leave the command-catalog seam obvious. It will be an allowlisted action in plugins/endpoints/commands.py and must run in the USER session (a SYSTEM remap creates a mapping the operator cannot see, and per the elevation note above an elevated remap has the same problem).

Phase 2 - Cheap system-side telemetry (agent service)

All service-side, no session needed. Group these because they share the inventory-snapshot plumbing and one migration.

• Time sync offset and source - CUT. Do not collect, do not surface. w32time is Stopped on provisioned endpoints, confirmed on more than one box (get-service w32time -> Stopped). The service appears to be stopped by server-side policy after provisioning, so w32tm /query /status returns 0x80070426 "service has not been started" fleet-wide and there is no offset to read. Collecting it would return null everywhere and produce an alerting signal that never fires - worse than nothing, because a permanently-null health check reads as "no problem found".

  Implemented and then removed in this phase. If time skew becomes a real suspect later, the prerequisite is getting w32time RUNNING on the fleet; the collection is the easy part. One implementation note worth keeping: parse the offset by LABEL, not by "first value ending in s". w32tm prints Root Delay and Root Dispersion in the same seconds format BEFORE Phase Offset, so a first-match parser silently reports network latency as clock skew. That bug was written and caught here; do not rewrite it.

• Disk errors / SMART predictive failure - DONE. Landed as its own "Disk health" card rather than folded into the Storage card as originally planned: Storage is per VOLUME (C:, D:) and this is per PHYSICAL device, and one drive can back several volumes or none the operator sees, so the two do not share a key and could not be one table. The important finding: the classic ATA SMART class (MSStorageDriver_FailurePredictStatus) returns ZERO instances on an NVMe machine. Verified on a two-NVMe box where both drives were present and healthy. An agent collecting only that class would have reported an empty disk health list on every modern endpoint and read as "no problems" - the exact silent-healthy failure mode this telemetry exists to prevent. So the agent collects BOTH MSFT_PhysicalDisk (the only NVMe-aware source) and Win32_DiskDrive.Status (ATA/SATA, and hosts where the Storage namespace is unavailable), and the server does not reconcile them - same reasoning as the Phase 1 mapped-drive merge, where two disagreeing sources was the signal. Note MSFT_PhysicalDisk returns raw enum INTEGERS through ManagementObjectSearcher, not the friendly strings Get-PhysicalDisk prints (Health=0 Op=2 Media=4 Bus=17 for a healthy NVMe); the mapping lives in the collector so the wire stays self-describing.

• NIC link speed and duplex - DONE. Paired onto the existing interfaces entries by MAC from MSFT_NetAdapter (Win32_NetworkAdapterConfiguration carries no link state, so it is a second query). Both fields are null when unreadable: verified on the live box that a DISCONNECTED adapter returns empty Speed and FullDuplex, so defaulting would have reported every unplugged NIC as half-duplex. Not surfaced.

• Boot type - DONE. Two facts kept separate: hiberbootenabled from the registry (fast startup configured) and lastboottype from Kernel-Boot event 27 (what the last boot did). Neither substitutes for the other - the registry cannot say what the last boot did, and the event can age out of the log. Verified live: BootType=0 with HiberbootEnabled=0. Not surfaced; placement still undecided.

• Windows Update history, last 5 - PINNED, deferred out of Phase 2. The COM update searcher is not cheap: it routinely takes seconds and sometimes tens of seconds, which would make it the most expensive thing in an otherwise WMI-and-registry Collect() running on the heartbeat path.

  The blocking concern is broader than cost. Some endpoints host COM server applications, and driving the update searcher through COM on those boxes risks interfering with collections - or with the hosted application itself. Needs its own investigation before any implementation, and if it lands it wants its own cadence (hourly/daily), never the every-cycle path.

• Defender/AV health - DONE, collected and persisted, NOT surfaced, exactly as specified. Note the ages are in days and a real 0 means updated today, so the collector sends null rather than 0 when Defender did not report. A box with third-party AV or a stripped image has no Defender WMI provider at all, which is normal here and not a fault.

Explicitly dropped after discussion, do not collect: certificates (corporate-managed), BitLocker (network join already requires it), display topology (single monitor on the floor, trained users), session type/RDP (RDP-in not permitted on the shop floor), logon server (local users).

Pending file-rename operations - DONE, but NOT as a new collection. It was already being read: GetPendingReboot() OR-s it together with the CBS and Windows Update keys into one bool. The gap was that the bool cannot say WHICH signal fired, so the "stronger signal" was already being flattened into the weaker one. Shipped as rebootreasons, a list naming the signals that fired, leaving pendingreboot untouched for compatibility. Empty list = checked, nothing pending; null = the check failed.

Schema: these add columns to endpointstates. This landed as 0015, not the 0014 reserved here - 0014 was taken by the disk health column earlier in Phase 2.

Phase 3 - Service/Process Monitors tab split

Tab churn is exactly what has silently eaten cards twice before, so this phase is deliberately isolated from any collection work.

• Rename the existing Monitors tab to System Monitors.
• Add a new Service/Process Monitors tab.
• Move Tracked Services (position 12) and Tracked Processes (13) into it.

Must land atomically across all three places:

1. group on those panels in plugins/endpoints/plugin.py
2. RENDERED_GROUPS in PluginAssetPanels.vue
3. the tab label map in PCDetail.vue

Also update the hardcoded panel-count assertion in tests/test_plugins/test_endpoints_plugin.py (test_endpoints_asset_panels_are_computer_scoped) - it has already needed bumping once (7 -> 12) and will fail again here.

Printers: a location-based printer auto-detect collection is in progress elsewhere. Reserve space and a position band for a printer card in the monitors family so this work does not have to be redone when it lands. Do not build the card now.

Phase 4 - Card content additions

Depends on Phase 2 data existing.

• Mapped drives card - Get-SmbMapping-shaped. Disconnected in red, Ok in green. Reuse existing badge/severity styling rather than inventing new CSS. Room reserved for the future one-click fix action.
• Storage card - superseded. Disk health shipped in Phase 2 as its own "Disk health" card (position 18, directly under Storage) instead of being folded in: Storage is per volume, disk health is per physical device, so they do not share a key. Rows sort failure-first, then unknown, then OK - a failing drive must not sit below the fold on a multi-disk box, and a drive that stopped answering is a weaker claim of health than one that said OK.
• Windows Update card - BLOCKED, not scheduled. Depends on Phase 2 data that is pinned pending a COM-interference investigation. Do not build the card until the collection question is settled.
• Boot type placement to be decided when the data is visible.
• Storage + Disk health into one card with two collapsible sections - DONE. VISUAL GROUPING ONLY. These two tables are not merged into shared rows: Storage is per volume, Disk health is per physical device, and one drive can back several volumes or none the operator sees. This is emphatically NOT the Phase 1 mapped-drive treatment, where two sources described the same drive and the worse status won. Nothing here is reconciled. Built as predicted: tabs was extended to support table sections (columns/ rows/empty per section) in the shared renderer, and the two routes collapsed into one /storage returning sections. /diskhealth is kept as a route - it is a stable read for non-panel callers, and removing it for a UI-only change would break them. Two things only found by building it:
  • The empty-section filter had to grow a table branch. A table section declares its columns up front and so must survive having no rows; dropping it takes its TAB with it, and a missing tab reads as "this machine has no disks" rather than "nothing was reported".
  • Full-width span was keyed off render === 'table'. A tabs panel holding a six-column table needs the same span, and whether it does is only knowable from FETCHED DATA, not the declaration - hence spansFullWidth(). Both are mutation-tested in tabSectionsTable.spec.js. Checked and cleared: the "inline label/value pairs, not nested cards" density rule in Phase 5 does NOT block this. That rule is scoped to the PC Details HEADER block, where the constraint is adding zero height so the tab strip is not pushed toward the fold. It says nothing about panel bodies in the Monitors tab.

Phase 5 - Header right-hand block (DEFERRED - revisit before building)

Status: on hold at user request. The original ask assumed pinned cards were a config value; they are not, and building a pinning system was not what the user wanted to sign up for. Not scheduled. Nothing below is committed.

The cascade worry that motivated pinning is explicitly parked - do not build anything for it unless reflow problems actually surface.

Revised direction: use the empty right-hand half of the existing header. The PC Details header is a full-width card whose right side is currently dead space - the hostname, badges, PC type, OS and location all sit left, and the Status / Remote access / Refresh / Restart row runs along the bottom. The block to the right of the identity lines is empty at every width.

This is a better target than either pinned grid cells or extending the header downward:

• It consumes space that already exists, so it costs ZERO additional vertical height and cannot push the tab strip toward the fold. That was the hard constraint, and this approach satisfies it structurally rather than by budgeting.
• It sidesteps the Network interfaces width problem entirely (that table is full-width by necessity and would fight a half-width pinned slot).
• Header content lives outside the panel grid, so it cannot be reflowed by cards appearing or disappearing below it.

Candidates for the block, unresolved:

• Entra/Intune deep links (currently a Details-tab card). Strongest candidate: they are navigation, not data, and the user has called out their outbound usability value.
• IP / network summary (currently the hand-written Network card).

Density rules if this proceeds: inline label/value pairs, not nested cards. Do not let the right block grow taller than the left identity block, or it starts adding height after all and the zero-cost property is lost.

Scope note: that "not nested cards" rule is about THIS header block only, where the whole value is adding zero height. It is not a general ban on sections inside a panel body - the Phase 4 Storage + Disk health grouping was checked against it and is not blocked.

Still true and still unbuilt if this proceeds:

• Collectors is not a current tab or group. If IP History moves to a new tab, that is the same three-file atomic change as Phase 3.
• Entra empty state, requested: show Please provide DeviceID / Please provide ObjectID when the backend has no value, rather than an empty or broken card.

LAPS is NOT a feature request - clarified by the user as an illustration of why these outbound portal links matter to usability. No LAPS work is scoped.

Phase 6 - Intune/Entra link bug (independent, can be done anytime)

Reported: the INTUNE and ENTRA buttons sometimes do not appear until the page is refreshed. Diagnosed, root cause found, NOT yet fixed.

Cause: a cold-cache race, not a render problem.

frontend/src/utils/siteSettings.js:9 caches only the RESOLVED settings object:

let settingsCache = null
async function loadSettings() {
  if (settingsCache) return settingsCache
  ...

The in-flight promise is not cached, so every caller during the first load fires its own settingsApi.list().

In plugins/computers/frontend/views/PCDetail.vue:513, getEntraDeviceUrls() is deliberately fire-and-forget (.then(...), not awaited) so a settings failure cannot block the page. Meanwhile computersApi.get() on line 519 IS awaited. On a cold cache the settings response can therefore land AFTER the computer data has rendered. Until it does, entraUrlTemplate / intuneUrlTemplate are still '', so fillTemplate() returns null (line 426: if (!template || !value) return null), and the v-if on the anchors hides both links. A later refresh hits the warm cache and they appear.

That also explains the intermittency: it depends on which response wins, so it varies by network timing and by whether another view already warmed the cache.

Fix: cache the in-flight promise, not just the result. Assign the promise to settingsCache immediately so concurrent callers await the same request. This is a one-line-shaped change in siteSettings.js and fixes every consumer of site settings, not just this card - the same race exists for any view that reads settings during mount.

Watch for: the existing catch sets settingsCache = {} on failure, which permanently caches the empty result for the page session. Confirm whether a failed settings load should stay cached (currently it does) or be retryable. Do not change that behavior silently while fixing the race.

Verify after fixing that the links appear on a hard reload with a cold cache and throttled network, which is the condition that reproduces it.

DONE. Fixed in siteSettings.js: the in-flight promise is now cached alongside the result, and a failed load is cached for 30s rather than the page session. Covered by frontend/src/utils/siteSettings.spec.js (6 tests). Open question 5 resolved in favour of the short TTL.

Measured side benefit: the old code assigned settingsCache only AFTER the response resolved, so each sequential await re-entered loadSettings() before the cache existed. applyBranding() awaits five settings in a row at bootstrap, getServicenowUrls() four, getEntraDeviceUrls() two - all of which were separate GETs on a cold load. Now one.

Phase 7 - PCForm "Failed to load data" (NOT the same bug)

Separate from Phase 6 and NOT caused by it - PCForm.vue does not import siteSettings.js. Surfaced on the F3KKSGH4 edit page.

Partially addressed, root cause still UNKNOWN.

Evidence: the form renders with reference data present (printer list populated, dropdowns built) but record fields empty. That means the eight-way Promise.all of reference endpoints SUCCEEDED and the failure is in the edit-only block, PCForm.vue lines 589-635. The only unprotected call there is computersApi.get(route.params.id); ensureModelLoaded() already swallows its own errors, so it is not the source.

Two candidates remain, with different fixes and different layers:

• the computersApi.get request fails (backend/route problem), or
• a TypeError while mapping the response into form.value (frontend shape mismatch, plausibly from recent schema work).

Both land in the same catch and both printed the same generic string, which is why this could not be diagnosed from the UI.

Done so far: the catch now uses apiError(err, ...) so the real server message surfaces instead of a generic one, and the stale comment claiming a call "handles its own failure" was removed - no such call remains in that list.

NEXT: reload the edit page and read the now-specific error. An axios status code points at the backend; a TypeError names the property and points at the mapping block. Do not guess between them.

Also worth doing regardless: a single failed reference list currently blanks the whole form via Promise.all rejection. Degrading one dropdown instead would be the behaviour the removed comment falsely claimed.

---

Open questions - answer before the relevant phase

1. LAPS - RESOLVED: not a feature request. The user was illustrating why the outbound Intune/Entra links matter to usability. No LAPS work scoped. The security concerns noted earlier stand if it is ever revisited.
2. Pending file-rename operations - RESOLVED: collect. Folded into Phase 2.
3. Phase 5 pinning / cascade symptom - RESOLVED: not worrying about the cascade unless it surfaces. Pinned cards dropped in favour of the header's empty right-hand block; Phase 5 deferred.
4. Collectors tab - only relevant if Phase 5 proceeds. New tab for one moved card, or does IP History belong in an existing tab?
5. Failed settings load caching - Phase 6. Should a failed settingsApi.list() stay cached for the page session (current behavior) or be retryable? Decide while fixing the race, do not change it silently.

Sequencing

1 -> 2 -> 3 -> 4. Phase 6 is independent and can be slotted in anytime; it is small, self-contained, and fixes a user-visible bug, so it is a good candidate to do early. Phase 5 is deferred and unscheduled.

Phases 1 and 2 are additive and independently deployable. Phase 3 is a contained tab move. Phase 4 needs 2.

Agent/tray version bumps: Phase 1 touches the probe (tray behavior) and Phase 2 touches the service collector. Both need csproj changelog entries. Note the existing coupling hazard - Program.cs has no unknown-arg guard, so an older agent receiving a newer tray's verb starts a full second host rather than no-oping. Fix that guard before any deployment where tray and agent could skew.

Phase 8 - fetchAllPages returns a bare array (LAST, and it affects the above)

Deliberately last: it is a cross-cutting API-surface change, and every earlier phase should land against the current shape rather than a moving one.

frontend/src/api/index.js:131 - fetchAllPages() returns the ROWS:

return items

Every other method on every API object returns the axios RESPONSE. So callers must remember, per call site, which of two shapes they are holding:

const rows = await vendorsApi.listAll()       // array
const response = await vendorsApi.list()      // { data: { data: [...] } }

Reading .data.data off the array yields Cannot read properties of undefined (reading 'data'). Inside a Promise.all that rejects the whole batch, so one wrong unwrap blanks an entire form - and if it throws before the record fetch, NO request is made at all, which makes it look like a backend or schema fault. That is exactly how the PC edit page presented, and why it was misdiagnosed twice before the log showed /api/computers/729 was never requested.

Known history - this class of bug has now hit at least four times:

• MachineForm.vue pcsRes (fixed previously, comment left in place)
• PCForm.vue vendorRes (fixed now)
• MachineForm.vue vendorRes and buRes (fixed now - the warning comment was already sitting three lines away and the adjacent defects were still missed)
• MeasuringToolForm.vue buResponse (fixed now)

ConfigurationPoliciesList.vue:187 defends with Array.isArray(value) ? value : ..., which is evidence the shape is not trustworthy at the call site.

Why this matters to the phases above: Phase 1/2 UI work adds call sites that consume paged reference data, and Phase 5 (if revived) moves cards between components. Any new listAll() caller is another chance to hit this. Landing the phases first and this last means one migration pass, not a moving target.

Options, in order of preference:

1. Make fetchAllPages return { data: { data: items } } so every API method has ONE shape. Mechanical but touches every existing listAll() caller in the same commit, or they break in the opposite direction. Highest value, highest blast radius - needs the full grep and a single atomic change.
2. Rename to something the shape is obvious from (fetchAllRows, listAllRows). Cheap, no behavior change, keeps the trap but makes it visible at the call site. Good interim step if 1 is too large right now.
3. Leave it and rely on review. Rejected - four occurrences say review does not catch it, and one of them happened directly beneath a comment warning about it.

Whichever is chosen, do the full-codebase grep for listAll( first; there are callers in core views, print views, and at least seven plugin frontends.

---

Where we left off

Written mid-session when the test tooling stopped working. Read this first.

Phase 1 is code complete but NOT fully validated

What landed, all uncommitted in the working tree:

agent/ShopDbEndpointAgent.Core/UserContextProbe.cs - added Get-SmbMapping alongside the existing Win32_NetworkConnection capture, emitted as a separate smbmapping key with its own smbmappingerror. The two sources are NOT reconciled in the probe, deliberately - see the Phase 1 notes above. Still unelevated, still in the operator session.

plugins/endpoints/api/routes.py - added _drive_letter, _DRIVE_STATUS_RANK, _worse_drive_status, _mapped_drives, and pointed panel_usercontext() at _mapped_drives(). Merges the two views by drive letter and shows the WORSE status. Carries a long WARNING comment naming the three load-bearing behaviours; that comment is doing real work, leave it.

plugins/endpoints/commands.py - commented-out remap_drive seam with the user-session/elevation reasoning. Comment only, no behavior.

Validation state - the important part

Last KNOWN-GOOD full run: 8 failed, 2489 passed, 11 skipped. All 8 failures are pre-existing and unrelated:

• 3x test_endpoints_plugin.py timestamp tests - these fail ONLY on Windows. _in_zone() calls time.tzset(), which does not exist on Windows (Unix-only in CPython), so it raises AttributeError before the assertions ever run. _parse_dt itself is correct. These pass on CI. Consider a skipif(not hasattr(time, 'tzset')) so the suite is honest on Windows.
• 5x docs / security-headers / plugin-guard, all pre-existing.

TWO CHANGES LANDED AFTER THAT RUN AND ARE UNVERIFIED:

1. _mapped_drives() now renders (Unknown) for a drive neither source gave a status for. Previously it rendered blank, which reads as healthy - the exact silent-green failure the warning comment warns about. Small but real, behavioral, and untested.
2. The commands.py seam. Comment only, zero risk.

get_errors on both changed files is clean, but that only proves they parse.

Tooling failure that stopped validation

Every run_command_in_terminal call began failing with "A task was canceled", including a two-file pytest run that should finish in under a second, and a background launch that never wrote its Tee-Object output file. Two full-suite runs had succeeded earlier in the same session, so this is not slowness - the execution path went down partway through.

Test Explorer discovers all 2508 tests but returns zero results for every filter tried (TypeName, MethodName, FullyQualifiedName). Likely needs a VS restart.

Next actions, in order

1. Restart VS, then run python -m pytest tests/test_plugins/test_endpoints_plugin.py -q. Expect 3 failed (the tzset ones) / 23 passed. Anything else is ours.
2. Consider adding tests for _mapped_drives() - currently zero coverage on logic that decides whether a broken drive renders as broken. The cases that matter: sources disagree (worse wins), drive in only one source (kept), no status at all (Unknown, not blank), Get-SmbMapping absent entirely (degrades to netuse alone).
3. Only then start Phase 2. Do not begin new collection work on an unverified base - the untested (Unknown) delta would get buried under it.

Deferred, agreed explicitly

Agent and tray republish happens at the END of all phases, not per-phase. The server change is safe to deploy ahead of the agent: until the new probe ships, smbmapping is simply absent from the payload and _mapped_drives() falls through to netuse alone, rendering exactly as it does today. No ordering constraint between them.