Merge main into fix/1025

wasn't caught when we merged the API changes

.github/actions/typecheck/action.yml

@@ -1,12 +0,0 @@
-name: Type Check
-
-description: "Run type checker"
-
-runs:
-  using: "composite"
-  steps:
-    - name: Run type checker
-      run: |
-        nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just sync
-        nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just check
-      shell: bash

.github/workflows/pipeline.yml

@@ -26,73 +26,14 @@ jobs:
           name: exo
           authToken: "${{ secrets.CACHIX_AUTH_TOKEN }}"
 
-      - name: Configure git user
-        run: |
-          git config --local user.email "github-actions@users.noreply.github.com"
-          git config --local user.name  "github-actions bot"
-        shell: bash
+      - name: Load nix develop environment
+        run: nix run github:nicknovitski/nix-develop/v1
 
-      - name: Pull LFS files
-        run: |
-          echo "Pulling Git LFS files..."
-          git lfs pull
-        shell: bash
+      - name: Sync dependencies
+        run: uv sync --all-packages
 
-      - name: Setup Nix Environment
-        run: |
-          echo "Checking for nix installation..."
-          
-          # Check if nix binary exists directly
-          if [ -f /nix/var/nix/profiles/default/bin/nix ]; then
-            echo "Found nix binary at /nix/var/nix/profiles/default/bin/nix"
-            export PATH="/nix/var/nix/profiles/default/bin:$PATH"
-            echo "PATH=$PATH" >> $GITHUB_ENV
-            nix --version
-          elif [ -f /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh ]; then
-            echo "Found nix profile script, sourcing..."
-            source /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh
-            nix --version
-          elif command -v nix >/dev/null 2>&1; then
-            echo "Nix already in PATH"
-            nix --version
-          else
-            echo "Nix not found. Debugging info:"
-            echo "Contents of /nix/var/nix/profiles/default/:"
-            ls -la /nix/var/nix/profiles/default/ 2>/dev/null || echo "Directory not found"
-            echo "Contents of /nix/var/nix/profiles/default/bin/:"
-            ls -la /nix/var/nix/profiles/default/bin/ 2>/dev/null || echo "Directory not found"
-            exit 1
-          fi
-        shell: bash
-
-      - name: Configure basedpyright include for local MLX
-        run: |
-          RUNNER_LABELS='${{ toJSON(runner.labels) }}'
-          if echo "$RUNNER_LABELS" | grep -q "local_mlx"; then
-            if [ -d "/Users/Shared/mlx" ]; then
-              echo "Updating [tool.basedpyright].include to use /Users/Shared/mlx"
-              awk '
-                BEGIN { in=0 }
-                /^\[tool\.basedpyright\]/ { in=1; print; next }
-                in && /^\[/ { in=0 }  # next section
-                in && /^[ \t]*include[ \t]*=/ {
-                  print "include = [\"/Users/Shared/mlx\"]"
-                  next
-                }
-                { print }
-              ' pyproject.toml > pyproject.toml.tmp && mv pyproject.toml.tmp pyproject.toml
-
-              echo "New [tool.basedpyright] section:"
-              sed -n '/^\[tool\.basedpyright\]/,/^\[/p' pyproject.toml | sed '$d' || true
-            else
-              echo "local_mlx tag present but /Users/Shared/mlx not found; leaving pyproject unchanged."
-            fi
-          else
-            echo "Runner does not have 'local_mlx' tag; leaving pyproject unchanged."
-          fi
-        shell: bash
-
-      - uses: ./.github/actions/typecheck
+      - name: Run type checker
+        run: uv run basedpyright --project pyproject.toml
 
   nix:
     name: Build and check (${{ matrix.system }})
@@ -123,6 +64,63 @@ jobs:
           name: exo
           authToken: "${{ secrets.CACHIX_AUTH_TOKEN }}"
 
+      - name: Build Metal packages (macOS only)
+        if: runner.os == 'macOS'
+        run: |
+          # Try to build metal-toolchain first (may succeed via cachix cache hit)
+          if nix build .#metal-toolchain 2>/dev/null; then
+            echo "metal-toolchain built successfully (likely cache hit)"
+          else
+            echo "metal-toolchain build failed, extracting from Xcode..."
+
+            NAR_HASH="sha256-ayR5mXN4sZAddwKEG2OszGRF93k9ZFc7H0yi2xbylQw="
+            NAR_NAME="metal-toolchain-17C48.nar"
+
+            # Use RUNNER_TEMP to avoid /tmp symlink issues on macOS
+            WORK_DIR="${RUNNER_TEMP}/metal-work"
+            mkdir -p "$WORK_DIR"
+
+            # Download the Metal toolchain component
+            xcodebuild -downloadComponent MetalToolchain
+
+            # Find and mount the DMG
+            DMG_PATH=$(find /System/Library/AssetsV2/com_apple_MobileAsset_MetalToolchain -name '*.dmg' 2>/dev/null | head -1)
+            if [ -z "$DMG_PATH" ]; then
+              echo "Error: Could not find Metal toolchain DMG"
+              exit 1
+            fi
+
+            echo "Found DMG at: $DMG_PATH"
+            hdiutil attach "$DMG_PATH" -mountpoint "${WORK_DIR}/metal-dmg"
+
+            # Copy the toolchain
+            cp -R "${WORK_DIR}/metal-dmg/Metal.xctoolchain" "${WORK_DIR}/metal-export"
+            hdiutil detach "${WORK_DIR}/metal-dmg"
+
+            # Create NAR and add to store
+            nix nar pack "${WORK_DIR}/metal-export" > "${WORK_DIR}/${NAR_NAME}"
+            STORE_PATH=$(nix store add --mode flat "${WORK_DIR}/${NAR_NAME}")
+            echo "Added NAR to store: $STORE_PATH"
+
+            # Verify the hash matches
+            ACTUAL_HASH=$(nix hash file "${WORK_DIR}/${NAR_NAME}")
+            if [ "$ACTUAL_HASH" != "$NAR_HASH" ]; then
+              echo "Warning: NAR hash mismatch!"
+              echo "Expected: $NAR_HASH"
+              echo "Actual:   $ACTUAL_HASH"
+              echo "The metal-toolchain.nix may need updating"
+            fi
+
+            # Clean up
+            rm -rf "$WORK_DIR"
+
+            # Retry the build now that NAR is in store
+            nix build .#metal-toolchain
+          fi
+
+          # Build mlx (depends on metal-toolchain)
+          nix build .#mlx
+
       - name: Build all Nix outputs
         run: |
           nix flake show --json | jq -r '
@@ -134,3 +132,14 @@ jobs:
 
       - name: Run nix flake check
         run: nix flake check
+
+      - name: Run pytest (macOS only)
+        if: runner.os == 'macOS'
+        run: |
+          # Build the test environment (requires relaxed sandbox for uv2nix on macOS)
+          TEST_ENV=$(nix build '.#exo-test-env' --option sandbox relaxed --print-out-paths)
+
+          # Run pytest outside sandbox (needs GPU access for MLX)
+          export HOME="$RUNNER_TEMP"
+          export EXO_TESTS=1
+          $TEST_ENV/bin/python -m pytest src -m "not slow" --import-mode=importlib

.gitignore

@@ -28,3 +28,6 @@ target/
 dashboard/build/
 dashboard/node_modules/
 dashboard/.svelte-kit/
+
+# host config snapshots
+hosts_*.json

README.md

@@ -5,7 +5,7 @@
   <img alt="exo logo" src="/docs/imgs/exo-logo-transparent.png" width="50%" height="50%">
 </picture>
 
-exo: Run your own AI cluster at home with everyday devices. Maintained by [exo labs](https://x.com/exolabs).
+exo: Run frontier AI locally. Maintained by [exo labs](https://x.com/exolabs).
 
 <p align="center">
   <a href="https://discord.gg/TJ4P57arEm" target="_blank" rel="noopener noreferrer"><img src="https://img.shields.io/badge/Discord-Join%20Server-5865F2?logo=discord&logoColor=white" alt="Discord"></a>
@@ -107,6 +107,10 @@ uv run exo
 
 This starts the exo dashboard and API at http://localhost:52415/
 
+
+*Please view the section on RDMA to enable this feature on MacOS >=26.2!*
+
+
 ### Run from Source (Linux)
 
 **Prerequisites:**
@@ -230,7 +234,7 @@ This removes:
 
 RDMA is a new capability added to macOS 26.2. It works on any Mac with Thunderbolt 5 (M4 Pro Mac Mini, M4 Max Mac Studio, M4 Max MacBook Pro, M3 Ultra Mac Studio).
 
-Note that on Mac Studio, you cannot use the Thunderbolt 5 port next to the Ethernet port.
+Please refer to the caveats for immediate troubleshooting.
 
 To enable RDMA on macOS, follow these steps:
 
@@ -247,6 +251,14 @@ To enable RDMA on macOS, follow these steps:
 
 After that, RDMA will be enabled in macOS and exo will take care of the rest.
 
+**Important Caveats**
+
+1. Devices that wish to be part of an RDMA cluster must be connected to all other devices in the cluster.
+2. The cables must support TB5.
+3. On a Mac Studio, you cannot use the Thunderbolt 5 port next to the Ethernet port.
+4. If running from source, please use the script found at `tmp/set_rdma_network_config.sh`, which will disable Thunderbolt Bridge and set dhcp on each RDMA port.
+5. RDMA ports may be unable to discover each other on different versions of MacOS. Please ensure that OS versions match exactly (even beta version numbers) on all devices.
+
 ---
 
 ### Using the API

app/EXO/EXO.xcodeproj/project.pbxproj

@@ -342,6 +342,8 @@
 				SDKROOT = macosx;
 				SWIFT_ACTIVE_COMPILATION_CONDITIONS = "DEBUG $(inherited)";
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+				SWIFT_TREAT_WARNINGS_AS_ERRORS = YES;
+				GCC_TREAT_WARNINGS_AS_ERRORS = YES;
 			};
 			name = Debug;
 		};
@@ -397,6 +399,8 @@
 				MTL_FAST_MATH = YES;
 				SDKROOT = macosx;
 				SWIFT_COMPILATION_MODE = wholemodule;
+				SWIFT_TREAT_WARNINGS_AS_ERRORS = YES;
+				GCC_TREAT_WARNINGS_AS_ERRORS = YES;
 			};
 			name = Release;
 		};

app/EXO/EXO.xcodeproj/project.xcworkspace/xcshareddata/swiftpm/Package.resolved

@@ -6,8 +6,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/sparkle-project/Sparkle.git",
       "state" : {
-        "revision" : "e641adb41915a8409895e2e30666aa64e487b637",
-        "version" : "2.9.0-beta.1"
+        "revision" : "afea2cda87819c960114f26e26f369a1a0945b17",
+        "version" : "2.9.0-beta.2"
       }
     }
   ],

app/EXO/EXO/ContentView.swift

@@ -14,6 +14,7 @@ struct ContentView: View {
     @EnvironmentObject private var networkStatusService: NetworkStatusService
     @EnvironmentObject private var localNetworkChecker: LocalNetworkChecker
     @EnvironmentObject private var updater: SparkleUpdater
+    @EnvironmentObject private var thunderboltBridgeService: ThunderboltBridgeService
     @State private var focusedNode: NodeViewModel?
     @State private var deletingInstanceIDs: Set<String> = []
     @State private var showAllNodes = false
@@ -24,6 +25,8 @@ struct ContentView: View {
     @State private var bugReportMessage: String?
     @State private var uninstallInProgress = false
     @State private var pendingNamespace: String = ""
+    @State private var pendingHFToken: String = ""
+    @State private var pendingEnableImageModels = false
 
     var body: some View {
         VStack(alignment: .leading, spacing: 12) {
@@ -303,6 +306,49 @@ struct ContentView: View {
                             .disabled(pendingNamespace == controller.customNamespace)
                         }
                     }
+                    VStack(alignment: .leading, spacing: 4) {
+                        Text("HuggingFace Token")
+                            .font(.caption2)
+                            .foregroundColor(.secondary)
+                        HStack {
+                            SecureField("optional", text: $pendingHFToken)
+                                .textFieldStyle(.roundedBorder)
+                                .font(.caption2)
+                                .onAppear {
+                                    pendingHFToken = controller.hfToken
+                                }
+                            Button("Save & Restart") {
+                                controller.hfToken = pendingHFToken
+                                if controller.status == .running || controller.status == .starting {
+                                    controller.restart()
+                                }
+                            }
+                            .font(.caption2)
+                            .disabled(pendingHFToken == controller.hfToken)
+                        }
+                    }
+                    Divider()
+                    HStack {
+                        Toggle(
+                            "Enable Image Models (experimental)", isOn: $pendingEnableImageModels
+                        )
+                        .toggleStyle(.switch)
+                        .font(.caption2)
+                        .onAppear {
+                            pendingEnableImageModels = controller.enableImageModels
+                        }
+
+                        Spacer()
+
+                        Button("Save & Restart") {
+                            controller.enableImageModels = pendingEnableImageModels
+                            if controller.status == .running || controller.status == .starting {
+                                controller.restart()
+                            }
+                        }
+                        .font(.caption2)
+                        .disabled(pendingEnableImageModels == controller.enableImageModels)
+                    }
                     HoverButton(title: "Check for Updates", small: true) {
                         updater.checkForUpdates()
                     }
@@ -423,6 +469,44 @@ struct ContentView: View {
         }
     }
 
+    /// Shows TB bridge status for all nodes from exo cluster state
+    private var clusterThunderboltBridgeView: some View {
+        let bridgeStatuses = stateService.latestSnapshot?.nodeThunderboltBridge ?? [:]
+        let localNodeId = stateService.localNodeId
+        let nodeProfiles = stateService.latestSnapshot?.nodeProfiles ?? [:]
+
+        return VStack(alignment: .leading, spacing: 1) {
+            if bridgeStatuses.isEmpty {
+                Text("Cluster TB Bridge: No data")
+                    .font(.caption2)
+                    .foregroundColor(.secondary)
+            } else {
+                Text("Cluster TB Bridge Status:")
+                    .font(.caption2)
+                    .foregroundColor(.secondary)
+                ForEach(Array(bridgeStatuses.keys.sorted()), id: \.self) { nodeId in
+                    if let status = bridgeStatuses[nodeId] {
+                        let nodeName =
+                            nodeProfiles[nodeId]?.friendlyName ?? String(nodeId.prefix(8))
+                        let isLocal = nodeId == localNodeId
+                        let prefix = isLocal ? "  \(nodeName) (local):" : "  \(nodeName):"
+                        let statusText =
+                            !status.exists
+                            ? "N/A"
+                            : (status.enabled ? "Enabled" : "Disabled")
+                        let color: Color =
+                            !status.exists
+                            ? .secondary
+                            : (status.enabled ? .red : .green)
+                        Text("\(prefix) \(statusText)")
+                            .font(.caption2)
+                            .foregroundColor(color)
+                    }
+                }
+            }
+        }
+    }
+
     private var interfaceIpList: some View {
         let statuses = networkStatusService.status.interfaceStatuses
         return VStack(alignment: .leading, spacing: 1) {
@@ -465,6 +549,7 @@ struct ContentView: View {
                     Text(thunderboltStatusText)
                         .font(.caption2)
                         .foregroundColor(thunderboltStatusColor)
+                    clusterThunderboltBridgeView
                     interfaceIpList
                     rdmaStatusView
                     sendBugReportButton

app/EXO/EXO/EXOApp.swift

@@ -21,6 +21,7 @@ struct EXOApp: App {
     @StateObject private var networkStatusService: NetworkStatusService
     @StateObject private var localNetworkChecker: LocalNetworkChecker
     @StateObject private var updater: SparkleUpdater
+    @StateObject private var thunderboltBridgeService: ThunderboltBridgeService
     private let terminationObserver: TerminationObserver
     private let ciContext = CIContext(options: nil)
 
@@ -41,10 +42,13 @@ struct EXOApp: App {
         let localNetwork = LocalNetworkChecker()
         _localNetworkChecker = StateObject(wrappedValue: localNetwork)
         _updater = StateObject(wrappedValue: updater)
+        let thunderboltBridge = ThunderboltBridgeService(clusterStateService: service)
+        _thunderboltBridgeService = StateObject(wrappedValue: thunderboltBridge)
         enableLaunchAtLoginIfNeeded()
-        NetworkSetupHelper.ensureLaunchDaemonInstalled()
-        // Check local network access BEFORE launching exo
-        localNetwork.check()
+        // Install LaunchDaemon to disable Thunderbolt Bridge on startup (prevents network loops)
+        NetworkSetupHelper.promptAndInstallIfNeeded()
+        // Check local network access periodically (warning disappears when user grants permission)
+        localNetwork.startPeriodicChecking(interval: 10)
         controller.scheduleLaunch(after: 15)
         service.startPolling()
         networkStatus.startPolling()
@@ -58,6 +62,7 @@ struct EXOApp: App {
                 .environmentObject(networkStatusService)
                 .environmentObject(localNetworkChecker)
                 .environmentObject(updater)
+                .environmentObject(thunderboltBridgeService)
         } label: {
             menuBarIcon
         }
@@ -130,6 +135,7 @@ struct EXOApp: App {
                 "Failed to register EXO for launch at login: \(error.localizedDescription)")
         }
     }
+
 }
 
 /// Helper for managing EXO's launch-at-login registration
@@ -219,7 +225,7 @@ private final class ExoUpdaterDelegate: NSObject, SPUUpdaterDelegate {
         }
     }
 
-    private func showNotification(title: String, body: String) {
+    nonisolated private func showNotification(title: String, body: String) {
         let center = UNUserNotificationCenter.current()
         let content = UNMutableNotificationContent()
         content.title = title

app/EXO/EXO/ExoProcessController.swift

@@ -3,6 +3,8 @@ import Combine
 import Foundation
 
 private let customNamespaceKey = "EXOCustomNamespace"
+private let hfTokenKey = "EXOHFToken"
+private let enableImageModelsKey = "EXOEnableImageModels"
 
 @MainActor
 final class ExoProcessController: ObservableObject {
@@ -37,6 +39,22 @@ final class ExoProcessController: ObservableObject {
             UserDefaults.standard.set(customNamespace, forKey: customNamespaceKey)
         }
     }
+    @Published var hfToken: String = {
+        return UserDefaults.standard.string(forKey: hfTokenKey) ?? ""
+    }()
+    {
+        didSet {
+            UserDefaults.standard.set(hfToken, forKey: hfTokenKey)
+        }
+    }
+    @Published var enableImageModels: Bool = {
+        return UserDefaults.standard.bool(forKey: enableImageModelsKey)
+    }()
+    {
+        didSet {
+            UserDefaults.standard.set(enableImageModels, forKey: enableImageModelsKey)
+        }
+    }
 
     private var process: Process?
     private var runtimeDirectoryURL: URL?
@@ -191,6 +209,12 @@ final class ExoProcessController: ObservableObject {
         var environment = ProcessInfo.processInfo.environment
         environment["EXO_RUNTIME_DIR"] = runtimeURL.path
         environment["EXO_LIBP2P_NAMESPACE"] = computeNamespace()
+        if !hfToken.isEmpty {
+            environment["HF_TOKEN"] = hfToken
+        }
+        if enableImageModels {
+            environment["EXO_ENABLE_IMAGE_MODELS"] = "true"
+        }
 
         var paths: [String] = []
         if let existing = environment["PATH"], !existing.isEmpty {

app/EXO/EXO/Models/ClusterState.swift

@@ -5,17 +5,43 @@ import Foundation
 struct ClusterState: Decodable {
     let instances: [String: ClusterInstance]
     let runners: [String: RunnerStatusSummary]
-    let nodeProfiles: [String: NodeProfile]
     let tasks: [String: ClusterTask]
     let topology: Topology?
     let downloads: [String: [NodeDownloadStatus]]
+    let thunderboltBridgeCycles: [[String]]
+
+    // Granular node state (split from the old nodeProfiles)
+    let nodeIdentities: [String: NodeIdentity]
+    let nodeMemory: [String: MemoryInfo]
+    let nodeSystem: [String: SystemInfo]
+    let nodeThunderboltBridge: [String: ThunderboltBridgeStatus]
+
+    /// Computed property for backwards compatibility - merges granular state into NodeProfile
+    var nodeProfiles: [String: NodeProfile] {
+        var profiles: [String: NodeProfile] = [:]
+        let allNodeIds = Set(nodeIdentities.keys)
+            .union(nodeMemory.keys)
+            .union(nodeSystem.keys)
+        for nodeId in allNodeIds {
+            let identity = nodeIdentities[nodeId]
+            let memory = nodeMemory[nodeId]
+            let system = nodeSystem[nodeId]
+            profiles[nodeId] = NodeProfile(
+                modelId: identity?.modelId,
+                chipId: identity?.chipId,
+                friendlyName: identity?.friendlyName,
+                memory: memory,
+                system: system
+            )
+        }
+        return profiles
+    }
 
     init(from decoder: Decoder) throws {
         let container = try decoder.container(keyedBy: CodingKeys.self)
         let rawInstances = try container.decode([String: TaggedInstance].self, forKey: .instances)
         self.instances = rawInstances.mapValues(\.instance)
         self.runners = try container.decode([String: RunnerStatusSummary].self, forKey: .runners)
-        self.nodeProfiles = try container.decode([String: NodeProfile].self, forKey: .nodeProfiles)
         let rawTasks =
             try container.decodeIfPresent([String: TaggedTask].self, forKey: .tasks) ?? [:]
         self.tasks = rawTasks.compactMapValues(\.task)
@@ -24,15 +50,34 @@ struct ClusterState: Decodable {
             try container.decodeIfPresent([String: [TaggedNodeDownload]].self, forKey: .downloads)
             ?? [:]
         self.downloads = rawDownloads.mapValues { $0.compactMap(\.status) }
+        self.thunderboltBridgeCycles =
+            try container.decodeIfPresent([[String]].self, forKey: .thunderboltBridgeCycles) ?? []
+
+        // Granular node state
+        self.nodeIdentities =
+            try container.decodeIfPresent([String: NodeIdentity].self, forKey: .nodeIdentities)
+            ?? [:]
+        self.nodeMemory =
+            try container.decodeIfPresent([String: MemoryInfo].self, forKey: .nodeMemory) ?? [:]
+        self.nodeSystem =
+            try container.decodeIfPresent([String: SystemInfo].self, forKey: .nodeSystem) ?? [:]
+        self.nodeThunderboltBridge =
+            try container.decodeIfPresent(
+                [String: ThunderboltBridgeStatus].self, forKey: .nodeThunderboltBridge
+            ) ?? [:]
     }
 
     private enum CodingKeys: String, CodingKey {
         case instances
         case runners
-        case nodeProfiles
         case topology
         case tasks
         case downloads
+        case thunderboltBridgeCycles
+        case nodeIdentities
+        case nodeMemory
+        case nodeSystem
+        case nodeThunderboltBridge
     }
 }
 
@@ -102,6 +147,18 @@ struct NodeProfile: Decodable {
     let system: SystemInfo?
 }
 
+struct NodeIdentity: Decodable {
+    let modelId: String?
+    let chipId: String?
+    let friendlyName: String?
+}
+
+struct ThunderboltBridgeStatus: Decodable {
+    let enabled: Bool
+    let exists: Bool
+    let serviceName: String?
+}
+
 struct MemoryInfo: Decodable {
     let ramTotal: MemoryValue?
     let ramAvailable: MemoryValue?
@@ -120,16 +177,51 @@ struct SystemInfo: Decodable {
 }
 
 struct Topology: Decodable {
-    let nodes: [TopologyNode]
-    let connections: [TopologyConnection]?
+    /// Node IDs in the topology
+    let nodes: [String]
+    /// Flattened list of connections (source -> sink pairs)
+    let connections: [TopologyConnection]
+
+    init(from decoder: Decoder) throws {
+        let container = try decoder.container(keyedBy: CodingKeys.self)
+        self.nodes = try container.decodeIfPresent([String].self, forKey: .nodes) ?? []
+
+        // Connections come as nested map: { source: { sink: [edges] } }
+        // We flatten to array of (source, sink) pairs
+        var flatConnections: [TopologyConnection] = []
+        if let nested = try container.decodeIfPresent(
+            [String: [String: [AnyCodable]]].self, forKey: .connections
+        ) {
+            for (source, sinks) in nested {
+                for sink in sinks.keys {
+                    flatConnections.append(
+                        TopologyConnection(localNodeId: source, sendBackNodeId: sink))
+                }
+            }
+        }
+        self.connections = flatConnections
+    }
+
+    private enum CodingKeys: String, CodingKey {
+        case nodes
+        case connections
+    }
 }
 
-struct TopologyNode: Decodable {
-    let nodeId: String
-    let nodeProfile: NodeProfile
+/// Placeholder for decoding arbitrary JSON values we don't need to inspect
+private struct AnyCodable: Decodable {
+    init(from decoder: Decoder) throws {
+        // Just consume the value without storing it
+        _ = try? decoder.singleValueContainer().decode(Bool.self)
+        _ = try? decoder.singleValueContainer().decode(Int.self)
+        _ = try? decoder.singleValueContainer().decode(Double.self)
+        _ = try? decoder.singleValueContainer().decode(String.self)
+        _ = try? decoder.singleValueContainer().decode([AnyCodable].self)
+        _ = try? decoder.singleValueContainer().decode([String: AnyCodable].self)
+    }
 }
 
-struct TopologyConnection: Decodable {
+struct TopologyConnection {
     let localNodeId: String
     let sendBackNodeId: String
 }
@@ -201,7 +293,7 @@ struct ClusterTask {
     let modelName: String?
     let promptPreview: String?
     let errorMessage: String?
-    let parameters: ChatCompletionTaskParameters?
+    let parameters: TextGenerationTaskParameters?
 
     var sortPriority: Int {
         switch status {
@@ -238,12 +330,12 @@ struct ClusterTaskPayload: Decodable {
     let taskStatus: TaskStatus?
     let instanceId: String?
     let commandId: String?
-    let taskParams: ChatCompletionTaskParameters?
+    let taskParams: TextGenerationTaskParameters?
     let errorType: String?
     let errorMessage: String?
 }
 
-struct ChatCompletionTaskParameters: Decodable, Equatable {
+struct TextGenerationTaskParameters: Decodable, Equatable {
     let model: String?
     let messages: [ChatCompletionMessage]?
     let maxTokens: Int?
@@ -282,7 +374,7 @@ extension ClusterTask {
         guard let id = payload.taskId else { return nil }
         let status = payload.taskStatus ?? .unknown
         switch kindKey {
-        case "ChatCompletion":
+        case "TextGeneration":
             self.init(
                 id: id,
                 status: status,

app/EXO/EXO/Services/BugReportService.swift

@@ -55,12 +55,16 @@ struct BugReportService {
         let stateData = try await stateResult
         let eventsData = try await eventsResult
 
+        // Extract cluster TB bridge status from exo state
+        let clusterTbBridgeStatus = extractClusterTbBridgeStatus(from: stateData)
+
         let reportJSON = makeReportJson(
             timestamp: timestamp,
             hostName: hostName,
             ifconfig: ifconfigText,
             debugInfo: debugInfo,
-            isManual: isManual
+            isManual: isManual,
+            clusterTbBridgeStatus: clusterTbBridgeStatus
         )
 
         let uploads: [(path: String, data: Data?)] = [
@@ -178,18 +182,19 @@ struct BugReportService {
     }
 
     private func readThunderboltBridgeDisabled() -> Bool? {
-        let result = runCommand([
-            "/usr/sbin/networksetup", "-getnetworkserviceenabled", "Thunderbolt Bridge",
-        ])
-        guard result.exitCode == 0 else { return nil }
-        let output = result.output.lowercased()
-        if output.contains("enabled") {
-            return false
+        // Dynamically find the Thunderbolt Bridge service (don't assume the name)
+        guard let serviceName = ThunderboltBridgeDetector.findThunderboltBridgeServiceName() else {
+            // No bridge containing Thunderbolt interfaces exists
+            return nil
         }
-        if output.contains("disabled") {
-            return true
+
+        guard let isEnabled = ThunderboltBridgeDetector.isServiceEnabled(serviceName: serviceName)
+        else {
+            return nil
         }
-        return nil
+
+        // Return true if disabled, false if enabled
+        return !isEnabled
     }
 
     private func readInterfaces() -> [DebugInfo.InterfaceStatus] {
@@ -268,11 +273,12 @@ struct BugReportService {
         hostName: String,
         ifconfig: String,
         debugInfo: DebugInfo,
-        isManual: Bool
+        isManual: Bool,
+        clusterTbBridgeStatus: [[String: Any]]?
     ) -> Data? {
         let system = readSystemMetadata()
         let exo = readExoMetadata()
-        let payload: [String: Any] = [
+        var payload: [String: Any] = [
             "timestamp": timestamp,
             "host": hostName,
             "ifconfig": ifconfig,
@@ -282,9 +288,38 @@ struct BugReportService {
             "exo_commit": exo.commit as Any,
             "report_type": isManual ? "manual" : "automated",
         ]
+        if let tbStatus = clusterTbBridgeStatus {
+            payload["cluster_thunderbolt_bridge"] = tbStatus
+        }
         return try? JSONSerialization.data(withJSONObject: payload, options: [.prettyPrinted])
     }
 
+    /// Extracts cluster-wide Thunderbolt Bridge status from exo state JSON
+    private func extractClusterTbBridgeStatus(from stateData: Data?) -> [[String: Any]]? {
+        guard let data = stateData,
+            let json = try? JSONSerialization.jsonObject(with: data) as? [String: Any],
+            let nodeThunderboltBridge = json["node_thunderbolt_bridge"] as? [String: [String: Any]]
+        else {
+            return nil
+        }
+
+        var result: [[String: Any]] = []
+        for (nodeId, status) in nodeThunderboltBridge {
+            var entry: [String: Any] = ["node_id": nodeId]
+            if let enabled = status["enabled"] as? Bool {
+                entry["enabled"] = enabled
+            }
+            if let exists = status["exists"] as? Bool {
+                entry["exists"] = exists
+            }
+            if let serviceName = status["service_name"] as? String {
+                entry["service_name"] = serviceName
+            }
+            result.append(entry)
+        }
+        return result.isEmpty ? nil : result
+    }
+
     private func readSystemMetadata() -> [String: Any] {
         let hostname = safeRunCommand(["/bin/hostname"])
         let computerName = safeRunCommand(["/usr/sbin/scutil", "--get", "ComputerName"])

app/EXO/EXO/Services/LocalNetworkChecker.swift

@@ -41,6 +41,7 @@ final class LocalNetworkChecker: ObservableObject {
 
     private var connection: NWConnection?
     private var checkTask: Task<Void, Never>?
+    private var periodicTask: Task<Void, Never>?
 
     /// Whether we've completed at least one check (stored in UserDefaults)
     private var hasCompletedInitialCheck: Bool {
@@ -48,10 +49,39 @@ final class LocalNetworkChecker: ObservableObject {
         set { UserDefaults.standard.set(newValue, forKey: Self.hasCompletedInitialCheckKey) }
     }
 
-    /// Checks if local network access is working.
+    /// Checks if local network access is working (one-time check).
     func check() {
+        performCheck()
+    }
+
+    /// Starts periodic checking of local network access.
+    /// Re-checks every `interval` seconds so the warning disappears when user grants permission.
+    func startPeriodicChecking(interval: TimeInterval = 10) {
+        stopPeriodicChecking()
+        // Do an immediate check first
+        performCheck()
+        // Then schedule periodic checks
+        periodicTask = Task { [weak self] in
+            while !Task.isCancelled {
+                try? await Task.sleep(nanoseconds: UInt64(interval * 1_000_000_000))
+                guard !Task.isCancelled else { break }
+                self?.performCheck()
+            }
+        }
+    }
+
+    /// Stops periodic checking.
+    func stopPeriodicChecking() {
+        periodicTask?.cancel()
+        periodicTask = nil
+    }
+
+    private func performCheck() {
         checkTask?.cancel()
-        status = .checking
+        // Only show "checking" status on first check to avoid UI flicker
+        if status == .unknown {
+            status = .checking
+        }
 
         // Use longer timeout on first launch to allow time for permission prompt
         let isFirstCheck = !hasCompletedInitialCheck
@@ -60,12 +90,15 @@ final class LocalNetworkChecker: ObservableObject {
         checkTask = Task { [weak self] in
             guard let self else { return }
 
-            Self.logger.info("Checking local network connectivity (first check: \(isFirstCheck))")
+            Self.logger.debug("Checking local network connectivity (first check: \(isFirstCheck))")
             let result = await self.checkConnectivity(timeout: timeout)
             self.status = result
             self.hasCompletedInitialCheck = true
 
-            Self.logger.info("Local network check complete: \(result.displayText)")
+            // Only log on state changes or first check to reduce noise
+            if isFirstCheck || result != self.status {
+                Self.logger.info("Local network check: \(result.displayText)")
+            }
         }
     }
 
@@ -141,6 +174,7 @@ final class LocalNetworkChecker: ObservableObject {
     }
 
     func stop() {
+        stopPeriodicChecking()
         checkTask?.cancel()
         checkTask = nil
         connection?.cancel()

app/EXO/EXO/Services/NetworkSetupHelper.swift

@@ -7,14 +7,20 @@ enum NetworkSetupHelper {
     private static let daemonLabel = "io.exo.networksetup"
     private static let scriptDestination =
         "/Library/Application Support/EXO/disable_bridge.sh"
+    // Legacy script path from older versions
+    private static let legacyScriptDestination =
+        "/Library/Application Support/EXO/disable_bridge_enable_dhcp.sh"
     private static let plistDestination = "/Library/LaunchDaemons/io.exo.networksetup.plist"
-    private static let requiredStartInterval: Int = 1791
+    private static let requiredStartInterval: Int = 1786
 
     private static let setupScript = """
         #!/usr/bin/env bash
 
         set -euo pipefail
 
+        # Wait for macOS to finish network setup after boot
+        sleep 20
+
         PREFS="/Library/Preferences/SystemConfiguration/preferences.plist"
 
         # Remove bridge0 interface
@@ -28,19 +34,69 @@ enum NetworkSetupHelper {
         # Remove Thunderbolt Bridge from VirtualNetworkInterfaces in preferences.plist
         /usr/libexec/PlistBuddy -c "Delete :VirtualNetworkInterfaces:Bridge:bridge0" "$PREFS" 2>/dev/null || true
 
+        networksetup -listlocations | grep -q exo || {
+          networksetup -createlocation exo
+        }
+
+        networksetup -switchtolocation exo
+        networksetup -listallhardwareports \\
+          | awk -F': ' '/Hardware Port: / {print $2}' \\
+          | while IFS=":" read -r name; do
+              case "$name" in
+                "Ethernet Adapter"*)
+                        ;;
+                "Thunderbolt Bridge")
+                        ;;
+                "Thunderbolt "*)
+                  networksetup -listallnetworkservices \\
+                    | grep -q "EXO $name" \\
+                      || networksetup -createnetworkservice "EXO $name" "$name" 2>/dev/null \\
+                      || continue
+                  networksetup -setdhcp "EXO $name"
+                        ;;
+                *)
+                  networksetup -listallnetworkservices \\
+                    | grep -q "$name" \\
+                      || networksetup -createnetworkservice "$name" "$name" 2>/dev/null \\
+                      || continue
+                        ;;
+              esac
+            done
+
         networksetup -listnetworkservices | grep -q "Thunderbolt Bridge" && {
           networksetup -setnetworkserviceenabled "Thunderbolt Bridge" off
         } || true
         """
 
-    static func ensureLaunchDaemonInstalled() {
+    /// Prompts user and installs the LaunchDaemon if not already installed.
+    /// Shows an alert explaining what will be installed before requesting admin privileges.
+    static func promptAndInstallIfNeeded() {
         // Use .utility priority to match NSAppleScript's internal QoS and avoid priority inversion
         Task.detached(priority: .utility) {
+            // If already correctly installed, skip
+            if daemonAlreadyInstalled() {
+                return
+            }
+
+            // Show alert on main thread
+            let shouldInstall = await MainActor.run {
+                let alert = NSAlert()
+                alert.messageText = "EXO Network Configuration"
+                alert.informativeText =
+                    "EXO needs to install a system service to configure local networking. This will disable Thunderbolt Bridge (preventing packet storms) and install a Network Location.\n\nYou will be prompted for your password."
+                alert.alertStyle = .informational
+                alert.addButton(withTitle: "Install")
+                alert.addButton(withTitle: "Not Now")
+                return alert.runModal() == .alertFirstButtonReturn
+            }
+
+            guard shouldInstall else {
+                logger.info("User deferred network setup daemon installation")
+                return
+            }
+
             do {
-                if daemonAlreadyInstalled() {
-                    return
-                }
-                try await installLaunchDaemon()
+                try installLaunchDaemon()
                 logger.info("Network setup launch daemon installed and started")
             } catch {
                 logger.error(
@@ -63,48 +119,9 @@ enum NetworkSetupHelper {
     static func hasInstalledComponents() -> Bool {
         let manager = FileManager.default
         let scriptExists = manager.fileExists(atPath: scriptDestination)
+        let legacyScriptExists = manager.fileExists(atPath: legacyScriptDestination)
         let plistExists = manager.fileExists(atPath: plistDestination)
-        return scriptExists || plistExists
-    }
-
-    private static func makeUninstallScript() -> String {
-        """
-        set -euo pipefail
-
-        LABEL="\(daemonLabel)"
-        SCRIPT_DEST="\(scriptDestination)"
-        PLIST_DEST="\(plistDestination)"
-        LOG_OUT="/var/log/\(daemonLabel).log"
-        LOG_ERR="/var/log/\(daemonLabel).err.log"
-
-        # Unload the LaunchDaemon if running
-        launchctl bootout system/"$LABEL" 2>/dev/null || true
-
-        # Remove LaunchDaemon plist
-        rm -f "$PLIST_DEST"
-
-        # Remove the script and parent directory if empty
-        rm -f "$SCRIPT_DEST"
-        rmdir "$(dirname "$SCRIPT_DEST")" 2>/dev/null || true
-
-        # Remove log files
-        rm -f "$LOG_OUT" "$LOG_ERR"
-
-        # Switch back to Automatic network location
-        networksetup -switchtolocation Automatic 2>/dev/null || true
-
-        # Delete the exo network location if it exists
-        networksetup -listlocations | grep -q '^exo$' && {
-          networksetup -deletelocation exo 2>/dev/null || true
-        } || true
-
-        # Re-enable Thunderbolt Bridge if it exists
-        networksetup -listnetworkservices | grep -q "Thunderbolt Bridge" && {
-          networksetup -setnetworkserviceenabled "Thunderbolt Bridge" on 2>/dev/null || true
-        } || true
-
-        echo "EXO network components removed successfully"
-        """
+        return scriptExists || legacyScriptExists || plistExists
     }
 
     private static func daemonAlreadyInstalled() -> Bool {
@@ -140,7 +157,7 @@ enum NetworkSetupHelper {
         return true
     }
 
-    private static func installLaunchDaemon() async throws {
+    private static func installLaunchDaemon() throws {
         let installerScript = makeInstallerScript()
         try runShellAsAdmin(installerScript)
     }
@@ -151,8 +168,19 @@ enum NetworkSetupHelper {
 
         LABEL="\(daemonLabel)"
         SCRIPT_DEST="\(scriptDestination)"
+        LEGACY_SCRIPT_DEST="\(legacyScriptDestination)"
         PLIST_DEST="\(plistDestination)"
+        LOG_OUT="/var/log/\(daemonLabel).log"
+        LOG_ERR="/var/log/\(daemonLabel).err.log"
 
+        # First, completely remove any existing installation
+        launchctl bootout system/"$LABEL" 2>/dev/null || true
+        rm -f "$PLIST_DEST"
+        rm -f "$SCRIPT_DEST"
+        rm -f "$LEGACY_SCRIPT_DEST"
+        rm -f "$LOG_OUT" "$LOG_ERR"
+
+        # Install fresh
         mkdir -p "$(dirname "$SCRIPT_DEST")"
 
         cat > "$SCRIPT_DEST" <<'EOF_SCRIPT'
@@ -184,13 +212,82 @@ enum NetworkSetupHelper {
         </plist>
         EOF_PLIST
 
-        launchctl bootout system/"$LABEL" >/dev/null 2>&1 || true
         launchctl bootstrap system "$PLIST_DEST"
         launchctl enable system/"$LABEL"
         launchctl kickstart -k system/"$LABEL"
         """
     }
 
+    private static func makeUninstallScript() -> String {
+        """
+        set -euo pipefail
+
+        LABEL="\(daemonLabel)"
+        SCRIPT_DEST="\(scriptDestination)"
+        LEGACY_SCRIPT_DEST="\(legacyScriptDestination)"
+        PLIST_DEST="\(plistDestination)"
+        LOG_OUT="/var/log/\(daemonLabel).log"
+        LOG_ERR="/var/log/\(daemonLabel).err.log"
+
+        # Unload the LaunchDaemon if running
+        launchctl bootout system/"$LABEL" 2>/dev/null || true
+
+        # Remove LaunchDaemon plist
+        rm -f "$PLIST_DEST"
+
+        # Remove the script (current and legacy paths) and parent directory if empty
+        rm -f "$SCRIPT_DEST"
+        rm -f "$LEGACY_SCRIPT_DEST"
+        rmdir "$(dirname "$SCRIPT_DEST")" 2>/dev/null || true
+
+        # Remove log files
+        rm -f "$LOG_OUT" "$LOG_ERR"
+
+        # Switch back to Automatic network location
+        networksetup -switchtolocation Automatic >/dev/null 2>&1 || true
+
+        # Delete the exo network location if it exists
+        networksetup -listlocations 2>/dev/null | grep -q '^exo$' && {
+          networksetup -deletelocation exo >/dev/null 2>&1 || true
+        } || true
+
+        # Re-enable any Thunderbolt Bridge service if it exists
+        # We find it dynamically by looking for bridges containing Thunderbolt interfaces
+        find_and_enable_thunderbolt_bridge() {
+          # Get Thunderbolt interface devices from hardware ports
+          tb_devices=$(networksetup -listallhardwareports 2>/dev/null | awk '
+            /^Hardware Port:/ { port = tolower(substr($0, 16)) }
+            /^Device:/ { if (port ~ /thunderbolt/) print substr($0, 9) }
+          ') || true
+          [ -z "$tb_devices" ] && return 0
+
+          # For each bridge device, check if it contains Thunderbolt interfaces
+          for bridge in bridge0 bridge1 bridge2; do
+            members=$(ifconfig "$bridge" 2>/dev/null | awk '/member:/ {print $2}') || true
+            [ -z "$members" ] && continue
+
+            for tb_dev in $tb_devices; do
+              if echo "$members" | grep -qx "$tb_dev"; then
+                # Find the service name for this bridge device
+                service_name=$(networksetup -listnetworkserviceorder 2>/dev/null | awk -v dev="$bridge" '
+                  /^\\([0-9*]/ { gsub(/^\\([0-9*]+\\) /, ""); svc = $0 }
+                  /Device:/ && $0 ~ dev { print svc; exit }
+                ') || true
+                if [ -n "$service_name" ]; then
+                  networksetup -setnetworkserviceenabled "$service_name" on 2>/dev/null || true
+                  return 0
+                fi
+              fi
+            done
+          done
+          return 0
+        }
+        find_and_enable_thunderbolt_bridge || true
+
+        echo "EXO network components removed successfully"
+        """
+    }
+
     private static func runShellAsAdmin(_ script: String) throws {
         let escapedScript =
             script

app/EXO/EXO/Services/NetworkStatusService.swift

@@ -153,22 +153,18 @@ private struct NetworkStatusFetcher {
     }
 
     private func readThunderboltBridgeState() -> ThunderboltState? {
-        let result = runCommand(["networksetup", "-getnetworkserviceenabled", "Thunderbolt Bridge"])
-        guard result.exitCode == 0 else {
-            let lower = result.output.lowercased() + result.error.lowercased()
-            if lower.contains("not a recognized network service") {
-                return .deleted
-            }
+        // Dynamically find the Thunderbolt Bridge service (don't assume the name)
+        guard let serviceName = ThunderboltBridgeDetector.findThunderboltBridgeServiceName() else {
+            // No bridge containing Thunderbolt interfaces exists
+            return .deleted
+        }
+
+        guard let isEnabled = ThunderboltBridgeDetector.isServiceEnabled(serviceName: serviceName)
+        else {
             return nil
         }
-        let output = result.output.lowercased()
-        if output.contains("enabled") {
-            return .enabled
-        }
-        if output.contains("disabled") {
-            return .disabled
-        }
-        return nil
+
+        return isEnabled ? .enabled : .disabled
     }
 
     private func readBridgeInactive() -> Bool? {

app/EXO/EXO/Services/ThunderboltBridgeDetector.swift

@@ -0,0 +1,194 @@
+import Foundation
+import os.log
+
+/// Utility for dynamically detecting Thunderbolt Bridge network services.
+/// This mirrors the Python logic in info_gatherer.py - we never assume the service
+/// is named "Thunderbolt Bridge", instead we find bridges containing Thunderbolt interfaces.
+enum ThunderboltBridgeDetector {
+    private static let logger = Logger(
+        subsystem: "io.exo.EXO", category: "ThunderboltBridgeDetector")
+
+    struct CommandResult {
+        let exitCode: Int32
+        let output: String
+        let error: String
+    }
+
+    /// Find the network service name of a bridge containing Thunderbolt interfaces.
+    /// Returns nil if no such bridge exists.
+    static func findThunderboltBridgeServiceName() -> String? {
+        // 1. Get all Thunderbolt interface devices (e.g., en2, en3)
+        guard let thunderboltDevices = getThunderboltDevices(), !thunderboltDevices.isEmpty else {
+            logger.debug("No Thunderbolt devices found")
+            return nil
+        }
+        logger.debug("Found Thunderbolt devices: \(thunderboltDevices.joined(separator: ", "))")
+
+        // 2. Get bridge services from network service order
+        guard let bridgeServices = getBridgeServices(), !bridgeServices.isEmpty else {
+            logger.debug("No bridge services found")
+            return nil
+        }
+        logger.debug("Found bridge services: \(bridgeServices.keys.joined(separator: ", "))")
+
+        // 3. Find a bridge that contains Thunderbolt interfaces
+        for (bridgeDevice, serviceName) in bridgeServices {
+            let members = getBridgeMembers(bridgeDevice: bridgeDevice)
+            logger.debug(
+                "Bridge \(bridgeDevice) (\(serviceName)) has members: \(members.joined(separator: ", "))"
+            )
+
+            // Check if any Thunderbolt device is a member of this bridge
+            if !members.isDisjoint(with: thunderboltDevices) {
+                logger.info(
+                    "Found Thunderbolt Bridge service: '\(serviceName)' (device: \(bridgeDevice))")
+                return serviceName
+            }
+        }
+
+        logger.debug("No bridge found containing Thunderbolt interfaces")
+        return nil
+    }
+
+    /// Get Thunderbolt interface device names (e.g., en2, en3) from hardware ports.
+    private static func getThunderboltDevices() -> Set<String>? {
+        let result = runCommand(["networksetup", "-listallhardwareports"])
+        guard result.exitCode == 0 else {
+            logger.warning("networksetup -listallhardwareports failed: \(result.error)")
+            return nil
+        }
+
+        var thunderboltDevices: Set<String> = []
+        var currentPort: String?
+
+        for line in result.output.components(separatedBy: .newlines) {
+            let trimmed = line.trimmingCharacters(in: .whitespaces)
+            if trimmed.hasPrefix("Hardware Port:") {
+                currentPort = String(trimmed.dropFirst("Hardware Port:".count)).trimmingCharacters(
+                    in: .whitespaces)
+            } else if trimmed.hasPrefix("Device:"), let port = currentPort {
+                let device = String(trimmed.dropFirst("Device:".count)).trimmingCharacters(
+                    in: .whitespaces)
+                if port.lowercased().contains("thunderbolt") {
+                    thunderboltDevices.insert(device)
+                }
+                currentPort = nil
+            }
+        }
+
+        return thunderboltDevices
+    }
+
+    /// Get mapping of bridge device -> service name from network service order.
+    private static func getBridgeServices() -> [String: String]? {
+        let result = runCommand(["networksetup", "-listnetworkserviceorder"])
+        guard result.exitCode == 0 else {
+            logger.warning("networksetup -listnetworkserviceorder failed: \(result.error)")
+            return nil
+        }
+
+        // Parse service order to find bridge devices and their service names
+        // Format: "(1) Service Name\n(Hardware Port: ..., Device: bridge0)\n"
+        var bridgeServices: [String: String] = [:]
+        var currentService: String?
+
+        for line in result.output.components(separatedBy: .newlines) {
+            let trimmed = line.trimmingCharacters(in: .whitespaces)
+
+            // Match "(N) Service Name" or "(*) Service Name" (disabled)
+            // but NOT "(Hardware Port: ...)" lines
+            if trimmed.hasPrefix("("), trimmed.contains(")"),
+                !trimmed.hasPrefix("(Hardware Port:")
+            {
+                if let parenEnd = trimmed.firstIndex(of: ")") {
+                    let afterParen = trimmed.index(after: parenEnd)
+                    if afterParen < trimmed.endIndex {
+                        currentService =
+                            String(trimmed[afterParen...])
+                            .trimmingCharacters(in: .whitespaces)
+                    }
+                }
+            }
+            // Match "(Hardware Port: ..., Device: bridgeX)"
+            else if let service = currentService, trimmed.contains("Device: bridge") {
+                // Extract device name from "..., Device: bridge0)"
+                if let deviceRange = trimmed.range(of: "Device: ") {
+                    let afterDevice = trimmed[deviceRange.upperBound...]
+                    if let parenIndex = afterDevice.firstIndex(of: ")") {
+                        let device = String(afterDevice[..<parenIndex])
+                        bridgeServices[device] = service
+                    }
+                }
+            }
+        }
+
+        return bridgeServices
+    }
+
+    /// Get member interfaces of a bridge device via ifconfig.
+    private static func getBridgeMembers(bridgeDevice: String) -> Set<String> {
+        let result = runCommand(["ifconfig", bridgeDevice])
+        guard result.exitCode == 0 else {
+            logger.debug("ifconfig \(bridgeDevice) failed")
+            return []
+        }
+
+        var members: Set<String> = []
+        for line in result.output.components(separatedBy: .newlines) {
+            let trimmed = line.trimmingCharacters(in: .whitespaces)
+            if trimmed.hasPrefix("member:") {
+                let parts = trimmed.split(separator: " ")
+                if parts.count > 1 {
+                    members.insert(String(parts[1]))
+                }
+            }
+        }
+
+        return members
+    }
+
+    /// Check if a network service is enabled.
+    static func isServiceEnabled(serviceName: String) -> Bool? {
+        let result = runCommand(["networksetup", "-getnetworkserviceenabled", serviceName])
+        guard result.exitCode == 0 else {
+            logger.warning("Failed to check if '\(serviceName)' is enabled: \(result.error)")
+            return nil
+        }
+
+        let output = result.output.lowercased().trimmingCharacters(in: .whitespacesAndNewlines)
+        if output.contains("enabled") {
+            return true
+        }
+        if output.contains("disabled") {
+            return false
+        }
+        return nil
+    }
+
+    private static func runCommand(_ arguments: [String]) -> CommandResult {
+        let process = Process()
+        process.launchPath = "/usr/bin/env"
+        process.arguments = arguments
+
+        let stdout = Pipe()
+        let stderr = Pipe()
+        process.standardOutput = stdout
+        process.standardError = stderr
+
+        do {
+            try process.run()
+        } catch {
+            return CommandResult(exitCode: -1, output: "", error: error.localizedDescription)
+        }
+        process.waitUntilExit()
+
+        let outputData = stdout.fileHandleForReading.readDataToEndOfFile()
+        let errorData = stderr.fileHandleForReading.readDataToEndOfFile()
+
+        return CommandResult(
+            exitCode: process.terminationStatus,
+            output: String(decoding: outputData, as: UTF8.self),
+            error: String(decoding: errorData, as: UTF8.self)
+        )
+    }
+}

app/EXO/EXO/Services/ThunderboltBridgeService.swift

@@ -0,0 +1,261 @@
+import AppKit
+import Combine
+import Foundation
+import Security
+import SystemConfiguration
+import os.log
+
+@MainActor
+final class ThunderboltBridgeService: ObservableObject {
+    private static let logger = Logger(subsystem: "io.exo.EXO", category: "ThunderboltBridge")
+
+    @Published private(set) var detectedCycle: [String]?
+    @Published private(set) var hasPromptedForCurrentCycle = false
+    @Published private(set) var lastError: String?
+
+    private weak var clusterStateService: ClusterStateService?
+    private var cancellables = Set<AnyCancellable>()
+    private var previousCycleSignature: String?
+
+    init(clusterStateService: ClusterStateService) {
+        self.clusterStateService = clusterStateService
+        setupObserver()
+    }
+
+    private func setupObserver() {
+        guard let service = clusterStateService else { return }
+
+        service.$latestSnapshot
+            .compactMap { $0 }
+            .sink { [weak self] snapshot in
+                self?.checkForCycles(snapshot: snapshot)
+            }
+            .store(in: &cancellables)
+    }
+
+    private func checkForCycles(snapshot: ClusterState) {
+        let cycles = snapshot.thunderboltBridgeCycles
+
+        // Only consider cycles with more than 2 nodes
+        guard let firstCycle = cycles.first, firstCycle.count > 2 else {
+            // No problematic cycles detected, reset state
+            if detectedCycle != nil {
+                detectedCycle = nil
+                previousCycleSignature = nil
+                hasPromptedForCurrentCycle = false
+            }
+            return
+        }
+
+        // Create a signature for this cycle to detect if it changed
+        let cycleSignature = firstCycle.sorted().joined(separator: ",")
+
+        // If this is a new/different cycle, reset the prompt state
+        if cycleSignature != previousCycleSignature {
+            previousCycleSignature = cycleSignature
+            hasPromptedForCurrentCycle = false
+        }
+
+        detectedCycle = firstCycle
+
+        // Only prompt once per cycle
+        if !hasPromptedForCurrentCycle {
+            showDisableBridgePrompt(nodeIds: firstCycle)
+        }
+    }
+
+    private func showDisableBridgePrompt(nodeIds: [String]) {
+        hasPromptedForCurrentCycle = true
+
+        // Get friendly names for the nodes if available
+        let nodeNames = nodeIds.map { nodeId -> String in
+            if let snapshot = clusterStateService?.latestSnapshot,
+                let profile = snapshot.nodeProfiles[nodeId],
+                let friendlyName = profile.friendlyName, !friendlyName.isEmpty
+            {
+                return friendlyName
+            }
+            return String(nodeId.prefix(8))  // Use first 8 chars of node ID as fallback
+        }
+        let machineNames = nodeNames.joined(separator: ", ")
+
+        let alert = NSAlert()
+        alert.messageText = "Thunderbolt Bridge Loop Detected"
+        alert.informativeText = """
+            A Thunderbolt Bridge loop has been detected between \(nodeNames.count) machines: \(machineNames).
+
+            This can cause network packet storms and connectivity issues. Would you like to disable Thunderbolt Bridge on this machine to break the loop?
+            """
+        alert.alertStyle = .warning
+        alert.addButton(withTitle: "Disable Bridge")
+        alert.addButton(withTitle: "Not Now")
+
+        let response = alert.runModal()
+
+        if response == .alertFirstButtonReturn {
+            Task {
+                await disableThunderboltBridge()
+            }
+        }
+    }
+
+    func disableThunderboltBridge() async {
+        Self.logger.info("Attempting to disable Thunderbolt Bridge via SCPreferences")
+        lastError = nil
+
+        do {
+            try await disableThunderboltBridgeWithSCPreferences()
+            Self.logger.info("Successfully disabled Thunderbolt Bridge")
+        } catch {
+            Self.logger.error(
+                "Failed to disable Thunderbolt Bridge: \(error.localizedDescription, privacy: .public)"
+            )
+            lastError = error.localizedDescription
+            showErrorAlert(message: error.localizedDescription)
+        }
+    }
+
+    private func disableThunderboltBridgeWithSCPreferences() async throws {
+        // 1. Create authorization reference
+        var authRef: AuthorizationRef?
+        var status = AuthorizationCreate(nil, nil, [], &authRef)
+        guard status == errAuthorizationSuccess, let authRef = authRef else {
+            throw ThunderboltBridgeError.authorizationFailed
+        }
+
+        defer { AuthorizationFree(authRef, [.destroyRights]) }
+
+        // 2. Request specific network configuration rights
+        let rightName = "system.services.systemconfiguration.network"
+        status = rightName.withCString { nameCString in
+            var item = AuthorizationItem(
+                name: nameCString,
+                valueLength: 0,
+                value: nil,
+                flags: 0
+            )
+            return withUnsafeMutablePointer(to: &item) { itemPointer in
+                var rights = AuthorizationRights(count: 1, items: itemPointer)
+                return AuthorizationCopyRights(
+                    authRef,
+                    &rights,
+                    nil,
+                    [.extendRights, .interactionAllowed],
+                    nil
+                )
+            }
+        }
+        guard status == errAuthorizationSuccess else {
+            if status == errAuthorizationCanceled {
+                throw ThunderboltBridgeError.authorizationCanceled
+            }
+            throw ThunderboltBridgeError.authorizationDenied
+        }
+
+        // 3. Create SCPreferences with authorization
+        guard
+            let prefs = SCPreferencesCreateWithAuthorization(
+                kCFAllocatorDefault,
+                "EXO" as CFString,
+                nil,
+                authRef
+            )
+        else {
+            throw ThunderboltBridgeError.preferencesCreationFailed
+        }
+
+        // 4. Lock, modify, commit
+        guard SCPreferencesLock(prefs, true) else {
+            throw ThunderboltBridgeError.lockFailed
+        }
+
+        defer {
+            SCPreferencesUnlock(prefs)
+        }
+
+        // 5. Find the Thunderbolt Bridge service dynamically (don't assume the name)
+        guard let targetServiceName = ThunderboltBridgeDetector.findThunderboltBridgeServiceName()
+        else {
+            throw ThunderboltBridgeError.serviceNotFound
+        }
+
+        guard let allServices = SCNetworkServiceCopyAll(prefs) as? [SCNetworkService] else {
+            throw ThunderboltBridgeError.servicesNotFound
+        }
+
+        var found = false
+        for service in allServices {
+            if let name = SCNetworkServiceGetName(service) as String?,
+                name == targetServiceName
+            {
+                guard SCNetworkServiceSetEnabled(service, false) else {
+                    throw ThunderboltBridgeError.disableFailed
+                }
+                found = true
+                Self.logger.info(
+                    "Found and disabled Thunderbolt Bridge service: '\(targetServiceName)'")
+                break
+            }
+        }
+
+        if !found {
+            throw ThunderboltBridgeError.serviceNotFound
+        }
+
+        // 6. Commit and apply
+        guard SCPreferencesCommitChanges(prefs) else {
+            throw ThunderboltBridgeError.commitFailed
+        }
+
+        guard SCPreferencesApplyChanges(prefs) else {
+            throw ThunderboltBridgeError.applyFailed
+        }
+    }
+
+    private func showErrorAlert(message: String) {
+        let alert = NSAlert()
+        alert.messageText = "Failed to Disable Thunderbolt Bridge"
+        alert.informativeText = message
+        alert.alertStyle = .critical
+        alert.addButton(withTitle: "OK")
+        alert.runModal()
+    }
+}
+
+enum ThunderboltBridgeError: LocalizedError {
+    case authorizationFailed
+    case authorizationCanceled
+    case authorizationDenied
+    case preferencesCreationFailed
+    case lockFailed
+    case servicesNotFound
+    case serviceNotFound
+    case disableFailed
+    case commitFailed
+    case applyFailed
+
+    var errorDescription: String? {
+        switch self {
+        case .authorizationFailed:
+            return "Failed to create authorization"
+        case .authorizationCanceled:
+            return "Authorization was canceled by user"
+        case .authorizationDenied:
+            return "Authorization was denied"
+        case .preferencesCreationFailed:
+            return "Failed to access network preferences"
+        case .lockFailed:
+            return "Failed to lock network preferences for modification"
+        case .servicesNotFound:
+            return "Could not retrieve network services"
+        case .serviceNotFound:
+            return "Thunderbolt Bridge service not found"
+        case .disableFailed:
+            return "Failed to disable Thunderbolt Bridge service"
+        case .commitFailed:
+            return "Failed to save network configuration changes"
+        case .applyFailed:
+            return "Failed to apply network configuration changes"
+        }
+    }
+}

app/EXO/EXO/ViewModels/InstanceViewModel.swift

@@ -216,7 +216,7 @@ struct InstanceTaskViewModel: Identifiable, Equatable {
     let promptPreview: String?
     let errorMessage: String?
     let subtitle: String?
-    let parameters: ChatCompletionTaskParameters?
+    let parameters: TextGenerationTaskParameters?
 
     var title: String {
         switch kind {

app/EXO/EXO/ViewModels/NodeViewModel.swift

@@ -86,7 +86,7 @@ struct TopologyViewModel {
 
 extension ClusterState {
     func topologyViewModel(localNodeId: String?) -> TopologyViewModel? {
-        let topologyNodeIds = Set(topology?.nodes.map(\.nodeId) ?? [])
+        let topologyNodeIds = Set(topology?.nodes ?? [])
         let allNodes = nodeViewModels().filter {
             topologyNodeIds.isEmpty || topologyNodeIds.contains($0.id)
         }
@@ -95,8 +95,8 @@ extension ClusterState {
         let nodesById = Dictionary(uniqueKeysWithValues: allNodes.map { ($0.id, $0) })
         var orderedNodes: [NodeViewModel] = []
         if let topologyNodes = topology?.nodes {
-            for topoNode in topologyNodes {
-                if let viewModel = nodesById[topoNode.nodeId] {
+            for nodeId in topologyNodes {
+                if let viewModel = nodesById[nodeId] {
                     orderedNodes.append(viewModel)
                 }
             }
@@ -116,7 +116,7 @@ extension ClusterState {
 
         let nodeIds = Set(orderedNodes.map(\.id))
         let edgesArray: [TopologyEdgeViewModel] =
-            topology?.connections?.compactMap { connection in
+            topology?.connections.compactMap { connection in
                 guard nodeIds.contains(connection.localNodeId),
                     nodeIds.contains(connection.sendBackNodeId)
                 else { return nil }

app/EXO/uninstall-exo.sh

@@ -29,21 +29,21 @@ YELLOW='\033[1;33m'
 NC='\033[0m' # No Color
 
 echo_info() {
-    echo -e "${GREEN}[INFO]${NC} $1"
+  echo -e "${GREEN}[INFO]${NC} $1"
 }
 
 echo_warn() {
-    echo -e "${YELLOW}[WARN]${NC} $1"
+  echo -e "${YELLOW}[WARN]${NC} $1"
 }
 
 echo_error() {
-    echo -e "${RED}[ERROR]${NC} $1"
+  echo -e "${RED}[ERROR]${NC} $1"
 }
 
 # Check if running as root
 if [[ $EUID -ne 0 ]]; then
-    echo_error "This script must be run as root (use sudo)"
-    exit 1
+  echo_error "This script must be run as root (use sudo)"
+  exit 1
 fi
 
 echo ""
@@ -55,64 +55,64 @@ echo ""
 # Unload the LaunchDaemon if running
 echo_info "Stopping network setup daemon..."
 if launchctl list | grep -q "$LABEL"; then
-    launchctl bootout system/"$LABEL" 2>/dev/null || true
-    echo_info "Daemon stopped"
+  launchctl bootout system/"$LABEL" 2>/dev/null || true
+  echo_info "Daemon stopped"
 else
-    echo_warn "Daemon was not running"
+  echo_warn "Daemon was not running"
 fi
 
 # Remove LaunchDaemon plist
-if [[ -f "$PLIST_DEST" ]]; then
-    rm -f "$PLIST_DEST"
-    echo_info "Removed LaunchDaemon plist"
+if [[ -f $PLIST_DEST ]]; then
+  rm -f "$PLIST_DEST"
+  echo_info "Removed LaunchDaemon plist"
 else
-    echo_warn "LaunchDaemon plist not found (already removed?)"
+  echo_warn "LaunchDaemon plist not found (already removed?)"
 fi
 
 # Remove the script and parent directory
-if [[ -f "$SCRIPT_DEST" ]]; then
-    rm -f "$SCRIPT_DEST"
-    echo_info "Removed network setup script"
+if [[ -f $SCRIPT_DEST ]]; then
+  rm -f "$SCRIPT_DEST"
+  echo_info "Removed network setup script"
 else
-    echo_warn "Network setup script not found (already removed?)"
+  echo_warn "Network setup script not found (already removed?)"
 fi
 
 # Remove EXO directory if empty
 if [[ -d "/Library/Application Support/EXO" ]]; then
-    rmdir "/Library/Application Support/EXO" 2>/dev/null && \
-        echo_info "Removed EXO support directory" || \
-        echo_warn "EXO support directory not empty, leaving in place"
+  rmdir "/Library/Application Support/EXO" 2>/dev/null &&
+    echo_info "Removed EXO support directory" ||
+    echo_warn "EXO support directory not empty, leaving in place"
 fi
 
 # Remove log files
-if [[ -f "$LOG_OUT" ]] || [[ -f "$LOG_ERR" ]]; then
-    rm -f "$LOG_OUT" "$LOG_ERR"
-    echo_info "Removed log files"
+if [[ -f $LOG_OUT ]] || [[ -f $LOG_ERR ]]; then
+  rm -f "$LOG_OUT" "$LOG_ERR"
+  echo_info "Removed log files"
 else
-    echo_warn "Log files not found (already removed?)"
+  echo_warn "Log files not found (already removed?)"
 fi
 
 # Switch back to Automatic network location
 echo_info "Restoring network configuration..."
 if networksetup -listlocations | grep -q "^Automatic$"; then
-    networksetup -switchtolocation Automatic 2>/dev/null || true
-    echo_info "Switched to Automatic network location"
+  networksetup -switchtolocation Automatic 2>/dev/null || true
+  echo_info "Switched to Automatic network location"
 else
-    echo_warn "Automatic network location not found"
+  echo_warn "Automatic network location not found"
 fi
 
 # Delete the exo network location if it exists
 if networksetup -listlocations | grep -q "^exo$"; then
-    networksetup -deletelocation exo 2>/dev/null || true
-    echo_info "Deleted 'exo' network location"
+  networksetup -deletelocation exo 2>/dev/null || true
+  echo_info "Deleted 'exo' network location"
 else
-    echo_warn "'exo' network location not found (already removed?)"
+  echo_warn "'exo' network location not found (already removed?)"
 fi
 
 # Re-enable Thunderbolt Bridge if it exists
 if networksetup -listnetworkservices 2>/dev/null | grep -q "Thunderbolt Bridge"; then
-    networksetup -setnetworkserviceenabled "Thunderbolt Bridge" on 2>/dev/null || true
-    echo_info "Re-enabled Thunderbolt Bridge"
+  networksetup -setnetworkserviceenabled "Thunderbolt Bridge" on 2>/dev/null || true
+  echo_info "Re-enabled Thunderbolt Bridge"
 fi
 
 # Note about launch at login registration
@@ -124,14 +124,14 @@ echo_warn "  System Settings → General → Login Items → Remove EXO"
 # Check if EXO.app exists in common locations
 APP_FOUND=false
 for app_path in "/Applications/EXO.app" "$HOME/Applications/EXO.app"; do
-    if [[ -d "$app_path" ]]; then
-        if [[ "$APP_FOUND" == false ]]; then
-            echo ""
-            APP_FOUND=true
-        fi
-        echo_warn "EXO.app found at: $app_path"
-        echo_warn "You may want to move it to Trash manually."
+  if [[ -d $app_path ]]; then
+    if [[ $APP_FOUND == false ]]; then
+      echo ""
+      APP_FOUND=true
     fi
+    echo_warn "EXO.app found at: $app_path"
+    echo_warn "You may want to move it to Trash manually."
+  fi
 done
 
 echo ""
@@ -151,4 +151,3 @@ echo ""
 echo "Manual step required:"
 echo "  Remove EXO from Login Items in System Settings → General → Login Items"
 echo ""
-

bench/exo_bench.py

@@ -5,10 +5,13 @@ from __future__ import annotations
 import argparse
 import contextlib
 import http.client
+import itertools
 import json
 import os
+import sys
 import time
 from collections.abc import Callable
+from pathlib import Path
 from statistics import mean
 from typing import Any
 from urllib.parse import urlencode
@@ -16,6 +19,84 @@ from urllib.parse import urlencode
 from loguru import logger
 from transformers import AutoTokenizer
 
+# Monkey-patch for transformers 5.x compatibility
+# Kimi's tokenization_kimi.py imports bytes_to_unicode from the old location
+# which was moved in transformers 5.0.0rc2
+try:
+    import transformers.models.gpt2.tokenization_gpt2 as gpt2_tokenization
+    from transformers.convert_slow_tokenizer import bytes_to_unicode
+
+    if not hasattr(gpt2_tokenization, "bytes_to_unicode"):
+        gpt2_tokenization.bytes_to_unicode = bytes_to_unicode  # type: ignore[attr-defined]
+except ImportError:
+    pass  # transformers < 5.0 or bytes_to_unicode not available
+
+
+def load_tokenizer_for_bench(model_id: str) -> Any:
+    """
+    Load tokenizer for benchmarking, with special handling for Kimi models.
+
+    Kimi uses a custom TikTokenTokenizer that transformers 5.x can't load via AutoTokenizer.
+    This function replicates the logic from utils_mlx.py for bench compatibility.
+    """
+    model_id_lower = model_id.lower()
+
+    if "kimi-k2" in model_id_lower:
+        import importlib.util
+        import types
+
+        from huggingface_hub import snapshot_download
+
+        # Download/get the model path
+        model_path = Path(
+            snapshot_download(
+                model_id,
+                allow_patterns=["*.json", "*.py", "*.tiktoken"],
+            )
+        )
+
+        sys.path.insert(0, str(model_path))
+
+        # Load tool_declaration_ts first (tokenization_kimi imports it with relative import)
+        tool_decl_path = model_path / "tool_declaration_ts.py"
+        if tool_decl_path.exists():
+            spec = importlib.util.spec_from_file_location(
+                "tool_declaration_ts", tool_decl_path
+            )
+            if spec and spec.loader:
+                tool_decl_module = importlib.util.module_from_spec(spec)
+                sys.modules["tool_declaration_ts"] = tool_decl_module
+                spec.loader.exec_module(tool_decl_module)
+
+        # Load tokenization_kimi with patched source (convert relative to absolute import)
+        tok_path = model_path / "tokenization_kimi.py"
+        source = tok_path.read_text()
+        source = source.replace("from .tool_declaration_ts", "from tool_declaration_ts")
+        spec = importlib.util.spec_from_file_location("tokenization_kimi", tok_path)
+        if spec:
+            tok_module = types.ModuleType("tokenization_kimi")
+            tok_module.__file__ = str(tok_path)
+            sys.modules["tokenization_kimi"] = tok_module
+            exec(compile(source, tok_path, "exec"), tok_module.__dict__)  # noqa: S102
+            TikTokenTokenizer = tok_module.TikTokenTokenizer  # noqa: N806
+        else:
+            from tokenization_kimi import TikTokenTokenizer  # type: ignore[import-not-found]  # noqa: I001
+
+        hf_tokenizer: Any = TikTokenTokenizer.from_pretrained(model_path)
+
+        # Patch encode to use internal tiktoken model directly
+        # transformers 5.x has a bug in the encode->pad path for slow tokenizers
+        def _patched_encode(text: str, **kwargs: object) -> list[int]:
+            # Pass allowed_special="all" to handle special tokens like <|im_user|>
+            return list(hf_tokenizer.model.encode(text, allowed_special="all"))
+
+        hf_tokenizer.encode = _patched_encode
+
+        return hf_tokenizer
+
+    # Default: use AutoTokenizer
+    return AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+
 
 class ExoHttpError(RuntimeError):
     def __init__(self, status: int, reason: str, body_preview: str):
@@ -24,7 +105,7 @@ class ExoHttpError(RuntimeError):
 
 
 class ExoClient:
-    def __init__(self, host: str, port: int, timeout_s: float = 600.0):
+    def __init__(self, host: str, port: int, timeout_s: float = 7200.0):
         self.host = host
         self.port = port
         self.timeout_s = timeout_s
@@ -180,14 +261,7 @@ def parse_int_list(values: list[str]) -> list[int]:
             part = part.strip()
             if part:
                 items.append(int(part))
-
-    seen: set[int] = set()
-    out: list[int] = []
-    for x in items:
-        if x not in seen:
-            out.append(x)
-            seen.add(x)
-    return out
+    return items
 
 
 def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]:
@@ -240,7 +314,11 @@ def run_one_completion(
 
     stats = out.get("generation_stats")
 
-    preview = (out.get("choices") or [{}])[0]["message"]["content"][:200]
+    # Extract preview, handling None content (common for thinking models)
+    choices = out.get("choices") or [{}]
+    message = choices[0].get("message", {}) if choices else {}
+    content = message.get("content") or ""
+    preview = content[:200] if content else ""
 
     return {
         "elapsed_s": elapsed,
@@ -277,12 +355,29 @@ class PromptSizer:
                 f"Target ({target}) is smaller than template overhead ({self.base_tokens})."
             )
 
-        content = ""
-        tok = self.count_fn(content)
+        # Estimate tokens per atom using a sample
+        sample_count = 100
+        sample_content = self.atom * sample_count
+        sample_tokens = self.count_fn(sample_content) - self.base_tokens
+        tokens_per_atom = sample_tokens / sample_count
 
-        while tok < target:
-            content += self.atom
-            tok = self.count_fn(content)
+        # Estimate starting point
+        needed_tokens = target - self.base_tokens
+        estimated_atoms = int(needed_tokens / tokens_per_atom)
+
+        # Binary search to find exact atom count
+        low, high = 0, estimated_atoms * 2 + 100
+        while low < high:
+            mid = (low + high) // 2
+            tok = self.count_fn(self.atom * mid)
+            if tok < target:
+                low = mid + 1
+            else:
+                high = mid
+
+        content = self.atom * low
+        tok = self.count_fn(content)
+        logger.info(f"{tok=}")
 
         if tok != target:
             raise RuntimeError(
@@ -348,7 +443,7 @@ def main() -> int:
         help="Warmup runs per placement (uses first pp/tg).",
     )
     ap.add_argument(
-        "--timeout", type=float, default=600.0, help="HTTP timeout (seconds)."
+        "--timeout", type=float, default=7200.0, help="HTTP timeout (seconds)."
     )
     ap.add_argument(
         "--json-out",
@@ -358,6 +453,11 @@ def main() -> int:
     ap.add_argument(
         "--dry-run", action="store_true", help="List selected placements and exit."
     )
+    ap.add_argument(
+        "--all-combinations",
+        action="store_true",
+        help="Force all pp×tg combinations (cartesian product) even when lists have equal length.",
+    )
     args = ap.parse_args()
 
     pp_list = parse_int_list(args.pp)
@@ -369,6 +469,15 @@ def main() -> int:
         logger.error("--repeat must be >= 1")
         return 2
 
+    # Log pairing mode
+    use_combinations = args.all_combinations or len(pp_list) != len(tg_list)
+    if use_combinations:
+        logger.info(
+            f"pp/tg mode: combinations (product) - {len(pp_list) * len(tg_list)} pairs"
+        )
+    else:
+        logger.info(f"pp/tg mode: tandem (zip) - {len(pp_list)} pairs")
+
     client = ExoClient(args.host, args.port, timeout_s=args.timeout)
     short_id, full_model_id = resolve_model_short_id(client, args.model)
 
@@ -377,10 +486,7 @@ def main() -> int:
     )
     previews = previews_resp.get("previews") or []
 
-    tokenizer = AutoTokenizer.from_pretrained(
-        full_model_id,
-        trust_remote_code=True,
-    )
+    tokenizer = load_tokenizer_for_bench(full_model_id)
     if tokenizer is None:
         raise RuntimeError("[exo-bench] tokenizer load failed")
 
@@ -486,60 +592,55 @@ def main() -> int:
                 )
                 logger.debug(f"  warmup {i + 1}/{args.warmup} done")
 
-            for pp in pp_list:
-                # if (
-                #     pp * n_nodes > 2048
-                #     and "ring" in instance_meta.lower()
-                #     and "tensor" in sharding.lower()
-                # ):
-                #     model_card = MODEL_CARDS[short_id]
-                #     if model_card.metadata.storage_size > Memory.from_gb(10):
-                #         logger.info(
-                #             f"Skipping tensor ring as this is too slow for model of size {model_card.metadata.storage_size} on {n_nodes=}"
-                #         )
-                #         continue
-                for tg in tg_list:
-                    runs: list[dict[str, Any]] = []
-                    for r in range(args.repeat):
-                        time.sleep(3)
-                        try:
-                            row, actual_pp_tokens = run_one_completion(
-                                client, full_model_id, pp, tg, prompt_sizer
-                            )
-                        except Exception as e:
-                            logger.error(e)
-                            continue
-                        row.update(
-                            {
-                                "model_short_id": short_id,
-                                "model_id": full_model_id,
-                                "placement_sharding": sharding,
-                                "placement_instance_meta": instance_meta,
-                                "placement_nodes": n_nodes,
-                                "instance_id": instance_id,
-                                "pp_tokens": actual_pp_tokens,
-                                "tg": tg,
-                                "repeat_index": r,
-                            }
-                        )
-                        runs.append(row)
-                        all_rows.append(row)
+            # If pp and tg lists have same length, run in tandem (zip)
+            # Otherwise (or if --all-combinations), run all combinations (cartesian product)
+            if use_combinations:
+                pp_tg_pairs = list(itertools.product(pp_list, tg_list))
+            else:
+                pp_tg_pairs = list(zip(pp_list, tg_list, strict=True))
 
-                    if runs:
-                        prompt_tps = mean(x["stats"]["prompt_tps"] for x in runs)
-                        gen_tps = mean(x["stats"]["generation_tps"] for x in runs)
-                        ptok = mean(x["stats"]["prompt_tokens"] for x in runs)
-                        gtok = mean(x["stats"]["generation_tokens"] for x in runs)
-                        peak = mean(
-                            x["stats"]["peak_memory_usage"]["inBytes"] for x in runs
+            for pp, tg in pp_tg_pairs:
+                runs: list[dict[str, Any]] = []
+                for r in range(args.repeat):
+                    time.sleep(3)
+                    try:
+                        row, actual_pp_tokens = run_one_completion(
+                            client, full_model_id, pp, tg, prompt_sizer
                         )
+                    except Exception as e:
+                        logger.error(e)
+                        continue
+                    row.update(
+                        {
+                            "model_short_id": short_id,
+                            "model_id": full_model_id,
+                            "placement_sharding": sharding,
+                            "placement_instance_meta": instance_meta,
+                            "placement_nodes": n_nodes,
+                            "instance_id": instance_id,
+                            "pp_tokens": actual_pp_tokens,
+                            "tg": tg,
+                            "repeat_index": r,
+                        }
+                    )
+                    runs.append(row)
+                    all_rows.append(row)
 
-                        logger.info(
-                            f"prompt_tps={prompt_tps:.2f} gen_tps={gen_tps:.2f}    "
-                            f"prompt_tokens={ptok} gen_tokens={gtok}    "
-                            f"peak_memory={format_peak_memory(peak)}\n"
-                        )
-                    time.sleep(2)
+                if runs:
+                    prompt_tps = mean(x["stats"]["prompt_tps"] for x in runs)
+                    gen_tps = mean(x["stats"]["generation_tps"] for x in runs)
+                    ptok = mean(x["stats"]["prompt_tokens"] for x in runs)
+                    gtok = mean(x["stats"]["generation_tokens"] for x in runs)
+                    peak = mean(
+                        x["stats"]["peak_memory_usage"]["inBytes"] for x in runs
+                    )
+
+                    logger.info(
+                        f"prompt_tps={prompt_tps:.2f} gen_tps={gen_tps:.2f}    "
+                        f"prompt_tokens={ptok} gen_tokens={gtok}    "
+                        f"peak_memory={format_peak_memory(peak)}\n"
+                    )
+                time.sleep(2)
         finally:
             try:
                 client.request_json("DELETE", f"/instance/{instance_id}")

dashboard/parts.nix

@@ -3,6 +3,61 @@
   perSystem =
     { pkgs, lib, ... }:
     let
+      # Filter source to ONLY include package.json and package-lock.json
+      # This ensures prettier-svelte only rebuilds when lockfiles change
+      dashboardLockfileSrc = lib.cleanSourceWith {
+        src = inputs.self;
+        filter =
+          path: type:
+          let
+            baseName = builtins.baseNameOf path;
+            isDashboardDir = baseName == "dashboard" && type == "directory";
+            isPackageFile =
+              (lib.hasInfix "/dashboard/" path || lib.hasSuffix "/dashboard" (builtins.dirOf path))
+              && (baseName == "package.json" || baseName == "package-lock.json");
+          in
+          isDashboardDir || isPackageFile;
+      };
+
+      # Stub source with lockfiles and minimal files for build to succeed
+      # This allows prettier-svelte to avoid rebuilding when dashboard source changes
+      dashboardStubSrc = pkgs.runCommand "dashboard-stub-src" { } ''
+        mkdir -p $out
+        cp ${dashboardLockfileSrc}/dashboard/package.json $out/
+        cp ${dashboardLockfileSrc}/dashboard/package-lock.json $out/
+        # Minimal files so vite build succeeds (produces empty output)
+        echo '<!DOCTYPE html><html><head></head><body></body></html>' > $out/index.html
+        mkdir -p $out/src
+        touch $out/src/app.html
+      '';
+
+      # Deps-only build using stub source (for prettier-svelte)
+      # Only rebuilds when package.json or package-lock.json change
+      dashboardDeps = inputs.dream2nix.lib.evalModules {
+        packageSets.nixpkgs = pkgs;
+        modules = [
+          ./dashboard.nix
+          {
+            paths.projectRoot = inputs.self;
+            paths.projectRootFile = "flake.nix";
+            paths.package = inputs.self + "/dashboard";
+          }
+          {
+            deps.dashboardSrc = lib.mkForce dashboardStubSrc;
+          }
+          # Override build phases to skip the actual build - just need node_modules
+          {
+            mkDerivation = {
+              buildPhase = lib.mkForce "true";
+              installPhase = lib.mkForce ''
+                runHook preInstall
+                runHook postInstall
+              '';
+            };
+          }
+        ];
+      };
+
       # Filter source to only include dashboard directory
       dashboardSrc = lib.cleanSourceWith {
         src = inputs.self;
@@ -42,11 +97,12 @@
       '';
 
       # Prettier with svelte plugin for treefmt
+      # Uses dashboardDeps instead of dashboardFull to avoid rebuilding on source changes
       packages.prettier-svelte = pkgs.writeShellScriptBin "prettier-svelte" ''
-        export NODE_PATH="${dashboardFull}/lib/node_modules/exo-dashboard/node_modules"
+        export NODE_PATH="${dashboardDeps}/lib/node_modules/exo-dashboard/node_modules"
         exec ${pkgs.nodejs}/bin/node \
-          ${dashboardFull}/lib/node_modules/exo-dashboard/node_modules/prettier/bin/prettier.cjs \
-          --plugin "${dashboardFull}/lib/node_modules/exo-dashboard/node_modules/prettier-plugin-svelte/plugin.js" \
+          ${dashboardDeps}/lib/node_modules/exo-dashboard/node_modules/prettier/bin/prettier.cjs \
+          --plugin "${dashboardDeps}/lib/node_modules/exo-dashboard/node_modules/prettier-plugin-svelte/plugin.js" \
           "$@"
       '';
     };

dashboard/src/lib/components/ChatForm.svelte

@@ -89,7 +89,10 @@
 
   const isImageModel = $derived(() => {
     if (!currentModel) return false;
-    return modelSupportsTextToImage(currentModel);
+    return (
+      modelSupportsTextToImage(currentModel) ||
+      modelSupportsImageEditing(currentModel)
+    );
   });
 
   const isEditOnlyWithoutImage = $derived(
@@ -646,6 +649,23 @@
             </svg>
             <span>EDIT</span>
           </span>
+        {:else if isEditOnlyWithoutImage}
+          <span class="inline-flex items-center gap-1.5">
+            <svg
+              class="w-3.5 h-3.5"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+              stroke-width="2"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z"
+              />
+            </svg>
+            <span>EDIT</span>
+          </span>
         {:else if isImageModel()}
           <span class="inline-flex items-center gap-1.5">
             <svg

dashboard/src/lib/components/ImageParamsPanel.svelte

@@ -110,6 +110,36 @@
     setImageGenerationParams({ negativePrompt: value || null });
   }
 
+  function handleNumImagesChange(event: Event) {
+    const input = event.target as HTMLInputElement;
+    const value = input.value.trim();
+    if (value === "") {
+      setImageGenerationParams({ numImages: 1 });
+    } else {
+      const num = parseInt(value, 10);
+      if (!isNaN(num) && num >= 1) {
+        setImageGenerationParams({ numImages: num });
+      }
+    }
+  }
+
+  function handleStreamChange(enabled: boolean) {
+    setImageGenerationParams({ stream: enabled });
+  }
+
+  function handlePartialImagesChange(event: Event) {
+    const input = event.target as HTMLInputElement;
+    const value = input.value.trim();
+    if (value === "") {
+      setImageGenerationParams({ partialImages: 0 });
+    } else {
+      const num = parseInt(value, 10);
+      if (!isNaN(num) && num >= 0) {
+        setImageGenerationParams({ partialImages: num });
+      }
+    }
+  }
+
   function clearSteps() {
     setImageGenerationParams({ numInferenceSteps: null });
   }
@@ -134,90 +164,92 @@
 <div class="border-b border-exo-medium-gray/30 px-3 py-2">
   <!-- Basic params row -->
   <div class="flex items-center gap-3 flex-wrap">
-    <!-- Size -->
-    <div class="flex items-center gap-1.5">
-      <span class="text-xs text-exo-light-gray uppercase tracking-wider"
-        >SIZE:</span
-      >
-      <div class="relative">
-        <button
-          bind:this={sizeButtonRef}
-          type="button"
-          onclick={() => (isSizeDropdownOpen = !isSizeDropdownOpen)}
-          class="bg-exo-medium-gray/50 border border-exo-yellow/30 rounded pl-2 pr-6 py-1 text-xs font-mono text-exo-yellow cursor-pointer transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70 {isSizeDropdownOpen
-            ? 'border-exo-yellow/70'
-            : ''}"
+    <!-- Size (hidden in edit mode - output size comes from input image) -->
+    {#if !isEditMode}
+      <div class="flex items-center gap-1.5">
+        <span class="text-xs text-exo-light-gray uppercase tracking-wider"
+          >SIZE:</span
         >
-          {params.size}
-        </button>
-        <div
-          class="absolute right-1.5 top-1/2 -translate-y-1/2 pointer-events-none transition-transform duration-200 {isSizeDropdownOpen
-            ? 'rotate-180'
-            : ''}"
-        >
-          <svg
-            class="w-3 h-3 text-exo-yellow/60"
-            fill="none"
-            viewBox="0 0 24 24"
-            stroke="currentColor"
+        <div class="relative">
+          <button
+            bind:this={sizeButtonRef}
+            type="button"
+            onclick={() => (isSizeDropdownOpen = !isSizeDropdownOpen)}
+            class="bg-exo-medium-gray/50 border border-exo-yellow/30 rounded pl-2 pr-6 py-1 text-xs font-mono text-exo-yellow cursor-pointer transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70 {isSizeDropdownOpen
+              ? 'border-exo-yellow/70'
+              : ''}"
           >
-            <path
-              stroke-linecap="round"
-              stroke-linejoin="round"
-              stroke-width="2"
-              d="M19 9l-7 7-7-7"
-            />
-          </svg>
-        </div>
-      </div>
-
-      {#if isSizeDropdownOpen}
-        <!-- Backdrop to close dropdown -->
-        <button
-          type="button"
-          class="fixed inset-0 z-[9998] cursor-default"
-          onclick={() => (isSizeDropdownOpen = false)}
-          aria-label="Close dropdown"
-        ></button>
-
-        <!-- Dropdown Panel - fixed positioning to escape overflow:hidden -->
-        <div
-          class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto min-w-max"
-          style="bottom: calc(100vh - {sizeDropdownPosition()
-            .top}px + 4px); left: {sizeDropdownPosition().left}px;"
-        >
-          <div class="py-1">
-            {#each sizeOptions as size}
-              <button
-                type="button"
-                onclick={() => selectSize(size)}
-                class="w-full px-3 py-1.5 text-left text-xs font-mono tracking-wide transition-colors duration-100 flex items-center gap-2 {params.size ===
-                size
-                  ? 'bg-transparent text-exo-yellow'
-                  : 'text-exo-light-gray hover:text-exo-yellow'}"
-              >
-                {#if params.size === size}
-                  <svg
-                    class="w-3 h-3 flex-shrink-0"
-                    fill="currentColor"
-                    viewBox="0 0 20 20"
-                  >
-                    <path
-                      fill-rule="evenodd"
-                      d="M16.707 5.293a1 1 0 010 1.414l-8 8a1 1 0 01-1.414 0l-4-4a1 1 0 011.414-1.414L8 12.586l7.293-7.293a1 1 0 011.414 0z"
-                      clip-rule="evenodd"
-                    />
-                  </svg>
-                {:else}
-                  <span class="w-3"></span>
-                {/if}
-                <span>{size}</span>
-              </button>
-            {/each}
+            {params.size}
+          </button>
+          <div
+            class="absolute right-1.5 top-1/2 -translate-y-1/2 pointer-events-none transition-transform duration-200 {isSizeDropdownOpen
+              ? 'rotate-180'
+              : ''}"
+          >
+            <svg
+              class="w-3 h-3 text-exo-yellow/60"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                stroke-width="2"
+                d="M19 9l-7 7-7-7"
+              />
+            </svg>
           </div>
         </div>
-      {/if}
-    </div>
+
+        {#if isSizeDropdownOpen}
+          <!-- Backdrop to close dropdown -->
+          <button
+            type="button"
+            class="fixed inset-0 z-[9998] cursor-default"
+            onclick={() => (isSizeDropdownOpen = false)}
+            aria-label="Close dropdown"
+          ></button>
+
+          <!-- Dropdown Panel - fixed positioning to escape overflow:hidden -->
+          <div
+            class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto min-w-max"
+            style="bottom: calc(100vh - {sizeDropdownPosition()
+              .top}px + 4px); left: {sizeDropdownPosition().left}px;"
+          >
+            <div class="py-1">
+              {#each sizeOptions as size}
+                <button
+                  type="button"
+                  onclick={() => selectSize(size)}
+                  class="w-full px-3 py-1.5 text-left text-xs font-mono tracking-wide transition-colors duration-100 flex items-center gap-2 {params.size ===
+                  size
+                    ? 'bg-transparent text-exo-yellow'
+                    : 'text-exo-light-gray hover:text-exo-yellow'}"
+                >
+                  {#if params.size === size}
+                    <svg
+                      class="w-3 h-3 flex-shrink-0"
+                      fill="currentColor"
+                      viewBox="0 0 20 20"
+                    >
+                      <path
+                        fill-rule="evenodd"
+                        d="M16.707 5.293a1 1 0 010 1.414l-8 8a1 1 0 01-1.414 0l-4-4a1 1 0 011.414-1.414L8 12.586l7.293-7.293a1 1 0 011.414 0z"
+                        clip-rule="evenodd"
+                      />
+                    </svg>
+                  {:else}
+                    <span class="w-3"></span>
+                  {/if}
+                  <span>{size}</span>
+                </button>
+              {/each}
+            </div>
+          </div>
+        {/if}
+      </div>
+    {/if}
 
     <!-- Quality -->
     <div class="flex items-center gap-1.5">
@@ -325,6 +357,59 @@
       </div>
     </div>
 
+    <!-- Number of Images (not in edit mode) -->
+    {#if !isEditMode}
+      <div class="flex items-center gap-1.5">
+        <span class="text-xs text-exo-light-gray uppercase tracking-wider"
+          >IMAGES:</span
+        >
+        <input
+          type="number"
+          min="1"
+          value={params.numImages}
+          oninput={handleNumImagesChange}
+          class="w-12 bg-exo-medium-gray/50 border border-exo-yellow/30 rounded px-2 py-1 text-xs font-mono text-exo-yellow text-center transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70"
+        />
+      </div>
+    {/if}
+
+    <!-- Stream toggle -->
+    <div class="flex items-center gap-1.5">
+      <span class="text-xs text-exo-light-gray uppercase tracking-wider"
+        >STREAM:</span
+      >
+      <button
+        type="button"
+        onclick={() => handleStreamChange(!params.stream)}
+        class="w-8 h-4 rounded-full transition-all duration-200 cursor-pointer relative {params.stream
+          ? 'bg-exo-yellow'
+          : 'bg-exo-medium-gray/50 border border-exo-yellow/30'}"
+        title={params.stream ? "Streaming enabled" : "Streaming disabled"}
+      >
+        <div
+          class="absolute top-0.5 w-3 h-3 rounded-full transition-all duration-200 {params.stream
+            ? 'right-0.5 bg-exo-black'
+            : 'left-0.5 bg-exo-light-gray'}"
+        ></div>
+      </button>
+    </div>
+
+    <!-- Partial Images (only when streaming) -->
+    {#if params.stream}
+      <div class="flex items-center gap-1.5">
+        <span class="text-xs text-exo-light-gray uppercase tracking-wider"
+          >PARTIALS:</span
+        >
+        <input
+          type="number"
+          min="0"
+          value={params.partialImages}
+          oninput={handlePartialImagesChange}
+          class="w-12 bg-exo-medium-gray/50 border border-exo-yellow/30 rounded px-2 py-1 text-xs font-mono text-exo-yellow text-center transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70"
+        />
+      </div>
+    {/if}
+
     <!-- Input Fidelity (edit mode only) -->
     {#if isEditMode}
       <div class="flex items-center gap-1.5">

dashboard/src/lib/components/ModelCard.svelte

@@ -1003,16 +1003,5 @@
 </div>
 
 <style>
-  @keyframes pulse-slow {
-    0%,
-    100% {
-      opacity: 0.8;
-    }
-    50% {
-      opacity: 1;
-    }
-  }
-  .animate-pulse-slow {
-    animation: pulse-slow 1.5s ease-in-out infinite;
-  }
+  /* Styles removed - animations were causing GPU overhead */
 </style>

dashboard/src/routes/+page.svelte

@@ -19,6 +19,9 @@
     selectedPreviewModelId,
     isLoadingPreviews,
     selectPreviewModel,
+    togglePreviewNodeFilter,
+    clearPreviewNodeFilter,
+    previewNodeFilter,
     createConversation,
     setSelectedChatModel,
     selectedChatModel,
@@ -28,6 +31,8 @@
     toggleTopologyOnlyMode,
     chatSidebarVisible,
     toggleChatSidebarVisible,
+    thunderboltBridgeCycles,
+    nodeThunderboltBridge,
     type DownloadProgress,
     type PlacementPreview,
   } from "$lib/stores/app.svelte";
@@ -49,6 +54,41 @@
   const debugEnabled = $derived(debugMode());
   const topologyOnlyEnabled = $derived(topologyOnlyMode());
   const sidebarVisible = $derived(chatSidebarVisible());
+  const tbBridgeCycles = $derived(thunderboltBridgeCycles());
+  const tbBridgeData = $derived(nodeThunderboltBridge());
+  const nodeFilter = $derived(previewNodeFilter());
+
+  // Helper to get friendly node name from node ID
+  function getNodeName(nodeId: string): string {
+    const node = data?.nodes?.[nodeId];
+    return node?.friendly_name || nodeId.slice(0, 8) + "...";
+  }
+
+  // Helper to get the thunderbolt bridge service name from a cycle
+  function getTbBridgeServiceName(cycle: string[]): string {
+    // Try to find service name from any node in the cycle
+    for (const nodeId of cycle) {
+      const nodeData = tbBridgeData?.[nodeId];
+      if (nodeData?.serviceName) {
+        return nodeData.serviceName;
+      }
+    }
+    return "Thunderbolt Bridge"; // Fallback if no service name found
+  }
+
+  // Copy to clipboard state and function
+  let copiedCommand = $state(false);
+  async function copyToClipboard(text: string) {
+    try {
+      await navigator.clipboard.writeText(text);
+      copiedCommand = true;
+      setTimeout(() => {
+        copiedCommand = false;
+      }, 2000);
+    } catch (err) {
+      console.error("Failed to copy:", err);
+    }
+  }
 
   let mounted = $state(false);
 
@@ -90,6 +130,15 @@
       model.tasks.includes("ImageToImage")
     );
   }
+
+  // Helper to check if a model supports image editing
+  function modelSupportsImageEditing(modelId: string): boolean {
+    const model = models.find(
+      (m) => m.id === modelId || m.hugging_face_id === modelId,
+    );
+    if (!model?.tasks) return false;
+    return model.tasks.includes("ImageToImage");
+  }
   let selectedSharding = $state<"Pipeline" | "Tensor">("Pipeline");
   type InstanceMeta = "MlxRing" | "MlxIbv" | "MlxJaccl";
 
@@ -181,6 +230,9 @@
   // Preview card hover state for highlighting nodes in topology
   let hoveredPreviewNodes = $state<Set<string>>(new Set());
 
+  // Computed: Check if filter is active (from store)
+  const isFilterActive = $derived(() => nodeFilter.size > 0);
+
   // Helper to unwrap tagged instance for hover highlighting
   function unwrapInstanceNodes(instanceWrapped: unknown): Set<string> {
     if (!instanceWrapped || typeof instanceWrapped !== "object")
@@ -732,6 +784,8 @@
     instanceWrapped: unknown,
   ): {
     isDownloading: boolean;
+    isFailed: boolean;
+    errorMessage: string | null;
     progress: DownloadProgress | null;
     statusText: string;
     perNode: Array<{
@@ -743,6 +797,8 @@
     if (!downloadsData || Object.keys(downloadsData).length === 0) {
       return {
         isDownloading: false,
+        isFailed: false,
+        errorMessage: null,
         progress: null,
         statusText: "RUNNING",
         perNode: [],
@@ -754,6 +810,8 @@
     if (!instance || typeof instance !== "object") {
       return {
         isDownloading: false,
+        isFailed: false,
+        errorMessage: null,
         progress: null,
         statusText: "PREPARING",
         perNode: [],
@@ -809,6 +867,26 @@
           downloadKind
         ] as Record<string, unknown>;
 
+        // Handle DownloadFailed - return immediately with error info
+        if (downloadKind === "DownloadFailed") {
+          const downloadModelId = extractModelIdFromDownload(downloadPayload);
+          if (
+            instanceModelId &&
+            downloadModelId &&
+            downloadModelId === instanceModelId
+          ) {
+            return {
+              isDownloading: false,
+              isFailed: true,
+              errorMessage:
+                (downloadPayload.errorMessage as string) || "Download failed",
+              progress: null,
+              statusText: "FAILED",
+              perNode: [],
+            };
+          }
+        }
+
         if (downloadKind !== "DownloadOngoing") continue;
         if (!downloadPayload) continue;
 
@@ -844,6 +922,8 @@
       const statusInfo = deriveInstanceStatus(instanceWrapped);
       return {
         isDownloading: false,
+        isFailed: statusInfo.statusText === "FAILED",
+        errorMessage: null,
         progress: null,
         statusText: statusInfo.statusText,
         perNode: [],
@@ -856,6 +936,8 @@
 
     return {
       isDownloading: true,
+      isFailed: false,
+      errorMessage: null,
       progress: {
         totalBytes,
         downloadedBytes,
@@ -1451,7 +1533,7 @@
 
   // Get ALL filtered previews based on current settings (matching minimum nodes)
   // Note: previewsData already contains previews for the selected model (fetched via API)
-  // We filter by sharding/instance type and min nodes, returning ALL eligible previews
+  // Backend handles node_ids filtering, we filter by sharding/instance type and min nodes
   const filteredPreviews = $derived(() => {
     if (!selectedModelId || previewsData.length === 0) return [];
 
@@ -1584,7 +1666,86 @@
           <TopologyGraph
             class="w-full h-full"
             highlightedNodes={highlightedNodes()}
+            filteredNodes={nodeFilter}
+            onNodeClick={togglePreviewNodeFilter}
           />
+
+          <!-- Thunderbolt Bridge Cycle Warning -->
+          {#if tbBridgeCycles.length > 0}
+            {@const cycle = tbBridgeCycles[0]}
+            {@const serviceName = getTbBridgeServiceName(cycle)}
+            {@const disableCmd = `sudo networksetup -setnetworkserviceenabled "${serviceName}" off`}
+            <div class="absolute top-4 left-4 group" role="alert">
+              <div
+                class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+              >
+                <svg
+                  class="w-5 h-5 text-yellow-400 flex-shrink-0"
+                  fill="none"
+                  viewBox="0 0 24 24"
+                  stroke="currentColor"
+                  stroke-width="2"
+                >
+                  <path
+                    stroke-linecap="round"
+                    stroke-linejoin="round"
+                    d="M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z"
+                  />
+                </svg>
+                <span class="text-sm font-mono text-yellow-200">
+                  THUNDERBOLT BRIDGE CYCLE DETECTED
+                </span>
+              </div>
+
+              <!-- Tooltip on hover -->
+              <div
+                class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
+              >
+                <p class="text-xs text-white/80 mb-2">
+                  A network routing cycle was detected between nodes connected
+                  via Thunderbolt Bridge. This can cause connectivity issues.
+                </p>
+                <p class="text-xs text-white/60 mb-2">
+                  <span class="text-yellow-300">Affected nodes:</span>
+                  {cycle.map(getNodeName).join(" → ")}
+                </p>
+                <p class="text-xs text-white/60 mb-1">
+                  <span class="text-yellow-300">To fix:</span> Disable the Thunderbolt
+                  Bridge on one of the affected nodes:
+                </p>
+                <button
+                  type="button"
+                  onclick={() => copyToClipboard(disableCmd)}
+                  class="w-full flex items-center gap-2 text-[10px] font-mono bg-exo-black/60 px-2 py-1.5 rounded text-exo-yellow break-all text-left hover:bg-exo-black/80 transition-colors cursor-pointer group/copy"
+                  title="Click to copy"
+                >
+                  <span class="flex-1">{disableCmd}</span>
+                  <svg
+                    class="w-3.5 h-3.5 flex-shrink-0 text-white/40 group-hover/copy:text-exo-yellow transition-colors"
+                    fill="none"
+                    viewBox="0 0 24 24"
+                    stroke="currentColor"
+                    stroke-width="2"
+                  >
+                    {#if copiedCommand}
+                      <path
+                        stroke-linecap="round"
+                        stroke-linejoin="round"
+                        d="M5 13l4 4L19 7"
+                      />
+                    {:else}
+                      <path
+                        stroke-linecap="round"
+                        stroke-linejoin="round"
+                        d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z"
+                      />
+                    {/if}
+                  </svg>
+                </button>
+              </div>
+            </div>
+          {/if}
+
           <!-- Exit topology-only mode button -->
           <button
             type="button"
@@ -1624,7 +1785,111 @@
             <TopologyGraph
               class="w-full h-full"
               highlightedNodes={highlightedNodes()}
+              filteredNodes={nodeFilter}
+              onNodeClick={togglePreviewNodeFilter}
             />
+
+            <!-- Thunderbolt Bridge Cycle Warning -->
+            {#if tbBridgeCycles.length > 0}
+              {@const cycle = tbBridgeCycles[0]}
+              {@const serviceName = getTbBridgeServiceName(cycle)}
+              {@const disableCmd = `sudo networksetup -setnetworkserviceenabled "${serviceName}" off`}
+              <div class="absolute top-4 left-4 group" role="alert">
+                <div
+                  class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+                >
+                  <svg
+                    class="w-5 h-5 text-yellow-400 flex-shrink-0"
+                    fill="none"
+                    viewBox="0 0 24 24"
+                    stroke="currentColor"
+                    stroke-width="2"
+                  >
+                    <path
+                      stroke-linecap="round"
+                      stroke-linejoin="round"
+                      d="M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z"
+                    />
+                  </svg>
+                  <span class="text-sm font-mono text-yellow-200">
+                    THUNDERBOLT BRIDGE CYCLE DETECTED
+                  </span>
+                </div>
+
+                <!-- Tooltip on hover -->
+                <div
+                  class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
+                >
+                  <p class="text-xs text-white/80 mb-2">
+                    A network routing cycle was detected between nodes connected
+                    via Thunderbolt Bridge. This can cause connectivity issues.
+                  </p>
+                  <p class="text-xs text-white/60 mb-2">
+                    <span class="text-yellow-300">Affected nodes:</span>
+                    {cycle.map(getNodeName).join(" → ")}
+                  </p>
+                  <p class="text-xs text-white/60 mb-1">
+                    <span class="text-yellow-300">To fix:</span> Disable the Thunderbolt
+                    Bridge on one of the affected nodes:
+                  </p>
+                  <button
+                    type="button"
+                    onclick={() => copyToClipboard(disableCmd)}
+                    class="w-full flex items-center gap-2 text-[10px] font-mono bg-exo-black/60 px-2 py-1.5 rounded text-exo-yellow break-all text-left hover:bg-exo-black/80 transition-colors cursor-pointer group/copy"
+                    title="Click to copy"
+                  >
+                    <span class="flex-1">{disableCmd}</span>
+                    <svg
+                      class="w-3.5 h-3.5 flex-shrink-0 text-white/40 group-hover/copy:text-exo-yellow transition-colors"
+                      fill="none"
+                      viewBox="0 0 24 24"
+                      stroke="currentColor"
+                      stroke-width="2"
+                    >
+                      {#if copiedCommand}
+                        <path
+                          stroke-linecap="round"
+                          stroke-linejoin="round"
+                          d="M5 13l4 4L19 7"
+                        />
+                      {:else}
+                        <path
+                          stroke-linecap="round"
+                          stroke-linejoin="round"
+                          d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z"
+                        />
+                      {/if}
+                    </svg>
+                  </button>
+                </div>
+              </div>
+            {/if}
+
+            <!-- Node Filter Indicator (top-right corner) -->
+            {#if isFilterActive()}
+              <button
+                onclick={clearPreviewNodeFilter}
+                class="absolute top-2 right-2 flex items-center gap-1.5 px-2 py-1 bg-exo-dark-gray/80 border border-exo-yellow/40 rounded text-exo-yellow hover:border-exo-yellow/60 transition-colors cursor-pointer backdrop-blur-sm"
+                title="Clear filter"
+              >
+                <span class="text-[10px] font-mono tracking-wider">
+                  FILTER: {nodeFilter.size}
+                </span>
+                <svg
+                  class="w-3 h-3"
+                  fill="none"
+                  viewBox="0 0 24 24"
+                  stroke="currentColor"
+                  stroke-width="2"
+                >
+                  <path
+                    stroke-linecap="round"
+                    stroke-linejoin="round"
+                    d="M6 18L18 6M6 6l12 12"
+                  />
+                </svg>
+              </button>
+            {/if}
           </div>
 
           <!-- Chat Input - Below topology -->
@@ -2061,6 +2326,13 @@
                           >
                             {downloadInfo.statusText}
                           </div>
+                          {#if downloadInfo.isFailed && downloadInfo.errorMessage}
+                            <div
+                              class="text-xs text-red-400/80 font-mono mt-1 break-words"
+                            >
+                              {downloadInfo.errorMessage}
+                            </div>
+                          {/if}
                         {/if}
                       </div>
                     </div>
@@ -2106,6 +2378,9 @@
                     {@const isImageModel = modelSupportsImageGeneration(
                       foundModel.id,
                     )}
+                    {@const isImageEditModel = modelSupportsImageEditing(
+                      foundModel.id,
+                    )}
                     <span
                       class="flex items-center justify-between gap-2 w-full pr-4"
                     >
@@ -2132,6 +2407,22 @@
                             <polyline points="21 15 16 10 5 21" />
                           </svg>
                         {/if}
+                        {#if isImageEditModel}
+                          <svg
+                            class="w-4 h-4 flex-shrink-0 text-exo-yellow"
+                            fill="none"
+                            viewBox="0 0 24 24"
+                            stroke="currentColor"
+                            stroke-width="2"
+                          >
+                            <path
+                              d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7"
+                            />
+                            <path
+                              d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z"
+                            />
+                          </svg>
+                        {/if}
                         <span class="truncate"
                           >{foundModel.name || foundModel.id}</span
                         >
@@ -2204,6 +2495,9 @@
                       {@const isImageModel = modelSupportsImageGeneration(
                         model.id,
                       )}
+                      {@const isImageEditModel = modelSupportsImageEditing(
+                        model.id,
+                      )}
                       <button
                         type="button"
                         onclick={() => {
@@ -2244,6 +2538,23 @@
                               <polyline points="21 15 16 10 5 21" />
                             </svg>
                           {/if}
+                          {#if isImageEditModel}
+                            <svg
+                              class="w-4 h-4 flex-shrink-0 text-exo-yellow"
+                              fill="none"
+                              viewBox="0 0 24 24"
+                              stroke="currentColor"
+                              stroke-width="2"
+                              aria-label="Image editing model"
+                            >
+                              <path
+                                d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7"
+                              />
+                              <path
+                                d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z"
+                              />
+                            </svg>
+                          {/if}
                           <span class="truncate">{model.name || model.id}</span>
                         </span>
                         <span
@@ -2564,7 +2875,36 @@
               <div
                 class="relative aspect-square bg-exo-dark-gray rounded-lg overflow-hidden"
               >
-                <TopologyGraph highlightedNodes={highlightedNodes()} />
+                <TopologyGraph
+                  highlightedNodes={highlightedNodes()}
+                  filteredNodes={nodeFilter}
+                  onNodeClick={togglePreviewNodeFilter}
+                />
+
+                <!-- Thunderbolt Bridge Cycle Warning (compact) -->
+                {#if tbBridgeCycles.length > 0}
+                  <div
+                    class="absolute top-2 left-2 flex items-center gap-1.5 px-2 py-1 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm"
+                    title="Thunderbolt Bridge cycle detected - click to view details"
+                  >
+                    <svg
+                      class="w-3.5 h-3.5 text-yellow-400"
+                      fill="none"
+                      viewBox="0 0 24 24"
+                      stroke="currentColor"
+                      stroke-width="2"
+                    >
+                      <path
+                        stroke-linecap="round"
+                        stroke-linejoin="round"
+                        d="M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z"
+                      />
+                    </svg>
+                    <span class="text-[10px] font-mono text-yellow-200"
+                      >TB CYCLE</span
+                    >
+                  </div>
+                {/if}
               </div>
             </button>
 
@@ -2993,6 +3333,13 @@
                             >
                               {downloadInfo.statusText}
                             </div>
+                            {#if downloadInfo.isFailed && downloadInfo.errorMessage}
+                              <div
+                                class="text-xs text-red-400/80 font-mono mt-1 break-words"
+                              >
+                                {downloadInfo.errorMessage}
+                              </div>
+                            {/if}
                           {/if}
                         </div>
                       </div>

dashboard/src/routes/downloads/+page.svelte

@@ -6,6 +6,8 @@
     type DownloadProgress,
     refreshState,
     lastUpdate as lastUpdateStore,
+    startDownload,
+    deleteDownload,
   } from "$lib/stores/app.svelte";
   import HeaderNav from "$lib/components/HeaderNav.svelte";
 
@@ -28,6 +30,7 @@
     etaMs: number;
     status: "completed" | "downloading";
     files: FileProgress[];
+    shardMetadata?: Record<string, unknown>;
   };
 
   type NodeEntry = {
@@ -172,33 +175,6 @@
   }
 
   let downloadOverview = $state<NodeEntry[]>([]);
-  let models = $state<Array<{ id: string; storage_size_megabytes?: number }>>(
-    [],
-  );
-
-  async function fetchModels() {
-    try {
-      const response = await fetch("/models");
-      if (response.ok) {
-        const data = await response.json();
-        models = data.data || [];
-      }
-    } catch (error) {
-      console.error("Failed to fetch models:", error);
-    }
-  }
-
-  function getModelTotalBytes(
-    modelId: string,
-    downloadTotalBytes: number,
-  ): number {
-    if (downloadTotalBytes > 0) return downloadTotalBytes;
-    const model = models.find((m) => m.id === modelId);
-    if (model?.storage_size_megabytes) {
-      return model.storage_size_megabytes * 1024 * 1024;
-    }
-    return 0;
-  }
 
   $effect(() => {
     try {
@@ -296,6 +272,12 @@
             }
           }
 
+          // Extract shard_metadata for use with download actions
+          const shardMetadata = (downloadPayload.shard_metadata ??
+            downloadPayload.shardMetadata) as
+            | Record<string, unknown>
+            | undefined;
+
           const entry: ModelEntry = {
             modelId,
             prettyName,
@@ -312,6 +294,7 @@
                 ? "completed"
                 : "downloading",
             files,
+            shardMetadata,
           };
 
           const existing = modelMap.get(modelId);
@@ -373,7 +356,6 @@
   onMount(() => {
     // Ensure we fetch at least once when visiting downloads directly
     refreshState();
-    fetchModels();
   });
 </script>
 
@@ -482,7 +464,7 @@
                     {#if model.status !== "completed"}
                       <div class="text-[11px] text-exo-light-gray font-mono">
                         {formatBytes(model.downloadedBytes)} / {formatBytes(
-                          getModelTotalBytes(model.modelId, model.totalBytes),
+                          model.totalBytes,
                         )}
                       </div>
                     {/if}
@@ -497,6 +479,52 @@
                     >
                       {pct.toFixed(1)}%
                     </span>
+                    {#if model.status !== "completed" && model.shardMetadata}
+                      <button
+                        type="button"
+                        class="text-exo-light-gray hover:text-exo-yellow transition-colors"
+                        onclick={() =>
+                          startDownload(node.nodeId, model.shardMetadata!)}
+                        title="Start download"
+                      >
+                        <svg
+                          class="w-4 h-4"
+                          viewBox="0 0 20 20"
+                          fill="none"
+                          stroke="currentColor"
+                          stroke-width="2"
+                        >
+                          <path
+                            d="M10 3v10m0 0l-3-3m3 3l3-3M3 17h14"
+                            stroke-linecap="round"
+                            stroke-linejoin="round"
+                          ></path>
+                        </svg>
+                      </button>
+                    {/if}
+                    {#if model.status === "completed"}
+                      <button
+                        type="button"
+                        class="text-exo-light-gray hover:text-red-400 transition-colors"
+                        onclick={() =>
+                          deleteDownload(node.nodeId, model.modelId)}
+                        title="Delete download"
+                      >
+                        <svg
+                          class="w-4 h-4"
+                          viewBox="0 0 20 20"
+                          fill="none"
+                          stroke="currentColor"
+                          stroke-width="2"
+                        >
+                          <path
+                            d="M4 6h12M8 6V4h4v2m1 0v10a1 1 0 01-1 1H8a1 1 0 01-1-1V6h6"
+                            stroke-linecap="round"
+                            stroke-linejoin="round"
+                          ></path>
+                        </svg>
+                      </button>
+                    {/if}
                     <button
                       type="button"
                       class="text-exo-light-gray hover:text-exo-yellow transition-colors"

dashboard/src/routes/traces/+page.svelte

@@ -0,0 +1,190 @@
+<script lang="ts">
+  import { onMount } from "svelte";
+  import {
+    listTraces,
+    getTraceRawUrl,
+    type TraceListItem,
+  } from "$lib/stores/app.svelte";
+  import HeaderNav from "$lib/components/HeaderNav.svelte";
+
+  let traces = $state<TraceListItem[]>([]);
+  let loading = $state(true);
+  let error = $state<string | null>(null);
+
+  function formatBytes(bytes: number): string {
+    if (!bytes || bytes <= 0) return "0B";
+    const units = ["B", "KB", "MB", "GB"];
+    const i = Math.min(
+      Math.floor(Math.log(bytes) / Math.log(1024)),
+      units.length - 1,
+    );
+    const val = bytes / Math.pow(1024, i);
+    return `${val.toFixed(val >= 10 ? 0 : 1)}${units[i]}`;
+  }
+
+  function formatDate(isoString: string): string {
+    const date = new Date(isoString);
+    return date.toLocaleString();
+  }
+
+  async function downloadTrace(taskId: string) {
+    const response = await fetch(getTraceRawUrl(taskId));
+    const blob = await response.blob();
+    const url = URL.createObjectURL(blob);
+    const a = document.createElement("a");
+    a.href = url;
+    a.download = `trace_${taskId}.json`;
+    a.click();
+    URL.revokeObjectURL(url);
+  }
+
+  async function openInPerfetto(taskId: string) {
+    // Fetch trace data from our local API
+    const response = await fetch(getTraceRawUrl(taskId));
+    const traceData = await response.arrayBuffer();
+
+    // Open Perfetto UI
+    const perfettoWindow = window.open("https://ui.perfetto.dev");
+    if (!perfettoWindow) {
+      alert("Failed to open Perfetto. Please allow popups.");
+      return;
+    }
+
+    // Wait for Perfetto to be ready, then send trace via postMessage
+    const onMessage = (e: MessageEvent) => {
+      if (e.data === "PONG") {
+        window.removeEventListener("message", onMessage);
+        perfettoWindow.postMessage(
+          {
+            perfetto: {
+              buffer: traceData,
+              title: `Trace ${taskId}`,
+            },
+          },
+          "https://ui.perfetto.dev",
+        );
+      }
+    };
+    window.addEventListener("message", onMessage);
+
+    // Ping Perfetto until it responds
+    const pingInterval = setInterval(() => {
+      perfettoWindow.postMessage("PING", "https://ui.perfetto.dev");
+    }, 50);
+
+    // Clean up after 10 seconds
+    setTimeout(() => {
+      clearInterval(pingInterval);
+      window.removeEventListener("message", onMessage);
+    }, 10000);
+  }
+
+  async function refresh() {
+    loading = true;
+    error = null;
+    try {
+      const response = await listTraces();
+      traces = response.traces;
+    } catch (e) {
+      error = e instanceof Error ? e.message : "Failed to load traces";
+    } finally {
+      loading = false;
+    }
+  }
+
+  onMount(() => {
+    refresh();
+  });
+</script>
+
+<div class="min-h-screen bg-exo-dark-gray text-white">
+  <HeaderNav showHome={true} />
+  <div class="max-w-7xl mx-auto px-4 lg:px-8 py-6 space-y-6">
+    <div class="flex items-center justify-between gap-4 flex-wrap">
+      <div>
+        <h1
+          class="text-2xl font-mono tracking-[0.2em] uppercase text-exo-yellow"
+        >
+          Traces
+        </h1>
+      </div>
+      <div class="flex items-center gap-3">
+        <button
+          type="button"
+          class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"
+          onclick={refresh}
+          disabled={loading}
+        >
+          Refresh
+        </button>
+      </div>
+    </div>
+
+    {#if loading}
+      <div
+        class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-6 text-center text-exo-light-gray"
+      >
+        <div class="text-sm">Loading traces...</div>
+      </div>
+    {:else if error}
+      <div
+        class="rounded border border-red-500/30 bg-red-500/10 p-6 text-center text-red-400"
+      >
+        <div class="text-sm">{error}</div>
+      </div>
+    {:else if traces.length === 0}
+      <div
+        class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-6 text-center text-exo-light-gray space-y-2"
+      >
+        <div class="text-sm">No traces found.</div>
+        <div class="text-xs text-exo-light-gray/70">
+          Run exo with EXO_TRACING_ENABLED=1 to collect traces.
+        </div>
+      </div>
+    {:else}
+      <div class="space-y-3">
+        {#each traces as trace}
+          <div
+            class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 flex items-center justify-between gap-4"
+          >
+            <div class="min-w-0 flex-1">
+              <a
+                href="#/traces/{trace.taskId}"
+                class="text-sm font-mono text-white hover:text-exo-yellow transition-colors truncate block"
+              >
+                {trace.taskId}
+              </a>
+              <div class="text-xs text-exo-light-gray font-mono mt-1">
+                {formatDate(trace.createdAt)} &bull; {formatBytes(
+                  trace.fileSize,
+                )}
+              </div>
+            </div>
+            <div class="flex items-center gap-2 shrink-0">
+              <a
+                href="#/traces/{trace.taskId}"
+                class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"
+              >
+                View Stats
+              </a>
+              <button
+                type="button"
+                class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"
+                onclick={() => downloadTrace(trace.taskId)}
+              >
+                Download
+              </button>
+              <button
+                type="button"
+                class="text-xs font-mono text-exo-dark-gray bg-exo-yellow hover:bg-exo-yellow/90 transition-colors uppercase px-2 py-1 rounded font-semibold"
+                onclick={() => openInPerfetto(trace.taskId)}
+              >
+                View Trace
+              </button>
+            </div>
+          </div>
+        {/each}
+      </div>
+    {/if}
+  </div>
+</div>

dashboard/src/routes/traces/[taskId]/+page.svelte

@@ -0,0 +1,367 @@
+<script lang="ts">
+  import { page } from "$app/stores";
+  import { onMount } from "svelte";
+  import {
+    fetchTraceStats,
+    getTraceRawUrl,
+    type TraceStatsResponse,
+    type TraceCategoryStats,
+  } from "$lib/stores/app.svelte";
+  import HeaderNav from "$lib/components/HeaderNav.svelte";
+
+  const taskId = $derived($page.params.taskId);
+
+  let stats = $state<TraceStatsResponse | null>(null);
+  let loading = $state(true);
+  let error = $state<string | null>(null);
+
+  function formatDuration(us: number): string {
+    if (us < 1000) return `${us.toFixed(0)}us`;
+    if (us < 1_000_000) return `${(us / 1000).toFixed(2)}ms`;
+    return `${(us / 1_000_000).toFixed(2)}s`;
+  }
+
+  function formatPercentage(part: number, total: number): string {
+    if (total === 0) return "0.0%";
+    return `${((part / total) * 100).toFixed(1)}%`;
+  }
+
+  // Parse hierarchical categories like "sync/compute" into phases
+  type PhaseData = {
+    name: string;
+    subcategories: { name: string; stats: TraceCategoryStats }[];
+    totalUs: number; // From outer span (e.g., "sync" category)
+    stepCount: number; // Count of outer span events
+  };
+
+  function parsePhases(
+    byCategory: Record<string, TraceCategoryStats>,
+  ): PhaseData[] {
+    const phases = new Map<
+      string,
+      {
+        subcats: Map<string, TraceCategoryStats>;
+        outerStats: TraceCategoryStats | null;
+      }
+    >();
+
+    for (const [category, catStats] of Object.entries(byCategory)) {
+      if (category.includes("/")) {
+        const [phase, subcat] = category.split("/", 2);
+        if (!phases.has(phase)) {
+          phases.set(phase, { subcats: new Map(), outerStats: null });
+        }
+        phases.get(phase)!.subcats.set(subcat, catStats);
+      } else {
+        // Outer span - this IS the phase total
+        if (!phases.has(category)) {
+          phases.set(category, { subcats: new Map(), outerStats: null });
+        }
+        phases.get(category)!.outerStats = catStats;
+      }
+    }
+
+    return Array.from(phases.entries())
+      .filter(([_, data]) => data.outerStats !== null) // Only phases with outer spans
+      .map(([name, data]) => ({
+        name,
+        subcategories: Array.from(data.subcats.entries())
+          .map(([subName, subStats]) => ({ name: subName, stats: subStats }))
+          .sort((a, b) => b.stats.totalUs - a.stats.totalUs),
+        totalUs: data.outerStats!.totalUs, // Outer span total
+        stepCount: data.outerStats!.count, // Number of steps
+      }))
+      .sort((a, b) => b.totalUs - a.totalUs);
+  }
+
+  async function downloadTrace() {
+    if (!taskId) return;
+    const response = await fetch(getTraceRawUrl(taskId));
+    const blob = await response.blob();
+    const url = URL.createObjectURL(blob);
+    const a = document.createElement("a");
+    a.href = url;
+    a.download = `trace_${taskId}.json`;
+    a.click();
+    URL.revokeObjectURL(url);
+  }
+
+  async function openInPerfetto() {
+    if (!taskId) return;
+
+    // Fetch trace data from our local API
+    const response = await fetch(getTraceRawUrl(taskId));
+    const traceData = await response.arrayBuffer();
+
+    // Open Perfetto UI
+    const perfettoWindow = window.open("https://ui.perfetto.dev");
+    if (!perfettoWindow) {
+      alert("Failed to open Perfetto. Please allow popups.");
+      return;
+    }
+
+    // Wait for Perfetto to be ready, then send trace via postMessage
+    const onMessage = (e: MessageEvent) => {
+      if (e.data === "PONG") {
+        window.removeEventListener("message", onMessage);
+        perfettoWindow.postMessage(
+          {
+            perfetto: {
+              buffer: traceData,
+              title: `Trace ${taskId}`,
+            },
+          },
+          "https://ui.perfetto.dev",
+        );
+      }
+    };
+    window.addEventListener("message", onMessage);
+
+    // Ping Perfetto until it responds
+    const pingInterval = setInterval(() => {
+      perfettoWindow.postMessage("PING", "https://ui.perfetto.dev");
+    }, 50);
+
+    // Clean up after 10 seconds
+    setTimeout(() => {
+      clearInterval(pingInterval);
+      window.removeEventListener("message", onMessage);
+    }, 10000);
+  }
+
+  onMount(async () => {
+    if (!taskId) {
+      error = "No task ID provided";
+      loading = false;
+      return;
+    }
+
+    try {
+      stats = await fetchTraceStats(taskId);
+    } catch (e) {
+      error = e instanceof Error ? e.message : "Failed to load trace";
+    } finally {
+      loading = false;
+    }
+  });
+
+  const phases = $derived(stats ? parsePhases(stats.byCategory) : []);
+  const sortedRanks = $derived(
+    stats
+      ? Object.keys(stats.byRank)
+          .map(Number)
+          .sort((a, b) => a - b)
+      : [],
+  );
+  const nodeCount = $derived(sortedRanks.length || 1);
+</script>
+
+<div class="min-h-screen bg-exo-dark-gray text-white">
+  <HeaderNav showHome={true} />
+  <div class="max-w-7xl mx-auto px-4 lg:px-8 py-6 space-y-6">
+    <div class="flex items-center justify-between gap-4 flex-wrap">
+      <div>
+        <h1
+          class="text-2xl font-mono tracking-[0.2em] uppercase text-exo-yellow"
+        >
+          Trace
+        </h1>
+        <p class="text-sm text-exo-light-gray font-mono truncate max-w-lg">
+          {taskId}
+        </p>
+      </div>
+      <div class="flex items-center gap-3">
+        <a
+          href="#/traces"
+          class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-3 py-1.5 rounded"
+        >
+          All Traces
+        </a>
+        <button
+          type="button"
+          class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-3 py-1.5 rounded"
+          onclick={downloadTrace}
+          disabled={loading || !!error}
+        >
+          Download
+        </button>
+        <button
+          type="button"
+          class="text-xs font-mono text-exo-dark-gray bg-exo-yellow hover:bg-exo-yellow/90 transition-colors uppercase px-3 py-1.5 rounded font-semibold"
+          onclick={openInPerfetto}
+          disabled={loading || !!error}
+        >
+          View Trace
+        </button>
+      </div>
+    </div>
+
+    {#if loading}
+      <div
+        class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-6 text-center text-exo-light-gray"
+      >
+        <div class="text-sm">Loading trace data...</div>
+      </div>
+    {:else if error}
+      <div
+        class="rounded border border-red-500/30 bg-red-500/10 p-6 text-center text-red-400"
+      >
+        <div class="text-sm">{error}</div>
+      </div>
+    {:else if stats}
+      <!-- Wall Time Summary -->
+      <div
+        class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 space-y-2"
+      >
+        <h2
+          class="text-sm font-mono uppercase tracking-wider text-exo-light-gray"
+        >
+          Summary
+        </h2>
+        <div class="text-3xl font-mono text-exo-yellow">
+          {formatDuration(stats.totalWallTimeUs)}
+        </div>
+        <div class="text-xs text-exo-light-gray">Total wall time</div>
+      </div>
+
+      <!-- By Phase -->
+      {#if phases.length > 0}
+        <div
+          class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 space-y-4"
+        >
+          <h2
+            class="text-sm font-mono uppercase tracking-wider text-exo-light-gray"
+          >
+            By Phase <span class="text-exo-light-gray/50">(avg per node)</span>
+          </h2>
+          <div class="space-y-4">
+            {#each phases as phase}
+              {@const normalizedTotal = phase.totalUs / nodeCount}
+              {@const normalizedStepCount = phase.stepCount / nodeCount}
+              <div class="space-y-2">
+                <div class="flex items-center justify-between">
+                  <span class="text-sm font-mono text-white">{phase.name}</span>
+                  <span class="text-sm font-mono">
+                    <span class="text-exo-yellow"
+                      >{formatDuration(normalizedTotal)}</span
+                    >
+                    <span class="text-exo-light-gray ml-2">
+                      ({normalizedStepCount} steps, {formatDuration(
+                        normalizedTotal / normalizedStepCount,
+                      )}/step)
+                    </span>
+                  </span>
+                </div>
+                {#if phase.subcategories.length > 0}
+                  <div class="pl-4 space-y-1.5">
+                    {#each phase.subcategories as subcat}
+                      {@const normalizedSubcat =
+                        subcat.stats.totalUs / nodeCount}
+                      {@const pct = formatPercentage(
+                        normalizedSubcat,
+                        normalizedTotal,
+                      )}
+                      {@const perStep = normalizedSubcat / normalizedStepCount}
+                      <div
+                        class="flex items-center justify-between text-xs font-mono"
+                      >
+                        <span class="text-exo-light-gray">{subcat.name}</span>
+                        <span class="text-white">
+                          {formatDuration(normalizedSubcat)}
+                          <span class="text-exo-light-gray ml-2">({pct})</span>
+                          <span class="text-exo-light-gray/60 ml-2"
+                            >{formatDuration(perStep)}/step</span
+                          >
+                        </span>
+                      </div>
+                      <!-- Progress bar -->
+                      <div
+                        class="relative h-1.5 bg-exo-black/60 rounded-sm overflow-hidden"
+                      >
+                        <div
+                          class="absolute inset-y-0 left-0 bg-gradient-to-r from-exo-yellow to-exo-yellow/70 transition-all duration-300"
+                          style="width: {pct}"
+                        ></div>
+                      </div>
+                    {/each}
+                  </div>
+                {/if}
+              </div>
+            {/each}
+          </div>
+        </div>
+      {/if}
+
+      <!-- By Rank -->
+      {#if sortedRanks.length > 0}
+        <div
+          class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 space-y-4"
+        >
+          <h2
+            class="text-sm font-mono uppercase tracking-wider text-exo-light-gray"
+          >
+            By Rank
+          </h2>
+          <div class="grid grid-cols-1 md:grid-cols-2 lg:grid-cols-3 gap-4">
+            {#each sortedRanks as rank}
+              {@const rankStats = stats.byRank[rank]}
+              {@const rankPhases = parsePhases(rankStats.byCategory)}
+              <div
+                class="rounded border border-exo-medium-gray/20 bg-exo-dark-gray/60 p-3 space-y-3"
+              >
+                <div class="text-sm font-mono text-exo-yellow">
+                  Rank {rank}
+                </div>
+                <div class="space-y-2">
+                  {#each rankPhases as phase}
+                    <div class="space-y-1">
+                      <div class="flex items-center justify-between text-xs">
+                        <span class="font-mono text-exo-light-gray"
+                          >{phase.name}</span
+                        >
+                        <span class="font-mono text-white">
+                          {formatDuration(phase.totalUs)}
+                          <span class="text-exo-light-gray/50 ml-1">
+                            ({phase.stepCount}x)
+                          </span>
+                        </span>
+                      </div>
+                      {#if phase.subcategories.length > 0}
+                        <div class="pl-2 space-y-0.5">
+                          {#each phase.subcategories as subcat}
+                            {@const pct = formatPercentage(
+                              subcat.stats.totalUs,
+                              phase.totalUs,
+                            )}
+                            {@const perStep =
+                              subcat.stats.totalUs / phase.stepCount}
+                            <div
+                              class="flex items-center justify-between text-[10px] font-mono"
+                            >
+                              <span class="text-exo-light-gray/70"
+                                >{subcat.name}</span
+                              >
+                              <span class="text-exo-light-gray">
+                                {formatDuration(subcat.stats.totalUs)}
+                                <span class="text-exo-light-gray/50"
+                                  >({pct})</span
+                                >
+                                <span class="text-exo-light-gray/30 ml-1"
+                                  >{formatDuration(perStep)}/step</span
+                                >
+                              </span>
+                            </div>
+                          {/each}
+                        </div>
+                      {/if}
+                    </div>
+                  {/each}
+                </div>
+              </div>
+            {/each}
+          </div>
+        </div>
+      {/if}
+    {/if}
+  </div>
+</div>

flake.lock

@@ -21,7 +21,9 @@
           "nixpkgs"
         ],
         "purescript-overlay": "purescript-overlay",
-        "pyproject-nix": "pyproject-nix"
+        "pyproject-nix": [
+          "pyproject-nix"
+        ]
       },
       "locked": {
         "lastModified": 1765953015,
@@ -149,19 +151,44 @@
         "type": "github"
       }
     },
+    "pyproject-build-systems": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ],
+        "pyproject-nix": [
+          "pyproject-nix"
+        ],
+        "uv2nix": [
+          "uv2nix"
+        ]
+      },
+      "locked": {
+        "lastModified": 1763662255,
+        "narHash": "sha256-4bocaOyLa3AfiS8KrWjZQYu+IAta05u3gYZzZ6zXbT0=",
+        "owner": "pyproject-nix",
+        "repo": "build-system-pkgs",
+        "rev": "042904167604c681a090c07eb6967b4dd4dae88c",
+        "type": "github"
+      },
+      "original": {
+        "owner": "pyproject-nix",
+        "repo": "build-system-pkgs",
+        "type": "github"
+      }
+    },
     "pyproject-nix": {
       "inputs": {
         "nixpkgs": [
-          "dream2nix",
           "nixpkgs"
         ]
       },
       "locked": {
-        "lastModified": 1763017646,
-        "narHash": "sha256-Z+R2lveIp6Skn1VPH3taQIuMhABg1IizJd8oVdmdHsQ=",
+        "lastModified": 1764134915,
+        "narHash": "sha256-xaKvtPx6YAnA3HQVp5LwyYG1MaN4LLehpQI8xEdBvBY=",
         "owner": "pyproject-nix",
         "repo": "pyproject.nix",
-        "rev": "47bd6f296502842643078d66128f7b5e5370790c",
+        "rev": "2c8df1383b32e5443c921f61224b198a2282a657",
         "type": "github"
       },
       "original": {
@@ -178,7 +205,10 @@
         "flake-parts": "flake-parts",
         "nixpkgs": "nixpkgs",
         "nixpkgs-swift": "nixpkgs-swift",
-        "treefmt-nix": "treefmt-nix"
+        "pyproject-build-systems": "pyproject-build-systems",
+        "pyproject-nix": "pyproject-nix",
+        "treefmt-nix": "treefmt-nix",
+        "uv2nix": "uv2nix"
       }
     },
     "rust-analyzer-src": {
@@ -239,6 +269,29 @@
         "repo": "treefmt-nix",
         "type": "github"
       }
+    },
+    "uv2nix": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ],
+        "pyproject-nix": [
+          "pyproject-nix"
+        ]
+      },
+      "locked": {
+        "lastModified": 1767701098,
+        "narHash": "sha256-CJhKZnWb3gumR9oTRjFvCg/6lYTGbZRU7xtvcyWIRwU=",
+        "owner": "pyproject-nix",
+        "repo": "uv2nix",
+        "rev": "9d357f0d2ce6f5f35ec7959d7e704452352eb4da",
+        "type": "github"
+      },
+      "original": {
+        "owner": "pyproject-nix",
+        "repo": "uv2nix",
+        "type": "github"
+      }
     }
   },
   "root": "root",

flake.nix

@@ -24,6 +24,26 @@
     dream2nix = {
       url = "github:nix-community/dream2nix";
       inputs.nixpkgs.follows = "nixpkgs";
+      inputs.pyproject-nix.follows = "pyproject-nix";
+    };
+
+    # Python packaging with uv2nix
+    pyproject-nix = {
+      url = "github:pyproject-nix/pyproject.nix";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
+
+    uv2nix = {
+      url = "github:pyproject-nix/uv2nix";
+      inputs.pyproject-nix.follows = "pyproject-nix";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
+
+    pyproject-build-systems = {
+      url = "github:pyproject-nix/build-system-pkgs";
+      inputs.pyproject-nix.follows = "pyproject-nix";
+      inputs.uv2nix.follows = "uv2nix";
+      inputs.nixpkgs.follows = "nixpkgs";
     };
 
     # Pinned nixpkgs for swift-format (swift is broken on x86_64-linux in newer nixpkgs)
@@ -48,6 +68,7 @@
         inputs.treefmt-nix.flakeModule
         ./dashboard/parts.nix
         ./rust/parts.nix
+        ./python/parts.nix
       ];
 
       perSystem =
@@ -58,6 +79,11 @@
           pkgsSwift = import inputs.nixpkgs-swift { inherit system; };
         in
         {
+          # Allow unfree for metal-toolchain (needed for Darwin Metal packages)
+          _module.args.pkgs = import inputs.nixpkgs {
+            inherit system;
+            config.allowUnfreePredicate = pkg: (pkg.pname or "") == "metal-toolchain";
+          };
           treefmt = {
             projectRootFile = "flake.nix";
             programs = {
@@ -79,14 +105,24 @@
                 enable = true;
                 package = pkgsSwift.swiftPackages.swift-format;
               };
+              shfmt.enable = true;
             };
           };
 
-          checks.lint = pkgs.runCommand "lint-check" { } ''
-            export RUFF_CACHE_DIR="$TMPDIR/ruff-cache"
-            ${pkgs.ruff}/bin/ruff check ${inputs.self}/
-            touch $out
-          '';
+          packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin (
+            let
+              uvLock = builtins.fromTOML (builtins.readFile ./uv.lock);
+              mlxPackage = builtins.head (builtins.filter (p: p.name == "mlx") uvLock.package);
+              uvLockMlxVersion = mlxPackage.version;
+            in
+            {
+              metal-toolchain = pkgs.callPackage ./nix/metal-toolchain.nix { };
+              mlx = pkgs.callPackage ./nix/mlx.nix {
+                metal-toolchain = self'.packages.metal-toolchain;
+                inherit uvLockMlxVersion;
+              };
+            }
+          );
 
           devShells.default = with pkgs; pkgs.mkShell {
             inputsFrom = [ self'.checks.cargo-build ];

justfile

@@ -1,7 +1,7 @@
 export NIX_CONFIG := "extra-experimental-features = nix-command flakes"
 
 fmt:
-    nix fmt
+    treefmt || nix fmt
 
 lint:
     uv run ruff check --fix

nix/darwin-build-fixes.patch

@@ -0,0 +1,79 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 0ed30932..d8528132 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -177,11 +177,7 @@ if(MLX_BUILD_METAL)
+     add_compile_definitions(MLX_METAL_DEBUG)
+   endif()
+
+-  # Throw an error if xcrun not found
+-  execute_process(
+-    COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
+-    OUTPUT_VARIABLE MACOS_SDK_VERSION
+-    OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
++  set(MACOS_SDK_VERSION @sdkVersion@)
+
+   if(${MACOS_SDK_VERSION} LESS 14.0)
+     message(
+@@ -199,11 +195,8 @@ if(MLX_BUILD_METAL)
+     endif()
+     set(XCRUN_FLAGS "-mmacosx-version-min=${CMAKE_OSX_DEPLOYMENT_TARGET}")
+   endif()
+-  execute_process(
+-    COMMAND
+-      zsh "-c"
+-      "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal ${XCRUN_FLAGS} -E -x metal -P - | tail -1 | tr -d '\n'"
+-    OUTPUT_VARIABLE MLX_METAL_VERSION COMMAND_ERROR_IS_FATAL ANY)
++  set(
++    MLX_METAL_VERSION @metalVersion@)
+   FetchContent_Declare(metal_cpp URL ${METAL_CPP_URL})
+   FetchContent_MakeAvailable(metal_cpp)
+   target_include_directories(
+diff --git a/cmake/extension.cmake b/cmake/extension.cmake
+index 13db804a..5b385132 100644
+--- a/cmake/extension.cmake
++++ b/cmake/extension.cmake
+@@ -36,7 +36,7 @@ macro(mlx_build_metallib)
+   add_custom_command(
+     OUTPUT ${MTLLIB_BUILD_TARGET}
+     COMMAND
+-      xcrun -sdk macosx metal
++      metal -fmodules-cache-path=${CMAKE_BINARY_DIR}/metal-cache
+       "$<LIST:TRANSFORM,${MTLLIB_INCLUDE_DIRS},PREPEND,-I>"
+       ${MTLLIB_COMPILE_OPTIONS} ${MTLLIB_SOURCES} -o ${MTLLIB_BUILD_TARGET}
+     DEPENDS ${MTLLIB_DEPS} ${MTLLIB_SOURCES}
+diff --git a/mlx/backend/metal/kernels/CMakeLists.txt b/mlx/backend/metal/kernels/CMakeLists.txt
+index 262b0495..5c7446ad 100644
+--- a/mlx/backend/metal/kernels/CMakeLists.txt
++++ b/mlx/backend/metal/kernels/CMakeLists.txt
+@@ -29,7 +29,7 @@ function(build_kernel_base TARGET SRCFILE DEPS)
+                     "-mmacosx-version-min=${CMAKE_OSX_DEPLOYMENT_TARGET}")
+   endif()
+   add_custom_command(
+-    COMMAND xcrun -sdk macosx metal ${METAL_FLAGS} -c ${SRCFILE}
++    COMMAND metal -fmodules-cache-path=${CMAKE_BINARY_DIR}/metal-cache ${METAL_FLAGS} -c ${SRCFILE}
+             -I${PROJECT_SOURCE_DIR} -o ${TARGET}.air
+     DEPENDS ${SRCFILE} ${DEPS} ${BASE_HEADERS}
+     OUTPUT ${TARGET}.air
+@@ -170,7 +170,7 @@ endif()
+
+ add_custom_command(
+   OUTPUT ${MLX_METAL_PATH}/mlx.metallib
+-  COMMAND xcrun -sdk macosx metallib ${KERNEL_AIR} -o
++  COMMAND metallib ${KERNEL_AIR} -o
+           ${MLX_METAL_PATH}/mlx.metallib
+   DEPENDS ${KERNEL_AIR}
+   COMMENT "Building mlx.metallib"
+diff --git a/mlx/backend/metal/make_compiled_preamble.sh b/mlx/backend/metal/make_compiled_preamble.sh
+index bb55ed3a..94ea7dd7 100644
+--- a/mlx/backend/metal/make_compiled_preamble.sh
++++ b/mlx/backend/metal/make_compiled_preamble.sh
+@@ -31,7 +31,7 @@ OUTPUT_FILE=${OUTPUT_DIR}/${SRC_NAME}.cpp
+ mkdir -p "$OUTPUT_DIR"
+
+ # Use the metal compiler to get a list of headers (with depth)
+-CCC="xcrun -sdk macosx metal -x metal"
++CCC="metal -x metal -fmodules-cache-path=${OUTPUT_DIR}/metal-cache"
+ HDRS=$( $CCC -I"$SRC_DIR" -I"$JIT_INCLUDES" -DMLX_METAL_JIT -E -P -CC -C -H "$INPUT_FILE" $CFLAGS -w 2>&1 1>/dev/null )
+
+ # Remove any included system frameworks (for MetalPerformancePrimitive headers)

nix/metal-toolchain.nix

@@ -0,0 +1,56 @@
+{ lib, stdenvNoCC, requireFile, nix }:
+
+let
+  narFile = requireFile {
+    name = "metal-toolchain-17C48.nar";
+    message = ''
+      The Metal Toolchain NAR must be available.
+
+      If you have cachix configured for exo.cachix.org, this should be automatic.
+
+      Otherwise:
+        1. Install Xcode 26+ from the App Store
+        2. Run: xcodebuild -downloadComponent MetalToolchain
+        3. Export the toolchain:
+           hdiutil attach "$(find /System/Library/AssetsV2/com_apple_MobileAsset_MetalToolchain -name '*.dmg' | head -1)" -mountpoint /tmp/metal-dmg
+           cp -R /tmp/metal-dmg/Metal.xctoolchain /tmp/metal-export
+           hdiutil detach /tmp/metal-dmg
+        4. Create NAR and add to store:
+           nix nar pack /tmp/metal-export > /tmp/metal-toolchain-17C48.nar
+           nix store add --mode flat /tmp/metal-toolchain-17C48.nar
+    '';
+    hash = "sha256-ayR5mXN4sZAddwKEG2OszGRF93k9ZFc7H0yi2xbylQw=";
+  };
+in
+stdenvNoCC.mkDerivation {
+  pname = "metal-toolchain";
+  version = "17C48";
+
+  dontUnpack = true;
+  dontBuild = true;
+  dontFixup = true;
+
+  nativeBuildInputs = [ nix ];
+
+  installPhase = ''
+    runHook preInstall
+
+    nix-store --restore $out < ${narFile}
+
+    # Create bin directory with symlinks for PATH
+    mkdir -p $out/bin
+    ln -s $out/usr/bin/metal $out/bin/metal
+    ln -s $out/usr/bin/metallib $out/bin/metallib
+
+    runHook postInstall
+  '';
+
+  # Metal language version for CMake (from: echo __METAL_VERSION__ | metal -E -x metal -P -)
+  passthru.metalVersion = "400";
+
+  meta = {
+    description = "Apple Metal compiler toolchain";
+    platforms = [ "aarch64-darwin" ];
+    license = lib.licenses.unfree;
+  };
+}

nix/mlx.nix

@@ -0,0 +1,158 @@
+{ stdenv
+, lib
+, fetchFromGitHub
+, replaceVars
+, fetchzip
+, cmake
+, nlohmann_json
+, apple-sdk_26
+, metal-toolchain
+, runCommand
+, fmt
+, python313Packages
+, uvLockMlxVersion
+}:
+
+assert stdenv.isDarwin;
+
+let
+  python = python313Packages.python;
+
+  # Static dependencies included directly during compilation
+  gguf-tools = fetchFromGitHub {
+    owner = "antirez";
+    repo = "gguf-tools";
+    rev = "8fa6eb65236618e28fd7710a0fba565f7faa1848";
+    hash = "sha256-15FvyPOFqTOr5vdWQoPnZz+mYH919++EtghjozDlnSA=";
+  };
+
+  metal_cpp = fetchzip {
+    url = "https://developer.apple.com/metal/cpp/files/metal-cpp_26.zip";
+    hash = "sha256-7n2eI2lw/S+Us6l7YPAATKwcIbRRpaQ8VmES7S8ZjY8=";
+  };
+
+  nanobind = fetchFromGitHub {
+    owner = "wjakob";
+    repo = "nanobind";
+    rev = "v2.10.2";
+    hash = "sha256-io44YhN+VpfHFWyvvLWSanRgbzA0whK8WlDNRi3hahU=";
+    fetchSubmodules = true;
+  };
+
+  mlx = stdenv.mkDerivation rec {
+    pname = "mlx";
+    version = let v = "0.30.4"; in
+      assert v == uvLockMlxVersion || throw "MLX version mismatch: nix/mlx.nix has ${v} but uv.lock has ${uvLockMlxVersion}. Update both the version and hash in nix/mlx.nix.";
+      v;
+    pyproject = true;
+
+    src = fetchFromGitHub {
+      owner = "ml-explore";
+      repo = "mlx";
+      tag = "v${version}";
+      hash = "sha256-OJk6jPlbaSlsUdk3ADz3tWcRzTWXRof3/q8Soe1AO6w=";
+    };
+
+    patches = [
+      (replaceVars ./darwin-build-fixes.patch {
+        sdkVersion = apple-sdk_26.version;
+        metalVersion = metal-toolchain.metalVersion;
+      })
+    ];
+
+    postPatch = ''
+      substituteInPlace mlx/backend/cpu/jit_compiler.cpp \
+        --replace-fail "g++" "$CXX"
+    '';
+
+    dontUseCmakeConfigure = true;
+
+    enableParallelBuilding = true;
+
+    # Allows multiple cores to be used in Python builds.
+    postUnpack = ''
+      export MAKEFLAGS+="''${enableParallelBuilding:+-j$NIX_BUILD_CORES}"
+    '';
+
+    # Updates the wrong fetcher rev attribute
+    passthru.skipBulkUpdate = true;
+
+    env = {
+      DEV_RELEASE = 1;
+      CMAKE_ARGS = toString [
+        (lib.cmakeBool "USE_SYSTEM_FMT" true)
+        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_GGUFLIB" "${gguf-tools}")
+        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_JSON" "${nlohmann_json.src}")
+        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_NANOBIND" "${nanobind}")
+        (lib.cmakeBool "FETCHCONTENT_FULLY_DISCONNECTED" true)
+        (lib.cmakeBool "MLX_BUILD_METAL" true)
+        (lib.cmakeOptionType "filepath" "FETCHCONTENT_SOURCE_DIR_METAL_CPP" "${metal_cpp}")
+        (lib.cmakeOptionType "string" "CMAKE_OSX_DEPLOYMENT_TARGET" "${apple-sdk_26.version}")
+        (lib.cmakeOptionType "filepath" "CMAKE_OSX_SYSROOT" "${apple-sdk_26.passthru.sdkroot}")
+      ];
+      SDKROOT = apple-sdk_26.passthru.sdkroot;
+      MACOSX_DEPLOYMENT_TARGET = apple-sdk_26.version;
+    };
+
+    build-system = [
+      python313Packages.setuptools
+    ];
+
+    nativeBuildInputs = [
+      cmake
+      metal-toolchain
+      python313Packages.pypaBuildHook
+      python313Packages.pypaInstallHook
+      python313Packages.setuptools
+      python313Packages.typing-extensions
+      python313Packages.wheel
+      python313Packages.cmake
+      python313Packages.ninja
+    ];
+
+    buildInputs = [
+      fmt
+      gguf-tools
+      python313Packages.nanobind
+      python313Packages.pybind11
+      apple-sdk_26
+    ];
+
+    # Tests require Metal GPU access which isn't available in the Nix sandbox.
+    # To run tests, build with: nix build --option sandbox false .#mlx.passthru.tests.mlxTest
+    doCheck = false;
+
+    pythonImportsCheck = [ "mlx" ];
+
+    passthru.tests = {
+      # Runs example scripts to verify MLX works. Requires --option sandbox false
+      # since Metal GPU access is needed.
+      mlxTest =
+        runCommand "run-mlx-examples"
+          {
+            buildInputs = [ mlx ];
+            nativeBuildInputs = [ python ];
+          }
+          ''
+            cp ${src}/examples/python/logistic_regression.py .
+            ${python.interpreter} logistic_regression.py
+            rm logistic_regression.py
+
+            cp ${src}/examples/python/linear_regression.py .
+            ${python.interpreter} linear_regression.py
+            rm linear_regression.py
+
+            touch $out
+          '';
+    };
+
+    meta = {
+      homepage = "https://github.com/ml-explore/mlx";
+      description = "Array framework for Apple silicon";
+      changelog = "https://github.com/ml-explore/mlx/releases/tag/${src.tag}";
+      license = lib.licenses.mit;
+      platforms = [ "aarch64-darwin" ];
+    };
+  };
+in
+mlx

pyproject.toml

@@ -17,16 +17,16 @@ dependencies = [
     "loguru>=0.7.3",
     "exo_pyo3_bindings", # rust bindings
     "anyio==4.11.0",
-    "mlx==0.30.3; sys_platform == 'darwin'",
-    "mlx[cpu]==0.30.3; sys_platform == 'linux'",
-    "mlx-lm @ git+https://github.com/AlexCheema/mlx-lm.git@fix-transformers-5.0.0rc2",
+    "mlx==0.30.4; sys_platform == 'darwin'",
+    "mlx[cpu]==0.30.4; sys_platform == 'linux'",
+    "mlx-lm",
     "tiktoken>=0.12.0", # required for kimi k2 tokenizer
     "hypercorn>=0.18.0",
     "openai-harmony>=0.0.8",
     "httpx>=0.28.1",
     "tomlkit>=0.14.0",
     "pillow>=11.0,<12.0", # compatibility with mflux
-    "mflux>=0.14.2",
+    "mflux==0.15.4",
     "python-multipart>=0.0.21",
 ]
 
@@ -34,7 +34,6 @@ dependencies = [
 exo-master = "exo.master.main:main"
 exo-worker = "exo.worker.main:main"
 exo = "exo.main:main"
-exo-rsh = "exo.rsh.client:main"
 
 # dependencies only required for development
 [dependency-groups]
@@ -64,6 +63,7 @@ members = [
 
 [tool.uv.sources]
 exo_pyo3_bindings = { workspace = true }
+mlx-lm = { git = "https://github.com/ml-explore/mlx-lm", branch = "main" }
 # Uncomment to use local mlx/mlx-lm development versions:
 # mlx = { path = "/Users/Shared/mlx", editable=true }
 # mlx-lm = { path = "/Users/Shared/mlx-lm", editable=true }

python/parts.nix

@@ -0,0 +1,94 @@
+{ inputs, ... }:
+{
+  perSystem =
+    { config, self', pkgs, lib, system, ... }:
+    let
+      # Load workspace from uv.lock
+      workspace = inputs.uv2nix.lib.workspace.loadWorkspace {
+        workspaceRoot = inputs.self;
+      };
+
+      # Create overlay from workspace
+      # Use wheels from PyPI for most packages; we override mlx with our pure Nix Metal build
+      overlay = workspace.mkPyprojectOverlay { sourcePreference = "wheel"; };
+
+      # Override overlay to inject Nix-built components
+      exoOverlay = final: prev: {
+        # Replace workspace exo_pyo3_bindings with Nix-built wheel
+        exo-pyo3-bindings = pkgs.stdenv.mkDerivation {
+          pname = "exo-pyo3-bindings";
+          version = "0.1.0";
+          src = self'.packages.exo_pyo3_bindings;
+          # Install from pre-built wheel
+          nativeBuildInputs = [ final.pyprojectWheelHook ];
+          dontStrip = true;
+        };
+      };
+
+      python = pkgs.python313;
+
+      # Overlay to provide build systems and custom packages
+      buildSystemsOverlay = final: prev: {
+        # Use our pure Nix-built MLX with Metal support
+        mlx = self'.packages.mlx;
+
+        # mlx-lm is a git dependency that needs setuptools
+        mlx-lm = prev.mlx-lm.overrideAttrs (old: {
+          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
+            final.setuptools
+          ];
+        });
+      };
+
+      pythonSet = (pkgs.callPackage inputs.pyproject-nix.build.packages {
+        inherit python;
+      }).overrideScope (
+        lib.composeManyExtensions [
+          inputs.pyproject-build-systems.overlays.default
+          overlay
+          exoOverlay
+          buildSystemsOverlay
+        ]
+      );
+      exoVenv = pythonSet.mkVirtualEnv "exo-env" workspace.deps.default;
+
+      # Virtual environment with dev dependencies for testing
+      testVenv = pythonSet.mkVirtualEnv "exo-test-env" (
+        workspace.deps.default // {
+          exo = [ "dev" ]; # Include pytest, pytest-asyncio, pytest-env
+        }
+      );
+
+      exoPackage = pkgs.runCommand "exo"
+        {
+          nativeBuildInputs = [ pkgs.makeWrapper ];
+        }
+        ''
+          mkdir -p $out/bin
+
+          # Create wrapper scripts
+          for script in exo exo-master exo-worker; do
+            makeWrapper ${exoVenv}/bin/$script $out/bin/$script \
+              --set DASHBOARD_DIR ${self'.packages.dashboard} \
+              ${lib.optionalString pkgs.stdenv.isDarwin "--prefix PATH : ${pkgs.macmon}/bin"}
+          done
+        '';
+    in
+    {
+      # Python package only available on macOS (requires MLX/Metal)
+      packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin {
+        exo = exoPackage;
+        # Test environment for running pytest outside of Nix sandbox (needs GPU access)
+        exo-test-env = testVenv;
+      };
+
+      checks = {
+        # Ruff linting (works on all platforms)
+        lint = pkgs.runCommand "ruff-lint" { } ''
+          export RUFF_CACHE_DIR="$TMPDIR/ruff-cache"
+          ${pkgs.ruff}/bin/ruff check ${inputs.self}/
+          touch $out
+        '';
+      };
+    };
+}

src/exo/cli/__init__.py

@@ -1,32 +0,0 @@
-"""Exo CLI - SLURM-compatible job management commands."""
-
-
-def run_subcommand(command: str, args: list[str]) -> int:
-    """Route to the appropriate subcommand handler.
-
-    Args:
-        command: The subcommand name (sbatch, squeue, scancel, salloc)
-        args: Command line arguments for the subcommand
-
-    Returns:
-        Exit code from the subcommand
-    """
-    if command == "sbatch":
-        from exo.cli.sbatch import main
-
-        return main(args)
-    elif command == "squeue":
-        from exo.cli.squeue import main
-
-        return main(args)
-    elif command == "scancel":
-        from exo.cli.scancel import main
-
-        return main(args)
-    elif command == "salloc":
-        from exo.cli.salloc import main
-
-        return main(args)
-    else:
-        print(f"Unknown subcommand: {command}")
-        return 1

src/exo/cli/common.py

@@ -1,118 +0,0 @@
-"""Common utilities for Exo CLI commands."""
-
-import json
-import os
-import urllib.request
-from typing import Any
-from urllib.error import HTTPError, URLError
-
-# Default API endpoint
-DEFAULT_API_HOST = "localhost"
-DEFAULT_API_PORT = 52415
-
-
-def get_api_base() -> str:
-    """Get the API base URL from environment or defaults."""
-    host = os.environ.get("EXO_API_HOST", DEFAULT_API_HOST)
-    port = os.environ.get("EXO_API_PORT", str(DEFAULT_API_PORT))
-    return f"http://{host}:{port}"
-
-
-def api_request(
-    method: str,
-    path: str,
-    data: dict[str, Any] | None = None,
-) -> dict[str, Any] | list[Any]:
-    """Make an API request to the Exo server.
-
-    Args:
-        method: HTTP method (GET, POST, DELETE, etc.)
-        path: API path (e.g., "/flash/instances")
-        data: Optional JSON data for POST/PUT requests
-
-    Returns:
-        Parsed JSON response
-
-    Raises:
-        SystemExit: On connection or HTTP errors
-    """
-    url = f"{get_api_base()}{path}"
-
-    request_data = None
-    if data is not None:
-        request_data = json.dumps(data).encode("utf-8")
-
-    req = urllib.request.Request(
-        url,
-        data=request_data,
-        method=method,
-    )
-    req.add_header("Content-Type", "application/json")
-
-    try:
-        with urllib.request.urlopen(req, timeout=30) as response:  # pyright: ignore[reportAny]
-            body: str = response.read().decode("utf-8")  # pyright: ignore[reportAny]
-            if body:
-                return json.loads(body)  # pyright: ignore[reportAny]
-            return {}
-    except HTTPError as e:
-        error_body = e.read().decode("utf-8") if e.fp else ""
-        print(f"API error: {e.code} {e.reason}")
-        if error_body:
-            try:
-                error_json: dict[str, str] = json.loads(error_body)  # pyright: ignore[reportAny]
-                if "detail" in error_json:
-                    print(f"  {error_json['detail']}")
-            except json.JSONDecodeError:
-                print(f"  {error_body}")
-        raise SystemExit(1) from None
-    except URLError as e:
-        print(f"Connection error: {e.reason}")
-        print(f"Is Exo running at {get_api_base()}?")
-        raise SystemExit(1) from None
-
-
-def truncate_id(instance_id: str, length: int = 8) -> str:
-    """Truncate a UUID for display.
-
-    Args:
-        instance_id: Full UUID string
-        length: Number of characters to keep
-
-    Returns:
-        Truncated ID without hyphens
-    """
-    return instance_id.replace("-", "")[:length]
-
-
-def format_table(headers: list[str], rows: list[list[str]]) -> str:
-    """Format data as a simple text table.
-
-    Args:
-        headers: Column headers
-        rows: List of rows, each row is a list of column values
-
-    Returns:
-        Formatted table string
-    """
-    if not rows:
-        return "  ".join(f"{h:<10}" for h in headers)
-
-    # Calculate column widths
-    widths = [len(h) for h in headers]
-    for row in rows:
-        for i, cell in enumerate(row):
-            if i < len(widths):
-                widths[i] = max(widths[i], len(cell))
-
-    # Build format string
-    fmt = "  ".join(f"{{:<{w}}}" for w in widths)
-
-    # Format output
-    lines = [fmt.format(*headers)]
-    for row in rows:
-        # Pad row if needed
-        padded = row + [""] * (len(headers) - len(row))
-        lines.append(fmt.format(*padded[: len(headers)]))
-
-    return "\n".join(lines)

src/exo/cli/salloc.py

@@ -1,100 +0,0 @@
-"""salloc - Allocate nodes for interactive use.
-
-Usage:
-    exo salloc [options] [-- command [args...]]
-
-Options:
-    -N, --nodes N       Number of nodes to allocate (default: 1)
-    --hosts HOSTS       Comma-separated list of hostnames
-
-If a command is provided after --, it will be executed with
-SLURM-like environment variables set:
-    SLURM_JOB_NODELIST  - Comma-separated list of allocated nodes
-    SLURM_NNODES        - Number of allocated nodes
-
-Examples:
-    exo salloc --nodes=2 --hosts=node1,node2 -- mpirun ./my_program
-    exo salloc --hosts=localhost -- bash
-"""
-
-import argparse
-import os
-import subprocess
-import sys
-
-
-def main(args: list[str]) -> int:
-    """Main entry point for salloc command."""
-    # Split args at -- if present
-    cmd_args: list[str] = []
-    salloc_args = args
-
-    if "--" in args:
-        idx = args.index("--")
-        salloc_args = args[:idx]
-        cmd_args = args[idx + 1 :]
-
-    parser = argparse.ArgumentParser(
-        prog="exo salloc",
-        description="Allocate nodes for interactive use",
-    )
-    parser.add_argument(
-        "-N",
-        "--nodes",
-        type=int,
-        default=1,
-        help="Number of nodes to allocate (default: 1)",
-    )
-    parser.add_argument(
-        "--hosts",
-        help="Comma-separated list of hostnames (required)",
-    )
-
-    parsed = parser.parse_args(salloc_args)
-
-    nodes: int = parsed.nodes  # pyright: ignore[reportAny]
-    hosts: str | None = parsed.hosts  # pyright: ignore[reportAny]
-
-    # Require explicit hosts since we can't discover them from topology
-    if not hosts:
-        print("Error: --hosts is required (e.g., --hosts=node1,node2)", file=sys.stderr)
-        print("       The Exo topology doesn't expose hostnames.", file=sys.stderr)
-        return 1
-
-    host_list = [h.strip() for h in hosts.split(",") if h.strip()]
-
-    if len(host_list) < nodes:
-        print(
-            f"Error: Requested {nodes} nodes but only {len(host_list)} hosts provided",
-            file=sys.stderr,
-        )
-        return 1
-
-    # Use first N hosts
-    allocated_hosts = host_list[:nodes]
-    nodelist = ",".join(allocated_hosts)
-
-    # Set environment variables
-    env = os.environ.copy()
-    env["SLURM_JOB_NODELIST"] = nodelist
-    env["SLURM_NNODES"] = str(nodes)
-
-    print(f"salloc: Granted job allocation on {nodes} node(s)")
-    print(f"salloc: Nodes: {nodelist}")
-
-    if cmd_args:
-        # Run the command
-        print(f"salloc: Running: {' '.join(cmd_args)}")
-        result = subprocess.run(cmd_args, env=env)
-        return result.returncode
-    else:
-        # Start interactive shell
-        shell = os.environ.get("SHELL", "/bin/bash")
-        print(f"salloc: Starting shell {shell}")
-        print("salloc: Use 'exit' to release allocation")
-        result = subprocess.run([shell], env=env)
-        return result.returncode
-
-
-if __name__ == "__main__":
-    sys.exit(main(sys.argv[1:]))

src/exo/cli/sbatch.py

@@ -1,233 +0,0 @@
-"""sbatch - Submit a batch job to Exo.
-
-Usage:
-    exo sbatch [options] <script|executable>
-    exo sbatch --job-name=NAME --nodes=N <executable>
-
-Options:
-    -J, --job-name NAME         Job name
-    -N, --nodes N               Number of nodes (default: 1)
-    --ntasks-per-node N         Tasks per node (default: 1)
-    -D, --chdir DIR             Working directory
-    --hosts HOSTS               Comma-separated list of hostnames
-
-Job scripts can contain #SBATCH directives:
-    #!/bin/bash
-    #SBATCH --job-name=Sod2D
-    #SBATCH --nodes=2
-    #SBATCH --chdir=/path/to/workdir
-
-    /path/to/flash4
-"""
-
-import argparse
-import os
-import re
-import sys
-
-from exo.cli.common import api_request, truncate_id
-
-
-def parse_job_script(script_path: str) -> tuple[dict[str, str], str | None]:
-    """Parse a job script for #SBATCH directives and executable.
-
-    Args:
-        script_path: Path to the job script
-
-    Returns:
-        Tuple of (directives dict, executable path or None)
-    """
-    directives: dict[str, str] = {}
-    executable: str | None = None
-
-    with open(script_path, "r") as f:
-        for line in f:
-            line = line.strip()
-
-            # Parse #SBATCH directives
-            if line.startswith("#SBATCH"):
-                # Handle both --option=value and --option value formats
-                match = re.match(r"#SBATCH\s+(-\w|--[\w-]+)(?:=|\s+)(.+)", line)
-                if match:
-                    opt, val = match.groups()
-                    directives[opt.lstrip("-")] = val.strip()
-                continue
-
-            # Skip comments and empty lines
-            if line.startswith("#") or not line:
-                continue
-
-            # First non-comment, non-directive line is the executable
-            if executable is None:
-                # Handle lines like "/path/to/flash4" or "srun /path/to/flash4"
-                parts = line.split()
-                if parts:
-                    # Skip srun/mpirun prefixes if present
-                    for part in parts:
-                        if not part.startswith("-") and "/" in part:
-                            executable = part
-                            break
-                    if executable is None and parts:
-                        executable = parts[-1]  # Last token
-
-    return directives, executable
-
-
-def main(args: list[str]) -> int:
-    """Main entry point for sbatch command."""
-    parser = argparse.ArgumentParser(
-        prog="exo sbatch",
-        description="Submit a batch job to Exo",
-    )
-    parser.add_argument(
-        "script",
-        help="Job script or executable path",
-    )
-    parser.add_argument(
-        "-J",
-        "--job-name",
-        dest="job_name",
-        help="Job name",
-    )
-    parser.add_argument(
-        "-N",
-        "--nodes",
-        type=int,
-        default=1,
-        help="Number of nodes (default: 1)",
-    )
-    parser.add_argument(
-        "--ntasks-per-node",
-        type=int,
-        default=1,
-        help="Tasks per node (default: 1)",
-    )
-    parser.add_argument(
-        "-D",
-        "--chdir",
-        help="Working directory",
-    )
-    parser.add_argument(
-        "--hosts",
-        help="Comma-separated list of hostnames",
-    )
-
-    parsed = parser.parse_args(args)
-
-    # Extract typed values from namespace
-    script_path: str = parsed.script  # pyright: ignore[reportAny]
-    arg_job_name: str | None = parsed.job_name  # pyright: ignore[reportAny]
-    arg_nodes: int = parsed.nodes  # pyright: ignore[reportAny]
-    arg_ntasks: int = parsed.ntasks_per_node  # pyright: ignore[reportAny]
-    arg_chdir: str | None = parsed.chdir  # pyright: ignore[reportAny]
-    arg_hosts: str | None = parsed.hosts  # pyright: ignore[reportAny]
-
-    # Determine if input is a script or direct executable
-    executable: str | None = None
-    directives: dict[str, str] = {}
-
-    if os.path.isfile(script_path):
-        # Check if it's a binary file (executable) or text script
-        is_binary = False
-        try:
-            with open(script_path, "rb") as f:
-                chunk = f.read(512)
-                # Binary files typically contain null bytes
-                is_binary = b"\x00" in chunk
-        except OSError:
-            pass
-
-        if is_binary:
-            # It's a binary executable
-            executable = script_path
-        else:
-            # Try to read as text
-            try:
-                with open(script_path, "r") as f:
-                    first_line = f.readline()
-                    f.seek(0)
-                    content = f.read(1024)
-
-                if first_line.startswith("#!") or "#SBATCH" in content:
-                    # It's a job script - parse it
-                    directives, executable = parse_job_script(script_path)
-                else:
-                    # It's an executable (text but no shebang/directives)
-                    executable = script_path
-            except UnicodeDecodeError:
-                # Can't read as text - treat as binary executable
-                executable = script_path
-    else:
-        # Not a file - treat as executable path
-        executable = script_path
-
-    if executable is None:
-        print("Error: No executable found in job script", file=sys.stderr)
-        return 1
-
-    # Build job parameters - CLI args override script directives
-    job_name = arg_job_name or directives.get("job-name") or directives.get("J")
-    if not job_name:
-        # Generate name from executable
-        job_name = os.path.basename(executable).replace(".", "_")
-
-    nodes = arg_nodes
-    if "nodes" in directives:
-        nodes = int(directives["nodes"])
-    if "N" in directives:
-        nodes = int(directives["N"])
-    if arg_nodes != 1:  # CLI override
-        nodes = arg_nodes
-
-    ntasks = arg_ntasks
-    if "ntasks-per-node" in directives:
-        ntasks = int(directives["ntasks-per-node"])
-    if arg_ntasks != 1:  # CLI override
-        ntasks = arg_ntasks
-
-    workdir = arg_chdir or directives.get("chdir") or directives.get("D")
-    if not workdir:
-        workdir = os.getcwd()
-
-    hosts = arg_hosts or directives.get("hosts") or ""
-
-    # Resolve executable to absolute path
-    if not os.path.isabs(executable):
-        executable = os.path.abspath(os.path.join(workdir, executable))
-
-    # Submit job via API using query parameters
-    from urllib.parse import urlencode
-
-    params = {
-        "simulation_name": job_name,
-        "flash_executable_path": executable,
-        "parameter_file_path": "",  # FLASH par file - use default
-        "working_directory": workdir,
-        "ranks_per_node": str(ntasks),
-        "min_nodes": str(nodes),
-        "hosts": hosts,
-    }
-
-    query_string = urlencode(params)
-    result = api_request("POST", f"/flash/launch?{query_string}")
-
-    # Print job submission confirmation
-    if isinstance(result, dict):
-        instance_id_val = result.get("instance_id")
-
-        if instance_id_val is not None:
-            job_id = truncate_id(str(instance_id_val))  # pyright: ignore[reportAny]
-            print(f"Submitted batch job {job_id}")
-        else:
-            # Instance created asynchronously - user should check squeue
-            print("Job submitted successfully")
-            print("Use 'exo squeue' to view job ID")
-    else:
-        print("Job submitted successfully")
-        print("Use 'exo squeue' to view job ID")
-
-    return 0
-
-
-if __name__ == "__main__":
-    sys.exit(main(sys.argv[1:]))

src/exo/cli/scancel.py

@@ -1,95 +0,0 @@
-"""scancel - Cancel jobs in the Exo queue.
-
-Usage:
-    exo scancel <jobid> [<jobid>...]
-
-Arguments:
-    jobid   Job ID (or prefix) to cancel. Can specify multiple.
-
-Examples:
-    exo scancel abc123          # Cancel job starting with abc123
-    exo scancel abc123 def456   # Cancel multiple jobs
-"""
-
-import argparse
-import sys
-from typing import Any, cast
-
-from exo.cli.common import api_request, truncate_id
-
-
-def main(args: list[str]) -> int:
-    """Main entry point for scancel command."""
-    parser = argparse.ArgumentParser(
-        prog="exo scancel",
-        description="Cancel jobs in the Exo queue",
-    )
-    parser.add_argument(
-        "jobids",
-        nargs="+",
-        help="Job ID(s) to cancel",
-    )
-
-    parsed = parser.parse_args(args)
-    jobids: list[str] = parsed.jobids  # pyright: ignore[reportAny]
-
-    # Fetch current jobs to resolve partial IDs
-    result = api_request("GET", "/flash/instances")
-    if isinstance(result, list):
-        instances = cast(list[dict[str, Any]], result)
-    else:
-        instances = cast(list[dict[str, Any]], result.get("instances", []))
-
-    # Build lookup of full IDs
-    id_map: dict[str, str] = {}
-    for inst in instances:
-        iid = inst.get("instance_id", "")  # pyright: ignore[reportAny]
-        full_id = str(iid) if iid else ""  # pyright: ignore[reportAny]
-        if full_id:
-            # Map both full ID and truncated versions
-            normalized = full_id.replace("-", "").lower()
-            id_map[normalized] = full_id
-            # Also map prefixes
-            for length in range(4, len(normalized) + 1):
-                prefix = normalized[:length]
-                if prefix not in id_map:
-                    id_map[prefix] = full_id
-
-    cancelled = 0
-    errors = 0
-
-    for jobid in jobids:
-        search = jobid.lower().replace("-", "")
-
-        # Find matching full ID
-        full_id = id_map.get(search)
-        if not full_id:
-            # Try prefix match
-            matches = [fid for key, fid in id_map.items() if key.startswith(search)]
-            if len(matches) == 1:
-                full_id = matches[0]
-            elif len(matches) > 1:
-                print(f"Ambiguous job ID: {jobid} matches multiple jobs")
-                errors += 1
-                continue
-            else:
-                print(f"Job not found: {jobid}")
-                errors += 1
-                continue
-
-        # Cancel the job
-        try:
-            api_request("DELETE", f"/flash/{full_id}")
-            print(f"Job {truncate_id(full_id)} cancelled")
-            cancelled += 1
-        except SystemExit:
-            print(f"Failed to cancel job {truncate_id(full_id)}")
-            errors += 1
-
-    if errors > 0 and cancelled == 0:
-        return 1
-    return 0
-
-
-if __name__ == "__main__":
-    sys.exit(main(sys.argv[1:]))

src/exo/cli/squeue.py

@@ -1,165 +0,0 @@
-"""squeue - View the Exo job queue.
-
-Usage:
-    exo squeue [options]
-
-Options:
-    -l, --long      Show detailed output
-    -j, --job ID    Show only this job
-
-Output columns:
-    JOBID   - Job identifier (truncated UUID)
-    NAME    - Job name
-    NODES   - Number of nodes
-    STATE   - Job state (PENDING, RUNNING, FAILED, etc.)
-"""
-
-import argparse
-import sys
-from typing import Any, cast
-
-from exo.cli.common import api_request, format_table, truncate_id
-
-# Map Exo runner statuses to SLURM-like states
-STATUS_MAP: dict[str, str] = {
-    "RunnerIdle": "PENDING",
-    "RunnerConnecting": "CONFIGURING",
-    "RunnerConnected": "CONFIGURING",
-    "RunnerLoading": "CONFIGURING",
-    "RunnerLoaded": "CONFIGURING",
-    "RunnerWarmingUp": "CONFIGURING",
-    "RunnerReady": "COMPLETING",
-    "RunnerRunning": "RUNNING",
-    "RunnerShuttingDown": "COMPLETING",
-    "RunnerShutdown": "COMPLETED",
-    "RunnerFailed": "FAILED",
-}
-
-
-def get_job_state(runner_statuses: dict[str, Any]) -> str:
-    """Determine overall job state from runner statuses."""
-    if not runner_statuses:
-        return "PENDING"
-
-    states: set[str] = set()
-    for status_val in runner_statuses.values():  # pyright: ignore[reportAny]
-        if isinstance(status_val, dict):
-            # Extract status type from discriminated union
-            type_val = status_val.get("type", "RunnerIdle")  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
-            status_type = str(type_val) if type_val else "RunnerIdle"  # pyright: ignore[reportUnknownArgumentType]
-        elif isinstance(status_val, str):
-            status_type = status_val
-        else:
-            status_type = "RunnerIdle"
-        # Strip parentheses from status strings like "RunnerRunning()"
-        if status_type.endswith("()"):
-            status_type = status_type[:-2]
-        states.add(STATUS_MAP.get(status_type, "UNKNOWN"))
-
-    # Priority order for overall state
-    if "FAILED" in states:
-        return "FAILED"
-    if "RUNNING" in states:
-        return "RUNNING"
-    if "CONFIGURING" in states:
-        return "CONFIGURING"
-    if "COMPLETING" in states:
-        return "COMPLETING"
-    if "COMPLETED" in states:
-        return "COMPLETED"
-    if "PENDING" in states:
-        return "PENDING"
-    return "UNKNOWN"
-
-
-def main(args: list[str]) -> int:
-    """Main entry point for squeue command."""
-    parser = argparse.ArgumentParser(
-        prog="exo squeue",
-        description="View the Exo job queue",
-    )
-    parser.add_argument(
-        "-l",
-        "--long",
-        action="store_true",
-        help="Show detailed output",
-    )
-    parser.add_argument(
-        "-j",
-        "--job",
-        help="Show only this job ID",
-    )
-
-    parsed = parser.parse_args(args)
-
-    # Extract typed values
-    long_format: bool = parsed.long  # pyright: ignore[reportAny]
-    job_filter: str | None = parsed.job  # pyright: ignore[reportAny]
-
-    # Fetch jobs from API - returns list directly
-    result = api_request("GET", "/flash/instances")
-    # API returns list directly, not {"instances": [...]}
-    if isinstance(result, list):
-        instances = cast(list[dict[str, Any]], result)
-    else:
-        instances = cast(list[dict[str, Any]], result.get("instances", []))
-
-    if not instances:
-        # No jobs - just print header
-        if long_format:
-            print("JOBID           NAME            NODES  RANKS  STATE        WORKDIR")
-        else:
-            print("JOBID       NAME            NODES  STATE")
-        return 0
-
-    # Filter by job ID if specified
-    if job_filter:
-        search = job_filter.lower()
-        filtered: list[dict[str, Any]] = []
-        for i in instances:
-            iid = i.get("instance_id", "")  # pyright: ignore[reportAny]
-            if search in str(iid).lower().replace("-", ""):  # pyright: ignore[reportAny]
-                filtered.append(i)
-        instances = filtered
-
-    # Build table
-    rows: list[list[str]] = []
-
-    if long_format:
-        headers = ["JOBID", "NAME", "NODES", "RANKS", "STATE", "WORKDIR"]
-        for inst in instances:
-            iid_val = inst.get("instance_id", "")  # pyright: ignore[reportAny]
-            instance_id = str(iid_val) if iid_val else ""  # pyright: ignore[reportAny]
-            job_id = truncate_id(instance_id, 12)
-            name_val = inst.get("simulation_name", "")  # pyright: ignore[reportAny]
-            name = (str(name_val) if name_val else "")[:15]  # pyright: ignore[reportAny]
-            runner_statuses = cast(dict[str, Any], inst.get("runner_statuses", {}))
-            nodes = str(len(runner_statuses))
-            ranks_val = inst.get("total_ranks", 0)  # pyright: ignore[reportAny]
-            ranks = str(ranks_val) if ranks_val else "0"  # pyright: ignore[reportAny]
-            state = get_job_state(runner_statuses)
-            workdir_val = inst.get("working_directory", "")  # pyright: ignore[reportAny]
-            workdir = str(workdir_val) if workdir_val else ""  # pyright: ignore[reportAny]
-            # Truncate workdir for display
-            if len(workdir) > 30:
-                workdir = "..." + workdir[-27:]
-            rows.append([job_id, name, nodes, ranks, state, workdir])
-    else:
-        headers = ["JOBID", "NAME", "NODES", "STATE"]
-        for inst in instances:
-            iid_val = inst.get("instance_id", "")  # pyright: ignore[reportAny]
-            instance_id = str(iid_val) if iid_val else ""  # pyright: ignore[reportAny]
-            job_id = truncate_id(instance_id, 8)
-            name_val = inst.get("simulation_name", "")  # pyright: ignore[reportAny]
-            name = (str(name_val) if name_val else "")[:15]  # pyright: ignore[reportAny]
-            runner_statuses = cast(dict[str, Any], inst.get("runner_statuses", {}))
-            nodes = str(len(runner_statuses))
-            state = get_job_state(runner_statuses)
-            rows.append([job_id, name, nodes, state])
-
-    print(format_table(headers, rows))
-    return 0
-
-
-if __name__ == "__main__":
-    sys.exit(main(sys.argv[1:]))

src/exo/download/coordinator.py

@@ -0,0 +1,284 @@
+import asyncio
+from dataclasses import dataclass, field
+from typing import Iterator
+
+import anyio
+from anyio import current_time
+from anyio.abc import TaskGroup
+from loguru import logger
+
+from exo.download.download_utils import (
+    RepoDownloadProgress,
+    delete_model,
+    map_repo_download_progress_to_download_progress_data,
+)
+from exo.download.shard_downloader import ShardDownloader
+from exo.shared.models.model_cards import ModelId
+from exo.shared.types.commands import (
+    DeleteDownload,
+    ForwarderDownloadCommand,
+    StartDownload,
+)
+from exo.shared.types.common import NodeId, SessionId
+from exo.shared.types.events import (
+    Event,
+    ForwarderEvent,
+    NodeDownloadProgress,
+)
+from exo.shared.types.worker.downloads import (
+    DownloadCompleted,
+    DownloadFailed,
+    DownloadOngoing,
+    DownloadPending,
+    DownloadProgress,
+)
+from exo.shared.types.worker.shards import ShardMetadata
+from exo.utils.channels import Receiver, Sender, channel
+
+
+@dataclass
+class DownloadCoordinator:
+    node_id: NodeId
+    session_id: SessionId
+    shard_downloader: ShardDownloader
+    download_command_receiver: Receiver[ForwarderDownloadCommand]
+    local_event_sender: Sender[ForwarderEvent]
+    event_index_counter: Iterator[int]
+
+    # Local state
+    download_status: dict[ModelId, DownloadProgress] = field(default_factory=dict)
+    active_downloads: dict[ModelId, asyncio.Task[None]] = field(default_factory=dict)
+
+    # Internal event channel for forwarding (initialized in __post_init__)
+    event_sender: Sender[Event] = field(init=False)
+    event_receiver: Receiver[Event] = field(init=False)
+    _tg: TaskGroup = field(init=False)
+
+    def __post_init__(self) -> None:
+        self.event_sender, self.event_receiver = channel[Event]()
+        self._tg = anyio.create_task_group()
+
+    async def run(self) -> None:
+        logger.info("Starting DownloadCoordinator")
+        async with self._tg as tg:
+            tg.start_soon(self._command_processor)
+            tg.start_soon(self._forward_events)
+            tg.start_soon(self._emit_existing_download_progress)
+
+    def shutdown(self) -> None:
+        self._tg.cancel_scope.cancel()
+
+    async def _command_processor(self) -> None:
+        with self.download_command_receiver as commands:
+            async for cmd in commands:
+                # Only process commands targeting this node
+                if cmd.command.target_node_id != self.node_id:
+                    continue
+
+                match cmd.command:
+                    case StartDownload(shard_metadata=shard):
+                        await self._start_download(shard)
+                    case DeleteDownload(model_id=model_id):
+                        await self._delete_download(model_id)
+
+    async def _start_download(self, shard: ShardMetadata) -> None:
+        model_id = shard.model_card.model_id
+
+        # Check if already downloading or complete
+        if model_id in self.download_status:
+            status = self.download_status[model_id]
+            if isinstance(status, (DownloadOngoing, DownloadCompleted)):
+                logger.debug(
+                    f"Download for {model_id} already in progress or complete, skipping"
+                )
+                return
+
+        # Emit pending status
+        progress = DownloadPending(shard_metadata=shard, node_id=self.node_id)
+        self.download_status[model_id] = progress
+        await self.event_sender.send(NodeDownloadProgress(download_progress=progress))
+
+        # Check initial status from downloader
+        initial_progress = (
+            await self.shard_downloader.get_shard_download_status_for_shard(shard)
+        )
+
+        if initial_progress.status == "complete":
+            completed = DownloadCompleted(
+                shard_metadata=shard,
+                node_id=self.node_id,
+                total_bytes=initial_progress.total_bytes,
+            )
+            self.download_status[model_id] = completed
+            await self.event_sender.send(
+                NodeDownloadProgress(download_progress=completed)
+            )
+            return
+
+        # Start actual download
+        self._start_download_task(shard, initial_progress)
+
+    def _start_download_task(
+        self, shard: ShardMetadata, initial_progress: RepoDownloadProgress
+    ) -> None:
+        model_id = shard.model_card.model_id
+
+        # Emit ongoing status
+        status = DownloadOngoing(
+            node_id=self.node_id,
+            shard_metadata=shard,
+            download_progress=map_repo_download_progress_to_download_progress_data(
+                initial_progress
+            ),
+        )
+        self.download_status[model_id] = status
+        self.event_sender.send_nowait(NodeDownloadProgress(download_progress=status))
+
+        last_progress_time = 0.0
+        throttle_interval_secs = 1.0
+
+        async def download_progress_callback(
+            callback_shard: ShardMetadata, progress: RepoDownloadProgress
+        ) -> None:
+            nonlocal last_progress_time
+
+            if progress.status == "complete":
+                completed = DownloadCompleted(
+                    shard_metadata=callback_shard,
+                    node_id=self.node_id,
+                    total_bytes=progress.total_bytes,
+                )
+                self.download_status[callback_shard.model_card.model_id] = completed
+                await self.event_sender.send(
+                    NodeDownloadProgress(download_progress=completed)
+                )
+                # Clean up active download tracking
+                if callback_shard.model_card.model_id in self.active_downloads:
+                    del self.active_downloads[callback_shard.model_card.model_id]
+            elif (
+                progress.status == "in_progress"
+                and current_time() - last_progress_time > throttle_interval_secs
+            ):
+                ongoing = DownloadOngoing(
+                    node_id=self.node_id,
+                    shard_metadata=callback_shard,
+                    download_progress=map_repo_download_progress_to_download_progress_data(
+                        progress
+                    ),
+                )
+                self.download_status[callback_shard.model_card.model_id] = ongoing
+                await self.event_sender.send(
+                    NodeDownloadProgress(download_progress=ongoing)
+                )
+                last_progress_time = current_time()
+
+        self.shard_downloader.on_progress(download_progress_callback)
+
+        async def download_wrapper() -> None:
+            try:
+                await self.shard_downloader.ensure_shard(shard)
+            except Exception as e:
+                logger.error(f"Download failed for {model_id}: {e}")
+                failed = DownloadFailed(
+                    shard_metadata=shard,
+                    node_id=self.node_id,
+                    error_message=str(e),
+                )
+                self.download_status[model_id] = failed
+                await self.event_sender.send(
+                    NodeDownloadProgress(download_progress=failed)
+                )
+            finally:
+                if model_id in self.active_downloads:
+                    del self.active_downloads[model_id]
+
+        task = asyncio.create_task(download_wrapper())
+        self.active_downloads[model_id] = task
+
+    async def _delete_download(self, model_id: ModelId) -> None:
+        # Cancel if active
+        if model_id in self.active_downloads:
+            logger.info(f"Cancelling active download for {model_id} before deletion")
+            self.active_downloads[model_id].cancel()
+            del self.active_downloads[model_id]
+
+        # Delete from disk
+        logger.info(f"Deleting model files for {model_id}")
+        deleted = await delete_model(model_id)
+
+        if deleted:
+            logger.info(f"Successfully deleted model {model_id}")
+        else:
+            logger.warning(f"Model {model_id} was not found on disk")
+
+        # Emit pending status to reset UI state, then remove from local tracking
+        if model_id in self.download_status:
+            current_status = self.download_status[model_id]
+            pending = DownloadPending(
+                shard_metadata=current_status.shard_metadata,
+                node_id=self.node_id,
+            )
+            await self.event_sender.send(
+                NodeDownloadProgress(download_progress=pending)
+            )
+            del self.download_status[model_id]
+
+    async def _forward_events(self) -> None:
+        with self.event_receiver as events:
+            async for event in events:
+                idx = next(self.event_index_counter)
+                fe = ForwarderEvent(
+                    origin_idx=idx,
+                    origin=self.node_id,
+                    session=self.session_id,
+                    event=event,
+                )
+                logger.debug(
+                    f"DownloadCoordinator published event {idx}: {str(event)[:100]}"
+                )
+                await self.local_event_sender.send(fe)
+
+    async def _emit_existing_download_progress(self) -> None:
+        try:
+            while True:
+                logger.info(
+                    "DownloadCoordinator: Fetching and emitting existing download progress..."
+                )
+                async for (
+                    _,
+                    progress,
+                ) in self.shard_downloader.get_shard_download_status():
+                    if progress.status == "complete":
+                        status: DownloadProgress = DownloadCompleted(
+                            node_id=self.node_id,
+                            shard_metadata=progress.shard,
+                            total_bytes=progress.total_bytes,
+                        )
+                    elif progress.status in ["in_progress", "not_started"]:
+                        if progress.downloaded_bytes_this_session.in_bytes == 0:
+                            status = DownloadPending(
+                                node_id=self.node_id, shard_metadata=progress.shard
+                            )
+                        else:
+                            status = DownloadOngoing(
+                                node_id=self.node_id,
+                                shard_metadata=progress.shard,
+                                download_progress=map_repo_download_progress_to_download_progress_data(
+                                    progress
+                                ),
+                            )
+                    else:
+                        continue
+
+                    self.download_status[progress.shard.model_card.model_id] = status
+                    await self.event_sender.send(
+                        NodeDownloadProgress(download_progress=status)
+                    )
+                logger.info(
+                    "DownloadCoordinator: Done emitting existing download progress."
+                )
+                await anyio.sleep(5 * 60)  # 5 minutes
+        except Exception as e:
+            logger.error(
+                f"DownloadCoordinator: Error emitting existing download progress: {e}"
+            )

src/exo/download/download_utils.py

@@ -24,7 +24,15 @@ from pydantic import (
     TypeAdapter,
 )
 
+from exo.download.huggingface_utils import (
+    filter_repo_objects,
+    get_allow_patterns,
+    get_auth_headers,
+    get_hf_endpoint,
+    get_hf_token,
+)
 from exo.shared.constants import EXO_MODELS_DIR
+from exo.shared.models.model_cards import ModelTask
 from exo.shared.types.common import ModelId
 from exo.shared.types.memory import Memory
 from exo.shared.types.worker.downloads import (
@@ -35,12 +43,27 @@ from exo.shared.types.worker.downloads import (
     RepoFileDownloadProgress,
 )
 from exo.shared.types.worker.shards import ShardMetadata
-from exo.worker.download.huggingface_utils import (
-    filter_repo_objects,
-    get_allow_patterns,
-    get_auth_headers,
-    get_hf_endpoint,
-)
+
+
+class HuggingFaceAuthenticationError(Exception):
+    """Raised when HuggingFace returns 401/403 for a model download."""
+
+
+async def _build_auth_error_message(status_code: int, model_id: ModelId) -> str:
+    token = await get_hf_token()
+    if status_code == 401 and token is None:
+        return (
+            f"Model '{model_id}' requires authentication. "
+            f"Set HF_TOKEN in the app's Advanced settings, set the HF_TOKEN environment variable, or run `hf auth login`. "
+            f"Get a token at https://huggingface.co/settings/tokens"
+        )
+    elif status_code == 403:
+        return (
+            f"Access denied to '{model_id}'. "
+            f"Please accept the model terms at https://huggingface.co/{model_id}"
+        )
+    else:
+        return f"Authentication failed for '{model_id}' (HTTP {status_code})"
 
 
 def trim_etag(etag: str) -> str:
@@ -98,11 +121,20 @@ async def ensure_models_dir() -> Path:
 
 
 async def delete_model(model_id: ModelId) -> bool:
-    model_dir = await ensure_models_dir() / model_id.normalize()
-    if not await aios.path.exists(model_dir):
-        return False
-    await asyncio.to_thread(shutil.rmtree, model_dir, ignore_errors=False)
-    return True
+    models_dir = await ensure_models_dir()
+    model_dir = models_dir / model_id.normalize()
+    cache_dir = models_dir / "caches" / model_id.normalize()
+
+    deleted = False
+    if await aios.path.exists(model_dir):
+        await asyncio.to_thread(shutil.rmtree, model_dir, ignore_errors=False)
+        deleted = True
+
+    # Also clear cache
+    if await aios.path.exists(cache_dir):
+        await asyncio.to_thread(shutil.rmtree, cache_dir, ignore_errors=False)
+
+    return deleted
 
 
 async def seed_models(seed_dir: str | Path):
@@ -128,16 +160,28 @@ async def fetch_file_list_with_cache(
     target_dir = (await ensure_models_dir()) / "caches" / model_id.normalize()
     await aios.makedirs(target_dir, exist_ok=True)
     cache_file = target_dir / f"{model_id.normalize()}--{revision}--file_list.json"
-    if await aios.path.exists(cache_file):
-        async with aiofiles.open(cache_file, "r") as f:
-            return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
-    file_list = await fetch_file_list_with_retry(
-        model_id, revision, recursive=recursive
-    )
-    await aios.makedirs(cache_file.parent, exist_ok=True)
-    async with aiofiles.open(cache_file, "w") as f:
-        await f.write(TypeAdapter(list[FileListEntry]).dump_json(file_list).decode())
-    return file_list
+
+    # Always try fresh first
+    try:
+        file_list = await fetch_file_list_with_retry(
+            model_id, revision, recursive=recursive
+        )
+        # Update cache with fresh data
+        async with aiofiles.open(cache_file, "w") as f:
+            await f.write(
+                TypeAdapter(list[FileListEntry]).dump_json(file_list).decode()
+            )
+        return file_list
+    except Exception as e:
+        # Fetch failed - try cache fallback
+        if await aios.path.exists(cache_file):
+            logger.warning(
+                f"Failed to fetch file list for {model_id}, using cached data: {e}"
+            )
+            async with aiofiles.open(cache_file, "r") as f:
+                return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
+        # No cache available, propagate the error
+        raise
 
 
 async def fetch_file_list_with_retry(
@@ -147,6 +191,8 @@ async def fetch_file_list_with_retry(
     for attempt in range(n_attempts):
         try:
             return await _fetch_file_list(model_id, revision, path, recursive)
+        except HuggingFaceAuthenticationError:
+            raise
         except Exception as e:
             if attempt == n_attempts - 1:
                 raise e
@@ -167,6 +213,9 @@ async def _fetch_file_list(
         create_http_session(timeout_profile="short") as session,
         session.get(url, headers=headers) as response,
     ):
+        if response.status in [401, 403]:
+            msg = await _build_auth_error_message(response.status, model_id)
+            raise HuggingFaceAuthenticationError(msg)
         if response.status == 200:
             data_json = await response.text()
             data = TypeAdapter(list[FileListEntry]).validate_json(data_json)
@@ -256,6 +305,9 @@ async def file_meta(
             # Otherwise, follow the redirect to get authoritative size/hash
             redirected_location = r.headers.get("location")
             return await file_meta(model_id, revision, path, redirected_location)
+        if r.status in [401, 403]:
+            msg = await _build_auth_error_message(r.status, model_id)
+            raise HuggingFaceAuthenticationError(msg)
         content_length = int(
             r.headers.get("x-linked-size") or r.headers.get("content-length") or 0
         )
@@ -279,6 +331,8 @@ async def download_file_with_retry(
             return await _download_file(
                 model_id, revision, path, target_dir, on_progress
             )
+        except HuggingFaceAuthenticationError:
+            raise
         except Exception as e:
             if isinstance(e, FileNotFoundError) or attempt == n_attempts - 1:
                 raise e
@@ -299,8 +353,28 @@ async def _download_file(
     target_dir: Path,
     on_progress: Callable[[int, int, bool], None] = lambda _, __, ___: None,
 ) -> Path:
-    if await aios.path.exists(target_dir / path):
-        return target_dir / path
+    target_path = target_dir / path
+
+    if await aios.path.exists(target_path):
+        local_size = (await aios.stat(target_path)).st_size
+
+        # Try to verify against remote, but allow offline operation
+        try:
+            remote_size, _ = await file_meta(model_id, revision, path)
+            if local_size != remote_size:
+                logger.info(
+                    f"File {path} size mismatch (local={local_size}, remote={remote_size}), re-downloading"
+                )
+                await aios.remove(target_path)
+            else:
+                return target_path
+        except Exception as e:
+            # Offline or network error - trust local file
+            logger.debug(
+                f"Could not verify {path} against remote (offline?): {e}, using local file"
+            )
+            return target_path
+
     await aios.makedirs((target_dir / path).parent, exist_ok=True)
     length, etag = await file_meta(model_id, revision, path)
     remote_hash = etag[:-5] if etag.endswith("-gzip") else etag
@@ -322,6 +396,9 @@ async def _download_file(
         ):
             if r.status == 404:
                 raise FileNotFoundError(f"File not found: {url}")
+            if r.status in [401, 403]:
+                msg = await _build_auth_error_message(r.status, model_id)
+                raise HuggingFaceAuthenticationError(msg)
             assert r.status in [200, 206], (
                 f"Failed to download {path} from {url}: {r.status}"
             )
@@ -446,6 +523,11 @@ async def resolve_allow_patterns(shard: ShardMetadata) -> list[str]:
         return ["*"]
 
 
+def is_image_model(shard: ShardMetadata) -> bool:
+    tasks = shard.model_card.tasks
+    return ModelTask.TextToImage in tasks or ModelTask.ImageToImage in tasks
+
+
 async def get_downloaded_size(path: Path) -> int:
     partial_path = path.with_suffix(path.suffix + ".partial")
     if await aios.path.exists(path):
@@ -463,7 +545,7 @@ async def download_shard(
     allow_patterns: list[str] | None = None,
 ) -> tuple[Path, RepoDownloadProgress]:
     if not skip_download:
-        logger.info(f"Downloading {shard.model_card.model_id=}")
+        logger.debug(f"Downloading {shard.model_card.model_id=}")
 
     revision = "main"
     target_dir = await ensure_models_dir() / str(shard.model_card.model_id).replace(
@@ -476,7 +558,7 @@ async def download_shard(
         allow_patterns = await resolve_allow_patterns(shard)
 
     if not skip_download:
-        logger.info(f"Downloading {shard.model_card.model_id=} with {allow_patterns=}")
+        logger.debug(f"Downloading {shard.model_card.model_id=} with {allow_patterns=}")
 
     all_start_time = time.time()
     file_list = await fetch_file_list_with_cache(
@@ -487,22 +569,40 @@ async def download_shard(
             file_list, allow_patterns=allow_patterns, key=lambda x: x.path
         )
     )
+
+    # For image models, skip root-level safetensors files since weights
+    # are stored in component subdirectories (e.g., transformer/, vae/)
+    if is_image_model(shard):
+        filtered_file_list = [
+            f
+            for f in filtered_file_list
+            if "/" in f.path or not f.path.endswith(".safetensors")
+        ]
     file_progress: dict[str, RepoFileDownloadProgress] = {}
 
     async def on_progress_wrapper(
         file: FileListEntry, curr_bytes: int, total_bytes: int, is_renamed: bool
     ) -> None:
-        start_time = (
-            file_progress[file.path].start_time
-            if file.path in file_progress
-            else time.time()
-        )
-        downloaded_this_session = (
-            file_progress[file.path].downloaded_this_session.in_bytes
-            + (curr_bytes - file_progress[file.path].downloaded.in_bytes)
-            if file.path in file_progress
-            else curr_bytes
+        previous_progress = file_progress.get(file.path)
+
+        # Detect re-download: curr_bytes < previous downloaded means file was deleted and restarted
+        is_redownload = (
+            previous_progress is not None
+            and curr_bytes < previous_progress.downloaded.in_bytes
         )
+
+        if is_redownload or previous_progress is None:
+            # Fresh download or re-download: reset tracking
+            start_time = time.time()
+            downloaded_this_session = curr_bytes
+        else:
+            # Continuing download: accumulate
+            start_time = previous_progress.start_time
+            downloaded_this_session = (
+                previous_progress.downloaded_this_session.in_bytes
+                + (curr_bytes - previous_progress.downloaded.in_bytes)
+            )
+
         speed = (
             downloaded_this_session / (time.time() - start_time)
             if time.time() - start_time > 0

src/exo/download/huggingface_utils.py

@@ -68,7 +68,11 @@ def get_hf_home() -> Path:
 
 
 async def get_hf_token() -> str | None:
-    """Retrieve the Hugging Face token from the user's HF_HOME directory."""
+    """Retrieve the Hugging Face token from HF_TOKEN env var or HF_HOME directory."""
+    # Check environment variable first
+    if token := os.environ.get("HF_TOKEN"):
+        return token
+    # Fall back to file-based token
     token_path = get_hf_home() / "token"
     if await aios.path.exists(token_path):
         async with aiofiles.open(token_path, "r") as f:

src/exo/download/impl_shard_downloader.py

@@ -3,13 +3,15 @@ from collections.abc import Awaitable
 from pathlib import Path
 from typing import AsyncIterator, Callable
 
+from loguru import logger
+
+from exo.download.download_utils import RepoDownloadProgress, download_shard
+from exo.download.shard_downloader import ShardDownloader
 from exo.shared.models.model_cards import MODEL_CARDS, ModelCard, ModelId
 from exo.shared.types.worker.shards import (
     PipelineShardMetadata,
     ShardMetadata,
 )
-from exo.worker.download.download_utils import RepoDownloadProgress, download_shard
-from exo.worker.download.shard_downloader import ShardDownloader
 
 
 def exo_shard_downloader(max_parallel_downloads: int = 8) -> ShardDownloader:
@@ -164,9 +166,8 @@ class ResumableShardDownloader(ShardDownloader):
         for task in asyncio.as_completed(tasks):
             try:
                 yield await task
-            # TODO: except Exception
             except Exception as e:
-                print("Error downloading shard:", e)
+                logger.warning(f"Error downloading shard: {type(e).__name__}")
 
     async def get_shard_download_status_for_shard(
         self, shard: ShardMetadata

src/exo/download/shard_downloader.py

@@ -5,13 +5,13 @@ from datetime import timedelta
 from pathlib import Path
 from typing import AsyncIterator, Callable
 
+from exo.download.download_utils import RepoDownloadProgress
 from exo.shared.models.model_cards import ModelCard, ModelId, ModelTask
 from exo.shared.types.memory import Memory
 from exo.shared.types.worker.shards import (
     PipelineShardMetadata,
     ShardMetadata,
 )
-from exo.worker.download.download_utils import RepoDownloadProgress
 
 
 # TODO: the PipelineShardMetadata getting reinstantiated is a bit messy. Should this be a classmethod?

src/exo/download/tests/test_download_verification.py

@@ -0,0 +1,451 @@
+"""Tests for download verification and cache behavior."""
+
+import time
+from collections.abc import AsyncIterator
+from datetime import timedelta
+from pathlib import Path
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import aiofiles
+import aiofiles.os as aios
+import pytest
+from pydantic import TypeAdapter
+
+from exo.download.download_utils import (
+    delete_model,
+    fetch_file_list_with_cache,
+)
+from exo.shared.types.common import ModelId
+from exo.shared.types.memory import Memory
+from exo.shared.types.worker.downloads import FileListEntry, RepoFileDownloadProgress
+
+
+@pytest.fixture
+def model_id() -> ModelId:
+    return ModelId("test-org/test-model")
+
+
+@pytest.fixture
+async def temp_models_dir(tmp_path: Path) -> AsyncIterator[Path]:
+    """Set up a temporary models directory for testing."""
+    models_dir = tmp_path / "models"
+    await aios.makedirs(models_dir, exist_ok=True)
+    with patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir):
+        yield models_dir
+
+
+class TestFileVerification:
+    """Tests for file size verification in _download_file."""
+
+    async def test_redownload_when_file_size_changes_upstream(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that files with mismatched sizes are re-downloaded."""
+        # Import inside test to allow patching
+        from exo.download.download_utils import (
+            _download_file,  # pyright: ignore[reportPrivateUsage]
+        )
+
+        target_dir = tmp_path / "downloads"
+        await aios.makedirs(target_dir, exist_ok=True)
+
+        # Create a local file with wrong size
+        local_file = target_dir / "test.safetensors"
+        async with aiofiles.open(local_file, "wb") as f:
+            await f.write(b"local content")  # 13 bytes
+
+        remote_size = 1000  # Different from local
+        remote_hash = "abc123"
+
+        with (
+            patch(
+                "exo.download.download_utils.file_meta",
+                new_callable=AsyncMock,
+                return_value=(remote_size, remote_hash),
+            ) as mock_file_meta,
+            patch(
+                "exo.download.download_utils.create_http_session"
+            ) as mock_session_factory,
+        ):
+            # Set up mock HTTP response for re-download
+            mock_response = MagicMock()
+            mock_response.status = 200
+            mock_response.content.read = AsyncMock(  # pyright: ignore[reportAny]
+                side_effect=[b"x" * remote_size, b""]
+            )
+
+            mock_session = MagicMock()
+            mock_session.get.return_value.__aenter__ = AsyncMock(  # pyright: ignore[reportAny]
+                return_value=mock_response
+            )
+            mock_session.get.return_value.__aexit__ = AsyncMock(  # pyright: ignore[reportAny]
+                return_value=None
+            )
+            mock_session_factory.return_value.__aenter__ = AsyncMock(  # pyright: ignore[reportAny]
+                return_value=mock_session
+            )
+            mock_session_factory.return_value.__aexit__ = AsyncMock(  # pyright: ignore[reportAny]
+                return_value=None
+            )
+
+            # Mock calc_hash to return the expected hash
+            with patch(
+                "exo.download.download_utils.calc_hash",
+                new_callable=AsyncMock,
+                return_value=remote_hash,
+            ):
+                await _download_file(model_id, "main", "test.safetensors", target_dir)
+
+            # file_meta should be called twice: once for verification, once for download
+            assert mock_file_meta.call_count == 2
+
+    async def test_skip_download_when_file_size_matches(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that files with matching sizes are not re-downloaded."""
+        from exo.download.download_utils import (
+            _download_file,  # pyright: ignore[reportPrivateUsage]
+        )
+
+        target_dir = tmp_path / "downloads"
+        await aios.makedirs(target_dir, exist_ok=True)
+
+        # Create a local file
+        local_file = target_dir / "test.safetensors"
+        local_content = b"local content"
+        async with aiofiles.open(local_file, "wb") as f:
+            await f.write(local_content)
+
+        remote_size = len(local_content)  # Same as local
+        remote_hash = "abc123"
+
+        with (
+            patch(
+                "exo.download.download_utils.file_meta",
+                new_callable=AsyncMock,
+                return_value=(remote_size, remote_hash),
+            ) as mock_file_meta,
+            patch(
+                "exo.download.download_utils.create_http_session"
+            ) as mock_session_factory,
+        ):
+            result = await _download_file(
+                model_id, "main", "test.safetensors", target_dir
+            )
+
+            # Should return immediately without downloading
+            assert result == local_file
+            mock_file_meta.assert_called_once()
+            mock_session_factory.assert_not_called()
+
+    async def test_offline_fallback_uses_local_file(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that local files are used when network is unavailable."""
+        from exo.download.download_utils import (
+            _download_file,  # pyright: ignore[reportPrivateUsage]
+        )
+
+        target_dir = tmp_path / "downloads"
+        await aios.makedirs(target_dir, exist_ok=True)
+
+        # Create a local file
+        local_file = target_dir / "test.safetensors"
+        async with aiofiles.open(local_file, "wb") as f:
+            await f.write(b"local content")
+
+        with (
+            patch(
+                "exo.download.download_utils.file_meta",
+                new_callable=AsyncMock,
+                side_effect=Exception("Network error"),
+            ),
+            patch(
+                "exo.download.download_utils.create_http_session"
+            ) as mock_session_factory,
+        ):
+            result = await _download_file(
+                model_id, "main", "test.safetensors", target_dir
+            )
+
+            # Should return local file without attempting download
+            assert result == local_file
+            mock_session_factory.assert_not_called()
+
+
+class TestFileListCache:
+    """Tests for file list caching behavior."""
+
+    async def test_fetch_fresh_and_update_cache(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that fresh data is fetched and cache is updated."""
+        models_dir = tmp_path / "models"
+
+        file_list = [
+            FileListEntry(type="file", path="model.safetensors", size=1000),
+            FileListEntry(type="file", path="config.json", size=100),
+        ]
+
+        with (
+            patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir),
+            patch(
+                "exo.download.download_utils.fetch_file_list_with_retry",
+                new_callable=AsyncMock,
+                return_value=file_list,
+            ) as mock_fetch,
+        ):
+            result = await fetch_file_list_with_cache(model_id, "main")
+
+            assert result == file_list
+            mock_fetch.assert_called_once()
+
+            # Verify cache was written
+            cache_file = (
+                models_dir
+                / "caches"
+                / model_id.normalize()
+                / f"{model_id.normalize()}--main--file_list.json"
+            )
+            assert await aios.path.exists(cache_file)
+
+            async with aiofiles.open(cache_file, "r") as f:
+                cached_data = TypeAdapter(list[FileListEntry]).validate_json(
+                    await f.read()
+                )
+            assert cached_data == file_list
+
+    async def test_fallback_to_cache_when_fetch_fails(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that cached data is used when fetch fails."""
+        models_dir = tmp_path / "models"
+        cache_dir = models_dir / "caches" / model_id.normalize()
+        await aios.makedirs(cache_dir, exist_ok=True)
+
+        # Create cache file
+        cached_file_list = [
+            FileListEntry(type="file", path="model.safetensors", size=1000),
+        ]
+        cache_file = cache_dir / f"{model_id.normalize()}--main--file_list.json"
+        async with aiofiles.open(cache_file, "w") as f:
+            await f.write(
+                TypeAdapter(list[FileListEntry]).dump_json(cached_file_list).decode()
+            )
+
+        with (
+            patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir),
+            patch(
+                "exo.download.download_utils.fetch_file_list_with_retry",
+                new_callable=AsyncMock,
+                side_effect=Exception("Network error"),
+            ),
+        ):
+            result = await fetch_file_list_with_cache(model_id, "main")
+
+            assert result == cached_file_list
+
+    async def test_error_propagates_when_no_cache(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that errors propagate when fetch fails and no cache exists."""
+        models_dir = tmp_path / "models"
+
+        with (
+            patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir),
+            patch(
+                "exo.download.download_utils.fetch_file_list_with_retry",
+                new_callable=AsyncMock,
+                side_effect=Exception("Network error"),
+            ),
+            pytest.raises(Exception, match="Network error"),
+        ):
+            await fetch_file_list_with_cache(model_id, "main")
+
+
+class TestModelDeletion:
+    """Tests for model deletion including cache cleanup."""
+
+    async def test_delete_model_clears_cache(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that deleting a model also deletes its cache."""
+        models_dir = tmp_path / "models"
+        model_dir = models_dir / model_id.normalize()
+        cache_dir = models_dir / "caches" / model_id.normalize()
+
+        # Create model and cache directories
+        await aios.makedirs(model_dir, exist_ok=True)
+        await aios.makedirs(cache_dir, exist_ok=True)
+
+        # Add some files
+        async with aiofiles.open(model_dir / "model.safetensors", "w") as f:
+            await f.write("model data")
+        async with aiofiles.open(cache_dir / "file_list.json", "w") as f:
+            await f.write("[]")
+
+        with patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir):
+            result = await delete_model(model_id)
+
+            assert result is True
+            assert not await aios.path.exists(model_dir)
+            assert not await aios.path.exists(cache_dir)
+
+    async def test_delete_model_only_cache_exists(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test deleting when only cache exists (model already deleted)."""
+        models_dir = tmp_path / "models"
+        cache_dir = models_dir / "caches" / model_id.normalize()
+
+        # Only create cache directory
+        await aios.makedirs(cache_dir, exist_ok=True)
+        async with aiofiles.open(cache_dir / "file_list.json", "w") as f:
+            await f.write("[]")
+
+        with patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir):
+            result = await delete_model(model_id)
+
+            # Returns False because model dir didn't exist
+            assert result is False
+            # But cache should still be cleaned up
+            assert not await aios.path.exists(cache_dir)
+
+    async def test_delete_nonexistent_model(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test deleting a model that doesn't exist."""
+        models_dir = tmp_path / "models"
+        await aios.makedirs(models_dir, exist_ok=True)
+
+        with patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir):
+            result = await delete_model(model_id)
+
+            assert result is False
+
+
+class TestProgressResetOnRedownload:
+    """Tests for progress tracking when files are re-downloaded."""
+
+    async def test_progress_resets_correctly_on_redownload(
+        self, model_id: ModelId
+    ) -> None:
+        """Test that progress tracking resets when a file is re-downloaded.
+
+        When a file is deleted and re-downloaded (due to size mismatch),
+        the progress tracking should reset rather than calculating negative
+        downloaded_this_session values.
+        """
+        # Simulate file_progress dict as it exists in download_shard
+        file_progress: dict[str, RepoFileDownloadProgress] = {}
+
+        # Initialize with old file progress (simulating existing large file)
+        old_file_size = 1_500_000_000  # 1.5 GB
+        file_progress["model.safetensors"] = RepoFileDownloadProgress(
+            repo_id=model_id,
+            repo_revision="main",
+            file_path="model.safetensors",
+            downloaded=Memory.from_bytes(old_file_size),
+            downloaded_this_session=Memory.from_bytes(0),
+            total=Memory.from_bytes(old_file_size),
+            speed=0,
+            eta=timedelta(0),
+            status="not_started",
+            start_time=time.time() - 10,  # Started 10 seconds ago
+        )
+
+        # Simulate the logic from on_progress_wrapper after re-download starts
+        # This is the exact logic from the fixed on_progress_wrapper
+        curr_bytes = 100_000  # 100 KB - new download just started
+        previous_progress = file_progress.get("model.safetensors")
+
+        # Detect re-download: curr_bytes < previous downloaded
+        is_redownload = (
+            previous_progress is not None
+            and curr_bytes < previous_progress.downloaded.in_bytes
+        )
+
+        if is_redownload or previous_progress is None:
+            # Fresh download or re-download: reset tracking
+            start_time = time.time()
+            downloaded_this_session = curr_bytes
+        else:
+            # Continuing download: accumulate
+            start_time = previous_progress.start_time
+            downloaded_this_session = (
+                previous_progress.downloaded_this_session.in_bytes
+                + (curr_bytes - previous_progress.downloaded.in_bytes)
+            )
+
+        # Key assertions
+        assert is_redownload is True, "Should detect re-download scenario"
+        assert downloaded_this_session == curr_bytes, (
+            "downloaded_this_session should equal curr_bytes on re-download"
+        )
+        assert downloaded_this_session > 0, (
+            "downloaded_this_session should be positive, not negative"
+        )
+
+        # Calculate speed (should be positive)
+        elapsed = time.time() - start_time
+        speed = downloaded_this_session / elapsed if elapsed > 0 else 0
+        assert speed >= 0, "Speed should be non-negative"
+
+    async def test_progress_accumulates_on_continuing_download(
+        self, model_id: ModelId
+    ) -> None:
+        """Test that progress accumulates correctly for continuing downloads.
+
+        When a download continues from where it left off (resume),
+        the progress should accumulate correctly.
+        """
+        file_progress: dict[str, RepoFileDownloadProgress] = {}
+
+        # Initialize with partial download progress
+        initial_downloaded = 500_000  # 500 KB already downloaded
+        start_time = time.time() - 5  # Started 5 seconds ago
+        file_progress["model.safetensors"] = RepoFileDownloadProgress(
+            repo_id=model_id,
+            repo_revision="main",
+            file_path="model.safetensors",
+            downloaded=Memory.from_bytes(initial_downloaded),
+            downloaded_this_session=Memory.from_bytes(initial_downloaded),
+            total=Memory.from_bytes(1_000_000),
+            speed=100_000,
+            eta=timedelta(seconds=5),
+            status="in_progress",
+            start_time=start_time,
+        )
+
+        # Progress callback with more bytes downloaded
+        curr_bytes = 600_000  # 600 KB - continuing download
+        previous_progress = file_progress.get("model.safetensors")
+
+        # This is NOT a re-download (curr_bytes > previous downloaded)
+        is_redownload = (
+            previous_progress is not None
+            and curr_bytes < previous_progress.downloaded.in_bytes
+        )
+
+        if is_redownload or previous_progress is None:
+            downloaded_this_session = curr_bytes
+            used_start_time = time.time()
+        else:
+            used_start_time = previous_progress.start_time
+            downloaded_this_session = (
+                previous_progress.downloaded_this_session.in_bytes
+                + (curr_bytes - previous_progress.downloaded.in_bytes)
+            )
+
+        # Key assertions
+        assert is_redownload is False, (
+            "Should NOT detect re-download for continuing download"
+        )
+        assert used_start_time == start_time, "Should preserve original start_time"
+        expected_session = initial_downloaded + (curr_bytes - initial_downloaded)
+        assert downloaded_this_session == expected_session, (
+            f"Should accumulate: {downloaded_this_session} == {expected_session}"
+        )
+        assert downloaded_this_session == 600_000, (
+            "downloaded_this_session should equal total downloaded so far"
+        )

src/exo/main.py

@@ -1,10 +1,11 @@
 import argparse
+import itertools
 import multiprocessing as mp
 import os
 import resource
 import signal
 from dataclasses import dataclass, field
-from typing import Self
+from typing import Iterator, Self
 
 import anyio
 from anyio.abc import TaskGroup
@@ -12,6 +13,8 @@ from loguru import logger
 from pydantic import PositiveInt
 
 import exo.routing.topics as topics
+from exo.download.coordinator import DownloadCoordinator
+from exo.download.impl_shard_downloader import exo_shard_downloader
 from exo.master.api import API  # TODO: should API be in master?
 from exo.master.main import Master
 from exo.routing.router import Router, get_node_id_keypair
@@ -21,7 +24,6 @@ from exo.shared.logging import logger_cleanup, logger_setup
 from exo.shared.types.common import NodeId, SessionId
 from exo.utils.channels import Receiver, channel
 from exo.utils.pydantic_ext import CamelCaseModel
-from exo.worker.download.impl_shard_downloader import exo_shard_downloader
 from exo.worker.main import Worker
 
 
@@ -29,6 +31,7 @@ from exo.worker.main import Worker
 @dataclass
 class Node:
     router: Router
+    download_coordinator: DownloadCoordinator | None
     worker: Worker | None
     election: Election  # Every node participates in election, as we do want a node to become master even if it isn't a master candidate if no master candidates are present.
     election_result_receiver: Receiver[ElectionResult]
@@ -36,6 +39,7 @@ class Node:
     api: API | None
 
     node_id: NodeId
+    event_index_counter: Iterator[int]
     _tg: TaskGroup = field(init=False, default_factory=anyio.create_task_group)
 
     @classmethod
@@ -49,8 +53,26 @@ class Node:
         await router.register_topic(topics.COMMANDS)
         await router.register_topic(topics.ELECTION_MESSAGES)
         await router.register_topic(topics.CONNECTION_MESSAGES)
+        await router.register_topic(topics.DOWNLOAD_COMMANDS)
 
         logger.info(f"Starting node {node_id}")
+
+        # Create shared event index counter for Worker and DownloadCoordinator
+        event_index_counter = itertools.count()
+
+        # Create DownloadCoordinator (unless --no-downloads)
+        if not args.no_downloads:
+            download_coordinator = DownloadCoordinator(
+                node_id,
+                session_id,
+                exo_shard_downloader(),
+                download_command_receiver=router.receiver(topics.DOWNLOAD_COMMANDS),
+                local_event_sender=router.sender(topics.LOCAL_EVENTS),
+                event_index_counter=event_index_counter,
+            )
+        else:
+            download_coordinator = None
+
         if args.spawn_api:
             api = API(
                 node_id,
@@ -58,6 +80,7 @@ class Node:
                 port=args.api_port,
                 global_event_receiver=router.receiver(topics.GLOBAL_EVENTS),
                 command_sender=router.sender(topics.COMMANDS),
+                download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
                 election_receiver=router.receiver(topics.ELECTION_MESSAGES),
             )
         else:
@@ -67,11 +90,11 @@ class Node:
             worker = Worker(
                 node_id,
                 session_id,
-                exo_shard_downloader(),
-                connection_message_receiver=router.receiver(topics.CONNECTION_MESSAGES),
                 global_event_receiver=router.receiver(topics.GLOBAL_EVENTS),
                 local_event_sender=router.sender(topics.LOCAL_EVENTS),
                 command_sender=router.sender(topics.COMMANDS),
+                download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
+                event_index_counter=event_index_counter,
             )
         else:
             worker = None
@@ -99,13 +122,25 @@ class Node:
             election_result_sender=er_send,
         )
 
-        return cls(router, worker, election, er_recv, master, api, node_id)
+        return cls(
+            router,
+            download_coordinator,
+            worker,
+            election,
+            er_recv,
+            master,
+            api,
+            node_id,
+            event_index_counter,
+        )
 
     async def run(self):
         async with self._tg as tg:
             signal.signal(signal.SIGINT, lambda _, __: self.shutdown())
             tg.start_soon(self.router.run)
             tg.start_soon(self.election.run)
+            if self.download_coordinator:
+                tg.start_soon(self.download_coordinator.run)
             if self.worker:
                 tg.start_soon(self.worker.run)
             if self.master:
@@ -170,21 +205,36 @@ class Node:
                     )
                 if result.is_new_master:
                     await anyio.sleep(0)
+                    # Fresh counter for new session (buffer expects indices from 0)
+                    self.event_index_counter = itertools.count()
+                    if self.download_coordinator:
+                        self.download_coordinator.shutdown()
+                        self.download_coordinator = DownloadCoordinator(
+                            self.node_id,
+                            result.session_id,
+                            exo_shard_downloader(),
+                            download_command_receiver=self.router.receiver(
+                                topics.DOWNLOAD_COMMANDS
+                            ),
+                            local_event_sender=self.router.sender(topics.LOCAL_EVENTS),
+                            event_index_counter=self.event_index_counter,
+                        )
+                        self._tg.start_soon(self.download_coordinator.run)
                     if self.worker:
                         self.worker.shutdown()
                         # TODO: add profiling etc to resource monitor
                         self.worker = Worker(
                             self.node_id,
                             result.session_id,
-                            exo_shard_downloader(),
-                            connection_message_receiver=self.router.receiver(
-                                topics.CONNECTION_MESSAGES
-                            ),
                             global_event_receiver=self.router.receiver(
                                 topics.GLOBAL_EVENTS
                             ),
                             local_event_sender=self.router.sender(topics.LOCAL_EVENTS),
                             command_sender=self.router.sender(topics.COMMANDS),
+                            download_command_sender=self.router.sender(
+                                topics.DOWNLOAD_COMMANDS
+                            ),
+                            event_index_counter=self.event_index_counter,
                         )
                         self._tg.start_soon(self.worker.run)
                     if self.api:
@@ -195,14 +245,6 @@ class Node:
 
 
 def main():
-    # Check for SLURM-compatible subcommands first
-    import sys
-
-    if len(sys.argv) > 1 and sys.argv[1] in ("sbatch", "squeue", "scancel", "salloc"):
-        from exo.cli import run_subcommand
-
-        sys.exit(run_subcommand(sys.argv[1], sys.argv[2:]))
-
     args = Args.parse()
     soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
     resource.setrlimit(resource.RLIMIT_NOFILE, (max(soft, 65535), hard))
@@ -213,11 +255,6 @@ def main():
     logger.info("Starting EXO")
     logger.info(f"EXO_LIBP2P_NAMESPACE: {os.getenv('EXO_LIBP2P_NAMESPACE')}")
 
-    # Discover and register plugins
-    from exo.plugins.registry import discover_plugins
-
-    discover_plugins()
-
     # Set FAST_SYNCH override env var for runner subprocesses
     if args.fast_synch is True:
         os.environ["EXO_FAST_SYNCH"] = "on"
@@ -239,6 +276,7 @@ class Args(CamelCaseModel):
     api_port: PositiveInt = 52415
     tb_only: bool = False
     no_worker: bool = False
+    no_downloads: bool = False
     fast_synch: bool | None = None  # None = auto, True = force on, False = force off
 
     @classmethod
@@ -281,6 +319,11 @@ class Args(CamelCaseModel):
             "--no-worker",
             action="store_true",
         )
+        parser.add_argument(
+            "--no-downloads",
+            action="store_true",
+            help="Disable the download coordinator (node won't download models)",
+        )
         fast_synch_group = parser.add_mutually_exclusive_group()
         fast_synch_group.add_argument(
             "--fast-synch",

src/exo/master/adapters/__init__.py

@@ -0,0 +1 @@
+"""API adapters for different API formats (Claude, OpenAI Responses, etc.)."""

src/exo/master/adapters/chat_completions.py

@@ -0,0 +1,212 @@
+"""OpenAI Chat Completions API adapter for converting requests/responses."""
+
+import time
+from collections.abc import AsyncGenerator
+from typing import Any
+from uuid import uuid4
+
+from exo.shared.types.api import (
+    ChatCompletionChoice,
+    ChatCompletionMessage,
+    ChatCompletionMessageText,
+    ChatCompletionRequest,
+    ChatCompletionResponse,
+    ErrorInfo,
+    ErrorResponse,
+    FinishReason,
+    StreamingChoiceResponse,
+    ToolCall,
+)
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
+from exo.shared.types.common import CommandId
+from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
+
+
+def chat_request_to_text_generation(
+    request: ChatCompletionRequest,
+) -> TextGenerationTaskParams:
+    instructions: str | None = None
+    input_messages: list[InputMessage] = []
+    chat_template_messages: list[dict[str, Any]] = []
+
+    for msg in request.messages:
+        # Normalize content to string
+        content: str
+        if msg.content is None:
+            content = ""
+        elif isinstance(msg.content, str):
+            content = msg.content
+        elif isinstance(msg.content, ChatCompletionMessageText):
+            content = msg.content.text
+        else:
+            # List of ChatCompletionMessageText
+            content = "\n".join(item.text for item in msg.content)
+
+        # Extract system message as instructions
+        if msg.role == "system":
+            if instructions is None:
+                instructions = content
+            else:
+                # Append additional system messages
+                instructions = f"{instructions}\n{content}"
+            chat_template_messages.append({"role": "system", "content": content})
+        else:
+            # Skip messages with no meaningful content
+            if msg.content is None and msg.thinking is None and msg.tool_calls is None:
+                continue
+
+            if msg.role in ("user", "assistant", "developer"):
+                input_messages.append(InputMessage(role=msg.role, content=content))
+
+            # Build full message dict for chat template (preserves tool_calls etc.)
+            # Normalize content for model_dump
+            msg_copy = msg.model_copy(update={"content": content})
+            dumped: dict[str, Any] = msg_copy.model_dump(exclude_none=True)
+            chat_template_messages.append(dumped)
+
+    return TextGenerationTaskParams(
+        model=request.model,
+        input=input_messages if input_messages else "",
+        instructions=instructions,
+        max_output_tokens=request.max_tokens,
+        temperature=request.temperature,
+        top_p=request.top_p,
+        top_k=request.top_k,
+        stop=request.stop,
+        seed=request.seed,
+        stream=request.stream,
+        tools=request.tools,
+        chat_template_messages=chat_template_messages
+        if chat_template_messages
+        else None,
+    )
+
+
+def chunk_to_response(
+    chunk: TokenChunk, command_id: CommandId
+) -> ChatCompletionResponse:
+    """Convert a TokenChunk to a streaming ChatCompletionResponse."""
+    return ChatCompletionResponse(
+        id=command_id,
+        created=int(time.time()),
+        model=chunk.model,
+        choices=[
+            StreamingChoiceResponse(
+                index=0,
+                delta=ChatCompletionMessage(role="assistant", content=chunk.text),
+                finish_reason=chunk.finish_reason,
+            )
+        ],
+    )
+
+
+async def generate_chat_stream(
+    command_id: CommandId,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> AsyncGenerator[str, None]:
+    """Generate Chat Completions API streaming events from chunks."""
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            error_response = ErrorResponse(
+                error=ErrorInfo(
+                    message=chunk.error_message or "Internal server error",
+                    type="InternalServerError",
+                    code=500,
+                )
+            )
+            yield f"data: {error_response.model_dump_json()}\n\n"
+            yield "data: [DONE]\n\n"
+            return
+
+        if isinstance(chunk, ToolCallChunk):
+            tool_call_deltas = [
+                ToolCall(
+                    id=str(uuid4()),
+                    index=i,
+                    function=tool,
+                )
+                for i, tool in enumerate(chunk.tool_calls)
+            ]
+            tool_response = ChatCompletionResponse(
+                id=command_id,
+                created=int(time.time()),
+                model=chunk.model,
+                choices=[
+                    StreamingChoiceResponse(
+                        index=0,
+                        delta=ChatCompletionMessage(
+                            role="assistant",
+                            tool_calls=tool_call_deltas,
+                        ),
+                        finish_reason="tool_calls",
+                    )
+                ],
+            )
+            yield f"data: {tool_response.model_dump_json()}\n\n"
+            yield "data: [DONE]\n\n"
+            return
+
+        chunk_response = chunk_to_response(chunk, command_id)
+        yield f"data: {chunk_response.model_dump_json()}\n\n"
+
+        if chunk.finish_reason is not None:
+            yield "data: [DONE]\n\n"
+
+
+async def collect_chat_response(
+    command_id: CommandId,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> ChatCompletionResponse:
+    """Collect all token chunks and return a single ChatCompletionResponse."""
+    text_parts: list[str] = []
+    tool_calls: list[ToolCall] = []
+    model: str | None = None
+    finish_reason: FinishReason | None = None
+    error_message: str | None = None
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            error_message = chunk.error_message or "Internal server error"
+            break
+
+        if model is None:
+            model = chunk.model
+
+        if isinstance(chunk, TokenChunk):
+            text_parts.append(chunk.text)
+
+        if isinstance(chunk, ToolCallChunk):
+            tool_calls.extend(
+                ToolCall(
+                    id=str(uuid4()),
+                    index=i,
+                    function=tool,
+                )
+                for i, tool in enumerate(chunk.tool_calls)
+            )
+
+        if chunk.finish_reason is not None:
+            finish_reason = chunk.finish_reason
+
+    if error_message is not None:
+        raise ValueError(error_message)
+
+    combined_text = "".join(text_parts)
+    assert model is not None
+
+    return ChatCompletionResponse(
+        id=command_id,
+        created=int(time.time()),
+        model=model,
+        choices=[
+            ChatCompletionChoice(
+                index=0,
+                message=ChatCompletionMessage(
+                    role="assistant",
+                    content=combined_text,
+                    tool_calls=tool_calls if tool_calls else None,
+                ),
+                finish_reason=finish_reason,
+            )
+        ],
+    )

src/exo/master/adapters/claude.py

@@ -0,0 +1,321 @@
+"""Claude Messages API adapter for converting requests/responses."""
+
+import json
+from collections.abc import AsyncGenerator
+from typing import Any
+from uuid import uuid4
+
+from exo.shared.types.api import FinishReason
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
+from exo.shared.types.claude_api import (
+    ClaudeContentBlock,
+    ClaudeContentBlockDeltaEvent,
+    ClaudeContentBlockStartEvent,
+    ClaudeContentBlockStopEvent,
+    ClaudeInputJsonDelta,
+    ClaudeMessageDelta,
+    ClaudeMessageDeltaEvent,
+    ClaudeMessageDeltaUsage,
+    ClaudeMessagesRequest,
+    ClaudeMessagesResponse,
+    ClaudeMessageStart,
+    ClaudeMessageStartEvent,
+    ClaudeMessageStopEvent,
+    ClaudeStopReason,
+    ClaudeTextBlock,
+    ClaudeTextDelta,
+    ClaudeToolResultBlock,
+    ClaudeToolUseBlock,
+    ClaudeUsage,
+)
+from exo.shared.types.common import CommandId
+from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
+
+
+def finish_reason_to_claude_stop_reason(
+    finish_reason: FinishReason | None,
+) -> ClaudeStopReason | None:
+    """Map OpenAI finish_reason to Claude stop_reason."""
+    if finish_reason is None:
+        return None
+    mapping: dict[FinishReason, ClaudeStopReason] = {
+        "stop": "end_turn",
+        "length": "max_tokens",
+        "tool_calls": "tool_use",
+        "content_filter": "end_turn",
+        "function_call": "tool_use",
+    }
+    return mapping.get(finish_reason, "end_turn")
+
+
+def _extract_tool_result_text(block: ClaudeToolResultBlock) -> str:
+    """Extract plain text from a tool_result content field."""
+    if block.content is None:
+        return ""
+    if isinstance(block.content, str):
+        return block.content
+    return "".join(sub_block.text for sub_block in block.content)
+
+
+def claude_request_to_text_generation(
+    request: ClaudeMessagesRequest,
+) -> TextGenerationTaskParams:
+    # Handle system message
+    instructions: str | None = None
+    chat_template_messages: list[dict[str, Any]] = []
+
+    if request.system:
+        if isinstance(request.system, str):
+            instructions = request.system
+        else:
+            instructions = "".join(block.text for block in request.system)
+        chat_template_messages.append({"role": "system", "content": instructions})
+
+    # Convert messages to input
+    input_messages: list[InputMessage] = []
+    for msg in request.messages:
+        if isinstance(msg.content, str):
+            input_messages.append(InputMessage(role=msg.role, content=msg.content))
+            chat_template_messages.append({"role": msg.role, "content": msg.content})
+            continue
+
+        # Process structured content blocks
+        text_parts: list[str] = []
+        tool_calls: list[dict[str, Any]] = []
+        tool_results: list[ClaudeToolResultBlock] = []
+
+        for block in msg.content:
+            if isinstance(block, ClaudeTextBlock):
+                text_parts.append(block.text)
+            elif isinstance(block, ClaudeToolUseBlock):
+                tool_calls.append(
+                    {
+                        "id": block.id,
+                        "type": "function",
+                        "function": {
+                            "name": block.name,
+                            "arguments": json.dumps(block.input),
+                        },
+                    }
+                )
+            elif isinstance(block, ClaudeToolResultBlock):
+                tool_results.append(block)
+
+        content = "".join(text_parts)
+
+        # Build InputMessage from text content
+        if msg.role in ("user", "assistant"):
+            input_messages.append(InputMessage(role=msg.role, content=content))
+
+        # Build chat_template_messages preserving tool structure
+        if tool_calls:
+            chat_template_messages.append(
+                {"role": "assistant", "content": content, "tool_calls": tool_calls}
+            )
+        elif tool_results:
+            for tr in tool_results:
+                chat_template_messages.append(
+                    {
+                        "role": "tool",
+                        "tool_call_id": tr.tool_use_id,
+                        "content": _extract_tool_result_text(tr),
+                    }
+                )
+        else:
+            chat_template_messages.append({"role": msg.role, "content": content})
+
+    # Convert Claude tool definitions to OpenAI-style function tools
+    tools: list[dict[str, Any]] | None = None
+    if request.tools:
+        tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": tool.name,
+                    "description": tool.description or "",
+                    "parameters": tool.input_schema,
+                },
+            }
+            for tool in request.tools
+        ]
+
+    return TextGenerationTaskParams(
+        model=request.model,
+        input=input_messages if input_messages else "",
+        instructions=instructions,
+        max_output_tokens=request.max_tokens,
+        temperature=request.temperature,
+        top_p=request.top_p,
+        top_k=request.top_k,
+        stop=request.stop_sequences,
+        stream=request.stream,
+        tools=tools,
+        chat_template_messages=chat_template_messages
+        if chat_template_messages
+        else None,
+    )
+
+
+async def collect_claude_response(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> ClaudeMessagesResponse:
+    """Collect all token chunks and return a single ClaudeMessagesResponse."""
+    text_parts: list[str] = []
+    tool_use_blocks: list[ClaudeToolUseBlock] = []
+    stop_reason: ClaudeStopReason | None = None
+    last_stats = None
+    error_message: str | None = None
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            error_message = chunk.error_message or "Internal server error"
+            break
+
+        if isinstance(chunk, ToolCallChunk):
+            for tool in chunk.tool_calls:
+                tool_use_blocks.append(
+                    ClaudeToolUseBlock(
+                        id=f"toolu_{uuid4().hex[:24]}",
+                        name=tool.name,
+                        input=json.loads(tool.arguments),  # pyright: ignore[reportAny]
+                    )
+                )
+            last_stats = chunk.stats or last_stats
+            stop_reason = "tool_use"
+            continue
+
+        text_parts.append(chunk.text)
+        last_stats = chunk.stats or last_stats
+
+        if chunk.finish_reason is not None:
+            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)
+
+    if error_message is not None:
+        raise ValueError(error_message)
+
+    combined_text = "".join(text_parts)
+
+    # Build content blocks
+    content: list[ClaudeContentBlock] = []
+    if combined_text:
+        content.append(ClaudeTextBlock(text=combined_text))
+    content.extend(tool_use_blocks)
+
+    # If no content at all, include empty text block
+    if not content:
+        content.append(ClaudeTextBlock(text=""))
+
+    # Use actual usage data from stats if available
+    input_tokens = last_stats.prompt_tokens if last_stats else 0
+    output_tokens = last_stats.generation_tokens if last_stats else 0
+
+    return ClaudeMessagesResponse(
+        id=f"msg_{command_id}",
+        model=model,
+        content=content,
+        stop_reason=stop_reason,
+        usage=ClaudeUsage(
+            input_tokens=input_tokens,
+            output_tokens=output_tokens,
+        ),
+    )
+
+
+async def generate_claude_stream(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> AsyncGenerator[str, None]:
+    """Generate Claude Messages API streaming events from TokenChunks."""
+    # Initial message_start event
+    initial_message = ClaudeMessageStart(
+        id=f"msg_{command_id}",
+        model=model,
+        content=[],
+        stop_reason=None,
+        usage=ClaudeUsage(input_tokens=0, output_tokens=0),
+    )
+    start_event = ClaudeMessageStartEvent(message=initial_message)
+    yield f"event: message_start\ndata: {start_event.model_dump_json()}\n\n"
+
+    # content_block_start for text block at index 0
+    block_start = ClaudeContentBlockStartEvent(
+        index=0, content_block=ClaudeTextBlock(text="")
+    )
+    yield f"event: content_block_start\ndata: {block_start.model_dump_json()}\n\n"
+
+    output_tokens = 0
+    stop_reason: ClaudeStopReason | None = None
+    last_stats = None
+    next_block_index = 1  # text block is 0, tool blocks start at 1
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            # Close text block and bail
+            break
+
+        if isinstance(chunk, ToolCallChunk):
+            last_stats = chunk.stats or last_stats
+            stop_reason = "tool_use"
+
+            # Emit tool_use content blocks
+            for tool in chunk.tool_calls:
+                tool_id = f"toolu_{uuid4().hex[:24]}"
+                tool_input_json = tool.arguments
+
+                # content_block_start for tool_use
+                tool_block_start = ClaudeContentBlockStartEvent(
+                    index=next_block_index,
+                    content_block=ClaudeToolUseBlock(
+                        id=tool_id, name=tool.name, input={}
+                    ),
+                )
+                yield f"event: content_block_start\ndata: {tool_block_start.model_dump_json()}\n\n"
+
+                # content_block_delta with input_json_delta
+                tool_delta_event = ClaudeContentBlockDeltaEvent(
+                    index=next_block_index,
+                    delta=ClaudeInputJsonDelta(partial_json=tool_input_json),
+                )
+                yield f"event: content_block_delta\ndata: {tool_delta_event.model_dump_json()}\n\n"
+
+                # content_block_stop
+                tool_block_stop = ClaudeContentBlockStopEvent(index=next_block_index)
+                yield f"event: content_block_stop\ndata: {tool_block_stop.model_dump_json()}\n\n"
+
+                next_block_index += 1
+            continue
+
+        output_tokens += 1  # Count each chunk as one token
+        last_stats = chunk.stats or last_stats
+
+        # content_block_delta
+        delta_event = ClaudeContentBlockDeltaEvent(
+            index=0,
+            delta=ClaudeTextDelta(text=chunk.text),
+        )
+        yield f"event: content_block_delta\ndata: {delta_event.model_dump_json()}\n\n"
+
+        if chunk.finish_reason is not None:
+            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)
+
+    # Use actual token count from stats if available
+    if last_stats is not None:
+        output_tokens = last_stats.generation_tokens
+
+    # content_block_stop for text block
+    block_stop = ClaudeContentBlockStopEvent(index=0)
+    yield f"event: content_block_stop\ndata: {block_stop.model_dump_json()}\n\n"
+
+    # message_delta
+    message_delta = ClaudeMessageDeltaEvent(
+        delta=ClaudeMessageDelta(stop_reason=stop_reason),
+        usage=ClaudeMessageDeltaUsage(output_tokens=output_tokens),
+    )
+    yield f"event: message_delta\ndata: {message_delta.model_dump_json()}\n\n"
+
+    # message_stop
+    message_stop = ClaudeMessageStopEvent()
+    yield f"event: message_stop\ndata: {message_stop.model_dump_json()}\n\n"

src/exo/master/adapters/responses.py

@@ -0,0 +1,369 @@
+"""OpenAI Responses API adapter for converting requests/responses."""
+
+from collections.abc import AsyncGenerator
+from itertools import count
+from typing import Any
+from uuid import uuid4
+
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
+from exo.shared.types.common import CommandId
+from exo.shared.types.openai_responses import (
+    FunctionCallInputItem,
+    ResponseCompletedEvent,
+    ResponseContentPart,
+    ResponseContentPartAddedEvent,
+    ResponseContentPartDoneEvent,
+    ResponseCreatedEvent,
+    ResponseFunctionCallArgumentsDeltaEvent,
+    ResponseFunctionCallArgumentsDoneEvent,
+    ResponseFunctionCallItem,
+    ResponseInProgressEvent,
+    ResponseInputMessage,
+    ResponseItem,
+    ResponseMessageItem,
+    ResponseOutputItemAddedEvent,
+    ResponseOutputItemDoneEvent,
+    ResponseOutputText,
+    ResponsesRequest,
+    ResponsesResponse,
+    ResponseTextDeltaEvent,
+    ResponseTextDoneEvent,
+    ResponseUsage,
+)
+from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
+
+
+def _extract_content(content: str | list[ResponseContentPart]) -> str:
+    """Extract plain text from a content field that may be a string or list of parts."""
+    if isinstance(content, str):
+        return content
+    return "".join(part.text for part in content)
+
+
+def responses_request_to_text_generation(
+    request: ResponsesRequest,
+) -> TextGenerationTaskParams:
+    input_value: str | list[InputMessage]
+    built_chat_template: list[dict[str, Any]] | None = None
+    if isinstance(request.input, str):
+        input_value = request.input
+    else:
+        input_messages: list[InputMessage] = []
+        chat_template_messages: list[dict[str, Any]] = []
+
+        if request.instructions is not None:
+            chat_template_messages.append(
+                {"role": "system", "content": request.instructions}
+            )
+
+        for item in request.input:
+            if isinstance(item, ResponseInputMessage):
+                content = _extract_content(item.content)
+                if item.role in ("user", "assistant", "developer"):
+                    input_messages.append(InputMessage(role=item.role, content=content))
+                if item.role == "system":
+                    chat_template_messages.append(
+                        {"role": "system", "content": content}
+                    )
+                else:
+                    chat_template_messages.append(
+                        {"role": item.role, "content": content}
+                    )
+            elif isinstance(item, FunctionCallInputItem):
+                chat_template_messages.append(
+                    {
+                        "role": "assistant",
+                        "content": "",
+                        "tool_calls": [
+                            {
+                                "id": item.call_id,
+                                "type": "function",
+                                "function": {
+                                    "name": item.name,
+                                    "arguments": item.arguments,
+                                },
+                            }
+                        ],
+                    }
+                )
+            else:
+                chat_template_messages.append(
+                    {
+                        "role": "tool",
+                        "tool_call_id": item.call_id,
+                        "content": item.output,
+                    }
+                )
+
+        input_value = input_messages if input_messages else ""
+        built_chat_template = chat_template_messages if chat_template_messages else None
+
+    return TextGenerationTaskParams(
+        model=request.model,
+        input=input_value,
+        instructions=request.instructions,
+        max_output_tokens=request.max_output_tokens,
+        temperature=request.temperature,
+        top_p=request.top_p,
+        stream=request.stream,
+        tools=request.tools,
+        top_k=request.top_k,
+        stop=request.stop,
+        seed=request.seed,
+        chat_template_messages=built_chat_template or request.chat_template_messages,
+    )
+
+
+async def collect_responses_response(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> ResponsesResponse:
+    """Collect all token chunks and return a single ResponsesResponse."""
+    response_id = f"resp_{command_id}"
+    item_id = f"item_{command_id}"
+    accumulated_text = ""
+    function_call_items: list[ResponseFunctionCallItem] = []
+    last_stats = None
+    error_message: str | None = None
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            error_message = chunk.error_message or "Internal server error"
+            break
+
+        if isinstance(chunk, ToolCallChunk):
+            for tool in chunk.tool_calls:
+                function_call_items.append(
+                    ResponseFunctionCallItem(
+                        id=f"fc_{uuid4().hex[:24]}",
+                        call_id=f"call_{uuid4().hex[:24]}",
+                        name=tool.name,
+                        arguments=tool.arguments,
+                    )
+                )
+            last_stats = chunk.stats or last_stats
+            continue
+
+        accumulated_text += chunk.text
+        last_stats = chunk.stats or last_stats
+
+    if error_message is not None:
+        raise ValueError(error_message)
+
+    # Create usage from stats if available
+    usage = None
+    if last_stats is not None:
+        usage = ResponseUsage(
+            input_tokens=last_stats.prompt_tokens,
+            output_tokens=last_stats.generation_tokens,
+            total_tokens=last_stats.prompt_tokens + last_stats.generation_tokens,
+        )
+
+    output: list[ResponseItem] = [
+        ResponseMessageItem(
+            id=item_id,
+            content=[ResponseOutputText(text=accumulated_text)],
+            status="completed",
+        )
+    ]
+    output.extend(function_call_items)
+
+    return ResponsesResponse(
+        id=response_id,
+        model=model,
+        status="completed",
+        output=output,
+        output_text=accumulated_text,
+        usage=usage,
+    )
+
+
+async def generate_responses_stream(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> AsyncGenerator[str, None]:
+    """Generate OpenAI Responses API streaming events from TokenChunks."""
+    response_id = f"resp_{command_id}"
+    item_id = f"item_{command_id}"
+    seq = count(1)
+
+    # response.created
+    initial_response = ResponsesResponse(
+        id=response_id,
+        model=model,
+        status="in_progress",
+        output=[],
+        output_text="",
+    )
+    created_event = ResponseCreatedEvent(
+        sequence_number=next(seq), response=initial_response
+    )
+    yield f"event: response.created\ndata: {created_event.model_dump_json()}\n\n"
+
+    # response.in_progress
+    in_progress_event = ResponseInProgressEvent(
+        sequence_number=next(seq), response=initial_response
+    )
+    yield f"event: response.in_progress\ndata: {in_progress_event.model_dump_json()}\n\n"
+
+    # response.output_item.added
+    initial_item = ResponseMessageItem(
+        id=item_id,
+        content=[ResponseOutputText(text="")],
+        status="in_progress",
+    )
+    item_added = ResponseOutputItemAddedEvent(
+        sequence_number=next(seq), output_index=0, item=initial_item
+    )
+    yield f"event: response.output_item.added\ndata: {item_added.model_dump_json()}\n\n"
+
+    # response.content_part.added
+    initial_part = ResponseOutputText(text="")
+    part_added = ResponseContentPartAddedEvent(
+        sequence_number=next(seq),
+        item_id=item_id,
+        output_index=0,
+        content_index=0,
+        part=initial_part,
+    )
+    yield f"event: response.content_part.added\ndata: {part_added.model_dump_json()}\n\n"
+
+    accumulated_text = ""
+    function_call_items: list[ResponseFunctionCallItem] = []
+    last_stats = None
+    next_output_index = 1  # message item is at 0
+
+    async for chunk in chunk_stream:
+        if isinstance(chunk, ErrorChunk):
+            break
+
+        if isinstance(chunk, ToolCallChunk):
+            last_stats = chunk.stats or last_stats
+            for tool in chunk.tool_calls:
+                fc_id = f"fc_{uuid4().hex[:24]}"
+                call_id = f"call_{uuid4().hex[:24]}"
+
+                # response.output_item.added for function_call
+                fc_item = ResponseFunctionCallItem(
+                    id=fc_id,
+                    call_id=call_id,
+                    name=tool.name,
+                    arguments="",
+                    status="in_progress",
+                )
+                fc_added = ResponseOutputItemAddedEvent(
+                    sequence_number=next(seq),
+                    output_index=next_output_index,
+                    item=fc_item,
+                )
+                yield f"event: response.output_item.added\ndata: {fc_added.model_dump_json()}\n\n"
+
+                # response.function_call_arguments.delta
+                args_delta = ResponseFunctionCallArgumentsDeltaEvent(
+                    sequence_number=next(seq),
+                    item_id=fc_id,
+                    output_index=next_output_index,
+                    delta=tool.arguments,
+                )
+                yield f"event: response.function_call_arguments.delta\ndata: {args_delta.model_dump_json()}\n\n"
+
+                # response.function_call_arguments.done
+                args_done = ResponseFunctionCallArgumentsDoneEvent(
+                    sequence_number=next(seq),
+                    item_id=fc_id,
+                    output_index=next_output_index,
+                    name=tool.name,
+                    arguments=tool.arguments,
+                )
+                yield f"event: response.function_call_arguments.done\ndata: {args_done.model_dump_json()}\n\n"
+
+                # response.output_item.done
+                fc_done_item = ResponseFunctionCallItem(
+                    id=fc_id,
+                    call_id=call_id,
+                    name=tool.name,
+                    arguments=tool.arguments,
+                    status="completed",
+                )
+                fc_item_done = ResponseOutputItemDoneEvent(
+                    sequence_number=next(seq),
+                    output_index=next_output_index,
+                    item=fc_done_item,
+                )
+                yield f"event: response.output_item.done\ndata: {fc_item_done.model_dump_json()}\n\n"
+
+                function_call_items.append(fc_done_item)
+                next_output_index += 1
+            continue
+
+        accumulated_text += chunk.text
+        last_stats = chunk.stats or last_stats
+
+        # response.output_text.delta
+        delta_event = ResponseTextDeltaEvent(
+            sequence_number=next(seq),
+            item_id=item_id,
+            output_index=0,
+            content_index=0,
+            delta=chunk.text,
+        )
+        yield f"event: response.output_text.delta\ndata: {delta_event.model_dump_json()}\n\n"
+
+    # response.output_text.done
+    text_done = ResponseTextDoneEvent(
+        sequence_number=next(seq),
+        item_id=item_id,
+        output_index=0,
+        content_index=0,
+        text=accumulated_text,
+    )
+    yield f"event: response.output_text.done\ndata: {text_done.model_dump_json()}\n\n"
+
+    # response.content_part.done
+    final_part = ResponseOutputText(text=accumulated_text)
+    part_done = ResponseContentPartDoneEvent(
+        sequence_number=next(seq),
+        item_id=item_id,
+        output_index=0,
+        content_index=0,
+        part=final_part,
+    )
+    yield f"event: response.content_part.done\ndata: {part_done.model_dump_json()}\n\n"
+
+    # response.output_item.done
+    final_message_item = ResponseMessageItem(
+        id=item_id,
+        content=[ResponseOutputText(text=accumulated_text)],
+        status="completed",
+    )
+    item_done = ResponseOutputItemDoneEvent(
+        sequence_number=next(seq), output_index=0, item=final_message_item
+    )
+    yield f"event: response.output_item.done\ndata: {item_done.model_dump_json()}\n\n"
+
+    # Create usage from stats if available
+    usage = None
+    if last_stats is not None:
+        usage = ResponseUsage(
+            input_tokens=last_stats.prompt_tokens,
+            output_tokens=last_stats.generation_tokens,
+            total_tokens=last_stats.prompt_tokens + last_stats.generation_tokens,
+        )
+
+    # response.completed
+    output: list[ResponseItem] = [final_message_item]
+    output.extend(function_call_items)
+    final_response = ResponsesResponse(
+        id=response_id,
+        model=model,
+        status="completed",
+        output=output,
+        output_text=accumulated_text,
+        usage=usage,
+    )
+    completed_event = ResponseCompletedEvent(
+        sequence_number=next(seq), response=final_response
+    )
+    yield f"event: response.completed\ndata: {completed_event.model_dump_json()}\n\n"

src/exo/master/main.py

@@ -10,10 +10,9 @@ from exo.master.placement import (
     get_transition_events,
     place_instance,
 )
-from exo.plugins.registry import PluginRegistry
 from exo.shared.apply import apply
+from exo.shared.constants import EXO_TRACING_ENABLED
 from exo.shared.types.commands import (
-    ChatCompletion,
     CreateInstance,
     DeleteInstance,
     ForwarderCommand,
@@ -24,10 +23,10 @@ from exo.shared.types.commands import (
     SendInputChunk,
     TaskFinished,
     TestCommand,
+    TextGeneration,
 )
 from exo.shared.types.common import CommandId, NodeId, SessionId
 from exo.shared.types.events import (
-    BaseEvent,
     Event,
     ForwarderEvent,
     IndexedEvent,
@@ -37,11 +36,11 @@ from exo.shared.types.events import (
     NodeTimedOut,
     TaskCreated,
     TaskDeleted,
+    TraceEventData,
+    TracesCollected,
+    TracesMerged,
 )
 from exo.shared.types.state import State
-from exo.shared.types.tasks import (
-    ChatCompletion as ChatCompletionTask,
-)
 from exo.shared.types.tasks import (
     ImageEdits as ImageEditsTask,
 )
@@ -52,6 +51,9 @@ from exo.shared.types.tasks import (
     TaskId,
     TaskStatus,
 )
+from exo.shared.types.tasks import (
+    TextGeneration as TextGenerationTask,
+)
 from exo.shared.types.worker.instances import InstanceId
 from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.event_buffer import MultiSourceBuffer
@@ -85,9 +87,11 @@ class Master:
         self._loopback_event_sender: Sender[ForwarderEvent] = (
             local_event_receiver.clone_sender()
         )
-        self._multi_buffer = MultiSourceBuffer[NodeId, BaseEvent]()
+        self._multi_buffer = MultiSourceBuffer[NodeId, Event]()
         # TODO: not have this
-        self._event_log: list[BaseEvent] = []
+        self._event_log: list[Event] = []
+        self._pending_traces: dict[TaskId, dict[int, list[TraceEventData]]] = {}
+        self._expected_ranks: dict[TaskId, set[int]] = {}
 
     async def run(self):
         logger.info("Starting Master")
@@ -119,11 +123,11 @@ class Master:
                     match command:
                         case TestCommand():
                             pass
-                        case ChatCompletion():
+                        case TextGeneration():
                             for instance in self.state.instances.values():
                                 if (
                                     instance.shard_assignments.model_id
-                                    == command.request_params.model
+                                    == command.task_params.model
                                 ):
                                     task_count = sum(
                                         1
@@ -136,7 +140,7 @@ class Master:
 
                             if not instance_task_counts:
                                 raise ValueError(
-                                    f"No instance found for model {command.request_params.model}"
+                                    f"No instance found for model {command.task_params.model}"
                                 )
 
                             available_instance_ids = sorted(
@@ -150,12 +154,12 @@ class Master:
                             generated_events.append(
                                 TaskCreated(
                                     task_id=task_id,
-                                    task=ChatCompletionTask(
+                                    task=TextGenerationTask(
                                         task_id=task_id,
                                         command_id=command.command_id,
                                         instance_id=available_instance_ids[0],
                                         task_status=TaskStatus.Pending,
-                                        task_params=command.request_params,
+                                        task_params=command.task_params,
                                     ),
                                 )
                             )
@@ -165,7 +169,7 @@ class Master:
                             for instance in self.state.instances.values():
                                 if (
                                     instance.shard_assignments.model_id
-                                    == command.request_params.model
+                                    == command.task_params.model
                                 ):
                                     task_count = sum(
                                         1
@@ -178,7 +182,7 @@ class Master:
 
                             if not instance_task_counts:
                                 raise ValueError(
-                                    f"No instance found for model {command.request_params.model}"
+                                    f"No instance found for model {command.task_params.model}"
                                 )
 
                             available_instance_ids = sorted(
@@ -189,25 +193,37 @@ class Master:
                             )
 
                             task_id = TaskId()
+                            selected_instance_id = available_instance_ids[0]
                             generated_events.append(
                                 TaskCreated(
                                     task_id=task_id,
                                     task=ImageGenerationTask(
                                         task_id=task_id,
                                         command_id=command.command_id,
-                                        instance_id=available_instance_ids[0],
+                                        instance_id=selected_instance_id,
                                         task_status=TaskStatus.Pending,
-                                        task_params=command.request_params,
+                                        task_params=command.task_params,
                                     ),
                                 )
                             )
 
                             self.command_task_mapping[command.command_id] = task_id
+
+                            if EXO_TRACING_ENABLED:
+                                selected_instance = self.state.instances.get(
+                                    selected_instance_id
+                                )
+                                if selected_instance:
+                                    ranks = set(
+                                        shard.device_rank
+                                        for shard in selected_instance.shard_assignments.runner_to_shard.values()
+                                    )
+                                    self._expected_ranks[task_id] = ranks
                         case ImageEdits():
                             for instance in self.state.instances.values():
                                 if (
                                     instance.shard_assignments.model_id
-                                    == command.request_params.model
+                                    == command.task_params.model
                                 ):
                                     task_count = sum(
                                         1
@@ -220,7 +236,7 @@ class Master:
 
                             if not instance_task_counts:
                                 raise ValueError(
-                                    f"No instance found for model {command.request_params.model}"
+                                    f"No instance found for model {command.task_params.model}"
                                 )
 
                             available_instance_ids = sorted(
@@ -231,20 +247,32 @@ class Master:
                             )
 
                             task_id = TaskId()
+                            selected_instance_id = available_instance_ids[0]
                             generated_events.append(
                                 TaskCreated(
                                     task_id=task_id,
                                     task=ImageEditsTask(
                                         task_id=task_id,
                                         command_id=command.command_id,
-                                        instance_id=available_instance_ids[0],
+                                        instance_id=selected_instance_id,
                                         task_status=TaskStatus.Pending,
-                                        task_params=command.request_params,
+                                        task_params=command.task_params,
                                     ),
                                 )
                             )
 
                             self.command_task_mapping[command.command_id] = task_id
+
+                            if EXO_TRACING_ENABLED:
+                                selected_instance = self.state.instances.get(
+                                    selected_instance_id
+                                )
+                                if selected_instance:
+                                    ranks = set(
+                                        shard.device_rank
+                                        for shard in selected_instance.shard_assignments.runner_to_shard.values()
+                                    )
+                                    self._expected_ranks[task_id] = ranks
                         case DeleteInstance():
                             placement = delete_instance(command, self.state.instances)
                             transition_events = get_transition_events(
@@ -298,17 +326,6 @@ class Master:
                                 await self._send_event(
                                     IndexedEvent(idx=i, event=self._event_log[i])
                                 )
-                        case _:
-                            # Check if a plugin handles this command
-                            registry = PluginRegistry.get()
-                            plugin = registry.get_plugin_for_command(command)
-                            if plugin is not None:
-                                events = plugin.process_command(
-                                    command,
-                                    self.state.topology,
-                                    self.state.instances,
-                                )
-                                generated_events.extend(events)
                     for event in generated_events:
                         await self.event_sender.send(event)
                 except ValueError as e:
@@ -348,6 +365,10 @@ class Master:
                     local_event.origin,
                 )
                 for event in self._multi_buffer.drain():
+                    if isinstance(event, TracesCollected):
+                        await self._handle_traces_collected(event)
+                        continue
+
                     logger.debug(f"Master indexing event: {str(event)[:100]}")
                     indexed = IndexedEvent(event=event, idx=len(self._event_log))
                     self.state = apply(self.state, indexed)
@@ -386,3 +407,29 @@ class Master:
                 event=event.event,
             )
         )
+
+    async def _handle_traces_collected(self, event: TracesCollected) -> None:
+        task_id = event.task_id
+        if task_id not in self._pending_traces:
+            self._pending_traces[task_id] = {}
+        self._pending_traces[task_id][event.rank] = event.traces
+
+        if (
+            task_id in self._expected_ranks
+            and set(self._pending_traces[task_id].keys())
+            >= self._expected_ranks[task_id]
+        ):
+            await self._merge_and_save_traces(task_id)
+
+    async def _merge_and_save_traces(self, task_id: TaskId) -> None:
+        all_trace_data: list[TraceEventData] = []
+        for trace_data in self._pending_traces[task_id].values():
+            all_trace_data.extend(trace_data)
+
+        await self.event_sender.send(
+            TracesMerged(task_id=task_id, traces=all_trace_data)
+        )
+
+        del self._pending_traces[task_id]
+        if task_id in self._expected_ranks:
+            del self._expected_ranks[task_id]

src/exo/master/placement.py

@@ -24,7 +24,7 @@ from exo.shared.types.events import Event, InstanceCreated, InstanceDeleted
 from exo.shared.types.memory import Memory
 from exo.shared.types.profiling import MemoryUsage, NodeNetworkInfo
 from exo.shared.types.worker.instances import (
-    BaseInstance,
+    Instance,
     InstanceId,
     InstanceMeta,
     MlxJacclInstance,
@@ -35,14 +35,14 @@ from exo.shared.types.worker.shards import Sharding
 
 def random_ephemeral_port() -> int:
     port = random.randint(49153, 65535)
-    return port - 1 if port <= 52415 else 52414
+    return port - 1 if port <= 52415 else port
 
 
 def add_instance_to_placements(
     command: CreateInstance,
     topology: Topology,
-    current_instances: Mapping[InstanceId, BaseInstance],
-) -> Mapping[InstanceId, BaseInstance]:
+    current_instances: Mapping[InstanceId, Instance],
+) -> Mapping[InstanceId, Instance]:
     # TODO: validate against topology
 
     return {**current_instances, command.instance.instance_id: command.instance}
@@ -51,12 +51,21 @@ def add_instance_to_placements(
 def place_instance(
     command: PlaceInstance,
     topology: Topology,
-    current_instances: Mapping[InstanceId, BaseInstance],
+    current_instances: Mapping[InstanceId, Instance],
     node_memory: Mapping[NodeId, MemoryUsage],
     node_network: Mapping[NodeId, NodeNetworkInfo],
-) -> dict[InstanceId, BaseInstance]:
+    required_nodes: set[NodeId] | None = None,
+) -> dict[InstanceId, Instance]:
     cycles = topology.get_cycles()
     candidate_cycles = list(filter(lambda it: len(it) >= command.min_nodes, cycles))
+
+    # Filter to cycles containing all required nodes (subset matching)
+    if required_nodes:
+        candidate_cycles = [
+            cycle
+            for cycle in candidate_cycles
+            if required_nodes.issubset(cycle.node_ids)
+        ]
     cycles_with_sufficient_memory = filter_cycles_by_memory(
         candidate_cycles, node_memory, command.model_card.storage_size
     )
@@ -87,12 +96,12 @@ def place_instance(
 
     smallest_cycles = get_smallest_cycles(cycles_with_sufficient_memory)
 
-    smallest_tb_cycles = [
-        cycle for cycle in smallest_cycles if topology.is_thunderbolt_cycle(cycle)
+    smallest_rdma_cycles = [
+        cycle for cycle in smallest_cycles if topology.is_rdma_cycle(cycle)
     ]
 
-    if smallest_tb_cycles != []:
-        smallest_cycles = smallest_tb_cycles
+    if command.instance_meta == InstanceMeta.MlxJaccl and smallest_rdma_cycles != []:
+        smallest_cycles = smallest_rdma_cycles
 
     cycles_with_leaf_nodes: list[Cycle] = [
         cycle
@@ -159,8 +168,8 @@ def place_instance(
 
 def delete_instance(
     command: DeleteInstance,
-    current_instances: Mapping[InstanceId, BaseInstance],
-) -> dict[InstanceId, BaseInstance]:
+    current_instances: Mapping[InstanceId, Instance],
+) -> dict[InstanceId, Instance]:
     target_instances = dict(deepcopy(current_instances))
     if command.instance_id in target_instances:
         del target_instances[command.instance_id]
@@ -169,8 +178,8 @@ def delete_instance(
 
 
 def get_transition_events(
-    current_instances: Mapping[InstanceId, BaseInstance],
-    target_instances: Mapping[InstanceId, BaseInstance],
+    current_instances: Mapping[InstanceId, Instance],
+    target_instances: Mapping[InstanceId, Instance],
 ) -> Sequence[Event]:
     events: list[Event] = []
 

src/exo/master/placement_utils.py

@@ -197,49 +197,6 @@ def get_shard_assignments(
             )
 
 
-def get_hosts_from_subgraph(cycle_digraph: Topology) -> list[Host]:
-    cycles = cycle_digraph.get_cycles()
-    expected_length = len(list(cycle_digraph.list_nodes()))
-    cycles = [cycle for cycle in cycles if len(cycle) == expected_length]
-    if not cycles:
-        if expected_length > 1:
-            logger.warning(
-                f"No cycles of length {expected_length} found even though chosen subgraph contained {expected_length} nodes"
-            )
-        return []
-
-    cycle = cycles[0]
-
-    get_thunderbolt = False
-    if cycle_digraph.is_thunderbolt_cycle(cycle):
-        get_thunderbolt = True
-
-    logger.debug(f"Using thunderbolt cycle: {get_thunderbolt}")
-
-    hosts: list[Host] = []
-    for i in range(len(cycle)):
-        current_node = cycle.node_ids[i]
-        next_node = cycle.node_ids[(i + 1) % len(cycle)]
-
-        for connection in cycle_digraph.get_all_connections_between(
-            source=current_node, sink=next_node
-        ):
-            if not isinstance(connection, SocketConnection):
-                continue
-
-            if get_thunderbolt and not connection.is_thunderbolt():
-                continue
-
-            host = Host(
-                ip=connection.sink_multiaddr.ip_address,
-                port=connection.sink_multiaddr.port,
-            )
-            hosts.append(host)
-            break
-
-    return hosts
-
-
 def get_mlx_jaccl_devices_matrix(
     selected_cycle: list[NodeId],
     cycle_digraph: Topology,
@@ -265,9 +222,6 @@ def get_mlx_jaccl_devices_matrix(
                     matrix[i][j] = conn.source_rdma_iface
                     break
             else:
-                logger.warning(
-                    f"Failed to find interface name between {node_i} and {node_j}"
-                )
                 raise ValueError(
                     "Current jaccl backend requires all-to-all RDMA connections"
                 )
@@ -279,22 +233,11 @@ def _find_connection_ip(
     node_i: NodeId,
     node_j: NodeId,
     cycle_digraph: Topology,
-) -> Generator[tuple[str, bool]]:
+) -> Generator[str, None, None]:
     """Find all IP addresses that connect node i to node j."""
     for connection in cycle_digraph.get_all_connections_between(node_i, node_j):
         if isinstance(connection, SocketConnection):
-            yield connection.sink_multiaddr.ip_address, connection.is_thunderbolt()
-
-
-def _find_interface_name_for_ip(
-    ip_address: str, node_network: NodeNetworkInfo
-) -> str | None:
-    """Find the interface name for an IP address on a node (any interface)."""
-    for interface in node_network.interfaces:
-        if interface.ip_address == ip_address:
-            return interface.name
-
-    return None
+            yield connection.sink_multiaddr.ip_address
 
 
 def _find_ip_prioritised(
@@ -303,43 +246,25 @@ def _find_ip_prioritised(
     cycle_digraph: Topology,
     node_network: Mapping[NodeId, NodeNetworkInfo],
 ) -> str | None:
-    # TODO: Actually prioritize in the correct Ethernet > Wifi > Non-TB > TB order.
     """Find an IP address between nodes with prioritization.
 
-    Priority order:
-    1. en0 (Ethernet on Mac Studio, WiFi on MacBook)
-    2. en1 (WiFi on Mac Studio, Ethernet on MacBook)
-    3. Non-Thunderbolt connections
-    4. Any other IP address
+    Priority: ethernet > wifi > unknown > thunderbolt
     """
     ips = list(_find_connection_ip(node_id, other_node_id, cycle_digraph))
-    # We expect a unique iface -> ip mapping
-    iface_map = {
-        _find_interface_name_for_ip(
-            ip, node_network.get(other_node_id, NodeNetworkInfo())
-        ): ip
-        for ip, _ in ips
+    if not ips:
+        return None
+    other_network = node_network.get(other_node_id, NodeNetworkInfo())
+    ip_to_type = {
+        iface.ip_address: iface.interface_type for iface in other_network.interfaces
     }
-
-    en0_ip = iface_map.get("en0")
-    if en0_ip:
-        return en0_ip
-
-    en1_ip = iface_map.get("en1")
-    if en1_ip:
-        return en1_ip
-
-    non_thunderbolt_ip = next(
-        (ip for (ip, is_thunderbolt) in ips if not is_thunderbolt), None
-    )
-
-    if non_thunderbolt_ip:
-        return non_thunderbolt_ip
-
-    if ips:
-        return ips[0][0]
-
-    return None
+    priority = {
+        "ethernet": 0,
+        "wifi": 1,
+        "unknown": 2,
+        "maybe_ethernet": 3,
+        "thunderbolt": 4,
+    }
+    return min(ips, key=lambda ip: priority.get(ip_to_type.get(ip, "unknown"), 2))
 
 
 def get_mlx_ring_hosts_by_node(
@@ -381,9 +306,6 @@ def get_mlx_ring_hosts_by_node(
                 node_id, other_node_id, cycle_digraph, node_network
             )
             if connection_ip is None:
-                logger.warning(
-                    f"Failed to find prioritised connection IP between {node_id} and {other_node_id}"
-                )
                 raise ValueError(
                     "MLX ring backend requires connectivity between neighbouring nodes"
                 )
@@ -416,9 +338,6 @@ def get_mlx_jaccl_coordinators(
         if ip is not None:
             return ip
 
-        logger.warning(
-            f"Failed to find directly connected ip between {n} and {coordinator}"
-        )
         raise ValueError(
             "Current jaccl backend requires all participating devices to be able to communicate"
         )

src/exo/master/tests/test_api_error_handling.py

@@ -1,13 +1,9 @@
 # pyright: reportUnusedFunction=false, reportAny=false
-from typing import Any, get_args
+from typing import Any
 
 from fastapi import FastAPI, HTTPException
 from fastapi.testclient import TestClient
 
-from exo.shared.types.api import ErrorInfo, ErrorResponse, FinishReason
-from exo.shared.types.chunks import ImageChunk, TokenChunk
-from exo.worker.tests.constants import MODEL_A_ID
-
 
 def test_http_exception_handler_formats_openai_style() -> None:
     """Test that HTTPException is converted to OpenAI-style error format."""
@@ -48,95 +44,3 @@ def test_http_exception_handler_formats_openai_style() -> None:
     assert data["error"]["message"] == "Resource not found"
     assert data["error"]["type"] == "Not Found"
     assert data["error"]["code"] == 404
-
-
-def test_finish_reason_includes_error() -> None:
-    valid_reasons = get_args(FinishReason)
-    assert "error" in valid_reasons
-
-
-def test_token_chunk_with_error_fields() -> None:
-    chunk = TokenChunk(
-        idx=0,
-        model=MODEL_A_ID,
-        text="",
-        token_id=0,
-        finish_reason="error",
-        error_message="Something went wrong",
-    )
-
-    assert chunk.finish_reason == "error"
-    assert chunk.error_message == "Something went wrong"
-
-
-def test_token_chunk_without_error() -> None:
-    chunk = TokenChunk(
-        idx=1,
-        model=MODEL_A_ID,
-        text="Hello",
-        token_id=42,
-        finish_reason=None,
-    )
-
-    assert chunk.finish_reason is None
-    assert chunk.error_message is None
-
-
-def test_error_response_construction() -> None:
-    error_response = ErrorResponse(
-        error=ErrorInfo(
-            message="Generation failed",
-            type="InternalServerError",
-            code=500,
-        )
-    )
-
-    assert error_response.error.message == "Generation failed"
-    assert error_response.error.code == 500
-
-
-def test_normal_finish_reasons_still_work() -> None:
-    for reason in ["stop", "length", "tool_calls", "content_filter", "function_call"]:
-        chunk = TokenChunk(
-            idx=0,
-            model=MODEL_A_ID,
-            text="done",
-            token_id=100,
-            finish_reason=reason,  # type: ignore[arg-type]
-        )
-        assert chunk.finish_reason == reason
-
-
-def test_image_chunk_with_error_fields() -> None:
-    chunk = ImageChunk(
-        idx=0,
-        model=MODEL_A_ID,
-        data="",
-        chunk_index=0,
-        total_chunks=1,
-        image_index=0,
-        finish_reason="error",
-        error_message="Image generation failed",
-    )
-
-    assert chunk.finish_reason == "error"
-    assert chunk.error_message == "Image generation failed"
-    assert chunk.data == ""
-    assert chunk.chunk_index == 0
-    assert chunk.total_chunks == 1
-    assert chunk.image_index == 0
-
-
-def test_image_chunk_without_error() -> None:
-    chunk = ImageChunk(
-        idx=0,
-        model=MODEL_A_ID,
-        data="base64encodeddata",
-        chunk_index=0,
-        total_chunks=1,
-        image_index=0,
-    )
-
-    assert chunk.finish_reason is None
-    assert chunk.error_message is None
-    assert chunk.data == "base64encodeddata"

src/exo/master/tests/test_claude_api.py

@@ -0,0 +1,182 @@
+"""Tests for Claude Messages API conversion functions and types."""
+
+import pydantic
+import pytest
+
+from exo.master.adapters.claude import (
+    claude_request_to_text_generation,
+    finish_reason_to_claude_stop_reason,
+)
+from exo.shared.types.claude_api import (
+    ClaudeMessage,
+    ClaudeMessagesRequest,
+    ClaudeTextBlock,
+)
+from exo.shared.types.common import ModelId
+
+
+class TestFinishReasonToClaudeStopReason:
+    """Tests for finish_reason to Claude stop_reason mapping."""
+
+    def test_stop_maps_to_end_turn(self):
+        assert finish_reason_to_claude_stop_reason("stop") == "end_turn"
+
+    def test_length_maps_to_max_tokens(self):
+        assert finish_reason_to_claude_stop_reason("length") == "max_tokens"
+
+    def test_tool_calls_maps_to_tool_use(self):
+        assert finish_reason_to_claude_stop_reason("tool_calls") == "tool_use"
+
+    def test_function_call_maps_to_tool_use(self):
+        assert finish_reason_to_claude_stop_reason("function_call") == "tool_use"
+
+    def test_content_filter_maps_to_end_turn(self):
+        assert finish_reason_to_claude_stop_reason("content_filter") == "end_turn"
+
+    def test_none_returns_none(self):
+        assert finish_reason_to_claude_stop_reason(None) is None
+
+
+class TestClaudeRequestToInternal:
+    """Tests for converting Claude Messages API requests to TextGenerationTaskParams."""
+
+    def test_basic_request_conversion(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert params.model == "claude-3-opus"
+        assert params.max_output_tokens == 100
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+        assert params.input[0].role == "user"
+        assert params.input[0].content == "Hello"
+        assert params.instructions is None
+
+    def test_request_with_system_string(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            system="You are a helpful assistant.",
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert params.instructions == "You are a helpful assistant."
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+        assert params.input[0].role == "user"
+        assert params.input[0].content == "Hello"
+
+    def test_request_with_system_text_blocks(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            system=[
+                ClaudeTextBlock(text="You are helpful. "),
+                ClaudeTextBlock(text="Be concise."),
+            ],
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert params.instructions == "You are helpful. Be concise."
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+
+    def test_request_with_content_blocks(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            messages=[
+                ClaudeMessage(
+                    role="user",
+                    content=[
+                        ClaudeTextBlock(text="First part. "),
+                        ClaudeTextBlock(text="Second part."),
+                    ],
+                ),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+        assert params.input[0].content == "First part. Second part."
+
+    def test_request_with_multi_turn_conversation(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+                ClaudeMessage(role="assistant", content="Hi there!"),
+                ClaudeMessage(role="user", content="How are you?"),
+            ],
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert isinstance(params.input, list)
+        assert len(params.input) == 3
+        assert params.input[0].role == "user"
+        assert params.input[1].role == "assistant"
+        assert params.input[2].role == "user"
+
+    def test_request_with_optional_parameters(self):
+        request = ClaudeMessagesRequest(
+            model=ModelId("claude-3-opus"),
+            max_tokens=100,
+            messages=[ClaudeMessage(role="user", content="Hello")],
+            temperature=0.7,
+            top_p=0.9,
+            top_k=40,
+            stop_sequences=["STOP", "END"],
+            stream=True,
+        )
+        params = claude_request_to_text_generation(request)
+
+        assert params.temperature == 0.7
+        assert params.top_p == 0.9
+        assert params.top_k == 40
+        assert params.stop == ["STOP", "END"]
+        assert params.stream is True
+
+
+class TestClaudeMessagesRequestValidation:
+    """Tests for Claude Messages API request validation."""
+
+    def test_request_requires_model(self):
+        with pytest.raises(pydantic.ValidationError):
+            ClaudeMessagesRequest.model_validate(
+                {
+                    "max_tokens": 100,
+                    "messages": [{"role": "user", "content": "Hello"}],
+                }
+            )
+
+    def test_request_requires_max_tokens(self):
+        with pytest.raises(pydantic.ValidationError):
+            ClaudeMessagesRequest.model_validate(
+                {
+                    "model": "claude-3-opus",
+                    "messages": [{"role": "user", "content": "Hello"}],
+                }
+            )
+
+    def test_request_requires_messages(self):
+        with pytest.raises(pydantic.ValidationError):
+            ClaudeMessagesRequest.model_validate(
+                {
+                    "model": "claude-3-opus",
+                    "max_tokens": 100,
+                }
+            )

src/exo/master/tests/test_claude_tool_use.py

@@ -0,0 +1,265 @@
+"""Tests for Claude Messages API tool_use support in the adapter."""
+
+import json
+from collections.abc import AsyncGenerator
+from typing import Any, cast
+
+from exo.master.adapters.claude import collect_claude_response, generate_claude_stream
+from exo.shared.types.api import ToolCallItem
+from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
+from exo.shared.types.common import CommandId, ModelId
+
+
+async def _chunks_to_stream(
+    chunks: list[ErrorChunk | ToolCallChunk | TokenChunk],
+) -> AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None]:
+    for chunk in chunks:
+        yield chunk
+
+
+MODEL = ModelId("test-model")
+COMMAND_ID = CommandId("cmd_test123")
+
+
+def _parse_sse_events(events: list[str]) -> list[dict[str, Any]]:
+    """Parse SSE event strings into JSON dicts."""
+    parsed: list[dict[str, Any]] = []
+    for event_str in events:
+        for line in event_str.strip().split("\n"):
+            if line.startswith("data: "):
+                parsed.append(cast(dict[str, Any], json.loads(line[6:])))
+    return parsed
+
+
+class TestCollectClaudeResponseToolUse:
+    """Tests for non-streaming tool_use response collection."""
+
+    async def test_tool_call_chunk_produces_tool_use_blocks(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="get_weather",
+                        arguments='{"location": "San Francisco"}',
+                    )
+                ],
+            ),
+        ]
+        response = await collect_claude_response(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        )
+
+        assert response.stop_reason == "tool_use"
+        tool_blocks = [b for b in response.content if b.type == "tool_use"]
+        assert len(tool_blocks) == 1
+        block = tool_blocks[0]
+        assert block.type == "tool_use"
+        assert block.name == "get_weather"
+        assert block.input == {"location": "San Francisco"}
+        assert block.id.startswith("toolu_")
+
+    async def test_multiple_tool_calls(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="get_weather",
+                        arguments='{"location": "SF"}',
+                    ),
+                    ToolCallItem(
+                        name="get_time",
+                        arguments='{"timezone": "PST"}',
+                    ),
+                ],
+            ),
+        ]
+        response = await collect_claude_response(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        )
+
+        assert response.stop_reason == "tool_use"
+        tool_blocks = [b for b in response.content if b.type == "tool_use"]
+        assert len(tool_blocks) == 2
+        assert tool_blocks[0].name == "get_weather"
+        assert tool_blocks[1].name == "get_time"
+
+    async def test_mixed_text_and_tool_use(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            TokenChunk(model=MODEL, text="Let me check ", token_id=1, usage=None),
+            TokenChunk(model=MODEL, text="the weather.", token_id=2, usage=None),
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="get_weather",
+                        arguments='{"location": "NYC"}',
+                    )
+                ],
+            ),
+        ]
+        response = await collect_claude_response(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        )
+
+        assert response.stop_reason == "tool_use"
+        text_blocks = [b for b in response.content if b.type == "text"]
+        tool_blocks = [b for b in response.content if b.type == "tool_use"]
+        assert len(text_blocks) == 1
+        assert text_blocks[0].text == "Let me check the weather."
+        assert len(tool_blocks) == 1
+        assert tool_blocks[0].name == "get_weather"
+
+    async def test_no_content_produces_empty_text_block(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = []
+        response = await collect_claude_response(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        )
+        assert len(response.content) == 1
+        assert response.content[0].type == "text"
+
+
+class TestGenerateClaudeStreamToolUse:
+    """Tests for streaming tool_use event generation."""
+
+    async def test_tool_call_emits_tool_use_events(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="get_weather",
+                        arguments='{"location": "SF"}',
+                    )
+                ],
+            ),
+        ]
+        events: list[str] = []
+        async for event in generate_claude_stream(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        ):
+            events.append(event)
+
+        parsed = _parse_sse_events(events)
+
+        # Find tool_use content_block_start
+        tool_starts = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_start"
+            and cast(dict[str, Any], e.get("content_block", {})).get("type")
+            == "tool_use"
+        ]
+        assert len(tool_starts) == 1
+        content_block = cast(dict[str, Any], tool_starts[0]["content_block"])
+        assert content_block["name"] == "get_weather"
+        assert content_block["input"] == {}
+        assert cast(str, content_block["id"]).startswith("toolu_")
+
+        # Find input_json_delta
+        json_deltas = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_delta"
+            and cast(dict[str, Any], e.get("delta", {})).get("type")
+            == "input_json_delta"
+        ]
+        assert len(json_deltas) == 1
+        delta = cast(dict[str, Any], json_deltas[0]["delta"])
+        assert json.loads(cast(str, delta["partial_json"])) == {"location": "SF"}
+
+        # Find message_delta with tool_use stop reason
+        msg_deltas = [e for e in parsed if e.get("type") == "message_delta"]
+        assert len(msg_deltas) == 1
+        assert cast(dict[str, Any], msg_deltas[0]["delta"])["stop_reason"] == "tool_use"
+
+    async def test_streaming_mixed_text_and_tool_use(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            TokenChunk(model=MODEL, text="Hello ", token_id=1, usage=None),
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(
+                        name="search",
+                        arguments='{"query": "test"}',
+                    )
+                ],
+            ),
+        ]
+        events: list[str] = []
+        async for event in generate_claude_stream(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        ):
+            events.append(event)
+
+        parsed = _parse_sse_events(events)
+
+        # Should have text delta at index 0
+        text_deltas = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_delta"
+            and cast(dict[str, Any], e.get("delta", {})).get("type") == "text_delta"
+        ]
+        assert len(text_deltas) == 1
+        assert text_deltas[0]["index"] == 0
+        assert cast(dict[str, Any], text_deltas[0]["delta"])["text"] == "Hello "
+
+        # Tool block at index 1
+        tool_starts = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_start"
+            and cast(dict[str, Any], e.get("content_block", {})).get("type")
+            == "tool_use"
+        ]
+        assert len(tool_starts) == 1
+        assert tool_starts[0]["index"] == 1
+
+        # Stop reason should be tool_use
+        msg_deltas = [e for e in parsed if e.get("type") == "message_delta"]
+        assert cast(dict[str, Any], msg_deltas[0]["delta"])["stop_reason"] == "tool_use"
+
+    async def test_streaming_tool_block_stop_events(self):
+        chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = [
+            ToolCallChunk(
+                model=MODEL,
+                usage=None,
+                tool_calls=[
+                    ToolCallItem(name="fn1", arguments="{}"),
+                    ToolCallItem(name="fn2", arguments='{"a": 1}'),
+                ],
+            ),
+        ]
+        events: list[str] = []
+        async for event in generate_claude_stream(
+            COMMAND_ID, "test-model", _chunks_to_stream(chunks)
+        ):
+            events.append(event)
+
+        parsed = _parse_sse_events(events)
+
+        # Two tool block starts (at indices 1 and 2)
+        tool_starts = [
+            e
+            for e in parsed
+            if e.get("type") == "content_block_start"
+            and cast(dict[str, Any], e.get("content_block", {})).get("type")
+            == "tool_use"
+        ]
+        assert len(tool_starts) == 2
+        assert tool_starts[0]["index"] == 1
+        assert tool_starts[1]["index"] == 2
+
+        # Two tool block stops (at indices 1 and 2), plus text block stop at 0
+        block_stops = [e for e in parsed if e.get("type") == "content_block_stop"]
+        stop_indices = [e["index"] for e in block_stops]
+        assert 0 in stop_indices
+        assert 1 in stop_indices
+        assert 2 in stop_indices

src/exo/master/tests/test_master.py

@@ -7,15 +7,14 @@ from loguru import logger
 
 from exo.master.main import Master
 from exo.routing.router import get_node_id_keypair
-from exo.shared.models.model_cards import ModelCard, ModelId, ModelTask
-from exo.shared.types.api import ChatCompletionMessage, ChatCompletionTaskParams
+from exo.shared.models.model_cards import ModelCard, ModelTask
 from exo.shared.types.commands import (
-    ChatCompletion,
     CommandId,
     ForwarderCommand,
     PlaceInstance,
+    TextGeneration,
 )
-from exo.shared.types.common import NodeId, SessionId
+from exo.shared.types.common import ModelId, NodeId, SessionId
 from exo.shared.types.events import (
     ForwarderEvent,
     IndexedEvent,
@@ -27,8 +26,9 @@ from exo.shared.types.memory import Memory
 from exo.shared.types.profiling import (
     MemoryUsage,
 )
-from exo.shared.types.tasks import ChatCompletion as ChatCompletionTask
 from exo.shared.types.tasks import TaskStatus
+from exo.shared.types.tasks import TextGeneration as TextGenerationTask
+from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import (
     InstanceMeta,
     MlxRingInstance,
@@ -127,20 +127,16 @@ async def test_master():
         logger.info("wait for an instance")
         while len(master.state.instances.keys()) == 0:
             await anyio.sleep(0.001)
-        logger.info("inject a ChatCompletion Command")
+        logger.info("inject a TextGeneration Command")
         await command_sender.send(
             ForwarderCommand(
                 origin=node_id,
                 command=(
-                    ChatCompletion(
+                    TextGeneration(
                         command_id=CommandId(),
-                        request_params=ChatCompletionTaskParams(
-                            model="llama-3.2-1b",
-                            messages=[
-                                ChatCompletionMessage(
-                                    role="user", content="Hello, how are you?"
-                                )
-                            ],
+                        task_params=TextGenerationTaskParams(
+                            model=ModelId("llama-3.2-1b"),
+                            input="Hello, how are you?",
                         ),
                     )
                 ),
@@ -190,12 +186,10 @@ async def test_master():
         assert created_instance.ephemeral_port > 0
         assert isinstance(events[2].event, TaskCreated)
         assert events[2].event.task.task_status == TaskStatus.Pending
-        assert isinstance(events[2].event.task, ChatCompletionTask)
-        assert events[2].event.task.task_params == ChatCompletionTaskParams(
-            model="llama-3.2-1b",
-            messages=[
-                ChatCompletionMessage(role="user", content="Hello, how are you?")
-            ],
+        assert isinstance(events[2].event.task, TextGenerationTask)
+        assert events[2].event.task.task_params == TextGenerationTaskParams(
+            model=ModelId("llama-3.2-1b"),
+            input="Hello, how are you?",
         )
 
         await master.shutdown()

src/exo/master/tests/test_openai_responses_api.py

@@ -0,0 +1,48 @@
+"""Tests for OpenAI Responses API wire types.
+
+ResponsesRequest is the API wire type for the Responses endpoint.
+The responses adapter converts it to TextGenerationTaskParams for the pipeline.
+"""
+
+import pydantic
+import pytest
+
+from exo.shared.types.common import ModelId
+from exo.shared.types.openai_responses import (
+    ResponseInputMessage,
+    ResponsesRequest,
+)
+
+
+class TestResponsesRequestValidation:
+    """Tests for OpenAI Responses API request validation."""
+
+    def test_request_requires_model(self):
+        with pytest.raises(pydantic.ValidationError):
+            ResponsesRequest.model_validate(
+                {
+                    "input": "Hello",
+                }
+            )
+
+    def test_request_requires_input(self):
+        with pytest.raises(pydantic.ValidationError):
+            ResponsesRequest.model_validate(
+                {
+                    "model": "gpt-4o",
+                }
+            )
+
+    def test_request_accepts_string_input(self):
+        request = ResponsesRequest(
+            model=ModelId("gpt-4o"),
+            input="Hello",
+        )
+        assert request.input == "Hello"
+
+    def test_request_accepts_message_array_input(self):
+        request = ResponsesRequest(
+            model=ModelId("gpt-4o"),
+            input=[ResponseInputMessage(role="user", content="Hello")],
+        )
+        assert len(request.input) == 1

src/exo/master/tests/test_placement_utils.py

@@ -3,7 +3,6 @@ import pytest
 from exo.master.placement_utils import (
     allocate_layers_proportionally,
     filter_cycles_by_memory,
-    get_hosts_from_subgraph,
     get_mlx_jaccl_coordinators,
     get_shard_assignments,
     get_smallest_cycles,
@@ -14,7 +13,7 @@ from exo.master.tests.conftest import (
 )
 from exo.shared.models.model_cards import ModelCard, ModelId, ModelTask
 from exo.shared.topology import Topology
-from exo.shared.types.common import Host, NodeId
+from exo.shared.types.common import NodeId
 from exo.shared.types.memory import Memory
 from exo.shared.types.profiling import (
     NetworkInterfaceInfo,
@@ -273,45 +272,6 @@ def test_get_shard_assignments(
     )
 
 
-def test_get_hosts_from_subgraph():
-    # arrange
-    node_a_id = NodeId()
-    node_b_id = NodeId()
-    node_c_id = NodeId()
-    topology = Topology()
-
-    topology.add_node(node_a_id)
-    topology.add_node(node_b_id)
-    topology.add_node(node_c_id)
-
-    connection1 = Connection(
-        source=node_a_id, sink=node_b_id, edge=create_socket_connection(1)
-    )
-    connection2 = Connection(
-        source=node_b_id, sink=node_c_id, edge=create_socket_connection(2)
-    )
-    connection3 = Connection(
-        source=node_c_id, sink=node_a_id, edge=create_socket_connection(3)
-    )
-
-    topology.add_connection(connection1)
-    topology.add_connection(connection2)
-    topology.add_connection(connection3)
-
-    # act
-    hosts = get_hosts_from_subgraph(topology)
-
-    # assert
-    assert len(hosts) == 3
-    expected_hosts = [
-        Host(ip="169.254.0.1", port=1234),
-        Host(ip="169.254.0.2", port=1234),
-        Host(ip="169.254.0.3", port=1234),
-    ]
-    for expected_host in expected_hosts:
-        assert expected_host in hosts
-
-
 def test_get_mlx_jaccl_coordinators():
     # arrange
     node_a_id = NodeId()

src/exo/plugins/__init__.py

@@ -1,24 +0,0 @@
-"""Exo Plugin System.
-
-This module provides the plugin architecture for extending exo with custom
-workload types (simulations, ML frameworks, etc.) without modifying core code.
-"""
-
-from exo.plugins.base import EXOPlugin, PluginCommand, PluginInstance
-from exo.plugins.registry import PluginRegistry, discover_plugins
-from exo.plugins.type_registry import (
-    command_registry,
-    event_registry,
-    instance_registry,
-)
-
-__all__ = [
-    "EXOPlugin",
-    "PluginCommand",
-    "PluginInstance",
-    "PluginRegistry",
-    "discover_plugins",
-    "command_registry",
-    "event_registry",
-    "instance_registry",
-]

src/exo/plugins/base.py

@@ -1,171 +0,0 @@
-"""Base classes and protocols for Exo plugins."""
-
-from abc import ABC, abstractmethod
-from collections.abc import Callable, Mapping, Sequence
-from typing import TYPE_CHECKING, Any
-
-from pydantic import Field
-
-from exo.shared.types.common import CommandId
-from exo.shared.types.events import Event
-from exo.shared.types.tasks import Task
-from exo.shared.types.worker.instances import InstanceId
-from exo.shared.types.worker.runners import RunnerId
-from exo.utils.pydantic_ext import TaggedModel
-
-if TYPE_CHECKING:
-    from exo.shared.topology import Topology
-    from exo.shared.types.worker.instances import BaseInstance, BoundInstance
-    from exo.utils.channels import MpReceiver, MpSender
-    from exo.worker.runner.runner_supervisor import RunnerSupervisor
-
-
-class PluginCommand(TaggedModel):
-    """Base class for plugin-defined commands.
-
-    All plugin commands must inherit from this class. Commands are serialized
-    with their class name as a tag for routing.
-    """
-
-    command_id: CommandId = Field(default_factory=CommandId)
-
-
-class PluginInstance(TaggedModel):
-    """Base class for plugin-defined instances.
-
-    All plugin instances must inherit from this class. Plugins are expected
-    to define their own instance type with workload-specific fields.
-    """
-
-    instance_id: InstanceId
-
-
-class EXOPlugin(ABC):
-    """Protocol that all exo plugins must implement.
-
-    A plugin provides:
-    - Custom command types for API -> Master communication
-    - Custom instance types representing running workloads
-    - Placement logic for distributing work across nodes
-    - Planning logic for local task scheduling
-    - Runner implementation for executing work
-    """
-
-    @property
-    @abstractmethod
-    def name(self) -> str:
-        """Unique identifier for this plugin (e.g., 'flash', 'pytorch', 'mpi')."""
-        ...
-
-    @property
-    @abstractmethod
-    def version(self) -> str:
-        """Semantic version string (e.g., '1.0.0')."""
-        ...
-
-    # ========== Type Registration ==========
-
-    @abstractmethod
-    def get_command_types(self) -> Sequence[type]:
-        """Return command types this plugin handles.
-
-        These commands are routed to this plugin's process_command method.
-        Can return core BaseCommand types or PluginCommand types.
-        """
-        ...
-
-    @abstractmethod
-    def get_instance_type(self) -> type:
-        """Return the instance type this plugin creates.
-
-        This instance type is used for routing in planning and runner bootstrap.
-        Can return core Instance types or PluginInstance types.
-        """
-        ...
-
-    # ========== API Routes ==========
-
-    @abstractmethod
-    def get_api_routes(
-        self,
-    ) -> Sequence[tuple[str, str, Callable[..., Any]]]:
-        """Return FastAPI routes to register.
-
-        Each tuple: (method, path, handler)
-        Example: [('post', '/flash/launch', self.launch_handler)]
-
-        Handlers receive a PluginContext with access to:
-        - state: Current State object
-        - send_command: Async function to send commands
-        - node_id: Current node's ID
-        """
-        ...
-
-    # ========== Master Command Handling ==========
-
-    @abstractmethod
-    def handles_command(self, command: Any) -> bool:  # pyright: ignore[reportAny]
-        """Return True if this plugin handles the given command type."""
-        ...
-
-    @abstractmethod
-    def process_command(
-        self,
-        command: Any,  # pyright: ignore[reportAny]
-        topology: "Topology",
-        current_instances: Mapping[InstanceId, "BaseInstance"],
-    ) -> Sequence[Event]:
-        """Process a command and return events to emit.
-
-        Typically creates placement and returns InstanceCreated/InstanceDeleted events.
-
-        Args:
-            command: The command to process
-            topology: Current cluster topology
-            current_instances: Currently running instances
-
-        Returns:
-            Sequence of events to emit (e.g., InstanceCreated, InstanceDeleted)
-        """
-        ...
-
-    # ========== Worker Planning ==========
-
-    @abstractmethod
-    def handles_instance(self, instance: object) -> bool:
-        """Return True if this plugin manages the given instance type."""
-        ...
-
-    @abstractmethod
-    def plan_task(
-        self,
-        runners: Mapping[RunnerId, "RunnerSupervisor"],
-        instances: Mapping[InstanceId, "BaseInstance"],
-    ) -> Task | None:
-        """Plan the next task for plugin instances.
-
-        Called during each planning cycle.
-        Return None if no task is needed.
-        """
-        ...
-
-    @abstractmethod
-    def should_skip_download(self, instance: object) -> bool:
-        """Return True if this instance type doesn't need model downloads."""
-        ...
-
-    # ========== Runner Bootstrap ==========
-
-    @abstractmethod
-    def create_runner(
-        self,
-        bound_instance: "BoundInstance",
-        event_sender: "MpSender[Event]",
-        task_receiver: "MpReceiver[Task]",
-    ) -> None:
-        """Entry point for the runner process.
-
-        Called in a subprocess to execute the actual workload.
-        This function should block until the workload completes.
-        """
-        ...

src/exo/plugins/context.py

@@ -1,21 +0,0 @@
-"""Context objects passed to plugin handlers."""
-
-from collections.abc import Awaitable, Callable
-from dataclasses import dataclass
-
-from exo.shared.types.commands import BaseCommand
-from exo.shared.types.common import NodeId
-from exo.shared.types.state import State
-
-
-@dataclass
-class PluginContext:
-    """Context provided to plugin API handlers.
-
-    This gives plugins access to the current state and the ability to send
-    commands without direct access to internal API components.
-    """
-
-    state: State
-    send_command: Callable[[BaseCommand], Awaitable[None]]
-    node_id: NodeId

src/exo/plugins/implementations/__init__.py

@@ -1,5 +0,0 @@
-"""Plugin implementations directory.
-
-Each subdirectory should contain a plugin with a register() function
-that returns an EXOPlugin instance.
-"""

src/exo/plugins/implementations/flash/__init__.py

@@ -1,15 +0,0 @@
-"""FLASH Plugin - MPI-based simulation support for Exo."""
-
-from exo.plugins.implementations.flash.plugin import FLASHPlugin
-from exo.plugins.implementations.flash.types import (
-    FLASHInstance,
-    LaunchFLASH,
-    StopFLASH,
-)
-
-__all__ = ["FLASHPlugin", "FLASHInstance", "LaunchFLASH", "StopFLASH", "register"]
-
-
-def register() -> FLASHPlugin:
-    """Entry point for plugin discovery."""
-    return FLASHPlugin()

src/exo/plugins/implementations/flash/api_handlers.py

@@ -1,109 +0,0 @@
-"""FLASH plugin API handlers."""
-
-from typing import Any
-
-from fastapi import HTTPException
-
-from exo.plugins.context import PluginContext
-from exo.plugins.implementations.flash.types import (
-    FLASHInstance,
-    LaunchFLASH,
-    StopFLASH,
-)
-
-
-async def handle_launch_flash(
-    ctx: PluginContext,
-    simulation_name: str,
-    flash_executable_path: str,
-    working_directory: str,
-    parameter_file_path: str = "",
-    ranks_per_node: int = 1,
-    min_nodes: int = 1,
-    hosts: str = "",
-) -> dict[str, str]:
-    """Launch a FLASH MPI simulation across the cluster.
-
-    Args:
-        ctx: Plugin context with state and send_command
-        simulation_name: Name of the simulation
-        flash_executable_path: Path to the FLASH executable
-        working_directory: Working directory for the simulation
-        parameter_file_path: Path to parameter file (optional)
-        ranks_per_node: Number of MPI ranks per node
-        min_nodes: Minimum number of nodes required
-        hosts: Optional comma-separated hostnames (e.g., "s14,james21-1").
-               If not provided, IPs are discovered from topology edges.
-    """
-    command = LaunchFLASH(
-        simulation_name=simulation_name,
-        flash_executable_path=flash_executable_path,
-        parameter_file_path=parameter_file_path,
-        working_directory=working_directory,
-        ranks_per_node=ranks_per_node,
-        min_nodes=min_nodes,
-        hosts=hosts,
-    )
-    await ctx.send_command(command)
-
-    return {
-        "message": "FLASH launch command received",
-        "command_id": str(command.command_id),
-        "simulation_name": simulation_name,
-    }
-
-
-async def handle_stop_flash(
-    ctx: PluginContext,
-    instance_id: str,
-) -> dict[str, str]:
-    """Stop a running FLASH simulation."""
-    from exo.shared.types.worker.instances import InstanceId
-
-    inst_id = InstanceId(instance_id)
-
-    if inst_id not in ctx.state.instances:
-        raise HTTPException(status_code=404, detail="Instance not found")
-
-    instance = ctx.state.instances[inst_id]
-    if not isinstance(instance, FLASHInstance):
-        raise HTTPException(
-            status_code=400, detail="Instance is not a FLASH simulation"
-        )
-
-    command = StopFLASH(instance_id=inst_id)
-    await ctx.send_command(command)
-
-    return {
-        "message": "Stop command received",
-        "command_id": str(command.command_id),
-        "instance_id": str(instance_id),
-    }
-
-
-async def handle_list_flash_instances(ctx: PluginContext) -> list[dict[str, Any]]:
-    """List all FLASH simulation instances."""
-    flash_instances: list[dict[str, Any]] = []
-    for instance_id, instance in ctx.state.instances.items():
-        if isinstance(instance, FLASHInstance):
-            # Get runner statuses for this instance
-            runner_statuses: dict[str, str | None] = {}
-            for (
-                node_id,
-                runner_id,
-            ) in instance.shard_assignments.node_to_runner.items():
-                runner_status = ctx.state.runners.get(runner_id)
-                runner_statuses[str(node_id)] = (
-                    str(runner_status) if runner_status else None
-                )
-
-            flash_instances.append(
-                {
-                    "instance_id": str(instance_id),
-                    "simulation_name": instance.simulation_name,
-                    "total_ranks": instance.total_ranks,
-                    "working_directory": instance.working_directory,
-                    "runner_statuses": runner_statuses,
-                }
-            )
-    return flash_instances

src/exo/plugins/implementations/flash/placement.py

@@ -1,167 +0,0 @@
-"""FLASH plugin placement logic."""
-
-from collections.abc import Mapping
-from copy import deepcopy
-
-from loguru import logger
-
-from exo.plugins.implementations.flash.types import FLASHInstance, LaunchFLASH
-from exo.shared.models.model_cards import ModelCard
-from exo.shared.topology import Topology
-from exo.shared.types.common import Host, ModelId, NodeId
-from exo.shared.types.memory import Memory
-from exo.shared.types.topology import SocketConnection
-from exo.shared.types.worker.instances import BaseInstance, InstanceId
-from exo.shared.types.worker.runners import (
-    RunnerId,
-    ShardAssignments,
-)
-from exo.shared.types.worker.shards import PipelineShardMetadata
-
-
-def place_flash_instance(
-    command: LaunchFLASH,
-    topology: Topology,
-    current_instances: Mapping[InstanceId, BaseInstance],
-) -> dict[InstanceId, BaseInstance]:
-    """Place a FLASH simulation instance across available nodes.
-
-    Unlike MLX instances which use ring/JACCL topology for tensor parallelism,
-    FLASH instances use MPI for communication. We just need to provide the
-    node IPs so the runner can generate an MPI hostfile.
-    """
-    instance_id = InstanceId()
-    target_instances: dict[InstanceId, BaseInstance] = dict(deepcopy(current_instances))
-
-    all_nodes = list(topology.list_nodes())
-
-    if len(all_nodes) < command.min_nodes:
-        raise ValueError(
-            f"Not enough nodes: need {command.min_nodes}, have {len(all_nodes)}"
-        )
-
-    # Select nodes (take the first min_nodes)
-    selected_nodes = all_nodes[: command.min_nodes]
-
-    logger.info(
-        f"Placing FLASH instance '{command.simulation_name}' on {len(selected_nodes)} nodes"
-    )
-
-    # Build shard assignments (one runner per node for FLASH)
-    runner_to_shard: dict[RunnerId, PipelineShardMetadata] = {}
-    node_to_runner: dict[NodeId, RunnerId] = {}
-
-    # Create a dummy ModelCard for FLASH (required by ShardMetadata interface)
-    flash_model_card = ModelCard(
-        model_id=ModelId(command.simulation_name),
-        storage_size=Memory(in_bytes=0),
-        n_layers=1,
-        hidden_size=1,
-        supports_tensor=False,
-        tasks=[],
-    )
-
-    for i, node_id in enumerate(selected_nodes):
-        runner_id = RunnerId()
-        node_to_runner[node_id] = runner_id
-        runner_to_shard[runner_id] = PipelineShardMetadata(
-            device_rank=i,
-            world_size=len(selected_nodes),
-            model_card=flash_model_card,
-            start_layer=0,
-            end_layer=1,
-            n_layers=1,
-        )
-
-    shard_assignments = ShardAssignments(
-        model_id=ModelId(command.simulation_name),
-        runner_to_shard=runner_to_shard,
-        node_to_runner=node_to_runner,
-    )
-
-    # Build hosts_by_node - get hostnames/IPs for MPI hostfile generation
-    hosts_by_node: dict[NodeId, list[Host]] = {}
-
-    # If explicit hosts are provided, use them directly
-    if command.hosts:
-        explicit_hosts = [h.strip() for h in command.hosts.split(",") if h.strip()]
-        logger.info(f"FLASH placement: explicit hosts provided: {explicit_hosts}")
-        for i, node_id in enumerate(selected_nodes):
-            if i < len(explicit_hosts):
-                hosts_by_node[node_id] = [Host(ip=explicit_hosts[i], port=0)]
-                logger.info(
-                    f"FLASH placement: node {node_id} (rank {i}) -> IP {explicit_hosts[i]}"
-                )
-            else:
-                logger.warning(
-                    f"Not enough hosts provided for node {i}, using localhost"
-                )
-                hosts_by_node[node_id] = [Host(ip="127.0.0.1", port=0)]
-        logger.info(
-            f"FLASH placement: coordinator will be rank 0 at IP {explicit_hosts[0]}"
-        )
-    else:
-        # Get each node's own IP by looking at how OTHER nodes see it.
-        # out_edges(A) gives edges A->B with B's IP in sink_multiaddr,
-        # so to find A's IP we look at edges B->A from any other node.
-        for node_id in selected_nodes:
-            candidate_ips: set[str] = set()
-
-            for other_node in all_nodes:
-                if other_node == node_id:
-                    continue
-                for conn in topology.out_edges(other_node):
-                    if conn.sink == node_id and isinstance(conn.edge, SocketConnection):
-                        ip = conn.edge.sink_multiaddr.ip_address
-                        # Skip link-local and localhost addresses
-                        if not ip.startswith("169.254.") and not ip.startswith("127."):
-                            candidate_ips.add(ip)
-
-            # Prefer private network IPs (10.x, 192.168.x) over Tailscale CGNAT (100.64-127.x)
-            chosen_ip: str | None = None
-            for ip in candidate_ips:
-                if ip.startswith(("10.", "192.168.")):
-                    chosen_ip = ip
-                    break
-            if chosen_ip is None and candidate_ips:
-                chosen_ip = next(iter(candidate_ips))
-
-            if chosen_ip:
-                hosts_by_node[node_id] = [Host(ip=chosen_ip, port=0)]
-            else:
-                logger.warning(
-                    f"Could not determine IP for node {node_id}, using localhost"
-                )
-                hosts_by_node[node_id] = [Host(ip="127.0.0.1", port=0)]
-
-            logger.info(
-                f"FLASH placement: node {node_id} -> IP {hosts_by_node[node_id][0].ip}"
-                f" (candidates: {candidate_ips})"
-            )
-
-    total_ranks = len(selected_nodes) * command.ranks_per_node
-
-    # Determine coordinator IP - first node's first host IP
-    first_node_id: NodeId = next(iter(hosts_by_node.keys()))
-    coordinator_ip: str = (
-        hosts_by_node[first_node_id][0].ip
-        if hosts_by_node[first_node_id]
-        else "127.0.0.1"
-    )
-
-    target_instances[instance_id] = FLASHInstance(
-        instance_id=instance_id,
-        shard_assignments=shard_assignments,
-        hosts_by_node=hosts_by_node,
-        flash_executable_path=command.flash_executable_path,
-        parameter_file_path=command.parameter_file_path,
-        working_directory=command.working_directory,
-        ranks_per_node=command.ranks_per_node,
-        total_ranks=total_ranks,
-        simulation_name=command.simulation_name,
-        coordinator_ip=coordinator_ip,
-    )
-
-    logger.info(f"Created FLASH instance {instance_id} with {total_ranks} total ranks")
-
-    return target_instances

src/exo/plugins/implementations/flash/planning.py

@@ -1,37 +0,0 @@
-"""FLASH plugin planning logic."""
-
-from collections.abc import Mapping
-
-from exo.plugins.implementations.flash.types import FLASHInstance
-from exo.shared.types.tasks import LoadModel, Task
-from exo.shared.types.worker.instances import BaseInstance, InstanceId
-from exo.shared.types.worker.runners import RunnerId, RunnerIdle
-from exo.worker.runner.runner_supervisor import RunnerSupervisor
-
-
-def plan_flash(
-    runners: Mapping[RunnerId, RunnerSupervisor],
-    instances: Mapping[InstanceId, BaseInstance],
-) -> Task | None:
-    """Plan tasks specifically for FLASH instances.
-
-    FLASH instances have a simpler lifecycle:
-    - CreateRunner (handled by core _create_runner)
-    - LoadModel (starts the simulation immediately)
-    - Shutdown (handled by core _kill_runner)
-
-    This function handles the LoadModel step for FLASH instances,
-    skipping the MLX-specific download/init/warmup steps.
-    """
-    for runner in runners.values():
-        instance = runner.bound_instance.instance
-
-        # Only handle FLASH instances
-        if not isinstance(instance, FLASHInstance):
-            continue
-
-        # If runner is idle, emit LoadModel to start the simulation
-        if isinstance(runner.status, RunnerIdle):
-            return LoadModel(instance_id=instance.instance_id)
-
-    return None

src/exo/plugins/implementations/flash/plugin.py

@@ -1,98 +0,0 @@
-"""FLASH Plugin - Main plugin class."""
-
-from collections.abc import Callable, Mapping, Sequence
-from typing import Any
-
-from exo.plugins.base import EXOPlugin
-from exo.plugins.implementations.flash.api_handlers import (
-    handle_launch_flash,
-    handle_list_flash_instances,
-    handle_stop_flash,
-)
-from exo.plugins.implementations.flash.placement import place_flash_instance
-from exo.plugins.implementations.flash.planning import plan_flash
-from exo.plugins.implementations.flash.runner import main as flash_runner_main
-from exo.plugins.implementations.flash.types import (
-    FLASHInstance,
-    LaunchFLASH,
-    StopFLASH,
-)
-from exo.shared.topology import Topology
-from exo.shared.types.commands import DeleteInstance
-from exo.shared.types.events import Event
-from exo.shared.types.tasks import Task
-from exo.shared.types.worker.instances import BaseInstance, BoundInstance, InstanceId
-from exo.shared.types.worker.runners import RunnerId
-from exo.utils.channels import MpReceiver, MpSender
-from exo.worker.runner.runner_supervisor import RunnerSupervisor
-
-
-class FLASHPlugin(EXOPlugin):
-    """Plugin for FLASH MPI simulations."""
-
-    @property
-    def name(self) -> str:
-        return "flash"
-
-    @property
-    def version(self) -> str:
-        return "1.0.0"
-
-    def get_command_types(self) -> Sequence[type]:
-        return [LaunchFLASH, StopFLASH]
-
-    def get_instance_type(self) -> type:
-        return FLASHInstance
-
-    def get_api_routes(
-        self,
-    ) -> Sequence[tuple[str, str, Callable[..., Any]]]:
-        return [
-            ("post", "/flash/launch", handle_launch_flash),
-            ("delete", "/flash/{instance_id}", handle_stop_flash),
-            ("get", "/flash/instances", handle_list_flash_instances),
-        ]
-
-    def handles_command(self, command: Any) -> bool:  # pyright: ignore[reportAny]
-        return isinstance(command, (LaunchFLASH, StopFLASH))
-
-    def process_command(
-        self,
-        command: Any,  # pyright: ignore[reportAny]
-        topology: Topology,
-        current_instances: Mapping[InstanceId, BaseInstance],
-    ) -> Sequence[Event]:
-        from exo.master.placement import delete_instance, get_transition_events
-
-        if isinstance(command, LaunchFLASH):
-            placement = place_flash_instance(command, topology, current_instances)
-            return list(get_transition_events(current_instances, placement))
-        elif isinstance(command, StopFLASH):
-            placement = delete_instance(
-                DeleteInstance(instance_id=command.instance_id),
-                current_instances,
-            )
-            return list(get_transition_events(current_instances, placement))
-        return []
-
-    def handles_instance(self, instance: object) -> bool:
-        return isinstance(instance, FLASHInstance)
-
-    def plan_task(
-        self,
-        runners: Mapping[RunnerId, RunnerSupervisor],
-        instances: Mapping[InstanceId, BaseInstance],
-    ) -> Task | None:
-        return plan_flash(runners, instances)
-
-    def should_skip_download(self, instance: object) -> bool:
-        # FLASH instances don't need model downloads
-        return True
-
-    def create_runner(
-        self,
-        bound_instance: BoundInstance,
-        event_sender: MpSender[Event],
-        task_receiver: MpReceiver[Task],
-    ) -> None:
-        flash_runner_main(bound_instance, event_sender, task_receiver)

src/exo/plugins/implementations/flash/runner.py

@@ -1,305 +0,0 @@
-"""FLASH MPI Runner - spawns and monitors FLASH simulations.
-
-Exo-native distributed MPI:
-- Exo handles node discovery and coordination
-- Coordinator generates hostfile from Exo topology
-- mpirun uses exo-rsh (no SSH required) to spawn on remote nodes
-- exo-rsh connects to each node's Exo API (/execute endpoint) for remote execution
-- Workers just report ready and wait
-"""
-# ruff: noqa: I001 - Import order intentional (plugin types before shared types)
-
-import os
-import shutil
-import socket
-import subprocess
-import threading
-
-from loguru import logger
-
-from exo.shared.types.events import (
-    Event,
-    RunnerStatusUpdated,
-    TaskAcknowledged,
-    TaskStatusUpdated,
-)
-from exo.shared.types.tasks import (
-    LoadModel,
-    Shutdown,
-    Task,
-    TaskStatus,
-)
-from exo.plugins.implementations.flash.types import FLASHInstance
-from exo.shared.types.worker.instances import BoundInstance
-from exo.shared.types.worker.runners import (
-    RunnerFailed,
-    RunnerIdle,
-    RunnerLoading,
-    RunnerReady,
-    RunnerRunning,
-    RunnerShutdown,
-    RunnerShuttingDown,
-    RunnerStatus,
-)
-from exo.utils.channels import MpReceiver, MpSender
-
-# Find mpirun in PATH, fallback to common locations
-MPIRUN_PATH = shutil.which("mpirun") or "/opt/homebrew/bin/mpirun"
-
-# exo-rsh is installed as console script by exo package
-_exo_rsh_path = shutil.which("exo-rsh")
-if not _exo_rsh_path:
-    raise RuntimeError("exo-rsh not found in PATH - this should be installed with exo")
-EXO_RSH_PATH: str = _exo_rsh_path
-
-
-def get_my_rank(instance: FLASHInstance, my_node_id: str) -> int:
-    """Determine this node's rank based on position in hosts_by_node."""
-    for i, node_id in enumerate(instance.hosts_by_node.keys()):
-        if str(node_id) == str(my_node_id):
-            return i
-    return -1
-
-
-def get_coordinator_host(instance: FLASHInstance) -> str:
-    """Get the IP of the coordinator node."""
-    return instance.coordinator_ip
-
-
-def resolve_host(host: str) -> str:
-    """Resolve host string to a usable hostname for MPI hostfile.
-
-    Accepts either an IP address or hostname. For IPs, attempts to resolve
-    to a hostname via DNS/mDNS. Hostnames are returned as-is after validation.
-    """
-    # Check if input is already a hostname (not an IP)
-    try:
-        socket.inet_aton(host)
-        is_ip = True
-    except socket.error:
-        is_ip = False
-
-    if not is_ip:
-        # Already a hostname, verify it resolves and return as-is
-        try:
-            socket.gethostbyname(host)
-            return host
-        except socket.gaierror:
-            logger.warning(f"Hostname {host} does not resolve, using anyway")
-            return host
-
-    # It's an IP address, try to resolve to hostname
-    try:
-        hostname, _, _ = socket.gethostbyaddr(host)
-        hostname = hostname.split(".")[0]
-        logger.info(f"Resolved {host} to {hostname}")
-        return hostname
-    except socket.herror:
-        pass
-
-    # Fall back to IP
-    logger.warning(f"Could not resolve {host} to hostname, using IP directly")
-    return host
-
-
-def generate_hostfile(instance: FLASHInstance, working_dir: str) -> str:
-    """Generate MPI hostfile from instance topology."""
-    hostfile_path = os.path.join(working_dir, "flash_hosts.txt")
-    with open(hostfile_path, "w") as f:
-        for _node_id, hosts in instance.hosts_by_node.items():
-            if hosts:
-                host = resolve_host(hosts[0].ip)
-                f.write(f"{host} slots={instance.ranks_per_node}\n")
-    logger.info(f"Generated hostfile at {hostfile_path}")
-    with open(hostfile_path, "r") as f:
-        logger.info(f"Hostfile contents:\n{f.read()}")
-    return hostfile_path
-
-
-def main(
-    bound_instance: BoundInstance,
-    event_sender: MpSender[Event],
-    task_receiver: MpReceiver[Task],
-) -> None:
-    """Main FLASH runner loop.
-
-    Coordinator: generates hostfile and runs mpirun (uses exo-rsh instead of SSH)
-    Workers: just report ready and wait for mpirun to spawn processes on them
-    """
-    assert isinstance(bound_instance.instance, FLASHInstance)
-    instance = bound_instance.instance
-    runner_id = bound_instance.bound_runner_id
-    my_node_id = str(bound_instance.bound_node_id)
-
-    logger.info(f"FLASH runner starting for simulation: {instance.simulation_name}")
-
-    my_rank = get_my_rank(instance, my_node_id)
-    world_size = len(instance.hosts_by_node)
-    is_coordinator = my_rank == 0
-    coordinator_ip = get_coordinator_host(instance)
-
-    logger.info(
-        f"FLASH node: rank={my_rank}, world_size={world_size}, coordinator={is_coordinator}"
-    )
-    logger.info(f"FLASH coordinator IP: {coordinator_ip}")
-
-    process: subprocess.Popen[bytes] | None = None
-    current_status: RunnerStatus = RunnerIdle()
-    shutdown_requested = False
-
-    event_sender.send(
-        RunnerStatusUpdated(runner_id=runner_id, runner_status=current_status)
-    )
-
-    def monitor_output(proc: subprocess.Popen[bytes]) -> None:
-        """Monitor FLASH stdout for progress updates."""
-        if proc.stdout is None:
-            return
-        for line in iter(proc.stdout.readline, b""):
-            if shutdown_requested:
-                break
-            try:
-                decoded: str = line.decode("utf-8", errors="replace").strip()
-                if decoded:
-                    logger.info(f"[FLASH] {decoded}")
-            except Exception as e:
-                logger.warning(f"Error parsing FLASH output: {e}")
-
-    with task_receiver as tasks:
-        for task in tasks:
-            event_sender.send(
-                TaskStatusUpdated(task_id=task.task_id, task_status=TaskStatus.Running)
-            )
-            event_sender.send(TaskAcknowledged(task_id=task.task_id))
-
-            match task:
-                case LoadModel() if isinstance(current_status, RunnerIdle):
-                    current_status = RunnerLoading()
-                    logger.info("Starting FLASH simulation")
-                    event_sender.send(
-                        RunnerStatusUpdated(
-                            runner_id=runner_id, runner_status=current_status
-                        )
-                    )
-
-                    try:
-                        if is_coordinator:
-                            # Coordinator: generate hostfile and run mpirun
-                            hostfile = generate_hostfile(
-                                instance, instance.working_directory
-                            )
-
-                            iface = instance.network_interface
-                            cmd = [
-                                MPIRUN_PATH,
-                                "-np",
-                                str(instance.total_ranks),
-                                "--hostfile",
-                                hostfile,
-                                "--wdir",
-                                instance.working_directory,
-                                "--oversubscribe",
-                                "--allow-run-as-root",
-                                "--mca",
-                                "btl",
-                                "tcp,self",
-                                "--mca",
-                                "btl_tcp_if_include",
-                                iface,
-                                "--mca",
-                                "oob_tcp_if_include",
-                                iface,
-                                "--mca",
-                                "plm_rsh_no_tree_spawn",
-                                "1",
-                            ]
-
-                            # Use exo-rsh for remote execution (no SSH needed)
-                            cmd.extend(["--mca", "plm_rsh_agent", EXO_RSH_PATH])
-
-                            cmd.append(instance.flash_executable_path)
-
-                            logger.info(f"FLASH distributed launch: {' '.join(cmd)}")
-
-                            process = subprocess.Popen(
-                                cmd,
-                                cwd=instance.working_directory,
-                                stdout=subprocess.PIPE,
-                                stderr=subprocess.STDOUT,
-                            )
-
-                            monitor_thread = threading.Thread(
-                                target=monitor_output, args=(process,), daemon=True
-                            )
-                            monitor_thread.start()
-
-                            current_status = RunnerRunning()
-                            logger.info(
-                                f"FLASH running on {world_size} nodes with {instance.total_ranks} ranks"
-                            )
-
-                        else:
-                            # Worker: mpirun on coordinator will use exo-rsh to spawn processes here
-                            logger.info(
-                                f"Worker {my_rank}: Ready for mpirun to spawn processes via exo-rsh"
-                            )
-                            current_status = RunnerRunning()
-
-                    except Exception as e:
-                        logger.error(f"Failed to start FLASH: {e}")
-                        import traceback
-
-                        logger.error(traceback.format_exc())
-                        current_status = RunnerFailed(error_message=str(e))
-
-                case Shutdown():
-                    shutdown_requested = True
-                    current_status = RunnerShuttingDown()
-                    logger.info("FLASH runner shutting down")
-                    event_sender.send(
-                        RunnerStatusUpdated(
-                            runner_id=runner_id, runner_status=current_status
-                        )
-                    )
-
-                    if process and process.poll() is None:
-                        logger.info("Terminating FLASH simulation")
-                        process.terminate()
-                        try:
-                            process.wait(timeout=10)
-                        except subprocess.TimeoutExpired:
-                            logger.warning("FLASH didn't terminate, killing")
-                            process.kill()
-                            process.wait()
-
-                    current_status = RunnerShutdown()
-
-                case _:
-                    if process and process.poll() is not None:
-                        exit_code = process.returncode
-                        if exit_code == 0:
-                            logger.info("FLASH simulation completed successfully")
-                            current_status = RunnerReady()
-                        else:
-                            logger.error(
-                                f"FLASH simulation failed with code {exit_code}"
-                            )
-                            current_status = RunnerFailed(
-                                error_message=f"Exit code {exit_code}"
-                            )
-
-            event_sender.send(
-                TaskStatusUpdated(task_id=task.task_id, task_status=TaskStatus.Complete)
-            )
-            event_sender.send(
-                RunnerStatusUpdated(runner_id=runner_id, runner_status=current_status)
-            )
-
-            if isinstance(current_status, RunnerShutdown):
-                break
-
-    if process and process.poll() is None:
-        process.terminate()
-        process.wait(timeout=5)
-
-    logger.info("FLASH runner exiting")

src/exo/plugins/implementations/flash/types.py

@@ -1,62 +0,0 @@
-"""FLASH plugin types - commands and instances."""
-
-from exo.plugins.type_registry import command_registry, instance_registry
-from exo.shared.types.commands import BaseCommand
-from exo.shared.types.common import Host, NodeId
-from exo.shared.types.worker.instances import BaseInstance, InstanceId
-from exo.shared.types.worker.runners import RunnerId
-from exo.shared.types.worker.shards import ShardMetadata
-
-# ============================================================================
-# Commands
-# ============================================================================
-
-
-@command_registry.register
-class LaunchFLASH(BaseCommand):
-    """Command to launch a FLASH MPI simulation."""
-
-    simulation_name: str
-    flash_executable_path: str
-    parameter_file_path: str
-    working_directory: str
-    ranks_per_node: int = 1
-    min_nodes: int = 1
-    # Optional: explicit hostnames for MPI (e.g., "s14,james21-1")
-    # Used when topology edges don't contain IP addresses
-    hosts: str = ""
-
-
-@command_registry.register
-class StopFLASH(BaseCommand):
-    """Command to stop a running FLASH simulation."""
-
-    instance_id: InstanceId
-
-
-# ============================================================================
-# Instances
-# ============================================================================
-
-
-@instance_registry.register
-class FLASHInstance(BaseInstance):
-    """Instance for FLASH MPI simulation.
-
-    Unlike MLX instances which do tensor parallelism, FLASH instances
-    coordinate MPI processes across nodes. Each node runs one or more
-    MPI ranks of the FLASH simulation.
-    """
-
-    hosts_by_node: dict[NodeId, list[Host]]
-    flash_executable_path: str
-    parameter_file_path: str
-    working_directory: str
-    ranks_per_node: int = 1
-    total_ranks: int
-    simulation_name: str
-    coordinator_ip: str
-    network_interface: str = "en0"  # Network interface for MPI (e.g., en0, eth0)
-
-    def shard(self, runner_id: RunnerId) -> ShardMetadata | None:
-        return self.shard_assignments.runner_to_shard.get(runner_id, None)

src/exo/plugins/registry.py

@@ -1,110 +0,0 @@
-"""Plugin registry for discovering and managing plugins."""
-
-from collections.abc import Callable, Sequence
-from typing import Any
-
-from loguru import logger
-
-from exo.plugins.base import EXOPlugin
-
-
-class PluginRegistry:
-    """Central registry for all plugins."""
-
-    _instance: "PluginRegistry | None" = None
-
-    def __init__(self) -> None:
-        self._plugins: dict[str, EXOPlugin] = {}
-        self._command_handlers: dict[type, EXOPlugin] = {}
-        self._instance_handlers: dict[type, EXOPlugin] = {}
-
-    @classmethod
-    def get(cls) -> "PluginRegistry":
-        """Get the singleton registry instance."""
-        if cls._instance is None:
-            cls._instance = cls()
-        return cls._instance
-
-    @classmethod
-    def reset(cls) -> None:
-        """Reset the singleton instance (useful for testing)."""
-        cls._instance = None
-
-    def register(self, plugin: EXOPlugin) -> None:
-        """Register a plugin and its types."""
-        if plugin.name in self._plugins:
-            raise ValueError(f"Plugin '{plugin.name}' already registered")
-
-        logger.info(f"Registering plugin: {plugin.name} v{plugin.version}")
-
-        self._plugins[plugin.name] = plugin
-
-        # Register command handlers
-        for cmd_type in plugin.get_command_types():
-            self._command_handlers[cmd_type] = plugin
-            logger.debug(f"  Registered command: {cmd_type.__name__}")
-
-        # Register instance handler
-        instance_type = plugin.get_instance_type()
-        self._instance_handlers[instance_type] = plugin
-        logger.debug(f"  Registered instance: {instance_type.__name__}")
-
-    def get_plugin(self, name: str) -> EXOPlugin | None:
-        """Get a plugin by name."""
-        return self._plugins.get(name)
-
-    def get_plugin_for_command(self, command: object) -> EXOPlugin | None:
-        """Get the plugin that handles a command."""
-        for plugin in self._plugins.values():
-            if plugin.handles_command(command):
-                return plugin
-        return None
-
-    def get_plugin_for_instance(self, instance: object) -> EXOPlugin | None:
-        """Get the plugin that manages an instance."""
-        for plugin in self._plugins.values():
-            if plugin.handles_instance(instance):
-                return plugin
-        return None
-
-    def all_plugins(self) -> Sequence[EXOPlugin]:
-        """Get all registered plugins."""
-        return list(self._plugins.values())
-
-    def get_all_api_routes(
-        self,
-    ) -> Sequence[tuple[str, str, Callable[..., Any], EXOPlugin]]:
-        """Get all API routes from all plugins."""
-        routes: list[tuple[str, str, Callable[..., Any], EXOPlugin]] = []
-        for plugin in self._plugins.values():
-            for method, path, handler in plugin.get_api_routes():
-                routes.append((method, path, handler, plugin))
-        return routes
-
-
-def discover_plugins() -> None:
-    """Auto-discover and register plugins from the implementations directory.
-
-    Plugins should have a register() function that returns an EXOPlugin instance.
-    """
-    import importlib
-    import pkgutil
-
-    registry = PluginRegistry.get()
-
-    try:
-        import exo.plugins.implementations as impl_package
-
-        for _, module_name, _ in pkgutil.iter_modules(impl_package.__path__):
-            try:
-                module = importlib.import_module(
-                    f"exo.plugins.implementations.{module_name}"
-                )
-                if hasattr(module, "register"):
-                    plugin = module.register()  # pyright: ignore[reportAny]
-                    if plugin is not None:
-                        registry.register(plugin)  # pyright: ignore[reportAny]
-            except Exception as e:
-                logger.warning(f"Failed to load plugin {module_name}: {e}")
-    except ImportError:
-        logger.debug("No plugin implementations package found")

src/exo/plugins/type_registry.py

@@ -1,84 +0,0 @@
-"""Dynamic type registry for plugin types.
-
-This module provides a registry system that allows plugins to register their
-command and instance types dynamically, eliminating the need for static union
-types and avoiding circular imports.
-"""
-
-from typing import TypeVar
-
-from loguru import logger
-
-from exo.utils.pydantic_ext import CamelCaseModel
-
-# TypeVar for preserving exact types through the register decorator
-_TCls = TypeVar("_TCls", bound=type[CamelCaseModel])
-
-
-class TypeRegistry[T: CamelCaseModel]:
-    """Registry for dynamically registered Pydantic types.
-
-    Enables plugins to register their types at import time. Deserialization
-    uses the class name from the tagged JSON format to look up the correct type.
-    """
-
-    def __init__(self, name: str) -> None:
-        self._name = name
-        self._types: dict[str, type[T]] = {}
-
-    def register(self, cls: _TCls) -> _TCls:
-        """Decorator to register a type with this registry.
-
-        Preserves the exact type through the decorator for proper type checking.
-        """
-        self._types[cls.__name__] = cls  # type: ignore[assignment]
-        logger.debug(f"{self._name}: registered {cls.__name__}")
-        return cls
-
-    def get(self, name: str) -> type[T] | None:
-        """Look up a type by class name."""
-        return self._types.get(name)
-
-    def all_types(self) -> dict[str, type[T]]:
-        """Return all registered types."""
-        return dict(self._types)
-
-    def deserialize(self, data: dict[str, dict[str, object]] | CamelCaseModel) -> T:
-        """Deserialize dict to the appropriate registered type.
-
-        Supports two formats:
-        1. Tagged format: {"ClassName": {...fields...}} - used for network serialization
-        2. Flat format: {...fields...} - used for API requests, tries each type
-        """
-        # If already deserialized (e.g., from Pydantic), return as-is
-        if isinstance(data, CamelCaseModel):
-            return data  # type: ignore[return-value]
-
-        # Check for tagged format: single key that matches a registered type
-        if len(data) == 1:
-            class_name: str = next(iter(data.keys()))
-            cls = self._types.get(class_name)
-            if cls is not None:
-                return cls.model_validate(data[class_name], strict=False)
-
-        # Flat format: try each registered type, use first that validates
-        errors: list[str] = []
-        for type_name, cls in self._types.items():
-            try:
-                return cls.model_validate(data, strict=False)
-            except Exception as e:  # noqa: BLE001
-                errors.append(f"{type_name}: {e}")
-
-        # None matched - provide helpful error
-        available = ", ".join(self._types.keys())
-        raise ValueError(
-            f"{self._name}: could not deserialize data. "
-            f"Available types: {available}. Errors: {'; '.join(errors[:3])}"
-        )
-
-
-# Global registries for commands, instances, events, and tasks
-command_registry: TypeRegistry[CamelCaseModel] = TypeRegistry("CommandRegistry")
-instance_registry: TypeRegistry[CamelCaseModel] = TypeRegistry("InstanceRegistry")
-event_registry: TypeRegistry[CamelCaseModel] = TypeRegistry("EventRegistry")
-task_registry: TypeRegistry[CamelCaseModel] = TypeRegistry("TaskRegistry")

src/exo/routing/router.py

@@ -216,6 +216,8 @@ def get_node_id_keypair(
     Obtains the :class:`Keypair` associated with this node-ID.
     Obtain the :class:`PeerId` by from it.
     """
+    # TODO(evan): bring back node id persistence once we figure out how to deal with duplicates
+    return Keypair.generate_ed25519()
 
     def lock_path(path: str | bytes | PathLike[str] | PathLike[bytes]) -> Path:
         return Path(str(path) + ".lock")

src/exo/routing/topics.py

@@ -3,7 +3,7 @@ from enum import Enum
 
 from exo.routing.connection_message import ConnectionMessage
 from exo.shared.election import ElectionMessage
-from exo.shared.types.commands import ForwarderCommand
+from exo.shared.types.commands import ForwarderCommand, ForwarderDownloadCommand
 from exo.shared.types.events import (
     ForwarderEvent,
 )
@@ -30,7 +30,7 @@ class TypedTopic[T: CamelCaseModel]:
 
     @staticmethod
     def serialize(t: T) -> bytes:
-        return t.model_dump_json(by_alias=True, serialize_as_any=True).encode("utf-8")
+        return t.model_dump_json().encode("utf-8")
 
     def deserialize(self, b: bytes) -> T:
         return self.model_type.model_validate_json(b.decode("utf-8"))
@@ -45,3 +45,6 @@ ELECTION_MESSAGES = TypedTopic(
 CONNECTION_MESSAGES = TypedTopic(
     "connection_messages", PublishPolicy.Never, ConnectionMessage
 )
+DOWNLOAD_COMMANDS = TypedTopic(
+    "download_commands", PublishPolicy.Always, ForwarderDownloadCommand
+)

src/exo/rsh/__init__.py

@@ -1,13 +0,0 @@
-"""Exo RSH - Remote Shell for MPI without SSH.
-
-This module provides a remote execution mechanism that allows mpirun to spawn
-processes on remote nodes without requiring SSH setup. It works by:
-
-1. Each Exo node runs an API server on port 52415 with an /execute endpoint
-2. The exo-rsh script acts as a drop-in replacement for ssh
-3. When mpirun calls "exo-rsh hostname command", it HTTP POSTs to the target's /execute
-4. The target executes the command and returns output
-
-Usage:
-    mpirun --mca plm_rsh_agent exo-rsh -np 4 --hostfile hosts.txt ./program
-"""

src/exo/rsh/client.py

@@ -1,101 +0,0 @@
-#!/usr/bin/env python3
-"""exo-rsh - Remote shell client for MPI.
-
-This script is called by mpirun as a replacement for ssh.
-Usage: exo-rsh [ssh-options...] hostname command [args...]
-
-It connects to the target node's Exo API (port 52415) and executes the command.
-"""
-
-import json
-import socket
-import sys
-from typing import Any, cast
-from urllib.error import URLError
-from urllib.request import Request, urlopen
-
-# Use the same port as Exo's API server
-EXO_API_PORT = 52415
-
-
-def resolve_hostname(hostname: str) -> str:
-    """Resolve hostname to IP address."""
-    try:
-        return socket.gethostbyname(hostname)
-    except socket.gaierror:
-        # If resolution fails, try using the hostname directly
-        return hostname
-
-
-def main():
-    # Parse arguments - mpirun calls us like: exo-rsh [options] hostname command [args...]
-    # SSH options we might see: -x (disable X11), -o options, etc.
-    args = sys.argv[1:]
-
-    # Skip SSH-style options
-    hostname = None
-    command_start = 0
-
-    i = 0
-    while i < len(args):
-        arg = args[i]
-        if arg.startswith("-"):
-            # Skip option and its value if needed
-            if arg in ("-o", "-i", "-l", "-p", "-F"):
-                i += 2  # Skip option and its argument
-                continue
-            i += 1
-            continue
-        else:
-            # First non-option is the hostname
-            hostname = arg
-            command_start = i + 1
-            break
-        i += 1
-
-    if hostname is None or command_start >= len(args):
-        print("Usage: exo-rsh [options] hostname command [args...]", file=sys.stderr)
-        sys.exit(1)
-
-    command = args[command_start:]
-
-    # Resolve hostname to IP
-    ip = resolve_hostname(hostname)
-
-    # Make request to Exo API
-    url = f"http://{ip}:{EXO_API_PORT}/execute"
-    data = json.dumps({"command": command}).encode("utf-8")
-
-    try:
-        req = Request(url, data=data, headers={"Content-Type": "application/json"})
-        with urlopen(req, timeout=300) as response:  # pyright: ignore[reportAny]
-            response_body: bytes = cast(bytes, response.read())  # pyright: ignore[reportAny]
-            result: dict[str, Any] = json.loads(response_body.decode("utf-8"))  # pyright: ignore[reportAny]
-
-        # Output stdout/stderr
-        stdout: str = cast(str, result.get("stdout", ""))
-        stderr: str = cast(str, result.get("stderr", ""))
-        exit_code: int = cast(int, result.get("exit_code", 0))
-
-        if stdout:
-            sys.stdout.write(stdout)
-            sys.stdout.flush()
-        if stderr:
-            sys.stderr.write(stderr)
-            sys.stderr.flush()
-
-        sys.exit(exit_code)
-
-    except URLError as e:
-        print(
-            f"exo-rsh: Failed to connect to {hostname}:{EXO_API_PORT}: {e}",
-            file=sys.stderr,
-        )
-        sys.exit(255)
-    except Exception as e:
-        print(f"exo-rsh: Error: {e}", file=sys.stderr)
-        sys.exit(1)
-
-
-if __name__ == "__main__":
-    main()

src/exo/shared/apply.py

@@ -6,8 +6,8 @@ from loguru import logger
 
 from exo.shared.types.common import NodeId
 from exo.shared.types.events import (
-    BaseEvent,
     ChunkGenerated,
+    Event,
     IndexedEvent,
     InputChunkReceived,
     InstanceCreated,
@@ -25,17 +25,20 @@ from exo.shared.types.events import (
     TestEvent,
     TopologyEdgeCreated,
     TopologyEdgeDeleted,
+    TracesCollected,
+    TracesMerged,
 )
 from exo.shared.types.profiling import (
     NodeIdentity,
     NodeNetworkInfo,
     NodeThunderboltInfo,
+    ThunderboltBridgeStatus,
 )
 from exo.shared.types.state import State
-from exo.shared.types.tasks import BaseTask, TaskId, TaskStatus
+from exo.shared.types.tasks import Task, TaskId, TaskStatus
 from exo.shared.types.topology import Connection, RDMAConnection
 from exo.shared.types.worker.downloads import DownloadProgress
-from exo.shared.types.worker.instances import BaseInstance, InstanceId
+from exo.shared.types.worker.instances import Instance, InstanceId
 from exo.shared.types.worker.runners import RunnerId, RunnerStatus
 from exo.utils.info_gatherer.info_gatherer import (
     MacmonMetrics,
@@ -46,14 +49,20 @@ from exo.utils.info_gatherer.info_gatherer import (
     NodeConfig,
     NodeNetworkInterfaces,
     StaticNodeInformation,
+    ThunderboltBridgeInfo,
 )
 
 
-def event_apply(event: BaseEvent, state: State) -> State:
+def event_apply(event: Event, state: State) -> State:
     """Apply an event to state."""
     match event:
         case (
-            TestEvent() | ChunkGenerated() | TaskAcknowledged() | InputChunkReceived()
+            TestEvent()
+            | ChunkGenerated()
+            | TaskAcknowledged()
+            | InputChunkReceived()
+            | TracesCollected()
+            | TracesMerged()
         ):  # Pass-through events that don't modify state
             return state
         case InstanceCreated():
@@ -82,10 +91,6 @@ def event_apply(event: BaseEvent, state: State) -> State:
             return apply_topology_edge_created(event, state)
         case TopologyEdgeDeleted():
             return apply_topology_edge_deleted(event, state)
-        case _:
-            # Unknown event types from plugins are ignored
-            logger.debug(f"Ignoring unknown event type: {type(event).__name__}")
-            return state
 
 
 def apply(state: State, event: IndexedEvent) -> State:
@@ -126,12 +131,12 @@ def apply_node_download_progress(event: NodeDownloadProgress, state: State) -> S
 
 
 def apply_task_created(event: TaskCreated, state: State) -> State:
-    new_tasks: Mapping[TaskId, BaseTask] = {**state.tasks, event.task_id: event.task}
+    new_tasks: Mapping[TaskId, Task] = {**state.tasks, event.task_id: event.task}
     return state.model_copy(update={"tasks": new_tasks})
 
 
 def apply_task_deleted(event: TaskDeleted, state: State) -> State:
-    new_tasks: Mapping[TaskId, BaseTask] = {
+    new_tasks: Mapping[TaskId, Task] = {
         tid: task for tid, task in state.tasks.items() if tid != event.task_id
     }
     return state.model_copy(update={"tasks": new_tasks})
@@ -150,7 +155,7 @@ def apply_task_status_updated(event: TaskStatusUpdated, state: State) -> State:
         update["error_message"] = None
 
     updated_task = state.tasks[event.task_id].model_copy(update=update)
-    new_tasks: Mapping[TaskId, BaseTask] = {**state.tasks, event.task_id: updated_task}
+    new_tasks: Mapping[TaskId, Task] = {**state.tasks, event.task_id: updated_task}
     return state.model_copy(update={"tasks": new_tasks})
 
 
@@ -162,13 +167,13 @@ def apply_task_failed(event: TaskFailed, state: State) -> State:
     updated_task = state.tasks[event.task_id].model_copy(
         update={"error_type": event.error_type, "error_message": event.error_message}
     )
-    new_tasks: Mapping[TaskId, BaseTask] = {**state.tasks, event.task_id: updated_task}
+    new_tasks: Mapping[TaskId, Task] = {**state.tasks, event.task_id: updated_task}
     return state.model_copy(update={"tasks": new_tasks})
 
 
 def apply_instance_created(event: InstanceCreated, state: State) -> State:
     instance = event.instance
-    new_instances: Mapping[InstanceId, BaseInstance] = {
+    new_instances: Mapping[InstanceId, Instance] = {
         **state.instances,
         instance.instance_id: instance,
     }
@@ -176,7 +181,7 @@ def apply_instance_created(event: InstanceCreated, state: State) -> State:
 
 
 def apply_instance_deleted(event: InstanceDeleted, state: State) -> State:
-    new_instances: Mapping[InstanceId, BaseInstance] = {
+    new_instances: Mapping[InstanceId, Instance] = {
         iid: inst for iid, inst in state.instances.items() if iid != event.instance_id
     }
     return state.model_copy(update={"instances": new_instances})
@@ -229,6 +234,21 @@ def apply_node_timed_out(event: NodeTimedOut, state: State) -> State:
         for key, value in state.node_thunderbolt.items()
         if key != event.node_id
     }
+    node_thunderbolt_bridge = {
+        key: value
+        for key, value in state.node_thunderbolt_bridge.items()
+        if key != event.node_id
+    }
+    # Only recompute cycles if the leaving node had TB bridge enabled
+    leaving_node_status = state.node_thunderbolt_bridge.get(event.node_id)
+    leaving_node_had_tb_enabled = (
+        leaving_node_status is not None and leaving_node_status.enabled
+    )
+    thunderbolt_bridge_cycles = (
+        topology.get_thunderbolt_bridge_cycles(node_thunderbolt_bridge, node_network)
+        if leaving_node_had_tb_enabled
+        else [list(cycle) for cycle in state.thunderbolt_bridge_cycles]
+    )
     return state.model_copy(
         update={
             "downloads": downloads,
@@ -239,6 +259,8 @@ def apply_node_timed_out(event: NodeTimedOut, state: State) -> State:
             "node_system": node_system,
             "node_network": node_network,
             "node_thunderbolt": node_thunderbolt,
+            "node_thunderbolt_bridge": node_thunderbolt_bridge,
+            "thunderbolt_bridge_cycles": thunderbolt_bridge_cycles,
         }
     )
 
@@ -316,6 +338,22 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:
                 if tb_conn.sink_uuid in conn_map
             ]
             topology.replace_all_out_rdma_connections(event.node_id, as_rdma_conns)
+        case ThunderboltBridgeInfo():
+            new_tb_bridge: dict[NodeId, ThunderboltBridgeStatus] = {
+                **state.node_thunderbolt_bridge,
+                event.node_id: info.status,
+            }
+            update["node_thunderbolt_bridge"] = new_tb_bridge
+            # Only recompute cycles if the enabled status changed
+            old_status = state.node_thunderbolt_bridge.get(event.node_id)
+            old_enabled = old_status.enabled if old_status else False
+            new_enabled = info.status.enabled
+            if old_enabled != new_enabled:
+                update["thunderbolt_bridge_cycles"] = (
+                    topology.get_thunderbolt_bridge_cycles(
+                        new_tb_bridge, state.node_network
+                    )
+                )
 
     return state.model_copy(update=update)
 

src/exo/shared/constants.py

@@ -49,3 +49,10 @@ LIBP2P_COMMANDS_TOPIC = "commands"
 EXO_MAX_CHUNK_SIZE = 512 * 1024
 
 EXO_IMAGE_CACHE_DIR = EXO_CACHE_HOME / "images"
+EXO_TRACING_CACHE_DIR = EXO_CACHE_HOME / "traces"
+
+EXO_ENABLE_IMAGE_MODELS = (
+    os.getenv("EXO_ENABLE_IMAGE_MODELS", "false").lower() == "true"
+)
+
+EXO_TRACING_ENABLED = os.getenv("EXO_TRACING_ENABLED", "false").lower() == "true"

src/exo/shared/models/model_cards.py

@@ -1,5 +1,5 @@
 from enum import Enum
-from typing import Annotated
+from typing import Annotated, Any
 
 import aiofiles
 import aiofiles.os as aios
@@ -7,8 +7,16 @@ import tomlkit
 from anyio import Path, open_file
 from huggingface_hub import model_info
 from loguru import logger
-from pydantic import BaseModel, Field, PositiveInt, field_validator
+from pydantic import (
+    AliasChoices,
+    BaseModel,
+    Field,
+    PositiveInt,
+    field_validator,
+    model_validator,
+)
 
+from exo.shared.constants import EXO_ENABLE_IMAGE_MODELS
 from exo.shared.types.common import ModelId
 from exo.shared.types.memory import Memory
 from exo.utils.pydantic_ext import CamelCaseModel
@@ -120,6 +128,14 @@ MODEL_CARDS: dict[str, ModelCard] = {
         supports_tensor=True,
         tasks=[ModelTask.TextGeneration],
     ),
+    "kimi-k2.5": ModelCard(
+        model_id=ModelId("mlx-community/Kimi-K2.5"),
+        storage_size=Memory.from_gb(617),
+        n_layers=61,
+        hidden_size=7168,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
     # llama-3.1
     "llama-3.1-8b": ModelCard(
         model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"),
@@ -254,7 +270,7 @@ MODEL_CARDS: dict[str, ModelCard] = {
     ),
     "qwen3-80b-a3B-thinking-4bit": ModelCard(
         model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"),
-        storage_size=Memory.from_mb(84700),
+        storage_size=Memory.from_mb(44900),
         n_layers=48,
         hidden_size=2048,
         supports_tensor=True,
@@ -410,174 +426,310 @@ MODEL_CARDS: dict[str, ModelCard] = {
         supports_tensor=True,
         tasks=[ModelTask.TextGeneration],
     ),
-    # Image models commented out - feature not stable (see https://github.com/exo-explore/exo/issues/1242)
-    # "flux1-schnell": ModelCard(
-    #     model_id=ModelId("black-forest-labs/FLUX.1-schnell"),
-    #     storage_size=Memory.from_bytes(23782357120 + 9524621312),
-    #     n_layers=57,
-    #     hidden_size=1,
-    #     supports_tensor=False,
-    #     tasks=[ModelTask.TextToImage],
-    #     components=[
-    #         ComponentInfo(
-    #             component_name="text_encoder",
-    #             component_path="text_encoder/",
-    #             storage_size=Memory.from_kb(0),
-    #             n_layers=12,
-    #             can_shard=False,
-    #             safetensors_index_filename=None,  # Single file
-    #         ),
-    #         ComponentInfo(
-    #             component_name="text_encoder_2",
-    #             component_path="text_encoder_2/",
-    #             storage_size=Memory.from_bytes(9524621312),
-    #             n_layers=24,
-    #             can_shard=False,
-    #             safetensors_index_filename="model.safetensors.index.json",
-    #         ),
-    #         ComponentInfo(
-    #             component_name="transformer",
-    #             component_path="transformer/",
-    #             storage_size=Memory.from_bytes(23782357120),
-    #             n_layers=57,  # 19 transformer_blocks + 38 single_transformer_blocks
-    #             can_shard=True,
-    #             safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
-    #         ),
-    #         ComponentInfo(
-    #             component_name="vae",
-    #             component_path="vae/",
-    #             storage_size=Memory.from_kb(0),
-    #             n_layers=None,
-    #             can_shard=False,
-    #             safetensors_index_filename=None,
-    #         ),
-    #     ],
-    # ),
-    # "flux1-dev": ModelCard(
-    #     model_id=ModelId("black-forest-labs/FLUX.1-dev"),
-    #     storage_size=Memory.from_bytes(23782357120 + 9524621312),
-    #     n_layers=57,
-    #     hidden_size=1,
-    #     supports_tensor=False,
-    #     tasks=[ModelTask.TextToImage, ModelTask.ImageToImage],
-    #     components=[
-    #         ComponentInfo(
-    #             component_name="text_encoder",
-    #             component_path="text_encoder/",
-    #             storage_size=Memory.from_kb(0),
-    #             n_layers=12,
-    #             can_shard=False,
-    #             safetensors_index_filename=None,  # Single file
-    #         ),
-    #         ComponentInfo(
-    #             component_name="text_encoder_2",
-    #             component_path="text_encoder_2/",
-    #             storage_size=Memory.from_bytes(9524621312),
-    #             n_layers=24,
-    #             can_shard=False,
-    #             safetensors_index_filename="model.safetensors.index.json",
-    #         ),
-    #         ComponentInfo(
-    #             component_name="transformer",
-    #             component_path="transformer/",
-    #             storage_size=Memory.from_bytes(23802816640),
-    #             n_layers=57,  # 19 transformer_blocks + 38 single_transformer_blocks
-    #             can_shard=True,
-    #             safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
-    #         ),
-    #         ComponentInfo(
-    #             component_name="vae",
-    #             component_path="vae/",
-    #             storage_size=Memory.from_kb(0),
-    #             n_layers=None,
-    #             can_shard=False,
-    #             safetensors_index_filename=None,
-    #         ),
-    #     ],
-    # ),
-    # "qwen-image": ModelCard(
-    #     model_id=ModelId("Qwen/Qwen-Image"),
-    #     storage_size=Memory.from_bytes(16584333312 + 40860802176),
-    #     n_layers=60,  # Qwen has 60 transformer blocks (all joint-style)
-    #     hidden_size=1,
-    #     supports_tensor=False,
-    #     tasks=[ModelTask.TextToImage, ModelTask.ImageToImage],
-    #     components=[
-    #         ComponentInfo(
-    #             component_name="text_encoder",
-    #             component_path="text_encoder/",
-    #             storage_size=Memory.from_kb(16584333312),
-    #             n_layers=12,
-    #             can_shard=False,
-    #             safetensors_index_filename=None,  # Single file
-    #         ),
-    #         ComponentInfo(
-    #             component_name="transformer",
-    #             component_path="transformer/",
-    #             storage_size=Memory.from_bytes(40860802176),
-    #             n_layers=60,
-    #             can_shard=True,
-    #             safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
-    #         ),
-    #         ComponentInfo(
-    #             component_name="vae",
-    #             component_path="vae/",
-    #             storage_size=Memory.from_kb(0),
-    #             n_layers=None,
-    #             can_shard=False,
-    #             safetensors_index_filename=None,
-    #         ),
-    #     ],
-    # ),
-    # "qwen-image-edit-2509": ModelCard(
-    #     model_id=ModelId("Qwen/Qwen-Image-Edit-2509"),
-    #     storage_size=Memory.from_bytes(16584333312 + 40860802176),
-    #     n_layers=60,  # Qwen has 60 transformer blocks (all joint-style)
-    #     hidden_size=1,
-    #     supports_tensor=False,
-    #     tasks=[ModelTask.ImageToImage],
-    #     components=[
-    #         ComponentInfo(
-    #             component_name="text_encoder",
-    #             component_path="text_encoder/",
-    #             storage_size=Memory.from_kb(16584333312),
-    #             n_layers=12,
-    #             can_shard=False,
-    #             safetensors_index_filename=None,  # Single file
-    #         ),
-    #         ComponentInfo(
-    #             component_name="transformer",
-    #             component_path="transformer/",
-    #             storage_size=Memory.from_bytes(40860802176),
-    #             n_layers=60,
-    #             can_shard=True,
-    #             safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
-    #         ),
-    #         ComponentInfo(
-    #             component_name="vae",
-    #             component_path="vae/",
-    #             storage_size=Memory.from_kb(0),
-    #             n_layers=None,
-    #             can_shard=False,
-    #             safetensors_index_filename=None,
-    #         ),
-    #     ],
-    # ),
 }
 
+_IMAGE_BASE_MODEL_CARDS: dict[str, ModelCard] = {
+    "flux1-schnell": ModelCard(
+        model_id=ModelId("exolabs/FLUX.1-schnell"),
+        storage_size=Memory.from_bytes(23782357120 + 9524621312),
+        n_layers=57,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.TextToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_kb(0),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="text_encoder_2",
+                component_path="text_encoder_2/",
+                storage_size=Memory.from_bytes(9524621312),
+                n_layers=24,
+                can_shard=False,
+                safetensors_index_filename="model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(23782357120),
+                n_layers=57,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+    "flux1-dev": ModelCard(
+        model_id=ModelId("exolabs/FLUX.1-dev"),
+        storage_size=Memory.from_bytes(23782357120 + 9524621312),
+        n_layers=57,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.TextToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_kb(0),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="text_encoder_2",
+                component_path="text_encoder_2/",
+                storage_size=Memory.from_bytes(9524621312),
+                n_layers=24,
+                can_shard=False,
+                safetensors_index_filename="model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(23802816640),
+                n_layers=57,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+    "flux1-krea-dev": ModelCard(
+        model_id=ModelId("exolabs/FLUX.1-Krea-dev"),
+        storage_size=Memory.from_bytes(23802816640 + 9524621312),  # Same as dev
+        n_layers=57,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.TextToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_kb(0),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="text_encoder_2",
+                component_path="text_encoder_2/",
+                storage_size=Memory.from_bytes(9524621312),
+                n_layers=24,
+                can_shard=False,
+                safetensors_index_filename="model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(23802816640),
+                n_layers=57,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+    "qwen-image": ModelCard(
+        model_id=ModelId("exolabs/Qwen-Image"),
+        storage_size=Memory.from_bytes(16584333312 + 40860802176),
+        n_layers=60,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.TextToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_bytes(16584333312),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(40860802176),
+                n_layers=60,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+    "qwen-image-edit-2509": ModelCard(
+        model_id=ModelId("exolabs/Qwen-Image-Edit-2509"),
+        storage_size=Memory.from_bytes(16584333312 + 40860802176),
+        n_layers=60,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.ImageToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_bytes(16584333312),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(40860802176),
+                n_layers=60,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+}
+
+
+def _generate_image_model_quant_variants(
+    base_name: str,
+    base_card: ModelCard,
+) -> dict[str, ModelCard]:
+    """Create quantized variants of an image model card.
+
+    Only the transformer component is quantized; text encoders stay at bf16.
+    Sizes are calculated exactly from the base card's component sizes.
+    """
+    if base_card.components is None:
+        raise ValueError(f"Image model {base_name} must have components defined")
+
+    # quantizations = [8, 6, 5, 4, 3]
+    quantizations = [8, 4]
+
+    num_transformer_bytes = next(
+        c.storage_size.in_bytes
+        for c in base_card.components
+        if c.component_name == "transformer"
+    )
+
+    transformer_bytes = Memory.from_bytes(num_transformer_bytes)
+
+    remaining_bytes = Memory.from_bytes(
+        sum(
+            c.storage_size.in_bytes
+            for c in base_card.components
+            if c.component_name != "transformer"
+        )
+    )
+
+    def with_transformer_size(new_size: Memory) -> list[ComponentInfo]:
+        assert base_card.components is not None
+        return [
+            ComponentInfo(
+                component_name=c.component_name,
+                component_path=c.component_path,
+                storage_size=new_size
+                if c.component_name == "transformer"
+                else c.storage_size,
+                n_layers=c.n_layers,
+                can_shard=c.can_shard,
+                safetensors_index_filename=c.safetensors_index_filename,
+            )
+            for c in base_card.components
+        ]
+
+    variants = {
+        base_name: ModelCard(
+            model_id=base_card.model_id,
+            storage_size=transformer_bytes + remaining_bytes,
+            n_layers=base_card.n_layers,
+            hidden_size=base_card.hidden_size,
+            supports_tensor=base_card.supports_tensor,
+            tasks=base_card.tasks,
+            components=with_transformer_size(transformer_bytes),
+        )
+    }
+
+    for quant in quantizations:
+        quant_transformer_bytes = Memory.from_bytes(
+            (num_transformer_bytes * quant) // 16
+        )
+        total_bytes = remaining_bytes + quant_transformer_bytes
+
+        model_id = ModelId(base_card.model_id + f"-{quant}bit")
+
+        variants[f"{base_name}-{quant}bit"] = ModelCard(
+            model_id=model_id,
+            storage_size=total_bytes,
+            n_layers=base_card.n_layers,
+            hidden_size=base_card.hidden_size,
+            supports_tensor=base_card.supports_tensor,
+            tasks=base_card.tasks,
+            components=with_transformer_size(quant_transformer_bytes),
+        )
+
+    return variants
+
+
+_image_model_cards: dict[str, ModelCard] = {}
+for _base_name, _base_card in _IMAGE_BASE_MODEL_CARDS.items():
+    _image_model_cards |= _generate_image_model_quant_variants(_base_name, _base_card)
+_IMAGE_MODEL_CARDS = _image_model_cards
+
+if EXO_ENABLE_IMAGE_MODELS:
+    MODEL_CARDS.update(_IMAGE_MODEL_CARDS)
+
 
 class ConfigData(BaseModel):
     model_config = {"extra": "ignore"}  # Allow unknown fields
 
-    # Common field names for number of layers across different architectures
-    num_hidden_layers: Annotated[int, Field(ge=0)] | None = None
-    num_layers: Annotated[int, Field(ge=0)] | None = None
-    n_layer: Annotated[int, Field(ge=0)] | None = None
-    n_layers: Annotated[int, Field(ge=0)] | None = None  # Sometimes used
-    num_decoder_layers: Annotated[int, Field(ge=0)] | None = None  # Transformer models
-    decoder_layers: Annotated[int, Field(ge=0)] | None = None  # Some architectures
-    hidden_size: Annotated[int, Field(ge=0)] | None = None
     architectures: list[str] | None = None
+    hidden_size: Annotated[int, Field(ge=0)] | None = None
+    layer_count: int = Field(
+        validation_alias=AliasChoices(
+            "num_hidden_layers",
+            "num_layers",
+            "n_layer",
+            "n_layers",
+            "num_decoder_layers",
+            "decoder_layers",
+        )
+    )
 
     @property
     def supports_tensor(self) -> bool:
@@ -592,30 +744,32 @@ class ConfigData(BaseModel):
             ["GptOssForCausalLM"],
         ]
 
-    @property
-    def layer_count(self) -> int:
-        # Check common field names for layer count
-        layer_fields = [
-            self.num_hidden_layers,
-            self.num_layers,
-            self.n_layer,
-            self.n_layers,
-            self.num_decoder_layers,
-            self.decoder_layers,
-        ]
+    @model_validator(mode="before")
+    @classmethod
+    def defer_to_text_config(cls, data: dict[str, Any]):
+        text_config = data.get("text_config")
+        if text_config is None:
+            return data
 
-        for layer_count in layer_fields:
-            if layer_count is not None:
-                return layer_count
+        for field in [
+            "architectures",
+            "hidden_size",
+            "num_hidden_layers",
+            "num_layers",
+            "n_layer",
+            "n_layers",
+            "num_decoder_layers",
+            "decoder_layers",
+        ]:
+            if (val := text_config.get(field)) is not None:  # pyright: ignore[reportAny]
+                data[field] = val
 
-        raise ValueError(
-            f"No layer count found in config.json: {self.model_dump_json()}"
-        )
+        return data
 
 
 async def get_config_data(model_id: ModelId) -> ConfigData:
     """Downloads and parses config.json for a model."""
-    from exo.worker.download.download_utils import (
+    from exo.download.download_utils import (
         download_file_with_retry,
         ensure_models_dir,
     )
@@ -627,7 +781,7 @@ async def get_config_data(model_id: ModelId) -> ConfigData:
         "main",
         "config.json",
         target_dir,
-        lambda curr_bytes, total_bytes, is_renamed: logger.info(
+        lambda curr_bytes, total_bytes, is_renamed: logger.debug(
             f"Downloading config.json for {model_id}: {curr_bytes}/{total_bytes} ({is_renamed=})"
         ),
     )
@@ -637,11 +791,11 @@ async def get_config_data(model_id: ModelId) -> ConfigData:
 
 async def get_safetensors_size(model_id: ModelId) -> Memory:
     """Gets model size from safetensors index or falls back to HF API."""
-    from exo.shared.types.worker.downloads import ModelSafetensorsIndex
-    from exo.worker.download.download_utils import (
+    from exo.download.download_utils import (
         download_file_with_retry,
         ensure_models_dir,
     )
+    from exo.shared.types.worker.downloads import ModelSafetensorsIndex
 
     target_dir = (await ensure_models_dir()) / model_id.normalize()
     await aios.makedirs(target_dir, exist_ok=True)
@@ -650,7 +804,7 @@ async def get_safetensors_size(model_id: ModelId) -> Memory:
         "main",
         "model.safetensors.index.json",
         target_dir,
-        lambda curr_bytes, total_bytes, is_renamed: logger.info(
+        lambda curr_bytes, total_bytes, is_renamed: logger.debug(
             f"Downloading model.safetensors.index.json for {model_id}: {curr_bytes}/{total_bytes} ({is_renamed=})"
         ),
     )

src/exo/shared/tests/test_node_id_persistence.py

@@ -8,7 +8,7 @@ from multiprocessing.synchronize import Event as EventT
 from multiprocessing.synchronize import Semaphore as SemaphoreT
 
 from loguru import logger
-from pytest import LogCaptureFixture
+from pytest import LogCaptureFixture, mark
 
 from exo.routing.router import get_node_id_keypair
 from exo.shared.constants import EXO_NODE_ID_KEYPAIR
@@ -74,6 +74,7 @@ def _delete_if_exists(p: str | bytes | os.PathLike[str] | os.PathLike[bytes]):
         os.remove(p)
 
 
+@mark.skip(reason="this functionality is currently disabled but may return in future")
 def test_node_id_fetching(caplog: LogCaptureFixture):
     reps = 10
 

src/exo/shared/topology.py

@@ -7,6 +7,11 @@ import rustworkx as rx
 from pydantic import BaseModel, ConfigDict
 
 from exo.shared.types.common import NodeId
+from exo.shared.types.profiling import (
+    InterfaceType,
+    NodeNetworkInfo,
+    ThunderboltBridgeStatus,
+)
 from exo.shared.types.topology import (
     Connection,
     Cycle,
@@ -188,24 +193,25 @@ class Topology:
             cycles.append(Cycle(node_ids=[node_id]))
         return cycles
 
-    def get_cycles_tb(self) -> list[Cycle]:
-        tb_edges = [
+    def get_rdma_cycles(self) -> list[Cycle]:
+        rdma_edges = [
             (u, v, conn)
             for u, v, conn in self._graph.weighted_edge_list()
-            if conn.is_thunderbolt()
+            if isinstance(conn, RDMAConnection)
         ]
 
-        tb_graph: rx.PyDiGraph[NodeId, SocketConnection] = rx.PyDiGraph()
-        tb_graph.add_nodes_from(self._graph.nodes())
+        rdma_graph: rx.PyDiGraph[NodeId, SocketConnection | RDMAConnection] = (
+            rx.PyDiGraph()
+        )
+        rdma_graph.add_nodes_from(self._graph.nodes())
 
-        for u, v, conn in tb_edges:
-            if isinstance(conn, SocketConnection):
-                tb_graph.add_edge(u, v, conn)
+        for u, v, conn in rdma_edges:
+            rdma_graph.add_edge(u, v, conn)
 
-        cycle_idxs = rx.simple_cycles(tb_graph)
+        cycle_idxs = rx.simple_cycles(rdma_graph)
         cycles: list[Cycle] = []
         for cycle_idx in cycle_idxs:
-            cycle = Cycle(node_ids=[tb_graph[idx] for idx in cycle_idx])
+            cycle = Cycle(node_ids=[rdma_graph[idx] for idx in cycle_idx])
             cycles.append(cycle)
 
         return cycles
@@ -219,18 +225,83 @@ class Topology:
                 topology.add_connection(connection)
         return topology
 
-    def is_thunderbolt_cycle(self, cycle: Cycle) -> bool:
+    def is_rdma_cycle(self, cycle: Cycle) -> bool:
         node_idxs = [node for node in cycle]
         rx_idxs = [self._vertex_indices[idx] for idx in node_idxs]
         for rid in rx_idxs:
             for neighbor_rid in self._graph.neighbors(rid):
                 if neighbor_rid not in rx_idxs:
                     continue
-                has_tb = False
+                has_rdma = False
                 for edge in self._graph.get_all_edge_data(rid, neighbor_rid):
-                    if edge.is_thunderbolt():
-                        has_tb = True
+                    if isinstance(edge, RDMAConnection):
+                        has_rdma = True
                         break
-                if not has_tb:
+                if not has_rdma:
                     return False
         return True
+
+    def get_thunderbolt_bridge_cycles(
+        self,
+        node_tb_bridge_status: Mapping[NodeId, ThunderboltBridgeStatus],
+        node_network: Mapping[NodeId, NodeNetworkInfo],
+    ) -> list[list[NodeId]]:
+        """
+        Find cycles in the Thunderbolt topology where all nodes have TB bridge enabled.
+        Only returns cycles with >=2 nodes (2+ machines in a loop), as
+        1 node doesn't cause the broadcast storm problem.
+        """
+        enabled_nodes = {
+            node_id
+            for node_id, status in node_tb_bridge_status.items()
+            if status.enabled
+        }
+
+        if len(enabled_nodes) < 2:
+            return []
+
+        thunderbolt_ips = _get_ips_with_interface_type(
+            enabled_nodes, node_network, "thunderbolt"
+        )
+
+        # Build subgraph with only TB bridge enabled nodes and thunderbolt connections
+        graph: rx.PyDiGraph[NodeId, SocketConnection | RDMAConnection] = rx.PyDiGraph()
+        node_to_idx: dict[NodeId, int] = {}
+
+        for node_id in enabled_nodes:
+            if node_id in self._vertex_indices:
+                node_to_idx[node_id] = graph.add_node(node_id)
+
+        for u, v, conn in self._graph.weighted_edge_list():
+            source_id, sink_id = self._graph[u], self._graph[v]
+            if source_id not in node_to_idx or sink_id not in node_to_idx:
+                continue
+            # Include connection if it's over a thunderbolt interface
+            if (
+                isinstance(conn, SocketConnection)
+                and conn.sink_multiaddr.ip_address in thunderbolt_ips
+            ):
+                graph.add_edge(node_to_idx[source_id], node_to_idx[sink_id], conn)
+            if isinstance(conn, RDMAConnection):
+                graph.add_edge(node_to_idx[source_id], node_to_idx[sink_id], conn)
+
+        return [
+            [graph[idx] for idx in cycle]
+            for cycle in rx.simple_cycles(graph)
+            if len(cycle) >= 2
+        ]
+
+
+def _get_ips_with_interface_type(
+    node_ids: set[NodeId],
+    node_network: Mapping[NodeId, NodeNetworkInfo],
+    interface_type: InterfaceType,
+) -> set[str]:
+    """Get all IP addresses on interfaces of the specified type for the given nodes."""
+    ips: set[str] = set()
+    for node_id in node_ids:
+        network_info = node_network.get(node_id, NodeNetworkInfo())
+        for iface in network_info.interfaces:
+            if iface.interface_type == interface_type:
+                ips.add(iface.ip_address)
+    return ips

src/exo/shared/tracing.py

@@ -0,0 +1,238 @@
+from __future__ import annotations
+
+import json
+import time
+from collections import defaultdict
+from collections.abc import Generator
+from contextlib import contextmanager
+from contextvars import ContextVar
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import cast, final
+
+from exo.shared.constants import EXO_TRACING_ENABLED
+from exo.worker.runner.bootstrap import logger
+
+# Context variable to track the current trace category for hierarchical nesting
+_current_category: ContextVar[str | None] = ContextVar("current_category", default=None)
+
+
+@final
+@dataclass(frozen=True)
+class TraceEvent:
+    name: str
+    start_us: int
+    duration_us: int
+    rank: int
+    category: str
+
+
+@final
+@dataclass
+class CategoryStats:
+    total_us: int = 0
+    count: int = 0
+    min_us: int = 0
+    max_us: int = 0
+
+    def add(self, duration_us: int) -> None:
+        if self.count == 0:
+            self.min_us = duration_us
+            self.max_us = duration_us
+        else:
+            self.min_us = min(self.min_us, duration_us)
+            self.max_us = max(self.max_us, duration_us)
+        self.total_us += duration_us
+        self.count += 1
+
+    @property
+    def avg_us(self) -> float:
+        return self.total_us / self.count if self.count > 0 else 0.0
+
+
+@final
+@dataclass
+class TraceStats:
+    total_wall_time_us: int = 0
+    by_category: dict[str, CategoryStats] = field(default_factory=dict)
+    by_rank: dict[int, dict[str, CategoryStats]] = field(default_factory=dict)
+
+
+# Global trace buffer - each rank accumulates traces here
+_trace_buffer: list[TraceEvent] = []
+
+
+def _record_span(
+    name: str, start_us: int, duration_us: int, rank: int, category: str
+) -> None:
+    _trace_buffer.append(
+        TraceEvent(
+            name=name,
+            start_us=start_us,
+            duration_us=duration_us,
+            rank=rank,
+            category=category,
+        )
+    )
+
+
+@contextmanager
+def trace(
+    name: str,
+    rank: int,
+    category: str = "compute",
+) -> Generator[None, None, None]:
+    """Context manager to trace any operation.
+
+    Nested traces automatically inherit the parent category, creating hierarchical
+    categories like "sync/compute" or "async/comms".
+
+    Args:
+        name: Name of the operation (e.g., "recv 0", "send 1", "joint_blocks")
+        rank: This rank's ID
+        category: Category for grouping in trace viewer ("comm", "compute", "step")
+
+    Example:
+        with trace(f"sync {t}", rank, "sync"):
+            with trace("joint_blocks", rank, "compute"):
+                # Recorded with category "sync/compute"
+                hidden_states = some_computation(...)
+    """
+    if not EXO_TRACING_ENABLED:
+        yield
+        return
+
+    # Combine with parent category if nested
+    parent = _current_category.get()
+    full_category = f"{parent}/{category}" if parent else category
+
+    # Set as current for nested traces
+    token = _current_category.set(full_category)
+
+    try:
+        start_us = int(time.time() * 1_000_000)
+        start_perf = time.perf_counter()
+        yield
+        duration_us = int((time.perf_counter() - start_perf) * 1_000_000)
+        _record_span(name, start_us, duration_us, rank, full_category)
+    finally:
+        _current_category.reset(token)
+
+
+def get_trace_buffer() -> list[TraceEvent]:
+    return list(_trace_buffer)
+
+
+def clear_trace_buffer() -> None:
+    _trace_buffer.clear()
+
+
+def export_trace(traces: list[TraceEvent], output_path: Path) -> None:
+    trace_events: list[dict[str, object]] = []
+
+    for event in traces:
+        # Chrome trace format uses "X" for complete events (with duration)
+        chrome_event: dict[str, object] = {
+            "name": event.name,
+            "cat": event.category,
+            "ph": "X",
+            "ts": event.start_us,
+            "dur": event.duration_us,
+            "pid": 0,
+            "tid": event.rank,
+            "args": {"rank": event.rank},
+        }
+        trace_events.append(chrome_event)
+
+    ranks_seen = set(t.rank for t in traces)
+    for rank in ranks_seen:
+        trace_events.append(
+            {
+                "name": "thread_name",
+                "ph": "M",  # Metadata event
+                "pid": 0,
+                "tid": rank,
+                "args": {"name": f"Rank {rank}"},
+            }
+        )
+
+    chrome_trace = {"traceEvents": trace_events}
+
+    try:
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+        with open(output_path, "w") as f:
+            json.dump(chrome_trace, f, indent=2)
+    except OSError as e:
+        logger.warning("Failed to export trace to %s: %s", output_path, e)
+
+
+def load_trace_file(path: Path) -> list[TraceEvent]:
+    with open(path) as f:
+        data = cast(dict[str, list[dict[str, object]]], json.load(f))
+
+    events = data.get("traceEvents", [])
+    traces: list[TraceEvent] = []
+
+    for event in events:
+        # Skip metadata events
+        if event.get("ph") == "M":
+            continue
+
+        name = str(event.get("name", ""))
+        category = str(event.get("cat", ""))
+        ts_value = event.get("ts", 0)
+        dur_value = event.get("dur", 0)
+        tid_value = event.get("tid", 0)
+        start_us = int(ts_value) if isinstance(ts_value, (int, float, str)) else 0
+        duration_us = int(dur_value) if isinstance(dur_value, (int, float, str)) else 0
+
+        # Get rank from tid or args
+        rank = int(tid_value) if isinstance(tid_value, (int, float, str)) else 0
+        args = event.get("args")
+        if isinstance(args, dict):
+            args_dict = cast(dict[str, object], args)
+            rank_from_args = args_dict.get("rank")
+            if isinstance(rank_from_args, (int, float, str)):
+                rank = int(rank_from_args)
+
+        traces.append(
+            TraceEvent(
+                name=name,
+                start_us=start_us,
+                duration_us=duration_us,
+                rank=rank,
+                category=category,
+            )
+        )
+
+    return traces
+
+
+def compute_stats(traces: list[TraceEvent]) -> TraceStats:
+    stats = TraceStats()
+
+    if not traces:
+        return stats
+
+    # Calculate wall time from earliest start to latest end
+    min_start = min(t.start_us for t in traces)
+    max_end = max(t.start_us + t.duration_us for t in traces)
+    stats.total_wall_time_us = max_end - min_start
+
+    # Initialize nested dicts
+    by_category: dict[str, CategoryStats] = defaultdict(CategoryStats)
+    by_rank: dict[int, dict[str, CategoryStats]] = defaultdict(
+        lambda: defaultdict(CategoryStats)
+    )
+
+    for event in traces:
+        # By category
+        by_category[event.category].add(event.duration_us)
+
+        # By rank and category
+        by_rank[event.rank][event.category].add(event.duration_us)
+
+    stats.by_category = dict(by_category)
+    stats.by_rank = {k: dict(v) for k, v in by_rank.items()}
+
+    return stats

src/exo/shared/types/api.py

@@ -1,22 +1,16 @@
 import time
 from collections.abc import Generator
-from typing import Annotated, Any, Literal, cast
+from typing import Annotated, Any, Literal
 
-from fastapi import UploadFile
 from pydantic import BaseModel, Field, field_validator
 from pydantic_core import PydanticUseDefault
 
-from exo.plugins.type_registry import instance_registry
 from exo.shared.models.model_cards import ModelCard, ModelId
-from exo.shared.types.common import CommandId
+from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.memory import Memory
-from exo.shared.types.worker.instances import (
-    BaseInstance,
-    Instance,
-    InstanceId,
-    InstanceMeta,
-)
-from exo.shared.types.worker.shards import Sharding
+from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
+from exo.shared.types.worker.shards import Sharding, ShardMetadata
+from exo.utils.pydantic_ext import CamelCaseModel
 
 FinishReason = Literal[
     "stop", "length", "tool_calls", "content_filter", "function_call", "error"
@@ -60,6 +54,18 @@ class ChatCompletionMessageText(BaseModel):
     text: str
 
 
+class ToolCallItem(BaseModel):
+    name: str
+    arguments: str
+
+
+class ToolCall(BaseModel):
+    id: str
+    index: int | None = None
+    type: Literal["function"] = "function"
+    function: ToolCallItem
+
+
 class ChatCompletionMessage(BaseModel):
     role: Literal["system", "user", "assistant", "developer", "tool", "function"]
     content: (
@@ -67,7 +73,7 @@ class ChatCompletionMessage(BaseModel):
     ) = None
     thinking: str | None = None  # Added for GPT-OSS harmony format support
     name: str | None = None
-    tool_calls: list[dict[str, Any]] | None = None
+    tool_calls: list[ToolCall] | None = None
     tool_call_id: str | None = None
     function_call: dict[str, Any] | None = None
 
@@ -109,8 +115,8 @@ class Usage(BaseModel):
     prompt_tokens: int
     completion_tokens: int
     total_tokens: int
-    prompt_tokens_details: PromptTokensDetails | None = None
-    completion_tokens_details: CompletionTokensDetails | None = None
+    prompt_tokens_details: PromptTokensDetails
+    completion_tokens_details: CompletionTokensDetails
 
 
 class StreamingChoiceResponse(BaseModel):
@@ -163,8 +169,12 @@ class BenchChatCompletionResponse(ChatCompletionResponse):
     generation_stats: GenerationStats | None = None
 
 
-class ChatCompletionTaskParams(BaseModel):
-    model: str
+class StreamOptions(BaseModel):
+    include_usage: bool = False
+
+
+class ChatCompletionRequest(BaseModel):
+    model: ModelId
     frequency_penalty: float | None = None
     messages: list[ChatCompletionMessage]
     logit_bias: dict[str, int] | None = None
@@ -177,15 +187,17 @@ class ChatCompletionTaskParams(BaseModel):
     seed: int | None = None
     stop: str | list[str] | None = None
     stream: bool = False
+    stream_options: StreamOptions | None = None
     temperature: float | None = None
     top_p: float | None = None
+    top_k: int | None = None
     tools: list[dict[str, Any]] | None = None
     tool_choice: str | dict[str, Any] | None = None
     parallel_tool_calls: bool | None = None
     user: str | None = None
 
 
-class BenchChatCompletionTaskParams(ChatCompletionTaskParams):
+class BenchChatCompletionRequest(ChatCompletionRequest):
     pass
 
 
@@ -206,12 +218,6 @@ class PlaceInstanceParams(BaseModel):
 class CreateInstanceParams(BaseModel):
     instance: Instance
 
-    @field_validator("instance", mode="before")
-    @classmethod
-    def validate_instance(cls, v: Any) -> BaseInstance:  # noqa: ANN401  # pyright: ignore[reportAny]
-        """Validate instance using registry to handle both tagged and flat formats."""
-        return cast(BaseInstance, instance_registry.deserialize(v))  # pyright: ignore[reportAny]
-
 
 class PlacementPreview(BaseModel):
     model_id: ModelId
@@ -275,28 +281,7 @@ class BenchImageGenerationTaskParams(ImageGenerationTaskParams):
 
 
 class ImageEditsTaskParams(BaseModel):
-    image: UploadFile
-    prompt: str
-    background: str | None = None
-    input_fidelity: float | None = None
-    mask: UploadFile | None = None
-    model: str
-    n: int | None = 1
-    output_compression: int | None = None
-    output_format: Literal["png", "jpeg", "webp"] = "png"
-    partial_images: int | None = 0
-    quality: Literal["high", "medium", "low"] | None = "medium"
-    response_format: Literal["url", "b64_json"] | None = "b64_json"
-    size: str | None = "1024x1024"
-    stream: bool | None = False
-    user: str | None = None
-    advanced_params: AdvancedImageParams | None = None
-    # Internal flag for benchmark mode - set by API, preserved through serialization
-    bench: bool = False
-
-
-class ImageEditsInternalParams(BaseModel):
-    """Serializable version of ImageEditsTaskParams for distributed task execution."""
+    """Internal task params for image-editing requests."""
 
     image_data: str = ""  # Base64-encoded image (empty when using chunked transfer)
     total_input_chunks: int = 0
@@ -352,3 +337,58 @@ class ImageListItem(BaseModel, frozen=True):
 
 class ImageListResponse(BaseModel, frozen=True):
     data: list[ImageListItem]
+
+
+class StartDownloadParams(CamelCaseModel):
+    target_node_id: NodeId
+    shard_metadata: ShardMetadata
+
+
+class StartDownloadResponse(CamelCaseModel):
+    command_id: CommandId
+
+
+class DeleteDownloadResponse(CamelCaseModel):
+    command_id: CommandId
+
+
+class TraceEventResponse(CamelCaseModel):
+    name: str
+    start_us: int
+    duration_us: int
+    rank: int
+    category: str
+
+
+class TraceResponse(CamelCaseModel):
+    task_id: str
+    traces: list[TraceEventResponse]
+
+
+class TraceCategoryStats(CamelCaseModel):
+    total_us: int
+    count: int
+    min_us: int
+    max_us: int
+    avg_us: float
+
+
+class TraceRankStats(CamelCaseModel):
+    by_category: dict[str, TraceCategoryStats]
+
+
+class TraceStatsResponse(CamelCaseModel):
+    task_id: str
+    total_wall_time_us: int
+    by_category: dict[str, TraceCategoryStats]
+    by_rank: dict[int, TraceRankStats]
+
+
+class TraceListItem(CamelCaseModel):
+    task_id: str
+    created_at: str
+    file_size: int
+
+
+class TraceListResponse(CamelCaseModel):
+    traces: list[TraceListItem]

src/exo/shared/types/chunks.py

@@ -1,31 +1,37 @@
 from collections.abc import Generator
-from enum import Enum
 from typing import Any, Literal
 
 from exo.shared.models.model_cards import ModelId
-from exo.shared.types.api import GenerationStats, ImageGenerationStats
+from exo.shared.types.api import GenerationStats, ImageGenerationStats, Usage
 from exo.utils.pydantic_ext import TaggedModel
 
 from .api import FinishReason
 from .common import CommandId
-
-
-class ChunkType(str, Enum):
-    Token = "Token"
-    Image = "Image"
+from .worker.runner_response import ToolCallItem
 
 
 class BaseChunk(TaggedModel):
-    idx: int
     model: ModelId
 
 
 class TokenChunk(BaseChunk):
     text: str
     token_id: int
-    finish_reason: FinishReason | None = None
+    usage: Usage | None
+    finish_reason: Literal["stop", "length", "content_filter"] | None = None
+    stats: GenerationStats | None = None
+
+
+class ErrorChunk(BaseChunk):
+    error_message: str
+    finish_reason: Literal["error"] = "error"
+
+
+class ToolCallChunk(BaseChunk):
+    tool_calls: list[ToolCallItem]
+    usage: Usage | None
+    finish_reason: Literal["tool_calls"] = "tool_calls"
     stats: GenerationStats | None = None
-    error_message: str | None = None
 
 
 class ImageChunk(BaseChunk):
@@ -63,4 +69,4 @@ class InputImageChunk(BaseChunk):
                 yield name, value
 
 
-GenerationChunk = TokenChunk | ImageChunk
+GenerationChunk = TokenChunk | ImageChunk | ToolCallChunk | ErrorChunk

src/exo/shared/types/claude_api.py

@@ -0,0 +1,214 @@
+"""Claude Messages API types for request/response conversion."""
+
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+from exo.shared.types.common import ModelId
+
+# Tool definition types
+ClaudeToolInputSchema = dict[str, Any]
+
+
+class ClaudeToolDefinition(BaseModel, frozen=True):
+    """Tool definition in Claude Messages API request."""
+
+    name: str
+    description: str | None = None
+    input_schema: ClaudeToolInputSchema
+
+
+# Type aliases
+ClaudeRole = Literal["user", "assistant"]
+ClaudeStopReason = Literal["end_turn", "max_tokens", "stop_sequence", "tool_use"]
+
+
+# Content block types
+class ClaudeTextBlock(BaseModel, frozen=True):
+    """Text content block in Claude Messages API."""
+
+    type: Literal["text"] = "text"
+    text: str
+
+
+class ClaudeImageSource(BaseModel, frozen=True):
+    """Image source for Claude image blocks."""
+
+    type: Literal["base64", "url"]
+    media_type: str | None = None
+    data: str | None = None
+    url: str | None = None
+
+
+class ClaudeImageBlock(BaseModel, frozen=True):
+    """Image content block in Claude Messages API."""
+
+    type: Literal["image"] = "image"
+    source: ClaudeImageSource
+
+
+class ClaudeToolUseBlock(BaseModel, frozen=True):
+    """Tool use content block in Claude Messages API."""
+
+    type: Literal["tool_use"] = "tool_use"
+    id: str
+    name: str
+    input: dict[str, Any]
+
+
+class ClaudeToolResultBlock(BaseModel, frozen=True):
+    """Tool result content block in Claude Messages API request."""
+
+    type: Literal["tool_result"] = "tool_result"
+    tool_use_id: str
+    content: str | list[ClaudeTextBlock] | None = None
+    is_error: bool | None = None
+    cache_control: dict[str, str] | None = None
+
+
+ClaudeContentBlock = ClaudeTextBlock | ClaudeImageBlock | ClaudeToolUseBlock
+
+# Input content blocks can also include tool_result (sent by user after tool_use)
+ClaudeInputContentBlock = (
+    ClaudeTextBlock | ClaudeImageBlock | ClaudeToolUseBlock | ClaudeToolResultBlock
+)
+
+
+# Request types
+class ClaudeMessage(BaseModel, frozen=True):
+    """Message in Claude Messages API request."""
+
+    role: ClaudeRole
+    content: str | list[ClaudeInputContentBlock]
+
+
+class ClaudeMessagesRequest(BaseModel):
+    """Request body for Claude Messages API."""
+
+    model: ModelId
+    max_tokens: int
+    messages: list[ClaudeMessage]
+    system: str | list[ClaudeTextBlock] | None = None
+    stop_sequences: list[str] | None = None
+    stream: bool = False
+    temperature: float | None = None
+    top_p: float | None = None
+    top_k: int | None = None
+    tools: list[ClaudeToolDefinition] | None = None
+    metadata: dict[str, str] | None = None
+
+
+# Response types
+class ClaudeUsage(BaseModel, frozen=True):
+    """Token usage in Claude Messages API response."""
+
+    input_tokens: int
+    output_tokens: int
+
+
+class ClaudeMessagesResponse(BaseModel, frozen=True):
+    """Response body for Claude Messages API."""
+
+    id: str
+    type: Literal["message"] = "message"
+    role: Literal["assistant"] = "assistant"
+    content: list[ClaudeContentBlock]
+    model: str
+    stop_reason: ClaudeStopReason | None = None
+    stop_sequence: str | None = None
+    usage: ClaudeUsage
+
+
+# Streaming event types
+class ClaudeMessageStart(BaseModel, frozen=True):
+    """Partial message in message_start event."""
+
+    id: str
+    type: Literal["message"] = "message"
+    role: Literal["assistant"] = "assistant"
+    content: list[ClaudeTextBlock] = Field(default_factory=list)
+    model: str
+    stop_reason: ClaudeStopReason | None = None
+    stop_sequence: str | None = None
+    usage: ClaudeUsage
+
+
+class ClaudeMessageStartEvent(BaseModel, frozen=True):
+    """Event sent at start of message stream."""
+
+    type: Literal["message_start"] = "message_start"
+    message: ClaudeMessageStart
+
+
+class ClaudeContentBlockStartEvent(BaseModel, frozen=True):
+    """Event sent at start of a content block."""
+
+    type: Literal["content_block_start"] = "content_block_start"
+    index: int
+    content_block: ClaudeTextBlock | ClaudeToolUseBlock
+
+
+class ClaudeTextDelta(BaseModel, frozen=True):
+    """Delta for text content block."""
+
+    type: Literal["text_delta"] = "text_delta"
+    text: str
+
+
+class ClaudeInputJsonDelta(BaseModel, frozen=True):
+    """Delta for tool use input JSON content block."""
+
+    type: Literal["input_json_delta"] = "input_json_delta"
+    partial_json: str
+
+
+class ClaudeContentBlockDeltaEvent(BaseModel, frozen=True):
+    """Event sent for content block delta."""
+
+    type: Literal["content_block_delta"] = "content_block_delta"
+    index: int
+    delta: ClaudeTextDelta | ClaudeInputJsonDelta
+
+
+class ClaudeContentBlockStopEvent(BaseModel, frozen=True):
+    """Event sent at end of a content block."""
+
+    type: Literal["content_block_stop"] = "content_block_stop"
+    index: int
+
+
+class ClaudeMessageDeltaUsage(BaseModel, frozen=True):
+    """Usage in message_delta event."""
+
+    output_tokens: int
+
+
+class ClaudeMessageDelta(BaseModel, frozen=True):
+    """Delta in message_delta event."""
+
+    stop_reason: ClaudeStopReason | None = None
+    stop_sequence: str | None = None
+
+
+class ClaudeMessageDeltaEvent(BaseModel, frozen=True):
+    """Event sent with final message delta."""
+
+    type: Literal["message_delta"] = "message_delta"
+    delta: ClaudeMessageDelta
+    usage: ClaudeMessageDeltaUsage
+
+
+class ClaudeMessageStopEvent(BaseModel, frozen=True):
+    """Event sent at end of message stream."""
+
+    type: Literal["message_stop"] = "message_stop"
+
+
+ClaudeStreamEvent = (
+    ClaudeMessageStartEvent
+    | ClaudeContentBlockStartEvent
+    | ClaudeContentBlockDeltaEvent
+    | ClaudeContentBlockStopEvent
+    | ClaudeMessageDeltaEvent
+    | ClaudeMessageStopEvent
+)

src/exo/shared/types/commands.py

@@ -1,54 +1,38 @@
-"""Command types for exo.
+from pydantic import Field
 
-Commands are registered dynamically via the command_registry, allowing plugins
-to add their own command types without modifying this file.
-"""
-
-from typing import Any, cast
-
-from pydantic import Field, field_validator
-
-from exo.plugins.type_registry import command_registry
-from exo.shared.models.model_cards import ModelCard
+from exo.shared.models.model_cards import ModelCard, ModelId
 from exo.shared.types.api import (
-    ChatCompletionTaskParams,
-    ImageEditsInternalParams,
+    ImageEditsTaskParams,
     ImageGenerationTaskParams,
 )
 from exo.shared.types.chunks import InputImageChunk
 from exo.shared.types.common import CommandId, NodeId
+from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
-from exo.shared.types.worker.shards import Sharding
+from exo.shared.types.worker.shards import Sharding, ShardMetadata
 from exo.utils.pydantic_ext import CamelCaseModel, TaggedModel
 
 
 class BaseCommand(TaggedModel):
-    """Base class for all commands."""
-
     command_id: CommandId = Field(default_factory=CommandId)
 
 
-@command_registry.register
 class TestCommand(BaseCommand):
     __test__ = False
 
 
-@command_registry.register
-class ChatCompletion(BaseCommand):
-    request_params: ChatCompletionTaskParams
+class TextGeneration(BaseCommand):
+    task_params: TextGenerationTaskParams
 
 
-@command_registry.register
 class ImageGeneration(BaseCommand):
-    request_params: ImageGenerationTaskParams
+    task_params: ImageGenerationTaskParams
 
 
-@command_registry.register
 class ImageEdits(BaseCommand):
-    request_params: ImageEditsInternalParams
+    task_params: ImageEditsTaskParams
 
 
-@command_registry.register
 class PlaceInstance(BaseCommand):
     model_card: ModelCard
     sharding: Sharding
@@ -56,38 +40,45 @@ class PlaceInstance(BaseCommand):
     min_nodes: int
 
 
-@command_registry.register
 class CreateInstance(BaseCommand):
     instance: Instance
 
 
-@command_registry.register
 class DeleteInstance(BaseCommand):
     instance_id: InstanceId
 
 
-@command_registry.register
 class TaskFinished(BaseCommand):
     finished_command_id: CommandId
 
 
-@command_registry.register
 class SendInputChunk(BaseCommand):
     """Command to send an input image chunk (converted to event by master)."""
 
     chunk: InputImageChunk
 
 
-@command_registry.register
 class RequestEventLog(BaseCommand):
     since_idx: int
 
 
-# Union type for core commands - used by ForwarderCommand for network deserialization
+class StartDownload(BaseCommand):
+    target_node_id: NodeId
+    shard_metadata: ShardMetadata
+
+
+class DeleteDownload(BaseCommand):
+    target_node_id: NodeId
+    model_id: ModelId
+
+
+DownloadCommand = StartDownload | DeleteDownload
+
+
 Command = (
     TestCommand
     | RequestEventLog
-    | ChatCompletion
+    | TextGeneration
     | ImageGeneration
     | ImageEdits
     | PlaceInstance
@@ -99,14 +90,10 @@ Command = (
 
 
 class ForwarderCommand(CamelCaseModel):
-    """Wrapper for commands that includes origin node."""
-
     origin: NodeId
-    command: BaseCommand
+    command: Command
 
-    @field_validator("command", mode="before")
-    @classmethod
-    def validate_command(cls, v: Any) -> BaseCommand:  # noqa: ANN401  # pyright: ignore[reportAny]
-        """Validate command, using registry for plugin commands not in Command union."""
-        # First try the registry (handles both core and plugin commands)
-        return cast(BaseCommand, command_registry.deserialize(v))  # pyright: ignore[reportAny]
+
+class ForwarderDownloadCommand(CamelCaseModel):
+    origin: NodeId
+    command: DownloadCommand

src/exo/shared/types/events.py

@@ -1,18 +1,17 @@
 from datetime import datetime
-from typing import Any, cast
+from typing import final
 
-from pydantic import Field, field_validator
+from pydantic import Field
 
-from exo.plugins.type_registry import event_registry, instance_registry, task_registry
 from exo.shared.topology import Connection
 from exo.shared.types.chunks import GenerationChunk, InputImageChunk
 from exo.shared.types.common import CommandId, Id, NodeId, SessionId
-from exo.shared.types.tasks import BaseTask, TaskId, TaskStatus
+from exo.shared.types.tasks import Task, TaskId, TaskStatus
 from exo.shared.types.worker.downloads import DownloadProgress
-from exo.shared.types.worker.instances import BaseInstance, InstanceId
+from exo.shared.types.worker.instances import Instance, InstanceId
 from exo.shared.types.worker.runners import RunnerId, RunnerStatus
 from exo.utils.info_gatherer.info_gatherer import GatheredInfo
-from exo.utils.pydantic_ext import CamelCaseModel, TaggedModel
+from exo.utils.pydantic_ext import CamelCaseModel, FrozenModel, TaggedModel
 
 
 class EventId(Id):
@@ -27,53 +26,36 @@ class BaseEvent(TaggedModel):
     _master_time_stamp: None | datetime = None
 
 
-@event_registry.register
 class TestEvent(BaseEvent):
     __test__ = False
 
 
-@event_registry.register
 class TaskCreated(BaseEvent):
     task_id: TaskId
-    task: BaseTask
-
-    @field_validator("task", mode="before")
-    @classmethod
-    def validate_task(cls, v: Any) -> BaseTask:  # noqa: ANN401  # pyright: ignore[reportAny]
-        return cast(BaseTask, task_registry.deserialize(v))  # pyright: ignore[reportAny]
+    task: Task
 
 
-@event_registry.register
 class TaskAcknowledged(BaseEvent):
     task_id: TaskId
 
 
-@event_registry.register
 class TaskDeleted(BaseEvent):
     task_id: TaskId
 
 
-@event_registry.register
 class TaskStatusUpdated(BaseEvent):
     task_id: TaskId
     task_status: TaskStatus
 
 
-@event_registry.register
 class TaskFailed(BaseEvent):
     task_id: TaskId
     error_type: str
     error_message: str
 
 
-@event_registry.register
 class InstanceCreated(BaseEvent):
-    instance: BaseInstance
-
-    @field_validator("instance", mode="before")
-    @classmethod
-    def validate_instance(cls, v: Any) -> BaseInstance:  # noqa: ANN401  # pyright: ignore[reportAny]
-        return cast(BaseInstance, instance_registry.deserialize(v))  # pyright: ignore[reportAny]
+    instance: Instance
 
     def __eq__(self, other: object) -> bool:
         if isinstance(other, InstanceCreated):
@@ -82,63 +64,74 @@ class InstanceCreated(BaseEvent):
         return False
 
 
-@event_registry.register
 class InstanceDeleted(BaseEvent):
     instance_id: InstanceId
 
 
-@event_registry.register
 class RunnerStatusUpdated(BaseEvent):
     runner_id: RunnerId
     runner_status: RunnerStatus
 
 
-@event_registry.register
 class RunnerDeleted(BaseEvent):
     runner_id: RunnerId
 
 
-@event_registry.register
 class NodeTimedOut(BaseEvent):
     node_id: NodeId
 
 
 # TODO: bikeshed this name
-@event_registry.register
 class NodeGatheredInfo(BaseEvent):
     node_id: NodeId
     when: str  # this is a manually cast datetime overrode by the master when the event is indexed, rather than the local time on the device
     info: GatheredInfo
 
 
-@event_registry.register
 class NodeDownloadProgress(BaseEvent):
     download_progress: DownloadProgress
 
 
-@event_registry.register
 class ChunkGenerated(BaseEvent):
     command_id: CommandId
     chunk: GenerationChunk
 
 
-@event_registry.register
 class InputChunkReceived(BaseEvent):
     command_id: CommandId
     chunk: InputImageChunk
 
 
-@event_registry.register
 class TopologyEdgeCreated(BaseEvent):
     conn: Connection
 
 
-@event_registry.register
 class TopologyEdgeDeleted(BaseEvent):
     conn: Connection
 
 
-# Union type for Pydantic validation - tries each type in order
+@final
+class TraceEventData(FrozenModel):
+    name: str
+    start_us: int
+    duration_us: int
+    rank: int
+    category: str
+
+
+@final
+class TracesCollected(BaseEvent):
+    task_id: TaskId
+    rank: int
+    traces: list[TraceEventData]
+
+
+@final
+class TracesMerged(BaseEvent):
+    task_id: TaskId
+    traces: list[TraceEventData]
+
+
 Event = (
     TestEvent
     | TaskCreated
@@ -157,6 +150,8 @@ Event = (
     | InputChunkReceived
     | TopologyEdgeCreated
     | TopologyEdgeDeleted
+    | TracesCollected
+    | TracesMerged
 )
 
 
@@ -164,12 +159,7 @@ class IndexedEvent(CamelCaseModel):
     """An event indexed by the master, with a globally unique index"""
 
     idx: int = Field(ge=0)
-    event: BaseEvent
-
-    @field_validator("event", mode="before")
-    @classmethod
-    def validate_event(cls, v: Any) -> BaseEvent:  # noqa: ANN401  # pyright: ignore[reportAny]
-        return cast(BaseEvent, event_registry.deserialize(v))  # pyright: ignore[reportAny]
+    event: Event
 
 
 class ForwarderEvent(CamelCaseModel):
@@ -178,9 +168,4 @@ class ForwarderEvent(CamelCaseModel):
     origin_idx: int = Field(ge=0)
     origin: NodeId
     session: SessionId
-    event: BaseEvent
-
-    @field_validator("event", mode="before")
-    @classmethod
-    def validate_event(cls, v: Any) -> BaseEvent:  # noqa: ANN401  # pyright: ignore[reportAny]
-        return cast(BaseEvent, event_registry.deserialize(v))  # pyright: ignore[reportAny]
+    event: Event

src/exo/shared/types/mlx.py

@@ -0,0 +1,12 @@
+"""Shared types for MLX-related functionality."""
+
+from collections.abc import Sequence
+
+from mlx_lm.models.cache import (
+    KVCache,
+    QuantizedKVCache,
+    RotatingKVCache,
+)
+
+# This list contains one cache entry per transformer layer
+KVCacheType = Sequence[KVCache | RotatingKVCache | QuantizedKVCache]

src/exo/shared/types/openai_responses.py

@@ -0,0 +1,296 @@
+"""OpenAI Responses API wire types.
+
+These types model the OpenAI Responses API request/response format.
+ResponsesRequest is the API-level wire type; for the canonical internal
+task params type used by the inference pipeline, see
+``exo.shared.types.text_generation.TextGenerationTaskParams``.
+"""
+
+import time
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+from exo.shared.types.common import ModelId
+
+# Type aliases
+ResponseStatus = Literal["completed", "failed", "in_progress", "incomplete"]
+ResponseRole = Literal["user", "assistant", "system", "developer"]
+
+
+# Request input content part types
+class ResponseInputTextPart(BaseModel, frozen=True):
+    """Text content part in a Responses API input message."""
+
+    type: Literal["input_text"] = "input_text"
+    text: str
+
+
+class ResponseOutputTextPart(BaseModel, frozen=True):
+    """Output text content part (used when replaying assistant messages in input)."""
+
+    type: Literal["output_text"] = "output_text"
+    text: str
+
+
+ResponseContentPart = ResponseInputTextPart | ResponseOutputTextPart
+
+
+# Request input item types
+class ResponseInputMessage(BaseModel, frozen=True):
+    """Input message for Responses API."""
+
+    role: ResponseRole
+    content: str | list[ResponseContentPart]
+    type: Literal["message"] = "message"
+
+
+class FunctionCallInputItem(BaseModel, frozen=True):
+    """Function call item replayed in input (from a previous assistant response)."""
+
+    type: Literal["function_call"] = "function_call"
+    id: str | None = None
+    call_id: str
+    name: str
+    arguments: str
+    status: ResponseStatus | None = None
+
+
+class FunctionCallOutputInputItem(BaseModel, frozen=True):
+    """Function call output item in input (user providing tool results)."""
+
+    type: Literal["function_call_output"] = "function_call_output"
+    call_id: str
+    output: str
+    id: str | None = None
+    status: ResponseStatus | None = None
+
+
+ResponseInputItem = (
+    ResponseInputMessage | FunctionCallInputItem | FunctionCallOutputInputItem
+)
+
+
+class ResponsesRequest(BaseModel, frozen=True):
+    """Request body for OpenAI Responses API.
+
+    This is the API wire type for the Responses endpoint. The canonical
+    internal task params type is ``TextGenerationTaskParams``; see the
+    ``responses_request_to_text_generation`` adapter for conversion.
+    """
+
+    # --- OpenAI Responses API standard fields ---
+    model: ModelId
+    input: str | list[ResponseInputItem]
+    instructions: str | None = None
+    max_output_tokens: int | None = None
+    temperature: float | None = None
+    top_p: float | None = None
+    stream: bool = False
+    tools: list[dict[str, Any]] | None = None
+    metadata: dict[str, str] | None = None
+
+    # --- exo extensions (not in OpenAI Responses API spec) ---
+    top_k: int | None = Field(
+        default=None,
+        description="[exo extension] Top-k sampling parameter. Not part of the OpenAI Responses API.",
+        json_schema_extra={"x-exo-extension": True},
+    )
+    stop: str | list[str] | None = Field(
+        default=None,
+        description="[exo extension] Stop sequence(s). Not part of the OpenAI Responses API.",
+        json_schema_extra={"x-exo-extension": True},
+    )
+    seed: int | None = Field(
+        default=None,
+        description="[exo extension] Seed for deterministic sampling. Not part of the OpenAI Responses API.",
+        json_schema_extra={"x-exo-extension": True},
+    )
+
+    # --- Internal fields (preserved during serialization, hidden from OpenAPI schema) ---
+    chat_template_messages: list[dict[str, Any]] | None = Field(
+        default=None,
+        description="Internal: pre-formatted messages for tokenizer chat template. Not part of the OpenAI Responses API.",
+        json_schema_extra={"x-exo-internal": True},
+    )
+
+
+# Response types
+class ResponseOutputText(BaseModel, frozen=True):
+    """Text content in response output."""
+
+    type: Literal["output_text"] = "output_text"
+    text: str
+    annotations: list[dict[str, str]] = Field(default_factory=list)
+
+
+class ResponseMessageItem(BaseModel, frozen=True):
+    """Message item in response output array."""
+
+    type: Literal["message"] = "message"
+    id: str
+    role: Literal["assistant"] = "assistant"
+    content: list[ResponseOutputText]
+    status: ResponseStatus = "completed"
+
+
+class ResponseFunctionCallItem(BaseModel, frozen=True):
+    """Function call item in response output array."""
+
+    type: Literal["function_call"] = "function_call"
+    id: str
+    call_id: str
+    name: str
+    arguments: str
+    status: ResponseStatus = "completed"
+
+
+ResponseItem = ResponseMessageItem | ResponseFunctionCallItem
+
+
+class ResponseUsage(BaseModel, frozen=True):
+    """Token usage in Responses API response."""
+
+    input_tokens: int
+    output_tokens: int
+    total_tokens: int
+
+
+class ResponsesResponse(BaseModel, frozen=True):
+    """Response body for OpenAI Responses API."""
+
+    id: str
+    object: Literal["response"] = "response"
+    created_at: int = Field(default_factory=lambda: int(time.time()))
+    status: ResponseStatus = "completed"
+    model: str
+    output: list[ResponseItem]
+    output_text: str
+    usage: ResponseUsage | None = None
+
+
+# Streaming event types
+class ResponseCreatedEvent(BaseModel, frozen=True):
+    """Event sent when response is created."""
+
+    type: Literal["response.created"] = "response.created"
+    sequence_number: int
+    response: ResponsesResponse
+
+
+class ResponseInProgressEvent(BaseModel, frozen=True):
+    """Event sent when response starts processing."""
+
+    type: Literal["response.in_progress"] = "response.in_progress"
+    sequence_number: int
+    response: ResponsesResponse
+
+
+class ResponseOutputItemAddedEvent(BaseModel, frozen=True):
+    """Event sent when an output item is added."""
+
+    type: Literal["response.output_item.added"] = "response.output_item.added"
+    sequence_number: int
+    output_index: int
+    item: ResponseItem
+
+
+class ResponseContentPartAddedEvent(BaseModel, frozen=True):
+    """Event sent when a content part is added."""
+
+    type: Literal["response.content_part.added"] = "response.content_part.added"
+    sequence_number: int
+    item_id: str
+    output_index: int
+    content_index: int
+    part: ResponseOutputText
+
+
+class ResponseTextDeltaEvent(BaseModel, frozen=True):
+    """Event sent for text delta during streaming."""
+
+    type: Literal["response.output_text.delta"] = "response.output_text.delta"
+    sequence_number: int
+    item_id: str
+    output_index: int
+    content_index: int
+    delta: str
+
+
+class ResponseTextDoneEvent(BaseModel, frozen=True):
+    """Event sent when text content is done."""
+
+    type: Literal["response.output_text.done"] = "response.output_text.done"
+    sequence_number: int
+    item_id: str
+    output_index: int
+    content_index: int
+    text: str
+
+
+class ResponseContentPartDoneEvent(BaseModel, frozen=True):
+    """Event sent when a content part is done."""
+
+    type: Literal["response.content_part.done"] = "response.content_part.done"
+    sequence_number: int
+    item_id: str
+    output_index: int
+    content_index: int
+    part: ResponseOutputText
+
+
+class ResponseOutputItemDoneEvent(BaseModel, frozen=True):
+    """Event sent when an output item is done."""
+
+    type: Literal["response.output_item.done"] = "response.output_item.done"
+    sequence_number: int
+    output_index: int
+    item: ResponseItem
+
+
+class ResponseFunctionCallArgumentsDeltaEvent(BaseModel, frozen=True):
+    """Event sent for function call arguments delta during streaming."""
+
+    type: Literal["response.function_call_arguments.delta"] = (
+        "response.function_call_arguments.delta"
+    )
+    sequence_number: int
+    item_id: str
+    output_index: int
+    delta: str
+
+
+class ResponseFunctionCallArgumentsDoneEvent(BaseModel, frozen=True):
+    """Event sent when function call arguments are complete."""
+
+    type: Literal["response.function_call_arguments.done"] = (
+        "response.function_call_arguments.done"
+    )
+    sequence_number: int
+    item_id: str
+    output_index: int
+    name: str
+    arguments: str
+
+
+class ResponseCompletedEvent(BaseModel, frozen=True):
+    """Event sent when response is completed."""
+
+    type: Literal["response.completed"] = "response.completed"
+    sequence_number: int
+    response: ResponsesResponse
+
+
+ResponsesStreamEvent = (
+    ResponseCreatedEvent
+    | ResponseInProgressEvent
+    | ResponseOutputItemAddedEvent
+    | ResponseContentPartAddedEvent
+    | ResponseTextDeltaEvent
+    | ResponseTextDoneEvent
+    | ResponseContentPartDoneEvent
+    | ResponseOutputItemDoneEvent
+    | ResponseFunctionCallArgumentsDeltaEvent
+    | ResponseFunctionCallArgumentsDoneEvent
+    | ResponseCompletedEvent
+)

src/exo/shared/types/profiling.py

@@ -1,5 +1,5 @@
 from collections.abc import Sequence
-from typing import Self
+from typing import Literal, Self
 
 import psutil
 
@@ -48,9 +48,13 @@ class SystemPerformanceProfile(CamelCaseModel):
     ecpu_usage: float = 0.0
 
 
+InterfaceType = Literal["wifi", "ethernet", "maybe_ethernet", "thunderbolt", "unknown"]
+
+
 class NetworkInterfaceInfo(CamelCaseModel):
     name: str
     ip_address: str
+    interface_type: InterfaceType = "unknown"
 
 
 class NodeIdentity(CamelCaseModel):
@@ -71,3 +75,11 @@ class NodeThunderboltInfo(CamelCaseModel):
     """Thunderbolt interface identifiers for a node."""
 
     interfaces: Sequence[ThunderboltIdentifier] = []
+
+
+class ThunderboltBridgeStatus(CamelCaseModel):
+    """Whether the Thunderbolt Bridge network service is enabled on this node."""
+
+    enabled: bool
+    exists: bool
+    service_name: str | None = None

src/exo/shared/types/state.py

@@ -13,10 +13,11 @@ from exo.shared.types.profiling import (
     NodeNetworkInfo,
     NodeThunderboltInfo,
     SystemPerformanceProfile,
+    ThunderboltBridgeStatus,
 )
-from exo.shared.types.tasks import BaseTask, TaskId
+from exo.shared.types.tasks import Task, TaskId
 from exo.shared.types.worker.downloads import DownloadProgress
-from exo.shared.types.worker.instances import BaseInstance, InstanceId
+from exo.shared.types.worker.instances import Instance, InstanceId
 from exo.shared.types.worker.runners import RunnerId, RunnerStatus
 from exo.utils.pydantic_ext import CamelCaseModel
 
@@ -37,10 +38,10 @@ class State(CamelCaseModel):
         strict=True,
         arbitrary_types_allowed=True,
     )
-    instances: Mapping[InstanceId, BaseInstance] = {}
+    instances: Mapping[InstanceId, Instance] = {}
     runners: Mapping[RunnerId, RunnerStatus] = {}
     downloads: Mapping[NodeId, Sequence[DownloadProgress]] = {}
-    tasks: Mapping[TaskId, BaseTask] = {}
+    tasks: Mapping[TaskId, Task] = {}
     last_seen: Mapping[NodeId, datetime] = {}
     topology: Topology = Field(default_factory=Topology)
     last_event_applied_idx: int = Field(default=-1, ge=-1)
@@ -51,16 +52,10 @@ class State(CamelCaseModel):
     node_system: Mapping[NodeId, SystemPerformanceProfile] = {}
     node_network: Mapping[NodeId, NodeNetworkInfo] = {}
     node_thunderbolt: Mapping[NodeId, NodeThunderboltInfo] = {}
+    node_thunderbolt_bridge: Mapping[NodeId, ThunderboltBridgeStatus] = {}
 
-    @field_serializer("instances", mode="plain")
-    def _encode_instances(
-        self, value: Mapping[InstanceId, BaseInstance]
-    ) -> dict[str, Any]:
-        """Serialize instances with full subclass fields."""
-        return {
-            str(k): v.model_dump(by_alias=True, serialize_as_any=True)
-            for k, v in value.items()
-        }
+    # Detected cycles where all nodes have Thunderbolt bridge enabled (>2 nodes)
+    thunderbolt_bridge_cycles: Sequence[Sequence[NodeId]] = []
 
     @field_serializer("topology", mode="plain")
     def _encode_topology(self, value: Topology) -> TopologySnapshot:

src/exo/shared/types/tasks.py

@@ -2,13 +2,12 @@ from enum import Enum
 
 from pydantic import Field
 
-from exo.plugins.type_registry import task_registry
 from exo.shared.types.api import (
-    ChatCompletionTaskParams,
-    ImageEditsInternalParams,
+    ImageEditsTaskParams,
     ImageGenerationTaskParams,
 )
 from exo.shared.types.common import CommandId, Id
+from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import BoundInstance, InstanceId
 from exo.shared.types.worker.runners import RunnerId
 from exo.shared.types.worker.shards import ShardMetadata
@@ -33,41 +32,34 @@ class BaseTask(TaggedModel):
     instance_id: InstanceId
 
 
-@task_registry.register
 class CreateRunner(BaseTask):  # emitted by Worker
     bound_instance: BoundInstance
 
 
-@task_registry.register
 class DownloadModel(BaseTask):  # emitted by Worker
     shard_metadata: ShardMetadata
 
 
-@task_registry.register
 class LoadModel(BaseTask):  # emitted by Worker
     pass
 
 
-@task_registry.register
 class ConnectToGroup(BaseTask):  # emitted by Worker
     pass
 
 
-@task_registry.register
 class StartWarmup(BaseTask):  # emitted by Worker
     pass
 
 
-@task_registry.register
-class ChatCompletion(BaseTask):  # emitted by Master
+class TextGeneration(BaseTask):  # emitted by Master
     command_id: CommandId
-    task_params: ChatCompletionTaskParams
+    task_params: TextGenerationTaskParams
 
     error_type: str | None = Field(default=None)
     error_message: str | None = Field(default=None)
 
 
-@task_registry.register
 class ImageGeneration(BaseTask):  # emitted by Master
     command_id: CommandId
     task_params: ImageGenerationTaskParams
@@ -76,28 +68,25 @@ class ImageGeneration(BaseTask):  # emitted by Master
     error_message: str | None = Field(default=None)
 
 
-@task_registry.register
 class ImageEdits(BaseTask):  # emitted by Master
     command_id: CommandId
-    task_params: ImageEditsInternalParams
+    task_params: ImageEditsTaskParams
 
     error_type: str | None = Field(default=None)
     error_message: str | None = Field(default=None)
 
 
-@task_registry.register
 class Shutdown(BaseTask):  # emitted by Worker
     runner_id: RunnerId
 
 
-# Union type for Pydantic validation - tries each type in order
 Task = (
     CreateRunner
     | DownloadModel
     | ConnectToGroup
     | LoadModel
     | StartWarmup
-    | ChatCompletion
+    | TextGeneration
     | ImageGeneration
     | ImageEdits
     | Shutdown

src/exo/shared/types/text_generation.py

@@ -0,0 +1,42 @@
+"""Canonical internal type for text generation task parameters.
+
+All external API formats (Chat Completions, Claude Messages, OpenAI Responses)
+are converted to TextGenerationTaskParams at the API boundary via adapters.
+"""
+
+from typing import Any, Literal
+
+from pydantic import BaseModel
+
+from exo.shared.types.common import ModelId
+
+MessageRole = Literal["user", "assistant", "system", "developer"]
+
+
+class InputMessage(BaseModel, frozen=True):
+    """Internal message for text generation pipelines."""
+
+    role: MessageRole
+    content: str
+
+
+class TextGenerationTaskParams(BaseModel, frozen=True):
+    """Canonical internal task params for text generation.
+
+    Every API adapter converts its wire type into this before handing
+    off to the master/worker pipeline.
+    """
+
+    model: ModelId
+    input: str | list[InputMessage]
+    instructions: str | None = None
+    max_output_tokens: int | None = None
+    temperature: float | None = None
+    top_p: float | None = None
+    stream: bool = False
+    tools: list[dict[str, Any]] | None = None
+    bench: bool = False
+    top_k: int | None = None
+    stop: str | list[str] | None = None
+    seed: int | None = None
+    chat_template_messages: list[dict[str, Any]] | None = None