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
@@ -66,46 +67,76 @@ func (p *harborForgePlugin) Init(ctx context.Context, host sdkplugin.HostAPI) er
 	}
 	p.cfg = hfcfg.Resolve(raw)
 	host.Log("info", "harbor-forge plugin initialized", map[string]any{
-		"version":           Version,
+		"version":              Version,
-		"backend":           p.cfg.BackendURL,
+		"backend":              p.cfg.BackendURL,
-		"identifier":        p.cfg.Identifier,
+		"identifier":           p.cfg.Identifier,
-		"monitor_port":      p.cfg.MonitorPort,
+		"monitor_port":         p.cfg.MonitorPort,
-		"calendar_enabled":  p.cfg.CalendarEnabled,
+		"monitor_push_enabled": p.cfg.MonitorPushEnabled,
+		"calendar_enabled":     p.cfg.CalendarEnabled,
 	})
 
-	collect := func() telemetry.Snapshot {
-		return telemetry.Collect(telemetry.CollectOpts{
-			Identifier: p.cfg.Identifier,
-			Version:    Version,
-			AgentLister: func() []telemetry.AgentInfo {
-				return p.listAgents(ctx, profileRoot)
-			},
-		})
-	}
-
-	p.bridge = monitor.New(p.cfg.MonitorPort, collect,
-		func(level, msg string, attrs map[string]any) { host.Log(level, msg, attrs) })
-
 	bgCtx, cancel := context.WithCancel(context.Background())
 	p.cancelBg = cancel
 
+	// 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.
+	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) })
+
 	if err := p.bridge.Start(bgCtx); err != nil {
 		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
 	}
-	bridge := calendar.New(calBackend, p.cfg.APIKey)
+	bridge := calendar.New(calBackend, p.cfg.Identifier)
 	p.sched = calendar.NewScheduler(
 		calendar.Config{
 			HeartbeatInterval: time.Duration(p.cfg.CalendarHeartbeatIntervalSeconds) * time.Second,
 		},
-		bridge, host, p.cfg.Identifier,
+		bridge, host,
 		calendar.PluginInfoTag{Name: "harbor-forge", Version: Version, Backend: "plexum"},
 		func() []calendar.ReportableAgent {
-			return p.listReportableAgents(ctx, profileRoot)
+			return p.listReportableAgents(bgCtx, profileRoot)
 		},
 	)
 	if p.cfg.CalendarEnabled {
@@ -125,18 +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()
 		},
 	}
 
@@ -203,21 +236,7 @@ func mapStateToCalendar(s string) calendar.AgentStatusValue {
 	case "offline":
 		return calendar.AgentStatusOffline
 	}
-	return calendar.AgentStatusUnknown
+	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). Returns the only active calendar
-// slot's agent if there's exactly one; 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 sch.Active[0].Slot.AgentID
-	}
-	return ""
 }
 
 func manifestFromDisk() sdkplugin.Manifest {

internal/calendar/bridge.go

@@ -1,7 +1,8 @@
-// Bridge — thin HTTP client for the HarborForge backend's Calendar API.
+// Bridge — typed HTTP client for HarborForge.Backend's calendar API.
-// All operations carry the API key as Authorization: Bearer; absent
+// Endpoint shapes verified via the backend's /openapi.json and against
-// key means missing-auth errors from the backend (caller should
+// HarborForge.OpenclawPlugin/plugin/calendar/calendar-bridge.ts so
-// handle them as transient and log).
+// the two plugins drop into the same backend without per-plugin
+// adapters.
 
 package calendar
 
@@ -13,30 +14,33 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"strconv"
 	"strings"
 	"time"
 )
 
-// Bridge is the typed wrapper around an HTTP client + backend URL.
+// Bridge is constructed once per scheduler and reused across heartbeats.
 type Bridge struct {
-	BackendURL string
+	BaseURL        string
-	APIKey     string
+	ClawIdentifier string
-	HTTP       *http.Client
+	HTTP           *http.Client
 }
 
-// New constructs a bridge with a sensible default timeout.
+// New constructs a bridge with a 20s default timeout.
-func New(backendURL, apiKey string) *Bridge {
+func New(baseURL, clawIdentifier string) *Bridge {
 	return &Bridge{
-		BackendURL: strings.TrimRight(backendURL, "/"),
+		BaseURL:        strings.TrimRight(baseURL, "/"),
-		APIKey:     apiKey,
+		ClawIdentifier: clawIdentifier,
-		HTTP:       &http.Client{Timeout: 20 * time.Second},
+		HTTP:           &http.Client{Timeout: 20 * time.Second},
 	}
 }
 
-// Heartbeat POSTs /calendar/agent/heartbeat. Returns the backend's
+// Heartbeat POSTs /calendar/agent/heartbeat. Per-agent: each running
-// reply or an error.
+// agent on this claw drives its own heartbeat (matches OpenClaw plugin
-func (b *Bridge) Heartbeat(ctx context.Context, payload HeartbeatPayload) (HeartbeatResponse, error) {
+// semantics).
-	raw, err := b.post(ctx, "/calendar/agent/heartbeat", payload)
+func (b *Bridge) Heartbeat(ctx context.Context, agentID string) (HeartbeatResponse, error) {
+	body := HeartbeatRequest{ClawIdentifier: b.ClawIdentifier, AgentID: agentID}
+	raw, err := b.doJSON(ctx, http.MethodPost, "/calendar/agent/heartbeat", agentID, body)
 	if err != nil {
 		return HeartbeatResponse{}, err
 	}
@@ -49,76 +53,64 @@ func (b *Bridge) Heartbeat(ctx context.Context, payload HeartbeatPayload) (Heart
 	return out, nil
 }
 
-// UpdateSlotStatus POSTs /calendar/slot/<id>/status to mark a slot
+// UpdateRealSlot PATCHes /calendar/slots/{id}/agent-update.
-// completed / aborted / paused / resumed.
+func (b *Bridge) UpdateRealSlot(ctx context.Context, agentID string, slotID int64, update SlotAgentUpdate) error {
-func (b *Bridge) UpdateSlotStatus(ctx context.Context, slotID string, update SlotUpdate) error {
+	path := "/calendar/slots/" + strconv.FormatInt(slotID, 10) + "/agent-update"
-	if slotID == "" {
+	_, err := b.doJSON(ctx, http.MethodPatch, path, agentID, update)
-		return errors.New("calendar: slot id required")
-	}
-	_, err := b.post(ctx, "/calendar/slot/"+slotID+"/status", update)
 	return err
 }
 
-// RestartPending GETs /restart/status — returns the backend's
+// UpdateVirtualSlot PATCHes /calendar/slots/virtual/{vid}/agent-update.
-// current restart-requested flag.
+// The backend materialises the virtual slot first; subsequent calls
-func (b *Bridge) RestartPending(ctx context.Context) (bool, error) {
+// against the same logical slot should use UpdateRealSlot with the
-	raw, err := b.get(ctx, "/restart/status")
+// id returned in the response — but for v1 we don't round-trip the
-	if err != nil {
+// materialised id back to the scheduler (would require a separate
-		return false, err
+// fetch); the agent-update path tolerates re-PATCHing a virtual id.
-	}
+func (b *Bridge) UpdateVirtualSlot(ctx context.Context, agentID string, virtualID string, update SlotAgentUpdate) error {
-	var out struct {
+	path := "/calendar/slots/virtual/" + virtualID + "/agent-update"
-		Pending bool `json:"pending"`
+	_, err := b.doJSON(ctx, http.MethodPatch, path, agentID, update)
-	}
+	return err
-	if len(raw) > 0 {
-		if err := json.Unmarshal(raw, &out); err != nil {
-			return false, fmt.Errorf("decode restart-status: %w", err)
-		}
-	}
-	return out.Pending, nil
 }
 
-// post serialises body as JSON, attaches Authorization, returns
+// PushAgentStatus POSTs /calendar/agent/status. Used to push idle ↔
-// response body bytes. Non-2xx becomes an error with the body
+// busy transitions out of the normal heartbeat cycle.
-// included for diagnostics.
+func (b *Bridge) PushAgentStatus(ctx context.Context, agentID string, status AgentStatusValue) error {
-func (b *Bridge) post(ctx context.Context, path string, body any) ([]byte, error) {
+	body := AgentStatusPush{
+		ClawIdentifier: b.ClawIdentifier, AgentID: agentID, Status: status,
+	}
+	_, err := b.doJSON(ctx, http.MethodPost, "/calendar/agent/status", agentID, body)
+	return err
+}
+
+// doJSON serialises body, attaches the two auth headers, and returns
+// the response bytes. Errors on non-2xx with truncated body.
+func (b *Bridge) doJSON(ctx context.Context, method, path, agentID string, body any) ([]byte, error) {
+	if agentID == "" {
+		return nil, errors.New("calendar: agent_id required for auth headers")
+	}
 	raw, err := json.Marshal(body)
 	if err != nil {
-		return nil, fmt.Errorf("marshal %s: %w", path, err)
+		return nil, fmt.Errorf("marshal %s %s: %w", method, path, err)
 	}
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, b.BackendURL+path, bytes.NewReader(raw))
+	req, err := http.NewRequestWithContext(ctx, method, b.BaseURL+path, bytes.NewReader(raw))
 	if err != nil {
 		return nil, err
 	}
 	req.Header.Set("Content-Type", "application/json")
-	if b.APIKey != "" {
+	req.Header.Set("X-Agent-ID", agentID)
-		req.Header.Set("Authorization", "Bearer "+b.APIKey)
+	req.Header.Set("X-Claw-Identifier", b.ClawIdentifier)
-	}
-	return b.do(req)
-}
 
-func (b *Bridge) get(ctx context.Context, path string) ([]byte, error) {
-	req, err := http.NewRequestWithContext(ctx, http.MethodGet, b.BackendURL+path, nil)
-	if err != nil {
-		return nil, err
-	}
-	if b.APIKey != "" {
-		req.Header.Set("Authorization", "Bearer "+b.APIKey)
-	}
-	return b.do(req)
-}
-
-func (b *Bridge) do(req *http.Request) ([]byte, error) {
 	res, err := b.HTTP.Do(req)
 	if err != nil {
-		return nil, fmt.Errorf("%s %s: %w", req.Method, req.URL.Path, err)
+		return nil, fmt.Errorf("%s %s: %w", method, path, err)
 	}
 	defer res.Body.Close()
-	body, _ := io.ReadAll(res.Body)
+	out, _ := io.ReadAll(res.Body)
 	if res.StatusCode < 200 || res.StatusCode >= 300 {
 		return nil, fmt.Errorf("%s %s → %d: %s",
-			req.Method, req.URL.Path, res.StatusCode, truncate(body, 300))
+			method, path, res.StatusCode, truncate(out, 300))
 	}
-	return body, nil
+	return out, nil
 }
 
 func truncate(b []byte, n int) string {

internal/calendar/scheduler.go

@@ -1,14 +1,16 @@
-// Scheduler — main loop that heartbeats the backend, dispatches
+// Scheduler — loops over every Plexum agent, heartbeats per-agent,
-// returned slots via Plexum's WakeAgent, and tracks per-agent active
+// picks the highest-priority pending slot for each, dispatches via
-// slot state for the calendar_* tools.
+// host.WakeAgent. Mirrors HarborForge.OpenclawPlugin's per-agent
+// scheduler loop (PLG-CAL-002).
 //
-// State is in-memory: a daemon restart drops everything. Next
+// In-memory state: per-agent active slot map. A daemon restart drops
-// heartbeat reconciles (backend keeps the canonical SlotStatus).
+// it; next heartbeat reconciles from the backend's canonical state.
 //
-// Concurrency:
+// Wake semantics: WakeAgent is fire-and-forget; the SDK's wake queue
-//   - one heartbeat ticker goroutine
+// (depth 1 replace-newest) handles state-aware dispatch. We mark the
-//   - per-slot dispatch is fire-and-forget via WakeAgent (queue-aware)
+// slot Ongoing optimistically the moment we call WakeAgent; agents
-//   - mu guards activeBySlot + activeByAgent maps
+// drive complete/abort/pause/resume via the harborforge_calendar_*
+// tools.
 
 package calendar
 
@@ -22,22 +24,20 @@ import (
 	sdkplugin "git.hangman-lab.top/hzhang/Plexum-sdk-go/plugin"
 )
 
-// Scheduler orchestrates the calendar loop.
+// Scheduler is the long-running calendar driver.
 type Scheduler struct {
-	cfg          Config
+	cfg         Config
-	bridge       *Bridge
+	bridge      *Bridge
-	host         sdkplugin.HostAPI
+	host        sdkplugin.HostAPI
-	agentLister  func() []ReportableAgent
+	agentLister func() []ReportableAgent
-	identifier   string
+	pluginInfo  PluginInfoTag
-	pluginInfo   PluginInfoTag
 
-	mu              sync.Mutex
+	mu                sync.Mutex
-	activeBySlotID  map[string]*ActiveSlot
+	activeByAgentID   map[string]*ActiveSlot
-	activeByAgentID map[string]*ActiveSlot
+	activeBySlotIdent map[string]*ActiveSlot
-	history         []HistoryEntry
+	history           []HistoryEntry
-	lastHeartbeat   time.Time
+	lastHeartbeats    map[string]time.Time
-	lastResponse    HeartbeatResponse
+	lastErrors        map[string]string
-	restartPending  bool
 }
 
 // Config bundles scheduler tunables.
@@ -46,36 +46,42 @@ type Config struct {
 	HistoryCap        int // bound on activity history; default 32
 }
 
-// ReportableAgent is the projection of a Plexum agent the scheduler
+// PluginInfoTag tags heartbeat reports so the backend knows which
-// needs for heartbeat — id + model + current sm state.
+// plugin / version is reporting.
-type ReportableAgent struct {
+type PluginInfoTag struct {
-	ID     string
+	Name    string
-	Model  string
+	Version string
-	State  AgentStatusValue
+	Backend string // "plexum"
 }
 
-// ActiveSlot tracks an in-flight slot (between WakeAgent dispatch and
+// ReportableAgent is the per-agent projection the scheduler needs for
-// terminal status update).
+// heartbeat enumeration.
+type ReportableAgent struct {
+	ID    string
+	Model string
+	State AgentStatusValue
+}
+
+// ActiveSlot tracks an in-flight slot from dispatch to terminal state.
 type ActiveSlot struct {
 	Slot          Slot
 	StartedAt     time.Time
 	LastHeartbeat time.Time
-	State         SlotStatus
 }
 
-// HistoryEntry is one resolved slot kept for the calendar_status tool.
+// HistoryEntry records one resolved slot for the calendar_status tool.
 type HistoryEntry struct {
-	SlotID    string
+	Ident       string
-	AgentID   string
+	AgentID     string
-	Status    SlotStatus
+	Status      SlotStatus
-	ResolvedAt time.Time
+	ResolvedAt  time.Time
-	Reason    string
+	Reason      string
-	Summary   string
+	Summary     string
 }
 
 // NewScheduler constructs a Scheduler in stopped state.
 func NewScheduler(cfg Config, bridge *Bridge, host sdkplugin.HostAPI,
-	identifier string, pluginInfo PluginInfoTag,
+	pluginInfo PluginInfoTag,
 	agentLister func() []ReportableAgent) *Scheduler {
 	if cfg.HeartbeatInterval <= 0 {
 		cfg.HeartbeatInterval = 30 * time.Second
@@ -84,193 +90,232 @@ func NewScheduler(cfg Config, bridge *Bridge, host sdkplugin.HostAPI,
 		cfg.HistoryCap = 32
 	}
 	return &Scheduler{
-		cfg:             cfg,
+		cfg:               cfg,
-		bridge:          bridge,
+		bridge:            bridge,
-		host:            host,
+		host:              host,
-		agentLister:     agentLister,
+		agentLister:       agentLister,
-		identifier:      identifier,
+		pluginInfo:        pluginInfo,
-		pluginInfo:      pluginInfo,
+		activeByAgentID:   map[string]*ActiveSlot{},
-		activeBySlotID:  map[string]*ActiveSlot{},
+		activeBySlotIdent: map[string]*ActiveSlot{},
-		activeByAgentID: map[string]*ActiveSlot{},
+		lastHeartbeats:    map[string]time.Time{},
+		lastErrors:        map[string]string{},
 	}
 }
 
-// Run blocks until ctx cancels, ticking heartbeats every
+// Run blocks until ctx cancels.
-// cfg.HeartbeatInterval. Returns nil on graceful shutdown.
 func (s *Scheduler) Run(ctx context.Context) error {
 	t := time.NewTicker(s.cfg.HeartbeatInterval)
 	defer t.Stop()
-	// First heartbeat immediately so initial state lands fast.
+	s.tick(ctx)
-	s.heartbeatOnce(ctx)
 	for {
 		select {
 		case <-ctx.Done():
 			return nil
 		case <-t.C:
-			s.heartbeatOnce(ctx)
+			s.tick(ctx)
 		}
 	}
 }
 
-func (s *Scheduler) heartbeatOnce(ctx context.Context) {
+func (s *Scheduler) tick(ctx context.Context) {
-	payload := HeartbeatPayload{
+	if s.agentLister == nil {
-		Identifier: s.identifier,
-		APIKey:     s.bridge.APIKey,
-		PluginInfo: s.pluginInfo,
-		CapturedAt: time.Now().UTC(),
-	}
-	if s.agentLister != nil {
-		for _, a := range s.agentLister() {
-			payload.AgentList = append(payload.AgentList, AgentReport{
-				ID: a.ID, Model: a.Model, Status: a.State,
-			})
-		}
-	}
-	resp, err := s.bridge.Heartbeat(ctx, payload)
-	s.mu.Lock()
-	s.lastHeartbeat = time.Now()
-	if err == nil {
-		s.lastResponse = resp
-		s.restartPending = resp.RestartPending
-	}
-	s.mu.Unlock()
-	if err != nil {
-		return // network blip; next tick retries
-	}
-	for _, slot := range resp.SlotsToFire {
-		s.dispatchSlot(ctx, slot)
-	}
-}
-
-// dispatchSlot fires the slot via host.WakeAgent and records it as
-// active. WakeAgent handles state-aware queueing — if the agent is
-// busy, our calendar slot enqueues at depth 1 and the previous wake
-// is dropped per replace-newest semantics. We mark the slot
-// in_progress optimistically when we ENQUEUED; backend reconciles on
-// its own watchdog.
-func (s *Scheduler) dispatchSlot(ctx context.Context, slot Slot) {
-	// Skip already-active slots (heartbeat may re-list a slot we
-	// already started — backend hasn't seen our optimistic update yet).
-	s.mu.Lock()
-	if _, ok := s.activeBySlotID[slot.ID]; ok {
-		s.mu.Unlock()
 		return
 	}
 	now := time.Now().UTC()
-	act := &ActiveSlot{
+	agents := s.agentLister()
-		Slot: slot, StartedAt: now, LastHeartbeat: now,
+	s.host.Log("info", "calendar tick", map[string]any{"agents": len(agents)})
-		State: SlotInProgress,
+	for _, agent := range agents {
+		s.tickForAgent(ctx, agent, now)
 	}
-	s.activeBySlotID[slot.ID] = act
+}
-	s.activeByAgentID[slot.AgentID] = act
-	s.mu.Unlock()
 
-	message := slot.WakeOptions.OverrideMessage
+func (s *Scheduler) tickForAgent(ctx context.Context, agent ReportableAgent, now time.Time) {
-	if message == "" {
+	resp, err := s.bridge.Heartbeat(ctx, agent.ID)
-		message = slot.PromptText
+	s.mu.Lock()
-	}
+	s.lastHeartbeats[agent.ID] = now
-	if message == "" {
+	if err != nil {
-		message = fmt.Sprintf("[calendar] slot %s: %s", slot.ID, slot.Title)
+		s.lastErrors[agent.ID] = err.Error()
-	}
+		s.mu.Unlock()
-	source := fmt.Sprintf("calendar:slot-%s", slot.ID)
+		s.host.Log("warn", "calendar heartbeat failed", map[string]any{
-	if err := s.host.WakeAgent(ctx, sdkplugin.WakeAgentRequest{
+			"agent": agent.ID, "err": err.Error(),
-		AgentID: slot.AgentID,
+		})
-		Message: message,
-		Source:  source,
-	}); err != nil {
-		// Wake itself failed (plumbing). Mark slot aborted +
-		// notify backend.
-		s.resolveSlot(ctx, slot.ID, SlotAborted, "", "wake-agent failed: "+err.Error())
 		return
 	}
+	delete(s.lastErrors, agent.ID)
+	s.mu.Unlock()
+	s.host.Log("info", "calendar heartbeat ok", map[string]any{
+		"agent": agent.ID, "slots": len(resp.Slots), "agent_status": string(resp.AgentStatus),
+	})
+
+	// Pick highest-priority NotStarted slot; defer the rest.
+	var chosen *Slot
+	for i := range resp.Slots {
+		slot := &resp.Slots[i]
+		if slot.Status != SlotNotStarted && slot.Status != SlotDeferred {
+			continue
+		}
+		if chosen == nil || slot.Priority > chosen.Priority {
+			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)
+	}
+	// Defer the other unchosen NotStarted/Deferred slots (priority +1)
+	// so they bubble up next heartbeat. We don't strictly need to push
+	// the update; the backend's priority bookkeeping survives without
+	// our nudge for v1. (OpenClaw plugin DOES push priority bumps —
+	// future v2 work if backend feedback shows starvation.)
 }
 
-// resolveSlot moves an active slot to a terminal status, records
+// dispatchSlot fires WakeAgent + records the slot active. Marks the
-// history, and tells the backend. Safe to call concurrently.
+// slot Ongoing on the backend so the dashboard reflects the
-func (s *Scheduler) resolveSlot(ctx context.Context, slotID string, status SlotStatus, summary, reason string) error {
+// 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()
-	act, ok := s.activeBySlotID[slotID]
+	if _, dup := s.activeBySlotIdent[ident]; dup {
-	if !ok {
 		s.mu.Unlock()
-		return fmt.Errorf("calendar: slot %s not active", slotID)
+		s.host.Log("info", "calendar dispatchSlot skipped (already active)", map[string]any{"slot": ident})
+		return
 	}
-	delete(s.activeBySlotID, slotID)
+	if _, agentBusy := s.activeByAgentID[agentID]; agentBusy {
-	delete(s.activeByAgentID, act.Slot.AgentID)
+		// Don't pick up another slot until the current one resolves.
-	s.appendHistoryLocked(HistoryEntry{
+		s.mu.Unlock()
-		SlotID: slotID, AgentID: act.Slot.AgentID, Status: status,
+		s.host.Log("info", "calendar dispatchSlot skipped (agent has active slot)", map[string]any{"agent": agentID})
+		return
+	}
+	now := time.Now().UTC()
+	active := &ActiveSlot{Slot: slot, StartedAt: now, LastHeartbeat: now}
+	s.activeBySlotIdent[ident] = active
+	s.activeByAgentID[agentID] = active
+	s.mu.Unlock()
+
+	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"),
+	}
+	s.pushUpdate(ctx, agentID, slot, update)
+}
+
+func buildWakeMessage(slot Slot) string {
+	// Backend EventData → prompt. v1 is intentionally simple; refine
+	// when the prompt-engineering side of the plugin matures.
+	if slot.EventType != nil {
+		switch *slot.EventType {
+		case EventTypeSystemEvent:
+			if ev, ok := slot.EventData["event"].(string); ok {
+				return fmt.Sprintf("[calendar system_event] %s", ev)
+			}
+		case EventTypeJob:
+			code, _ := slot.EventData["code"].(string)
+			typ, _ := slot.EventData["type"].(string)
+			if code != "" {
+				return fmt.Sprintf("[calendar job %s/%s] please handle this", typ, code)
+			}
+		}
+	}
+	return fmt.Sprintf("[calendar slot %s] scheduled work — please proceed", slot.SlotIdent())
+}
+
+// CompleteForAgent → terminal; pushes Finished to backend.
+func (s *Scheduler) CompleteForAgent(ctx context.Context, agentID, summary string) error {
+	act, ok := s.activeSlotForAgent(agentID)
+	if !ok {
+		return ErrNoActiveSlot
+	}
+	now := time.Now().UTC()
+	duration := int(now.Sub(act.StartedAt).Minutes())
+	if duration < 1 {
+		duration = 1
+	}
+	if err := s.pushUpdate(ctx, agentID, act.Slot, SlotAgentUpdate{
+		Status: SlotFinished, ActualDuration: duration,
+	}); err != nil {
+		return err
+	}
+	s.resolveLocally(act.Slot.SlotIdent(), agentID, SlotFinished, summary, "")
+	return nil
+}
+
+// AbortForAgent → terminal; pushes Aborted to backend.
+func (s *Scheduler) AbortForAgent(ctx context.Context, agentID, reason string) error {
+	act, ok := s.activeSlotForAgent(agentID)
+	if !ok {
+		return ErrNoActiveSlot
+	}
+	if err := s.pushUpdate(ctx, agentID, act.Slot, SlotAgentUpdate{Status: SlotAborted}); err != nil {
+		return err
+	}
+	s.resolveLocally(act.Slot.SlotIdent(), agentID, SlotAborted, "", reason)
+	return nil
+}
+
+// PauseForAgent → non-terminal; pushes Paused.
+func (s *Scheduler) PauseForAgent(ctx context.Context, agentID, reason string) error {
+	act, ok := s.activeSlotForAgent(agentID)
+	if !ok {
+		return ErrNoActiveSlot
+	}
+	return s.pushUpdate(ctx, agentID, act.Slot, SlotAgentUpdate{Status: SlotPaused})
+}
+
+// ResumeForAgent → non-terminal; pushes Ongoing.
+func (s *Scheduler) ResumeForAgent(ctx context.Context, agentID string) error {
+	act, ok := s.activeSlotForAgent(agentID)
+	if !ok {
+		return ErrNoActiveSlot
+	}
+	return s.pushUpdate(ctx, agentID, act.Slot, SlotAgentUpdate{Status: SlotOngoing})
+}
+
+func (s *Scheduler) pushUpdate(ctx context.Context, agentID string, slot Slot, update SlotAgentUpdate) error {
+	if slot.HasRealID() {
+		return s.bridge.UpdateRealSlot(ctx, agentID, *slot.ID, update)
+	}
+	if slot.VirtualID != nil {
+		return s.bridge.UpdateVirtualSlot(ctx, agentID, *slot.VirtualID, update)
+	}
+	return errors.New("calendar: slot has neither real id nor virtual id")
+}
+
+func (s *Scheduler) resolveLocally(ident, agentID string, status SlotStatus, summary, reason string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	delete(s.activeBySlotIdent, ident)
+	delete(s.activeByAgentID, agentID)
+	s.history = append(s.history, HistoryEntry{
+		Ident: ident, AgentID: agentID, Status: status,
 		ResolvedAt: time.Now().UTC(), Summary: summary, Reason: reason,
 	})
-	s.mu.Unlock()
-	return s.bridge.UpdateSlotStatus(ctx, slotID, SlotUpdate{
-		Status: status, Summary: summary, Reason: reason,
-	})
-}
-
-// SetSlotState is a non-terminal status change (paused/resumed).
-// Records the new state in-memory and tells the backend.
-func (s *Scheduler) SetSlotState(ctx context.Context, slotID string, status SlotStatus, reason string) error {
-	s.mu.Lock()
-	act, ok := s.activeBySlotID[slotID]
-	if !ok {
-		s.mu.Unlock()
-		return fmt.Errorf("calendar: slot %s not active", slotID)
-	}
-	act.State = status
-	act.LastHeartbeat = time.Now().UTC()
-	s.mu.Unlock()
-	return s.bridge.UpdateSlotStatus(ctx, slotID, SlotUpdate{
-		Status: status, Reason: reason,
-	})
-}
-
-func (s *Scheduler) appendHistoryLocked(entry HistoryEntry) {
-	s.history = append(s.history, entry)
 	if len(s.history) > s.cfg.HistoryCap {
 		s.history = s.history[len(s.history)-s.cfg.HistoryCap:]
 	}
 }
 
-// CompleteForAgent / AbortForAgent / PauseForAgent / ResumeForAgent
-// are the agent-facing tool entry points. They look up the agent's
-// active slot, transition or terminate it, and notify the backend.
-
-// CompleteForAgent terminates the agent's active slot as completed.
-func (s *Scheduler) CompleteForAgent(ctx context.Context, agentID, summary string) error {
-	slot, ok := s.activeSlotForAgent(agentID)
-	if !ok {
-		return ErrNoActiveSlot
-	}
-	return s.resolveSlot(ctx, slot.Slot.ID, SlotCompleted, summary, "")
-}
-
-// AbortForAgent terminates the agent's active slot as aborted.
-func (s *Scheduler) AbortForAgent(ctx context.Context, agentID, reason string) error {
-	slot, ok := s.activeSlotForAgent(agentID)
-	if !ok {
-		return ErrNoActiveSlot
-	}
-	return s.resolveSlot(ctx, slot.Slot.ID, SlotAborted, "", reason)
-}
-
-// PauseForAgent transitions the agent's slot to paused.
-func (s *Scheduler) PauseForAgent(ctx context.Context, agentID, reason string) error {
-	slot, ok := s.activeSlotForAgent(agentID)
-	if !ok {
-		return ErrNoActiveSlot
-	}
-	return s.SetSlotState(ctx, slot.Slot.ID, SlotPaused, reason)
-}
-
-// ResumeForAgent transitions the agent's slot back to in_progress.
-func (s *Scheduler) ResumeForAgent(ctx context.Context, agentID string) error {
-	slot, ok := s.activeSlotForAgent(agentID)
-	if !ok {
-		return ErrNoActiveSlot
-	}
-	return s.SetSlotState(ctx, slot.Slot.ID, SlotInProgress, "")
-}
-
-// activeSlotForAgent returns the per-agent active slot copy under lock.
 func (s *Scheduler) activeSlotForAgent(agentID string) (ActiveSlot, bool) {
 	s.mu.Lock()
 	defer s.mu.Unlock()
@@ -281,36 +326,59 @@ func (s *Scheduler) activeSlotForAgent(agentID string) (ActiveSlot, bool) {
 	return *act, true
 }
 
-// Status returns the introspection shape for the calendar_status tool.
+// Status is the introspection shape calendar_status returns.
-func (s *Scheduler) Status() SchedulerStatus {
+type Status struct {
+	Enabled        bool                 `json:"enabled"`
+	LastHeartbeats map[string]time.Time `json:"last_heartbeats"`
+	LastErrors     map[string]string    `json:"last_errors,omitempty"`
+	HeartbeatEvery time.Duration        `json:"heartbeat_every"`
+	Active         []ActiveSlot         `json:"active"`
+	History        []HistoryEntry       `json:"history"`
+}
+
+// SingleActiveAgentID returns the agent id when exactly one active
+// slot exists, empty otherwise. Used by the plugin's bestEffortAgentID
+// fallback for tool calls that don't carry agent context.
+func (s *Scheduler) SingleActiveAgentID() string {
 	s.mu.Lock()
 	defer s.mu.Unlock()
-	active := make([]ActiveSlot, 0, len(s.activeBySlotID))
+	if len(s.activeByAgentID) != 1 {
-	for _, a := range s.activeBySlotID {
+		return ""
+	}
+	for k := range s.activeByAgentID {
+		return k
+	}
+	return ""
+}
+
+// Status returns the introspection shape calendar_status returns.
+func (s *Scheduler) Status() Status {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	active := make([]ActiveSlot, 0, len(s.activeByAgentID))
+	for _, a := range s.activeByAgentID {
 		active = append(active, *a)
 	}
+	hb := make(map[string]time.Time, len(s.lastHeartbeats))
+	for k, v := range s.lastHeartbeats {
+		hb[k] = v
+	}
+	errs := make(map[string]string, len(s.lastErrors))
+	for k, v := range s.lastErrors {
+		errs[k] = v
+	}
 	history := make([]HistoryEntry, len(s.history))
 	copy(history, s.history)
-	return SchedulerStatus{
+	return Status{
 		Enabled:        true,
-		LastHeartbeat:  s.lastHeartbeat,
+		LastHeartbeats: hb,
+		LastErrors:     errs,
 		HeartbeatEvery: s.cfg.HeartbeatInterval,
 		Active:         active,
 		History:        history,
-		RestartPending: s.restartPending,
 	}
 }
 
-// SchedulerStatus is the shape calendar_status returns.
+// ErrNoActiveSlot is returned when an agent calls calendar_complete /
-type SchedulerStatus struct {
+// abort / pause / resume but has no slot active.
-	Enabled        bool          `json:"enabled"`
-	LastHeartbeat  time.Time     `json:"last_heartbeat"`
-	HeartbeatEvery time.Duration `json:"heartbeat_every"`
-	Active         []ActiveSlot  `json:"active"`
-	History        []HistoryEntry `json:"history"`
-	RestartPending bool          `json:"restart_pending"`
-}
-
-// ErrNoActiveSlot is returned by calendar_complete/abort/pause/resume
-// when the agent has no slot in progress.
 var ErrNoActiveSlot = errors.New("calendar: no active slot for agent")

internal/calendar/types.go

@@ -1,106 +1,154 @@
-// Package calendar talks to the HarborForge backend's Calendar API
+// Types matching HarborForge.Backend's actual calendar API contract
-// (heartbeat, slot fetch, status update, restart-pending check) and
+// (verified via /openapi.json on a running backend). Aligns 1:1 with
-// drives a scheduler loop that fires Plexum wake events when slots
+// HarborForge.OpenclawPlugin/plugin/calendar/types.ts so the two
-// come due. Types mirror HarborForge.OpenclawPlugin's calendar/types.ts
+// plugins can hit the same backend interchangeably.
-// so the backend doesn't need to know which plugin is reporting.
 
 package calendar
 
 import "time"
 
-// SlotStatus enumerates the slot lifecycle.
+// SlotStatus enumerates the lifecycle. String values match backend's
+// SlotStatus enum verbatim (snake_case — verified via heartbeat
+// response shape against running harborforge-backend).
 type SlotStatus string
 
 const (
 	SlotNotStarted SlotStatus = "not_started"
-	SlotInProgress SlotStatus = "in_progress"
+	SlotOngoing    SlotStatus = "ongoing"
-	SlotCompleted  SlotStatus = "completed"
+	SlotFinished   SlotStatus = "finished"
 	SlotAborted    SlotStatus = "aborted"
-	SlotPaused     SlotStatus = "paused"
 	SlotDeferred   SlotStatus = "deferred"
+	SlotPaused     SlotStatus = "paused"
+	SlotSkipped    SlotStatus = "skipped"
 )
 
-// AgentStatusValue mirrors the backend AgentStatus enum used in
+// SlotType: work vs on_call. Affects whether the agent flips to busy.
-// heartbeat responses (a hint about what the backend thinks the
+type SlotType string
-// agent is doing).
+
+const (
+	SlotTypeWork   SlotType = "work"
+	SlotTypeOnCall SlotType = "on_call"
+)
+
+// EventType categorises what the slot represents.
+type EventType string
+
+const (
+	EventTypeJob          EventType = "job"
+	EventTypeSystemEvent  EventType = "system_event"
+	EventTypeEntertainment EventType = "entertainment"
+)
+
+// AgentStatusValue mirrors the backend AgentStatus enum.
 type AgentStatusValue string
 
 const (
-	AgentStatusUnknown  AgentStatusValue = "unknown"
+	AgentStatusIdle      AgentStatusValue = "idle"
-	AgentStatusIdle     AgentStatusValue = "idle"
+	AgentStatusBusy      AgentStatusValue = "busy"
-	AgentStatusBusy     AgentStatusValue = "busy"
+	AgentStatusOffline   AgentStatusValue = "offline"
-	AgentStatusOffline  AgentStatusValue = "offline"
+	AgentStatusOnCall    AgentStatusValue = "on_call"
-	AgentStatusOnCall   AgentStatusValue = "on_call"
+	AgentStatusExhausted AgentStatusValue = "exhausted"
-	AgentStatusPaused   AgentStatusValue = "paused"
 )
 
-// SlotKind is "work" vs "on_call" — affects how the scheduler treats
+// HeartbeatRequest is the POST /calendar/agent/heartbeat body.
-// the slot (on_call slots don't move the agent into busy).
+type HeartbeatRequest struct {
-type SlotKind string
+	ClawIdentifier string `json:"claw_identifier"`
-
+	AgentID        string `json:"agent_id"`
-const (
-	SlotKindWork   SlotKind = "work"
-	SlotKindOnCall SlotKind = "on_call"
-)
-
-// Slot is one Calendar TimeSlot the backend serves.
-type Slot struct {
-	ID           string     `json:"id"`
-	VirtualID    string     `json:"virtual_id,omitempty"`
-	AgentID      string     `json:"agent_id"`
-	ClawID       string     `json:"claw_identifier,omitempty"`
-	Kind         SlotKind   `json:"slot_type"`
-	Title        string     `json:"title,omitempty"`
-	Description  string     `json:"description,omitempty"`
-	ScheduledAt  time.Time  `json:"scheduled_at"`
-	ExpiresAt    *time.Time `json:"expires_at,omitempty"`
-	Status       SlotStatus `json:"status"`
-	PromptText   string     `json:"prompt,omitempty"`
-	WakeOptions  WakeOpts   `json:"wake_options,omitempty"`
 }
 
-// WakeOpts customise how the scheduler should drive the agent. v1
+// HeartbeatResponse is the backend's reply.
-// honours only Force; the rest pass through as audit trail.
-type WakeOpts struct {
-	Force            bool   `json:"force,omitempty"`
-	OverrideMessage  string `json:"override_message,omitempty"`
-	ScopeSessionID   string `json:"scope_session_id,omitempty"`
-}
-
-// HeartbeatPayload is what the plugin POSTs every interval.
-type HeartbeatPayload struct {
-	Identifier  string         `json:"identifier"`
-	APIKey      string         `json:"api_key,omitempty"`
-	AgentList   []AgentReport  `json:"agents"`
-	PluginInfo  PluginInfoTag  `json:"plugin"`
-	CapturedAt  time.Time      `json:"captured_at"`
-}
-
-// AgentReport is one entry in HeartbeatPayload.AgentList.
-type AgentReport struct {
-	ID     string           `json:"agent_id"`
-	Status AgentStatusValue `json:"status"`
-	Model  string           `json:"model,omitempty"`
-}
-
-// PluginInfoTag identifies which plugin / version is heartbeating.
-type PluginInfoTag struct {
-	Name    string `json:"name"`    // "harbor-forge"
-	Version string `json:"version"` // e.g. 0.1.0
-	Backend string `json:"backend"` // "plexum"
-}
-
-// HeartbeatResponse is the backend's reply. SlotsToFire are slots
-// the scheduler should attempt to start.
 type HeartbeatResponse struct {
-	SlotsToFire    []Slot `json:"slots_to_fire,omitempty"`
+	Slots       []Slot           `json:"slots"`
-	RestartPending bool   `json:"restart_pending,omitempty"`
+	AgentStatus AgentStatusValue `json:"agent_status"`
-	ServerTime     time.Time `json:"server_time"`
+	Message     string           `json:"message,omitempty"`
 }
 
-// SlotUpdate is the body of POST /calendar/slot/<id>/status.
+// Slot is one calendar TimeSlot — real (has ID) or virtual
-type SlotUpdate struct {
+// (has VirtualID). Field names mirror the backend's
-	Status  SlotStatus `json:"status"`
+// CalendarSlotResponse schema.
-	Summary string     `json:"summary,omitempty"`
+type Slot struct {
-	Reason  string     `json:"reason,omitempty"`
+	ID                *int64       `json:"id"`                  // real slot db id; null for virtual
+	VirtualID         *string      `json:"virtual_id"`          // plan-{plan_id}-{date}; null for real
+	UserID            int64        `json:"user_id"`
+	Date              string       `json:"date"`                // YYYY-MM-DD
+	SlotType          SlotType     `json:"slot_type"`
+	EstimatedDuration int          `json:"estimated_duration"`  // minutes
+	ScheduledAt       string       `json:"scheduled_at"`        // HH:MM:SS
+	StartedAt         *string      `json:"started_at"`
+	Attended          bool         `json:"attended"`
+	ActualDuration    *int         `json:"actual_duration"`
+	EventType         *EventType   `json:"event_type"`
+	EventData         EventData    `json:"event_data"`
+	Priority          int          `json:"priority"`
+	Status            SlotStatus   `json:"status"`
+	PlanID            *int64       `json:"plan_id"`
+}
+
+// EventData is loosely-typed since the backend stores it as JSONB and
+// the shape varies by event_type. Plugin code does best-effort
+// unmarshal into JobData / SystemEventData when needed.
+type EventData map[string]any
+
+// JobData is the event_data shape when event_type=="job".
+type JobData struct {
+	Type            string   `json:"type"`             // Task|Support|Meeting|Essential
+	Code            string   `json:"code"`             // e.g. "TASK-42"
+	WorkingSessions []string `json:"working_sessions"` // arbitrary session ids
+}
+
+// SystemEventData is the event_data shape when event_type=="system_event".
+type SystemEventData struct {
+	Event string `json:"event"` // ScheduleToday | SummaryToday | ScheduledGatewayRestart
+}
+
+// SlotAgentUpdate is the body of PATCH /calendar/slots/{id}/agent-update
+// (and the virtual variant). started_at + actual_duration are set
+// depending on which status transition the agent is reporting.
+type SlotAgentUpdate struct {
+	Status         SlotStatus `json:"status"`
+	StartedAt      string     `json:"started_at,omitempty"`     // HH:MM:SS
+	ActualDuration int        `json:"actual_duration,omitempty"` // minutes
+}
+
+// AgentStatusPush is the body of POST /calendar/agent/status.
+type AgentStatusPush struct {
+	ClawIdentifier string           `json:"claw_identifier"`
+	AgentID        string           `json:"agent_id"`
+	Status         AgentStatusValue `json:"status"`
+}
+
+// HasRealID reports whether a Slot is the materialized (DB row) flavor.
+func (s Slot) HasRealID() bool { return s.ID != nil && *s.ID > 0 }
+
+// SlotIdent returns a stable string identifier for log + map keys —
+// "real:<id>" for materialized, "virtual:<vid>" for virtual.
+func (s Slot) SlotIdent() string {
+	if s.HasRealID() {
+		return formatInt("real", *s.ID)
+	}
+	if s.VirtualID != nil {
+		return "virtual:" + *s.VirtualID
+	}
+	return "unknown:" + time.Now().UTC().Format(time.RFC3339Nano)
+}
+
+func formatInt(prefix string, n int64) string {
+	// avoid pulling fmt for one call
+	const digits = "0123456789"
+	if n == 0 {
+		return prefix + ":0"
+	}
+	neg := n < 0
+	if neg {
+		n = -n
+	}
+	buf := make([]byte, 0, 20)
+	for n > 0 {
+		buf = append([]byte{digits[n%10]}, buf...)
+		n /= 10
+	}
+	if neg {
+		buf = append([]byte{'-'}, buf...)
+	}
+	return prefix + ":" + string(buf)
 }

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,
 		},
-		"calendar": sch,
+		"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{
-		"port":       bs.Port,
+		"bridge": map[string]any{
-		"listening":  bs.Listening,
+			"port":          bs.Port,
-		"queries":    bs.Queries,
+			"listening":     bs.Listening,
-		"last_query": bs.LastQuery,
+			"queries":       bs.Queries,
-		"last_snapshot": bs.LastSnap,
+			"last_query":    bs.LastQuery,
+			"last_snapshot": bs.LastSnap,
+		},
+		"push": monitorPushSummary(deps),
 	})
 }
 
@@ -176,11 +201,15 @@ func toolCalendarResume(ctx context.Context, deps Deps) (sdkplugin.ToolResult, e
 }
 
 func toolRestartStatus(deps Deps) (sdkplugin.ToolResult, error) {
+	// HarborForge backend doesn't expose a restart-pending endpoint
+	// (verified via /openapi.json) so we report the most recent
+	// heartbeat freshness instead. Useful for operators sanity-
+	// checking that the plugin's calendar loop is still alive.
 	sch := deps.Scheduler.Status()
 	return jsonResult(map[string]any{
-		"pending":        sch.RestartPending,
+		"pending":         false,
-		"last_heartbeat": sch.LastHeartbeat,
+		"last_heartbeats": sch.LastHeartbeats,
-		"observed_at":    time.Now().UTC(),
+		"observed_at":     time.Now().UTC(),
 	})
 }
 

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