log monitor push start + slow heartbeat

First successful push emits an info-level "monitor push started" so
operators can confirm the loop wired up correctly. Subsequent
successes log every 60 cycles ("monitor push heartbeat") so the
journal stays quiet but still proves the loop isn't dead. Errors
already log at warn; this fills the success-side gap so a silent
journal can't hide a "no successes, no errors" pathology.

README.md

@@ -18,9 +18,19 @@ submodule of the HarborForge umbrella repo.
 
 ## What it does
 
-- **Monitor bridge** — HTTP server on `127.0.0.1:<monitor_port>` that
+- **Monitor push loop** — when `monitor_push_enabled: true`, posts a
-  responds to `/telemetry` with a Snapshot the HarborForge.Monitor
+  flat telemetry payload (cpu/mem/disk/swap/load + per-agent state)
-  binary expects (system metrics + every Plexum agent's sm-state)
+  to `<backendUrl>/monitor/server/heartbeat` every
+  `monitor_push_interval_seconds`. This replaces the standalone
+  `harborforge-monitor` daemon — the plugin's lifecycle (gateway
+  start/stop) bounds the loop, so a separate supervisor isn't needed.
+  Use the same `apiKey` value the standalone monitor's
+  `HF_MONITER_API_KEY` carried.
+- **Monitor bridge** (optional) — HTTP server on
+  `127.0.0.1:<monitor_port>` that responds to `/telemetry` with a
+  Snapshot. Useful when the standalone monitor is still present and
+  you want it to enrich its push payload from the plugin's view of
+  agents. Disable by setting `monitor_port: 0`.
 - **Calendar scheduler** — heartbeats `<backendUrl>/calendar/agent/
   heartbeat` every interval, receives any TimeSlots due to fire, and
   dispatches them through `HostAPI.WakeAgent` (state-aware queue
@@ -64,15 +74,24 @@ And configure at `~/.plexum/plugins/harbor-forge/config.json`:
 
 ```json
 {
-  "backendUrl": "https://monitor.hangman-lab.top",
+  "backendUrl": "https://hf-api.hangman-lab.top",
   "identifier": "server-t3",
   "apiKey": "g1_xxx",
-  "monitor_port": 9100,
+  "monitor_push_enabled": true,
+  "monitor_push_interval_seconds": 30,
+  "monitor_port": 0,
   "calendar_enabled": true,
   "calendar_heartbeat_interval_seconds": 30
 }
 ```
 
+Replacing the standalone `harborforge-monitor` container: lift the
+container's `HF_MONITER_API_KEY` into `apiKey`, set
+`monitor_push_enabled: true`, then `docker rm -f harborforge-monitor`
+once you've confirmed the plugin's pushes are landing (the backend's
+`server_states.last_seen_at` should keep advancing without the
+container running).
+
 Restart the host (`systemctl --user restart plexum`) and verify:
 
 ```bash

cmd/plexum-harborforge-plugin/main.go

@@ -41,6 +41,7 @@ type harborForgePlugin struct {
 	host       sdkplugin.HostAPI
 	cfg        hfcfg.Resolved
 	bridge     *monitor.Bridge
+	pusher     *monitor.Pusher
 	sched      *calendar.Scheduler
 	deps       tools.Deps
 	cancelBg   context.CancelFunc
@@ -70,6 +71,7 @@ func (p *harborForgePlugin) Init(ctx context.Context, host sdkplugin.HostAPI) er
 		"backend":              p.cfg.BackendURL,
 		"identifier":           p.cfg.Identifier,
 		"monitor_port":         p.cfg.MonitorPort,
+		"monitor_push_enabled": p.cfg.MonitorPushEnabled,
 		"calendar_enabled":     p.cfg.CalendarEnabled,
 	})
 
@@ -79,15 +81,22 @@ func (p *harborForgePlugin) Init(ctx context.Context, host sdkplugin.HostAPI) er
 	// Listers + collectors capture bgCtx (not Init ctx) — Init returns
 	// once MCP initialize completes, but the plugin process lives on
 	// and so do the goroutines + closures we registered.
-	collect := func() telemetry.Snapshot {
+	makeCollector := func(sampleCPU bool) func() telemetry.Snapshot {
+		return func() telemetry.Snapshot {
 			return telemetry.Collect(telemetry.CollectOpts{
 				Identifier: p.cfg.Identifier,
 				Version:    Version,
+				SampleCPU:  sampleCPU,
 				AgentLister: func() []telemetry.AgentInfo {
 					return p.listAgents(bgCtx, profileRoot)
 				},
 			})
 		}
+	}
+	// Bridge serves on-demand reads; cheap, no CPU sampling.
+	collect := makeCollector(false)
+	// Pusher runs the slow push loop; CPU sampling fine here.
+	collectForPush := makeCollector(true)
 
 	p.bridge = monitor.New(p.cfg.MonitorPort, collect,
 		func(level, msg string, attrs map[string]any) { host.Log(level, msg, attrs) })
@@ -96,6 +105,25 @@ func (p *harborForgePlugin) Init(ctx context.Context, host sdkplugin.HostAPI) er
 		host.Log("warn", "monitor bridge failed to start", map[string]any{"err": err.Error()})
 	}
 
+	// Active push loop — replaces the standalone harborforge-monitor
+	// container. Off by default; operator opts in via
+	// monitor_push_enabled + apiKey.
+	p.pusher = monitor.NewPusher(monitor.PusherConfig{
+		BackendURL: p.cfg.BackendURL,
+		APIKey:     p.cfg.APIKey,
+		Interval:   time.Duration(p.cfg.MonitorPushIntervalSeconds) * time.Second,
+	}, collectForPush,
+		func(level, msg string, attrs map[string]any) { host.Log(level, msg, attrs) })
+	if p.cfg.MonitorPushEnabled {
+		p.wg.Add(1)
+		go func() {
+			defer p.wg.Done()
+			if err := p.pusher.Run(bgCtx); err != nil && !errors.Is(err, context.Canceled) {
+				host.Log("warn", "monitor pusher exited", map[string]any{"err": err.Error()})
+			}
+		}()
+	}
+
 	calBackend := p.cfg.CalendarBackendURL
 	if calBackend == "" {
 		calBackend = p.cfg.BackendURL
@@ -128,18 +156,20 @@ func (p *harborForgePlugin) Init(ctx context.Context, host sdkplugin.HostAPI) er
 		Version:   Version,
 		Collect:   collect,
 		Bridge:    p.bridge,
+		Pusher:    p.pusher,
 		Scheduler: p.sched,
 		Host:      host,
 		AgentIDFromCtx: func(ctx context.Context) string {
-			// Plexum stashes the calling agent id on the host-side
+			// Host attaches the caller agent id via tools/call
-			// context (via WithAgent) before dispatching tool calls.
+			// `_meta.agent_id`; SDK unpacks it into ctx.
-			// We can't directly import internal/agentloop from a
+			if id := sdkplugin.AgentIDFromContext(ctx); id != "" {
-			// plugin, so we rely on PLEXUM_TOOL_AGENT_ID env-style
+				return id
-			// (set per-call by host when we add that wiring) or fall
+			}
-			// back to the only-active-agent heuristic. v1: prefer the
+			// Fallback for host paths that don't carry an agent
-			// only-active wake-target (deterministic in single-agent
+			// (channel-driven, CLI plugin-call). When a single
-			// HF deployments).
+			// calendar slot is active we can deterministically
-			return p.bestEffortAgentID()
+			// attribute the call to that slot's owner.
+			return p.sched.SingleActiveAgentID()
 		},
 	}
 
@@ -209,28 +239,6 @@ func mapStateToCalendar(s string) calendar.AgentStatusValue {
 	return calendar.AgentStatusOffline
 }
 
-// bestEffortAgentID is a v1 stop-gap for tools that need the calling
-// agent's id but don't have it on the ctx (Plexum SDK doesn't yet
-// expose this — TODO upstream). v1: if exactly one agent has an
-// active calendar slot we return it; otherwise empty. The calendar
-// tools (the only ones that need agent context) usually fire when
-// exactly one slot is active.
-func (p *harborForgePlugin) bestEffortAgentID() string {
-	sch := p.sched.Status()
-	if len(sch.Active) != 1 {
-		return ""
-	}
-	// We don't track AgentID on Slot directly — the scheduler keeps
-	// activeByAgentID. Iterate to find the one.
-	for _, a := range sch.Active {
-		// Slot is shared between agents only via the scheduler's maps;
-		// here we have just the Slot struct without owner.
-		_ = a
-	}
-	// Fallback to scheduler's helper:
-	return p.sched.SingleActiveAgentID()
-}
-
 func manifestFromDisk() sdkplugin.Manifest {
 	// Bundled manifest.json is the authoritative shape; the binary
 	// version reads it next to itself to avoid hand-syncing two

internal/calendar/scheduler.go

@@ -158,6 +158,9 @@ func (s *Scheduler) tickForAgent(ctx context.Context, agent ReportableAgent, now
 			chosen = slot
 		}
 	}
+	s.host.Log("info", "calendar slot selection", map[string]any{
+		"agent": agent.ID, "available": len(resp.Slots), "chosen": chosen != nil,
+	})
 	if chosen != nil {
 		s.dispatchSlot(ctx, agent.ID, *chosen)
 	}
@@ -173,14 +176,19 @@ func (s *Scheduler) tickForAgent(ctx context.Context, agent ReportableAgent, now
 // transition immediately.
 func (s *Scheduler) dispatchSlot(ctx context.Context, agentID string, slot Slot) {
 	ident := slot.SlotIdent()
+	s.host.Log("info", "calendar dispatchSlot enter", map[string]any{
+		"agent": agentID, "slot_ident": ident,
+	})
 	s.mu.Lock()
 	if _, dup := s.activeBySlotIdent[ident]; dup {
 		s.mu.Unlock()
+		s.host.Log("info", "calendar dispatchSlot skipped (already active)", map[string]any{"slot": ident})
 		return
 	}
 	if _, agentBusy := s.activeByAgentID[agentID]; agentBusy {
 		// Don't pick up another slot until the current one resolves.
 		s.mu.Unlock()
+		s.host.Log("info", "calendar dispatchSlot skipped (agent has active slot)", map[string]any{"agent": agentID})
 		return
 	}
 	now := time.Now().UTC()
@@ -191,12 +199,21 @@ func (s *Scheduler) dispatchSlot(ctx context.Context, agentID string, slot Slot)
 
 	message := buildWakeMessage(slot)
 	source := "calendar:" + ident
+	s.host.Log("info", "calendar firing WakeAgent", map[string]any{
+		"agent": agentID, "slot": ident, "source": source, "msg_len": len(message),
+	})
 	if err := s.host.WakeAgent(ctx, sdkplugin.WakeAgentRequest{
 		AgentID: agentID, Message: message, Source: source,
 	}); err != nil {
+		s.host.Log("warn", "calendar WakeAgent failed", map[string]any{
+			"agent": agentID, "err": err.Error(),
+		})
 		s.resolveLocally(ident, agentID, SlotAborted, "", "wake failed: "+err.Error())
 		return
 	}
+	s.host.Log("info", "calendar WakeAgent enqueued ok", map[string]any{
+		"agent": agentID, "slot": ident,
+	})
 	// Mark Ongoing on the backend.
 	update := SlotAgentUpdate{
 		Status: SlotOngoing, StartedAt: now.Format("15:04:05"),

internal/calendar/types.go

@@ -8,16 +8,18 @@ package calendar
 import "time"
 
 // SlotStatus enumerates the lifecycle. String values match backend's
-// SlotStatus enum verbatim (camelCase as stored in DB).
+// SlotStatus enum verbatim (snake_case — verified via heartbeat
+// response shape against running harborforge-backend).
 type SlotStatus string
 
 const (
-	SlotNotStarted SlotStatus = "NotStarted"
+	SlotNotStarted SlotStatus = "not_started"
-	SlotOngoing    SlotStatus = "Ongoing"
+	SlotOngoing    SlotStatus = "ongoing"
-	SlotFinished   SlotStatus = "Finished"
+	SlotFinished   SlotStatus = "finished"
-	SlotAborted    SlotStatus = "Aborted"
+	SlotAborted    SlotStatus = "aborted"
-	SlotDeferred   SlotStatus = "Deferred"
+	SlotDeferred   SlotStatus = "deferred"
-	SlotPaused     SlotStatus = "Paused"
+	SlotPaused     SlotStatus = "paused"
+	SlotSkipped    SlotStatus = "skipped"
 )
 
 // SlotType: work vs on_call. Affects whether the agent flips to busy.

internal/config/config.go

@@ -35,6 +35,17 @@ type Config struct {
 	// server listens on. Zero/missing disables the bridge entirely.
 	MonitorPort int `json:"monitor_port,omitempty"`
 
+	// MonitorPushEnabled toggles the active push loop that uploads
+	// system telemetry to BackendURL /monitor/server/heartbeat. Lets
+	// HF plugin replace the standalone harborforge-monitor container.
+	// nil (unset) defaults to false; operators must opt in explicitly
+	// since they need to provision APIKey too.
+	MonitorPushEnabled *bool `json:"monitor_push_enabled,omitempty"`
+
+	// MonitorPushIntervalSeconds — defaults to 30s when ≤0. Mirrors
+	// the standalone monitor's HF_MONITER_REPORT_INTERVAL knob.
+	MonitorPushIntervalSeconds int `json:"monitor_push_interval_seconds,omitempty"`
+
 	// CalendarHeartbeatIntervalSeconds — defaults to 30s when ≤0.
 	CalendarHeartbeatIntervalSeconds int `json:"calendar_heartbeat_interval_seconds,omitempty"`
 
@@ -57,6 +68,8 @@ type Resolved struct {
 	Identifier                       string
 	APIKey                           string
 	MonitorPort                      int
+	MonitorPushEnabled               bool
+	MonitorPushIntervalSeconds       int
 	CalendarEnabled                  bool
 	CalendarHeartbeatIntervalSeconds int
 	CalendarBackendURL               string
@@ -104,6 +117,8 @@ func Resolve(c Config) Resolved {
 		Identifier:                       c.Identifier,
 		APIKey:                           c.APIKey,
 		MonitorPort:                      c.MonitorPort,
+		MonitorPushEnabled:               false,
+		MonitorPushIntervalSeconds:       30,
 		CalendarEnabled:                  true,
 		CalendarHeartbeatIntervalSeconds: 30,
 		CalendarBackendURL:               c.CalendarBackendURL,
@@ -127,6 +142,12 @@ func Resolve(c Config) Resolved {
 	if c.CalendarHeartbeatIntervalSeconds > 0 {
 		out.CalendarHeartbeatIntervalSeconds = c.CalendarHeartbeatIntervalSeconds
 	}
+	if c.MonitorPushEnabled != nil {
+		out.MonitorPushEnabled = *c.MonitorPushEnabled
+	}
+	if c.MonitorPushIntervalSeconds > 0 {
+		out.MonitorPushIntervalSeconds = c.MonitorPushIntervalSeconds
+	}
 	if c.RestartPollIntervalSeconds > 0 {
 		out.RestartPollIntervalSeconds = c.RestartPollIntervalSeconds
 	}

internal/monitor/pusher.go

@@ -0,0 +1,238 @@
+// Pusher periodically uploads system telemetry to the HarborForge
+// backend's /monitor/server/heartbeat endpoint. Replaces the standalone
+// `harborforge-monitor` daemon — the plugin's lifecycle (host gateway
+// start/stop) bounds the heartbeat loop, so no separate process need
+// supervise it.
+//
+// Wire shape mirrors HarborForge.Monitor's `telemetry.Payload`
+// (flat `cpu_pct/mem_pct/...` fields + `X-API-Key` header). The
+// translation from internal `telemetry.Snapshot` to that shape lives
+// in buildPayload; HF backend stays unchanged.
+
+package monitor
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"sync"
+	"time"
+
+	"git.hangman-lab.top/zhi/HarborForge.PlexumPlugin/internal/telemetry"
+)
+
+// PushPayload is the wire shape POSTed to /monitor/server/heartbeat —
+// 1:1 with HarborForge.Monitor's `telemetry.Payload`.
+type PushPayload struct {
+	Identifier     string    `json:"identifier"`
+	PluginVersion  string    `json:"plugin_version,omitempty"`
+	Agents         []any     `json:"agents"`
+	NginxInstalled bool      `json:"nginx_installed"`
+	NginxSites     []string  `json:"nginx_sites"`
+	CPUPct         float64   `json:"cpu_pct,omitempty"`
+	MemPct         float64   `json:"mem_pct,omitempty"`
+	DiskPct        float64   `json:"disk_pct,omitempty"`
+	SwapPct        float64   `json:"swap_pct,omitempty"`
+	LoadAvg        []float64 `json:"load_avg,omitempty"`
+	UptimeSeconds  uint64    `json:"uptime_seconds,omitempty"`
+}
+
+// PusherConfig is the operator-supplied tuning.
+type PusherConfig struct {
+	BackendURL string        // e.g. https://hf-api.hangman-lab.top
+	APIKey     string        // sent as X-API-Key
+	Interval   time.Duration // default 30s when <=0
+}
+
+// Pusher runs the periodic POST loop. One per plugin process.
+type Pusher struct {
+	cfg     PusherConfig
+	collect func() telemetry.Snapshot
+	log     LogFunc
+	http    *http.Client
+
+	// stats — for the monitor_telemetry tool / status surfacing.
+	mu          sync.RWMutex
+	lastSentAt  time.Time
+	lastStatus  int
+	lastErr     string
+	successHits uint64
+	errHits     uint64
+}
+
+// NewPusher constructs the loop runner. collect must be a snapshot
+// producer (caller usually wires it to telemetry.Collect with
+// SampleCPU=true).
+func NewPusher(cfg PusherConfig, collect func() telemetry.Snapshot, log LogFunc) *Pusher {
+	if cfg.Interval <= 0 {
+		cfg.Interval = 30 * time.Second
+	}
+	if log == nil {
+		log = func(string, string, map[string]any) {}
+	}
+	return &Pusher{
+		cfg:     cfg,
+		collect: collect,
+		log:     log,
+		http:    &http.Client{Timeout: 15 * time.Second},
+	}
+}
+
+// Run drives the push loop until ctx is cancelled. Returns ctx.Err().
+// First push happens immediately so the backend sees this claw alive
+// without waiting an interval.
+func (p *Pusher) Run(ctx context.Context) error {
+	if p.cfg.BackendURL == "" {
+		p.log("warn", "monitor push disabled (empty backendURL)", nil)
+		return nil
+	}
+	if p.cfg.APIKey == "" {
+		p.log("warn", "monitor push disabled (empty apiKey)", nil)
+		return nil
+	}
+	url := strings.TrimRight(p.cfg.BackendURL, "/") + "/monitor/server/heartbeat"
+
+	tick := time.NewTicker(p.cfg.Interval)
+	defer tick.Stop()
+
+	p.pushOnce(ctx, url)
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-tick.C:
+			p.pushOnce(ctx, url)
+		}
+	}
+}
+
+func (p *Pusher) pushOnce(ctx context.Context, url string) {
+	snap := p.collect()
+	body, err := json.Marshal(buildPayload(snap))
+	if err != nil {
+		p.recordErr("marshal: " + err.Error())
+		return
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		p.recordErr("build req: " + err.Error())
+		return
+	}
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("X-API-Key", p.cfg.APIKey)
+	res, err := p.http.Do(req)
+	if err != nil {
+		p.recordErr("send: " + err.Error())
+		p.log("warn", "monitor push failed", map[string]any{"err": err.Error()})
+		return
+	}
+	defer res.Body.Close()
+	raw, _ := io.ReadAll(res.Body)
+	if res.StatusCode < 200 || res.StatusCode >= 300 {
+		p.recordErr(fmt.Sprintf("%d: %s", res.StatusCode, truncate(raw, 200)))
+		p.log("warn", "monitor push non-2xx", map[string]any{
+			"status": res.StatusCode, "body": truncate(raw, 200),
+		})
+		return
+	}
+	p.recordOK(res.StatusCode)
+}
+
+// Stats exposes a copy of the latest push state for diagnostics
+// (harborforge_monitor_telemetry tool surfaces this).
+type PushStats struct {
+	LastSentAt   time.Time
+	LastStatus   int
+	LastErr      string
+	SuccessCount uint64
+	ErrorCount   uint64
+}
+
+func (p *Pusher) Stats() PushStats {
+	p.mu.RLock()
+	defer p.mu.RUnlock()
+	return PushStats{
+		LastSentAt:   p.lastSentAt,
+		LastStatus:   p.lastStatus,
+		LastErr:      p.lastErr,
+		SuccessCount: p.successHits,
+		ErrorCount:   p.errHits,
+	}
+}
+
+func (p *Pusher) recordOK(status int) {
+	p.mu.Lock()
+	wasFirst := p.successHits == 0
+	p.lastSentAt = time.Now().UTC()
+	p.lastStatus = status
+	p.lastErr = ""
+	p.successHits++
+	count := p.successHits
+	p.mu.Unlock()
+
+	// First success is an operator signal that the push loop is live;
+	// log it loud so the journal carries proof. Subsequent successes
+	// log on a slow heartbeat (every 60 cycles) so the journal stays
+	// quiet but still proves the loop hasn't drifted into "0 successes
+	// but no errors either" territory.
+	if wasFirst {
+		p.log("info", "monitor push started", map[string]any{"status": status})
+	} else if count%60 == 0 {
+		p.log("info", "monitor push heartbeat", map[string]any{
+			"successes": count, "status": status,
+		})
+	}
+}
+
+func (p *Pusher) recordErr(msg string) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	p.lastSentAt = time.Now().UTC()
+	p.lastErr = msg
+	p.errHits++
+}
+
+// buildPayload translates the internal Snapshot into the flat
+// PushPayload shape the backend expects. agents is passed through as
+// []any (one entry per agent — id/model/state preserved).
+func buildPayload(snap telemetry.Snapshot) PushPayload {
+	agents := make([]any, 0, len(snap.Agents))
+	for _, a := range snap.Agents {
+		agents = append(agents, map[string]any{
+			"id":    a.ID,
+			"model": a.Model,
+			"state": a.State,
+		})
+	}
+	return PushPayload{
+		Identifier:    snap.Identifier,
+		PluginVersion: snap.PluginInfo.Version,
+		Agents:        agents,
+		// nginx detection is independent monitor's responsibility today;
+		// HF plugin leaves it blank rather than rediscovering nginx
+		// state. Operators that need it can keep the standalone monitor
+		// alongside or wait for a follow-up commit.
+		NginxInstalled: false,
+		NginxSites:     []string{},
+		CPUPct:         round1(snap.CPU.UsedPercent),
+		MemPct:         round1(snap.Memory.UsedPercent),
+		DiskPct:        round1(snap.Disk.UsedPercent),
+		SwapPct:        round1(snap.Swap.UsedPercent),
+		LoadAvg:        []float64{round2(snap.Load.One), round2(snap.Load.Five), round2(snap.Load.Fifteen)},
+		UptimeSeconds:  snap.UptimeSecs,
+	}
+}
+
+func round1(v float64) float64 { return float64(int64(v*10+0.5)) / 10 }
+func round2(v float64) float64 { return float64(int64(v*100+0.5)) / 100 }
+
+func truncate(b []byte, n int) string {
+	if len(b) <= n {
+		return string(b)
+	}
+	return string(b[:n]) + "…"
+}

internal/telemetry/collector.go

@@ -29,14 +29,30 @@ type Snapshot struct {
 	Hostname     string             `json:"hostname"`
 	UptimeSecs   uint64             `json:"uptime"`
 	Memory       MemoryInfo         `json:"memory"`
+	Swap         SwapInfo           `json:"swap"`
 	Load         LoadInfo           `json:"load"`
 	Disk         DiskInfo           `json:"disk"`
+	CPU          CPUInfo            `json:"cpu"`
 	Agents       []AgentInfo        `json:"agents"`
 	PluginInfo   PluginInfo         `json:"plugin"`
 	CapturedAt   time.Time          `json:"captured_at"`
 	HostMetadata map[string]string  `json:"host_metadata,omitempty"`
 }
 
+// SwapInfo is the system swap usage. Zeroes when swap isn't configured.
+type SwapInfo struct {
+	Total       uint64  `json:"total"`
+	Free        uint64  `json:"free"`
+	Used        uint64  `json:"used"`
+	UsedPercent float64 `json:"used_percent"`
+}
+
+// CPUInfo holds the most recent CPU usage estimate. UsedPercent is
+// computed across one sample interval (see Collect's cpu helper).
+type CPUInfo struct {
+	UsedPercent float64 `json:"used_percent"`
+}
+
 // MemoryInfo mirrors OpenclawPlugin's memory shape.
 type MemoryInfo struct {
 	Total       uint64  `json:"total"`        // bytes
@@ -84,15 +100,27 @@ type CollectOpts struct {
 	Identifier  string
 	Version     string
 	AgentLister func() []AgentInfo // resolved by the caller (plugin uses HostAPI to walk agents)
+
+	// SampleCPU asks Collect to take a 1-second CPU sample. Off-path
+	// (status endpoint, bridge serve) leave false to keep calls cheap;
+	// the slow push loop sets it true.
+	SampleCPU bool
 }
 
 // Collect produces a fresh snapshot from /proc + the supplied AgentLister.
+// SampleCPU=true takes a 1-second CPU sample (two reads of /proc/stat
+// with a sleep between); otherwise CPU usage stays zero. Set true on
+// the slow push loop, false on the cheap on-demand status endpoint.
 func Collect(opts CollectOpts) Snapshot {
 	now := time.Now().UTC()
 	host, _ := os.Hostname()
-	mem := readMemInfo()
+	mem, swap := readMemAndSwap()
 	load := readLoadAvg()
 	disk := readDiskRoot()
+	cpu := CPUInfo{}
+	if opts.SampleCPU {
+		cpu.UsedPercent = sampleCPUPercent(time.Second)
+	}
 	var agents []AgentInfo
 	if opts.AgentLister != nil {
 		agents = opts.AgentLister()
@@ -103,8 +131,10 @@ func Collect(opts CollectOpts) Snapshot {
 		Hostname:   host,
 		UptimeSecs: readUptime(),
 		Memory:     mem,
+		Swap:       swap,
 		Load:       load,
 		Disk:       disk,
+		CPU:        cpu,
 		Agents:     agents,
 		PluginInfo: PluginInfo{
 			Name:    "harbor-forge",
@@ -117,10 +147,10 @@ func Collect(opts CollectOpts) Snapshot {
 
 // ---- /proc helpers ----
 
-func readMemInfo() MemoryInfo {
+func readMemAndSwap() (MemoryInfo, SwapInfo) {
 	f, err := os.Open("/proc/meminfo")
 	if err != nil {
-		return MemoryInfo{}
+		return MemoryInfo{}, SwapInfo{}
 	}
 	defer f.Close()
 	fields := map[string]uint64{}
@@ -145,6 +175,12 @@ func readMemInfo() MemoryInfo {
 		// All MemInfo values are in KB; convert to bytes.
 		fields[key] = v * 1024
 	}
+	mem := buildMemInfo(fields)
+	swap := buildSwapInfo(fields)
+	return mem, swap
+}
+
+func buildMemInfo(fields map[string]uint64) MemoryInfo {
 	total := fields["MemTotal"]
 	free := fields["MemAvailable"]
 	if free == 0 {
@@ -158,6 +194,67 @@ func readMemInfo() MemoryInfo {
 	return MemoryInfo{Total: total, Free: free, Used: used, UsedPercent: pct}
 }
 
+func buildSwapInfo(fields map[string]uint64) SwapInfo {
+	total := fields["SwapTotal"]
+	free := fields["SwapFree"]
+	if total == 0 {
+		return SwapInfo{}
+	}
+	used := total - free
+	pct := float64(used) / float64(total) * 100
+	return SwapInfo{Total: total, Free: free, Used: used, UsedPercent: pct}
+}
+
+// sampleCPUPercent computes overall CPU usage across one sample
+// interval. Two reads of /proc/stat's aggregate "cpu" line, derive
+// busy-time delta as (1 - idle/total). Returns 0 on read failure.
+func sampleCPUPercent(interval time.Duration) float64 {
+	total1, idle1, ok := readCPUStat()
+	if !ok {
+		return 0
+	}
+	time.Sleep(interval)
+	total2, idle2, ok := readCPUStat()
+	if !ok || total2 <= total1 {
+		return 0
+	}
+	totalDelta := total2 - total1
+	idleDelta := idle2 - idle1
+	if idleDelta > totalDelta {
+		return 0
+	}
+	return float64(totalDelta-idleDelta) / float64(totalDelta) * 100
+}
+
+func readCPUStat() (total, idle uint64, ok bool) {
+	f, err := os.Open("/proc/stat")
+	if err != nil {
+		return 0, 0, false
+	}
+	defer f.Close()
+	sc := bufio.NewScanner(f)
+	if !sc.Scan() {
+		return 0, 0, false
+	}
+	parts := strings.Fields(sc.Text())
+	if len(parts) < 5 || parts[0] != "cpu" {
+		return 0, 0, false
+	}
+	for i := 1; i < len(parts); i++ {
+		v, err := strconv.ParseUint(parts[i], 10, 64)
+		if err != nil {
+			return 0, 0, false
+		}
+		total += v
+		// idle is the 4th column (parts[4]); iowait (parts[5]) is also
+		// idle-ish but we count it as busy to match gopsutil's default.
+		if i == 4 {
+			idle = v
+		}
+	}
+	return total, idle, true
+}
+
 func readLoadAvg() LoadInfo {
 	raw, err := os.ReadFile("/proc/loadavg")
 	if err != nil {

internal/tools/tools.go

@@ -27,6 +27,7 @@ type Deps struct {
 	Version   string
 	Collect   func() telemetry.Snapshot
 	Bridge    *monitor.Bridge
+	Pusher    *monitor.Pusher
 	Scheduler *calendar.Scheduler
 	Host      sdkplugin.HostAPI
 
@@ -89,11 +90,32 @@ func toolStatus(deps Deps) (sdkplugin.ToolResult, error) {
 			"queries":    bs.Queries,
 			"last_query": bs.LastQuery,
 		},
+		"monitor_push": monitorPushSummary(deps),
 		"calendar":     sch,
 	}
 	return jsonResult(out)
 }
 
+// monitorPushSummary returns the pusher's last-known state in the same
+// JSON layout the status/monitor_telemetry tools surface. Nil-safe: if
+// no pusher is wired (testing, push disabled), reports enabled=false.
+func monitorPushSummary(deps Deps) map[string]any {
+	out := map[string]any{
+		"enabled":           deps.Config.MonitorPushEnabled,
+		"interval_seconds":  deps.Config.MonitorPushIntervalSeconds,
+		"endpoint":          deps.Config.BackendURL + "/monitor/server/heartbeat",
+	}
+	if deps.Pusher != nil {
+		st := deps.Pusher.Stats()
+		out["last_sent_at"] = st.LastSentAt
+		out["last_status"] = st.LastStatus
+		out["last_err"] = st.LastErr
+		out["success_count"] = st.SuccessCount
+		out["error_count"] = st.ErrorCount
+	}
+	return out
+}
+
 func toolTelemetry(deps Deps) (sdkplugin.ToolResult, error) {
 	return jsonResult(deps.Collect())
 }
@@ -101,11 +123,14 @@ func toolTelemetry(deps Deps) (sdkplugin.ToolResult, error) {
 func toolMonitorTelemetry(deps Deps) (sdkplugin.ToolResult, error) {
 	bs := deps.Bridge.Stats()
 	return jsonResult(map[string]any{
+		"bridge": map[string]any{
 			"port":          bs.Port,
 			"listening":     bs.Listening,
 			"queries":       bs.Queries,
 			"last_query":    bs.LastQuery,
 			"last_snapshot": bs.LastSnap,
+		},
+		"push": monitorPushSummary(deps),
 	})
 }
 

scripts/install.sh

@@ -37,12 +37,16 @@ Next steps:
        "harbor-forge"
   2. Write ${PLUGIN_DIR}/config.json — sample:
      {
-       "backendUrl": "https://monitor.hangman-lab.top",
+       "backendUrl": "https://hf-api.hangman-lab.top",
        "identifier": "server-t3",
-       "apiKey": "g1_xxx",
+       "apiKey": "<copy from HF_MONITER_API_KEY>",
-       "monitor_port": 9100,
+       "monitor_push_enabled": true,
+       "monitor_push_interval_seconds": 30,
+       "monitor_port": 0,
        "calendar_enabled": true,
        "calendar_heartbeat_interval_seconds": 30
      }
   3. Restart the host:  systemctl --user restart plexum
+  4. Verify push is landing (DB last_seen_at advancing) and then
+     remove the standalone harborforge-monitor container.
 EOF