rcourtman
780732a82a
Stacked on #2743 (Patrol resolves finding references before reading operator state), whose commit this branch carries until it lands. Builds on #2721 (link-fold references, merged). ## Problem An operator link joins agent A to a live resource V, for example an agent inside a vSphere VM. After a restart, until A reports again, A exists only through saved-host continuity. Three consumers then disagreed about what a reference to A means: - **Monitor read state** (`GetUnifiedReadStateOrSnapshot`, used by Patrol): continuity ingest skipped the link pass, so A stayed its own offline row and references resolved to A. - **Resources API** (`QueryService.buildRegistry`): seeding from that listing re-applied the link and folded A into V, merging A's stale telemetry into V. Operator-state writes naming A landed on V. - **Alert intent** (`installOperatorIntentResolver`): resolved through the published registry, where A is absent, so `agent:<host ID>` stayed literal. So maintenance or `NeverAutoRemediate` set on V reached neither A's alerts nor its Patrol findings in that window, while the API said it applied. ## Change - **Registry holds instead of merging.** Rows continuity introduces carry an in-memory marker (`Resource.continuityOnly`, rides clones, cleared when an observation merges in). The link pass never merges a link with a saved member. When the saved row is the member the fold would have taken in, it is held beside the primary: every reference that resolves to it (canonical ID, `agent:<host ID>`, machine ID, an alias it shares with the primary) resolves to the primary, through `ResolveReferenceID`, `GetByReference`, the alias index and availability-check links. Both rows keep their own telemetry. - The continuity pass now runs the link pass too. The resources API's registry, seeded from the read state, holds the saved row the same way. - A saved **primary** leaves a live member its own identity: a live row never resolves to a saved one. - A hold whose primary a later link in the same pass folds follows the fold. A saved row held under a saved primary follows that primary's hold, and a cycle of saved rows resolves each to itself. - **Alert intent** resolves references and ancestors through the published registry with saved hosts overlaid, the same read state Patrol uses. This also lets intent set on an unlinked saved host reach its offline alert. - The resolver runs under the alert manager's lock once per metric check, so the overlay is kept per published registry generation (`MonitorAdapter.Generation`: the registry plus its last change time, since two publications can share a timestamp). - Publication builds it before alerts are evaluated. - A build that overlaps a mock-mode switch is discarded inside the mock-mode fence. - In mock mode, lookups use the published registry as before. ## What lookups by A's ID do - **Follow the link to V:** reference lookups (resource detail, agent resource context), operator-state reads and writes, Patrol's finding projection and alert intent. These match what a live fold already does. - **Stay on A:** exact reads (`Get`). The saved agent's drawer still shows its own facet bundle and history, and its operator-state section shows V's state. - **Unchanged:** history bindings, `SupersededCanonicalIDs`, and action planning, which uses exact `Get`. Continuity rows advertise no capabilities. ## Known limits - The read state does not re-resolve availability checks after continuity, so a check naming the saved agent stays as the published registry left it. The resources API replays checks and attaches it to V. This was the same before the change. - When the saved agent is the link's effective primary (it stays primary unless it was linked into its own guest), intent set under its ID does not reach the live member until it reports. - **Chains are out of scope.** Take an agent linked into a VM that is itself linked into another resource. The saved agent is not held, because the published registry has already folded the VM when continuity adds the agent. It answers for itself until it reports. Live chains already depend on the order their links are stored in. I tried making chains order-independent here and every precedence rule picked a worse identity somewhere, so that is a separate follow-up. - A lookup that races ahead of publication, or the first one after leaving mock mode, still builds the overlay under the alert lock. ## Testing - `internal/unifiedresources/resolve_test.go`: - `TestContinuityAgentLinkedIntoLiveGuestAnswersToTheGuest`: read state and resources-API registry, including a hostname both rows carry, then the fold once the agent reports. - `TestContinuityPrimaryLeavesLiveLinkMemberItsOwnIdentity`. - `TestObservationMergedIntoASavedRowLetsItsLinkFold`. - `TestLinkHoldsFollowTheirPrimaryThroughThePass`: later fold, saved primary, three-member cycle. - `internal/monitoring/issue1913_host_continuity_test.go` (`TestManualLinkToSupplementalGuestHoldsWithAndWithoutContinuity`): with no continuity, an unrelated absent machine, and the agent's own saved enrollment, the read state, alert intent and a registry seeded from the read state resolve the agent to the guest without merging it. - `internal/monitoring/monitor_host_agent_removal_lifecycle_test.go` (`TestHostAgentRemovalLifecycleHonorsSavedHostIntentAfterRestart`): after a restart, the monitor's own host-offline check raises no alert for a saved host the operator marked intentionally offline. On main it raises a critical host-offline alert, even with no link involved. - `internal/unifiedresources/clone_test.go` (`TestCloneResourceKeepsContinuityMarker`). - `internal/monitoring/monitor_alert_intent_test.go`: - `TestOperatorIntentReachesSavedHostAlertsAfterRestart`: an unlinked saved host's intentionally-offline state reaches its offline alert. - `TestOperatorIntentIdentityKeepsOneOverlayPerGeneration`: reuse within a generation, a publication sharing or preceding the last timestamp, the mock switch, and a rebuild from inside an admitted fence call. Passes 3× under `-race`. - `internal/api/resources_link_continuity_test.go` (`TestResourcesAPIAndPatrolAgreeOnContinuityAgentLinkedIntoGuest`): API write and Patrol read on the same store, with zero listing clones. - Negative controls on pulse-dev: - With the base `registry.go`, the registry, 1913, and API tests fail. The resources API folds the saved agent into the guest when the guest is the link primary, and folds the live member into the saved agent when the saved agent is the primary. - With the base `monitor_alert_intent.go` (plus `monitor_alert_sync.go` and `monitor.go`), the 1913 continuity case, `TestOperatorIntentReachesSavedHostAlertsAfterRestart` and `TestHostAgentRemovalLifecycleHonorsSavedHostIntentAfterRestart` fail. - Full `./internal/unifiedresources/...`, `./internal/monitoring/...`, `./internal/alerts/...` and `./internal/api/...` on pulse-dev pass on this change over #2721's earlier head. The one exception is `TestServerInfoEndpointReportsDevelopment`, which fails in a tree rsynced without `.git`. - On the final stack (main after #2721 and #2742, then #2743, then this commit), the targeted proofs plus #2739's unlink and report-merge tests in `resourceapi` pass locally. The Go tree is byte-identical to that verified stack. pulse-dev's disk was stalled on I/O at the time, so every SQLite-backed test hung there in `fsync`, on `origin/main` too.
Pulse
Real-time Proxmox, Docker and Kubernetes monitoring that finds what needs attention.
Pulse is a self-hosted monitoring workspace for Proxmox, Docker, Kubernetes, TrueNAS, physical and virtual machines, and early-access VMware vSphere environments. It combines live infrastructure state, history, alerts, recovery visibility, and scheduled health checks without requiring a conventional enterprise monitoring stack.
Why Pulse
- It watches between visits. Alerts and Pulse Patrol find failed backups, capacity pressure, restart loops, unhealthy containers, clock drift, and other problems that dashboards cannot surface when nobody is looking.
- It keeps each platform familiar. Proxmox, Docker, Kubernetes, TrueNAS, vSphere, and machines have dedicated views, backed by one shared resource model for search, alerts, history, and investigation.
- It stays operator-controlled. Credentials are encrypted at rest, API tokens are scoped, agent commands are disabled by default, and governed fixes require the configured policy and approval path.
Platform coverage
| Platform | Coverage |
|---|---|
| Proxmox VE, PBS, and PMG | Nodes, guests, storage, backups, replication, Ceph, mail gateways, and alerts |
| Docker and Podman | Hosts, containers, Compose projects, Swarm services, health, images, and updates |
| Kubernetes | Clusters, nodes, workloads, pods, services, storage, and events through the unified agent |
| TrueNAS SCALE and CORE | Pools, datasets, disks, snapshots, replication tasks, apps, VMs, and alerts |
| Linux, Windows, and macOS machines | Host health, filesystems, networking, temperatures, RAID, and availability through the unified agent |
| VMware vSphere | Early-access inventory, hosts, clusters, VMs, datastores, networks, snapshots, and recovery context; validate against your own vCenter before production use |
Platform pages keep storage and recovery information beside the infrastructure it belongs to. Alerts, Actions, and Patrol remain cross-platform views.
Patrol: monitoring that does rounds
Pulse Patrol runs scheduled checks across the current state and recent history of your infrastructure. Community installations can use a local model or their own AI provider for watch-only analysis. Pulse Pro adds investigation and policy-bound fixes with approval, verification, and an audit trail.
Pulse also includes an interactive Assistant and an MCP adapter for external clients such as Claude Code and OpenCode. Both sit on top of the same scoped inventory, metrics, alert, storage, and governed-action contracts.
Quick start
Choose an exact version from the latest release and keep that version pinned during installation.
Docker
docker run -d \
--name pulse \
-p 7655:7655 \
-v pulse_data:/data \
-e PULSE_DEPLOYMENT_METHOD=docker_run \
--restart unless-stopped \
rcourtman/pulse:vX.Y.Z
Open http://<your-ip>:7655 and follow the bootstrap-token setup. Docker host
monitoring is provided by the unified agent; the Pulse server container does
not need the Docker socket.
Proxmox LXC, Linux, and Kubernetes
The installer is signed. Verify install.sh against the pinned
pulse-installer key before running it:
export PULSE_VERSION=vX.Y.Z
curl -fsSLO "https://github.com/rcourtman/Pulse/releases/download/${PULSE_VERSION}/install.sh" &&
curl -fsSLO "https://github.com/rcourtman/Pulse/releases/download/${PULSE_VERSION}/install.sh.sshsig" &&
ssh-keygen -Y verify \
-f <(printf '%s\n' 'pulse-installer namespaces="pulse-install" ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIMZd/DaH+BldzOkq1A8KVTcFk73nAyrE8aJOyf7i00jm pulse-installer') \
-I pulse-installer \
-n pulse-install \
-s install.sh.sshsig < install.sh &&
bash install.sh --version "${PULSE_VERSION}"
Paste the whole block: installation stops if either download or signature verification fails. Do not run the final command separately to bypass a failure.
The GitHub installer installs the Pulse server. Install and upgrade agents (including v5-to-v6 agent upgrades) with the per-host command generated under Settings → Infrastructure → Install on a host.
Do you need an agent?
Often not. Proxmox VE, PBS, and PMG are monitored through the Proxmox API with a read-only token, so nothing runs on the host and the generated setup script creates a privilege-separated monitoring user. Install the unified agent only where you want data the platform API cannot provide, such as host SMART health, temperatures, Docker hosts, or standalone machines.
The agent is operator-controlled by design. Command execution is off by
default, the local listener binds to localhost, and generated systemd units
ship hardened. On Linux, the installer's --least-privilege profile runs the
service as a dedicated non-root user, with optional scoped sudo grants for
the few collectors that need elevation. The
agent security model documents exactly what the
agent can and cannot do at each privilege level.
Important
GitHub release assets and
rcourtman/pulseimages are Community builds. Relay, Pro, and eligible legacy customers should use the private image or Linux archive provided by the Pulse download portal. Replacing a private Pro runtime with a public Community build removes its private runtime hooks.
Editions
- Community — self-hosted monitoring, seven days of metric history, core SSO, update alerts, and Patrol with your own provider or local model.
- Pro — Community plus Patrol investigation, governed fixes, ninety days of history, centralized agent profiles, RBAC, audit logging, and reporting.
- MSP — for managed service providers: one Pulse Account running many client workspaces, each with an isolated Pulse runtime — separate dashboards, alerts, users, audit history, and reports. Free sixty-day two-client evaluation at Pulse for MSPs.
Relay is no longer sold. Existing Relay subscribers receive Pro features at their current price for as long as their subscription continues. Existing paired phones keep working until 31 March 2027. Relay connects the app, not the web UI. Use your own VPN or tunnel for remote web access. For phone alerts, add an ntfy, Gotify or Pushover destination and open Pulse in your phone's browser.
Core self-hosted monitoring is not gated by monitored-system or child-resource volume. See the runtime-aligned capability reference and current plans for details.
Documentation
- Install and deployment
- Production deployment and security
- Upgrade from Pulse v5
- Releases and update channels
- Configuration
- Platform and agent guides
- Pulse Intelligence
- Security and privacy
- Code signing policy
- Troubleshooting
- API reference and architecture
Localized getting started guides: Deutsch · Español
Development
Pulse uses Go 1.26 and a SolidJS/TypeScript frontend. The managed development
runtime starts the frontend at http://127.0.0.1:5173 and proxies API and
WebSocket traffic to the backend on port 7655.
npm ci
npm --prefix frontend-modern ci
npm run dev
Useful checks:
go test ./...
npm --prefix frontend-modern test
npm --prefix frontend-modern run type-check
python3 scripts/check_public_docs.py
See CONTRIBUTING.md before investing in a code change. Pulse uses an issue-first contribution process and does not normally accept unsolicited pull requests.
Community and support
- Ask questions in GitHub Discussions.
- Report reproducible bugs through GitHub Issues.
- Home Assistant users can use the community-maintained Pulse add-ons.
- If Pulse is useful to you, support its development through GitHub Sponsors or Ko-fi.
License
Everything in this repository is licensed under the MIT License. There is no dual licensing and no commercial or source-available code here.
Pulse Pro is a separate commercial product, built from private sources and distributed through pulserelay.pro. Its use is governed by the Terms of Service, which apply to that product only, not to anything in this repository.