New-finding telemetry lost older activity once the run history reached its
100-entry cap. Persist a bounded daily finding tally with a separate
upgrade cursor, preserving run counts while backfilling retained findings.
Cover restart, repeated saves, upgrade, read failure and UTC-day retention.
Record the measurement boundary and retire the resolved coverage gap.
A three-week decline in investigations/new_findings (7.2% to 5.0%) looked
like a Patrol regression. It is not one. The ratio is not a rate at all:
the two counters come from different stores, cover different spans, and
are drawn from populations that barely overlap.
new_findings_30d sums run.NewFindings over history.Runs, which
SavePatrolRunHistory caps at MaxPatrolRunHistory, so it covers at most the
last hundred runs rather than thirty days. investigations_30d instead
scans the current findings store and counts surviving finding records
investigated in-window, including findings created before it, which is how
the paid cohort read 128.57% in the week to 2026-08-25. Finding
ShouldInvestigate returns false at monitor autonomy and effective autonomy
is licence-gated, so free installs produced 4384 findings and 1
investigation while 67 paid installs produced 242.
The decline was composition: flat in version-stable installs, and fleet
investigations rose once the single install that swung the total by 38 was
excluded. Finding-detection code is identical between v6.3.2 and v6.4.1.
The new test pins the asymmetry behind the bad denominator. Its twin
already asserts that runs_30d ignores the history cap after 63c40ebe5e;
nothing asserted that the findings loop immediately below it does not, so
the truncation could regress or be mistaken for a thirty-day total
unnoticed. Fixing it needs a per-day findings tally alongside DailyRuns,
which the coverage gap tracks as its own slice.
pkg/server tests boot the real server through Run() with the version
literal "test-version", which internal/updates normalizes to
0.0.0-test-version. Each test runs against its own t.TempDir(), so every
run minted a fresh install ID. The startup ping waits two minutes and so
never fired inside a short test, but the service-health failure reporter
added on 2026-08-29 sends synchronously from a deferred handler as soon
as Run() returns an error, so every CI shard containing pkg/server posted
one ping.
The licence server recorded 317 single-ping installs between 2026-08-29
and 2026-09-03 - 311 from linux/amd64 CI runners, 3 from a maintainer
workstation - still arriving at roughly 60 a day. The canonical clean
denominator excludes single-ping installs and was unaffected, but raw
install counts and the operator-evidence blocked-cause read counted them
as real installations.
A test binary is not an installation, which is the same reason mock mode
already suppresses pings, so the guard belongs beside it in the telemetry
package rather than at the four call sites: send() now refuses the
production endpoint whenever testing.Testing() reports true. The check
compares against productionPingEndpoint, so telemetry's own tests keep
asserting on real ping content through a redirected endpoint, and the
server tests additionally opt out at the config layer to say so locally.
The 2026-09-01 Patrol assessment could not answer four questions from
telemetry: what share of Patrol installs run local versus cloud models,
what Patrol costs an install per month (issue 1789 estimates ~104k input
tokens per full run at ~5.5 runs a day, above the Pro fee on non-Flash
tiers), how the 240 of 255 fleet investigations that produced no plan
actually ended, and which effective Patrol mode an install runs.
Schema v17 adds closed buckets only. ai_provider_class classifies the
Patrol model route (none, local, cloud_byok, cloud_subscription,
hosted_quickstart, unknown) with a syntactic private-host check on custom
OpenAI-compatible endpoints; it never resolves DNS or carries a provider,
model, endpoint, or account. pulse_intelligence_patrol_autonomy_level is
the effective level after licence and Autopilot gating. The two token
fields bucket the existing local usage ledger's Patrol events, so exact
counts and prices stay on the install. Thirteen investigation outcome
counters partition the findings already counted as investigated, one
bucket per finding, adding no finding, resource, or session identity.
The strings are never omitted, so an empty value can only mean a pre-v17
sender; the receiver stores those as unknown. PRIVACY.md, its shipped
mirror, the Settings preview interface, and the security-privacy and
api-contracts subsystem notes carry the disclosure, and the coverage gap
is registered in status.json.
The agent-lifecycle and storage-recovery contracts name internal/api under
their extension points, so both record that this telemetry is adjacent
adoption analytics and never agent or storage state. The stable E2E tier
pins the public schema version, so the disclosure spec moves to 17, and
the Settings preview interface change carries a browser receipt from a
scratch build of this tree at 1280x800 and 390x844. status.json also drops
an identical duplicate of the ai-provider-guided-setup coverage-gap record
that two merges each appended, which the private governance audit rejects.
The bucket for a legacy config still pointing at the retired Pulse-hosted
route is named hosted_legacy: the shipped privacy document discloses the
vocabulary verbatim and the frontend copy contract keeps that document
free of the retired hosted quickstart wording, so the telemetry package
now pins the vocabulary at the source.
Resolve the process memory-controller hierarchy before selecting the tightest v1 hard limit, so systemd and nested appliance limits actually inform the existing Go runtime headroom.
Contract-Neutral: runtime memory-limit detection only; no API or persistence contract change
Telemetry could see only saved connections, so an install that tried to
reach a node and could not was indistinguishable from one that never
opened the add-node dialog. Both report zero configured connections and
stall at the same activation stage. Fleet data shows that population is
real and concentrated three to one in container deployments, and nothing
recorded whether those installs attempted a connection at all.
Record node connection test attempts and failures in a bounded,
day-bucketed tally in the config directory, pruned to a 31-day retention
window, and report both over the install-ID rotation window as
node_test_attempts_30d and node_test_failures_30d.
Recording starts only once a request carries a target and credentials, so
an incomplete form is never counted as a node that could not be reached.
A host string that turns out to be unusable does count, because the
attempt was made and it failed. Only the add-node dialog endpoint is
instrumented: instrumenting the unused test-config endpoint as well would
double-count a single operator action.
The tally holds counts alone. Hosts, credentials, and error text never
enter it, which is why it is plain JSON rather than encrypted history.
An enabled Patrol that can never run and one that runs and finds nothing
were indistinguishable in the fleet: both presented as high run counts
with zero AI calls and zero findings. The install that motivated this
sat blocked for over a month because provider initialisation failed once
at boot and was never retried; the self-heal landed separately, but
telemetry still cannot see which blocked cause dominates in the field.
Schema v10 exports the fixed machine cause code (for example
provider_not_configured) only while Patrol is in the blocked runtime
state. The cause rides the router-owned Pulse Intelligence snapshot into
the outbound ping. Blocked-reason text, provider endpoints, model names,
and configuration stay on the install; an untyped blocked reason exports
nothing rather than free text, and a disabled, active, or mid-run Patrol
exports an empty value even when a stale cause is still recorded.
Two counters could not answer the question they exist for.
pulse_intelligence_patrol_runs_30d was counted from the operator-facing
patrol run history, which is capped at MaxPatrolRunHistory (100). On any
install patrolling on a normal schedule that cap is reached within hours
of a thirty-day window opening, so the field reported the cap rather than
the run count, and the calls-per-run ratio derived from it was inflated by
a censored denominator. Patrol run history now carries an uncapped daily
tally beside the capped list, advanced from a persisted high-water mark so
repeated full-list saves cannot double count, pruned to 31 days, and read
in preference to the list. An install with no tally yet falls back to the
list, so the counter never regresses on upgrade.
Schema v8 split agent-side pre-mutation refusals into target-change,
prerequisite and contract categories so they would stop hiding in "other".
In production all three are zero fleet-wide and "other" still absorbs every
refusal, because the dominant reason code is the legacy preflight_refused
aggregate recorded when an agent sends no machine reason code at all.
Agents older than the typed refusal contract report every refusal that way,
so the split is starved rather than broken, and folding the two together
made those two states indistinguishable. Schema v9 counts uncoded refusals
separately.
A handler that flushes before its first write commits the response
headers with an implicit 200. The gzip wrapper deferred its
compress-or-not decision to the first write, so such a handler would
send headers without Content-Encoding and then a compressed body,
which clients cannot decode. Every current flusher is an SSE handler
whose content type never compresses, so this was latent, but any
future streaming JSON endpoint would have corrupted silently. Flush
now runs the decision first, matching net/http header-commit
semantics.
Large estates that exceed the WebSocket frame ceiling recover over an
uncompressed /api/state poll loop, which made the biggest deployments
pay the heaviest transfer cost. Wrap the main HTTP handler in a gzip
middleware that compresses JSON and other text responses roughly an
order of magnitude, while leaving WebSocket upgrades, SSE streams,
Range requests, small declared bodies, and binary content untouched.
Contract-Neutral: transport-only gzip response middleware, no RBAC or persistence semantics change
The 2026-08-07 telemetry read showed installs at or above 5 PVE nodes, 10
Docker hosts, or 3 VMware hosts convert to paid at ~8x the rate of smaller
estates. The ping now carries that classification as a server-derived
boolean so receiver-side cohort queries keep a stable column even if the
thresholds move later; it is derived in the pkg/server snapshot closure
from the same AggregateInstallSnapshotCounts values the payload already
sends, so no new information leaves the install.
The thresholds move to internal/monitoring/business_estate.go as the
single definition; the session-capability surface behind the in-product
business-estate card delegates to it, and dropping the now-unneeded direct
pkg/licensing import there restores TestPkgLicensingImportBoundary, which
f0e2243b4 had left red. All three payload surfaces (Ping struct, private
receiver, TelemetryPingPreview) move together per
check_telemetry_schema_parity.py, and both PRIVACY.md copies document the
field.
Verified live on an isolated worktree backend with 6 mock PVE nodes: the
Settings telemetry preview renders schema_version 8 with business_estate
true, and /api/security/status still reports
sessionCapabilities.businessEstate true through the delegated thresholds.
Schema v6 shipped audit_logging_persistent and audit_events_30d as Pro adoption
signals. Neither discriminated. pkg/server installs the SQLite audit logger on
every install for defense in depth and gates only the read/export endpoints, so
the boolean was true on all 8 installs that had taken rc.8 and 0 rows in the
retained table have ever had it false. The event count measured that background
write volume: three of those eight unlicensed community installs were pegged at
the receiver's 100000 clamp ceiling, with the rest between 4863 and 67509.
Schema v7 replaces both with audit_reads_30d, a count of requests that cleared
the license gate on an audit read or export surface. A read requires a human
action, so unlike store presence or write volume it cannot settle into a
constant. The recorder is wrapped INSIDE RequireLicenseFeature so unentitled
requests never count, and the persisted marker carries a timestamp and a coarse
activity class from a fixed allowlist. Query filters, actors, ranges, and every
audit row read stay on the install.
The retired columns are left in the live database. They hold real rc.8 rows and
migrations only add, so dropping them would be a pointless risk; nothing writes
them once the receiver struct loses the fields.
Adds the guard this class needed. LicensedFeatureAdoptionFields registers every
field that exists to measure licensed-feature adoption, and
TestLicensedFeatureAdoptionFieldsDiscriminate builds an unused install through
the real production snapshot paths, installs a real SQLite audit logger exactly
as pkg/server does, records a baseline audit event, and fails if any registered
field is non-zero. Pinning a console logger there would have made the guard pass
while the payload lied, so it deliberately does not. The guard was verified by
reintroducing the v6 sourcing and confirming it fails with the field named.
A companion test pins the three retired fields so they cannot return under
their old names.
This is the third instance of one bug class. v6 removed
pulse_intelligence_patrol_autofixes_30d, hardcoded to zero with no increment
site, and then introduced two fields that were constant in the other direction.
Three occurrences is a guard, not a habit.
Verified end to end on a running unlicensed install: the payload that reported
audit_logging_persistent true under v6 now reports audit_reads_30d 0, and
seeding two in-window reads, one outside the window, and one with an invalid
activity class yields 2.
Six of the eight Pro-exclusive features had no telemetry field at all, so
there was no way to answer whether RBAC, audit logging, scheduled reporting,
agent profiles, alert-triggered AI, or Kubernetes AI were being used by the
installs paying for them. Schema v6 adds nine content-free adoption signals:
alert_ai_enabled AIConfig.IsAlertTriggeredAnalysisEnabled()
rbac_custom_roles non-built-in roles, per org
rbac_user_assignments user-to-role assignments, per org
audit_logging_persistent a persistent audit store is active, not console
audit_events_30d audit events retained inside the window
report_schedules configured scheduled reports
report_schedules_enabled scheduled reports switched on
report_schedules_run_30d schedules whose last run falls inside the window
agent_profiles configured agent profiles
Counts only. Role names, permissions, usernames, schedule names, delivery
recipients, report scope, profile names, and every audit event field stay on
the install. kubernetes_ai needs no field of its own: it is derivable at read
time from alert_ai_enabled combined with the existing kubernetes_clusters
count, and a dedicated field would be redundant.
Config-sourced signals are read through applyLicensedFeatureConfigSnapshot;
RBAC and audit live behind the router and are read through
Router.ApplyLicensedFeatureTelemetrySnapshot. The RBAC read goes through a new
TenantRBACProvider.PeekManager so a background telemetry read can never
provision an RBAC store for an org that has never used RBAC.
Also removes pulse_intelligence_patrol_autofixes_30d and the AutoFixCount
field behind it. patrol_run.go hardcoded AutoFixCount to 0 and no increment
site existed anywhere in the tree, so the counter was zero in all 233,364
retained production pings. That was a wiring bug, not evidence that nobody
uses Patrol fixes; governed fixes are delivered through the approved-action
pipeline, which is already instrumented. The field was plumbed through run
records, history persistence, the Assistant handoff, and telemetry while being
structurally incapable of holding a non-zero value.
Verified end to end against a running install rather than only in unit tests,
which is precisely the check the autofix counter never had: seeding three
report schedules (two enabled, one last run inside the window) and two agent
profiles produced report_schedules 3, report_schedules_enabled 2,
report_schedules_run_30d 1, agent_profiles 2 in the Settings telemetry
preview, and signing in moved audit_events_30d to 1.
The private receiver landed first in pulse-pro 78ff7dd so the new fields are
accepted on arrival.
Pulse under a systemd MemoryMax, docker --memory, or Kubernetes limit ran
with a GC that was blind to the cap: GOGC=100 lets total heap float to
roughly twice the live set, so a capped process grows until the kernel
OOM-kills it instead of collecting harder near the boundary. The demo
droplet's kill history (three OOM kills in July under an 800M cap) is
this mechanism, and any containerized install with a memory limit is
exposed the same way.
At startup, when GOMEMLIMIT is not set by the operator, resolve the
process cgroup (v2 ancestor walk taking the smallest memory.max on the
path, v1 limit_in_bytes fallback) and set the runtime soft limit to 90%
of it, leaving headroom for stacks, mmap, and CGO. Best effort: no
detectable limit leaves GC defaults untouched.
Contract-Neutral: runtime GC limit alignment: no payload or contract delta
Add availability_probe_targets and availability_probe_agents to the
telemetry ping - counts only, no agent names or addresses - with the
disclosure table updated in both privacy doc copies. Document the
feature in the availability-checks configuration guide and the
unified agent guide, including the ICMP capability caveat for
containerized probes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 30d window said 18 of 26 approved action attempts failed but
carried zero cause information. Every approved attempt that is not a
verified success now lands in exactly one content-free bucket:
- pre_dispatch: terminally refused before dispatch (plan drift, expiry,
emergency stop, policy authorization)
- execution: dispatched execution failed or ended inconclusive
- unverified: execution succeeded but outcome verification was not
confirmed (reads as success in the UI, counted as failure here)
- stuck_executing: still executing over an hour after dispatch
plus one sanitized machine reason code for the most recent failure.
Successes, failure buckets, and recent in-flight attempts partition the
attempt count, so the next window explains its own gap.
RefuseActionExecution now persists the specific refusal code
(plan_drift, action_plan_expired, ...) as the canonical execution
reason code instead of the generic pre_dispatch_refused, so audit truth
and telemetry distinguish refusal causes without message parsing.
Privacy disclosures updated in both PRIVACY.md copies; the adoption
report script surfaces the new counters.
Contract-Neutral: additive content-free telemetry counters + specific pre-dispatch refusal reason codes; privacy disclosures updated in-commit per security-privacy extension point
Refs #1510
Refs #1501
Refs #1507
Refs #1442
- persist scoped workloads status filters across platform navigation
- derive host memory pressure from available memory
- reapply system settings after every monitor reload path
- bound PBS backup snapshot polling workers during large backup scans
When a monitor reload was triggered by node auto-registration, the
reloadFunc in server.go recreated the monitor (and its notification
manager) but never re-applied system settings. The new notification
manager started with an empty webhook private CIDR allowlist, causing
webhook notifications to private IPs to fail until the allowlist was
manually re-saved in Settings.
Fix: call router.ReloadSystemSettings() at the end of reloadFunc,
after the new monitor references are set. This re-applies all
persisted system settings — including the webhook CIDR allowlist —
to the freshly created notification manager.
Refs #1507
Manifest-backed MCP tools, prompts, and resources with surface affordance contracts; agent capability manifest and governance projection; API contract tests and capability route projection; operations-loop and intelligence-funnel telemetry; release-control subsystem documentation, registry, and tooling; licensing and configuration.
golangci-lint run ./... failed on ~190 pre-existing errcheck violations and
5 unformatted files, burying any new regression in noise. Fix all of them:
- Test files that hand-rolled mock-mode set/restore (vmware, truenas, and
friends) now use the canonical setMockModeForTest/testutil.SetMockMode
helper instead of drift copies that ignored SetEnabled errors.
- internal/mock and internal/monitoring tests get package-local
mustSetEnabled/mustSetMockEnabled/mustSetMonitorMockMode helpers that
fail the test on toggle errors.
- pkg/auth/sqlite_manager.go, pkg/metrics/store.go, pkg/server/server.go:
rollbacks in defers use the explicit-discard idiom, migration renames and
rollup commits log failures, the hosted reaper goroutine logs an error
exit, shutdown mock-disable logs failures.
- Remaining test sites check errors with t.Fatalf/t.Errorf or explicitly
discard best-effort calls (restore-chmods, handler-closure unmarshals)
per existing repo style.
- gofmt: internal/api/maintenance_verification.go, internal/ai/demo.go and
three findings test files.
Only dupl findings remain (44 pre-existing production-code duplication
pairs) — those need real refactors, not mechanical fixes.
Full test suites pass for every touched package.
Operators can set PULSE_AGENT_INGEST_PORT to serve agent report and
management traffic (/api/agents/*) on a second listener, so that surface
can be placed on its own network or firewall boundary without exposing
the web UI or the rest of the REST API on that port. The dedicated port
serves only the /api/agents/* prefix and 404s everything else.
The option is additive and fail-closed: it is disabled at 0, the main
listener keeps serving agent ingest so existing single-port deployments
and agents are unaffected, and validation rejects out-of-range ports or
collisions with the frontend or HTTP redirect ports.
Retire runtime/API/UI monitored-system volume enforcement now that infrastructure monitoring is no longer capped.
Keep only legacy metadata scrubbing and purchase-start compatibility for old max_monitored_systems references.
Rename the remaining preview surface to monitored-system impact and make previews explanatory rather than save-blocking.
Update subsystem contracts and RA7 evidence for the caps-retired invariant.
rcourtman