Merge branch 'fix/ui-status-dispatch' into test/gui-memory-leak-fix

client/server/status_stream.go

@@ -1,20 +1,27 @@
 package server
 
 import (
+	"context"
+	"time"
+
 	log "github.com/sirupsen/logrus"
 
 	"github.com/netbirdio/netbird/client/proto"
 )
 
+const statusCoalesceWindow = 200 * time.Millisecond
+
 // SubscribeStatus pushes a fresh StatusResponse on every connection state
 // change. The first message is the current snapshot, so a re-subscribing
 // client doesn't need to also call Status. Subsequent messages fire when
 // the peer recorder reports any of: connected/disconnected/connecting,
 // management or signal flip, address change, or peers list change.
 //
-// The change channel coalesces bursts to a single tick. If the consumer
-// is slow the daemon drops extras (not blocks), and the next snapshot
-// the consumer pulls already reflects everything.
+// Bursts are coalesced deterministically: the first tick after a quiet
+// period is sent immediately, then a short window swallows the rest of the
+// burst and a single trailing snapshot covers whatever arrived meanwhile.
+// Every send is a full snapshot of the recorder's current state, so
+// swallowed ticks lose no information.
 func (s *Server) SubscribeStatus(req *proto.StatusRequest, stream proto.DaemonService_SubscribeStatusServer) error {
 	subID, ch := s.statusRecorder.SubscribeToStateChanges()
 	defer func() {
@@ -37,12 +44,50 @@ func (s *Server) SubscribeStatus(req *proto.StatusRequest, stream proto.DaemonSe
 			if err := s.sendStatusSnapshot(req, stream); err != nil {
 				return err
 			}
+			pending, open := collectStatusBurst(stream.Context(), ch)
+			if pending {
+				if err := s.sendStatusSnapshot(req, stream); err != nil {
+					return err
+				}
+			}
+			if !open {
+				return nil
+			}
 		case <-stream.Context().Done():
 			return nil
 		}
 	}
 }
 
+// collectStatusBurst waits out the coalesce window, absorbing further ticks.
+// pending reports whether any tick arrived; open is false when the channel
+// closed or the stream context ended.
+func collectStatusBurst(ctx context.Context, ch <-chan struct{}) (pending, open bool) {
+	timer := time.NewTimer(statusCoalesceWindow)
+	defer timer.Stop()
+	for {
+		select {
+		case _, ok := <-ch:
+			if !ok {
+				return pending, false
+			}
+			pending = true
+		case <-timer.C:
+			select {
+			case _, ok := <-ch:
+				if !ok {
+					return pending, false
+				}
+				pending = true
+			default:
+			}
+			return pending, true
+		case <-ctx.Done():
+			return false, false
+		}
+	}
+}
+
 func (s *Server) sendStatusSnapshot(req *proto.StatusRequest, stream proto.DaemonService_SubscribeStatusServer) error {
 	resp, err := s.buildStatusResponse(stream.Context(), req)
 	if err != nil {

client/ui/main.go

@@ -103,6 +103,18 @@ func main() {
 	updaterHolder := updater.NewHolder(app.Event)
 	update := services.NewUpdate(conn, updaterHolder)
 	daemonFeed := services.NewDaemonFeed(conn, app.Event, updaterHolder, debugLog)
+	// Status snapshots go only to visible windows — a snapshot is full state,
+	// so a hidden webview loses nothing by being skipped; it gets the cached
+	// one on show (SetShowReplay below). Every other event stays on the bus.
+	daemonFeed.SetWindowDispatcher(func(st services.Status) {
+		ev := &application.CustomEvent{Name: services.EventStatusSnapshot, Data: st}
+		for _, w := range app.Window.GetAll() {
+			if w == nil || !w.IsVisible() {
+				continue
+			}
+			w.DispatchWailsEvent(ev)
+		}
+	})
 	notifier := notifications.New()
 	compat := services.NewCompat(conn)
 	// macOS shows no toast until permission is requested. Run it after
@@ -152,8 +164,31 @@ func main() {
 	// re-centering on that environment; nil leaves placement to the WM on full
 	// desktops, macOS, and Windows.
 	windowManager.SetRecenterOnShow(recenterOnShowPredicate())
+	// Replay the latest snapshot into a window on (re)show, so a webview that
+	// was hidden while pushes flowed never paints stale state. ReplayLast keeps
+	// the feed's status lock across the dispatch, so a racing live push can't
+	// slip in between and then be overwritten by this older cached snapshot.
+	windowManager.SetShowReplay(func(w application.Window) {
+		daemonFeed.ReplayLast(func(st services.Status) {
+			w.DispatchWailsEvent(&application.CustomEvent{Name: services.EventStatusSnapshot, Data: st})
+		})
+	})
 	app.RegisterService(application.NewService(windowManager))
 
+	// On macOS, Wails' default applicationShouldHandleReopen handler Show()s
+	// every hidden window on dock-icon click, resurrecting hide-on-close
+	// surfaces like Settings. Cancel it in a hook (hooks run before listeners)
+	// and show only the main window. No-op elsewhere — the event never fires.
+	if runtime.GOOS == "darwin" {
+		app.Event.RegisterApplicationEventHook(events.Mac.ApplicationShouldHandleReopen, func(e *application.ApplicationEvent) {
+			e.Cancel()
+			if e.Context().HasVisibleWindows() {
+				return
+			}
+			windowManager.ShowMain()
+		})
+	}
+
 	// Welcome window, first launch only — Continue flips OnboardingCompleted
 	// so later launches skip it. ApplicationStarted hook so the Wails window
 	// machinery is fully up before the window is created.
@@ -377,20 +412,5 @@ func newMainWindow(app *application.App, prefStore *preferences.Store) *applicat
 		window.Hide()
 	})
 
-	// On macOS, Wails' default applicationShouldHandleReopen handler Show()s
-	// every hidden window on dock-icon click, resurrecting hide-on-close
-	// surfaces like Settings. Cancel it in a hook (hooks run before listeners)
-	// and show only the main window. No-op elsewhere — the event never fires.
-	if runtime.GOOS == "darwin" {
-		app.Event.RegisterApplicationEventHook(events.Mac.ApplicationShouldHandleReopen, func(e *application.ApplicationEvent) {
-			e.Cancel()
-			if e.Context().HasVisibleWindows() {
-				return
-			}
-			window.Show()
-			window.Focus()
-		})
-	}
-
 	return window
 }

client/ui/services/daemon_feed.go

@@ -5,6 +5,7 @@ package services
 import (
 	"context"
 	"fmt"
+	"slices"
 	"strings"
 	"sync"
 	"time"
@@ -167,6 +168,14 @@ type DaemonFeed struct {
 	cancel   context.CancelFunc
 	streamWg sync.WaitGroup
 
+	// statusSubs are Go-side snapshot consumers (the tray), fed directly so
+	// they don't ride the window event bus. Callbacks run synchronously on
+	// the stream goroutine, so pushes arrive in order.
+	statusSubsMu     sync.Mutex
+	statusSubs       []func(Status)
+	lastStatus       *Status
+	windowDispatcher func(Status)
+
 	switchMu              sync.Mutex
 	switchInProgress      bool
 	switchInProgressUntil time.Time
@@ -188,6 +197,55 @@ func NewDaemonFeed(conn DaemonConn, emitter Emitter, updaterHolder *updater.Hold
 	return &DaemonFeed{conn: conn, emitter: emitter, updater: updaterHolder, logCtl: logCtl}
 }
 
+// OnStatus registers a Go-side status subscriber. Not for the frontend —
+// React consumers subscribe to EventStatusSnapshot on the event bus.
+func (s *DaemonFeed) OnStatus(cb func(Status)) {
+	s.statusSubsMu.Lock()
+	s.statusSubs = append(s.statusSubs, cb)
+	s.statusSubsMu.Unlock()
+}
+
+// SetWindowDispatcher installs the frontend push path: it receives every
+// snapshot and decides which webview windows get it (visible ones). While
+// unset, pushStatus falls back to the event-bus broadcast.
+func (s *DaemonFeed) SetWindowDispatcher(fn func(Status)) {
+	s.statusSubsMu.Lock()
+	s.windowDispatcher = fn
+	s.statusSubsMu.Unlock()
+}
+
+// pushStatus delivers a snapshot to the Go-side subscribers and the frontend,
+// and caches it for ReplayLast. Hidden windows are skipped by the
+// window dispatcher; they catch up via the show replay (WindowManager).
+func (s *DaemonFeed) pushStatus(st Status) {
+	s.statusSubsMu.Lock()
+	s.lastStatus = &st
+	subs := slices.Clone(s.statusSubs)
+	dispatch := s.windowDispatcher
+	s.statusSubsMu.Unlock()
+	for _, cb := range subs {
+		cb(st)
+	}
+	if dispatch != nil {
+		dispatch(st)
+		return
+	}
+	s.emitter.Emit(EventStatusSnapshot, st)
+}
+
+// ReplayLast feeds the most recently pushed snapshot to dispatch; no-op before
+// the first push. The status lock is held across the dispatch so a concurrent
+// pushStatus cannot deliver a newer snapshot in between the read and the
+// dispatch — the replayed value is never older than anything already delivered.
+func (s *DaemonFeed) ReplayLast(dispatch func(Status)) {
+	s.statusSubsMu.Lock()
+	defer s.statusSubsMu.Unlock()
+	if s.lastStatus == nil {
+		return
+	}
+	dispatch(*s.lastStatus)
+}
+
 // BeginProfileSwitch arms suppression for a switch from Connected/Connecting,
 // where the daemon emits stale Connected updates during Down's teardown then an
 // Idle before the new Up; statusStreamLoop drops those, and a synthetic
@@ -201,7 +259,7 @@ func (s *DaemonFeed) BeginProfileSwitch() {
 	s.switchLoginWatch = true
 	s.switchLoginWatchUntil = now.Add(30 * time.Second)
 	s.switchMu.Unlock()
-	s.emitter.Emit(EventStatusSnapshot, Status{Status: StatusConnecting})
+	s.pushStatus(Status{Status: StatusConnecting})
 }
 
 // CancelProfileSwitch aborts a switch midway (tray Disconnect while Connecting):
@@ -343,7 +401,7 @@ func (s *DaemonFeed) statusStreamLoop(ctx context.Context) {
 			return
 		}
 		unavailable = true
-		s.emitter.Emit(EventStatusSnapshot, Status{Status: StatusDaemonUnavailable})
+		s.pushStatus(Status{Status: StatusDaemonUnavailable})
 	}
 
 	op := func() error {
@@ -403,7 +461,7 @@ func (s *DaemonFeed) emitStatus(st Status) {
 		log.Debugf("suppressing status=%q during profile switch", st.Status)
 		return
 	}
-	s.emitter.Emit(EventStatusSnapshot, st)
+	s.pushStatus(st)
 	if triggerLogin {
 		s.emitter.Emit(EventTriggerLogin)
 	}

client/ui/services/windowmanager.go

@@ -115,6 +115,9 @@ type WindowManager struct {
 	// recenterOnShow is set only on the minimal-WM/XEmbed path, where the WM neither centers nor
 	// restores position; nil on full desktops so re-centering can't fight a user-moved window.
 	recenterOnShow func() bool
+	// showReplay fires whenever a live-but-hidden window is (re)shown, so the
+	// caller can replay the latest status snapshot into its webview.
+	showReplay func(application.Window)
 }
 
 // NewWindowManager wires the manager to the main app; translator/prefs may be nil (tests). The
@@ -174,6 +177,7 @@ func (s *WindowManager) OpenSettings(tab string) {
 	s.app.Event.Emit(EventSettingsOpen, target)
 	s.settings.Show()
 	s.settings.Focus()
+	s.notifyShown(s.settings)
 	// Re-center (minimal-WM only; see centerWhenReady).
 	s.centerWhenReady(s.settings)
 }
@@ -220,6 +224,7 @@ func (s *WindowManager) OpenBrowserLogin(uri string) {
 	s.centerOnCursorScreen(s.browserLogin)
 	s.browserLogin.Show()
 	s.browserLogin.Focus()
+	s.notifyShown(s.browserLogin)
 }
 
 // BrowserLoginWindow returns the live SSO popup, or nil. While non-nil it is the
@@ -280,6 +285,7 @@ func (s *WindowManager) OpenSessionExpiration(seconds int) {
 	s.centerOnCursorScreen(s.sessionExpiration)
 	s.sessionExpiration.Show()
 	s.sessionExpiration.Focus()
+	s.notifyShown(s.sessionExpiration)
 }
 
 func (s *WindowManager) CloseSessionExpiration() {
@@ -347,6 +353,7 @@ func (s *WindowManager) OpenInstallProgress(version string) {
 	s.installProgress.SetURL(startURL)
 	s.installProgress.Show()
 	s.installProgress.Focus()
+	s.notifyShown(s.installProgress)
 	s.centerWhenReady(s.installProgress)
 }
 
@@ -380,6 +387,7 @@ func (s *WindowManager) OpenWelcome() {
 	}
 	s.welcome.Show()
 	s.welcome.Focus()
+	s.notifyShown(s.welcome)
 	s.centerWhenReady(s.welcome)
 }
 
@@ -417,6 +425,7 @@ func (s *WindowManager) OpenError(title, message string) {
 	s.errorDialog.SetURL(startURL)
 	s.errorDialog.Show()
 	s.errorDialog.Focus()
+	s.notifyShown(s.errorDialog)
 	s.centerWhenReady(s.errorDialog)
 }
 
@@ -443,6 +452,7 @@ func (s *WindowManager) ShowMain() {
 	}
 	s.mainWindow.Show()
 	s.mainWindow.Focus()
+	s.notifyShown(s.mainWindow)
 	// Re-center (minimal-WM only; see centerWhenReady).
 	s.centerWhenReady(s.mainWindow)
 }
@@ -452,6 +462,17 @@ func (s *WindowManager) SetRecenterOnShow(pred func() bool) {
 	s.recenterOnShow = pred
 }
 
+// SetShowReplay installs the shown-window hook (see the showReplay field).
+func (s *WindowManager) SetShowReplay(fn func(application.Window)) {
+	s.showReplay = fn
+}
+
+func (s *WindowManager) notifyShown(w application.Window) {
+	if s.showReplay != nil && w != nil {
+		s.showReplay(w)
+	}
+}
+
 // centerWhenReady centers w only on minimal WMs (recenterOnShow); elsewhere it
 // returns so it never fights a user-moved window. On GTK4 an inline Center()
 // no-ops until the GdkSurface is realized (async, after Show) and InvokeAsync
@@ -572,6 +593,7 @@ func (s *WindowManager) restoreHiddenWindowsLocked() {
 			continue
 		}
 		w.Show()
+		s.notifyShown(w)
 		if w == s.mainWindow {
 			mainRestored = true
 		}

client/ui/tray.go

@@ -67,9 +67,8 @@ type Tray struct {
 	loc       *Localizer
 
 	// menu and the *Item/*Submenu fields below are reassigned by buildMenu
-	// on every relayout — touch them only with menuMu held. Exceptions:
-	// the Connect/Disconnect OnClick closures capture their own item, and
-	// refreshSessionExpiresLabel snapshots its item under menuMu.
+	// on every relayout, which destroys the replaced tree — locate items and
+	// call them with menuMu held, so a relayout can't destroy one mid-call.
 	menu       *application.Menu
 	statusItem *application.MenuItem
 	// sessionExpiresItem shows the SSO deadline as a remaining-time label,
@@ -172,7 +171,7 @@ func NewTray(app *application.App, window *application.WebviewWindow, svc TraySe
 		// in the right locale — no English flash then re-paint.
 		loc: svc.Localizer,
 	}
-	t.updater = newTrayUpdater(app, window, svc.Update, svc.Notifier, t.loc, func() { t.applyIcon() }, func() { t.relayoutMenu() })
+	t.updater = newTrayUpdater(app, window, svc.Update, svc.Notifier, t.loc, func() { t.applyIcon() }, func() { t.relayoutMenu() }, func() { t.showMainWindow() })
 	t.tray = app.SystemTray.New()
 	// Seed panel-theme detection before the first paint so the initial icon
 	// matches the panel's light/dark scheme (Linux only).
@@ -196,7 +195,7 @@ func NewTray(app *application.App, window *application.WebviewWindow, svc TraySe
 	// menu (e.g. GNOME Shell AppIndicator).
 	bindTrayClick(t)
 
-	app.Event.On(services.EventStatusSnapshot, t.onStatusEvent)
+	svc.DaemonFeed.OnStatus(t.applyStatus)
 	app.Event.On(services.EventDaemonNotification, t.onSystemEvent)
 	// Refresh the Profiles submenu on ProfileSwitcher's change event. A
 	// switch on an idle daemon drives no status transition, so without this
@@ -253,6 +252,19 @@ func (t *Tray) ShowWindow() {
 	t.window.Focus()
 }
 
+// showMainWindow brings the main window forward through the WindowManager so
+// the status replay and re-centering apply; falls back to a bare Show in tests.
+func (t *Tray) showMainWindow() {
+	if t.svc.WindowManager != nil {
+		t.svc.WindowManager.ShowMain()
+		return
+	}
+	if t.window != nil {
+		t.window.Show()
+		t.window.Focus()
+	}
+}
+
 // applyLanguage re-renders every translated surface in the Localizer's current
 // language. Wails dispatches menu/tray APIs onto the UI thread internally, so
 // calling them from the Localizer's background goroutine is safe; profileLoadMu
@@ -286,6 +298,7 @@ func (t *Tray) relayoutMenu() {
 	t.menuMu.Lock()
 	defer t.menuMu.Unlock()
 
+	old := t.menu
 	t.menu = t.buildMenu()
 
 	t.statusMu.Lock()
@@ -356,6 +369,10 @@ func (t *Tray) relayoutMenu() {
 	// Single push of the whole tree: on Linux one LayoutUpdated with fresh
 	// container ids; on darwin an NSMenu rebuild against the cached pointer.
 	t.tray.SetMenu(t.menu)
+
+	if old != nil {
+		old.Destroy()
+	}
 }
 
 func (t *Tray) buildMenu() *application.Menu {
@@ -371,13 +388,11 @@ func (t *Tray) buildMenu() *application.Menu {
 
 	menu.AddSeparator()
 
-	// The OnClick closures capture the local item because t.upItem/t.downItem
-	// are menuMu-guarded and must not be read from the click goroutine.
 	upItem := menu.Add(t.loc.T("tray.menu.connect"))
-	upItem.OnClick(func(*application.Context) { t.handleConnect(upItem) })
+	upItem.OnClick(func(*application.Context) { t.handleConnect() })
 	t.upItem = upItem
 	downItem := menu.Add(t.loc.T("tray.menu.disconnect"))
-	downItem.OnClick(func(*application.Context) { t.handleDisconnect(downItem) })
+	downItem.OnClick(func(*application.Context) { t.handleDisconnect() })
 	downItem.SetHidden(true)
 	t.downItem = downItem
 
@@ -469,9 +484,7 @@ func (t *Tray) handleQuit() {
 	t.app.Quit()
 }
 
-// handleConnect receives the clicked item from the buildMenu closure —
-// t.upItem is menuMu-guarded and must not be read here.
-func (t *Tray) handleConnect(upItem *application.MenuItem) {
+func (t *Tray) handleConnect() {
 	// NeedsLogin/SessionExpired/LoginFailed won't honor a plain Up RPC — they
 	// need the Login → WaitSSOLogin → Up sequence. Emit EventTriggerLogin so
 	// the React startLogin() (which owns the BrowserLogin popup) drives it;
@@ -485,7 +498,7 @@ func (t *Tray) handleConnect(upItem *application.MenuItem) {
 		t.app.Event.Emit(services.EventTriggerLogin)
 		return
 	}
-	upItem.SetEnabled(false)
+	t.setItemEnabled(func() *application.MenuItem { return t.upItem }, false)
 	// Arm the SSO auto-handoff: Up() is async and the daemon may flip to
 	// NeedsLogin on an SSO peer with no cached token. applyStatus consumes the
 	// flag on that transition to trigger browser-login without a second Connect
@@ -500,18 +513,25 @@ func (t *Tray) handleConnect(upItem *application.MenuItem) {
 			t.statusMu.Lock()
 			t.pendingConnectLogin = false
 			t.statusMu.Unlock()
-			upItem.SetEnabled(true)
+			t.setItemEnabled(func() *application.MenuItem { return t.upItem }, true)
 		}
 	}()
 }
 
+func (t *Tray) setItemEnabled(get func() *application.MenuItem, enabled bool) {
+	t.menuMu.Lock()
+	defer t.menuMu.Unlock()
+	if item := get(); item != nil {
+		item.SetEnabled(enabled)
+	}
+}
+
 // handleDisconnect aborts any in-flight profile switch before sending Down —
 // otherwise the switcher's queued Up would reconnect right after, making the
 // click a no-op. Also clears Peers' optimistic-Connecting guard so the daemon's
 // Idle push paints through instead of being swallowed by the suppression filter.
-// Receives the clicked item from the buildMenu closure (see handleConnect).
-func (t *Tray) handleDisconnect(downItem *application.MenuItem) {
-	downItem.SetEnabled(false)
+func (t *Tray) handleDisconnect() {
+	t.setItemEnabled(func() *application.MenuItem { return t.downItem }, false)
 	t.profileMu.Lock()
 	if t.switchCancel != nil {
 		t.switchCancel()
@@ -523,7 +543,7 @@ func (t *Tray) handleDisconnect(downItem *application.MenuItem) {
 		if err := t.svc.Connection.Down(context.Background()); err != nil {
 			log.Errorf("disconnect: %v", err)
 			t.notifyError(t.loc.T("notify.error.disconnect"))
-			downItem.SetEnabled(true)
+			t.setItemEnabled(func() *application.MenuItem { return t.downItem }, true)
 		}
 	}()
 }

client/ui/tray_profiles.go

@@ -5,6 +5,7 @@ package main
 import (
 	"context"
 	"fmt"
+	"slices"
 	"sort"
 
 	log "github.com/sirupsen/logrus"
@@ -59,10 +60,11 @@ func (t *Tray) loadConfig() {
 	t.profileMu.Unlock()
 }
 
-// loadProfiles fetches the profile list and relayouts the menu. Also called
-// from applyStatus to catch flips from another channel (CLI, autoconnect),
-// since the daemon emits no active-profile event. Full relayout (not
-// Clear()+Add()) is required for KDE/Plasma — see relayoutMenu's doc comment.
+// loadProfiles fetches the profile list and relayouts the menu when the rows
+// changed. Also called from applyStatus to catch flips from another channel
+// (CLI, autoconnect), since the daemon emits no active-profile event. Full
+// relayout (not Clear()+Add()) is required for KDE/Plasma — see relayoutMenu's
+// doc comment.
 func (t *Tray) loadProfiles() {
 	t.profileLoadMu.Lock()
 	defer t.profileLoadMu.Unlock()
@@ -80,11 +82,14 @@ func (t *Tray) loadProfiles() {
 	}
 
 	t.profilesMu.Lock()
+	changed := username != t.profilesUser || !slices.Equal(profiles, t.profiles)
 	t.profiles = profiles
 	t.profilesUser = username
 	t.profilesMu.Unlock()
 
-	t.relayoutMenu()
+	if changed {
+		t.relayoutMenu()
+	}
 }
 
 // fillProfileSubmenu paints cached profile rows into the freshly built submenu.

client/ui/tray_session.go

@@ -30,11 +30,11 @@ const (
 // handleSessionExpired notifies and brings the window forward so the frontend's /login route drives renewal.
 func (t *Tray) handleSessionExpired() {
 	t.notify(t.loc.T("notify.sessionExpired.title"), t.loc.T("notify.sessionExpired.body"), notifyIDSessionExpired)
-	if t.window != nil {
-		t.window.SetURL("/#/login")
-		t.window.Show()
-		t.window.Focus()
+	if t.window == nil {
+		return
 	}
+	t.window.SetURL("/#/login")
+	t.showMainWindow()
 }
 
 // applySessionExpiry refreshes the cached SSO deadline and reports whether it changed.
@@ -104,21 +104,19 @@ func sessionRefreshInterval(remaining time.Duration) time.Duration {
 }
 
 // refreshSessionExpiresLabel updates only the countdown label, no relayout, to avoid disturbing an open menu.
-// The item is snapshotted under menuMu since buildMenu reassigns it on every relayout.
+// menuMu is held across the SetLabel so a relayout can't destroy the item mid-call.
 func (t *Tray) refreshSessionExpiresLabel() {
-	t.menuMu.Lock()
-	item := t.sessionExpiresItem
-	t.menuMu.Unlock()
-	if item == nil {
-		return
-	}
 	t.sessionMu.Lock()
 	deadline := t.sessionExpiresAt
 	t.sessionMu.Unlock()
 	if deadline.IsZero() {
 		return
 	}
-	item.SetLabel(t.sessionRowLabel(deadline))
+	t.menuMu.Lock()
+	defer t.menuMu.Unlock()
+	if t.sessionExpiresItem != nil {
+		t.sessionExpiresItem.SetLabel(t.sessionRowLabel(deadline))
+	}
 }
 
 func (t *Tray) sessionRowLabel(deadline time.Time) string {
@@ -310,8 +308,7 @@ func (t *Tray) openSessionExtendFlow() {
 	if seconds <= 0 {
 		if t.window != nil {
 			t.window.SetURL("/#/login")
-			t.window.Show()
-			t.window.Focus()
+			t.showMainWindow()
 		}
 		return
 	}

client/ui/tray_status.go

@@ -5,19 +5,9 @@ package main
 import (
 	"strings"
 
-	"github.com/wailsapp/wails/v3/pkg/application"
-
 	"github.com/netbirdio/netbird/client/ui/services"
 )
 
-func (t *Tray) onStatusEvent(ev *application.CustomEvent) {
-	st, ok := ev.Data.(services.Status)
-	if !ok {
-		return
-	}
-	t.applyStatus(st)
-}
-
 // applyStatus repaints the tray from a daemon snapshot. Icon refresh is skipped
 // when no icon-relevant input changed: the daemon emits rapid SubscribeStatus
 // bursts during health probes that would otherwise spam Shell_NotifyIcon.

client/ui/tray_update.go

@@ -28,6 +28,9 @@ type trayUpdater struct {
 	// About submenu, which KDE/Plasma caches on first open and never re-fetches
 	// on a plain SetLabel/SetHidden — only a relayout (fresh submenu ids) repaints.
 	onMenuChange func()
+	// showMain brings the main window forward via the WindowManager (status
+	// replay + centering); nil falls back to a bare window.Show.
+	showMain func()
 
 	mu                 sync.Mutex
 	item               *application.MenuItem
@@ -36,7 +39,7 @@ type trayUpdater struct {
 	progressWindowOpen bool
 }
 
-func newTrayUpdater(app *application.App, window *application.WebviewWindow, update *services.Update, notifier *notifications.NotificationService, loc *Localizer, onIconChange func(), onMenuChange func()) *trayUpdater {
+func newTrayUpdater(app *application.App, window *application.WebviewWindow, update *services.Update, notifier *notifications.NotificationService, loc *Localizer, onIconChange, onMenuChange, showMain func()) *trayUpdater {
 	u := &trayUpdater{
 		app:          app,
 		window:       window,
@@ -45,6 +48,7 @@ func newTrayUpdater(app *application.App, window *application.WebviewWindow, upd
 		loc:          loc,
 		onIconChange: onIconChange,
 		onMenuChange: onMenuChange,
+		showMain:     showMain,
 	}
 	app.Event.On(updater.EventStateChanged, u.onStateEvent)
 	// Seed from cached state to cover an event that fired before wiring completed.
@@ -193,6 +197,10 @@ func (u *trayUpdater) openProgressWindow(version string) {
 		url += "?version=" + version
 	}
 	u.window.SetURL(url)
+	if u.showMain != nil {
+		u.showMain()
+		return
+	}
 	u.window.Show()
 	u.window.Focus()
 }

management/cmd/management.go

@@ -23,7 +23,6 @@ import (
 	"github.com/netbirdio/netbird/management/server/types"
 
 	"github.com/netbirdio/netbird/formatter/hook"
-	agentnetworkpricing "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
 	"github.com/netbirdio/netbird/management/internals/server"
 	nbconfig "github.com/netbirdio/netbird/management/internals/server/config"
 	nbdomain "github.com/netbirdio/netbird/shared/management/domain"
@@ -113,23 +112,6 @@ var (
 				mgmtSingleAccModeDomain = ""
 			}
 
-			// Load the management-side LLM pricing defaults file: an
-			// explicitly configured path is required to load (a typo must
-			// fail startup — the operator believes those rates are live);
-			// otherwise <datadir>/defaults_llm_pricing.yaml is probed and
-			// may be absent (compiled-in defaults serve). Either way the
-			// path stays watched: the reloader picks up edits — and the
-			// file appearing later — without a restart.
-			pricingPath := config.AgentNetwork.PricingDefaultsFile
-			pricingRequired := pricingPath != ""
-			if !pricingRequired {
-				pricingPath = filepath.Join(config.Datadir, agentnetworkpricing.DefaultFileName)
-			}
-			if err := agentnetworkpricing.LoadFile(pricingPath, pricingRequired); err != nil {
-				return fmt.Errorf("load agent-network pricing defaults: %v", err)
-			}
-			agentnetworkpricing.StartReloader(ctx, agentnetworkpricing.ReloadInterval)
-
 			srv := newServer(&server.Config{
 				NbConfig:                    config,
 				DNSDomain:                   dnsDomain,

management/internals/modules/agentnetwork/catalog/catalog.go

@@ -7,28 +7,12 @@ package catalog
 import "github.com/netbirdio/netbird/shared/management/http/api"
 
 // Model is the in-memory representation of a catalog model.
-//
-// The three cache rates mirror the proxy cost meter's Entry semantics
-// (USD per 1k tokens; 0 = no rate configured, that bucket bills at
-// InputPer1k):
-//   - CachedInputPer1k: OpenAI-shape rate for cached prompt tokens
-//     (a SUBSET of input tokens). Typically 0.1-0.5x input.
-//   - CacheReadPer1k / CacheCreationPer1k: Anthropic-shape rates for
-//     the two ADDITIVE prompt-cache buckets. Typically 0.1x / 1.25x
-//     input.
-//
-// The catalog is the single default-pricing source: the agentnetwork
-// pricing package folds these models into per-surface tables that the
-// synthesizer ships to the proxy's cost_meter.
 type Model struct {
-	ID                 string
-	Label              string
-	InputPer1k         float64
-	OutputPer1k        float64
-	CachedInputPer1k   float64
-	CacheReadPer1k     float64
-	CacheCreationPer1k float64
-	ContextWindow      int
+	ID            string
+	Label         string
+	InputPer1k    float64
+	OutputPer1k   float64
+	ContextWindow int
 }
 
 // ProviderKind groups catalog entries for UI presentation. The split
@@ -81,17 +65,6 @@ type Provider struct {
 	// surface — the proxy middleware then falls back to URL sniffing
 	// or skips request-side enrichment.
 	ParserID string
-	// PricingSurfaces names the cost-meter pricing surfaces this
-	// provider's Models are priced under ("openai", "anthropic",
-	// "bedrock" — the llm.Parser surface the request parser stamps as
-	// llm.provider at billing time). NOT derivable from ParserID:
-	// bedrock_api and vertex_ai_api leave ParserID empty (URL-sniffed)
-	// yet price under "bedrock" / "anthropic", and kimi_api serves two
-	// body shapes so it prices under both. Nil for gateway/custom
-	// entries, which declare no models. Same (surface, model) pair
-	// contributed by two providers must carry identical rates — the
-	// pricing package's tests enforce that.
-	PricingSurfaces []string
 	// IdentityInjection, when non-nil, instructs the proxy to stamp
 	// the caller's NetBird identity onto upstream requests under the
 	// configured header names. Used for gateways like LiteLLM that
@@ -246,7 +219,6 @@ var providers = []Provider{
 		DefaultContentType: "application/json",
 		BrandColor:         "#10A37F",
 		ParserID:           "openai",
-		PricingSurfaces:    []string{"openai"},
 		// Pricing + context windows cross-checked against LiteLLM's
 		// model_prices_and_context_window.json. Notable corrections from
 		// earlier values: o4-mini repriced from $4/$16 to $1.10/$4.40
@@ -254,20 +226,20 @@ var providers = []Provider{
 		// family context windows split between 1.05M for full-size
 		// models and 272K for mini/nano/codex variants.
 		Models: []Model{
-			{ID: "gpt-5.5", Label: "GPT-5.5", InputPer1k: 0.005, OutputPer1k: 0.030, CachedInputPer1k: 0.0005, ContextWindow: 1050000},
-			{ID: "gpt-5.5-pro", Label: "GPT-5.5 Pro", InputPer1k: 0.030, OutputPer1k: 0.180, CachedInputPer1k: 0.003, ContextWindow: 1050000},
-			{ID: "gpt-5.4", Label: "GPT-5.4", InputPer1k: 0.0025, OutputPer1k: 0.015, CachedInputPer1k: 0.00025, ContextWindow: 1050000},
-			{ID: "gpt-5.4-pro", Label: "GPT-5.4 Pro", InputPer1k: 0.030, OutputPer1k: 0.180, CachedInputPer1k: 0.003, ContextWindow: 1050000},
-			{ID: "gpt-5.4-mini", Label: "GPT-5.4 Mini", InputPer1k: 0.00075, OutputPer1k: 0.0045, CachedInputPer1k: 0.000075, ContextWindow: 272000},
-			{ID: "gpt-5.4-nano", Label: "GPT-5.4 Nano", InputPer1k: 0.0002, OutputPer1k: 0.00125, CachedInputPer1k: 0.00002, ContextWindow: 272000},
-			{ID: "gpt-5.3-codex", Label: "GPT-5.3 Codex", InputPer1k: 0.00175, OutputPer1k: 0.014, CachedInputPer1k: 0.000175, ContextWindow: 272000},
-			{ID: "gpt-5.3-chat-latest", Label: "GPT-5.3 Chat", InputPer1k: 0.00175, OutputPer1k: 0.014, CachedInputPer1k: 0.000175, ContextWindow: 128000},
-			{ID: "o4-mini", Label: "o4-mini", InputPer1k: 0.0011, OutputPer1k: 0.0044, CachedInputPer1k: 0.000275, ContextWindow: 200000},
-			{ID: "gpt-4.1", Label: "GPT-4.1", InputPer1k: 0.002, OutputPer1k: 0.008, CachedInputPer1k: 0.0005, ContextWindow: 1047576},
-			{ID: "gpt-4.1-mini", Label: "GPT-4.1 mini", InputPer1k: 0.0004, OutputPer1k: 0.0016, CachedInputPer1k: 0.0001, ContextWindow: 1047576},
-			{ID: "gpt-4.1-nano", Label: "GPT-4.1 nano", InputPer1k: 0.0001, OutputPer1k: 0.0004, CachedInputPer1k: 0.000025, ContextWindow: 1047576},
-			{ID: "gpt-4o", Label: "GPT-4o", InputPer1k: 0.0025, OutputPer1k: 0.010, CachedInputPer1k: 0.00125, ContextWindow: 128000},
-			{ID: "gpt-4o-mini", Label: "GPT-4o mini", InputPer1k: 0.00015, OutputPer1k: 0.0006, CachedInputPer1k: 0.000075, ContextWindow: 128000},
+			{ID: "gpt-5.5", Label: "GPT-5.5", InputPer1k: 0.005, OutputPer1k: 0.030, ContextWindow: 1050000},
+			{ID: "gpt-5.5-pro", Label: "GPT-5.5 Pro", InputPer1k: 0.030, OutputPer1k: 0.180, ContextWindow: 1050000},
+			{ID: "gpt-5.4", Label: "GPT-5.4", InputPer1k: 0.0025, OutputPer1k: 0.015, ContextWindow: 1050000},
+			{ID: "gpt-5.4-pro", Label: "GPT-5.4 Pro", InputPer1k: 0.030, OutputPer1k: 0.180, ContextWindow: 1050000},
+			{ID: "gpt-5.4-mini", Label: "GPT-5.4 Mini", InputPer1k: 0.00075, OutputPer1k: 0.0045, ContextWindow: 272000},
+			{ID: "gpt-5.4-nano", Label: "GPT-5.4 Nano", InputPer1k: 0.0002, OutputPer1k: 0.00125, ContextWindow: 272000},
+			{ID: "gpt-5.3-codex", Label: "GPT-5.3 Codex", InputPer1k: 0.00175, OutputPer1k: 0.014, ContextWindow: 272000},
+			{ID: "gpt-5.3-chat-latest", Label: "GPT-5.3 Chat", InputPer1k: 0.00175, OutputPer1k: 0.014, ContextWindow: 128000},
+			{ID: "o4-mini", Label: "o4-mini", InputPer1k: 0.0011, OutputPer1k: 0.0044, ContextWindow: 200000},
+			{ID: "gpt-4.1", Label: "GPT-4.1", InputPer1k: 0.002, OutputPer1k: 0.008, ContextWindow: 1047576},
+			{ID: "gpt-4.1-mini", Label: "GPT-4.1 mini", InputPer1k: 0.0004, OutputPer1k: 0.0016, ContextWindow: 1047576},
+			{ID: "gpt-4.1-nano", Label: "GPT-4.1 nano", InputPer1k: 0.0001, OutputPer1k: 0.0004, ContextWindow: 1047576},
+			{ID: "gpt-4o", Label: "GPT-4o", InputPer1k: 0.0025, OutputPer1k: 0.010, ContextWindow: 128000},
+			{ID: "gpt-4o-mini", Label: "GPT-4o mini", InputPer1k: 0.00015, OutputPer1k: 0.0006, ContextWindow: 128000},
 			{ID: "gpt-4-turbo", Label: "GPT-4 Turbo", InputPer1k: 0.01, OutputPer1k: 0.03, ContextWindow: 128000},
 			{ID: "gpt-3.5-turbo", Label: "GPT-3.5 Turbo", InputPer1k: 0.0005, OutputPer1k: 0.0015, ContextWindow: 16385},
 			{ID: "text-embedding-3-large", Label: "text-embedding-3-large", InputPer1k: 0.00013, OutputPer1k: 0, ContextWindow: 8191},
@@ -285,7 +257,6 @@ var providers = []Provider{
 		DefaultContentType: "application/json",
 		BrandColor:         "#D97757",
 		ParserID:           "anthropic",
-		PricingSurfaces:    []string{"anthropic"},
 		// Per Anthropic's current model lineup. Pricing in USD per 1k
 		// tokens. Context windows: 4.6+ family is 1M; Haiku 4.5 stays at
 		// 200K. claude-3-7-sonnet and claude-3-5-haiku retired
@@ -296,14 +267,14 @@ var providers = []Provider{
 		// account to be on >= 30-day data retention or all requests
 		// 400.
 		Models: []Model{
-			{ID: "claude-fable-5", Label: "Claude Fable 5", InputPer1k: 0.010, OutputPer1k: 0.050, CacheReadPer1k: 0.001, CacheCreationPer1k: 0.0125, ContextWindow: 1000000},
-			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
-			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
-			{ID: "claude-opus-4-6", Label: "Claude Opus 4.6", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
-			{ID: "claude-opus-4-1", Label: "Claude Opus 4.1 (deprecated, retires 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, CacheReadPer1k: 0.0015, CacheCreationPer1k: 0.01875, ContextWindow: 200000},
-			{ID: "claude-sonnet-4-6", Label: "Claude Sonnet 4.6", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 1000000},
-			{ID: "claude-sonnet-4-5", Label: "Claude Sonnet 4.5", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 200000},
-			{ID: "claude-haiku-4-5", Label: "Claude Haiku 4.5", InputPer1k: 0.001, OutputPer1k: 0.005, CacheReadPer1k: 0.0001, CacheCreationPer1k: 0.00125, ContextWindow: 200000},
+			{ID: "claude-fable-5", Label: "Claude Fable 5", InputPer1k: 0.010, OutputPer1k: 0.050, ContextWindow: 1000000},
+			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
+			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
+			{ID: "claude-opus-4-6", Label: "Claude Opus 4.6", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
+			{ID: "claude-opus-4-1", Label: "Claude Opus 4.1 (deprecated, retires 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, ContextWindow: 200000},
+			{ID: "claude-sonnet-4-6", Label: "Claude Sonnet 4.6", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
+			{ID: "claude-sonnet-4-5", Label: "Claude Sonnet 4.5", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 200000},
+			{ID: "claude-haiku-4-5", Label: "Claude Haiku 4.5", InputPer1k: 0.001, OutputPer1k: 0.005, ContextWindow: 200000},
 		},
 	},
 	{
@@ -317,19 +288,18 @@ var providers = []Provider{
 		DefaultContentType: "application/json",
 		BrandColor:         "#0078D4",
 		ParserID:           "openai",
-		PricingSurfaces:    []string{"openai"},
 		// Mirrors openai_api pricing — Azure resells OpenAI models at the
 		// same per-token rates, just under different deployment names.
 		Models: []Model{
-			{ID: "gpt-5.5", Label: "GPT-5.5 (Azure)", InputPer1k: 0.005, OutputPer1k: 0.030, CachedInputPer1k: 0.0005, ContextWindow: 1050000},
-			{ID: "gpt-5.4", Label: "GPT-5.4 (Azure)", InputPer1k: 0.0025, OutputPer1k: 0.015, CachedInputPer1k: 0.00025, ContextWindow: 1050000},
-			{ID: "gpt-5.4-mini", Label: "GPT-5.4 Mini (Azure)", InputPer1k: 0.00075, OutputPer1k: 0.0045, CachedInputPer1k: 0.000075, ContextWindow: 272000},
-			{ID: "gpt-5.4-nano", Label: "GPT-5.4 Nano (Azure)", InputPer1k: 0.0002, OutputPer1k: 0.00125, CachedInputPer1k: 0.00002, ContextWindow: 272000},
-			{ID: "o4-mini", Label: "o4-mini (Azure)", InputPer1k: 0.0011, OutputPer1k: 0.0044, CachedInputPer1k: 0.000275, ContextWindow: 200000},
-			{ID: "gpt-4.1", Label: "GPT-4.1 (Azure)", InputPer1k: 0.002, OutputPer1k: 0.008, CachedInputPer1k: 0.0005, ContextWindow: 1047576},
-			{ID: "gpt-4.1-mini", Label: "GPT-4.1 mini (Azure)", InputPer1k: 0.0004, OutputPer1k: 0.0016, CachedInputPer1k: 0.0001, ContextWindow: 1047576},
-			{ID: "gpt-4o", Label: "GPT-4o (Azure)", InputPer1k: 0.0025, OutputPer1k: 0.010, CachedInputPer1k: 0.00125, ContextWindow: 128000},
-			{ID: "gpt-4o-mini", Label: "GPT-4o mini (Azure)", InputPer1k: 0.00015, OutputPer1k: 0.0006, CachedInputPer1k: 0.000075, ContextWindow: 128000},
+			{ID: "gpt-5.5", Label: "GPT-5.5 (Azure)", InputPer1k: 0.005, OutputPer1k: 0.030, ContextWindow: 1050000},
+			{ID: "gpt-5.4", Label: "GPT-5.4 (Azure)", InputPer1k: 0.0025, OutputPer1k: 0.015, ContextWindow: 1050000},
+			{ID: "gpt-5.4-mini", Label: "GPT-5.4 Mini (Azure)", InputPer1k: 0.00075, OutputPer1k: 0.0045, ContextWindow: 272000},
+			{ID: "gpt-5.4-nano", Label: "GPT-5.4 Nano (Azure)", InputPer1k: 0.0002, OutputPer1k: 0.00125, ContextWindow: 272000},
+			{ID: "o4-mini", Label: "o4-mini (Azure)", InputPer1k: 0.0011, OutputPer1k: 0.0044, ContextWindow: 200000},
+			{ID: "gpt-4.1", Label: "GPT-4.1 (Azure)", InputPer1k: 0.002, OutputPer1k: 0.008, ContextWindow: 1047576},
+			{ID: "gpt-4.1-mini", Label: "GPT-4.1 mini (Azure)", InputPer1k: 0.0004, OutputPer1k: 0.0016, ContextWindow: 1047576},
+			{ID: "gpt-4o", Label: "GPT-4o (Azure)", InputPer1k: 0.0025, OutputPer1k: 0.010, ContextWindow: 128000},
+			{ID: "gpt-4o-mini", Label: "GPT-4o mini (Azure)", InputPer1k: 0.00015, OutputPer1k: 0.0006, ContextWindow: 128000},
 			{ID: "gpt-35-turbo", Label: "GPT-3.5 Turbo (Azure)", InputPer1k: 0.0005, OutputPer1k: 0.0015, ContextWindow: 16385},
 		},
 	},
@@ -343,9 +313,6 @@ var providers = []Provider{
 		AuthHeaderTemplate: "Bearer ${API_KEY}",
 		DefaultContentType: "application/json",
 		BrandColor:         "#FF9900",
-		// ParserID stays empty (path-style dispatch via IsBedrockPathStyle);
-		// the request parser meters these under the "bedrock" surface.
-		PricingSurfaces: []string{"bedrock"},
 		// Anthropic models on Bedrock take the anthropic.* prefix and
 		// follow the same lineup / pricing as the first-party Anthropic
 		// catalog entry above. claude-3-7-sonnet and claude-3-5-haiku
@@ -355,13 +322,13 @@ var providers = []Provider{
 		// Llama 3.3 70B entry kept unchanged — LiteLLM tracks only
 		// per-region Llama 3 entries; standalone 3.3 not yet listed.
 		Models: []Model{
-			{ID: "anthropic.claude-opus-4-8", Label: "Claude Opus 4.8 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
-			{ID: "anthropic.claude-opus-4-7", Label: "Claude Opus 4.7 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
-			{ID: "anthropic.claude-opus-4-6", Label: "Claude Opus 4.6 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
-			{ID: "anthropic.claude-opus-4-1", Label: "Claude Opus 4.1 (Bedrock, deprecated 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, CacheReadPer1k: 0.0015, CacheCreationPer1k: 0.01875, ContextWindow: 200000},
-			{ID: "anthropic.claude-sonnet-4-6", Label: "Claude Sonnet 4.6 (Bedrock)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 1000000},
-			{ID: "anthropic.claude-sonnet-4-5", Label: "Claude Sonnet 4.5 (Bedrock)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 200000},
-			{ID: "anthropic.claude-haiku-4-5", Label: "Claude Haiku 4.5 (Bedrock)", InputPer1k: 0.001, OutputPer1k: 0.005, CacheReadPer1k: 0.0001, CacheCreationPer1k: 0.00125, ContextWindow: 200000},
+			{ID: "anthropic.claude-opus-4-8", Label: "Claude Opus 4.8 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
+			{ID: "anthropic.claude-opus-4-7", Label: "Claude Opus 4.7 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
+			{ID: "anthropic.claude-opus-4-6", Label: "Claude Opus 4.6 (Bedrock)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
+			{ID: "anthropic.claude-opus-4-1", Label: "Claude Opus 4.1 (Bedrock, deprecated 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, ContextWindow: 200000},
+			{ID: "anthropic.claude-sonnet-4-6", Label: "Claude Sonnet 4.6 (Bedrock)", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
+			{ID: "anthropic.claude-sonnet-4-5", Label: "Claude Sonnet 4.5 (Bedrock)", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 200000},
+			{ID: "anthropic.claude-haiku-4-5", Label: "Claude Haiku 4.5 (Bedrock)", InputPer1k: 0.001, OutputPer1k: 0.005, ContextWindow: 200000},
 			{ID: "meta.llama3-3-70b-instruct", Label: "Llama 3.3 70B (Bedrock)", InputPer1k: 0.00072, OutputPer1k: 0.00072, ContextWindow: 128000},
 			{ID: "amazon.nova-2-lite", Label: "Amazon Nova 2 Lite (Bedrock, preview)", InputPer1k: 0.0003, OutputPer1k: 0.0025, ContextWindow: 1000000},
 			{ID: "amazon.nova-pro", Label: "Amazon Nova Pro (Bedrock)", InputPer1k: 0.0008, OutputPer1k: 0.0032, ContextWindow: 300000},
@@ -391,10 +358,6 @@ var providers = []Provider{
 		AuthHeaderTemplate: "Bearer ${API_KEY}",
 		DefaultContentType: "application/json",
 		BrandColor:         "#4285F4",
-		// ParserID stays empty (path-style dispatch via IsVertexPathStyle);
-		// Anthropic-on-Vertex requests are metered under the "anthropic"
-		// surface with the bare, unversioned model id.
-		PricingSurfaces: []string{"anthropic"},
 		// Vertex carries the model in the URL path and authenticates with a
 		// service-account-minted OAuth token (api_key = "keyfile::<base64 SA>").
 		// Only Anthropic-on-Vertex is metered today: the request parser maps the
@@ -406,14 +369,14 @@ var providers = []Provider{
 		// exists — the router denies unmeterable publishers rather than forward
 		// them uncounted.
 		Models: []Model{
-			{ID: "claude-fable-5", Label: "Claude Fable 5 (Vertex)", InputPer1k: 0.010, OutputPer1k: 0.050, CacheReadPer1k: 0.001, CacheCreationPer1k: 0.0125, ContextWindow: 1000000},
-			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
-			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
-			{ID: "claude-opus-4-6", Label: "Claude Opus 4.6 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625, ContextWindow: 1000000},
-			{ID: "claude-opus-4-1", Label: "Claude Opus 4.1 (Vertex, deprecated 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, CacheReadPer1k: 0.0015, CacheCreationPer1k: 0.01875, ContextWindow: 200000},
-			{ID: "claude-sonnet-4-6", Label: "Claude Sonnet 4.6 (Vertex)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 1000000},
-			{ID: "claude-sonnet-4-5", Label: "Claude Sonnet 4.5 (Vertex)", InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003, CacheCreationPer1k: 0.00375, ContextWindow: 200000},
-			{ID: "claude-haiku-4-5", Label: "Claude Haiku 4.5 (Vertex)", InputPer1k: 0.001, OutputPer1k: 0.005, CacheReadPer1k: 0.0001, CacheCreationPer1k: 0.00125, ContextWindow: 200000},
+			{ID: "claude-fable-5", Label: "Claude Fable 5 (Vertex)", InputPer1k: 0.010, OutputPer1k: 0.050, ContextWindow: 1000000},
+			{ID: "claude-opus-4-8", Label: "Claude Opus 4.8 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
+			{ID: "claude-opus-4-7", Label: "Claude Opus 4.7 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
+			{ID: "claude-opus-4-6", Label: "Claude Opus 4.6 (Vertex)", InputPer1k: 0.005, OutputPer1k: 0.025, ContextWindow: 1000000},
+			{ID: "claude-opus-4-1", Label: "Claude Opus 4.1 (Vertex, deprecated 2026-08-05)", InputPer1k: 0.015, OutputPer1k: 0.075, ContextWindow: 200000},
+			{ID: "claude-sonnet-4-6", Label: "Claude Sonnet 4.6 (Vertex)", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
+			{ID: "claude-sonnet-4-5", Label: "Claude Sonnet 4.5 (Vertex)", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 200000},
+			{ID: "claude-haiku-4-5", Label: "Claude Haiku 4.5 (Vertex)", InputPer1k: 0.001, OutputPer1k: 0.005, ContextWindow: 200000},
 		},
 	},
 	{
@@ -427,7 +390,6 @@ var providers = []Provider{
 		DefaultContentType: "application/json",
 		BrandColor:         "#FF7000",
 		ParserID:           "openai",
-		PricingSurfaces:    []string{"openai"},
 		// Pricing + context windows cross-checked against LiteLLM. Key
 		// gotchas the marketing page hides:
 		//   - `mistral-medium-latest` aliases to Medium 3.1 ($0.40/$2),
@@ -486,10 +448,6 @@ var providers = []Provider{
 		// model id "k3") is account-bound seat licensing rather than a
 		// meterable platform key, so it's deliberately not the default.
 		ParserID: "",
-		// Both body shapes are metered: /v1/chat/completions under
-		// "openai", /anthropic/v1/messages under "anthropic" — so the
-		// K3 entry is priced on both surfaces.
-		PricingSurfaces: []string{"openai", "anthropic"},
 		// Pricing per Moonshot's platform rates at K3 launch (July 2026):
 		// $3/$15 per MTok with $0.30 cached input, flat across the 1M-token
 		// window. kimi-k3 is the ONLY model the platform serves newer
@@ -500,14 +458,7 @@ var providers = []Provider{
 		// The consumer app's "K3 Swarm Max" mode is not an API SKU, so it
 		// doesn't appear here.
 		Models: []Model{
-			// Carries both cache shapes: Moonshot reports cache hits
-			// OpenAI-style on /v1/chat/completions (CachedInputPer1k)
-			// and Anthropic-style on /anthropic/v1/messages
-			// (CacheReadPer1k) — $0.30/MTok either way. Each surface's
-			// cost formula reads only its own field, so the superset
-			// entry prices both endpoints correctly. No cache-creation
-			// rate published; writes bill at the input rate.
-			{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, CachedInputPer1k: 0.0003, CacheReadPer1k: 0.0003, ContextWindow: 1000000},
+			{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
 		},
 	},
 	{
@@ -807,28 +758,13 @@ func IsBedrockPathStyle(providerID string) bool {
 func (p Provider) ToAPIResponse() api.AgentNetworkCatalogProvider {
 	models := make([]api.AgentNetworkCatalogModel, 0, len(p.Models))
 	for _, m := range p.Models {
-		am := api.AgentNetworkCatalogModel{
+		models = append(models, api.AgentNetworkCatalogModel{
 			Id:            m.ID,
 			Label:         m.Label,
 			InputPer1k:    m.InputPer1k,
 			OutputPer1k:   m.OutputPer1k,
 			ContextWindow: m.ContextWindow,
-		}
-		// Cache rates are emitted only when configured so the dashboard
-		// can prefill them; 0 stays off the wire (absent = no rate).
-		if m.CachedInputPer1k > 0 {
-			v := m.CachedInputPer1k
-			am.CachedInputPer1k = &v
-		}
-		if m.CacheReadPer1k > 0 {
-			v := m.CacheReadPer1k
-			am.CacheReadPer1k = &v
-		}
-		if m.CacheCreationPer1k > 0 {
-			v := m.CacheCreationPer1k
-			am.CacheCreationPer1k = &v
-		}
-		models = append(models, am)
+		})
 	}
 	kind := api.AgentNetworkCatalogProviderKindProvider
 	switch p.Kind {
@@ -848,10 +784,6 @@ func (p Provider) ToAPIResponse() api.AgentNetworkCatalogProvider {
 		BrandColor:         p.BrandColor,
 		Models:             models,
 	}
-	if len(p.PricingSurfaces) > 0 {
-		surfaces := append([]string(nil), p.PricingSurfaces...)
-		resp.PricingSurfaces = &surfaces
-	}
 	if len(p.ExtraHeaders) > 0 {
 		extras := make([]api.AgentNetworkCatalogExtraHeader, 0, len(p.ExtraHeaders))
 		for _, h := range p.ExtraHeaders {

management/internals/modules/agentnetwork/handlers/providers_handler.go

@@ -7,7 +7,6 @@ package handlers
 
 import (
 	"encoding/json"
-	"math"
 	"net/http"
 	"net/url"
 	"strings"
@@ -16,7 +15,6 @@ import (
 
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 	nbcontext "github.com/netbirdio/netbird/management/server/context"
 	"github.com/netbirdio/netbird/shared/management/http/api"
@@ -54,45 +52,11 @@ func (h *handler) getCatalogProviders(w http.ResponseWriter, r *http.Request) {
 	entries := catalog.All()
 	out := make([]api.AgentNetworkCatalogProvider, 0, len(entries))
 	for _, e := range entries {
-		resp := e.ToAPIResponse()
-		applyDefaultPricing(e, &resp)
-		out = append(out, resp)
+		out = append(out, e.ToAPIResponse())
 	}
 	util.WriteJSONObject(r.Context(), w, out)
 }
 
-// applyDefaultPricing overwrites the catalog response's model rates with
-// the LIVE default pricing table, which may differ from the compiled-in
-// catalog rates when the operator provides a defaults_llm_pricing.yaml.
-// This keeps the dashboard's model-row prefill identical to what the
-// proxy will actually bill — the same table the synthesizer ships.
-func applyDefaultPricing(cp catalog.Provider, resp *api.AgentNetworkCatalogProvider) {
-	if len(cp.PricingSurfaces) == 0 {
-		return
-	}
-	for i := range resp.Models {
-		m := &resp.Models[i]
-		e, ok := pricing.LookupDefault(cp.PricingSurfaces, m.Id)
-		if !ok {
-			continue
-		}
-		m.InputPer1k = e.InputPer1k
-		m.OutputPer1k = e.OutputPer1k
-		m.CachedInputPer1k = positiveRatePtr(e.CachedInputPer1k)
-		m.CacheReadPer1k = positiveRatePtr(e.CacheReadPer1k)
-		m.CacheCreationPer1k = positiveRatePtr(e.CacheCreationPer1k)
-	}
-}
-
-// positiveRatePtr renders a cache rate for the API: absent (nil) when
-// unset, matching the catalog response convention.
-func positiveRatePtr(v float64) *float64 {
-	if v <= 0 {
-		return nil
-	}
-	return &v
-}
-
 func (h *handler) getAllProviders(w http.ResponseWriter, r *http.Request) {
 	userAuth, err := nbcontext.GetUserAuthFromContext(r.Context())
 	if err != nil {
@@ -249,38 +213,5 @@ func validate(req *api.AgentNetworkProviderRequest, requireAPIKey bool) error {
 	if requireAPIKey && (req.ApiKey == nil || strings.TrimSpace(*req.ApiKey) == "") {
 		return status.Errorf(status.InvalidArgument, "api_key is required")
 	}
-	if req.Models != nil {
-		for i, m := range *req.Models {
-			if err := validateModel(i, m); err != nil {
-				return err
-			}
-		}
-	}
-	return nil
-}
-
-// validateModel is the single ingress guard for operator-entered pricing:
-// these rates are synthesized into the proxy's cost_meter config verbatim,
-// and a negative or non-finite rate there would poison every cost the
-// proxy records, so reject at the API boundary.
-func validateModel(i int, m api.AgentNetworkProviderModel) error {
-	if strings.TrimSpace(m.Id) == "" {
-		return status.Errorf(status.InvalidArgument, "models[%d]: id is required", i)
-	}
-	rates := map[string]*float64{
-		"input_per_1k":          &m.InputPer1k,
-		"output_per_1k":         &m.OutputPer1k,
-		"cached_input_per_1k":   m.CachedInputPer1k,
-		"cache_read_per_1k":     m.CacheReadPer1k,
-		"cache_creation_per_1k": m.CacheCreationPer1k,
-	}
-	for field, v := range rates {
-		if v == nil {
-			continue
-		}
-		if *v < 0 || math.IsNaN(*v) || math.IsInf(*v, 0) {
-			return status.Errorf(status.InvalidArgument, "models[%d] (%s): %s must be a finite, non-negative USD rate", i, m.Id, field)
-		}
-	}
 	return nil
 }

management/internals/modules/agentnetwork/handlers/providers_handler_test.go

@@ -1,53 +0,0 @@
-package handlers
-
-import (
-	"math"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/shared/management/http/api"
-)
-
-func f(v float64) *float64 { return &v }
-
-// TestValidate_ModelRates guards the single ingress point for operator-entered
-// pricing. These rates flow verbatim into the proxy's cost_meter config at
-// synthesis time; the proxy treats a bad rate as a chain-build failure, so
-// rejecting here is what keeps an account's gateway from going down.
-func TestValidate_ModelRates(t *testing.T) {
-	base := func(models ...api.AgentNetworkProviderModel) *api.AgentNetworkProviderRequest {
-		key := "sk-test"
-		return &api.AgentNetworkProviderRequest{
-			ProviderId:  "openai_api",
-			Name:        "OpenAI",
-			UpstreamUrl: "https://api.openai.com",
-			ApiKey:      &key,
-			Models:      &models,
-		}
-	}
-
-	valid := api.AgentNetworkProviderModel{
-		Id: "gpt-4o", InputPer1k: 0.0025, OutputPer1k: 0.01,
-		CachedInputPer1k: f(0.00125),
-	}
-	require.NoError(t, validate(base(valid), true), "finite non-negative rates must pass")
-
-	zeroRates := api.AgentNetworkProviderModel{Id: "self-hosted-llama", InputPer1k: 0, OutputPer1k: 0}
-	require.NoError(t, validate(base(zeroRates), true), "explicit zero prices are allowed (free / self-hosted models)")
-
-	cases := map[string]api.AgentNetworkProviderModel{
-		"empty id":           {Id: "  ", InputPer1k: 0.001, OutputPer1k: 0.002},
-		"negative input":     {Id: "m", InputPer1k: -0.001, OutputPer1k: 0.002},
-		"negative output":    {Id: "m", InputPer1k: 0.001, OutputPer1k: -0.002},
-		"NaN input":          {Id: "m", InputPer1k: math.NaN(), OutputPer1k: 0.002},
-		"Inf output":         {Id: "m", InputPer1k: 0.001, OutputPer1k: math.Inf(1)},
-		"negative cached":    {Id: "m", InputPer1k: 0.001, OutputPer1k: 0.002, CachedInputPer1k: f(-1)},
-		"NaN cache read":     {Id: "m", InputPer1k: 0.001, OutputPer1k: 0.002, CacheReadPer1k: f(math.NaN())},
-		"Inf cache creation": {Id: "m", InputPer1k: 0.001, OutputPer1k: 0.002, CacheCreationPer1k: f(math.Inf(-1))},
-	}
-	for name, m := range cases {
-		assert.Error(t, validate(base(m), true), "case %q must be rejected", name)
-	}
-}

management/internals/modules/agentnetwork/pricing/defaults.go

@@ -1,156 +0,0 @@
-// Package pricing builds the default LLM pricing table the synthesizer
-// ships to the proxy's cost_meter middleware. The catalog is the single
-// source of default rates: every catalog provider's models are folded
-// into the pricing surfaces the provider declares (PricingSurfaces),
-// then a small supplemental list adds priced-but-not-operator-selectable
-// entries. Management is the sole pricing authority — the proxy carries
-// no embedded price list and bills exclusively from the table it is
-// sent.
-//go:generate go run gen.go
-
-package pricing
-
-import (
-	"sync"
-	"sync/atomic"
-
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
-)
-
-// Entry is a single model's pricing in USD per 1k tokens. This struct IS
-// the wire shape: the synthesizer marshals it verbatim into cost_meter's
-// ConfigJSON, and the proxy unmarshals the same field names.
-//
-// A zero rate means "no rate configured" — the proxy bills that cache
-// bucket at InputPer1k (identical semantics to the retired proxy-embedded
-// table). CachedInputPer1k is the OpenAI shape (cached prompt tokens are
-// a subset of input); CacheReadPer1k / CacheCreationPer1k are the
-// Anthropic shape (additive buckets).
-type Entry struct {
-	InputPer1k         float64 `json:"input_per_1k"`
-	OutputPer1k        float64 `json:"output_per_1k"`
-	CachedInputPer1k   float64 `json:"cached_input_per_1k,omitempty"`
-	CacheReadPer1k     float64 `json:"cache_read_per_1k,omitempty"`
-	CacheCreationPer1k float64 `json:"cache_creation_per_1k,omitempty"`
-}
-
-// supplementalDefaults are (surface, model) entries that are priced but
-// deliberately not operator-selectable in the catalog. Each carries a
-// reason; when one of these models joins a catalog lineup, delete the
-// row here — the collision test fails loudly if the rates ever disagree.
-var supplementalDefaults = map[string]map[string]Entry{
-	"openai": {
-		// GPT-5 (2025) family — kept for gateway requests using the
-		// unsuffixed ids; the dashboard offers only the 5.x lineup.
-		"gpt-5":      {InputPer1k: 0.00125, OutputPer1k: 0.01, CachedInputPer1k: 0.000125},
-		"gpt-5-mini": {InputPer1k: 0.00025, OutputPer1k: 0.002, CachedInputPer1k: 0.000025},
-		"gpt-5-nano": {InputPer1k: 0.00005, OutputPer1k: 0.0004, CachedInputPer1k: 0.000005},
-	},
-	"anthropic": {
-		// claude-opus-5 is not yet in the catalog lineup but gateway /
-		// grandfathered traffic uses it; priced so it isn't skipped.
-		"claude-opus-5": {InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625},
-		// "kimi-k3[1m]" is the 1M-context alias some Claude Code guides
-		// configure against Moonshot's Anthropic-compatible endpoint;
-		// priced identically to kimi-k3 so those requests aren't skipped.
-		"kimi-k3[1m]": {InputPer1k: 0.003, OutputPer1k: 0.015, CacheReadPer1k: 0.0003},
-	},
-	"bedrock": {
-		"anthropic.claude-opus-5": {InputPer1k: 0.005, OutputPer1k: 0.025, CacheReadPer1k: 0.0005, CacheCreationPer1k: 0.00625},
-	},
-}
-
-var (
-	compiledOnce  sync.Once
-	compiledTable map[string]map[string]Entry
-	// mergedTable holds the current live table when a pricing defaults
-	// file is loaded: the file merged entry-whole over the compiled-in
-	// base. Nil while no file is loaded (or after the file is removed),
-	// in which case the compiled-in table serves. Swapped atomically by
-	// the file loader/reloader; readers never block.
-	mergedTable atomic.Pointer[map[string]map[string]Entry]
-)
-
-// DefaultTable returns the current default pricing table keyed
-// surface -> model -> Entry: the management-side defaults file (see
-// LoadFile / StartReloader) when one is loaded, merged over the
-// compiled-in catalog table, which alone serves as the fallback when no
-// file exists. The snapshot may change between calls as the file is
-// re-read — consumers (the synthesizer on every reconcile, the catalog
-// endpoint on every request) pick up fresh rates automatically. Callers
-// must not mutate the returned maps.
-func DefaultTable() map[string]map[string]Entry {
-	if t := mergedTable.Load(); t != nil {
-		return *t
-	}
-	return compiledBase()
-}
-
-// compiledBase returns the compiled-in table (catalog + supplementals),
-// built once.
-func compiledBase() map[string]map[string]Entry {
-	compiledOnce.Do(func() {
-		compiledTable = buildDefaultTable()
-	})
-	return compiledTable
-}
-
-func buildDefaultTable() map[string]map[string]Entry {
-	out := make(map[string]map[string]Entry)
-	for _, p := range catalog.All() {
-		for _, surface := range p.PricingSurfaces {
-			inner, ok := out[surface]
-			if !ok {
-				inner = make(map[string]Entry)
-				out[surface] = inner
-			}
-			for _, m := range p.Models {
-				// First writer wins; providers contributing the same
-				// (surface, model) must agree on rates — enforced by
-				// TestDefaultTable_NoConflictingContributions.
-				if _, dup := inner[m.ID]; dup {
-					continue
-				}
-				inner[m.ID] = entryFromCatalogModel(m)
-			}
-		}
-	}
-	for surface, models := range supplementalDefaults {
-		inner, ok := out[surface]
-		if !ok {
-			inner = make(map[string]Entry)
-			out[surface] = inner
-		}
-		for id, e := range models {
-			if _, dup := inner[id]; dup {
-				continue
-			}
-			inner[id] = e
-		}
-	}
-	return out
-}
-
-func entryFromCatalogModel(m catalog.Model) Entry {
-	return Entry{
-		InputPer1k:         m.InputPer1k,
-		OutputPer1k:        m.OutputPer1k,
-		CachedInputPer1k:   m.CachedInputPer1k,
-		CacheReadPer1k:     m.CacheReadPer1k,
-		CacheCreationPer1k: m.CacheCreationPer1k,
-	}
-}
-
-// LookupDefault returns the default entry for model on the first of the
-// given surfaces that prices it. Used by the synthesizer to seed a
-// per-provider entry with default cache rates before overlaying the
-// operator's stored prices.
-func LookupDefault(surfaces []string, model string) (Entry, bool) {
-	table := DefaultTable()
-	for _, s := range surfaces {
-		if e, ok := table[s][model]; ok {
-			return e, true
-		}
-	}
-	return Entry{}, false
-}

management/internals/modules/agentnetwork/pricing/defaults_test.go

@@ -1,148 +0,0 @@
-package pricing
-
-import (
-	"math"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
-)
-
-// TestDefaultTable_CoversEveryCatalogModel replaces the proxy's old
-// hand-maintained coverage list: because the table is built FROM the
-// catalog, drift is impossible by construction — this test guards the
-// fold itself (every catalog model of every surfaced provider resolves,
-// with exactly the catalog's rates).
-func TestDefaultTable_CoversEveryCatalogModel(t *testing.T) {
-	table := DefaultTable()
-	for _, p := range catalog.All() {
-		if len(p.PricingSurfaces) == 0 {
-			assert.Empty(t, p.Models, "catalog entry %s declares models but no pricing surfaces — those models would never be priced", p.ID)
-			continue
-		}
-		for _, surface := range p.PricingSurfaces {
-			byModel, ok := table[surface]
-			require.True(t, ok, "surface %q (provider %s) missing from default table", surface, p.ID)
-			for _, m := range p.Models {
-				e, ok := byModel[m.ID]
-				require.True(t, ok, "%s/%s (provider %s) missing from default table", surface, m.ID, p.ID)
-				assert.Equal(t, m.InputPer1k, e.InputPer1k, "%s/%s input rate", surface, m.ID)
-				assert.Equal(t, m.OutputPer1k, e.OutputPer1k, "%s/%s output rate", surface, m.ID)
-			}
-		}
-	}
-}
-
-// TestDefaultTable_NoConflictingContributions enforces the collision rule
-// documented on catalog.Provider.PricingSurfaces: when two catalog
-// providers contribute the same (surface, model) pair — azure/vertex
-// mirroring openai/anthropic, kimi on both surfaces — their rates must be
-// identical, because the surface-keyed table can only hold one entry.
-// If a provider ever diverges (e.g. Azure reprices a model), this fails
-// and the divergence must move to per-provider-record pricing.
-func TestDefaultTable_NoConflictingContributions(t *testing.T) {
-	type contribution struct {
-		providerID string
-		entry      Entry
-	}
-	seen := map[string]map[string]contribution{}
-	for _, p := range catalog.All() {
-		for _, surface := range p.PricingSurfaces {
-			if seen[surface] == nil {
-				seen[surface] = map[string]contribution{}
-			}
-			for _, m := range p.Models {
-				e := entryFromCatalogModel(m)
-				if prev, dup := seen[surface][m.ID]; dup {
-					assert.Equal(t, prev.entry, e,
-						"%s/%s: %s and %s contribute different rates", surface, m.ID, prev.providerID, p.ID)
-					continue
-				}
-				seen[surface][m.ID] = contribution{providerID: p.ID, entry: e}
-			}
-		}
-	}
-	// Supplemental entries must never shadow a catalog-contributed model —
-	// they exist precisely because the catalog does NOT list them.
-	for surface, models := range supplementalDefaults {
-		for id := range models {
-			_, fromCatalog := seen[surface][id]
-			assert.False(t, fromCatalog, "supplemental %s/%s is now in the catalog — delete the supplemental row", surface, id)
-		}
-	}
-}
-
-// TestDefaultTable_AllRatesFiniteNonNegative mirrors the proxy-side
-// NewTable validation so a bad catalog edit is caught here, at unit-test
-// time, rather than as a chain-build failure on every proxy.
-func TestDefaultTable_AllRatesFiniteNonNegative(t *testing.T) {
-	for surface, models := range DefaultTable() {
-		for id, e := range models {
-			for field, v := range map[string]float64{
-				"input":          e.InputPer1k,
-				"output":         e.OutputPer1k,
-				"cached_input":   e.CachedInputPer1k,
-				"cache_read":     e.CacheReadPer1k,
-				"cache_creation": e.CacheCreationPer1k,
-			} {
-				assert.False(t, v < 0 || math.IsNaN(v) || math.IsInf(v, 0),
-					"%s/%s: %s rate %v must be finite and non-negative", surface, id, field, v)
-			}
-		}
-	}
-}
-
-// TestDefaultTable_PinnedRates pins rates that previously drifted or are
-// easy to mis-enter (carried over from the proxy's retired
-// defaults_coverage_test), plus the supplemental entries.
-func TestDefaultTable_PinnedRates(t *testing.T) {
-	table := DefaultTable()
-
-	gpt54 := table["openai"]["gpt-5.4"]
-	assert.InDelta(t, 0.0025, gpt54.InputPer1k, 1e-9, "gpt-5.4 input")
-	assert.InDelta(t, 0.015, gpt54.OutputPer1k, 1e-9, "gpt-5.4 output")
-	assert.InDelta(t, 0.00025, gpt54.CachedInputPer1k, 1e-9, "gpt-5.4 cached input")
-
-	sonnet := table["bedrock"]["anthropic.claude-sonnet-4-5"]
-	assert.InDelta(t, 0.003, sonnet.InputPer1k, 1e-9, "bedrock sonnet-4-5 input")
-	assert.InDelta(t, 0.015, sonnet.OutputPer1k, 1e-9, "bedrock sonnet-4-5 output")
-	assert.InDelta(t, 0.0003, sonnet.CacheReadPer1k, 1e-9, "bedrock sonnet-4-5 cache read")
-	assert.InDelta(t, 0.00375, sonnet.CacheCreationPer1k, 1e-9, "bedrock sonnet-4-5 cache creation")
-
-	// Vertex Claude prices under "anthropic" with the bare id.
-	fable := table["anthropic"]["claude-fable-5"]
-	assert.InDelta(t, 0.010, fable.InputPer1k, 1e-9, "claude-fable-5 input")
-	assert.InDelta(t, 0.0125, fable.CacheCreationPer1k, 1e-9, "claude-fable-5 cache creation")
-
-	// Supplementals present on their surfaces.
-	for surface, ids := range map[string][]string{
-		"openai":    {"gpt-5", "gpt-5-mini", "gpt-5-nano"},
-		"anthropic": {"claude-opus-5", "kimi-k3[1m]", "kimi-k3"},
-		"bedrock":   {"anthropic.claude-opus-5"},
-	} {
-		for _, id := range ids {
-			_, ok := table[surface][id]
-			assert.True(t, ok, "%s/%s must be priced", surface, id)
-		}
-	}
-
-	// Embeddings bill input-only — output stays zero.
-	emb := table["openai"]["text-embedding-3-large"]
-	assert.Zero(t, emb.OutputPer1k, "embedding output rate must be zero")
-	assert.Positive(t, emb.InputPer1k, "embedding input rate must be set")
-}
-
-func TestLookupDefault_SurfaceOrder(t *testing.T) {
-	// kimi-k3 exists on both surfaces; first surface in the slice wins.
-	e, ok := LookupDefault([]string{"openai", "anthropic"}, "kimi-k3")
-	require.True(t, ok)
-	assert.InDelta(t, 0.003, e.InputPer1k, 1e-9)
-
-	_, ok = LookupDefault([]string{"bedrock"}, "gpt-4o")
-	assert.False(t, ok, "gpt-4o is not a bedrock model")
-
-	_, ok = LookupDefault(nil, "gpt-4o")
-	assert.False(t, ok, "no surfaces, no match")
-}

management/internals/modules/agentnetwork/pricing/exampleyaml.go

@@ -1,109 +0,0 @@
-package pricing
-
-import (
-	"bytes"
-	"fmt"
-	"sort"
-	"strconv"
-)
-
-// MarshalDefaultsYAML renders the built-in default pricing table (catalog
-// + supplementals, WITHOUT any operator override) as the YAML schema
-// LoadOverrideFile consumes. It backs the generated
-// defaults_llm_pricing.example.yaml so operators start from a file that
-// matches the compiled-in rates exactly; a golden test keeps the two in
-// sync. Output is deterministic (sorted surfaces and models).
-func MarshalDefaultsYAML() []byte {
-	var b bytes.Buffer
-	b.WriteString(exampleHeader)
-
-	table := buildDefaultTable()
-	surfaces := make([]string, 0, len(table))
-	for s := range table {
-		surfaces = append(surfaces, s)
-	}
-	sort.Strings(surfaces)
-
-	for _, surface := range surfaces {
-		fmt.Fprintf(&b, "\n%s:\n", surface)
-		models := table[surface]
-		ids := make([]string, 0, len(models))
-		for id := range models {
-			ids = append(ids, id)
-		}
-		sort.Strings(ids)
-		for _, id := range ids {
-			e := models[id]
-			fmt.Fprintf(&b, "  %s:\n", yamlKey(id))
-			fmt.Fprintf(&b, "    input_per_1k: %s\n", rate(e.InputPer1k))
-			fmt.Fprintf(&b, "    output_per_1k: %s\n", rate(e.OutputPer1k))
-			if e.CachedInputPer1k > 0 {
-				fmt.Fprintf(&b, "    cached_input_per_1k: %s\n", rate(e.CachedInputPer1k))
-			}
-			if e.CacheReadPer1k > 0 {
-				fmt.Fprintf(&b, "    cache_read_per_1k: %s\n", rate(e.CacheReadPer1k))
-			}
-			if e.CacheCreationPer1k > 0 {
-				fmt.Fprintf(&b, "    cache_creation_per_1k: %s\n", rate(e.CacheCreationPer1k))
-			}
-		}
-	}
-	return b.Bytes()
-}
-
-// rate renders a USD-per-1k rate without float noise ("0.00015", not
-// "0.000150000000...").
-func rate(v float64) string {
-	return strconv.FormatFloat(v, 'f', -1, 64)
-}
-
-// yamlKey quotes model ids that YAML would otherwise misparse (e.g.
-// "kimi-k3[1m]" starts a flow sequence unquoted).
-func yamlKey(id string) string {
-	for _, r := range id {
-		switch r {
-		case '[', ']', '{', '}', ':', '#', ',', '&', '*', '!', '|', '>', '\'', '"', '%', '@', '`':
-			return strconv.Quote(id)
-		}
-	}
-	return id
-}
-
-const exampleHeader = `# Default LLM pricing used by NetBird's Agent Network cost metering.
-# GENERATED from the management catalog — do not edit this file in the
-# repository; regenerate with:
-#
-#   go generate ./management/internals/modules/agentnetwork/pricing
-#
-# Operators: copy this file to <datadir>/defaults_llm_pricing.yaml (or
-# any path configured via management.json:
-#
-#   { "AgentNetwork": { "PricingDefaultsFile": "/path/defaults_llm_pricing.yaml" } }
-#
-# ) and adjust the entries you want to change. Management re-reads the
-# file periodically (mtime poll, every minute): the live table feeds the
-# proxies' cost metering and the dashboard's model-price prefill, so
-# edits apply without a restart. Your file only needs the entries you
-# want to change — but each entry REPLACES the built-in entry for that
-# surface+model whole, so repeat the cache rates you want to keep.
-# Unknown fields and negative or non-finite rates are rejected: at
-# startup that fails boot (for an explicitly configured path); at
-# runtime the previous table is kept and a warning is logged. Deleting
-# the file reverts to the built-in defaults below.
-#
-# Top-level keys are pricing surfaces — the parser shape requests are
-# metered under: "openai" (also Azure, Mistral, and OpenAI-compatible
-# gateways), "anthropic" (also Anthropic-on-Vertex), "bedrock"
-# (normalized ids, e.g. anthropic.claude-sonnet-4-5). Model keys must be
-# the normalized id the proxy meters (version/region suffixes stripped).
-#
-# Values are USD per 1_000 tokens. Optional cache fields:
-#   cached_input_per_1k    OpenAI shape: rate for cached prompt tokens
-#                          (a SUBSET of input tokens). Absent -> cached
-#                          portion bills at input_per_1k.
-#   cache_read_per_1k      Anthropic shape: rate for cache_read tokens
-#                          (ADDITIVE to input). Absent -> input rate.
-#   cache_creation_per_1k  Anthropic shape: rate for cache_creation
-#                          tokens (ADDITIVE to input). Absent -> input
-#                          rate.
-`

management/internals/modules/agentnetwork/pricing/gen.go

@@ -1,20 +0,0 @@
-//go:build ignore
-
-// Regenerates defaults_llm_pricing.example.yaml from the compiled-in
-// default pricing table. Run via:
-//
-//	go generate ./management/internals/modules/agentnetwork/pricing
-package main
-
-import (
-	"log"
-	"os"
-
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
-)
-
-func main() {
-	if err := os.WriteFile("defaults_llm_pricing.example.yaml", pricing.MarshalDefaultsYAML(), 0o644); err != nil {
-		log.Fatalf("write defaults_llm_pricing.example.yaml: %v", err)
-	}
-}

management/internals/modules/agentnetwork/pricing/override.go

@@ -1,244 +0,0 @@
-package pricing
-
-import (
-	"bytes"
-	"context"
-	"errors"
-	"fmt"
-	"io"
-	"io/fs"
-	"math"
-	"os"
-	"sync"
-	"time"
-
-	log "github.com/sirupsen/logrus"
-	"gopkg.in/yaml.v3"
-)
-
-// DefaultFileName is the basename probed under management's datadir when
-// AgentNetwork.PricingDefaultsFile doesn't configure an explicit path.
-const DefaultFileName = "defaults_llm_pricing.yaml"
-
-// ReloadInterval is the cadence at which the pricing file's mtime is
-// polled for changes.
-const ReloadInterval = time.Minute
-
-// maxFileBytes bounds the pricing file read so a misconfigured path
-// (pointed at a huge file) cannot exhaust process memory.
-const maxFileBytes = 4 << 20
-
-// pricingFile mirrors the on-disk YAML schema — the same schema the
-// proxy's retired embedded defaults_pricing.yaml used, so files written
-// for it keep working. Keys are pricing surfaces ("openai", "anthropic",
-// "bedrock"); nested keys are normalized model ids.
-type pricingFile map[string]map[string]struct {
-	InputPer1k         float64 `yaml:"input_per_1k"`
-	OutputPer1k        float64 `yaml:"output_per_1k"`
-	CachedInputPer1k   float64 `yaml:"cached_input_per_1k"`
-	CacheReadPer1k     float64 `yaml:"cache_read_per_1k"`
-	CacheCreationPer1k float64 `yaml:"cache_creation_per_1k"`
-}
-
-// fileState tracks the watched pricing file across reloads.
-var fileState struct {
-	mu    sync.Mutex
-	path  string
-	mtime int64
-}
-
-// LoadFile loads the management-side pricing defaults file and makes it
-// the live table (merged entry-whole over the compiled-in fallback; see
-// DefaultTable). The path stays registered for the periodic reloader, so
-// later edits — or the file (re)appearing after deletion — are picked up
-// without a restart.
-//
-// required governs the missing-file case: true for an explicitly
-// configured path (a typo must fail startup rather than silently bill
-// with built-ins the operator believes they replaced), false for the
-// conventional <datadir>/defaults_llm_pricing.yaml probe (absent file =
-// compiled-in defaults, still watched in case it appears). A file that
-// exists but is malformed is always an error at load time.
-func LoadFile(path string, required bool) error {
-	if path == "" {
-		return nil
-	}
-	fileState.mu.Lock()
-	fileState.path = path
-	fileState.mu.Unlock()
-
-	table, mtime, err := readFile(path)
-	if err != nil {
-		if errors.Is(err, fs.ErrNotExist) && !required {
-			return nil
-		}
-		return err
-	}
-	storeFileTable(table, mtime)
-	return nil
-}
-
-// StartReloader launches the periodic mtime-poll goroutine for the file
-// registered by LoadFile. Runtime failures are lenient — a parse error
-// keeps the previously loaded table and logs a warning; a deleted file
-// reverts to the compiled-in defaults — so a mid-edit save can never
-// take pricing down. Returns immediately when no path was registered.
-func StartReloader(ctx context.Context, interval time.Duration) {
-	fileState.mu.Lock()
-	path := fileState.path
-	fileState.mu.Unlock()
-	if path == "" {
-		return
-	}
-	if interval <= 0 {
-		interval = ReloadInterval
-	}
-	go func() {
-		t := time.NewTicker(interval)
-		defer t.Stop()
-		for {
-			select {
-			case <-ctx.Done():
-				return
-			case <-t.C:
-				reload()
-			}
-		}
-	}()
-}
-
-// reload performs one mtime check + reload cycle.
-func reload() {
-	fileState.mu.Lock()
-	path, lastMtime := fileState.path, fileState.mtime
-	fileState.mu.Unlock()
-
-	st, err := os.Stat(path)
-	if err != nil {
-		if errors.Is(err, fs.ErrNotExist) {
-			// File removed (or not yet created): serve compiled-in
-			// defaults and reset mtime so a future (re)appearance loads.
-			if mergedTable.Swap(nil) != nil {
-				log.Warnf("agent-network pricing defaults file %s removed; reverting to built-in defaults", path)
-			}
-			setMtime(0)
-			return
-		}
-		log.Warnf("agent-network pricing defaults reload: stat %s: %v", path, err)
-		return
-	}
-	if st.ModTime().UnixNano() == lastMtime {
-		return
-	}
-
-	table, mtime, err := readFile(path)
-	if err != nil {
-		// Keep the previously loaded table — never blank prices because
-		// an operator saved mid-edit.
-		log.Warnf("agent-network pricing defaults reload failed for %s (keeping previous table): %v", path, err)
-		return
-	}
-	storeFileTable(table, mtime)
-	log.Infof("agent-network pricing defaults reloaded from %s", path)
-}
-
-func readFile(path string) (map[string]map[string]Entry, int64, error) {
-	f, err := os.Open(path)
-	if err != nil {
-		return nil, 0, fmt.Errorf("open pricing defaults %s: %w", path, err)
-	}
-	defer func() { _ = f.Close() }()
-
-	st, err := f.Stat()
-	if err != nil {
-		return nil, 0, fmt.Errorf("stat pricing defaults %s: %w", path, err)
-	}
-	data, err := io.ReadAll(io.LimitReader(f, maxFileBytes+1))
-	if err != nil {
-		return nil, 0, fmt.Errorf("read pricing defaults %s: %w", path, err)
-	}
-	if len(data) > maxFileBytes {
-		return nil, 0, fmt.Errorf("pricing defaults %s exceeds %d bytes", path, maxFileBytes)
-	}
-	table, err := parsePricingYAML(data)
-	if err != nil {
-		return nil, 0, fmt.Errorf("parse pricing defaults %s: %w", path, err)
-	}
-	return table, st.ModTime().UnixNano(), nil
-}
-
-// storeFileTable merges the parsed file over the compiled-in base and
-// publishes the result as the live table. File entries replace the
-// built-in entry for the same (surface, model) whole — they are not
-// field-merged — and surfaces/models the file doesn't mention keep the
-// built-in rates, so a partial file only needs the entries it changes.
-func storeFileTable(table map[string]map[string]Entry, mtime int64) {
-	base := compiledBase()
-	merged := make(map[string]map[string]Entry, len(base)+len(table))
-	for surface, models := range base {
-		inner := make(map[string]Entry, len(models))
-		for id, e := range models {
-			inner[id] = e
-		}
-		merged[surface] = inner
-	}
-	for surface, models := range table {
-		inner, ok := merged[surface]
-		if !ok {
-			inner = make(map[string]Entry, len(models))
-			merged[surface] = inner
-		}
-		for id, e := range models {
-			inner[id] = e
-		}
-	}
-	mergedTable.Store(&merged)
-	setMtime(mtime)
-}
-
-func setMtime(v int64) {
-	fileState.mu.Lock()
-	fileState.mtime = v
-	fileState.mu.Unlock()
-}
-
-// parsePricingYAML decodes and validates the pricing YAML. Unknown
-// fields are rejected (typos surface instead of silently pricing at 0)
-// and every rate must be a finite, non-negative USD amount — the same
-// constraints the HTTP API enforces on operator per-provider prices.
-func parsePricingYAML(data []byte) (map[string]map[string]Entry, error) {
-	dec := yaml.NewDecoder(bytes.NewReader(data))
-	dec.KnownFields(true)
-
-	var raw pricingFile
-	if err := dec.Decode(&raw); err != nil && !errors.Is(err, io.EOF) {
-		return nil, fmt.Errorf("decode yaml: %w", err)
-	}
-
-	out := make(map[string]map[string]Entry, len(raw))
-	for surface, models := range raw {
-		inner := make(map[string]Entry, len(models))
-		for model, e := range models {
-			for field, v := range map[string]float64{
-				"input_per_1k":          e.InputPer1k,
-				"output_per_1k":         e.OutputPer1k,
-				"cached_input_per_1k":   e.CachedInputPer1k,
-				"cache_read_per_1k":     e.CacheReadPer1k,
-				"cache_creation_per_1k": e.CacheCreationPer1k,
-			} {
-				if v < 0 || math.IsNaN(v) || math.IsInf(v, 0) {
-					return nil, fmt.Errorf("%s/%s: %s must be a finite, non-negative rate, got %v", surface, model, field, v)
-				}
-			}
-			inner[model] = Entry{
-				InputPer1k:         e.InputPer1k,
-				OutputPer1k:        e.OutputPer1k,
-				CachedInputPer1k:   e.CachedInputPer1k,
-				CacheReadPer1k:     e.CacheReadPer1k,
-				CacheCreationPer1k: e.CacheCreationPer1k,
-			}
-		}
-		out[surface] = inner
-	}
-	return out, nil
-}

management/internals/modules/agentnetwork/pricing/override_test.go

@@ -1,173 +0,0 @@
-package pricing
-
-import (
-	"os"
-	"path/filepath"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// resetFileState snapshots and restores the package-level file state so
-// tests stay order-independent.
-func resetFileState(t *testing.T) {
-	t.Helper()
-	prevMerged := mergedTable.Load()
-	fileState.mu.Lock()
-	prevPath, prevMtime := fileState.path, fileState.mtime
-	fileState.mu.Unlock()
-	t.Cleanup(func() {
-		mergedTable.Store(prevMerged)
-		fileState.mu.Lock()
-		fileState.path, fileState.mtime = prevPath, prevMtime
-		fileState.mu.Unlock()
-	})
-}
-
-func writePricing(t *testing.T, path, yml string) {
-	t.Helper()
-	require.NoError(t, os.WriteFile(path, []byte(yml), 0o600))
-}
-
-func TestLoadFile_MergesOverCompiledDefaults(t *testing.T) {
-	resetFileState(t)
-	path := filepath.Join(t.TempDir(), DefaultFileName)
-	writePricing(t, path, `
-openai:
-  # Reprice a built-in model. The entry replaces the built-in WHOLE:
-  # omitting the cache rate here drops the built-in 0.00125 discount.
-  gpt-4o:
-    input_per_1k: 0.9
-    output_per_1k: 1.8
-  # A model NetBird doesn't know at all.
-  my-private-ft:
-    input_per_1k: 0.01
-    output_per_1k: 0.02
-    cached_input_per_1k: 0.005
-gemini:
-  gemini-pro:
-    input_per_1k: 0.00125
-    output_per_1k: 0.005
-`)
-	require.NoError(t, LoadFile(path, true))
-	table := DefaultTable()
-
-	gpt4o := table["openai"]["gpt-4o"]
-	assert.InDelta(t, 0.9, gpt4o.InputPer1k, 1e-9, "file rate replaces the compiled-in rate")
-	assert.Zero(t, gpt4o.CachedInputPer1k, "entries replace whole — omitted cache rate is dropped, not inherited")
-
-	ft := table["openai"]["my-private-ft"]
-	assert.InDelta(t, 0.005, ft.CachedInputPer1k, 1e-9, "unknown models are added to the surface")
-	_, ok := table["gemini"]["gemini-pro"]
-	assert.True(t, ok, "a surface the catalog doesn't declare can be added")
-
-	// Untouched entries keep compiled-in rates (catalog, other surface,
-	// supplemental).
-	assert.InDelta(t, 0.00015, table["openai"]["gpt-4o-mini"].InputPer1k, 1e-9, "unlisted model keeps compiled rate")
-	assert.InDelta(t, 0.003, table["anthropic"]["claude-sonnet-4-5"].InputPer1k, 1e-9, "unlisted surface untouched")
-	assert.InDelta(t, 0.00125, table["openai"]["gpt-5"].InputPer1k, 1e-9, "supplemental entries untouched")
-
-	// The synthesizer-facing lookup reads the live table too.
-	e, ok := LookupDefault([]string{"openai"}, "gpt-4o")
-	require.True(t, ok)
-	assert.InDelta(t, 0.9, e.InputPer1k, 1e-9, "LookupDefault serves the file-backed rate")
-}
-
-func TestLoadFile_MissingPath(t *testing.T) {
-	resetFileState(t)
-	missing := filepath.Join(t.TempDir(), DefaultFileName)
-
-	require.Error(t, LoadFile(missing, true),
-		"explicitly configured path that doesn't exist must fail startup")
-
-	require.NoError(t, LoadFile(missing, false),
-		"conventional datadir probe tolerates an absent file (compiled-in defaults serve)")
-	assert.Nil(t, mergedTable.Load(), "no file, no merged table")
-	fileState.mu.Lock()
-	path := fileState.path
-	fileState.mu.Unlock()
-	assert.Equal(t, missing, path, "the path stays registered so the reloader picks the file up when it appears")
-}
-
-func TestLoadFile_RejectsInvalid(t *testing.T) {
-	resetFileState(t)
-	dir := t.TempDir()
-	cases := map[string]string{
-		"unknown field (typo)": "openai:\n  gpt-4o:\n    input_per1k: 0.1\n",
-		"negative rate":        "openai:\n  gpt-4o:\n    input_per_1k: -0.1\n",
-		"non-numeric rate":     "openai:\n  gpt-4o:\n    input_per_1k: cheap\n",
-		"not a mapping":        "- just\n- a\n- list\n",
-	}
-	for name, yml := range cases {
-		path := filepath.Join(dir, DefaultFileName)
-		writePricing(t, path, yml)
-		assert.Error(t, LoadFile(path, true), "case %q must be rejected", name)
-	}
-}
-
-// TestReload_LifeCycle drives the reloader's single-shot reload through
-// its full lifecycle: file edit picked up on mtime change, a broken save
-// keeps the previous table, and file removal reverts to the compiled-in
-// defaults (then a re-created file loads again).
-func TestReload_LifeCycle(t *testing.T) {
-	resetFileState(t)
-	path := filepath.Join(t.TempDir(), DefaultFileName)
-	writePricing(t, path, "openai:\n  gpt-4o:\n    input_per_1k: 0.5\n    output_per_1k: 1\n")
-	require.NoError(t, LoadFile(path, true))
-	require.InDelta(t, 0.5, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9)
-
-	// Edit: new mtime, new rates.
-	writePricing(t, path, "openai:\n  gpt-4o:\n    input_per_1k: 0.7\n    output_per_1k: 1.4\n")
-	bumpMtime(t, path)
-	reload()
-	assert.InDelta(t, 0.7, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9, "edit must be picked up")
-
-	// Broken save: previous table survives.
-	writePricing(t, path, "openai:\n  gpt-4o:\n    input_per_1k: -1\n")
-	bumpMtime(t, path)
-	reload()
-	assert.InDelta(t, 0.7, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9,
-		"a malformed save must keep the previously loaded table, never blank prices")
-
-	// Removal: compiled-in defaults serve again.
-	require.NoError(t, os.Remove(path))
-	reload()
-	assert.InDelta(t, 0.0025, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9,
-		"file removal reverts to the compiled-in rate")
-
-	// Re-created file loads without a restart.
-	writePricing(t, path, "openai:\n  gpt-4o:\n    input_per_1k: 0.9\n    output_per_1k: 1.8\n")
-	bumpMtime(t, path)
-	reload()
-	assert.InDelta(t, 0.9, DefaultTable()["openai"]["gpt-4o"].InputPer1k, 1e-9,
-		"a file appearing after removal (or after a missing-probe boot) must load")
-}
-
-// bumpMtime pushes the file's mtime forward past the previously recorded
-// value — timestamps can otherwise collide within the test's timescale.
-func bumpMtime(t *testing.T, path string) {
-	t.Helper()
-	st, err := os.Stat(path)
-	require.NoError(t, err)
-	next := st.ModTime().Add(2 * 1e9)
-	require.NoError(t, os.Chtimes(path, next, next))
-}
-
-// TestExampleYAML_InSyncWithBuiltins is the golden guard for
-// defaults_llm_pricing.example.yaml: the shipped example must stay
-// byte-identical to what the compiled-in table renders (catalog edits
-// require `go generate ./management/internals/modules/agentnetwork/pricing`)
-// and must round-trip through the same parser operators' files go
-// through, reproducing the compiled-in table exactly.
-func TestExampleYAML_InSyncWithBuiltins(t *testing.T) {
-	onDisk, err := os.ReadFile("defaults_llm_pricing.example.yaml")
-	require.NoError(t, err, "example file must exist next to the package")
-	require.Equal(t, string(MarshalDefaultsYAML()), string(onDisk),
-		"defaults_llm_pricing.example.yaml is stale — run: go generate ./management/internals/modules/agentnetwork/pricing")
-
-	parsed, err := parsePricingYAML(onDisk)
-	require.NoError(t, err, "the example must be a valid pricing defaults file")
-	assert.Equal(t, buildDefaultTable(), parsed,
-		"parsing the example must reproduce the compiled-in table exactly")
-}

management/internals/modules/agentnetwork/synthesizer.go

@@ -233,11 +233,6 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
 		return nil, err
 	}
 
-	costMeterJSON, err := buildCostMeterConfigJSON(enabledProviders, groupIndex)
-	if err != nil {
-		return nil, err
-	}
-
 	mergedGuardrails := mergeGuardrails(enabledPolicies, guardrailsByID)
 	applyAccountCollectionControls(&mergedGuardrails, settings)
 	// The proxy guardrail is a per-provider fail-closed backstop; the
@@ -253,7 +248,7 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
 	// Use the merged decision (account settings OR policy-required redaction),
 	// not the raw account flag, so a policy that mandates PII redaction is
 	// honored by the capture parsers even when the account toggle is off.
-	middlewares := buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON, costMeterJSON, mergedGuardrails.PromptCapture.RedactPii, mergedGuardrails.PromptCapture.Enabled)
+	middlewares := buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON, mergedGuardrails.PromptCapture.RedactPii, mergedGuardrails.PromptCapture.Enabled)
 
 	priv, pub, err := pickServiceSessionKeys(enabledProviders)
 	if err != nil {
@@ -705,7 +700,7 @@ func buildIdentityExtraHeaders(p *types.Provider, extras []catalog.ExtraHeader)
 // requests bound for gateways like LiteLLM that key budgets and
 // attribution off request headers. CanMutate is required so its
 // HeadersAdd / HeadersRemove pass the framework's mutation gate.
-func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON, costMeterJSON []byte, redactPii, capturePromptContent bool) []rpservice.MiddlewareConfig {
+func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON []byte, redactPii, capturePromptContent bool) []rpservice.MiddlewareConfig {
 	// Both parsers receive an explicit capture flag derived from the account's
 	// enable_prompt_collection toggle; nil/unset would default to the legacy
 	// "always emit" behavior in the middleware, which is precisely what we
@@ -774,13 +769,10 @@ func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON, cost
 			ConfigJSON: []byte("{}"),
 		},
 		{
-			// Carries the full pricing table (defaults + per-provider
-			// operator prices) so the proxy bills without an embedded
-			// price list; see buildCostMeterConfigJSON.
 			ID:         middlewareIDCostMeter,
 			Enabled:    true,
 			Slot:       rpservice.MiddlewareSlotOnResponse,
-			ConfigJSON: costMeterJSON,
+			ConfigJSON: []byte("{}"),
 		},
 		{
 			ID:         middlewareIDLLMResponseParser,

management/internals/modules/agentnetwork/synthesizer_pricing.go

@@ -1,131 +0,0 @@
-package agentnetwork
-
-import (
-	"encoding/json"
-	"fmt"
-
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/catalog"
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
-	sharedllm "github.com/netbirdio/netbird/shared/llm"
-)
-
-// costMeterConfig is the JSON shape the proxy-side cost_meter middleware
-// expects (mirror-type pattern, same as routerConfig). The top-level
-// "pricing" wrapper is the feature-detection signal: an old proxy's config
-// struct ignores it as an unknown field, and a new proxy treats its
-// absence as "old management" (skips every cost computation and warns).
-type costMeterConfig struct {
-	Pricing *costMeterPricing `json:"pricing,omitempty"`
-}
-
-// costMeterPricing carries the full pricing table:
-//   - Defaults: surface ("openai"/"anthropic"/"bedrock") -> normalized
-//     model id -> rates. The full default table ships to every account —
-//     it is small (~10 KB) and keeps gateway-style providers (which
-//     enumerate no models) priced for every catalog model.
-//   - Providers: provider record id (matched against the
-//     llm.resolved_provider_id metadata llm_router stamps) -> normalized
-//     model id -> rates. Entries are fully materialized here at synth
-//     time — default cache rates already folded in — so the proxy does
-//     two map lookups and no merging.
-type costMeterPricing struct {
-	Defaults  map[string]map[string]pricing.Entry `json:"defaults,omitempty"`
-	Providers map[string]map[string]pricing.Entry `json:"providers,omitempty"`
-}
-
-// buildCostMeterConfigJSON assembles the cost_meter middleware config
-// from the default pricing table plus the operator's stored per-provider
-// model prices. Same orphan rule as the router: a provider no enabled
-// policy authorises is unreachable, so its prices are not shipped.
-//
-// Overlay semantics per model row:
-//   - The row's model id is normalized exactly the way the proxy's
-//     request parser normalizes the ids it meters (bedrock ARN/region/
-//     version stripping, vertex "@version" stripping), so the per-record
-//     lookup key compares equal to llm.model at billing time.
-//   - The entry starts from the default entry for that model (when one
-//     exists) to inherit cache rates the operator didn't state.
-//   - Operator input/output overlay verbatim — including an explicit 0,
-//     which prices the model as free (self-hosted / internal endpoints)
-//     rather than silently reverting to list price.
-//   - Cache-rate pointers overlay only when non-nil: nil means "inherit
-//     the default", an explicit 0 means "no discount, bill this bucket
-//     at the input rate".
-func buildCostMeterConfigJSON(providers []*types.Provider, groupIndex map[string][]string) ([]byte, error) {
-	cfg := costMeterConfig{Pricing: &costMeterPricing{
-		Defaults: pricing.DefaultTable(),
-	}}
-
-	perRecord := make(map[string]map[string]pricing.Entry)
-	for _, p := range providers {
-		if _, hasPolicy := groupIndex[p.ID]; !hasPolicy {
-			// Orphan: unreachable via the router, so unpriceable.
-			continue
-		}
-		if len(p.Models) == 0 {
-			// Gateway-style "claim every model" provider: the defaults
-			// table is its price list.
-			continue
-		}
-		entry, _ := catalog.Lookup(p.ProviderID)
-		models := make(map[string]pricing.Entry, len(p.Models))
-		for _, m := range p.Models {
-			id := normalizePricingModelID(p.ProviderID, m.ID)
-			if id == "" {
-				continue
-			}
-			if _, dup := models[id]; dup {
-				// First occurrence wins on post-normalization duplicates,
-				// matching providerModelIDs' dedup order for routing.
-				continue
-			}
-			models[id] = materializeEntry(entry.PricingSurfaces, id, m)
-		}
-		if len(models) > 0 {
-			perRecord[p.ID] = models
-		}
-	}
-	if len(perRecord) > 0 {
-		cfg.Pricing.Providers = perRecord
-	}
-
-	out, err := json.Marshal(cfg)
-	if err != nil {
-		return nil, fmt.Errorf("marshal cost_meter middleware config: %w", err)
-	}
-	return out, nil
-}
-
-// normalizePricingModelID maps an operator-entered model id onto the
-// normalized id the proxy's request parser emits as llm.model — the key
-// the cost meter looks up at billing time.
-func normalizePricingModelID(catalogProviderID, modelID string) string {
-	switch {
-	case catalog.IsBedrockPathStyle(catalogProviderID):
-		return sharedllm.NormalizeBedrockModel(modelID)
-	case catalog.IsVertexPathStyle(catalogProviderID):
-		return sharedllm.NormalizeVertexModel(modelID)
-	default:
-		return modelID
-	}
-}
-
-// materializeEntry folds the default entry for (surfaces, model) — when
-// one exists — under the operator's stored prices, producing the fully
-// materialized wire entry.
-func materializeEntry(surfaces []string, normalizedID string, m types.ProviderModel) pricing.Entry {
-	e, _ := pricing.LookupDefault(surfaces, normalizedID) // zero Entry on miss
-	e.InputPer1k = m.InputPer1k
-	e.OutputPer1k = m.OutputPer1k
-	if m.CachedInputPer1k != nil {
-		e.CachedInputPer1k = *m.CachedInputPer1k
-	}
-	if m.CacheReadPer1k != nil {
-		e.CacheReadPer1k = *m.CacheReadPer1k
-	}
-	if m.CacheCreationPer1k != nil {
-		e.CacheCreationPer1k = *m.CacheCreationPer1k
-	}
-	return e
-}

management/internals/modules/agentnetwork/synthesizer_pricing_test.go

@@ -1,105 +0,0 @@
-package agentnetwork
-
-import (
-	"encoding/json"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
-)
-
-func fptr(v float64) *float64 { return &v }
-
-func decodeCostMeterConfig(t *testing.T, raw []byte) costMeterConfig {
-	t.Helper()
-	var cfg costMeterConfig
-	require.NoError(t, json.Unmarshal(raw, &cfg), "cost meter config must round-trip")
-	require.NotNil(t, cfg.Pricing, "pricing wrapper must be present")
-	return cfg
-}
-
-// TestBuildCostMeterConfig_BedrockModelNormalization: the operator may
-// paste region-prefixed, versioned, or ARN-wrapped Bedrock ids; the
-// per-record map must be keyed by the normalized id the request parser
-// emits as llm.model, or the lookup never hits at billing time.
-func TestBuildCostMeterConfig_BedrockModelNormalization(t *testing.T) {
-	bedrock := &types.Provider{
-		ID:         "prov-bedrock",
-		ProviderID: "bedrock_api",
-		Enabled:    true,
-		Models: []types.ProviderModel{
-			{ID: "eu.anthropic.claude-sonnet-4-5-20250929-v1:0", InputPer1k: 0.0033, OutputPer1k: 0.0165},
-			// Post-normalization duplicate of the row above under a
-			// different regional spelling — first occurrence wins.
-			{ID: "us.anthropic.claude-sonnet-4-5-20250929-v1:0", InputPer1k: 9.9, OutputPer1k: 9.9},
-		},
-	}
-	raw, err := buildCostMeterConfigJSON([]*types.Provider{bedrock}, map[string][]string{"prov-bedrock": {"grp"}})
-	require.NoError(t, err)
-	cfg := decodeCostMeterConfig(t, raw)
-
-	models := cfg.Pricing.Providers["prov-bedrock"]
-	require.Len(t, models, 1, "both spellings normalize to one model; first row wins")
-	e, ok := models["anthropic.claude-sonnet-4-5"]
-	require.True(t, ok, "key must be the normalized id the parser emits, not the operator's raw spelling")
-	assert.InDelta(t, 0.0033, e.InputPer1k, 1e-9, "first row's rate wins the dedup")
-	assert.InDelta(t, 0.0003, e.CacheReadPer1k, 1e-9, "cache read inherited from the bedrock default entry")
-	assert.InDelta(t, 0.00375, e.CacheCreationPer1k, 1e-9, "cache creation inherited from the bedrock default entry")
-}
-
-// TestBuildCostMeterConfig_CacheRateNilVsZero pins the pointer semantics:
-// nil inherits the default cache rate, explicit 0 clears it (that bucket
-// bills at the input rate on the proxy).
-func TestBuildCostMeterConfig_CacheRateNilVsZero(t *testing.T) {
-	p := &types.Provider{
-		ID:         "prov-oai",
-		ProviderID: "openai_api",
-		Enabled:    true,
-		Models: []types.ProviderModel{
-			{ID: "gpt-4o", InputPer1k: 0.002, OutputPer1k: 0.008},                                    // nil → inherit 0.00125
-			{ID: "gpt-4o-mini", InputPer1k: 0.0001, OutputPer1k: 0.0005, CachedInputPer1k: fptr(0)},  // explicit 0 → no discount
-			{ID: "my-custom-ft", InputPer1k: 0.01, OutputPer1k: 0.02, CachedInputPer1k: fptr(0.005)}, // unknown model, explicit rate
-		},
-	}
-	raw, err := buildCostMeterConfigJSON([]*types.Provider{p}, map[string][]string{"prov-oai": {"grp"}})
-	require.NoError(t, err)
-	cfg := decodeCostMeterConfig(t, raw)
-	models := cfg.Pricing.Providers["prov-oai"]
-
-	assert.InDelta(t, 0.00125, models["gpt-4o"].CachedInputPer1k, 1e-9, "nil cache pointer inherits the default rate")
-	assert.Zero(t, models["gpt-4o-mini"].CachedInputPer1k, "explicit 0 overrides the default (0.000075) — bucket bills at input rate")
-	custom := models["my-custom-ft"]
-	assert.InDelta(t, 0.005, custom.CachedInputPer1k, 1e-9, "unknown model keeps the operator's explicit cache rate")
-	assert.Zero(t, custom.CacheReadPer1k, "no default to inherit for a model outside the catalog")
-}
-
-// TestBuildCostMeterConfig_OrphanAndGatewayProviders: an orphan (no
-// authorising policy) is unreachable so its prices must not ship; a
-// gateway with no model rows relies on the defaults table and gets no
-// per-record entry.
-func TestBuildCostMeterConfig_OrphanAndGatewayProviders(t *testing.T) {
-	orphan := &types.Provider{
-		ID:         "prov-orphan",
-		ProviderID: "openai_api",
-		Enabled:    true,
-		Models:     []types.ProviderModel{{ID: "gpt-4o", InputPer1k: 1, OutputPer1k: 1}},
-	}
-	gateway := &types.Provider{
-		ID:         "prov-litellm",
-		ProviderID: "litellm_proxy",
-		Enabled:    true,
-		Models:     []types.ProviderModel{},
-	}
-	raw, err := buildCostMeterConfigJSON(
-		[]*types.Provider{orphan, gateway},
-		map[string][]string{"prov-litellm": {"grp"}}, // orphan has no policy
-	)
-	require.NoError(t, err)
-	cfg := decodeCostMeterConfig(t, raw)
-
-	assert.NotContains(t, cfg.Pricing.Providers, "prov-orphan", "orphan provider prices must not ship")
-	assert.NotContains(t, cfg.Pricing.Providers, "prov-litellm", "empty-models gateway needs no per-record entry")
-	assert.NotEmpty(t, cfg.Pricing.Defaults["openai"], "defaults still ship so the gateway's catalog-model traffic is priced")
-}

management/internals/modules/agentnetwork/synthesizer_test.go

@@ -33,17 +33,14 @@ func newSynthTestSettings() *types.Settings {
 
 func newSynthTestProvider() *types.Provider {
 	return &types.Provider{
-		ID:          "prov-1",
-		AccountID:   testAccountID,
-		ProviderID:  "openai_api",
-		Name:        "OpenAI",
-		UpstreamURL: "https://api.openai.com",
-		APIKey:      "sk-test-key",
-		Enabled:     true,
-		// Prices deliberately differ from the catalog's gpt-5.4 rates
-		// (0.0025/0.015) so pricing tests can prove the operator's
-		// stored price overlays the catalog default.
-		Models:            []types.ProviderModel{{ID: "gpt-5.4", InputPer1k: 0.004, OutputPer1k: 0.02}},
+		ID:                "prov-1",
+		AccountID:         testAccountID,
+		ProviderID:        "openai_api",
+		Name:              "OpenAI",
+		UpstreamURL:       "https://api.openai.com",
+		APIKey:            "sk-test-key",
+		Enabled:           true,
+		Models:            []types.ProviderModel{{ID: "gpt-5.4", InputPer1k: 0.0025, OutputPer1k: 0.015}},
 		SessionPrivateKey: "test-priv-key",
 		SessionPublicKey:  "test-pub-key",
 		CreatedAt:         time.Date(2026, 1, 1, 0, 0, 0, 0, time.UTC),
@@ -217,27 +214,7 @@ func TestSynthesizeServices_HappyPath(t *testing.T) {
 
 	assert.Equal(t, middlewareIDCostMeter, mws[6].ID, "seventh middleware is the cost meter")
 	assert.Equal(t, rpservice.MiddlewareSlotOnResponse, mws[6].Slot, "cost meter runs on_response")
-
-	var costCfg costMeterConfig
-	require.NoError(t, json.Unmarshal(mws[6].ConfigJSON, &costCfg), "cost meter config must unmarshal")
-	require.NotNil(t, costCfg.Pricing, "cost meter config must carry the pricing table — its absence tells the proxy management predates config-delivered pricing")
-
-	gpt4o, ok := costCfg.Pricing.Defaults["openai"]["gpt-4o"]
-	require.True(t, ok, "the full default table ships regardless of the account's providers")
-	assert.InDelta(t, 0.0025, gpt4o.InputPer1k, 1e-9, "default gpt-4o input rate comes from the catalog")
-
-	openaiPrices, ok := costCfg.Pricing.Providers[openai.ID]
-	require.True(t, ok, "operator-priced provider must have a per-record entry")
-	gpt54, ok := openaiPrices["gpt-5.4"]
-	require.True(t, ok, "operator's model row keys the per-record map")
-	assert.InDelta(t, 0.004, gpt54.InputPer1k, 1e-9, "operator input price overlays the catalog default (0.0025)")
-	assert.InDelta(t, 0.02, gpt54.OutputPer1k, 1e-9, "operator output price overlays the catalog default (0.015)")
-	assert.InDelta(t, 0.00025, gpt54.CachedInputPer1k, 1e-9, "cache rate the operator didn't state is inherited from the default entry")
-
-	opus, ok := costCfg.Pricing.Providers[anthropic.ID]["claude-opus-4-7"]
-	require.True(t, ok, "anthropic's model row keys its per-record map")
-	assert.Zero(t, opus.InputPer1k, "operator-stored zero prices ship verbatim — an explicit $0 model bills as free, it does not revert to list price")
-	assert.InDelta(t, 0.0005, opus.CacheReadPer1k, 1e-9, "cache rates still inherit from the default entry")
+	assert.Equal(t, []byte("{}"), mws[6].ConfigJSON, "cost meter carries an explicit empty config")
 
 	assert.Equal(t, middlewareIDLLMResponseParser, mws[7].ID, "eighth middleware is the response parser")
 	assert.Equal(t, rpservice.MiddlewareSlotOnResponse, mws[7].Slot, "response parser runs on_response")

management/internals/modules/agentnetwork/types/provider.go

@@ -14,24 +14,10 @@ import (
 // ProviderModel is one row in the provider's models list. The operator
 // pins the per-1k input/output price for cost tracking; ID is the
 // model identifier the upstream provider expects on the wire.
-//
-// The three cache rates are pointers because absence is meaningful: nil
-// means "inherit NetBird's default rate for this model" (folded in at
-// synthesis time), while an explicit 0 means "no discount — bill this
-// cache bucket at the input rate".
 type ProviderModel struct {
 	ID          string  `json:"id"`
 	InputPer1k  float64 `json:"input_per_1k"`
 	OutputPer1k float64 `json:"output_per_1k"`
-	// CachedInputPer1k is the OpenAI-shape rate for cached prompt tokens
-	// (a subset of input tokens).
-	CachedInputPer1k *float64 `json:"cached_input_per_1k,omitempty"`
-	// CacheReadPer1k is the Anthropic-shape rate for cache-read tokens
-	// (additive to input tokens).
-	CacheReadPer1k *float64 `json:"cache_read_per_1k,omitempty"`
-	// CacheCreationPer1k is the Anthropic-shape rate for cache-creation
-	// tokens (additive to input tokens).
-	CacheCreationPer1k *float64 `json:"cache_creation_per_1k,omitempty"`
 }
 
 // Provider is an Agent Network AI provider record persisted per account.
@@ -142,12 +128,9 @@ func (p *Provider) FromAPIRequest(req *api.AgentNetworkProviderRequest) {
 	if req.Models != nil {
 		for _, m := range *req.Models {
 			p.Models = append(p.Models, ProviderModel{
-				ID:                 m.Id,
-				InputPer1k:         m.InputPer1k,
-				OutputPer1k:        m.OutputPer1k,
-				CachedInputPer1k:   copyFloatPtr(m.CachedInputPer1k),
-				CacheReadPer1k:     copyFloatPtr(m.CacheReadPer1k),
-				CacheCreationPer1k: copyFloatPtr(m.CacheCreationPer1k),
+				ID:          m.Id,
+				InputPer1k:  m.InputPer1k,
+				OutputPer1k: m.OutputPer1k,
 			})
 		}
 	}
@@ -181,12 +164,9 @@ func (p *Provider) ToAPIResponse() *api.AgentNetworkProvider {
 	models := make([]api.AgentNetworkProviderModel, 0, len(p.Models))
 	for _, m := range p.Models {
 		models = append(models, api.AgentNetworkProviderModel{
-			Id:                 m.ID,
-			InputPer1k:         m.InputPer1k,
-			OutputPer1k:        m.OutputPer1k,
-			CachedInputPer1k:   copyFloatPtr(m.CachedInputPer1k),
-			CacheReadPer1k:     copyFloatPtr(m.CacheReadPer1k),
-			CacheCreationPer1k: copyFloatPtr(m.CacheCreationPer1k),
+			Id:          m.ID,
+			InputPer1k:  m.InputPer1k,
+			OutputPer1k: m.OutputPer1k,
 		})
 	}
 	created := p.CreatedAt
@@ -221,27 +201,11 @@ func (p *Provider) ToAPIResponse() *api.AgentNetworkProvider {
 	return resp
 }
 
-// copyFloatPtr returns a fresh pointer to the same value, or nil. Keeps
-// stored models and API payloads from aliasing each other's rate fields.
-func copyFloatPtr(v *float64) *float64 {
-	if v == nil {
-		return nil
-	}
-	out := *v
-	return &out
-}
-
 // Copy returns a deep copy of the provider.
 func (p *Provider) Copy() *Provider {
 	clone := *p
 	if p.Models != nil {
-		clone.Models = make([]ProviderModel, len(p.Models))
-		for i, m := range p.Models {
-			m.CachedInputPer1k = copyFloatPtr(m.CachedInputPer1k)
-			m.CacheReadPer1k = copyFloatPtr(m.CacheReadPer1k)
-			m.CacheCreationPer1k = copyFloatPtr(m.CacheCreationPer1k)
-			clone.Models[i] = m
-		}
+		clone.Models = append([]ProviderModel(nil), p.Models...)
 	}
 	if p.ExtraValues != nil {
 		clone.ExtraValues = make(map[string]string, len(p.ExtraValues))

management/internals/modules/agentnetwork/wire_shape_test.go

@@ -103,12 +103,6 @@ func TestSynthesizedService_WireShape(t *testing.T) {
 
 	assert.Equal(t, middlewareIDCostMeter, mws[6].GetId(), "seventh middleware id")
 	assert.Equal(t, proto.MiddlewareSlot_MIDDLEWARE_SLOT_ON_RESPONSE, mws[6].GetSlot(), "cost meter slot")
-	var costCfg costMeterConfig
-	require.NoError(t, json.Unmarshal(mws[6].GetConfigJson(), &costCfg), "cost meter config JSON must decode from the wire")
-	require.NotNil(t, costCfg.Pricing, "the pricing table must travel on the wire — the proxy has no embedded price list to fall back to")
-	assert.NotEmpty(t, costCfg.Pricing.Defaults["openai"], "default table rides in every mapping")
-	assert.NotEmpty(t, costCfg.Pricing.Defaults["anthropic"], "default table covers all surfaces")
-	assert.NotEmpty(t, costCfg.Pricing.Defaults["bedrock"], "default table covers all surfaces")
 
 	assert.Equal(t, middlewareIDLLMResponseParser, mws[7].GetId(), "eighth middleware id")
 	assert.Equal(t, proto.MiddlewareSlot_MIDDLEWARE_SLOT_ON_RESPONSE, mws[7].GetSlot(), "response parser slot")

management/internals/server/config/config.go

@@ -55,8 +55,6 @@ type Config struct {
 
 	ReverseProxy ReverseProxy
 
-	AgentNetwork AgentNetwork
-
 	// disable default all-to-all policy
 	DisableDefaultPolicy bool
 
@@ -187,24 +185,6 @@ type StoreConfig struct {
 	Engine types.Engine
 }
 
-// AgentNetwork contains agent-network (LLM gateway) configuration.
-type AgentNetwork struct {
-	// PricingDefaultsFile is the path to the YAML file holding the default
-	// LLM pricing table (defaults_llm_pricing.yaml). Empty falls back to
-	// probing <Datadir>/defaults_llm_pricing.yaml; with no file present the
-	// compiled-in defaults serve. Schema: surface ("openai"/"anthropic"/
-	// "bedrock") -> model -> rates in USD per 1k tokens (input_per_1k,
-	// output_per_1k, and the optional cached_input_per_1k /
-	// cache_read_per_1k / cache_creation_per_1k). File entries replace the
-	// compiled-in entry for the same surface+model whole; everything else
-	// keeps the compiled-in rates. The file is re-read periodically (mtime
-	// poll), and the live table feeds both the synthesizer (what proxies
-	// bill with) and the dashboard's catalog endpoint (what model rows
-	// prefill with). An explicitly configured path that fails to load
-	// fails startup; runtime reload errors keep the previous table.
-	PricingDefaultsFile string
-}
-
 // ReverseProxy contains reverse proxy configuration in front of management.
 type ReverseProxy struct {
 	// TrustedHTTPProxies represents a list of trusted HTTP proxies by their IP prefixes.

proxy/internal/llm/bedrock_model.go

@@ -0,0 +1,38 @@
+package llm
+
+import (
+	"regexp"
+	"strings"
+)
+
+// bedrockRegionPrefixes are the cross-region inference-profile prefixes that
+// front a Bedrock model id (e.g. "eu.anthropic.claude-...").
+var bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
+
+// bedrockVersionSuffix matches the trailing "-vN[:N]" or "-YYYYMMDD-vN[:N]"
+// version/throughput suffix of a Bedrock model id.
+var bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
+
+// NormalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
+// prefix, and the version/throughput suffix from a Bedrock model id so it
+// matches the catalog/pricing key, e.g.
+// "eu.anthropic.claude-sonnet-4-5-20250929-v1:0" -> "anthropic.claude-sonnet-4-5"
+// and the inference-profile ARN's last segment likewise. It is the single
+// source of truth shared by the request parser (which normalizes the request
+// model from the URL path) and the router (which normalizes the operator's
+// registered Bedrock model ids so both sides compare equal).
+func NormalizeBedrockModel(modelID string) string {
+	m := modelID
+	if strings.HasPrefix(m, "arn:") {
+		if i := strings.LastIndex(m, "/"); i >= 0 {
+			m = m[i+1:]
+		}
+	}
+	for _, p := range bedrockRegionPrefixes {
+		if strings.HasPrefix(m, p) {
+			m = m[len(p):]
+			break
+		}
+	}
+	return bedrockVersionSuffix.ReplaceAllString(m, "")
+}

proxy/internal/llm/bedrock_model_test.go

@@ -11,8 +11,6 @@ func TestNormalizeBedrockModel(t *testing.T) {
 		"eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
 		"us.anthropic.claude-haiku-4-5":                "anthropic.claude-haiku-4-5",
 		"us.anthropic.claude-opus-4-8-20250101-v1:0":   "anthropic.claude-opus-4-8",
-		"apac.anthropic.claude-haiku-4-5-v1:0":         "anthropic.claude-haiku-4-5",
-		"amazon.nova-2-lite-v1:0":                      "amazon.nova-2-lite",
 		"anthropic.claude-sonnet-4-5-20250929-v1:0":    "anthropic.claude-sonnet-4-5",
 		"meta.llama3-3-70b-instruct-v1:0":              "meta.llama3-3-70b-instruct",
 		"amazon.nova-pro-v1:0":                         "amazon.nova-pro",
@@ -23,14 +21,3 @@ func TestNormalizeBedrockModel(t *testing.T) {
 		require.Equal(t, want, NormalizeBedrockModel(in), "normalize %q", in)
 	}
 }
-
-func TestNormalizeVertexModel(t *testing.T) {
-	cases := map[string]string{
-		"claude-sonnet-4-5@20250929": "claude-sonnet-4-5",
-		"claude-haiku-4-5":           "claude-haiku-4-5",
-		"gpt-4o@2024-08-06":          "gpt-4o",
-	}
-	for in, want := range cases {
-		require.Equal(t, want, NormalizeVertexModel(in), "normalize %q", in)
-	}
-}

proxy/internal/llm/fixtures/pricing.yaml

@@ -0,0 +1,59 @@
+# Realistic-pricing starter for llm_observability. Drop this into the
+# directory you point the proxy at via --plugin-data-dir, then reference it
+# from the target's plugin config:
+#
+#   plugins:
+#     - id: llm_observability
+#       enabled: true
+#       params:
+#         pricing_path: pricing.yaml
+#
+# Values are USD per 1_000 tokens. Public list prices drift; treat this as a
+# starting point and keep your production copy current.
+
+openai:
+  # GPT-5 family
+  gpt-5:
+    input_per_1k: 0.00125
+    output_per_1k: 0.01
+  gpt-5-mini:
+    input_per_1k: 0.00025
+    output_per_1k: 0.002
+  gpt-5-nano:
+    input_per_1k: 0.00005
+    output_per_1k: 0.0004
+  gpt-5.4:
+    input_per_1k: 0.00125
+    output_per_1k: 0.01
+  # GPT-4o family
+  gpt-4o:
+    input_per_1k: 0.0025
+    output_per_1k: 0.01
+  gpt-4o-mini:
+    input_per_1k: 0.00015
+    output_per_1k: 0.0006
+  # Embeddings
+  text-embedding-3-large:
+    input_per_1k: 0.00013
+    output_per_1k: 0
+  text-embedding-3-small:
+    input_per_1k: 0.00002
+    output_per_1k: 0
+
+anthropic:
+  # Claude 4.x family
+  claude-opus-4-7:
+    input_per_1k: 0.015
+    output_per_1k: 0.075
+  claude-sonnet-4-7:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+  claude-sonnet-4-6:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+  claude-sonnet-4-5:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+  claude-haiku-4-5:
+    input_per_1k: 0.0008
+    output_per_1k: 0.004

proxy/internal/llm/model.go

@@ -1,21 +0,0 @@
-package llm
-
-import (
-	sharedllm "github.com/netbirdio/netbird/shared/llm"
-)
-
-// NormalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
-// prefix, and the version/throughput suffix from a Bedrock model id so it
-// matches the catalog/pricing key. Thin delegate to the shared implementation
-// (shared/llm), which management also uses at synthesis time so both sides of
-// the pricing / routing contract normalize identically.
-func NormalizeBedrockModel(modelID string) string {
-	return sharedllm.NormalizeBedrockModel(modelID)
-}
-
-// NormalizeVertexModel strips the "@version" suffix from a Vertex AI model id
-// so it matches the catalog/pricing key. Thin delegate to shared/llm, kept
-// beside NormalizeBedrockModel for the same contract reason.
-func NormalizeVertexModel(modelID string) string {
-	return sharedllm.NormalizeVertexModel(modelID)
-}

proxy/internal/llm/pricing/defaults_coverage_test.go

@@ -0,0 +1,65 @@
+package pricing
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// TestDefaultTable_FirstPartyModelCoverage guards the embedded defaults against
+// silent drift/gaps: every metered first-party model the management catalog
+// enumerates must resolve to a price, and a few rates that previously drifted
+// are pinned to their LiteLLM-validated values. Keep this list in step with the
+// catalog (management/server/agentnetwork/catalog) when adding models.
+func TestDefaultTable_FirstPartyModelCoverage(t *testing.T) {
+	tbl := DefaultTable()
+	require.NotNil(t, tbl, "embedded default pricing table must load")
+
+	mustPrice := map[string][]string{
+		// openai parser covers openai_api, azure_openai_api, and mistral_api.
+		"openai": {
+			"gpt-5.5", "gpt-5.5-pro", "gpt-5.4", "gpt-5.4-mini", "gpt-5.4-nano",
+			"gpt-5.3-codex", "gpt-5.3-chat-latest", "o4-mini",
+			"gpt-4.1", "gpt-4.1-mini", "gpt-4.1-nano", "gpt-4o", "gpt-4o-mini",
+			"gpt-4-turbo", "gpt-3.5-turbo", "gpt-35-turbo",
+			"text-embedding-3-large", "text-embedding-3-small",
+			"mistral-large-latest", "mistral-medium-3-5", "codestral-2508",
+			"ministral-8b-latest", "mistral-embed",
+		},
+		"anthropic": {
+			"claude-fable-5", "claude-opus-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6",
+			"claude-opus-4-1", "claude-sonnet-4-6", "claude-sonnet-4-5", "claude-haiku-4-5",
+		},
+		// bedrock keys are the normalized ids the request parser emits.
+		"bedrock": {
+			"anthropic.claude-opus-5", "anthropic.claude-opus-4-8", "anthropic.claude-opus-4-7", "anthropic.claude-opus-4-6",
+			"anthropic.claude-opus-4-1", "anthropic.claude-sonnet-4-6", "anthropic.claude-sonnet-4-5",
+			"anthropic.claude-haiku-4-5", "meta.llama3-3-70b-instruct",
+			"amazon.nova-pro", "amazon.nova-lite", "amazon.nova-micro", "amazon.nova-2-lite",
+		},
+	}
+	for provider, models := range mustPrice {
+		for _, m := range models {
+			_, ok := tbl.Cost(provider, m, 1000, 1000, 0, 0)
+			assert.True(t, ok, "%s/%s must be priced in the embedded defaults", provider, m)
+		}
+	}
+
+	// Pin per-direction rates independently (input-only then output-only) so a
+	// swap or skew of input<->output that preserves the combined total is still
+	// caught — these are rates that previously drifted or are easy to mis-enter.
+	in, ok := tbl.Cost("openai", "gpt-5.4", 1000, 0, 0, 0)
+	require.True(t, ok)
+	assert.InDelta(t, 0.0025, in, 1e-9, "gpt-5.4 input = 0.0025 per 1k")
+	out, ok := tbl.Cost("openai", "gpt-5.4", 0, 1000, 0, 0)
+	require.True(t, ok)
+	assert.InDelta(t, 0.015, out, 1e-9, "gpt-5.4 output = 0.015 per 1k")
+
+	in, ok = tbl.Cost("bedrock", "anthropic.claude-sonnet-4-5", 1000, 0, 0, 0)
+	require.True(t, ok)
+	assert.InDelta(t, 0.003, in, 1e-9, "bedrock sonnet-4-5 input = 0.003 per 1k")
+	out, ok = tbl.Cost("bedrock", "anthropic.claude-sonnet-4-5", 0, 1000, 0, 0)
+	require.True(t, ok)
+	assert.InDelta(t, 0.015, out, 1e-9, "bedrock sonnet-4-5 output = 0.015 per 1k")
+}

proxy/internal/llm/pricing/defaults_pricing.yaml

@@ -1,196 +1,66 @@
-# Default LLM pricing used by NetBird's Agent Network cost metering.
-# GENERATED from the management catalog — do not edit this file in the
-# repository; regenerate with:
+# Embedded default pricing for llm_observability. Compiled into the proxy
+# binary via go:embed in pricing.go; cost annotation works out of the box
+# without any operator action.
 #
-#   go generate ./management/internals/modules/agentnetwork/pricing
+# Operators override entries by dropping a pricing.yaml into --plugin-data-dir
+# (or whichever basename is given via params.pricing_path). The override file
+# only needs entries the operator wants to change; missing entries fall
+# through to these defaults.
 #
-# Operators: copy this file to <datadir>/defaults_llm_pricing.yaml (or
-# any path configured via management.json:
+# Values are USD per 1_000 tokens. Public list prices drift; ship a fresh
+# binary or override individual entries via the override file as needed.
 #
-#   { "AgentNetwork": { "PricingDefaultsFile": "/path/defaults_llm_pricing.yaml" } }
-#
-# ) and adjust the entries you want to change. Management re-reads the
-# file periodically (mtime poll, every minute): the live table feeds the
-# proxies' cost metering and the dashboard's model-price prefill, so
-# edits apply without a restart. Your file only needs the entries you
-# want to change — but each entry REPLACES the built-in entry for that
-# surface+model whole, so repeat the cache rates you want to keep.
-# Unknown fields and negative or non-finite rates are rejected: at
-# startup that fails boot (for an explicitly configured path); at
-# runtime the previous table is kept and a warning is logged. Deleting
-# the file reverts to the built-in defaults below.
-#
-# Top-level keys are pricing surfaces — the parser shape requests are
-# metered under: "openai" (also Azure, Mistral, and OpenAI-compatible
-# gateways), "anthropic" (also Anthropic-on-Vertex), "bedrock"
-# (normalized ids, e.g. anthropic.claude-sonnet-4-5). Model keys must be
-# the normalized id the proxy meters (version/region suffixes stripped).
-#
-# Values are USD per 1_000 tokens. Optional cache fields:
-#   cached_input_per_1k    OpenAI shape: rate for cached prompt tokens
-#                          (a SUBSET of input tokens). Absent -> cached
-#                          portion bills at input_per_1k.
-#   cache_read_per_1k      Anthropic shape: rate for cache_read tokens
-#                          (ADDITIVE to input). Absent -> input rate.
-#   cache_creation_per_1k  Anthropic shape: rate for cache_creation
-#                          tokens (ADDITIVE to input). Absent -> input
-#                          rate.
-
-anthropic:
-  claude-fable-5:
-    input_per_1k: 0.01
-    output_per_1k: 0.05
-    cache_read_per_1k: 0.001
-    cache_creation_per_1k: 0.0125
-  claude-haiku-4-5:
-    input_per_1k: 0.001
-    output_per_1k: 0.005
-    cache_read_per_1k: 0.0001
-    cache_creation_per_1k: 0.00125
-  claude-opus-4-1:
-    input_per_1k: 0.015
-    output_per_1k: 0.075
-    cache_read_per_1k: 0.0015
-    cache_creation_per_1k: 0.01875
-  claude-opus-4-6:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  claude-opus-4-7:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  claude-opus-4-8:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  claude-opus-5:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  claude-sonnet-4-5:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-  claude-sonnet-4-6:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-  kimi-k3:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cached_input_per_1k: 0.0003
-    cache_read_per_1k: 0.0003
-  "kimi-k3[1m]":
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-
-bedrock:
-  amazon.nova-2-lite:
-    input_per_1k: 0.0003
-    output_per_1k: 0.0025
-  amazon.nova-lite:
-    input_per_1k: 0.00006
-    output_per_1k: 0.00024
-  amazon.nova-micro:
-    input_per_1k: 0.000035
-    output_per_1k: 0.00014
-  amazon.nova-pro:
-    input_per_1k: 0.0008
-    output_per_1k: 0.0032
-  anthropic.claude-haiku-4-5:
-    input_per_1k: 0.001
-    output_per_1k: 0.005
-    cache_read_per_1k: 0.0001
-    cache_creation_per_1k: 0.00125
-  anthropic.claude-opus-4-1:
-    input_per_1k: 0.015
-    output_per_1k: 0.075
-    cache_read_per_1k: 0.0015
-    cache_creation_per_1k: 0.01875
-  anthropic.claude-opus-4-6:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  anthropic.claude-opus-4-7:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  anthropic.claude-opus-4-8:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  anthropic.claude-opus-5:
-    input_per_1k: 0.005
-    output_per_1k: 0.025
-    cache_read_per_1k: 0.0005
-    cache_creation_per_1k: 0.00625
-  anthropic.claude-sonnet-4-5:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-  anthropic.claude-sonnet-4-6:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cache_read_per_1k: 0.0003
-    cache_creation_per_1k: 0.00375
-  meta.llama3-3-70b-instruct:
-    input_per_1k: 0.00072
-    output_per_1k: 0.00072
+# Optional cache fields:
+#   cached_input_per_1k    OpenAI: rate for prompt_tokens_details.cached_tokens
+#                          (a SUBSET of prompt_tokens). Typically 0.5x input.
+#                          Absent → cached portion bills at input_per_1k.
+#   cache_read_per_1k      Anthropic: rate for cache_read_input_tokens
+#                          (ADDITIVE to input_tokens). Typically 0.1x input.
+#                          Absent → cache reads bill at input_per_1k.
+#   cache_creation_per_1k  Anthropic: rate for cache_creation_input_tokens
+#                          (ADDITIVE to input_tokens). Typically 1.25x input.
+#                          Absent → cache writes bill at input_per_1k.
 
 openai:
-  codestral-2508:
-    input_per_1k: 0.0003
-    output_per_1k: 0.0009
-  codestral-latest:
-    input_per_1k: 0.001
-    output_per_1k: 0.003
-  devstral-medium-latest:
-    input_per_1k: 0.0004
-    output_per_1k: 0.002
-  devstral-small-latest:
-    input_per_1k: 0.0001
-    output_per_1k: 0.0003
-  gpt-3.5-turbo:
-    input_per_1k: 0.0005
-    output_per_1k: 0.0015
-  gpt-35-turbo:
-    input_per_1k: 0.0005
-    output_per_1k: 0.0015
-  gpt-4-turbo:
-    input_per_1k: 0.01
+  # OpenAI + OpenAI-compatible providers (openai_api, azure_openai_api,
+  # mistral_api, and the openai-parser gateways) all emit llm.provider="openai",
+  # so their models are priced here. Kept in sync with the management catalog;
+  # rates cross-checked against LiteLLM model_prices_and_context_window.json.
+
+  # GPT-5.x family — cache reads 10% of input (0.1x).
+  gpt-5.5:
+    input_per_1k: 0.005
     output_per_1k: 0.03
-  gpt-4.1:
-    input_per_1k: 0.002
-    output_per_1k: 0.008
     cached_input_per_1k: 0.0005
-  gpt-4.1-mini:
-    input_per_1k: 0.0004
-    output_per_1k: 0.0016
-    cached_input_per_1k: 0.0001
-  gpt-4.1-nano:
-    input_per_1k: 0.0001
-    output_per_1k: 0.0004
-    cached_input_per_1k: 0.000025
-  gpt-4o:
+  gpt-5.5-pro:
+    input_per_1k: 0.03
+    output_per_1k: 0.18
+    cached_input_per_1k: 0.003
+  gpt-5.4:
     input_per_1k: 0.0025
-    output_per_1k: 0.01
-    cached_input_per_1k: 0.00125
-  gpt-4o-mini:
-    input_per_1k: 0.00015
-    output_per_1k: 0.0006
+    output_per_1k: 0.015
+    cached_input_per_1k: 0.00025
+  gpt-5.4-pro:
+    input_per_1k: 0.03
+    output_per_1k: 0.18
+    cached_input_per_1k: 0.003
+  gpt-5.4-mini:
+    input_per_1k: 0.00075
+    output_per_1k: 0.0045
     cached_input_per_1k: 0.000075
+  gpt-5.4-nano:
+    input_per_1k: 0.0002
+    output_per_1k: 0.00125
+    cached_input_per_1k: 0.00002
+  gpt-5.3-codex:
+    input_per_1k: 0.00175
+    output_per_1k: 0.014
+    cached_input_per_1k: 0.000175
+  gpt-5.3-chat-latest:
+    input_per_1k: 0.00175
+    output_per_1k: 0.014
+    cached_input_per_1k: 0.000175
+  # GPT-5 (2025) family — kept for gateway requests using the unsuffixed ids.
   gpt-5:
     input_per_1k: 0.00125
     output_per_1k: 0.01
@@ -203,80 +73,226 @@ openai:
     input_per_1k: 0.00005
     output_per_1k: 0.0004
     cached_input_per_1k: 0.000005
-  gpt-5.3-chat-latest:
-    input_per_1k: 0.00175
-    output_per_1k: 0.014
-    cached_input_per_1k: 0.000175
-  gpt-5.3-codex:
-    input_per_1k: 0.00175
-    output_per_1k: 0.014
-    cached_input_per_1k: 0.000175
-  gpt-5.4:
-    input_per_1k: 0.0025
-    output_per_1k: 0.015
-    cached_input_per_1k: 0.00025
-  gpt-5.4-mini:
-    input_per_1k: 0.00075
-    output_per_1k: 0.0045
-    cached_input_per_1k: 0.000075
-  gpt-5.4-nano:
-    input_per_1k: 0.0002
-    output_per_1k: 0.00125
-    cached_input_per_1k: 0.00002
-  gpt-5.4-pro:
-    input_per_1k: 0.03
-    output_per_1k: 0.18
-    cached_input_per_1k: 0.003
-  gpt-5.5:
-    input_per_1k: 0.005
-    output_per_1k: 0.03
-    cached_input_per_1k: 0.0005
-  gpt-5.5-pro:
-    input_per_1k: 0.03
-    output_per_1k: 0.18
-    cached_input_per_1k: 0.003
-  kimi-k3:
-    input_per_1k: 0.003
-    output_per_1k: 0.015
-    cached_input_per_1k: 0.0003
-    cache_read_per_1k: 0.0003
-  magistral-medium-latest:
-    input_per_1k: 0.002
-    output_per_1k: 0.005
-  magistral-small-latest:
-    input_per_1k: 0.0005
-    output_per_1k: 0.0015
-  ministral-3-14b-2512:
-    input_per_1k: 0.0002
-    output_per_1k: 0.0002
-  ministral-3-3b-2512:
-    input_per_1k: 0.0001
-    output_per_1k: 0.0001
-  ministral-8b-latest:
-    input_per_1k: 0.00015
-    output_per_1k: 0.00015
-  mistral-embed:
-    input_per_1k: 0.0001
-    output_per_1k: 0
-  mistral-large-latest:
-    input_per_1k: 0.0005
-    output_per_1k: 0.0015
-  mistral-medium-3-5:
-    input_per_1k: 0.0015
-    output_per_1k: 0.0075
-  mistral-medium-latest:
-    input_per_1k: 0.0004
-    output_per_1k: 0.002
-  mistral-small-latest:
-    input_per_1k: 0.00006
-    output_per_1k: 0.00018
   o4-mini:
     input_per_1k: 0.0011
     output_per_1k: 0.0044
     cached_input_per_1k: 0.000275
+  # GPT-4.1 family — cache reads 25% of input.
+  gpt-4.1:
+    input_per_1k: 0.002
+    output_per_1k: 0.008
+    cached_input_per_1k: 0.0005
+  gpt-4.1-mini:
+    input_per_1k: 0.0004
+    output_per_1k: 0.0016
+    cached_input_per_1k: 0.0001
+  gpt-4.1-nano:
+    input_per_1k: 0.0001
+    output_per_1k: 0.0004
+    cached_input_per_1k: 0.000025
+  # GPT-4o family — cache reads 50% of input (0.5x).
+  gpt-4o:
+    input_per_1k: 0.0025
+    output_per_1k: 0.01
+    cached_input_per_1k: 0.00125
+  gpt-4o-mini:
+    input_per_1k: 0.00015
+    output_per_1k: 0.0006
+    cached_input_per_1k: 0.000075
+  # Older GPT — no prompt caching.
+  gpt-4-turbo:
+    input_per_1k: 0.01
+    output_per_1k: 0.03
+  gpt-3.5-turbo:
+    input_per_1k: 0.0005
+    output_per_1k: 0.0015
+  gpt-35-turbo: # Azure deployment alias of gpt-3.5-turbo
+    input_per_1k: 0.0005
+    output_per_1k: 0.0015
+  # Embeddings — no caching, no output tokens.
   text-embedding-3-large:
     input_per_1k: 0.00013
     output_per_1k: 0
   text-embedding-3-small:
     input_per_1k: 0.00002
     output_per_1k: 0
+
+  # Mistral (mistral_api) — routed via the openai parser; no prompt caching.
+  mistral-large-latest:
+    input_per_1k: 0.0005
+    output_per_1k: 0.0015
+  mistral-medium-latest:
+    input_per_1k: 0.0004
+    output_per_1k: 0.002
+  mistral-medium-3-5:
+    input_per_1k: 0.0015
+    output_per_1k: 0.0075
+  mistral-small-latest:
+    input_per_1k: 0.00006
+    output_per_1k: 0.00018
+  magistral-medium-latest:
+    input_per_1k: 0.002
+    output_per_1k: 0.005
+  magistral-small-latest:
+    input_per_1k: 0.0005
+    output_per_1k: 0.0015
+  devstral-medium-latest:
+    input_per_1k: 0.0004
+    output_per_1k: 0.002
+  devstral-small-latest:
+    input_per_1k: 0.0001
+    output_per_1k: 0.0003
+  codestral-2508:
+    input_per_1k: 0.0003
+    output_per_1k: 0.0009
+  codestral-latest:
+    input_per_1k: 0.001
+    output_per_1k: 0.003
+  ministral-3-14b-2512:
+    input_per_1k: 0.0002
+    output_per_1k: 0.0002
+  ministral-8b-latest:
+    input_per_1k: 0.00015
+    output_per_1k: 0.00015
+  ministral-3-3b-2512:
+    input_per_1k: 0.0001
+    output_per_1k: 0.0001
+  mistral-embed:
+    input_per_1k: 0.0001
+    output_per_1k: 0
+
+  # Kimi / Moonshot AI (kimi_api) — OpenAI-compatible /v1 endpoint. Moonshot
+  # reports cache hits OpenAI-style when present; cached input is 10% of
+  # input ($0.30 vs $3.00 per MTok). kimi-k3 is the only model the platform
+  # serves newer accounts (K2-era ids and kimi-latest 404), matching the
+  # management catalog.
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cached_input_per_1k: 0.0003
+
+anthropic:
+  # Claude 4.x family — cache reads ≈10% of input, cache writes ≈125% of input.
+  # Pricing source: Anthropic's current published rates per million tokens,
+  # divided by 1000 for the per-1k figures stored here.
+  claude-fable-5:
+    input_per_1k: 0.010
+    output_per_1k: 0.050
+    cache_read_per_1k: 0.001
+    cache_creation_per_1k: 0.0125
+  claude-opus-5:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  claude-opus-4-8:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  claude-opus-4-7:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  claude-opus-4-6:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  claude-opus-4-1:
+    input_per_1k: 0.015
+    output_per_1k: 0.075
+    cache_read_per_1k: 0.0015
+    cache_creation_per_1k: 0.01875
+  claude-sonnet-4-6:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
+  claude-sonnet-4-5:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
+  claude-haiku-4-5:
+    input_per_1k: 0.001
+    output_per_1k: 0.005
+    cache_read_per_1k: 0.0001
+    cache_creation_per_1k: 0.00125
+
+  # Kimi / Moonshot AI (kimi_api) via the Anthropic-compatible endpoint
+  # (/anthropic/v1/messages — the official Claude Code setup). Same rates
+  # as the OpenAI-shape entry above. "kimi-k3[1m]" is the model id some
+  # Claude Code guides set for the 1M-context alias; priced identically so
+  # cost metering doesn't silently skip those requests.
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+  "kimi-k3[1m]":
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+
+bedrock:
+  # AWS Bedrock model ids, normalised by the request parser (cross-region
+  # inference-profile prefix + version/throughput suffix stripped), e.g.
+  # eu.anthropic.claude-sonnet-4-5-20250929-v1:0 -> anthropic.claude-sonnet-4-5.
+  # Anthropic-on-Bedrock keeps the additive cache buckets (read ≈0.1x input,
+  # write ≈1.25x input); Nova / Llama report no cache, so cost is input+output.
+  anthropic.claude-opus-5:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  anthropic.claude-opus-4-8:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  anthropic.claude-opus-4-7:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  anthropic.claude-opus-4-6:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
+  anthropic.claude-opus-4-1:
+    input_per_1k: 0.015
+    output_per_1k: 0.075
+    cache_read_per_1k: 0.0015
+    cache_creation_per_1k: 0.01875
+  anthropic.claude-sonnet-4-6:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
+  anthropic.claude-sonnet-4-5:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
+  anthropic.claude-haiku-4-5:
+    input_per_1k: 0.001
+    output_per_1k: 0.005
+    cache_read_per_1k: 0.0001
+    cache_creation_per_1k: 0.00125
+  meta.llama3-3-70b-instruct:
+    input_per_1k: 0.00072
+    output_per_1k: 0.00072
+  amazon.nova-2-lite:
+    input_per_1k: 0.0003
+    output_per_1k: 0.0025
+  amazon.nova-pro:
+    input_per_1k: 0.0008
+    output_per_1k: 0.0032
+  amazon.nova-lite:
+    input_per_1k: 0.00006
+    output_per_1k: 0.00024
+  amazon.nova-micro:
+    input_per_1k: 0.000035
+    output_per_1k: 0.00014

proxy/internal/llm/pricing/pricing.go

@@ -1,30 +1,102 @@
-// Package pricing implements the pricing table and cost formula the
-// cost_meter middleware uses to convert LLM token usage into a USD cost
-// estimate. The table's content arrives from the management server inside
-// cost_meter's middleware config (synthesized from the catalog plus the
-// operator's stored per-provider prices) — the proxy carries no embedded
-// price list. Price updates ride the ordinary mapping push: a chain
-// rebuild constructs a fresh table, so there is nothing to reload.
+// Package pricing implements the embedded-default + override pricing table
+// shared by middleware that converts LLM token usage into a USD cost
+// estimate. The table is hot-reloadable from a basename under the proxy
+// data directory; missing override files keep the embedded defaults so
+// cost annotation works without operator action.
 package pricing
 
 import (
+	"bytes"
+	"context"
+	_ "embed"
+	"errors"
 	"fmt"
+	"io"
+	"io/fs"
 	"math"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+	"go.opentelemetry.io/otel/attribute"
+	"go.opentelemetry.io/otel/metric"
+	"gopkg.in/yaml.v3"
 )
 
+//go:embed defaults_pricing.yaml
+var defaultPricingYAML []byte
+
+var (
+	defaultTableOnce sync.Once
+	defaultTablePtr  *Table
+)
+
+// DefaultTable returns the pricing table embedded in the binary. The result
+// is parsed once and shared; callers must not mutate the returned value.
+// Cost annotation works without any operator action because every loader
+// starts with this table.
+func DefaultTable() *Table {
+	defaultTableOnce.Do(func() {
+		t, err := parsePricingBytes(defaultPricingYAML)
+		if err != nil {
+			panic(fmt.Sprintf("llmobs: embedded default pricing failed to parse: %v", err))
+		}
+		defaultTablePtr = t
+	})
+	return defaultTablePtr
+}
+
+// mergeOver returns a new Table containing every entry from base, with any
+// matching entry from overlay replacing the base value. Either argument may
+// be nil. Result is a fresh allocation so callers can mutate / Store safely.
+func mergeOver(base, overlay *Table) *Table {
+	if overlay == nil || len(overlay.entries) == 0 {
+		return base
+	}
+	if base == nil || len(base.entries) == 0 {
+		return overlay
+	}
+	out := make(map[string]map[string]Entry, len(base.entries))
+	for provider, models := range base.entries {
+		inner := make(map[string]Entry, len(models))
+		for model, e := range models {
+			inner[model] = e
+		}
+		out[provider] = inner
+	}
+	for provider, models := range overlay.entries {
+		inner, ok := out[provider]
+		if !ok {
+			inner = make(map[string]Entry, len(models))
+			out[provider] = inner
+		}
+		for model, e := range models {
+			inner[model] = e
+		}
+	}
+	return &Table{entries: out}
+}
+
 // Entry is a single model's input and output pricing, expressed in USD per
 // 1000 tokens.
 //
 // CachedInputPer1K applies to OpenAI's cached prompt tokens, which are a
 // subset of input_tokens — when set, the cached portion is billed at this
 // rate and the non-cached remainder at InputPer1K. Zero means "no discount
-// configured", and cached tokens are billed at InputPer1K.
+// configured", and cached tokens are billed at InputPer1K (matches current
+// behaviour where cached counts weren't extracted at all).
 //
 // CacheReadPer1K and CacheCreationPer1K apply to Anthropic's two prompt-
 // cache fields, which are additive to input_tokens: cache_read is the
 // cheaper read-from-cache rate, cache_creation is the more expensive
 // write-to-cache rate. Zero means "no rate configured" and the
-// corresponding token bucket is billed at InputPer1K.
+// corresponding token bucket is billed at InputPer1K. This is more
+// accurate than today's behaviour, where Anthropic's cache tokens are
+// ignored and not charged at all.
 type Entry struct {
 	InputPer1K         float64
 	OutputPer1K        float64
@@ -33,102 +105,33 @@ type Entry struct {
 	CacheCreationPer1K float64
 }
 
-// EntryJSON is the wire shape of a pricing entry inside cost_meter's
-// middleware config. Field names are the management→proxy contract; the
-// management synthesizer marshals the same names (its pricing.Entry).
-type EntryJSON struct {
-	InputPer1K         float64 `json:"input_per_1k"`
-	OutputPer1K        float64 `json:"output_per_1k"`
-	CachedInputPer1K   float64 `json:"cached_input_per_1k"`
-	CacheReadPer1K     float64 `json:"cache_read_per_1k"`
-	CacheCreationPer1K float64 `json:"cache_creation_per_1k"`
-}
-
-// Table is a provider-surface-to-model pricing lookup. Instances are
-// immutable once built; a mapping update builds a whole new middleware
-// instance (and with it a new table) rather than mutating this one.
+// Table is a provider-to-model pricing lookup. Instances are immutable once
+// built and are swapped atomically by Loader.
 type Table struct {
 	entries map[string]map[string]Entry
 }
 
-// NewEntries validates and converts a wire-shape map (surface-or-record ->
-// model -> rates) into the internal representation. Every rate must be a
-// finite, non-negative USD amount; a violation is returned as an error so
-// a corrupt config fails the chain build loudly instead of mispricing.
-// Management validates the same constraints at its API boundary, so this
-// is defense-in-depth. Nil input yields an empty (never-matching) map.
-func NewEntries(raw map[string]map[string]EntryJSON) (map[string]map[string]Entry, error) {
-	out := make(map[string]map[string]Entry, len(raw))
-	for outer, models := range raw {
-		inner := make(map[string]Entry, len(models))
-		for model, e := range models {
-			for field, v := range map[string]float64{
-				"input_per_1k":          e.InputPer1K,
-				"output_per_1k":         e.OutputPer1K,
-				"cached_input_per_1k":   e.CachedInputPer1K,
-				"cache_read_per_1k":     e.CacheReadPer1K,
-				"cache_creation_per_1k": e.CacheCreationPer1K,
-			} {
-				if v < 0 || math.IsNaN(v) || math.IsInf(v, 0) {
-					return nil, fmt.Errorf("pricing %s/%s: %s must be a finite, non-negative rate, got %v", outer, model, field, v)
-				}
-			}
-			// EntryJSON and Entry are field-identical (tags aside), so a
-			// direct conversion carries all five rates.
-			inner[model] = Entry(e)
-		}
-		out[outer] = inner
-	}
-	return out, nil
-}
-
-// NewTable builds an immutable Table from the wire-shape defaults map.
-// See NewEntries for validation semantics.
-func NewTable(raw map[string]map[string]EntryJSON) (*Table, error) {
-	entries, err := NewEntries(raw)
-	if err != nil {
-		return nil, err
-	}
-	return &Table{entries: entries}, nil
-}
-
-// Lookup returns the entry for the given provider surface and model.
-func (t *Table) Lookup(provider, model string) (Entry, bool) {
-	if t == nil {
-		return Entry{}, false
-	}
-	byModel, ok := t.entries[provider]
-	if !ok {
-		return Entry{}, false
-	}
-	e, ok := byModel[model]
-	return e, ok
-}
-
-// Has reports whether the provider/model pair is present in the table.
-func (t *Table) Has(provider, model string) bool {
-	_, ok := t.Lookup(provider, model)
-	return ok
-}
-
 // Cost returns the estimated USD cost for the given token counts. ok is
 // false when the provider or model is not present in the table; the caller
 // can still emit token metrics with a model=unknown label.
+//
+// Provider-shape semantics for cached / cache-creation counts:
+//
+//   - OpenAI: cachedInput is a SUBSET of inTokens. The cached portion is
+//     billed at CachedInputPer1K (or InputPer1K when no override), and the
+//     non-cached remainder of inTokens at InputPer1K. cacheCreation is
+//     ignored (OpenAI has no analogue).
+//   - Anthropic: cachedInput (cache_read) and cacheCreation are ADDITIVE to
+//     inTokens. The three buckets are billed at CacheReadPer1K,
+//     CacheCreationPer1K, and InputPer1K respectively, each falling back
+//     to InputPer1K when the corresponding rate is zero.
+//   - Other providers: cached and cacheCreation are ignored; cost is
+//     inTokens*InputPer1K + outTokens*OutputPer1K.
 func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (float64, bool) {
 	c, ok := t.Costs(provider, model, inTokens, outTokens, cachedInput, cacheCreation)
 	return c.TotalUSD, ok
 }
 
-// Costs returns the estimated USD cost split for the given token counts.
-// The provider surface selects the cache formula; see EntryCosts.
-func (t *Table) Costs(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (Costs, bool) {
-	entry, ok := t.Lookup(provider, model)
-	if !ok {
-		return Costs{}, false
-	}
-	return EntryCosts(entry, provider, inTokens, outTokens, cachedInput, cacheCreation), true
-}
-
 // Costs is a per-request cost split. The four per-bucket fields are the base
 // of the breakdown — one per token bucket the provider bills separately — and
 // the two aggregates are derived from them:
@@ -162,25 +165,9 @@ func newCosts(input, cachedInput, cacheCreation, output float64) Costs {
 	}
 }
 
-// EntryCosts computes the USD cost split for the given entry and token
-// counts. The surface (the llm.provider value the request parser stamps)
-// selects the cache formula; the entry may come from the surface-keyed
-// defaults table or from a per-provider-record override — the math is
-// identical either way.
-//
-// Provider-shape semantics for cached / cache-creation counts:
-//
-//   - "openai": cachedInput is a SUBSET of inTokens. The cached portion is
-//     billed at CachedInputPer1K (or InputPer1K when no override), and the
-//     non-cached remainder of inTokens at InputPer1K. cacheCreation is
-//     ignored (OpenAI has no analogue).
-//   - "anthropic", "bedrock": cachedInput (cache_read) and cacheCreation are
-//     ADDITIVE to inTokens. The three buckets are billed at CacheReadPer1K,
-//     CacheCreationPer1K, and InputPer1K respectively, each falling back
-//     to InputPer1K when the corresponding rate is zero.
-//   - Other surfaces: cached and cacheCreation are ignored; cost is
-//     inTokens*InputPer1K + outTokens*OutputPer1K.
-func EntryCosts(entry Entry, surface string, inTokens, outTokens, cachedInput, cacheCreation int64) Costs {
+// Costs returns the estimated USD cost split for the given token counts, with
+// the same semantics as Cost.
+func (t *Table) Costs(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (Costs, bool) {
 	// Clamp negatives to zero before any pricing math so a malformed
 	// upstream count can never produce a negative cost.
 	if inTokens < 0 {
@@ -195,8 +182,19 @@ func EntryCosts(entry Entry, surface string, inTokens, outTokens, cachedInput, c
 	if cacheCreation < 0 {
 		cacheCreation = 0
 	}
+	if t == nil {
+		return Costs{}, false
+	}
+	byModel, ok := t.entries[provider]
+	if !ok {
+		return Costs{}, false
+	}
+	entry, ok := byModel[model]
+	if !ok {
+		return Costs{}, false
+	}
 	output := (float64(outTokens) / 1000.0) * entry.OutputPer1K
-	switch surface {
+	switch provider {
 	case "openai":
 		// cachedInput is a subset of inTokens; clamp so a malformed
 		// upstream (cached > total) can't produce a negative remainder.
@@ -210,7 +208,7 @@ func EntryCosts(entry Entry, surface string, inTokens, outTokens, cachedInput, c
 		}
 		nonCached := float64(inTokens-clamped) / 1000.0 * entry.InputPer1K
 		cached := float64(clamped) / 1000.0 * cachedRate
-		return newCosts(nonCached, cached, 0, output)
+		return newCosts(nonCached, cached, 0, output), true
 	case "anthropic", "bedrock":
 		// Bedrock-Anthropic returns the same additive cache buckets as
 		// first-party Anthropic; non-Anthropic Bedrock models simply report
@@ -226,9 +224,266 @@ func EntryCosts(entry Entry, surface string, inTokens, outTokens, cachedInput, c
 		input := float64(inTokens) / 1000.0 * entry.InputPer1K
 		read := float64(cachedInput) / 1000.0 * readRate
 		create := float64(cacheCreation) / 1000.0 * createRate
-		return newCosts(input, read, create, output)
+		return newCosts(input, read, create, output), true
 	default:
 		input := float64(inTokens) / 1000.0 * entry.InputPer1K
-		return newCosts(input, 0, 0, output)
+		return newCosts(input, 0, 0, output), true
 	}
 }
+
+// Has reports whether the provider/model pair is present in the table.
+func (t *Table) Has(provider, model string) bool {
+	if t == nil {
+		return false
+	}
+	byModel, ok := t.entries[provider]
+	if !ok {
+		return false
+	}
+	_, ok = byModel[model]
+	return ok
+}
+
+// pricingFile mirrors the on-disk YAML schema. Keys are provider names; the
+// nested map keys are model names.
+type pricingFile map[string]map[string]struct {
+	InputPer1K         float64 `yaml:"input_per_1k"`
+	OutputPer1K        float64 `yaml:"output_per_1k"`
+	CachedInputPer1K   float64 `yaml:"cached_input_per_1k"`
+	CacheReadPer1K     float64 `yaml:"cache_read_per_1k"`
+	CacheCreationPer1K float64 `yaml:"cache_creation_per_1k"`
+}
+
+const (
+	// ReloadInterval is the mtime-poll cadence for the background reloader.
+	ReloadInterval = 30 * time.Second
+
+	// errorBackoff bounds how often the loader logs a repeated parse error.
+	errorBackoff = 5 * time.Minute
+)
+
+var basenameRegex = regexp.MustCompile(`^[a-zA-Z0-9._-]+$`)
+
+// Loader is a confined, hot-reloadable pricing table reader. Construction
+// must succeed against the target file; subsequent reload failures keep the
+// previously-loaded table so callers never observe a blank price list.
+type Loader struct {
+	baseDir  string
+	fullPath string
+	pluginID string
+	table    atomic.Pointer[Table]
+	mtime    atomic.Int64
+	failures metric.Int64Counter
+	interval time.Duration
+}
+
+// NewLoader returns a pricing loader that overlays an optional file-based
+// table on top of the embedded defaults. Missing override file, baseDir, or
+// relPath is not an error: the loader keeps the embedded defaults so cost
+// metadata is still emitted for known models.
+//
+// Errors:
+//   - bad basename, traversal segment, or absolute relPath are rejected so a
+//     misconfigured target surfaces immediately.
+//   - permission errors and YAML parse errors keep the defaults but log a
+//     warning; cost annotation does not silently break.
+//
+// failures is optional; pass nil in tests that do not care about
+// reload-failure telemetry.
+func NewLoader(baseDir, relPath, pluginID string, failures metric.Int64Counter) (*Loader, error) {
+	defaults := DefaultTable()
+	l := &Loader{
+		baseDir:  baseDir,
+		pluginID: pluginID,
+		failures: failures,
+	}
+	l.table.Store(defaults)
+
+	if strings.TrimSpace(baseDir) == "" || strings.TrimSpace(relPath) == "" {
+		return l, nil
+	}
+
+	full, err := resolveMiddlewareDataPath(baseDir, relPath)
+	if err != nil {
+		return nil, err
+	}
+	l.fullPath = full
+
+	overlay, mtime, err := loadPricing(full)
+	if err != nil {
+		if errors.Is(err, fs.ErrNotExist) {
+			// Override file is optional. Defaults already stored.
+			return l, nil
+		}
+		// Symlink rejection, oversize file, parse failure, permission errors
+		// — surface so a misconfigured operator sees the problem instead of
+		// silently running with stale defaults.
+		return nil, fmt.Errorf("load pricing %s: %w", full, err)
+	}
+	l.table.Store(mergeOver(defaults, overlay))
+	l.mtime.Store(mtime.UnixNano())
+	return l, nil
+}
+
+// Get returns the current pricing table. The returned pointer is immutable;
+// callers must not mutate its contents.
+func (l *Loader) Get() *Table {
+	if l == nil {
+		return nil
+	}
+	return l.table.Load()
+}
+
+// WatchesFile reports whether this loader is bound to an override file on
+// disk. False for defaults-only loaders (no operator override given).
+// Callers use this to decide whether to spawn the mtime-poll goroutine.
+func (l *Loader) WatchesFile() bool {
+	if l == nil {
+		return false
+	}
+	return l.fullPath != ""
+}
+
+// SetReloadInterval overrides the mtime-poll cadence used by Reload. Calls
+// after Reload has started have no effect on the running loop. Intended for
+// tests; production code uses the default ReloadInterval.
+func (l *Loader) SetReloadInterval(d time.Duration) {
+	if l == nil || d <= 0 {
+		return
+	}
+	l.interval = d
+}
+
+// Reload runs a polling loop that checks the pricing file mtime every
+// ReloadInterval (or the value passed to SetReloadInterval). Returns when
+// ctx is cancelled.
+func (l *Loader) Reload(ctx context.Context) {
+	if l == nil {
+		return
+	}
+	interval := l.interval
+	if interval <= 0 {
+		interval = ReloadInterval
+	}
+	t := time.NewTicker(interval)
+	defer t.Stop()
+
+	var lastErrAt time.Time
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-t.C:
+			if err := l.reload(); err != nil {
+				if l.failures != nil {
+					l.failures.Add(ctx, 1, metric.WithAttributes(
+						attribute.String("plugin", l.pluginID),
+					))
+				}
+				now := time.Now()
+				if now.Sub(lastErrAt) >= errorBackoff {
+					log.Warnf("llmobs: pricing reload failed for %s: %v", l.fullPath, err)
+					lastErrAt = now
+				}
+			}
+		}
+	}
+}
+
+// reload performs a single-shot mtime check and reload. The reloaded
+// override file is merged on top of the embedded defaults; missing override
+// (e.g. operator deleted the file) is not an error and reverts to defaults.
+func (l *Loader) reload() error {
+	if l.fullPath == "" {
+		// Defaults-only loader; nothing on disk to reload.
+		return nil
+	}
+	mtime, err := statMtime(l.fullPath)
+	if err != nil {
+		if errors.Is(err, fs.ErrNotExist) {
+			// File was removed since startup. Drop back to defaults and
+			// reset mtime so a future re-creation triggers a reload.
+			l.table.Store(DefaultTable())
+			l.mtime.Store(0)
+			return nil
+		}
+		return err
+	}
+	if mtime.UnixNano() == l.mtime.Load() {
+		return nil
+	}
+
+	overlay, newMtime, err := loadPricing(l.fullPath)
+	if err != nil {
+		return err
+	}
+	l.table.Store(mergeOver(DefaultTable(), overlay))
+	l.mtime.Store(newMtime.UnixNano())
+	return nil
+}
+
+// resolveMiddlewareDataPath validates relPath is a safe basename and resolves
+// it under baseDir. An additional cleaned-prefix check guards against
+// CVE-style edge cases where Join is used with trailing path segments.
+func resolveMiddlewareDataPath(baseDir, relPath string) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		return "", errors.New("middleware-data-dir is not configured")
+	}
+	if relPath == "" {
+		return "", errors.New("pricing path is empty")
+	}
+	if !basenameRegex.MatchString(relPath) {
+		return "", fmt.Errorf("pricing path %q is not a safe basename", relPath)
+	}
+	if filepath.IsAbs(relPath) {
+		return "", fmt.Errorf("pricing path %q must be a basename, not absolute", relPath)
+	}
+
+	cleanBase, err := filepath.Abs(filepath.Clean(baseDir))
+	if err != nil {
+		return "", fmt.Errorf("resolve middleware-data-dir: %w", err)
+	}
+	full := filepath.Join(cleanBase, relPath)
+	cleanedFull := filepath.Clean(full)
+	if !strings.HasPrefix(cleanedFull, cleanBase+string(filepath.Separator)) && cleanedFull != cleanBase {
+		return "", fmt.Errorf("pricing path %q escapes middleware-data-dir", relPath)
+	}
+	return cleanedFull, nil
+}
+
+func parsePricingBytes(data []byte) (*Table, error) {
+	dec := yaml.NewDecoder(bytes.NewReader(data))
+	dec.KnownFields(true)
+
+	var raw pricingFile
+	if err := dec.Decode(&raw); err != nil && !errors.Is(err, io.EOF) {
+		return nil, fmt.Errorf("decode pricing yaml: %w", err)
+	}
+
+	out := make(map[string]map[string]Entry, len(raw))
+	for provider, models := range raw {
+		inner := make(map[string]Entry, len(models))
+		for model, entry := range models {
+			for field, v := range map[string]float64{
+				"input_per_1k":          entry.InputPer1K,
+				"output_per_1k":         entry.OutputPer1K,
+				"cached_input_per_1k":   entry.CachedInputPer1K,
+				"cache_read_per_1k":     entry.CacheReadPer1K,
+				"cache_creation_per_1k": entry.CacheCreationPer1K,
+			} {
+				if v < 0 || math.IsNaN(v) || math.IsInf(v, 0) {
+					return nil, fmt.Errorf("pricing %s/%s: %s must be a finite, non-negative rate, got %v", provider, model, field, v)
+				}
+			}
+			inner[model] = Entry{
+				InputPer1K:         entry.InputPer1K,
+				OutputPer1K:        entry.OutputPer1K,
+				CachedInputPer1K:   entry.CachedInputPer1K,
+				CacheReadPer1K:     entry.CacheReadPer1K,
+				CacheCreationPer1K: entry.CacheCreationPer1K,
+			}
+		}
+		out[provider] = inner
+	}
+	return &Table{entries: out}, nil
+}

proxy/internal/llm/pricing/pricing_other.go

@@ -0,0 +1,20 @@
+//go:build !unix
+
+package pricing
+
+import (
+	"fmt"
+	"time"
+)
+
+// loadPricing is unavailable on non-Unix platforms because O_NOFOLLOW and
+// fstat-from-FD are required to honour the spec's symlink-safety rules. The
+// proxy is only deployed on Linux today; a Windows port would need an
+// equivalent path-as-handle implementation.
+func loadPricing(path string) (*Table, time.Time, error) {
+	return nil, time.Time{}, fmt.Errorf("llmobs pricing loader is not supported on this platform: %s", path)
+}
+
+func statMtime(path string) (time.Time, error) {
+	return time.Time{}, fmt.Errorf("llmobs pricing loader is not supported on this platform: %s", path)
+}

proxy/internal/llm/pricing/pricing_test.go

@@ -1,13 +1,47 @@
+//go:build unix
+
 package pricing
 
 import (
-	"math"
+	"context"
+	"os"
+	"path/filepath"
 	"testing"
+	"time"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
 
+func copyFixture(t *testing.T, src, dst string) {
+	t.Helper()
+	data, err := os.ReadFile(src)
+	require.NoError(t, err, "read source fixture")
+	require.NoError(t, os.WriteFile(dst, data, 0o600), "write target fixture")
+}
+
+func TestNewLoader_HappyPath(t *testing.T) {
+	base := t.TempDir()
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
+
+	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.NoError(t, err, "NewLoader must succeed with a valid fixture")
+	table := l.Get()
+	require.NotNil(t, table, "table populated after load")
+
+	cost, ok := table.Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
+	require.True(t, ok, "known provider/model resolves")
+	assert.InDelta(t, 0.00075, cost, 1e-9, "cost = 0.00015 + 0.0006 per 1k tokens")
+
+	cost, ok = table.Cost("openai", "gpt-4o", 2000, 1000, 0, 0)
+	require.True(t, ok, "second known model resolves")
+	assert.InDelta(t, 0.015, cost, 1e-9, "cost for gpt-4o: 2*0.0025 + 1*0.01")
+
+	cost, ok = table.Cost("anthropic", "claude-sonnet-4-5", 1000, 1000, 0, 0)
+	require.True(t, ok, "anthropic model resolves")
+	assert.InDelta(t, 0.018, cost, 1e-9, "cost for claude-sonnet-4-5: 0.003 + 0.015")
+}
+
 // TestCost_OpenAICachedSubsetDiscount proves OpenAI's cached input
 // tokens are billed at the configured cached_input_per_1k rate while
 // the non-cached remainder of input_tokens is billed at the regular
@@ -31,9 +65,11 @@ func TestCost_OpenAICachedSubsetDiscount(t *testing.T) {
 		"cached subset must bill at the discount rate; non-cached remainder at regular rate")
 }
 
-// TestCost_OpenAICachedFallsBackToInputRate covers the fallback
-// contract: when CachedInputPer1K is unset (zero), cached tokens bill
-// at the regular input rate.
+// TestCost_OpenAICachedFallsBackToInputRate covers the operator
+// opt-in contract: when CachedInputPer1K is unset (zero), cached
+// tokens bill at the regular input rate. This matches today's
+// behaviour (cached counts weren't extracted at all so they
+// implicitly billed at the input rate via prompt_tokens).
 func TestCost_OpenAICachedFallsBackToInputRate(t *testing.T) {
 	tbl := &Table{entries: map[string]map[string]Entry{
 		"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01}},
@@ -42,7 +78,7 @@ func TestCost_OpenAICachedFallsBackToInputRate(t *testing.T) {
 	require.True(t, ok)
 	want := 0.0025 + (500.0/1000.0)*0.01
 	assert.InDelta(t, want, cost, 1e-12,
-		"absent cached_input_per_1k rate must fall back to input_per_1k")
+		"absent cached_input_per_1k rate must fall back to input_per_1k — same as pre-feature behaviour")
 }
 
 // TestCost_OpenAIClampsCachedToInputCount is the defensive guard
@@ -64,7 +100,10 @@ func TestCost_OpenAIClampsCachedToInputCount(t *testing.T) {
 // TestCost_AnthropicCacheReadAndCreationAreAdditive proves the
 // Anthropic shape: cache_read and cache_creation tokens are
 // ADDITIVE to input_tokens (not subset), each billed at its own
-// configured rate.
+// configured rate. The two rates pull in opposite directions —
+// cache_read is the cheaper read-from-cache rate (≈0.1× input),
+// cache_creation is the more expensive write-to-cache rate
+// (≈1.25× input).
 func TestCost_AnthropicCacheReadAndCreationAreAdditive(t *testing.T) {
 	tbl := &Table{entries: map[string]map[string]Entry{
 		"anthropic": {"claude-sonnet": {
@@ -86,9 +125,11 @@ func TestCost_AnthropicCacheReadAndCreationAreAdditive(t *testing.T) {
 		"each Anthropic input bucket must bill at its own configured rate")
 }
 
-// TestCost_AnthropicCacheRatesFallBackToInput covers the no-rate
+// TestCost_AnthropicCacheRatesFallBackToInput covers the no-opt-in
 // path: when neither CacheReadPer1K nor CacheCreationPer1K is set,
-// cache tokens bill at the regular input rate.
+// cache tokens bill at the regular input rate. This is more
+// accurate than today's behaviour (cache tokens ignored entirely)
+// without requiring operators to opt in via YAML.
 func TestCost_AnthropicCacheRatesFallBackToInput(t *testing.T) {
 	tbl := &Table{entries: map[string]map[string]Entry{
 		"anthropic": {"claude-sonnet": {InputPer1K: 0.003, OutputPer1K: 0.015}},
@@ -98,39 +139,259 @@ func TestCost_AnthropicCacheRatesFallBackToInput(t *testing.T) {
 	// Without overrides: every input bucket at input_per_1k.
 	want := ((256.0+768.0+512.0)/1000.0)*0.003 + (200.0/1000.0)*0.015
 	assert.InDelta(t, want, cost, 1e-12,
-		"absent cache rates must fall back to input_per_1k")
+		"absent cache rates must fall back to input_per_1k — Anthropic cache tokens were ignored before this change, billing at input rate is more accurate as a default")
 }
 
-// TestEntryCosts_SurfaceSelectsFormula pins that the formula branches on
-// the SURFACE, not on which table the entry came from: the same entry
-// bills a subset carve-out on "openai", additive buckets on
-// "anthropic"/"bedrock", and ignores cache counts everywhere else. This
-// is what keeps per-provider-record entries (looked up by record id)
-// mathematically identical to defaults-table entries.
-func TestEntryCosts_SurfaceSelectsFormula(t *testing.T) {
-	e := Entry{InputPer1K: 0.002, OutputPer1K: 0.01, CachedInputPer1K: 0.001, CacheReadPer1K: 0.0002, CacheCreationPer1K: 0.0025}
+func TestNewLoader_UnknownModel(t *testing.T) {
+	base := t.TempDir()
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
 
-	openai := EntryCosts(e, "openai", 1000, 0, 400, 300)
-	assert.InDelta(t, (600.0/1000.0)*0.002+(400.0/1000.0)*0.001, openai.TotalUSD, 1e-12,
-		"openai: cached is a subset, cacheCreation ignored")
+	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.NoError(t, err)
 
-	anthropic := EntryCosts(e, "anthropic", 1000, 0, 400, 300)
-	assert.InDelta(t, 0.002+(400.0/1000.0)*0.0002+(300.0/1000.0)*0.0025, anthropic.TotalUSD, 1e-12,
-		"anthropic: cache buckets are additive")
+	_, ok := l.Get().Cost("openai", "fantasy-model", 10, 10, 0, 0)
+	assert.False(t, ok, "unknown model returns ok=false")
 
-	bedrock := EntryCosts(e, "bedrock", 1000, 0, 400, 300)
-	assert.InDelta(t, anthropic.TotalUSD, bedrock.TotalUSD, 1e-12, "bedrock shares the anthropic formula")
-
-	other := EntryCosts(e, "gemini", 1000, 0, 400, 300)
-	assert.InDelta(t, 0.002, other.TotalUSD, 1e-12, "unknown surface: cache counts ignored")
+	_, ok = l.Get().Cost("cohere", "anything", 10, 10, 0, 0)
+	assert.False(t, ok, "unknown provider returns ok=false")
 }
 
-// TestEntryCosts_ClampsNegativeTokens: malformed upstream counts must
-// never produce a negative cost.
-func TestEntryCosts_ClampsNegativeTokens(t *testing.T) {
-	e := Entry{InputPer1K: 0.002, OutputPer1K: 0.01}
-	c := EntryCosts(e, "openai", -50, -10, -5, -3)
-	assert.Zero(t, c.TotalUSD, "all-negative counts clamp to zero cost")
+func TestNewLoader_InvalidYAMLRejected(t *testing.T) {
+	base := t.TempDir()
+	require.NoError(t, os.WriteFile(filepath.Join(base, "pricing.yaml"), []byte("\t- this is not: valid: yaml: :["), 0o600))
+
+	_, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.Error(t, err, "invalid YAML must surface as construction error")
+}
+
+func TestLoader_ReloadKeepsPreviousOnParseError(t *testing.T) {
+	base := t.TempDir()
+	target := filepath.Join(base, "pricing.yaml")
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
+
+	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.NoError(t, err)
+	require.NotNil(t, l.Get(), "initial table populated")
+
+	// Overwrite with content that violates the strict schema (extra field)
+	// plus a bumped mtime to trigger reload.
+	require.NoError(t, os.WriteFile(target, []byte("openai:\n  gpt-4o:\n    input_per_1k: 1.0\n    output_per_1k: 2.0\n    bogus_field: nope\n"), 0o600))
+	future := time.Now().Add(time.Hour)
+	require.NoError(t, os.Chtimes(target, future, future))
+
+	err = l.reload()
+	require.Error(t, err, "parse error surfaced by reload()")
+
+	cost, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
+	require.True(t, ok, "previous table still available after parse failure")
+	assert.InDelta(t, 0.00075, cost, 1e-9, "previous cost preserved")
+}
+
+func TestLoader_ReloadNoChangeIsNoOp(t *testing.T) {
+	base := t.TempDir()
+	target := filepath.Join(base, "pricing.yaml")
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
+
+	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.NoError(t, err)
+	ptrBefore := l.Get()
+
+	require.NoError(t, l.reload(), "no-change reload must not error")
+	ptrAfter := l.Get()
+	assert.Same(t, ptrBefore, ptrAfter, "table pointer unchanged when mtime unchanged")
+}
+
+func TestLoader_ReloadDetectsChange(t *testing.T) {
+	base := t.TempDir()
+	target := filepath.Join(base, "pricing.yaml")
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
+
+	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.NoError(t, err)
+
+	updated := []byte("openai:\n  gpt-4o-mini:\n    input_per_1k: 1.00\n    output_per_1k: 2.00\n")
+	require.NoError(t, os.WriteFile(target, updated, 0o600))
+	future := time.Now().Add(time.Hour)
+	require.NoError(t, os.Chtimes(target, future, future))
+
+	require.NoError(t, l.reload(), "reload must succeed on valid new content")
+
+	cost, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
+	require.True(t, ok, "updated model still present")
+	assert.InDelta(t, 3.0, cost, 0.0001, "new prices are applied: 1 + 2 per 1k")
+}
+
+// TestLoader_ReloadGoroutinePicksUpChanges proves the background goroutine
+// started via Reload actually swaps the pricing table when the file changes
+// on disk. Without that goroutine running, pricing edits would never reach
+// requests until a proxy restart.
+func TestLoader_ReloadGoroutinePicksUpChanges(t *testing.T) {
+	base := t.TempDir()
+	target := filepath.Join(base, "pricing.yaml")
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
+
+	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.NoError(t, err)
+	l.SetReloadInterval(20 * time.Millisecond)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	done := make(chan struct{})
+	go func() {
+		l.Reload(ctx)
+		close(done)
+	}()
+
+	// Before any rewrite, the loader holds the fixture's prices.
+	costBefore, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
+	require.True(t, ok, "fixture model must resolve initially")
+	assert.InDelta(t, 0.00075, costBefore, 1e-9, "fixture prices apply before rewrite")
+
+	updated := []byte("openai:\n  gpt-4o-mini:\n    input_per_1k: 1.00\n    output_per_1k: 2.00\n")
+	require.NoError(t, os.WriteFile(target, updated, 0o600))
+	future := time.Now().Add(time.Hour)
+	require.NoError(t, os.Chtimes(target, future, future))
+
+	deadline := time.Now().Add(2 * time.Second)
+	for {
+		cost, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
+		if ok && cost > 2.5 {
+			break
+		}
+		if time.Now().After(deadline) {
+			t.Fatalf("background reloader did not pick up rewrite within deadline")
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+
+	cancel()
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("Reload loop did not exit after cancel")
+	}
+}
+
+func TestLoader_ReloadBackgroundLoopCancellation(t *testing.T) {
+	base := t.TempDir()
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
+	l, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.NoError(t, err)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() {
+		l.Reload(ctx)
+		close(done)
+	}()
+	cancel()
+
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("Reload loop did not exit on context cancel")
+	}
+}
+
+func TestNewLoader_PathValidation(t *testing.T) {
+	base := t.TempDir()
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
+
+	cases := []struct {
+		name    string
+		relPath string
+	}{
+		{"traversal", "../../etc/passwd"},
+		{"absolute", "/etc/passwd"},
+		{"slash in basename", "sub/pricing.yaml"},
+		{"control chars", "pricing\x00.yaml"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := NewLoader(base, tc.relPath, "llm_observability", nil)
+			require.Error(t, err, "NewLoader must reject %q", tc.relPath)
+		})
+	}
+
+	// Empty relPath is no longer a validation error: the loader treats it
+	// as "no override file, defaults only" so cost metadata is still
+	// emitted for the embedded models out of the box.
+	t.Run("empty falls back to defaults", func(t *testing.T) {
+		l, err := NewLoader(base, "", "llm_observability", nil)
+		require.NoError(t, err, "empty relPath should yield a defaults-only loader")
+		require.NotNil(t, l, "loader must be returned")
+		require.False(t, l.WatchesFile(), "no file watching when no override is given")
+		_, ok := l.Get().Cost("openai", "gpt-4o-mini", 1000, 1000, 0, 0)
+		assert.True(t, ok, "embedded defaults should still resolve gpt-4o-mini")
+	})
+}
+
+// TestNewLoader_PathValidation_Extended covers the remaining attack shapes
+// called out in C2: dot references, embedded traversal segments, and a
+// newline in the basename. The basename regex must reject each one even
+// though filepath.Clean would otherwise collapse them.
+func TestNewLoader_PathValidation_Extended(t *testing.T) {
+	base := t.TempDir()
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing.yaml"))
+
+	cases := []struct {
+		name    string
+		relPath string
+	}{
+		{"dot", "."},
+		{"dotdot", ".."},
+		{"relative traversal", "../pricing.yaml"},
+		{"embedded slash", "pri/cing.yaml"},
+		{"newline", "pricing\n.yaml"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := NewLoader(base, tc.relPath, "llm_observability", nil)
+			require.Error(t, err, "NewLoader must reject %q", tc.relPath)
+		})
+	}
+}
+
+// TestNewLoader_ValidBasenameLoads proves the allowlist is exclusive: a
+// basename containing only safe characters under baseDir loads. Without this
+// a regression that over-tightened the regex would silently break valid
+// deployments.
+func TestNewLoader_ValidBasenameLoads(t *testing.T) {
+	base := t.TempDir()
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), filepath.Join(base, "pricing-v2_prod.yaml"))
+
+	l, err := NewLoader(base, "pricing-v2_prod.yaml", "llm_observability", nil)
+	require.NoError(t, err, "basename with _, -, . must load")
+	require.NotNil(t, l.Get(), "table populated")
+}
+
+// TestNewLoader_SymlinkOutsideBaseDirRejected constructs a symlink under
+// baseDir that points to a file outside it. O_NOFOLLOW must refuse to open
+// the symlink even though the symlink path itself is a valid basename under
+// baseDir.
+func TestNewLoader_SymlinkOutsideBaseDirRejected(t *testing.T) {
+	outside := t.TempDir()
+	target := filepath.Join(outside, "evil.yaml")
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), target)
+
+	base := t.TempDir()
+	link := filepath.Join(base, "pricing.yaml")
+	require.NoError(t, os.Symlink(target, link), "symlink setup")
+
+	_, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.Error(t, err, "O_NOFOLLOW must reject symlink even when it points outside baseDir")
+}
+
+func TestNewLoader_SymlinkRejected(t *testing.T) {
+	base := t.TempDir()
+	concrete := filepath.Join(base, "real.yaml")
+	copyFixture(t, filepath.Join("..", "fixtures", "pricing.yaml"), concrete)
+
+	link := filepath.Join(base, "pricing.yaml")
+	require.NoError(t, os.Symlink(concrete, link), "symlink setup")
+
+	_, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.Error(t, err, "O_NOFOLLOW must reject symlinked targets")
 }
 
 func TestTableCost_NilSafe(t *testing.T) {
@@ -141,37 +402,31 @@ func TestTableCost_NilSafe(t *testing.T) {
 	assert.False(t, t1.Has("x", "y"), "nil table has nothing")
 }
 
-func TestNewTable_ValidatesRates(t *testing.T) {
-	good := map[string]map[string]EntryJSON{
-		"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01, CachedInputPer1K: 0.00125}},
-	}
-	tbl, err := NewTable(good)
-	require.NoError(t, err)
-	cost, ok := tbl.Cost("openai", "gpt-4o", 1000, 1000, 0, 0)
-	require.True(t, ok, "entry survives the wire conversion")
-	assert.InDelta(t, 0.0125, cost, 1e-9)
-
-	_, ok = tbl.Cost("openai", "unknown-model", 1, 1, 0, 0)
-	assert.False(t, ok, "unknown model misses")
-
-	for name, bad := range map[string]EntryJSON{
-		"negative input":  {InputPer1K: -1, OutputPer1K: 0.01},
-		"NaN output":      {InputPer1K: 0.01, OutputPer1K: math.NaN()},
-		"Inf cache read":  {InputPer1K: 0.01, OutputPer1K: 0.01, CacheReadPer1K: math.Inf(1)},
-		"negative cached": {InputPer1K: 0.01, OutputPer1K: 0.01, CachedInputPer1K: -0.001},
-	} {
-		_, err := NewTable(map[string]map[string]EntryJSON{"openai": {"m": bad}})
-		assert.Error(t, err, "case %q must be rejected so a corrupt config fails the chain build instead of mispricing", name)
-	}
+func TestLoaderGet_NilSafe(t *testing.T) {
+	var l *Loader
+	assert.Nil(t, l.Get(), "nil loader returns nil table")
 }
 
-func TestNewTable_NilAndEmpty(t *testing.T) {
-	tbl, err := NewTable(nil)
-	require.NoError(t, err, "nil map builds an empty (never-matching) table")
-	_, ok := tbl.Cost("openai", "gpt-4o", 1, 1, 0, 0)
-	assert.False(t, ok, "empty table prices nothing")
+// TestNewLoader_RejectsOversizedFile_FixesM4 proves the loader bounds reads
+// at maxPricingBytes so a hostile file cannot exhaust process memory.
+func TestNewLoader_RejectsOversizedFile_FixesM4(t *testing.T) {
+	base := t.TempDir()
+	target := filepath.Join(base, "pricing.yaml")
 
-	entries, err := NewEntries(nil)
-	require.NoError(t, err)
-	assert.Empty(t, entries, "nil in, empty (never-matching) map out for the per-record map")
+	// Build a YAML payload larger than the cap. We pad with valid YAML
+	// comments so a partial read would still fail the size check rather
+	// than the parser.
+	header := "openai:\n"
+	bigComment := make([]byte, maxPricingBytes+1024)
+	for i := range bigComment {
+		bigComment[i] = ' '
+	}
+	bigComment[0] = '#'
+	bigComment[len(bigComment)-1] = '\n'
+	payload := append([]byte(header), bigComment...)
+	require.NoError(t, os.WriteFile(target, payload, 0o600))
+
+	_, err := NewLoader(base, "pricing.yaml", "llm_observability", nil)
+	require.Error(t, err, "oversized pricing file must be rejected")
+	assert.Contains(t, err.Error(), "exceeds", "rejection must reference the byte cap")
 }

proxy/internal/llm/pricing/pricing_unix.go

@@ -0,0 +1,68 @@
+//go:build unix
+
+package pricing
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"syscall"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// maxPricingBytes caps the size of the pricing YAML on read so a hostile or
+// runaway file cannot exhaust process memory during reload. 1 MiB is several
+// orders of magnitude larger than any reasonable pricing table.
+const maxPricingBytes int64 = 1 << 20
+
+// loadPricing opens the file with O_NOFOLLOW, fstats the open descriptor,
+// and parses from that same descriptor. Never re-opens by path so a
+// mid-read rename or symlink swap cannot substitute content. Bytes are
+// capped at maxPricingBytes so the loader cannot be coerced into reading an
+// unbounded file.
+func loadPricing(path string) (*Table, time.Time, error) {
+	f, err := os.OpenFile(path, os.O_RDONLY|syscall.O_NOFOLLOW, 0)
+	if err != nil {
+		return nil, time.Time{}, fmt.Errorf("open %s: %w", path, err)
+	}
+	defer func() {
+		if cerr := f.Close(); cerr != nil {
+			log.Debugf("close pricing file %s: %v", path, cerr)
+		}
+	}()
+
+	info, err := f.Stat()
+	if err != nil {
+		return nil, time.Time{}, fmt.Errorf("fstat %s: %w", path, err)
+	}
+	if !info.Mode().IsRegular() {
+		return nil, time.Time{}, fmt.Errorf("pricing file %s is not a regular file", path)
+	}
+
+	data, err := io.ReadAll(io.LimitReader(f, maxPricingBytes+1))
+	if err != nil {
+		return nil, time.Time{}, fmt.Errorf("read %s: %w", path, err)
+	}
+	if int64(len(data)) > maxPricingBytes {
+		return nil, time.Time{}, fmt.Errorf("pricing file %s exceeds %d bytes", path, maxPricingBytes)
+	}
+
+	table, err := parsePricingBytes(data)
+	if err != nil {
+		return nil, time.Time{}, err
+	}
+	return table, info.ModTime(), nil
+}
+
+// statMtime returns the mtime of the file at path. It uses lstat semantics
+// via os.Lstat so a symlink swap is detected even though O_NOFOLLOW will
+// later reject the open.
+func statMtime(path string) (time.Time, error) {
+	info, err := os.Lstat(path)
+	if err != nil {
+		return time.Time{}, fmt.Errorf("lstat %s: %w", path, err)
+	}
+	return info.ModTime(), nil
+}

proxy/internal/middleware/builtin/builtin.go

@@ -36,13 +36,15 @@ var defaultRegistry = middleware.NewRegistry()
 
 // FactoryContext is the per-process bag that concrete factories may
 // consult during construction. It carries the proxy-lifetime context,
-// the OTel meter, and the proxy logger.
+// the data directory used for static config files (pricing tables,
+// allowlists), the OTel meter, and the proxy logger.
 //
 // Configure must be called once at boot before any chain build calls
 // Resolve. Calling it twice overwrites the prior value; tests may rely
 // on this to reset state.
 type FactoryContext struct {
 	Context    context.Context
+	DataDir    string
 	Meter      metric.Meter
 	Logger     *log.Logger
 	MgmtClient MgmtClient
@@ -56,11 +58,12 @@ var (
 // Configure stores the per-process FactoryContext. Concrete factories
 // reach for it via Context(). mgmt may be nil on tests / standalone
 // builds with no management server; consumers must guard.
-func Configure(ctx context.Context, meter metric.Meter, logger *log.Logger, mgmt MgmtClient) {
+func Configure(ctx context.Context, dataDir string, meter metric.Meter, logger *log.Logger, mgmt MgmtClient) {
 	ctxMu.Lock()
 	defer ctxMu.Unlock()
 	ctxStore = FactoryContext{
 		Context:    ctx,
+		DataDir:    dataDir,
 		Meter:      meter,
 		Logger:     logger,
 		MgmtClient: mgmt,

proxy/internal/middleware/builtin/cost_calculation_matrix_test.go

@@ -12,7 +12,6 @@ import (
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 
-	mgmtpricing "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/pricing"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/cost_meter"
@@ -20,20 +19,17 @@ import (
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_response_parser"
 )
 
-// Drives the real pipeline (llm_request_parser → llm_response_parser → cost_meter) on the REAL default pricing
-// table management ships (mgmtpricing.DefaultTable, catalog-derived) and asserts exact USD amounts hardcoded from
-// the vendors' published prices, including the cache split. This is the cross-stack pricing contract test: the
-// management-side Entry JSON must decode into the proxy-side table and produce these exact costs.
+// Drives the real pipeline (llm_request_parser → llm_response_parser → cost_meter) on the embedded default pricing
+// table and asserts exact USD amounts hardcoded from the vendors' published prices, including the cache split.
 func TestCostCalculation_ProviderMatrix(t *testing.T) {
-	builtin.Configure(context.Background(), nil, nil, nil)
+	// Empty data dir → embedded defaults, like a proxy with no pricing override.
+	builtin.Configure(context.Background(), t.TempDir(), nil, nil, nil)
 
 	reqMW, err := llm_request_parser.Factory{}.New(nil)
 	require.NoError(t, err, "build llm_request_parser")
 	respMW, err := llm_response_parser.Factory{}.New(nil)
 	require.NoError(t, err, "build llm_response_parser")
-	costCfgJSON, err := json.Marshal(map[string]any{"pricing": map[string]any{"defaults": mgmtpricing.DefaultTable()}})
-	require.NoError(t, err, "marshal management default table into cost_meter config")
-	costMW, err := cost_meter.Factory{}.New(costCfgJSON)
+	costMW, err := cost_meter.Factory{}.New(nil)
 	require.NoError(t, err, "build cost_meter")
 	t.Cleanup(func() { _ = costMW.Close() })
 

proxy/internal/middleware/builtin/cost_meter/factory.go

@@ -2,6 +2,7 @@ package cost_meter
 
 import (
 	"bytes"
+	"context"
 	"encoding/json"
 	"fmt"
 
@@ -10,27 +11,16 @@ import (
 	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
 )
 
-// Config is the on-wire configuration for the middleware, synthesized by
-// management (buildCostMeterConfigJSON). The proxy has no embedded price
-// list: this payload is the only pricing source, and updates arrive as
-// ordinary mapping pushes that rebuild the chain (and with it this
-// middleware instance) — no per-request fetches, no reload loops.
-type Config struct {
-	Pricing *PricingConfig `json:"pricing"`
-}
+// defaultPricingFilename is the basename probed inside the proxy data
+// directory when no override is configured.
+const defaultPricingFilename = "pricing.yaml"
 
-// PricingConfig carries the full pricing table:
-//   - Defaults: parser surface ("openai"/"anthropic"/"bedrock") ->
-//     normalized model id -> rates, matched against llm.provider +
-//     llm.model.
-//   - Providers: provider record id -> normalized model id -> rates,
-//     matched against the llm.resolved_provider_id metadata llm_router
-//     stamps. Entries arrive fully materialized (management folds default
-//     cache rates in at synth time), so lookup order is simply
-//     per-record first, defaults second.
-type PricingConfig struct {
-	Defaults  map[string]map[string]pricing.EntryJSON `json:"defaults"`
-	Providers map[string]map[string]pricing.EntryJSON `json:"providers"`
+// Config is the on-wire configuration for the middleware.
+type Config struct {
+	// PricingPath optionally overrides the basename of the pricing
+	// file probed inside the proxy data directory. When empty the
+	// loader falls back to "pricing.yaml".
+	PricingPath string `json:"pricing_path"`
 }
 
 // Factory builds cost_meter instances from raw config bytes.
@@ -39,45 +29,45 @@ type Factory struct{}
 // ID returns the registry identifier.
 func (Factory) ID() string { return ID }
 
-// New constructs a middleware instance. Empty, null, and {} configs are
-// accepted for backward compatibility with a management server that
-// predates config-delivered pricing — the instance then skips every cost
-// computation (unknown_model) and a warning is logged once at build time.
-// Non-empty rawConfig that fails to unmarshal, or a table carrying a
-// non-finite / negative rate, is rejected so misconfigurations surface at
-// chain build time.
+// New constructs a middleware instance. Empty, null, and {} configs
+// are accepted; non-empty rawConfig that fails to unmarshal is
+// rejected so misconfigurations surface at chain build time. The
+// pricing loader is built once per instance and reused across
+// invocations.
 func (Factory) New(rawConfig []byte) (middleware.Middleware, error) {
 	cfg, err := decodeConfig(rawConfig)
 	if err != nil {
 		return nil, err
 	}
 
-	if cfg.Pricing == nil {
-		if logger := builtin.Context().Logger; logger != nil {
-			logger.Warnf("cost_meter: no pricing table in middleware config; management predates config-delivered pricing — every request will record cost.skipped=unknown_model ($0)")
-		}
-		return newMiddleware(mustEmptyTable(), nil), nil
+	fctx := builtin.Context()
+	pricingPath := cfg.PricingPath
+	if pricingPath == "" {
+		pricingPath = defaultPricingFilename
 	}
 
-	defaults, err := pricing.NewTable(cfg.Pricing.Defaults)
+	loader, err := pricing.NewLoader(fctx.DataDir, pricingPath, ID, nil)
 	if err != nil {
-		return nil, fmt.Errorf("cost_meter pricing defaults: %w", err)
+		return nil, fmt.Errorf("init pricing loader: %w", err)
 	}
-	perRecord, err := pricing.NewEntries(cfg.Pricing.Providers)
-	if err != nil {
-		return nil, fmt.Errorf("cost_meter per-provider pricing: %w", err)
-	}
-	return newMiddleware(defaults, perRecord), nil
+
+	cancel := startReloader(fctx.Context, loader)
+
+	return newMiddleware(loader, cancel), nil
 }
 
-// mustEmptyTable returns a valid empty table. NewTable on a nil map cannot
-// fail; the panic guard documents that invariant.
-func mustEmptyTable() *pricing.Table {
-	t, err := pricing.NewTable(nil)
-	if err != nil {
-		panic(fmt.Sprintf("cost_meter: empty pricing table must build: %v", err))
+// startReloader binds the loader's mtime-poll goroutine to a context
+// derived from the proxy-lifetime context and returns its cancel func so
+// the owning middleware can stop the goroutine on teardown. Returns nil
+// when there's nothing to watch (nil context or defaults-only loader), in
+// which case the middleware's Close is a no-op.
+func startReloader(ctx context.Context, loader *pricing.Loader) context.CancelFunc {
+	if ctx == nil || !loader.WatchesFile() {
+		return nil
 	}
-	return t
+	cctx, cancel := context.WithCancel(ctx)
+	go loader.Reload(cctx)
+	return cancel
 }
 
 // decodeConfig accepts empty, null, and {} configs, returning a

proxy/internal/middleware/builtin/cost_meter/middleware.go

@@ -1,9 +1,7 @@
 // Package cost_meter implements the SlotOnResponse middleware that
 // converts token-usage metadata emitted by llm_response_parser into a
-// per-request USD cost estimate. Pricing arrives from management inside
-// the middleware config: a per-provider-record table (the operator's
-// stored prices, matched via llm.resolved_provider_id) consulted first,
-// then the surface-keyed defaults table.
+// per-request USD cost estimate. The middleware uses the shared pricing
+// loader so operator pricing overrides apply to the chain.
 package cost_meter
 
 import (
@@ -19,9 +17,7 @@ import (
 const ID = "cost_meter"
 
 // Version is the implementation version emitted via the spec merge.
-// 1.1.0: pricing is config-delivered (defaults + per-provider-record
-// entries) instead of proxy-embedded.
-const Version = "1.1.0"
+const Version = "1.0.0"
 
 // Skip reasons emitted under KeyCostSkipped. The set is closed; the
 // dashboard surfaces these verbatim.
@@ -46,21 +42,19 @@ var metadataKeys = []string{
 }
 
 // Middleware computes a per-response cost estimate from the token
-// counts emitted upstream by llm_response_parser. Both tables are
-// immutable — a pricing change arrives as a mapping push that rebuilds
-// the chain with a fresh instance.
+// counts emitted upstream by llm_response_parser.
 type Middleware struct {
-	// defaults is the surface-keyed table (llm.provider x llm.model).
-	defaults *pricing.Table
-	// perRecord is keyed by provider record id (llm.resolved_provider_id)
-	// then normalized model id; entries arrive fully materialized from
-	// management. Consulted before defaults. May be nil.
-	perRecord map[string]map[string]pricing.Entry
+	loader *pricing.Loader
+	// cancel stops this instance's pricing-reload goroutine. Non-nil only
+	// when the loader watches an override file; Close calls it so a chain
+	// rebuild doesn't leak a poll goroutine per retired instance.
+	cancel context.CancelFunc
 }
 
-// newMiddleware constructs a Middleware over the given pricing tables.
-func newMiddleware(defaults *pricing.Table, perRecord map[string]map[string]pricing.Entry) *Middleware {
-	return &Middleware{defaults: defaults, perRecord: perRecord}
+// newMiddleware constructs a Middleware bound to the given pricing loader.
+// cancel may be nil (defaults-only loader with no reloader to stop).
+func newMiddleware(loader *pricing.Loader, cancel context.CancelFunc) *Middleware {
+	return &Middleware{loader: loader, cancel: cancel}
 }
 
 // ID returns the registry identifier.
@@ -85,9 +79,16 @@ func (m *Middleware) MetadataKeys() []string {
 // response.
 func (m *Middleware) MutationsSupported() bool { return false }
 
-// Close releases resources owned by the middleware. Stateless — the
-// pricing tables are plain maps owned by this instance.
-func (m *Middleware) Close() error { return nil }
+// Close stops this instance's pricing-reload goroutine, if any. Called by
+// the chain when a rebuild retires the instance, so the mtime-poll loop
+// doesn't outlive the chain it belonged to. Safe to call on a nil receiver
+// and on an instance with no reloader.
+func (m *Middleware) Close() error {
+	if m != nil && m.cancel != nil {
+		m.cancel()
+	}
+	return nil
+}
 
 // Invoke reads provider, model, and token metadata, looks up pricing,
 // and emits either KeyCostUSDTotal or KeyCostSkipped. The decision is
@@ -143,7 +144,8 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 		return out, nil
 	}
 
-	costs, ok := m.lookupCosts(in.Metadata, provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
+	table := m.loader.Get()
+	costs, ok := table.Costs(provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
 	if !ok {
 		out.Metadata = skip(skipUnknownModel)
 		return out, nil
@@ -162,26 +164,6 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 	return out, nil
 }
 
-// lookupCosts resolves the price for this request and computes the cost
-// split. Resolution order:
-//
-//  1. Per-provider-record entry: the operator's stored price for the
-//     provider route that served the request, keyed by the
-//     llm.resolved_provider_id metadata llm_router stamped on the allow
-//     path. Absent metadata (e.g. no router in the chain) skips this tier.
-//  2. Surface defaults: the catalog-derived table keyed by llm.provider.
-//
-// The surface always selects the cache formula — a per-record entry for an
-// Anthropic route still bills its cache buckets additively.
-func (m *Middleware) lookupCosts(md []middleware.KV, surface, model string, inTokens, outTokens, cachedTokens, cacheCreationTokens int64) (pricing.Costs, bool) {
-	if recordID := lookupKV(md, middleware.KeyLLMResolvedProviderID); recordID != "" {
-		if entry, ok := m.perRecord[recordID][model]; ok {
-			return pricing.EntryCosts(entry, surface, inTokens, outTokens, cachedTokens, cacheCreationTokens), true
-		}
-	}
-	return m.defaults.Costs(surface, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
-}
-
 // usd renders a cost as the fixed-precision string every cost.usd_* key
 // carries, so the per-bucket values and the aggregates round identically.
 //

proxy/internal/middleware/builtin/cost_meter/middleware_test.go

@@ -3,50 +3,39 @@ package cost_meter
 import (
 	"context"
 	"encoding/json"
+	"os"
+	"path/filepath"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 
-	"github.com/netbirdio/netbird/proxy/internal/llm/pricing"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
 )
 
-// fixtureConfig mirrors what management's buildCostMeterConfigJSON ships:
-// a surface-keyed defaults table. Rates match the retired YAML fixture so
-// every cost assertion below is byte-identical to the pre-feature values.
-func fixtureConfig(t *testing.T) []byte {
-	t.Helper()
-	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
-		Defaults: map[string]map[string]pricing.EntryJSON{
-			"openai": {
-				"gpt-4o":      {InputPer1K: 0.0025, OutputPer1K: 0.01},
-				"gpt-4o-mini": {InputPer1K: 0.00015, OutputPer1K: 0.0006},
-			},
-			"anthropic": {
-				"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015},
-			},
-		},
-	}})
-	require.NoError(t, err)
-	return raw
-}
+const fixturePricing = `openai:
+  gpt-4o:
+    input_per_1k: 0.0025
+    output_per_1k: 0.01
+  gpt-4o-mini:
+    input_per_1k: 0.00015
+    output_per_1k: 0.0006
+anthropic:
+  claude-sonnet-4-5:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+`
 
-// fixtureConfigWithCache adds the cache-rate fields.
-func fixtureConfigWithCache(t *testing.T) []byte {
+// configureBuiltin points the package-level FactoryContext at a tmp
+// directory containing the test pricing fixture. Returns the path so
+// callers can override files later if needed.
+func configureBuiltin(t *testing.T) string {
 	t.Helper()
-	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
-		Defaults: map[string]map[string]pricing.EntryJSON{
-			"openai": {
-				"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01, CachedInputPer1K: 0.00125},
-			},
-			"anthropic": {
-				"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015, CacheReadPer1K: 0.0003, CacheCreationPer1K: 0.00375},
-			},
-		},
-	}})
-	require.NoError(t, err)
-	return raw
+	dir := t.TempDir()
+	require.NoError(t, os.WriteFile(filepath.Join(dir, "pricing.yaml"), []byte(fixturePricing), 0o600), "write pricing fixture")
+	builtin.Configure(context.Background(), dir, nil, nil, nil)
+	return dir
 }
 
 func metaValue(t *testing.T, kvs []middleware.KV, key string) (string, bool) {
@@ -67,7 +56,8 @@ func buildMiddleware(t *testing.T, raw []byte) middleware.Middleware {
 }
 
 func TestMiddleware_StaticSurface(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
 	assert.Equal(t, ID, mw.ID(), "ID must match the registered constant")
 	assert.Equal(t, Version, mw.Version(), "Version must match the constant")
@@ -89,10 +79,8 @@ func TestMiddleware_StaticSurface(t *testing.T) {
 	assert.Equal(t, expected, keys, "metadata key allowlist must match the spec")
 }
 
-// TestFactory_AcceptsEmptyAndJSONConfig: empty/null/{} configs are what an
-// old management (pre config-delivered pricing) sends — they must build a
-// working (all-skip) instance, never fail the chain.
 func TestFactory_AcceptsEmptyAndJSONConfig(t *testing.T) {
+	configureBuiltin(t)
 	cases := [][]byte{nil, {}, []byte("null"), []byte("{}"), []byte("   ")}
 	for _, raw := range cases {
 		mw, err := Factory{}.New(raw)
@@ -102,57 +90,15 @@ func TestFactory_AcceptsEmptyAndJSONConfig(t *testing.T) {
 }
 
 func TestFactory_RejectsMalformedConfig(t *testing.T) {
+	configureBuiltin(t)
 	mw, err := Factory{}.New([]byte("{not json"))
 	require.Error(t, err, "malformed config must surface at construction")
 	assert.Nil(t, mw, "no instance is returned on error")
 }
 
-// TestFactory_RejectsInvalidRates: a non-finite or negative rate anywhere
-// in the table fails the chain build (defense-in-depth behind management's
-// API validation) rather than silently mispricing.
-func TestFactory_RejectsInvalidRates(t *testing.T) {
-	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
-		Defaults: map[string]map[string]pricing.EntryJSON{
-			"openai": {"gpt-4o": {InputPer1K: -0.0025, OutputPer1K: 0.01}},
-		},
-	}})
-	require.NoError(t, err)
-	mw, err := Factory{}.New(raw)
-	require.Error(t, err, "negative rate must fail the build")
-	assert.Nil(t, mw)
-
-	raw, err = json.Marshal(Config{Pricing: &PricingConfig{
-		Providers: map[string]map[string]pricing.EntryJSON{
-			"prov-1": {"m": {InputPer1K: 0.01, OutputPer1K: 0.01, CacheReadPer1K: -1}},
-		},
-	}})
-	require.NoError(t, err)
-	_, err = Factory{}.New(raw)
-	require.Error(t, err, "per-record tables validate too")
-}
-
-// TestFactory_NilPricingSkipsEverything is the version-skew contract: a
-// new proxy under an old management ({} config) must build, allow, and
-// skip with unknown_model — degraded but never broken.
-func TestFactory_NilPricingSkipsEverything(t *testing.T) {
-	mw := buildMiddleware(t, []byte("{}"))
-	out, err := mw.Invoke(context.Background(), &middleware.Input{
-		Metadata: []middleware.KV{
-			{Key: middleware.KeyLLMProvider, Value: "openai"},
-			{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
-			{Key: middleware.KeyLLMInputTokens, Value: "1000"},
-			{Key: middleware.KeyLLMOutputTokens, Value: "1000"},
-		},
-	})
-	require.NoError(t, err)
-	assert.Equal(t, middleware.DecisionAllow, out.Decision, "cost_meter always allows")
-	value, ok := metaValue(t, out.Metadata, middleware.KeyCostSkipped)
-	require.True(t, ok, "no pricing table means every request skips")
-	assert.Equal(t, skipUnknownModel, value)
-}
-
-func TestFactory_ConfigDefaultsPriceRequests(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
+func TestFactory_DefaultPricingPathLoadsFixture(t *testing.T) {
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -170,78 +116,34 @@ func TestFactory_ConfigDefaultsPriceRequests(t *testing.T) {
 	assert.Equal(t, "0.000750000", value, "0.00015 + 0.0006 per 1k tokens, 9-decimal format")
 }
 
-// TestInvoke_PerRecordEntryBeatsDefaults: when llm_router resolved a
-// provider record whose operator pinned a price for the model, that price
-// wins over the surface default.
-func TestInvoke_PerRecordEntryBeatsDefaults(t *testing.T) {
-	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
-		Defaults: map[string]map[string]pricing.EntryJSON{
-			"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01}},
-		},
-		Providers: map[string]map[string]pricing.EntryJSON{
-			"prov-azure": {"gpt-4o": {InputPer1K: 0.005, OutputPer1K: 0.02}},
-		},
-	}})
-	require.NoError(t, err)
-	mw := buildMiddleware(t, raw)
+func TestFactory_PricingPathOverride(t *testing.T) {
+	dir := t.TempDir()
+	require.NoError(t, os.WriteFile(filepath.Join(dir, "custom.yaml"), []byte(fixturePricing), 0o600), "write custom pricing")
+	builtin.Configure(context.Background(), dir, nil, nil, nil)
 
+	raw, err := json.Marshal(Config{PricingPath: "custom.yaml"})
+	require.NoError(t, err)
+
+	mw := buildMiddleware(t, raw)
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMProvider, Value: "openai"},
 			{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
-			{Key: middleware.KeyLLMResolvedProviderID, Value: "prov-azure"},
-			{Key: middleware.KeyLLMInputTokens, Value: "1000"},
+			{Key: middleware.KeyLLMInputTokens, Value: "2000"},
 			{Key: middleware.KeyLLMOutputTokens, Value: "1000"},
 		},
 	})
 	require.NoError(t, err)
+
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
-	require.True(t, ok)
-	assert.Equal(t, "0.025000000", value, "operator's per-record price (0.005+0.02) wins over the default (0.0025+0.01)")
+	require.True(t, ok, "cost.usd_total must be emitted with custom pricing path")
+	assert.Equal(t, "0.015000000", value, "2*0.0025 + 1*0.01 = 0.015 with 9-decimal format")
 }
 
-// TestInvoke_PerRecordMissFallsBackToDefaults: a resolved record with no
-// entry for this model (or no entries at all) falls through to the
-// surface defaults — gateway providers rely on exactly this.
-func TestInvoke_PerRecordMissFallsBackToDefaults(t *testing.T) {
-	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
-		Defaults: map[string]map[string]pricing.EntryJSON{
-			"openai": {"gpt-4o": {InputPer1K: 0.0025, OutputPer1K: 0.01}},
-		},
-		Providers: map[string]map[string]pricing.EntryJSON{
-			"prov-1": {"some-other-model": {InputPer1K: 1, OutputPer1K: 1}},
-		},
-	}})
-	require.NoError(t, err)
-	mw := buildMiddleware(t, raw)
+func TestInvoke_ComputesCostForKnownModel(t *testing.T) {
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
-	for name, recordID := range map[string]string{
-		"record with other models": "prov-1",
-		"record with no entries":   "prov-gateway",
-	} {
-		t.Run(name, func(t *testing.T) {
-			out, err := mw.Invoke(context.Background(), &middleware.Input{
-				Metadata: []middleware.KV{
-					{Key: middleware.KeyLLMProvider, Value: "openai"},
-					{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
-					{Key: middleware.KeyLLMResolvedProviderID, Value: recordID},
-					{Key: middleware.KeyLLMInputTokens, Value: "1000"},
-					{Key: middleware.KeyLLMOutputTokens, Value: "1000"},
-				},
-			})
-			require.NoError(t, err)
-			value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
-			require.True(t, ok, "per-record miss must fall back to the surface default, not skip")
-			assert.Equal(t, "0.012500000", value, "default rates apply")
-		})
-	}
-}
-
-// TestInvoke_NoResolvedProviderIDUsesDefaults: metadata without a
-// resolved provider id (router denied, or a chain without llm_router)
-// prices from the defaults table directly.
-func TestInvoke_NoResolvedProviderIDUsesDefaults(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMProvider, Value: "anthropic"},
@@ -251,15 +153,17 @@ func TestInvoke_NoResolvedProviderIDUsesDefaults(t *testing.T) {
 		},
 	})
 	require.NoError(t, err)
+
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
-	require.True(t, ok)
+	require.True(t, ok, "cost.usd_total must be emitted")
 	assert.Equal(t, "0.018000000", value, "0.003 + 0.015 = 0.018 with 9-decimal format")
 	_, skipped := metaValue(t, out.Metadata, middleware.KeyCostSkipped)
 	assert.False(t, skipped, "cost.skipped must not be set when cost is computed")
 }
 
 func TestInvoke_MissingProvider(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -275,7 +179,8 @@ func TestInvoke_MissingProvider(t *testing.T) {
 }
 
 func TestInvoke_MissingModel(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -291,7 +196,8 @@ func TestInvoke_MissingModel(t *testing.T) {
 }
 
 func TestInvoke_MissingTokens(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
 	cases := []struct {
 		name string
@@ -334,7 +240,8 @@ func TestInvoke_MissingTokens(t *testing.T) {
 }
 
 func TestInvoke_UnparseableTokens(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
 	cases := []struct {
 		name string
@@ -364,7 +271,8 @@ func TestInvoke_UnparseableTokens(t *testing.T) {
 }
 
 func TestInvoke_ZeroTokens(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -383,7 +291,8 @@ func TestInvoke_ZeroTokens(t *testing.T) {
 }
 
 func TestInvoke_UnknownModel(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -400,7 +309,8 @@ func TestInvoke_UnknownModel(t *testing.T) {
 }
 
 func TestInvoke_NilInput(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfig(t))
+	configureBuiltin(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), nil)
 	require.NoError(t, err)
@@ -409,12 +319,36 @@ func TestInvoke_NilInput(t *testing.T) {
 	assert.Empty(t, out.Metadata, "no metadata must be emitted on nil input")
 }
 
+const fixturePricingWithCache = `openai:
+  gpt-4o:
+    input_per_1k: 0.0025
+    output_per_1k: 0.01
+    cached_input_per_1k: 0.00125
+anthropic:
+  claude-sonnet-4-5:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+    cache_creation_per_1k: 0.00375
+`
+
+// configureBuiltinWithCacheRates points the package-level
+// FactoryContext at a tmp directory containing pricing entries that
+// include the cache rate fields.
+func configureBuiltinWithCacheRates(t *testing.T) {
+	t.Helper()
+	dir := t.TempDir()
+	require.NoError(t, os.WriteFile(filepath.Join(dir, "pricing.yaml"), []byte(fixturePricingWithCache), 0o600), "write cache-aware pricing fixture")
+	builtin.Configure(context.Background(), dir, nil, nil, nil)
+}
+
 // TestInvoke_OpenAICachedSubsetDiscount proves the OpenAI shape end
 // to end through the middleware: cached_input_tokens is treated as a
 // SUBSET of input_tokens and discounted at the configured rate, not
 // added on top.
 func TestInvoke_OpenAICachedSubsetDiscount(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfigWithCache(t))
+	configureBuiltinWithCacheRates(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -456,7 +390,8 @@ func TestInvoke_OpenAICachedSubsetDiscount(t *testing.T) {
 // shape: cache_read and cache_creation are additive to input_tokens
 // and each carries its own rate.
 func TestInvoke_AnthropicCacheBucketsAdditive(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfigWithCache(t))
+	configureBuiltinWithCacheRates(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -494,37 +429,8 @@ func TestInvoke_AnthropicCacheBucketsAdditive(t *testing.T) {
 		"output bucket bills 200 tokens at 0.015/1k")
 }
 
-// TestInvoke_PerRecordEntryUsesSurfaceFormula: a per-record entry for an
-// anthropic-surface request must bill its cache buckets additively — the
-// formula follows llm.provider, not which table the entry came from.
-func TestInvoke_PerRecordEntryUsesSurfaceFormula(t *testing.T) {
-	raw, err := json.Marshal(Config{Pricing: &PricingConfig{
-		Providers: map[string]map[string]pricing.EntryJSON{
-			"prov-ant": {"claude-sonnet-4-5": {InputPer1K: 0.003, OutputPer1K: 0.015, CacheReadPer1K: 0.0003, CacheCreationPer1K: 0.00375}},
-		},
-	}})
-	require.NoError(t, err)
-	mw := buildMiddleware(t, raw)
-
-	out, err := mw.Invoke(context.Background(), &middleware.Input{
-		Metadata: []middleware.KV{
-			{Key: middleware.KeyLLMProvider, Value: "anthropic"},
-			{Key: middleware.KeyLLMModel, Value: "claude-sonnet-4-5"},
-			{Key: middleware.KeyLLMResolvedProviderID, Value: "prov-ant"},
-			{Key: middleware.KeyLLMInputTokens, Value: "256"},
-			{Key: middleware.KeyLLMOutputTokens, Value: "200"},
-			{Key: middleware.KeyLLMCachedInputTokens, Value: "768"},
-			{Key: middleware.KeyLLMCacheCreationTokens, Value: "512"},
-		},
-	})
-	require.NoError(t, err)
-	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
-	require.True(t, ok)
-	assert.Equal(t, "0.005918400", value, "identical math to the defaults-table entry with the same rates")
-}
-
 // assertBucket asserts one per-bucket cost key carries the expected
-// 9-decimal value.
+// 6-decimal value.
 func assertBucket(t *testing.T, md []middleware.KV, key, want, msg string) {
 	t.Helper()
 	got, ok := metaValue(t, md, key)
@@ -533,10 +439,13 @@ func assertBucket(t *testing.T, md []middleware.KV, key, want, msg string) {
 }
 
 // TestInvoke_CachedTokensAbsentFallsBackToBaseFormula covers the
-// no-cache-metadata path: with no cached keys emitted, the meter must
-// produce exactly the input+output cost.
+// "operator hasn't opted in" path: with no cached metadata keys
+// emitted, the meter must produce exactly the same cost as before
+// the feature landed. Critical so operators with the new binary but
+// no YAML changes see no behavioural drift on OpenAI requests.
 func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfigWithCache(t))
+	configureBuiltinWithCacheRates(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -551,7 +460,7 @@ func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok)
 	// 1000 input * 0.0025 + 500 output * 0.01 = 0.0025 + 0.005 = 0.0075
-	assert.Equal(t, "0.007500000", value, "no cached metadata = plain input+output cost")
+	assert.Equal(t, "0.007500000", value, "no cached metadata = same cost as before the feature landed")
 }
 
 // TestInvoke_UnparseableCachedTokensSkippedSilently proves the
@@ -560,7 +469,8 @@ func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
 // regular formula. Cache buckets are a refinement, never a reason to
 // abort cost computation.
 func TestInvoke_UnparseableCachedTokensSkippedSilently(t *testing.T) {
-	mw := buildMiddleware(t, fixtureConfigWithCache(t))
+	configureBuiltinWithCacheRates(t)
+	mw := buildMiddleware(t, nil)
 
 	out, err := mw.Invoke(context.Background(), &middleware.Input{
 		Metadata: []middleware.KV{
@@ -577,10 +487,22 @@ func TestInvoke_UnparseableCachedTokensSkippedSilently(t *testing.T) {
 	assert.Equal(t, "0.007500000", value, "same as the no-cached-metadata path")
 }
 
-// TestMiddleware_CloseNilSafe confirms Close is a no-op (no panic) even
-// for a nil receiver.
+// TestMiddleware_CloseCancelsReloader proves Close stops the per-instance
+// pricing-reload goroutine: a chain rebuild retires the old instance and
+// calls Close, which must invoke the cancel func startReloader handed it so
+// the mtime-poll loop doesn't outlive the chain.
+func TestMiddleware_CloseCancelsReloader(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	m := newMiddleware(nil, cancel)
+
+	require.NoError(t, m.Close(), "Close must not error")
+	require.Error(t, ctx.Err(), "Close must cancel the reloader context so the poll goroutine exits")
+}
+
+// TestMiddleware_CloseNilSafe confirms Close is a no-op (no panic) for an
+// instance with no reloader and for a nil receiver.
 func TestMiddleware_CloseNilSafe(t *testing.T) {
-	require.NoError(t, newMiddleware(nil, nil).Close(), "Close must be a no-op")
+	require.NoError(t, newMiddleware(nil, nil).Close(), "no-reloader Close must be a no-op")
 	var m *Middleware
 	require.NoError(t, m.Close(), "nil-receiver Close must be safe")
 }

proxy/internal/middleware/builtin/llm_request_parser/bedrock_test.go

@@ -6,6 +6,23 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+func TestNormalizeBedrockModel(t *testing.T) {
+	cases := map[string]string{
+		"eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
+		"us.anthropic.claude-opus-4-8-20250101-v1:0":   "anthropic.claude-opus-4-8",
+		"apac.anthropic.claude-haiku-4-5-v1:0":         "anthropic.claude-haiku-4-5",
+		"anthropic.claude-sonnet-4-5-20250929-v1:0":    "anthropic.claude-sonnet-4-5",
+		"meta.llama3-3-70b-instruct-v1:0":              "meta.llama3-3-70b-instruct",
+		"amazon.nova-pro-v1:0":                         "amazon.nova-pro",
+		"amazon.nova-2-lite-v1:0":                      "amazon.nova-2-lite",
+		// Inference-profile ARN — model id lives in the last path segment.
+		"arn:aws:bedrock:eu-central-1:123456789012:inference-profile/eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
+	}
+	for in, want := range cases {
+		require.Equal(t, want, normalizeBedrockModel(in), "normalize %q", in)
+	}
+}
+
 func TestParseBedrockPath(t *testing.T) {
 	tests := []struct {
 		path   string

proxy/internal/middleware/builtin/llm_request_parser/middleware.go

@@ -8,6 +8,7 @@ package llm_request_parser
 import (
 	"context"
 	"net/url"
+	"regexp"
 	"strconv"
 	"strings"
 	"unicode/utf8"
@@ -252,7 +253,9 @@ func parseVertexPath(reqPath string) (vertexRequest, bool) {
 	if c := strings.LastIndex(rest, ":"); c >= 0 {
 		model, action = rest[:c], rest[c+1:]
 	}
-	model = llm.NormalizeVertexModel(model)
+	if at := strings.Index(model, "@"); at >= 0 {
+		model = model[:at]
+	}
 	if model == "" {
 		return vertexRequest{}, false
 	}
@@ -340,6 +343,14 @@ func trimBedrockNamespace(reqPath string) string {
 	return reqPath
 }
 
+// bedrockRegionPrefixes are the cross-region inference-profile prefixes that
+// front a Bedrock model id (e.g. "eu.anthropic.claude-...").
+var bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
+
+// bedrockVersionSuffix matches the trailing "-vN[:N]" or "-YYYYMMDD-vN[:N]"
+// version/throughput suffix of a Bedrock model id.
+var bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
+
 // parseBedrockPath extracts the model and streaming/converse flags from an AWS
 // Bedrock runtime model endpoint:
 //
@@ -364,7 +375,7 @@ func parseBedrockPath(reqPath string) (bedrockRequest, bool) {
 	if decoded, err := url.PathUnescape(rawModel); err == nil {
 		rawModel = decoded
 	}
-	model := llm.NormalizeBedrockModel(rawModel)
+	model := normalizeBedrockModel(rawModel)
 	if model == "" {
 		return bedrockRequest{}, false
 	}
@@ -378,6 +389,30 @@ func parseBedrockPath(reqPath string) (bedrockRequest, bool) {
 	}
 }
 
+// normalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
+// prefix, and the version/throughput suffix from a Bedrock model id so it
+// matches the catalog/pricing key, e.g.
+// "eu.anthropic.claude-sonnet-4-5-20250929-v1:0" -> "anthropic.claude-sonnet-4-5"
+// and "arn:aws:bedrock:eu-central-1:123:inference-profile/eu.anthropic.claude-sonnet-4-5-20250929-v1:0"
+// -> "anthropic.claude-sonnet-4-5".
+func normalizeBedrockModel(modelID string) string {
+	m := modelID
+	// A full ARN (inference-profile / provisioned-throughput / foundation-model)
+	// carries the model id in its last path segment.
+	if strings.HasPrefix(m, "arn:") {
+		if i := strings.LastIndex(m, "/"); i >= 0 {
+			m = m[i+1:]
+		}
+	}
+	for _, p := range bedrockRegionPrefixes {
+		if strings.HasPrefix(m, p) {
+			m = m[len(p):]
+			break
+		}
+	}
+	return bedrockVersionSuffix.ReplaceAllString(m, "")
+}
+
 // invokeBedrock emits the model/provider/session/prompt for an AWS Bedrock
 // request. Bedrock is metered under the dedicated "bedrock" parser, which reads
 // both the InvokeModel and Converse response shapes.

proxy/internal/proxy/agent_network_chain_realstack_test.go

@@ -18,10 +18,10 @@ import (
 	"google.golang.org/grpc/credentials/insecure"
 	"google.golang.org/grpc/test/bufconn"
 
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
-	agentNetworkTypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
 	mgmtgrpc "github.com/netbirdio/netbird/management/internals/shared/grpc"
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
+	agentNetworkTypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 	"github.com/netbirdio/netbird/management/server/store"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
 	"github.com/netbirdio/netbird/proxy/internal/middleware/bodytap"
@@ -134,18 +134,14 @@ func TestReverseProxy_AgentNetworkRequest_FullChain(t *testing.T) {
 		RedactPii:              true,
 	}))
 	require.NoError(t, st.SaveAgentNetworkProvider(ctx, &agentNetworkTypes.Provider{
-		ID:          providerID,
-		AccountID:   testAccountID,
-		ProviderID:  "openai_api",
-		Name:        "openai-fullchain-test",
-		UpstreamURL: upstream.URL, // router rewrites to this
-		APIKey:      "sk-test",
-		Enabled:     true,
-		// Operator-pinned prices deliberately differ from the catalog's
-		// gpt-5.4 rates (0.0025/0.015) so the cost assertion below proves
-		// the per-provider-record price — not the default table — billed
-		// this request: stored price → synth → wire → cost_meter.
-		Models:            []agentNetworkTypes.ProviderModel{{ID: "gpt-5.4", InputPer1k: 0.004, OutputPer1k: 0.02}},
+		ID:                providerID,
+		AccountID:         testAccountID,
+		ProviderID:        "openai_api",
+		Name:              "openai-fullchain-test",
+		UpstreamURL:       upstream.URL, // router rewrites to this
+		APIKey:            "sk-test",
+		Enabled:           true,
+		Models:            []agentNetworkTypes.ProviderModel{{ID: "gpt-5.4"}},
 		SessionPrivateKey: "priv",
 		SessionPublicKey:  "pub",
 	}))
@@ -180,7 +176,7 @@ func TestReverseProxy_AgentNetworkRequest_FullChain(t *testing.T) {
 
 	// ---- 5. Wire the middleware framework — same registry the proxy uses
 	// in production, configured with our bufconn-backed management client.
-	mwbuiltin.Configure(ctx, nil, testLogger, mgmtClient)
+	mwbuiltin.Configure(ctx, t.TempDir(), nil, testLogger, mgmtClient)
 	registry := mwbuiltin.DefaultRegistry()
 	mwMetrics, err := middleware.NewMetrics(nil)
 	require.NoError(t, err)
@@ -287,23 +283,13 @@ func TestReverseProxy_AgentNetworkRequest_FullChain(t *testing.T) {
 			if r.DimensionKind == agentNetworkTypes.DimensionGroup &&
 				r.DimensionID == adminGroupID &&
 				r.WindowSeconds == 60 &&
-				r.TokensInput+r.TokensOutput > 0 &&
-				r.CostUSD > 0 {
+				r.TokensInput+r.TokensOutput > 0 {
 				return true
 			}
 		}
 		return false
 	}, 5*time.Second, 50*time.Millisecond,
-		"Admins group consumption row must increment via the response leg WITH a non-zero cost — a zero cost means the operator's stored price never reached cost_meter (synth → wire → per-record lookup broken)")
-
-	// 8a-cost. Exact cost from the OPERATOR's stored price, not the catalog
-	// default: 12 prompt tokens × 0.004/1k + 40 completion tokens × 0.02/1k
-	// = 0.000048 + 0.0008 = 0.000848. With catalog rates it would be 0.00063
-	// — this assertion distinguishes the two, closing the loop on the whole
-	// dynamic-pricing feature (dashboard save → synth → gRPC wire →
-	// llm.resolved_provider_id lookup → billing).
-	assert.Equal(t, "0.000848000", cd.GetMetadata()["cost.usd_total"],
-		"cost must be computed from the provider record's operator-pinned price")
+		"Admins group consumption row must increment via the response leg — if this fails the proxy's respInput dropped UserGroups again or the parser/recorder wiring is broken")
 
 	// 8b. Both the captured prompt and the captured completion are
 	// redacted — proves the synth threads redact_pii=true into BOTH parser

proxy/server.go

@@ -246,6 +246,10 @@ type Server struct {
 	// in processMappings before the receive loop reconnects to resync.
 	// Zero uses defaultMappingBatchWatchdog.
 	MappingBatchWatchdog time.Duration
+	// MiddlewareDataDir is the base directory the middleware system uses to
+	// resolve file-backed configuration (e.g. the cost_meter pricing table).
+	// Empty means any middleware that requires a file fails at configure time.
+	MiddlewareDataDir string
 	// MiddlewareCaptureBudgetBytes overrides the proxy-wide in-flight capture
 	// budget passed to middleware.NewManager. Zero or negative values fall
 	// back to defaultMiddlewareCaptureBudgetBytes (256 MiB).
@@ -2089,7 +2093,7 @@ func (s *Server) initMiddlewareManager(ctx context.Context) error {
 		return fmt.Errorf("middleware manager requires metrics bundle")
 	}
 	otelMeter := s.meter.Meter()
-	mwbuiltin.Configure(ctx, otelMeter, s.Logger, s.mgmtClient)
+	mwbuiltin.Configure(ctx, s.MiddlewareDataDir, otelMeter, s.Logger, s.mgmtClient)
 
 	mwMetrics, err := middleware.NewMetrics(otelMeter)
 	if err != nil {

shared/llm/model.go

@@ -1,58 +0,0 @@
-// Package llm holds LLM model-identifier helpers shared by the proxy and
-// the management server. The proxy normalizes model ids parsed off inbound
-// requests; management normalizes the operator's registered model ids at
-// synthesis time so both sides of the pricing / routing contract compare
-// equal.
-package llm
-
-import (
-	"regexp"
-	"strings"
-)
-
-// bedrockRegionPrefixes are the cross-region inference-profile prefixes that
-// front a Bedrock model id (e.g. "eu.anthropic.claude-...").
-var bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
-
-// bedrockVersionSuffix matches the trailing "-vN[:N]" or "-YYYYMMDD-vN[:N]"
-// version/throughput suffix of a Bedrock model id.
-var bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
-
-// NormalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
-// prefix, and the version/throughput suffix from a Bedrock model id so it
-// matches the catalog/pricing key, e.g.
-// "eu.anthropic.claude-sonnet-4-5-20250929-v1:0" -> "anthropic.claude-sonnet-4-5"
-// and the inference-profile ARN's last segment likewise. It is the single
-// source of truth shared by the proxy's request parser (which normalizes the
-// request model from the URL path), the proxy's router (which normalizes the
-// operator's registered Bedrock model ids so both sides compare equal), and
-// the management synthesizer (which keys per-provider pricing entries by the
-// normalized id the parser will emit at billing time).
-func NormalizeBedrockModel(modelID string) string {
-	m := modelID
-	// A full ARN (inference-profile / provisioned-throughput / foundation-model)
-	// carries the model id in its last path segment.
-	if strings.HasPrefix(m, "arn:") {
-		if i := strings.LastIndex(m, "/"); i >= 0 {
-			m = m[i+1:]
-		}
-	}
-	for _, p := range bedrockRegionPrefixes {
-		if strings.HasPrefix(m, p) {
-			m = m[len(p):]
-			break
-		}
-	}
-	return bedrockVersionSuffix.ReplaceAllString(m, "")
-}
-
-// NormalizeVertexModel strips the "@version" suffix from a Vertex AI model id
-// (e.g. "claude-sonnet-4-5@20250929" -> "claude-sonnet-4-5") so it matches
-// the catalog/pricing key. Vertex publisher models are priced under their
-// vendor surface with the bare, unversioned id.
-func NormalizeVertexModel(modelID string) string {
-	if at := strings.Index(modelID, "@"); at >= 0 {
-		return modelID[:at]
-	}
-	return modelID
-}

shared/management/http/api/openapi.yml

@@ -5271,21 +5271,6 @@ components:
           format: double
           description: Cost per 1k output tokens, in USD.
           example: 0.0006
-        cached_input_per_1k:
-          type: number
-          format: double
-          description: OpenAI-shape cache rate — cost per 1k cached prompt tokens (a subset of input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means no discount (cached tokens bill at input_per_1k).
-          example: 0.000075
-        cache_read_per_1k:
-          type: number
-          format: double
-          description: Anthropic-shape cache rate — cost per 1k cache-read tokens (additive to input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means cache reads bill at input_per_1k.
-          example: 0.0003
-        cache_creation_per_1k:
-          type: number
-          format: double
-          description: Anthropic-shape cache rate — cost per 1k cache-creation tokens (additive to input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means cache writes bill at input_per_1k.
-          example: 0.00375
       required:
         - id
         - input_per_1k
@@ -5311,21 +5296,6 @@ components:
           format: double
           description: Output token price per 1k tokens, in USD.
           example: 0.015
-        cached_input_per_1k:
-          type: number
-          format: double
-          description: OpenAI-shape cache rate — default cost per 1k cached prompt tokens (a subset of input tokens), in USD. Absent when the model has no cached-input discount.
-          example: 0.000075
-        cache_read_per_1k:
-          type: number
-          format: double
-          description: Anthropic-shape cache rate — default cost per 1k cache-read tokens (additive to input tokens), in USD. Absent when the model has no cache-read rate.
-          example: 0.0003
-        cache_creation_per_1k:
-          type: number
-          format: double
-          description: Anthropic-shape cache rate — default cost per 1k cache-creation tokens (additive to input tokens), in USD. Absent when the model has no cache-creation rate.
-          example: 0.00375
         context_window:
           type: integer
           description: Maximum context window in tokens.
@@ -5384,13 +5354,6 @@ components:
             $ref: '#/components/schemas/AgentNetworkCatalogExtraHeader'
         identity_injection:
           $ref: '#/components/schemas/AgentNetworkCatalogIdentityInjection'
-        pricing_surfaces:
-          type: array
-          description: |
-            Cost-meter pricing surfaces this provider's traffic is metered under ("openai", "anthropic", "bedrock"). Tells the dashboard which cache-rate fields apply to this provider's models: "openai" → cached_input_per_1k (cached prompt tokens are a subset of input); "anthropic"/"bedrock" → cache_read_per_1k + cache_creation_per_1k (additive buckets). Absent/empty for gateway and custom entries, whose upstream shape NetBird cannot know ahead of time — surface all cache fields for those.
-          items:
-            type: string
-          example: ["openai"]
         models:
           type: array
           description: Catalog models available for this provider.

shared/management/http/api/types.gen.go

@@ -2017,15 +2017,6 @@ type AgentNetworkCatalogJSONMetadataInjection struct {
 
 // AgentNetworkCatalogModel defines model for AgentNetworkCatalogModel.
 type AgentNetworkCatalogModel struct {
-	// CacheCreationPer1k Anthropic-shape cache rate — default cost per 1k cache-creation tokens (additive to input tokens), in USD. Absent when the model has no cache-creation rate.
-	CacheCreationPer1k *float64 `json:"cache_creation_per_1k,omitempty"`
-
-	// CacheReadPer1k Anthropic-shape cache rate — default cost per 1k cache-read tokens (additive to input tokens), in USD. Absent when the model has no cache-read rate.
-	CacheReadPer1k *float64 `json:"cache_read_per_1k,omitempty"`
-
-	// CachedInputPer1k OpenAI-shape cache rate — default cost per 1k cached prompt tokens (a subset of input tokens), in USD. Absent when the model has no cached-input discount.
-	CachedInputPer1k *float64 `json:"cached_input_per_1k,omitempty"`
-
 	// ContextWindow Maximum context window in tokens.
 	ContextWindow int `json:"context_window"`
 
@@ -2079,9 +2070,6 @@ type AgentNetworkCatalogProvider struct {
 
 	// Name Display name for the provider.
 	Name string `json:"name"`
-
-	// PricingSurfaces Cost-meter pricing surfaces this provider's traffic is metered under ("openai", "anthropic", "bedrock"). Tells the dashboard which cache-rate fields apply to this provider's models: "openai" → cached_input_per_1k (cached prompt tokens are a subset of input); "anthropic"/"bedrock" → cache_read_per_1k + cache_creation_per_1k (additive buckets). Absent/empty for gateway and custom entries, whose upstream shape NetBird cannot know ahead of time — surface all cache fields for those.
-	PricingSurfaces *[]string `json:"pricing_surfaces,omitempty"`
 }
 
 // AgentNetworkCatalogProviderKind Presentation grouping for the provider Select on the dashboard.
@@ -2305,15 +2293,6 @@ type AgentNetworkProvider struct {
 
 // AgentNetworkProviderModel A model exposed by the provider, with the operator's per-1k input/output prices in USD.
 type AgentNetworkProviderModel struct {
-	// CacheCreationPer1k Anthropic-shape cache rate — cost per 1k cache-creation tokens (additive to input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means cache writes bill at input_per_1k.
-	CacheCreationPer1k *float64 `json:"cache_creation_per_1k,omitempty"`
-
-	// CacheReadPer1k Anthropic-shape cache rate — cost per 1k cache-read tokens (additive to input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means cache reads bill at input_per_1k.
-	CacheReadPer1k *float64 `json:"cache_read_per_1k,omitempty"`
-
-	// CachedInputPer1k OpenAI-shape cache rate — cost per 1k cached prompt tokens (a subset of input tokens), in USD. Omitted means inherit NetBird's default rate for this model when one exists; 0 means no discount (cached tokens bill at input_per_1k).
-	CachedInputPer1k *float64 `json:"cached_input_per_1k,omitempty"`
-
 	// Id Model identifier (e.g. "gpt-4o-mini").
 	Id string `json:"id"`