feat: add --bootstrap-peer flag to bypass mDNS for peer discovery

macOS TCC blocks mDNS multicast from SSH sessions, preventing
SSH-spawned exo processes from discovering peers. This adds a
--bootstrap-peer IP:PORT flag that directly dials a known peer,
bypassing mDNS. The flag is additive — mDNS still runs alongside.

Changes across Rust networking, PyO3 bindings, and Python CLI:
- discovery.rs: store bootstrap addresses, dial on startup and retry
- swarm.rs: thread bootstrap_peers through to Behaviour
- networking.rs: accept optional bootstrap_peers in PyO3 constructor
- router.py: pass bootstrap_peers to NetworkingHandle
- main.py: parse --bootstrap-peer CLI arg into multiaddr format

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

.github/workflows/build-app.yml

@@ -396,7 +396,7 @@ jobs:
           path: output/EXO-${{ env.RELEASE_VERSION }}.dmg
 
       - name: Upload to S3
-        if: env.SPARKLE_S3_BUCKET != '' && github.ref_type == 'tag'
+        if: env.SPARKLE_S3_BUCKET != ''
         env:
           AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
           AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
@@ -412,6 +412,12 @@ jobs:
             PREFIX="${PREFIX}/"
           fi
           DMG_NAME="EXO-${RELEASE_VERSION}.dmg"
+
+          if [[ "${{ github.ref_type }}" != "tag" ]]; then
+            aws s3 cp "$DMG_NAME" "s3://${SPARKLE_S3_BUCKET}/${PREFIX}EXO-${GITHUB_SHA}.dmg"
+            exit 0
+          fi
+
           aws s3 cp "$DMG_NAME" "s3://${SPARKLE_S3_BUCKET}/${PREFIX}${DMG_NAME}"
           if [[ "$IS_ALPHA" != "true" ]]; then
             aws s3 cp "$DMG_NAME" "s3://${SPARKLE_S3_BUCKET}/${PREFIX}EXO-latest.dmg"

.github/workflows/pipeline.yml

@@ -8,33 +8,6 @@ on:
       - main
 
 jobs:
-  typecheck:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-        with:
-          lfs: false
-
-      - uses: cachix/install-nix-action@v31
-        with:
-          nix_path: nixpkgs=channel:nixos-unstable
-
-      - uses: cachix/cachix-action@v14
-        name: Configure Cachix
-        with:
-          name: exo
-          authToken: "${{ secrets.CACHIX_AUTH_TOKEN }}"
-
-      - name: Load nix develop environment
-        run: nix run github:nicknovitski/nix-develop/v1
-
-      - name: Sync dependencies
-        run: uv sync --all-packages
-
-      - name: Run type checker
-        run: uv run basedpyright --project pyproject.toml
-
   nix:
     name: Build and check (${{ matrix.system }})
     runs-on: ${{ matrix.runner }}

.mlx_typings/mlx/core/__init__.pyi

@@ -2366,7 +2366,7 @@ class custom_function:
 def default_device() -> Device:
     """Get the default device."""
 
-def default_stream(device: Device) -> Stream:
+def default_stream(device: Device | DeviceType) -> Stream:
     """Get the device's default stream."""
 
 def degrees(a: array, /, *, stream: Stream | Device | None = ...) -> array:

.mlx_typings/mlx_lm/models/step3p5.pyi

@@ -0,0 +1,151 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+from .switch_layers import SwitchGLU
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    vocab_size: int
+    num_attention_heads: int
+    num_attention_groups: int
+    head_dim: int
+    intermediate_size: int
+    rms_norm_eps: float
+    rope_theta: float
+    rope_scaling: Optional[Dict[str, Any]]
+    max_position_embeddings: int
+    sliding_window: int
+    layer_types: Optional[List[str]]
+    yarn_only_types: Optional[List[str]]
+    partial_rotary_factors: Optional[List[float]]
+    attention_other_setting: Optional[Dict[str, Any]]
+    use_head_wise_attn_gate: bool
+    moe_num_experts: int
+    moe_top_k: int
+    moe_intermediate_size: int
+    share_expert_dim: int
+    moe_layers_enum: Optional[str]
+    moe_router_scaling_factor: float
+    norm_expert_weight: bool
+    swiglu_limits: Optional[List[float]]
+    swiglu_limits_shared: Optional[List[float]]
+    tie_word_embeddings: bool
+
+class Step3p5MLP(nn.Module):
+    hidden_size: int
+    intermediate_size: int
+    gate_proj: nn.Linear
+    up_proj: nn.Linear
+    down_proj: nn.Linear
+    limit: Optional[float]
+
+    def __init__(
+        self, args: ModelArgs, intermediate_size: int, swiglu_limit: float = 0
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Step3p5MoEGate(nn.Module):
+    top_k: int
+    n_routed_experts: int
+    routed_scaling_factor: float
+    norm_topk_prob: bool
+    gate: nn.Linear
+    router_bias: mx.array
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(self, x: mx.array) -> tuple[mx.array, mx.array]: ...
+
+class Step3p5MoE(nn.Module):
+    gate: Step3p5MoEGate
+    switch_mlp: SwitchGLU
+    share_expert: Step3p5MLP
+    sharding_group: Optional[mx.distributed.Group]
+
+    def __init__(self, args: ModelArgs, layer_idx: int) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Step3p5Attention(nn.Module):
+    is_sliding: bool
+    num_heads: int
+    num_kv_heads: int
+    head_dim: int
+    scale: float
+    q_proj: nn.Linear
+    k_proj: nn.Linear
+    v_proj: nn.Linear
+    o_proj: nn.Linear
+    q_norm: nn.Module
+    k_norm: nn.Module
+    use_head_wise_attn_gate: bool
+    g_proj: nn.Linear
+    rope: nn.Module
+
+    def __init__(self, args: ModelArgs, layer_idx: int) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Step3p5DecoderLayer(nn.Module):
+    self_attn: Step3p5Attention
+    is_sliding: bool
+    is_moe_layer: bool
+    mlp: Step3p5MLP | Step3p5MoE
+    input_layernorm: nn.Module
+    post_attention_layernorm: nn.Module
+
+    def __init__(self, args: ModelArgs, layer_idx: int) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Step3p5Model(nn.Module):
+    args: ModelArgs
+    vocab_size: int
+    num_layers: int
+    embed_tokens: nn.Embedding
+    layers: list[Step3p5DecoderLayer]
+    norm: nn.Module
+    _swa_idx: Optional[int]
+    _full_idx: Optional[int]
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        cache: Optional[List[Any]] = None,
+    ) -> mx.array: ...
+
+class Model(nn.Module):
+    args: ModelArgs
+    model_type: str
+    model: Step3p5Model
+    lm_head: nn.Linear
+
+    def __init__(self, args: ModelArgs) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[List[Any]] = None,
+    ) -> mx.array: ...
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...
+    def shard(self, group: Optional[mx.distributed.Group] = None) -> None: ...
+    @property
+    def layers(self) -> list[Step3p5DecoderLayer]: ...
+    def make_cache(self) -> list[Any]: ...
+    @property
+    def cast_predicate(self) -> Any: ...
+    @property
+    def quant_predicate(self) -> Any: ...

AGENTS.md

@@ -119,3 +119,78 @@ From .cursorrules:
 ## Testing
 
 Tests use pytest-asyncio with `asyncio_mode = "auto"`. Tests are in `tests/` subdirectories alongside the code they test. The `EXO_TESTS=1` env var is set during tests.
+
+## Dashboard UI Testing & Screenshots
+
+### Building and Running the Dashboard
+```bash
+# Build the dashboard (must be done before running exo)
+cd dashboard && npm install && npm run build && cd ..
+
+# Start exo (serves the dashboard at http://localhost:52415)
+uv run exo &
+sleep 8  # Wait for server to start
+```
+
+### Taking Headless Screenshots with Playwright
+Use Playwright with headless Chromium for programmatic screenshots — no manual browser interaction needed.
+
+**Setup (one-time):**
+```bash
+npx --yes playwright install chromium
+cd /tmp && npm init -y && npm install playwright
+```
+
+**Taking screenshots:**
+```javascript
+// Run from /tmp where playwright is installed: cd /tmp && node -e "..."
+const { chromium } = require('playwright');
+(async () => {
+  const browser = await chromium.launch({ headless: true });
+  const page = await browser.newPage({ viewport: { width: 1280, height: 800 } });
+  await page.goto('http://localhost:52415', { waitUntil: 'networkidle' });
+  await page.waitForTimeout(2000);
+
+  // Inject test data into localStorage if needed (e.g., recent models)
+  await page.evaluate(() => {
+    localStorage.setItem('exo-recent-models', JSON.stringify([
+      { modelId: 'mlx-community/Qwen3-30B-A3B-4bit', launchedAt: Date.now() },
+    ]));
+  });
+  await page.reload({ waitUntil: 'networkidle' });
+  await page.waitForTimeout(2000);
+
+  // Interact with UI elements
+  await page.locator('text=SELECT MODEL').click();
+  await page.waitForTimeout(1000);
+
+  // Take screenshot
+  await page.screenshot({ path: '/tmp/screenshot.png', fullPage: false });
+  await browser.close();
+})();
+```
+
+### Uploading Images to GitHub PRs
+GitHub's API doesn't support direct image upload for PR comments. Workaround:
+
+1. **Commit images to the branch** (temporarily):
+   ```bash
+   cp /tmp/screenshot.png .
+   git add screenshot.png
+   git commit -m "temp: add screenshots for PR"
+   git push origin <branch>
+   COMMIT_SHA=$(git rev-parse HEAD)
+   ```
+
+2. **Post PR comment** referencing the raw image URL (uses permanent commit SHA so images survive deletion):
+   ```bash
+   gh pr comment <PR_NUMBER> --body "![Screenshot](https://raw.githubusercontent.com/exo-explore/exo/${COMMIT_SHA}/screenshot.png)"
+   ```
+
+3. **Remove the images** from the branch:
+   ```bash
+   git rm screenshot.png
+   git commit -m "chore: remove temporary screenshot files"
+   git push origin <branch>
+   ```
+   The images still render in the PR comment because they reference the permanent commit SHA.

Cargo.lock

@@ -353,31 +353,6 @@ dependencies = [
  "generic-array",
 ]
 
-[[package]]
-name = "bon"
-version = "3.8.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ebeb9aaf9329dff6ceb65c689ca3db33dbf15f324909c60e4e5eef5701ce31b1"
-dependencies = [
- "bon-macros",
- "rustversion",
-]
-
-[[package]]
-name = "bon-macros"
-version = "3.8.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "77e9d642a7e3a318e37c2c9427b5a6a48aa1ad55dcd986f3034ab2239045a645"
-dependencies = [
- "darling",
- "ident_case",
- "prettyplease",
- "proc-macro2",
- "quote",
- "rustversion",
- "syn 2.0.111",
-]
-
 [[package]]
 name = "bs58"
 version = "0.5.1"
@@ -700,41 +675,6 @@ dependencies = [
  "syn 2.0.111",
 ]
 
-[[package]]
-name = "darling"
-version = "0.21.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9cdf337090841a411e2a7f3deb9187445851f91b309c0c0a29e05f74a00a48c0"
-dependencies = [
- "darling_core",
- "darling_macro",
-]
-
-[[package]]
-name = "darling_core"
-version = "0.21.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1247195ecd7e3c85f83c8d2a366e4210d588e802133e1e355180a9870b517ea4"
-dependencies = [
- "fnv",
- "ident_case",
- "proc-macro2",
- "quote",
- "strsim",
- "syn 2.0.111",
-]
-
-[[package]]
-name = "darling_macro"
-version = "0.21.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d38308df82d1080de0afee5d069fa14b0326a88c14f15c5ccda35b4a6c414c81"
-dependencies = [
- "darling_core",
- "quote",
- "syn 2.0.111",
-]
-
 [[package]]
 name = "data-encoding"
 version = "2.9.0"
@@ -1625,12 +1565,6 @@ dependencies = [
  "zerovec",
 ]
 
-[[package]]
-name = "ident_case"
-version = "1.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9e0384b61958566e926dc50660321d12159025e767c18e043daf26b70104c39"
-
 [[package]]
 name = "idna"
 version = "1.1.0"
@@ -1745,15 +1679,6 @@ dependencies = [
  "generic-array",
 ]
 
-[[package]]
-name = "internment"
-version = "0.8.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "636d4b0f6a39fd684effe2a73f5310df16a3fa7954c26d36833e98f44d1977a2"
-dependencies = [
- "hashbrown 0.15.5",
-]
-
 [[package]]
 name = "inventory"
 version = "0.3.21"
@@ -3307,16 +3232,6 @@ dependencies = [
  "zerocopy",
 ]
 
-[[package]]
-name = "prettyplease"
-version = "0.2.37"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b"
-dependencies = [
- "proc-macro2",
- "syn 2.0.111",
-]
-
 [[package]]
 name = "primeorder"
 version = "0.13.6"
@@ -3740,12 +3655,6 @@ dependencies = [
  "yasna",
 ]
 
-[[package]]
-name = "recursion"
-version = "0.5.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9dba2197bf7b1d87b4dd460c195f4edeb45a94e82e8054f8d5f317c1f0e93ca1"
-
 [[package]]
 name = "redox_syscall"
 version = "0.5.18"
@@ -4807,15 +4716,6 @@ checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
 [[package]]
 name = "util"
 version = "0.0.1"
-dependencies = [
- "bon",
- "derive_more",
- "extend",
- "internment",
- "once_cell",
- "recursion",
- "thiserror 2.0.17",
-]
 
 [[package]]
 name = "uuid"

MISSED_THINGS.md

@@ -5,21 +5,21 @@
 [X] Fetching download status of all models on start
 [X] Deduplication of tasks in plan_step.
 [X] resolve_allow_patterns should just be wildcard now.
-[] no mx_barrier in genreate.py mlx_generate at the end.
+[X] no mx_barrier in genreate.py mlx_generate at the end.
 [] cache assertion not needed in auto_parallel.py PipelineLastLayer.
-[] GPTOSS support dropped in auto_parallel.py.
-[] sharding changed "all-to-sharded" became _all_to_sharded in auto_parallel.py.
-[] same as above with "sharded-to-all" became _sharded_to_all in auto_parallel.py.
-[] Dropped support for Ministral3Model, DeepseekV32Model, Glm4MoeModel, Qwen3NextModel, GptOssMode in auto_parallel.py.
+[X] GPTOSS support dropped in auto_parallel.py.
+[X] sharding changed "all-to-sharded" became _all_to_sharded in auto_parallel.py.
+[X] same as above with "sharded-to-all" became _sharded_to_all in auto_parallel.py.
+[X] Dropped support for Ministral3Model, DeepseekV32Model, Glm4MoeModel, Qwen3NextModel, GptOssMode in auto_parallel.py.
 [] Dropped prefill/decode code in auto_parallel.py and utils_mlx.py.
 [X] KV_CACHE_BITS should be None to disable quantized KV cache.
-[] Dropped _set_nofile_limit in utils_mlx.py.
-[] We have group optional in load_mlx_items in utils_mlx.py.
-[] Dropped add_missing_chat_templates for GptOss in load_mlx_items in utils_mlx.py.
-[] Dropped model.make_cache in make_kv_cache in utils_mlx.py.
+[X] Dropped _set_nofile_limit in utils_mlx.py.
+[X] We have group optional in load_mlx_items in utils_mlx.py.
+[X] Dropped add_missing_chat_templates for GptOss in load_mlx_items in utils_mlx.py.
+[X] Dropped model.make_cache in make_kv_cache in utils_mlx.py.
 [X] We put cache limit back in utils_mlx.py.
-[] topology.py remove_node removes the connections after checking if node is is in self._node_id_to_rx_id_map. on beta_1 it checks after, so would remove stale connections I guess?
-[] Missing Glm 4.7 model cards (this isn't ready yet but should be picked up, probably create an issue... the blocker is transforemrs version doesn't support the tokenizer for Glm 4.7. rc-1 does but we can't upgrade as it breaks other things.)
+[X] topology.py remove_node removes the connections after checking if node is is in self._node_id_to_rx_id_map. on beta_1 it checks after, so would remove stale connections I guess?
+[X] Missing Glm 4.7 model cards (this isn't ready yet but should be picked up, probably create an issue... the blocker is transforemrs version doesn't support the tokenizer for Glm 4.7. rc-1 does but we can't upgrade as it breaks other things.)
 [] try-except in _command_processor only excepts ValueError. This was silently failing leading to un-debuggable errors (we had a KeyError that was happening ). Changed this to catch Exception instead of ValueError. See exo-v2 89ae38405e0052e3c22405daf094b065878aa873 and fb99fea69b5a39017efc90c5dad0072e677455f0.
 [X] In placement.py, place_instance no longer looks at model_meta.supports_tensor and check if this tensor parallel number of nodes is supported by the model's tensor dimensions.
 [X] In placement.py, place_instanec, we no longer have the special case to exclude DeepSeek v3.1 pipeline parallel (it doesn't work).

TODO.md

@@ -1,28 +1,14 @@
-2. Currently a lot of requests from the API are timing out, but we still process those requests internally. If an API request times out, we should cancel all corresponding tasks to that API request (why process a request with nobody listening).
 3. Task cancellation. When API http request gets cancelled, it should cancel corresponding task.
 4. I'd like to see profiled network latency / bandwidth.
 5. I'd like to see how much bandwidth each link is using.
-6. We should handle the case where one machine doesn't have the model downloaded and then other machines are waiting on it. In this case we get loads of timeout errors because the others are waiting for the one that needs to download the model.
 7. Solve the problem of in continuous batching when a new prompt comes in, it will block decode of the current batch until the prefill is complete.
 8. We want people to be able to copy models over to a new device without ever connecting EXO to the internet. Right now EXO require internet connection once to cache some files to check if a download is complete. Instead, we should simply check if there is a non-empty model folder locally with no .partial files. This indicates it's a fully downloaded model that can be loaded.
-10. More granular control over how to deploy instances.
-12. Nix is great but installing it is a pain and we have ended up in a lot of cases having PATH issues or installation issues. For example, after rebooting mike it seemed to no longer have a nix installation and needed reinstalling. It has a bunch of broken symlinks left over from nix that caused ssh to fail, making it even harder to debug. We need consistent environments (perhaps MDM) so we can guarantee nix is installed properly on each machine.
 13. Memory pressure instead of memory used.
 14. Show the type of each connection (TB5, Ethernet, etc.) in the UI. Refer to old exo: https://github.com/exo-explore/exo/blob/56f783b38dc6b08ce606b07a5386dc40dae00330/exo/helpers.py#L251
 15. Prioritise certain connection types (or by latency). TB5 > Ethernet > WiFi. Refer to old exo: https://github.com/exo-explore/exo/blob/56f783b38dc6b08ce606b07a5386dc40dae00330/exo/helpers.py#L251
 16. Dynamically switch to higher priority connection when it becomes available. Probably bring back InstanceReplacedAtomically.
 17. Faster model loads by streaming model from other devices in cluster.
 18. Add support for specifying the type of network connection to use in a test. Depends on 15/16.
-20. Add chat completion cancellations (e.g OpenWebUI has something for cancelling an ongoing request).
-23. Do we need cache_limit? We went back and forth on that a lot because we thought it might be causing issues. One problem is it sets it relative to model size. So if you have multiple models loaded in it will take the most recent model size for the cache_limit. This is problematic if you launch DeepSeek -> Llama for example.
-24. further openai/lmstudio api compatibility
 25. Rethink retry logic
-26. Task cancellation. When API http request gets cancelled, it should cancel corresponding task.
 27. Log cleanup - per-module log filters and default to DEBUG log levels
 28. Validate RDMA connections with ibv_devinfo in the info gatherer
-
-Potential refactors:
-
-2. Topology can be simplified
-
-Random errors we've run into:

app/EXO/EXO/ContentView.swift

@@ -563,21 +563,45 @@ struct ContentView: View {
     }
 
     private var rdmaStatusView: some View {
-        let rdma = networkStatusService.status.rdmaStatus
+        let rdmaStatuses = stateService.latestSnapshot?.nodeRdmaCtl ?? [:]
+        let localNodeId = stateService.localNodeId
+        let nodeProfiles = stateService.latestSnapshot?.nodeProfiles ?? [:]
+        let localDevices = networkStatusService.status.localRdmaDevices
+        let localPorts = networkStatusService.status.localRdmaActivePorts
+
         return VStack(alignment: .leading, spacing: 1) {
-            Text("RDMA: \(rdmaStatusText(rdma))")
-                .font(.caption2)
-                .foregroundColor(rdmaStatusColor(rdma))
-            if !rdma.devices.isEmpty {
-                Text("  Devices: \(rdma.devices.joined(separator: ", "))")
+            if rdmaStatuses.isEmpty {
+                Text("Cluster RDMA: No data")
+                    .font(.caption2)
+                    .foregroundColor(.secondary)
+            } else {
+                Text("Cluster RDMA Status:")
+                    .font(.caption2)
+                    .foregroundColor(.secondary)
+                ForEach(Array(rdmaStatuses.keys.sorted()), id: \.self) { nodeId in
+                    if let status = rdmaStatuses[nodeId] {
+                        let nodeName =
+                            nodeProfiles[nodeId]?.friendlyName ?? String(nodeId.prefix(8))
+                        let isLocal = nodeId == localNodeId
+                        let prefix = isLocal ? "  \(nodeName) (local):" : "  \(nodeName):"
+                        let statusText = status.enabled ? "Enabled" : "Disabled"
+                        let color: Color = status.enabled ? .green : .orange
+                        Text("\(prefix) \(statusText)")
+                            .font(.caption2)
+                            .foregroundColor(color)
+                    }
+                }
+            }
+            if !localDevices.isEmpty {
+                Text("  Local Devices: \(localDevices.joined(separator: ", "))")
                     .font(.caption2)
                     .foregroundColor(.secondary)
             }
-            if !rdma.activePorts.isEmpty {
-                Text("  Active Ports:")
+            if !localPorts.isEmpty {
+                Text("  Local Active Ports:")
                     .font(.caption2)
                     .foregroundColor(.secondary)
-                ForEach(rdma.activePorts, id: \.device) { port in
+                ForEach(localPorts, id: \.device) { port in
                     Text("    \(port.device) port \(port.port): \(port.state)")
                         .font(.caption2)
                         .foregroundColor(.green)
@@ -586,28 +610,6 @@ struct ContentView: View {
         }
     }
 
-    private func rdmaStatusText(_ rdma: RDMAStatus) -> String {
-        switch rdma.rdmaCtlEnabled {
-        case .some(true):
-            return "Enabled"
-        case .some(false):
-            return "Disabled"
-        case nil:
-            return rdma.devices.isEmpty ? "Not Available" : "Available"
-        }
-    }
-
-    private func rdmaStatusColor(_ rdma: RDMAStatus) -> Color {
-        switch rdma.rdmaCtlEnabled {
-        case .some(true):
-            return .green
-        case .some(false):
-            return .orange
-        case nil:
-            return rdma.devices.isEmpty ? .secondary : .green
-        }
-    }
-
     private var sendBugReportButton: some View {
         VStack(alignment: .leading, spacing: 4) {
             Button {

app/EXO/EXO/ExoProcessController.swift

@@ -28,6 +28,10 @@ final class ExoProcessController: ObservableObject {
         }
     }
 
+    static let exoDirectoryURL: URL = {
+        URL(fileURLWithPath: NSHomeDirectory()).appendingPathComponent(".exo")
+    }()
+
     @Published private(set) var status: Status = .stopped
     @Published private(set) var lastError: String?
     @Published private(set) var launchCountdownSeconds: Int?
@@ -78,7 +82,11 @@ final class ExoProcessController: ObservableObject {
 
             let child = Process()
             child.executableURL = executableURL
-            child.currentDirectoryURL = runtimeURL
+            let exoHomeURL = Self.exoDirectoryURL
+            try? FileManager.default.createDirectory(
+                at: exoHomeURL, withIntermediateDirectories: true
+            )
+            child.currentDirectoryURL = exoHomeURL
             child.environment = makeEnvironment(for: runtimeURL)
 
             child.standardOutput = FileHandle.nullDevice

app/EXO/EXO/Models/ClusterState.swift

@@ -15,6 +15,7 @@ struct ClusterState: Decodable {
     let nodeMemory: [String: MemoryInfo]
     let nodeSystem: [String: SystemInfo]
     let nodeThunderboltBridge: [String: ThunderboltBridgeStatus]
+    let nodeRdmaCtl: [String: NodeRdmaCtlStatus]
 
     /// Computed property for backwards compatibility - merges granular state into NodeProfile
     var nodeProfiles: [String: NodeProfile] {
@@ -65,6 +66,10 @@ struct ClusterState: Decodable {
             try container.decodeIfPresent(
                 [String: ThunderboltBridgeStatus].self, forKey: .nodeThunderboltBridge
             ) ?? [:]
+        self.nodeRdmaCtl =
+            try container.decodeIfPresent(
+                [String: NodeRdmaCtlStatus].self, forKey: .nodeRdmaCtl
+            ) ?? [:]
     }
 
     private enum CodingKeys: String, CodingKey {
@@ -78,6 +83,7 @@ struct ClusterState: Decodable {
         case nodeMemory
         case nodeSystem
         case nodeThunderboltBridge
+        case nodeRdmaCtl
     }
 }
 
@@ -159,6 +165,10 @@ struct ThunderboltBridgeStatus: Decodable {
     let serviceName: String?
 }
 
+struct NodeRdmaCtlStatus: Decodable {
+    let enabled: Bool
+}
+
 struct MemoryInfo: Decodable {
     let ramTotal: MemoryValue?
     let ramAvailable: MemoryValue?

app/EXO/EXO/Services/BugReportService.swift

@@ -44,16 +44,12 @@ struct BugReportService {
         let dayPrefix = Self.dayPrefixString(now)
         let prefix = "reports/\(dayPrefix)/\(timestamp)/"
 
-        let logData = readLog()
+        let logFiles = readAllLogs()
         let ifconfigText = try await captureIfconfig()
         let hostName = Host.current().localizedName ?? "unknown"
         let debugInfo = readDebugInfo()
 
-        async let stateResult = fetch(url: baseURL.appendingPathComponent("state"))
-        async let eventsResult = fetch(url: baseURL.appendingPathComponent("events"))
-
-        let stateData = try await stateResult
-        let eventsData = try await eventsResult
+        let stateData = try await fetch(url: baseURL.appendingPathComponent("state"))
 
         // Extract cluster TB bridge status from exo state
         let clusterTbBridgeStatus = extractClusterTbBridgeStatus(from: stateData)
@@ -67,12 +63,19 @@ struct BugReportService {
             clusterTbBridgeStatus: clusterTbBridgeStatus
         )
 
-        let uploads: [(path: String, data: Data?)] = [
-            ("\(prefix)exo.log", logData),
+        let eventLogFiles = readAllEventLogs()
+
+        var uploads: [(path: String, data: Data?)] = logFiles.map { (path, data) in
+            ("\(prefix)\(path)", data)
+        }
+        uploads.append(
+            contentsOf: eventLogFiles.map { (path, data) in
+                ("\(prefix)\(path)", data as Data?)
+            })
+        uploads.append(contentsOf: [
             ("\(prefix)state.json", stateData),
-            ("\(prefix)events.json", eventsData),
             ("\(prefix)report.json", reportJSON),
-        ]
+        ])
 
         let uploadItems: [(key: String, body: Data)] = uploads.compactMap { item in
             guard let body = item.data else { return nil }
@@ -149,11 +152,40 @@ struct BugReportService {
         return decoded.urls
     }
 
-    private func readLog() -> Data? {
-        let logURL = URL(fileURLWithPath: NSHomeDirectory())
+    private func readAllLogs() -> [(path: String, data: Data)] {
+        let dir = URL(fileURLWithPath: NSHomeDirectory())
             .appendingPathComponent(".exo")
-            .appendingPathComponent("exo.log")
-        return try? Data(contentsOf: logURL)
+            .appendingPathComponent("exo_log")
+        var results: [(path: String, data: Data)] = []
+
+        let contents = (try? FileManager.default.contentsOfDirectory(atPath: dir.path)) ?? []
+        for name in contents {
+            if let data = try? Data(contentsOf: dir.appendingPathComponent(name)) {
+                results.append(("exo_log/\(name)", data))
+            }
+        }
+
+        return results
+    }
+
+    private func readAllEventLogs() -> [(path: String, data: Data)] {
+        let eventLogDir = URL(fileURLWithPath: NSHomeDirectory())
+            .appendingPathComponent(".exo")
+            .appendingPathComponent("event_log")
+        var results: [(path: String, data: Data)] = []
+
+        for subdir in ["master", "api"] {
+            let dir = eventLogDir.appendingPathComponent(subdir)
+            let contents =
+                (try? FileManager.default.contentsOfDirectory(atPath: dir.path)) ?? []
+            for name in contents where name.hasPrefix("events.") {
+                if let data = try? Data(contentsOf: dir.appendingPathComponent(name)) {
+                    results.append(("event_log/\(subdir)/\(name)", data))
+                }
+            }
+        }
+
+        return results
     }
 
     private func captureIfconfig() async throws -> String {

app/EXO/EXO/Services/NetworkStatusService.swift

@@ -35,28 +35,18 @@ struct NetworkStatus: Equatable {
     let thunderboltBridgeState: ThunderboltState?
     let bridgeInactive: Bool?
     let interfaceStatuses: [InterfaceIpStatus]
-    let rdmaStatus: RDMAStatus
+    let localRdmaDevices: [String]
+    let localRdmaActivePorts: [RDMAPort]
 
     static let empty = NetworkStatus(
         thunderboltBridgeState: nil,
         bridgeInactive: nil,
         interfaceStatuses: [],
-        rdmaStatus: .empty
+        localRdmaDevices: [],
+        localRdmaActivePorts: []
     )
 }
 
-struct RDMAStatus: Equatable {
-    let rdmaCtlEnabled: Bool?
-    let devices: [String]
-    let activePorts: [RDMAPort]
-
-    var isAvailable: Bool {
-        rdmaCtlEnabled == true || !devices.isEmpty
-    }
-
-    static let empty = RDMAStatus(rdmaCtlEnabled: nil, devices: [], activePorts: [])
-}
-
 struct RDMAPort: Equatable {
     let device: String
     let port: String
@@ -80,31 +70,11 @@ private struct NetworkStatusFetcher {
             thunderboltBridgeState: readThunderboltBridgeState(),
             bridgeInactive: readBridgeInactive(),
             interfaceStatuses: readInterfaceStatuses(),
-            rdmaStatus: readRDMAStatus()
+            localRdmaDevices: readRDMADevices(),
+            localRdmaActivePorts: readRDMAActivePorts()
         )
     }
 
-    private func readRDMAStatus() -> RDMAStatus {
-        let rdmaCtlEnabled = readRDMACtlEnabled()
-        let devices = readRDMADevices()
-        let activePorts = readRDMAActivePorts()
-        return RDMAStatus(
-            rdmaCtlEnabled: rdmaCtlEnabled, devices: devices, activePorts: activePorts)
-    }
-
-    private func readRDMACtlEnabled() -> Bool? {
-        let result = runCommand(["rdma_ctl", "status"])
-        guard result.exitCode == 0 else { 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 func readRDMADevices() -> [String] {
         let result = runCommand(["ibv_devices"])
         guard result.exitCode == 0 else { return [] }

bench/exo_bench.py

@@ -19,6 +19,11 @@ from urllib.parse import urlencode
 from loguru import logger
 from transformers import AutoTokenizer
 
+# Backoff constants for cluster settling retry
+_SETTLE_INITIAL_BACKOFF_S = 1.0
+_SETTLE_MAX_BACKOFF_S = 60.0
+_SETTLE_BACKOFF_MULTIPLIER = 2.0
+
 # 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
@@ -283,6 +288,151 @@ def resolve_model_short_id(client: ExoClient, model_arg: str) -> tuple[str, str]
     raise ValueError(f"Model not found in /models: {model_arg}")
 
 
+def run_planning_phase(
+    client: ExoClient,
+    full_model_id: str,
+    preview: dict[str, Any],
+    danger_delete: bool,
+    timeout: float,
+    settle_deadline: float | None,
+) -> None:
+    """Check disk space and ensure model is downloaded before benchmarking."""
+    # Get model size from /models
+    models = client.request_json("GET", "/models") or {}
+    model_bytes = 0
+    for m in models.get("data", []):
+        if m.get("hugging_face_id") == full_model_id:
+            model_bytes = m.get("storage_size_megabytes", 0) * 1024 * 1024
+            break
+
+    if not model_bytes:
+        logger.warning(
+            f"Could not determine size for {full_model_id}, skipping disk check"
+        )
+        return
+
+    # Get nodes from preview
+    inner = unwrap_instance(preview["instance"])
+    node_ids = list(inner["shardAssignments"]["nodeToRunner"].keys())
+    runner_to_shard = inner["shardAssignments"]["runnerToShard"]
+
+    state = client.request_json("GET", "/state")
+    downloads = state.get("downloads", {})
+    node_disk = state.get("nodeDisk", {})
+
+    for node_id in node_ids:
+        node_downloads = downloads.get(node_id, [])
+
+        # Check if model already downloaded on this node
+        already_downloaded = any(
+            "DownloadCompleted" in p
+            and unwrap_instance(p["DownloadCompleted"]["shardMetadata"])["modelCard"][
+                "modelId"
+            ]
+            == full_model_id
+            for p in node_downloads
+        )
+        if already_downloaded:
+            continue
+
+        # Wait for disk info if settle_deadline is set
+        disk_info = node_disk.get(node_id, {})
+        backoff = _SETTLE_INITIAL_BACKOFF_S
+        while not disk_info and settle_deadline and time.monotonic() < settle_deadline:
+            remaining = settle_deadline - time.monotonic()
+            logger.info(
+                f"Waiting for disk info on {node_id} ({remaining:.0f}s remaining)..."
+            )
+            time.sleep(min(backoff, remaining))
+            backoff = min(backoff * _SETTLE_BACKOFF_MULTIPLIER, _SETTLE_MAX_BACKOFF_S)
+            state = client.request_json("GET", "/state")
+            node_disk = state.get("nodeDisk", {})
+            disk_info = node_disk.get(node_id, {})
+
+        if not disk_info:
+            logger.warning(f"No disk info for {node_id}, skipping space check")
+            continue
+
+        avail = disk_info.get("available", {}).get("inBytes", 0)
+        if avail >= model_bytes:
+            continue
+
+        if not danger_delete:
+            raise RuntimeError(
+                f"Insufficient disk on {node_id}: need {model_bytes // (1024**3)}GB, "
+                f"have {avail // (1024**3)}GB. Use --danger-delete-downloads to free space."
+            )
+
+        # Delete from smallest to largest
+        completed = [
+            (
+                unwrap_instance(p["DownloadCompleted"]["shardMetadata"])["modelCard"][
+                    "modelId"
+                ],
+                p["DownloadCompleted"]["totalBytes"]["inBytes"],
+            )
+            for p in node_downloads
+            if "DownloadCompleted" in p
+        ]
+        for del_model, size in sorted(completed, key=lambda x: x[1]):
+            logger.info(f"Deleting {del_model} from {node_id} ({size // (1024**2)}MB)")
+            client.request_json("DELETE", f"/download/{node_id}/{del_model}")
+            avail += size
+            if avail >= model_bytes:
+                break
+
+        if avail < model_bytes:
+            raise RuntimeError(f"Could not free enough space on {node_id}")
+
+    # Start downloads (idempotent)
+    for node_id in node_ids:
+        runner_id = inner["shardAssignments"]["nodeToRunner"][node_id]
+        shard = runner_to_shard[runner_id]
+        client.request_json(
+            "POST",
+            "/download/start",
+            body={
+                "targetNodeId": node_id,
+                "shardMetadata": shard,
+            },
+        )
+        logger.info(f"Started download on {node_id}")
+
+    # Wait for downloads
+    start = time.time()
+    while time.time() - start < timeout:
+        state = client.request_json("GET", "/state")
+        downloads = state.get("downloads", {})
+        all_done = True
+        for node_id in node_ids:
+            done = any(
+                "DownloadCompleted" in p
+                and unwrap_instance(p["DownloadCompleted"]["shardMetadata"])[
+                    "modelCard"
+                ]["modelId"]
+                == full_model_id
+                for p in downloads.get(node_id, [])
+            )
+            failed = [
+                p["DownloadFailed"]["errorMessage"]
+                for p in downloads.get(node_id, [])
+                if "DownloadFailed" in p
+                and unwrap_instance(p["DownloadFailed"]["shardMetadata"])["modelCard"][
+                    "modelId"
+                ]
+                == full_model_id
+            ]
+            if failed:
+                raise RuntimeError(f"Download failed on {node_id}: {failed[0]}")
+            if not done:
+                all_done = False
+        if all_done:
+            return
+        time.sleep(1)
+
+    raise TimeoutError("Downloads did not complete in time")
+
+
 def placement_filter(instance_meta: str, wanted: str) -> bool:
     s = (instance_meta or "").lower()
     if wanted == "both":
@@ -388,6 +538,66 @@ class PromptSizer:
         return content, tok
 
 
+def fetch_and_filter_placements(
+    client: ExoClient, full_model_id: str, args: argparse.Namespace
+) -> list[dict[str, Any]]:
+    previews_resp = client.request_json(
+        "GET", "/instance/previews", params={"model_id": full_model_id}
+    )
+    previews = previews_resp.get("previews") or []
+
+    selected: list[dict[str, Any]] = []
+    for p in previews:
+        if p.get("error") is not None:
+            continue
+        if not placement_filter(str(p.get("instance_meta", "")), args.instance_meta):
+            continue
+        if not sharding_filter(str(p.get("sharding", "")), args.sharding):
+            continue
+
+        instance = p.get("instance")
+        if not isinstance(instance, dict):
+            continue
+
+        n = nodes_used_in_instance(instance)
+        # Skip tensor ring single node as it is pointless when pipeline ring
+        if n == 1 and (
+            (args.sharding == "both" and "tensor" in p.get("sharding", "").lower())
+            or (
+                args.instance_meta == "both"
+                and "jaccl" in p.get("instance_meta", "").lower()
+            )
+        ):
+            continue
+
+        if (
+            args.skip_pipeline_jaccl
+            and (
+                args.instance_meta == "both"
+                and "jaccl" in p.get("instance_meta", "").lower()
+            )
+            and (
+                args.sharding == "both" and "pipeline" in p.get("sharding", "").lower()
+            )
+        ):
+            continue
+
+        if (
+            args.skip_tensor_ring
+            and (
+                args.instance_meta == "both"
+                and "ring" in p.get("instance_meta", "").lower()
+            )
+            and (args.sharding == "both" and "tensor" in p.get("sharding", "").lower())
+        ):
+            continue
+
+        if args.min_nodes <= n <= args.max_nodes:
+            selected.append(p)
+
+    return selected
+
+
 def main() -> int:
     ap = argparse.ArgumentParser(
         prog="exo-bench",
@@ -464,6 +674,17 @@ def main() -> int:
         action="store_true",
         help="Force all pp×tg combinations (cartesian product) even when lists have equal length.",
     )
+    ap.add_argument(
+        "--settle-timeout",
+        type=float,
+        default=0,
+        help="Max seconds to wait for the cluster to produce valid placements (0 = try once).",
+    )
+    ap.add_argument(
+        "--danger-delete-downloads",
+        action="store_true",
+        help="Delete existing models from smallest to largest to make room for benchmark model.",
+    )
     args = ap.parse_args()
 
     pp_list = parse_int_list(args.pp)
@@ -487,11 +708,6 @@ def main() -> int:
     client = ExoClient(args.host, args.port, timeout_s=args.timeout)
     short_id, full_model_id = resolve_model_short_id(client, args.model)
 
-    previews_resp = client.request_json(
-        "GET", "/instance/previews", params={"model_id": full_model_id}
-    )
-    previews = previews_resp.get("previews") or []
-
     tokenizer = load_tokenizer_for_bench(full_model_id)
     if tokenizer is None:
         raise RuntimeError("[exo-bench] tokenizer load failed")
@@ -503,54 +719,23 @@ def main() -> int:
         logger.error("[exo-bench] tokenizer usable but prompt sizing failed")
         raise
 
-    selected: list[dict[str, Any]] = []
-    for p in previews:
-        if p.get("error") is not None:
-            continue
-        if not placement_filter(str(p.get("instance_meta", "")), args.instance_meta):
-            continue
-        if not sharding_filter(str(p.get("sharding", "")), args.sharding):
-            continue
+    settle_deadline = (
+        time.monotonic() + args.settle_timeout if args.settle_timeout > 0 else None
+    )
 
-        instance = p.get("instance")
-        if not isinstance(instance, dict):
-            continue
+    selected = fetch_and_filter_placements(client, full_model_id, args)
 
-        n = nodes_used_in_instance(instance)
-        # Skip tensor ring single node as it is pointless when pipeline ring
-        if n == 1 and (
-            (args.sharding == "both" and "tensor" in p.get("sharding", "").lower())
-            or (
-                args.instance_meta == "both"
-                and "jaccl" in p.get("instance_meta", "").lower()
+    if not selected and settle_deadline:
+        backoff = _SETTLE_INITIAL_BACKOFF_S
+        while not selected and time.monotonic() < settle_deadline:
+            remaining = settle_deadline - time.monotonic()
+            logger.warning(
+                f"No valid placements yet (cluster may still be settling). "
+                f"Retrying in {backoff:.1f}s ({remaining:.0f}s remaining)..."
             )
-        ):
-            continue
-
-        if (
-            args.skip_pipeline_jaccl
-            and (
-                args.instance_meta == "both"
-                and "jaccl" in p.get("instance_meta", "").lower()
-            )
-            and (
-                args.sharding == "both" and "pipeline" in p.get("sharding", "").lower()
-            )
-        ):
-            continue
-
-        if (
-            args.skip_tensor_ring
-            and (
-                args.instance_meta == "both"
-                and "ring" in p.get("instance_meta", "").lower()
-            )
-            and (args.sharding == "both" and "tensor" in p.get("sharding", "").lower())
-        ):
-            continue
-
-        if args.min_nodes <= n <= args.max_nodes:
-            selected.append(p)
+            time.sleep(min(backoff, remaining))
+            backoff = min(backoff * _SETTLE_BACKOFF_MULTIPLIER, _SETTLE_MAX_BACKOFF_S)
+            selected = fetch_and_filter_placements(client, full_model_id, args)
 
     if not selected:
         logger.error("No valid placements matched your filters.")
@@ -575,6 +760,16 @@ def main() -> int:
     if args.dry_run:
         return 0
 
+    logger.info("Planning phase: checking downloads...")
+    run_planning_phase(
+        client,
+        full_model_id,
+        selected[0],
+        args.danger_delete_downloads,
+        args.timeout,
+        settle_deadline,
+    )
+
     all_rows: list[dict[str, Any]] = []
 
     for preview in selected:

bench/pyproject.toml

@@ -0,0 +1,16 @@
+[project]
+name = "exo-bench"
+version = "0.1.0"
+description = "Benchmarking tool for exo distributed inference"
+requires-python = ">=3.13"
+dependencies = [
+    "loguru>=0.7.3",
+    "transformers>=5.0.0",
+    "huggingface-hub>=0.33.4",
+    "tiktoken>=0.12.0",
+    "jinja2>=3.1.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"

dashboard/src/lib/components/ChatForm.svelte

@@ -12,6 +12,8 @@
     ttftMs,
     tps,
     totalTokens,
+    thinkingEnabled as thinkingEnabledStore,
+    setConversationThinking,
   } from "$lib/stores/app.svelte";
   import ChatAttachments from "./ChatAttachments.svelte";
   import ImageParamsPanel from "./ImageParamsPanel.svelte";
@@ -25,6 +27,7 @@
     autofocus?: boolean;
     showModelSelector?: boolean;
     modelTasks?: Record<string, string[]>;
+    modelCapabilities?: Record<string, string[]>;
   }
 
   let {
@@ -34,6 +37,7 @@
     autofocus = true,
     showModelSelector = false,
     modelTasks = {},
+    modelCapabilities = {},
   }: Props = $props();
 
   let message = $state("");
@@ -41,6 +45,7 @@
   let fileInputRef: HTMLInputElement | undefined = $state();
   let uploadedFiles = $state<ChatUploadedFile[]>([]);
   let isDragOver = $state(false);
+  const thinkingEnabled = $derived(thinkingEnabledStore());
   let loading = $derived(isLoading());
   const currentModel = $derived(selectedChatModel());
   const instanceData = $derived(instances());
@@ -95,6 +100,12 @@
     );
   });
 
+  const modelSupportsThinking = $derived(() => {
+    if (!currentModel) return false;
+    const caps = modelCapabilities[currentModel] || [];
+    return caps.includes("thinking") && caps.includes("text");
+  });
+
   const isEditOnlyWithoutImage = $derived(
     currentModel !== null &&
       modelSupportsOnlyImageEditing(currentModel) &&
@@ -254,6 +265,7 @@
 
   function handleSubmit() {
     if ((!message.trim() && uploadedFiles.length === 0) || loading) return;
+    if (isEditOnlyWithoutImage) return;
 
     const content = message.trim();
     const files = [...uploadedFiles];
@@ -278,11 +290,19 @@
       if (imageFile.preview) {
         editImage(content, imageFile.preview);
       }
-    } else if (isImageModel() && content) {
+    } else if (
+      currentModel &&
+      modelSupportsTextToImage(currentModel) &&
+      content
+    ) {
       // Use image generation for text-to-image models
       generateImage(content);
     } else {
-      sendMessage(content, files);
+      sendMessage(
+        content,
+        files,
+        modelSupportsThinking() ? thinkingEnabled : null,
+      );
     }
 
     // Refocus the textarea after sending
@@ -520,6 +540,35 @@
             </div>
           {/if}
         </div>
+        <!-- Thinking toggle -->
+        {#if modelSupportsThinking()}
+          <button
+            type="button"
+            onclick={() => setConversationThinking(!thinkingEnabled)}
+            class="flex items-center gap-1.5 px-2 py-1 rounded text-xs font-mono tracking-wide transition-all duration-200 flex-shrink-0 cursor-pointer border {thinkingEnabled
+              ? 'bg-exo-yellow/15 border-exo-yellow/40 text-exo-yellow'
+              : 'bg-exo-medium-gray/30 border-exo-medium-gray/50 text-exo-light-gray/60 hover:text-exo-light-gray'}"
+            title={thinkingEnabled
+              ? "Thinking enabled — click to disable"
+              : "Thinking disabled — click to enable"}
+          >
+            <svg
+              class="w-3.5 h-3.5"
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="1.5"
+            >
+              <path
+                d="M12 2a7 7 0 0 0-7 7c0 2.38 1.19 4.47 3 5.74V17a1 1 0 0 0 1 1h6a1 1 0 0 0 1-1v-2.26c1.81-1.27 3-3.36 3-5.74a7 7 0 0 0-7-7zM9 20h6M10 22h4"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
+            </svg>
+            <span>{thinkingEnabled ? "THINK" : "NO THINK"}</span>
+          </button>
+        {/if}
+
         <!-- Performance stats -->
         {#if currentTtft !== null || currentTps !== null}
           <div class="flex items-center gap-4 text-xs font-mono flex-shrink-0">
@@ -599,9 +648,8 @@
             : isImageModel()
               ? "Describe the image you want to generate..."
               : placeholder}
-        disabled={loading}
         rows={1}
-        class="flex-1 resize-none bg-transparent text-foreground placeholder:text-exo-light-gray/60 placeholder:text-sm placeholder:tracking-[0.15em] placeholder:leading-7 focus:outline-none focus:ring-0 focus:border-none disabled:opacity-50 text-sm leading-7 font-mono"
+        class="flex-1 resize-none bg-transparent text-foreground placeholder:text-exo-light-gray/60 placeholder:text-sm placeholder:tracking-[0.15em] placeholder:leading-7 focus:outline-none focus:ring-0 focus:border-none text-sm leading-7 font-mono"
         style="min-height: 28px; max-height: 150px;"
       ></textarea>
 

dashboard/src/lib/components/ChatMessages.svelte

@@ -13,6 +13,7 @@
   import type { MessageAttachment } from "$lib/stores/app.svelte";
   import MarkdownContent from "./MarkdownContent.svelte";
   import TokenHeatmap from "./TokenHeatmap.svelte";
+  import ImageLightbox from "./ImageLightbox.svelte";
 
   interface Props {
     class?: string;
@@ -101,6 +102,9 @@
   let copiedMessageId = $state<string | null>(null);
   let expandedThinkingMessageIds = $state<Set<string>>(new Set());
 
+  // Lightbox state
+  let expandedImageSrc = $state<string | null>(null);
+
   // Uncertainty heatmap toggle
   let heatmapMessageIds = $state<Set<string>>(new Set());
 
@@ -221,6 +225,7 @@
   }
 
   function handleDeleteClick(messageId: string) {
+    if (loading) return;
     deleteConfirmId = messageId;
   }
 
@@ -251,7 +256,7 @@
 </script>
 
 <div class="flex flex-col gap-4 sm:gap-6 {className}">
-  {#each messageList as message (message.id)}
+  {#each messageList as message, i (message.id)}
     <div
       class="group flex {message.role === 'user'
         ? 'justify-end'
@@ -313,9 +318,11 @@
           <!-- Delete confirmation -->
           <div class="bg-red-500/10 border border-red-500/30 rounded-lg p-3">
             <p class="text-xs text-red-400 mb-3">
-              Delete this message{message.role === "user"
-                ? " and all responses after it"
-                : ""}?
+              {#if i === messageList.length - 1}
+                Delete this message?
+              {:else}
+                Delete this message and all messages after it?
+              {/if}
             </p>
             <div class="flex gap-2 justify-end">
               <button
@@ -389,10 +396,15 @@
                         class="flex items-center gap-2 bg-exo-dark-gray/60 border border-exo-yellow/20 rounded px-2 py-1 text-xs font-mono"
                       >
                         {#if attachment.type === "image" && attachment.preview}
+                          <!-- svelte-ignore a11y_no_noninteractive_element_interactions, a11y_click_events_have_key_events -->
                           <img
                             src={attachment.preview}
                             alt={attachment.name}
-                            class="w-12 h-12 object-cover rounded border border-exo-yellow/20"
+                            class="w-12 h-12 object-cover rounded border border-exo-yellow/20 cursor-pointer hover:border-exo-yellow/50 transition-colors"
+                            onclick={() => {
+                              if (attachment.preview)
+                                expandedImageSrc = attachment.preview;
+                            }}
                           />
                         {:else}
                           <span>{getAttachmentIcon(attachment)}</span>
@@ -466,15 +478,44 @@
                   <div class="mb-3">
                     {#each message.attachments.filter((a) => a.type === "generated-image") as attachment}
                       <div class="relative group/img inline-block">
+                        <!-- svelte-ignore a11y_no_noninteractive_element_interactions, a11y_click_events_have_key_events -->
                         <img
                           src={attachment.preview}
                           alt=""
-                          class="max-w-full max-h-[512px] rounded-lg border border-exo-yellow/20 shadow-lg shadow-black/20"
+                          class="max-w-full max-h-[512px] rounded-lg border border-exo-yellow/20 shadow-lg shadow-black/20 cursor-pointer"
+                          onclick={() => {
+                            if (attachment.preview)
+                              expandedImageSrc = attachment.preview;
+                          }}
                         />
                         <!-- Button overlay -->
                         <div
                           class="absolute top-2 right-2 flex gap-1 opacity-0 group-hover/img:opacity-100 transition-opacity"
                         >
+                          <!-- Expand button -->
+                          <button
+                            type="button"
+                            class="p-2 rounded-lg bg-exo-dark-gray/80 border border-exo-yellow/30 text-exo-yellow hover:bg-exo-dark-gray hover:border-exo-yellow/50 cursor-pointer"
+                            onclick={() => {
+                              if (attachment.preview)
+                                expandedImageSrc = attachment.preview;
+                            }}
+                            title="Expand image"
+                          >
+                            <svg
+                              class="w-4 h-4"
+                              fill="none"
+                              viewBox="0 0 24 24"
+                              stroke="currentColor"
+                              stroke-width="2"
+                            >
+                              <path
+                                stroke-linecap="round"
+                                stroke-linejoin="round"
+                                d="M4 8V4m0 0h4M4 4l5 5m11-1V4m0 0h-4m4 0l-5 5M4 16v4m0 0h4m-4 0l5-5m11 5l-5-5m5 5v-4m0 4h-4"
+                              />
+                            </svg>
+                          </button>
                           <!-- Edit button -->
                           <button
                             type="button"
@@ -713,8 +754,13 @@
             <!-- Delete button -->
             <button
               onclick={() => handleDeleteClick(message.id)}
-              class="p-1.5 text-exo-light-gray hover:text-red-400 transition-colors rounded hover:bg-red-500/10 cursor-pointer"
-              title="Delete message"
+              disabled={loading}
+              class="p-1.5 transition-colors rounded {loading
+                ? 'text-exo-light-gray/30 cursor-not-allowed'
+                : 'text-exo-light-gray hover:text-red-400 hover:bg-red-500/10 cursor-pointer'}"
+              title={loading
+                ? "Cannot delete while generating"
+                : "Delete message"}
             >
               <svg
                 class="w-3.5 h-3.5"
@@ -789,3 +835,8 @@
     </button>
   {/if}
 </div>
+
+<ImageLightbox
+  src={expandedImageSrc}
+  onclose={() => (expandedImageSrc = null)}
+/>

dashboard/src/lib/components/ChatSidebar.svelte

@@ -185,11 +185,7 @@
 
     let instanceType: string | null = null;
     if (instanceTag === "MlxRingInstance") instanceType = "MLX Ring";
-    else if (
-      instanceTag === "MlxIbvInstance" ||
-      instanceTag === "MlxJacclInstance"
-    )
-      instanceType = "MLX RDMA";
+    else if (instanceTag === "MlxJacclInstance") instanceType = "MLX RDMA";
 
     let sharding: string | null = null;
     const inst = instance as {

dashboard/src/lib/components/FamilyLogos.svelte

@@ -13,6 +13,12 @@
       d="M12 2l3.09 6.26L22 9.27l-5 4.87 1.18 6.88L12 17.77l-6.18 3.25L7 14.14 2 9.27l6.91-1.01L12 2z"
     />
   </svg>
+{:else if family === "recents"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M13 3a9 9 0 0 0-9 9H1l3.89 3.89.07.14L9 12H6c0-3.87 3.13-7 7-7s7 3.13 7 7-3.13 7-7 7c-1.93 0-3.68-.79-4.94-2.06l-1.42 1.42A8.954 8.954 0 0 0 13 21a9 9 0 0 0 0-18zm-1 5v5l4.28 2.54.72-1.21-3.5-2.08V8H12z"
+    />
+  </svg>
 {:else if family === "llama" || family === "meta"}
   <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
     <path
@@ -58,6 +64,18 @@
       d="M17.962 1.844h-4.326l-3.425 7.81H5.369V1.878H1.5V22h3.87v-8.477h6.824a3.025 3.025 0 002.743-1.75V22h3.87v-8.477a3.87 3.87 0 00-3.588-3.86v-.01h-2.125a3.94 3.94 0 002.323-2.12l2.545-5.689z"
     />
   </svg>
+{:else if family === "flux"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M12 2L2 19h7.5l2.5-4.5L14.5 19H22L12 2zm0 4.5L16.5 17h-3l-1.5-2.7L10.5 17h-3L12 6.5z"
+    />
+  </svg>
+{:else if family === "qwen-image"}
+  <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
+    <path
+      d="M12.604 1.34c.393.69.784 1.382 1.174 2.075a.18.18 0 00.157.091h5.552c.174 0 .322.11.446.327l1.454 2.57c.19.337.24.478.024.837-.26.43-.513.864-.76 1.3l-.367.658c-.106.196-.223.28-.04.512l2.652 4.637c.172.301.111.494-.043.77-.437.785-.882 1.564-1.335 2.34-.159.272-.352.375-.68.37-.777-.016-1.552-.01-2.327.016a.099.099 0 00-.081.05 575.097 575.097 0 01-2.705 4.74c-.169.293-.38.363-.725.364-.997.003-2.002.004-3.017.002a.537.537 0 01-.465-.271l-1.335-2.323a.09.09 0 00-.083-.049H4.982c-.285.03-.553-.001-.805-.092l-1.603-2.77a.543.543 0 01-.002-.54l1.207-2.12a.198.198 0 000-.197 550.951 550.951 0 01-1.875-3.272l-.79-1.395c-.16-.31-.173-.496.095-.965.465-.813.927-1.625 1.387-2.436.132-.234.304-.334.584-.335a338.3 338.3 0 012.589-.001.124.124 0 00.107-.063l2.806-4.895a.488.488 0 01.422-.246c.524-.001 1.053 0 1.583-.006L11.704 1c.341-.003.724.032.9.34zm-3.432.403a.06.06 0 00-.052.03L6.254 6.788a.157.157 0 01-.135.078H3.253c-.056 0-.07.025-.041.074l5.81 10.156c.025.042.013.062-.034.063l-2.795.015a.218.218 0 00-.2.116l-1.32 2.31c-.044.078-.021.118.068.118l5.716.008c.046 0 .08.02.104.061l1.403 2.454c.046.081.092.082.139 0l5.006-8.76.783-1.382a.055.055 0 01.096 0l1.424 2.53a.122.122 0 00.107.062l2.763-.02a.04.04 0 00.035-.02.041.041 0 000-.04l-2.9-5.086a.108.108 0 010-.113l.293-.507 1.12-1.977c.024-.041.012-.062-.035-.062H9.2c-.059 0-.073-.026-.043-.077l1.434-2.505a.107.107 0 000-.114L9.225 1.774a.06.06 0 00-.053-.031zm6.29 8.02c.046 0 .058.02.034.06l-.832 1.465-2.613 4.585a.056.056 0 01-.05.029.058.058 0 01-.05-.029L8.498 9.841c-.02-.034-.01-.052.028-.054l.216-.012 6.722-.012z"
+    />
+  </svg>
 {:else if family === "huggingface"}
   <svg class="w-6 h-6 {className}" viewBox="0 0 24 24" fill="currentColor">
     <path

dashboard/src/lib/components/FamilySidebar.svelte

@@ -5,15 +5,22 @@
     families: string[];
     selectedFamily: string | null;
     hasFavorites: boolean;
+    hasRecents: boolean;
     onSelect: (family: string | null) => void;
   };
 
-  let { families, selectedFamily, hasFavorites, onSelect }: FamilySidebarProps =
-    $props();
+  let {
+    families,
+    selectedFamily,
+    hasFavorites,
+    hasRecents,
+    onSelect,
+  }: FamilySidebarProps = $props();
 
   // Family display names
   const familyNames: Record<string, string> = {
     favorites: "Favorites",
+    recents: "Recent",
     huggingface: "Hub",
     llama: "Meta",
     qwen: "Qwen",
@@ -22,6 +29,8 @@
     glm: "GLM",
     minimax: "MiniMax",
     kimi: "Kimi",
+    flux: "FLUX",
+    "qwen-image": "Qwen Img",
   };
 
   function getFamilyName(family: string): string {
@@ -32,7 +41,7 @@
 </script>
 
 <div
-  class="flex flex-col gap-1 py-2 px-1 border-r border-exo-yellow/10 bg-exo-medium-gray/30 min-w-[64px]"
+  class="flex flex-col gap-1 py-2 px-1 border-r border-exo-yellow/10 bg-exo-medium-gray/30 min-w-[64px] overflow-y-auto scrollbar-hide"
 >
   <!-- All models (no filter) -->
   <button
@@ -87,6 +96,31 @@
     </button>
   {/if}
 
+  <!-- Recent (only show if has recent models) -->
+  {#if hasRecents}
+    <button
+      type="button"
+      onclick={() => onSelect("recents")}
+      class="group flex flex-col items-center justify-center p-2 rounded transition-all duration-200 cursor-pointer {selectedFamily ===
+      'recents'
+        ? 'bg-exo-yellow/20 border-l-2 border-exo-yellow'
+        : 'hover:bg-white/5 border-l-2 border-transparent'}"
+      title="Recently launched models"
+    >
+      <FamilyLogos
+        family="recents"
+        class={selectedFamily === "recents"
+          ? "text-exo-yellow"
+          : "text-white/50 group-hover:text-white/70"}
+      />
+      <span
+        class="text-[9px] font-mono mt-0.5 {selectedFamily === 'recents'
+          ? 'text-exo-yellow'
+          : 'text-white/40 group-hover:text-white/60'}">Recent</span
+      >
+    </button>
+  {/if}
+
   <!-- HuggingFace Hub -->
   <button
     type="button"

dashboard/src/lib/components/ImageLightbox.svelte

@@ -0,0 +1,96 @@
+<script lang="ts">
+  import { fade, fly } from "svelte/transition";
+  import { cubicOut } from "svelte/easing";
+
+  interface Props {
+    src: string | null;
+    onclose: () => void;
+  }
+
+  let { src, onclose }: Props = $props();
+
+  function handleKeydown(e: KeyboardEvent) {
+    if (e.key === "Escape") {
+      onclose();
+    }
+  }
+
+  function extensionFromSrc(dataSrc: string): string {
+    const match = dataSrc.match(/^data:image\/(\w+)/);
+    if (match) return match[1] === "jpeg" ? "jpg" : match[1];
+    const urlMatch = dataSrc.match(/\.(\w+)(?:\?|$)/);
+    if (urlMatch) return urlMatch[1];
+    return "png";
+  }
+
+  function handleDownload(e: MouseEvent) {
+    e.stopPropagation();
+    if (!src) return;
+    const link = document.createElement("a");
+    link.href = src;
+    link.download = `image-${Date.now()}.${extensionFromSrc(src)}`;
+    link.click();
+  }
+
+  function handleClose(e: MouseEvent) {
+    e.stopPropagation();
+    onclose();
+  }
+</script>
+
+<svelte:window onkeydown={src ? handleKeydown : undefined} />
+
+{#if src}
+  <div
+    class="fixed inset-0 z-50 bg-black/90 backdrop-blur-sm flex items-center justify-center"
+    transition:fade={{ duration: 200 }}
+    onclick={onclose}
+    role="presentation"
+    onintrostart={() => (document.body.style.overflow = "hidden")}
+    onoutroend={() => (document.body.style.overflow = "")}
+  >
+    <div class="absolute top-4 right-4 flex gap-2 z-10">
+      <button
+        type="button"
+        class="p-2 rounded-lg bg-exo-dark-gray/80 border border-exo-yellow/30 text-exo-yellow hover:bg-exo-dark-gray hover:border-exo-yellow/50 cursor-pointer transition-colors"
+        onclick={handleDownload}
+        title="Download image"
+      >
+        <svg
+          class="w-5 h-5"
+          fill="none"
+          viewBox="0 0 24 24"
+          stroke="currentColor"
+          stroke-width="2"
+        >
+          <path
+            stroke-linecap="round"
+            stroke-linejoin="round"
+            d="M4 16v1a3 3 0 003 3h10a3 3 0 003-3v-1m-4-4l-4 4m0 0l-4-4m4 4V4"
+          />
+        </svg>
+      </button>
+      <button
+        type="button"
+        class="p-2 rounded-lg bg-exo-dark-gray/80 border border-exo-yellow/30 text-exo-yellow hover:bg-exo-dark-gray hover:border-exo-yellow/50 cursor-pointer transition-colors"
+        onclick={handleClose}
+        title="Close"
+      >
+        <svg class="w-5 h-5" viewBox="0 0 24 24" fill="currentColor">
+          <path
+            d="M19 6.41L17.59 5 12 10.59 6.41 5 5 6.41 10.59 12 5 17.59 6.41 19 12 13.41 17.59 19 19 17.59 13.41 12 19 6.41z"
+          />
+        </svg>
+      </button>
+    </div>
+
+    <!-- svelte-ignore a11y_no_noninteractive_element_interactions, a11y_click_events_have_key_events -->
+    <img
+      {src}
+      alt=""
+      class="max-w-[90vw] max-h-[90vh] object-contain rounded-lg shadow-2xl"
+      transition:fly={{ y: 20, duration: 300, easing: cubicOut }}
+      onclick={(e) => e.stopPropagation()}
+    />
+  </div>
+{/if}

dashboard/src/lib/components/ModelCard.svelte

@@ -21,7 +21,7 @@
     } | null;
     nodes?: Record<string, NodeInfo>;
     sharding?: "Pipeline" | "Tensor";
-    runtime?: "MlxRing" | "MlxIbv" | "MlxJaccl";
+    runtime?: "MlxRing" | "MlxJaccl";
     onLaunch?: () => void;
     tags?: string[];
     apiPreview?: PlacementPreview | null;
@@ -348,7 +348,7 @@
   // Debug mode state
   const isDebugMode = $derived(debugMode());
   const topology = $derived(topologyData());
-  const isRdma = $derived(runtime === "MlxIbv" || runtime === "MlxJaccl");
+  const isRdma = $derived(runtime === "MlxJaccl");
 
   // Get interface name for an IP from node data
   function getInterfaceForIp(nodeId: string, ip?: string): string | null {
@@ -422,9 +422,16 @@
     const bToACandidates: Array<{ ip: string; iface: string | null }> = [];
 
     for (const edge of topology.edges) {
-      const ip = edge.sendBackIp || "?";
-      const iface =
-        edge.sendBackInterface || getInterfaceForIp(edge.source, ip);
+      let ip: string;
+      let iface: string | null;
+
+      if (edge.sourceRdmaIface || edge.sinkRdmaIface) {
+        ip = "RDMA";
+        iface = `${edge.sourceRdmaIface || "?"} \u2192 ${edge.sinkRdmaIface || "?"}`;
+      } else {
+        ip = edge.sendBackIp || "?";
+        iface = edge.sendBackInterface || getInterfaceForIp(edge.source, ip);
+      }
 
       if (edge.source === nodeId1 && edge.target === nodeId2) {
         aToBCandidates.push({ ip, iface });
@@ -568,7 +575,7 @@
       >
         {runtime === "MlxRing"
           ? "MLX Ring"
-          : runtime === "MlxIbv" || runtime === "MlxJaccl"
+          : runtime === "MlxJaccl"
             ? "MLX RDMA"
             : runtime}
       </span>

dashboard/src/lib/components/ModelFilterPopover.svelte

@@ -24,6 +24,8 @@
     { id: "thinking", label: "Thinking" },
     { id: "code", label: "Code" },
     { id: "vision", label: "Vision" },
+    { id: "image_gen", label: "Image Gen" },
+    { id: "image_edit", label: "Image Edit" },
   ];
 
   // Size ranges
@@ -142,6 +144,45 @@
                   stroke-linejoin="round"
                 /><circle cx="12" cy="12" r="3" /></svg
               >
+            {:else if cap.id === "image_gen"}
+              <svg
+                class="w-3.5 h-3.5 inline-block"
+                viewBox="0 0 24 24"
+                fill="none"
+                stroke="currentColor"
+                stroke-width="1.5"
+                ><rect
+                  x="3"
+                  y="3"
+                  width="18"
+                  height="18"
+                  rx="2"
+                  ry="2"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /><circle cx="8.5" cy="8.5" r="1.5" /><path
+                  d="M21 15l-5-5L5 21"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /></svg
+              >
+            {:else if cap.id === "image_edit"}
+              <svg
+                class="w-3.5 h-3.5 inline-block"
+                viewBox="0 0 24 24"
+                fill="none"
+                stroke="currentColor"
+                stroke-width="1.5"
+                ><path
+                  d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /><path
+                  d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z"
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                /></svg
+              >
             {/if}
             <span class="ml-1">{cap.label}</span>
           </button>

dashboard/src/lib/components/ModelPickerGroup.svelte

@@ -26,6 +26,7 @@
     nodeNames: string[];
     nodeIds: string[];
   };
+  type ModelFitStatus = "fits_now" | "fits_cluster_capacity" | "too_large";
 
   type ModelPickerGroupProps = {
     group: ModelGroup;
@@ -33,11 +34,13 @@
     isFavorite: boolean;
     selectedModelId: string | null;
     canModelFit: (id: string) => boolean;
+    getModelFitStatus: (id: string) => ModelFitStatus;
     onToggleExpand: () => void;
     onSelectModel: (modelId: string) => void;
     onToggleFavorite: (baseModelId: string) => void;
     onShowInfo: (group: ModelGroup) => void;
     downloadStatusMap?: Map<string, DownloadAvailability>;
+    launchedAt?: number;
   };
 
   let {
@@ -46,11 +49,13 @@
     isFavorite,
     selectedModelId,
     canModelFit,
+    getModelFitStatus,
     onToggleExpand,
     onSelectModel,
     onToggleFavorite,
     onShowInfo,
     downloadStatusMap,
+    launchedAt,
   }: ModelPickerGroupProps = $props();
 
   // Group-level download status: show if any variant is downloaded
@@ -72,10 +77,45 @@
     return `${mb}MB`;
   }
 
+  function timeAgo(ts: number): string {
+    const seconds = Math.floor((Date.now() - ts) / 1000);
+    if (seconds < 60) return "just now";
+    const minutes = Math.floor(seconds / 60);
+    if (minutes < 60) return `${minutes}m ago`;
+    const hours = Math.floor(minutes / 60);
+    if (hours < 24) return `${hours}h ago`;
+    const days = Math.floor(hours / 24);
+    return `${days}d ago`;
+  }
+
   // Check if any variant can fit
   const anyVariantFits = $derived(
     group.variants.some((v) => canModelFit(v.id)),
   );
+  const groupFitStatus = $derived.by((): ModelFitStatus => {
+    let hasClusterCapacityOnly = false;
+    for (const variant of group.variants) {
+      const fitStatus = getModelFitStatus(variant.id);
+      if (fitStatus === "fits_now") {
+        return "fits_now";
+      }
+      if (fitStatus === "fits_cluster_capacity") {
+        hasClusterCapacityOnly = true;
+      }
+    }
+    return hasClusterCapacityOnly ? "fits_cluster_capacity" : "too_large";
+  });
+
+  function getSizeClassForFitStatus(fitStatus: ModelFitStatus): string {
+    switch (fitStatus) {
+      case "fits_now":
+        return "text-white/40";
+      case "fits_cluster_capacity":
+        return "text-orange-400/80";
+      case "too_large":
+        return "text-red-400/70";
+    }
+  }
 
   // Check if this group's model is currently selected (for single-variant groups)
   const isMainSelected = $derived(
@@ -200,9 +240,42 @@
               stroke-width="1.5"
               title="Supports image generation"
             >
-              <rect x="3" y="3" width="18" height="18" rx="2" ry="2" />
+              <rect
+                x="3"
+                y="3"
+                width="18"
+                height="18"
+                rx="2"
+                ry="2"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
               <circle cx="8.5" cy="8.5" r="1.5" />
-              <path d="M21 15l-5-5L5 21" />
+              <path
+                d="M21 15l-5-5L5 21"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
+            </svg>
+          {:else if cap === "image_edit"}
+            <svg
+              class="w-3.5 h-3.5 text-white/40 flex-shrink-0"
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="1.5"
+              title="Supports image editing"
+            >
+              <path
+                d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
+              <path
+                d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z"
+                stroke-linecap="round"
+                stroke-linejoin="round"
+              />
             </svg>
           {/if}
         {/each}
@@ -211,7 +284,14 @@
 
     <!-- Size indicator (smallest variant) -->
     {#if !group.hasMultipleVariants && group.smallestVariant?.storage_size_megabytes}
-      <span class="text-xs font-mono text-white/30 flex-shrink-0">
+      {@const singleVariantFitStatus = getModelFitStatus(
+        group.smallestVariant.id,
+      )}
+      <span
+        class="text-xs font-mono flex-shrink-0 {getSizeClassForFitStatus(
+          singleVariantFitStatus,
+        )}"
+      >
         {formatSize(group.smallestVariant.storage_size_megabytes)}
       </span>
     {/if}
@@ -222,13 +302,24 @@
         .map((v) => v.storage_size_megabytes || 0)
         .filter((s) => s > 0)
         .sort((a, b) => a - b)}
-      <span class="text-xs font-mono text-white/30 flex-shrink-0">
+      <span
+        class="text-xs font-mono flex-shrink-0 {getSizeClassForFitStatus(
+          groupFitStatus,
+        )}"
+      >
         {group.variants.length} variants{#if sizes.length >= 2}{" "}({formatSize(
             sizes[0],
           )}-{formatSize(sizes[sizes.length - 1])}){/if}
       </span>
     {/if}
 
+    <!-- Time ago (for recent models) -->
+    {#if launchedAt}
+      <span class="text-xs font-mono text-white/20 flex-shrink-0">
+        {timeAgo(launchedAt)}
+      </span>
+    {/if}
+
     <!-- Download availability indicator -->
     {#if groupDownloadStatus && groupDownloadStatus.nodeIds.length > 0}
       <span
@@ -327,6 +418,7 @@
   {#if isExpanded && group.hasMultipleVariants}
     <div class="bg-black/20 border-t border-white/5">
       {#each group.variants as variant}
+        {@const fitStatus = getModelFitStatus(variant.id)}
         {@const modelCanFit = canModelFit(variant.id)}
         {@const isSelected = selectedModelId === variant.id}
         <button
@@ -351,7 +443,11 @@
           </span>
 
           <!-- Size -->
-          <span class="text-xs font-mono text-white/40 flex-1">
+          <span
+            class="text-xs font-mono flex-1 {getSizeClassForFitStatus(
+              fitStatus,
+            )}"
+          >
             {formatSize(variant.storage_size_megabytes)}
           </span>
 

dashboard/src/lib/components/ModelPickerModal.svelte

@@ -6,6 +6,7 @@
   import ModelFilterPopover from "./ModelFilterPopover.svelte";
   import HuggingFaceResultItem from "./HuggingFaceResultItem.svelte";
   import { getNodesWithModelDownloaded } from "$lib/utils/downloads";
+  import { getRecentEntries } from "$lib/stores/recents.svelte";
 
   interface ModelInfo {
     id: string;
@@ -46,13 +47,18 @@
     tags: string[];
   }
 
+  type ModelFitStatus = "fits_now" | "fits_cluster_capacity" | "too_large";
+
   type ModelPickerModalProps = {
     isOpen: boolean;
     models: ModelInfo[];
     selectedModelId: string | null;
     favorites: Set<string>;
+    recentModelIds?: string[];
+    hasRecents?: boolean;
     existingModelIds: Set<string>;
     canModelFit: (modelId: string) => boolean;
+    getModelFitStatus: (modelId: string) => ModelFitStatus;
     onSelect: (modelId: string) => void;
     onClose: () => void;
     onToggleFavorite: (baseModelId: string) => void;
@@ -76,8 +82,11 @@
     models,
     selectedModelId,
     favorites,
+    recentModelIds = [],
+    hasRecents: hasRecentsTab = false,
     existingModelIds,
     canModelFit,
+    getModelFitStatus,
     onSelect,
     onClose,
     onToggleFavorite,
@@ -363,6 +372,8 @@
       "deepseek",
       "gpt-oss",
       "llama",
+      "flux",
+      "qwen-image",
     ];
     return Array.from(families).sort((a, b) => {
       const aIdx = familyOrder.indexOf(a);
@@ -381,7 +392,11 @@
     // Filter by family
     if (selectedFamily === "favorites") {
       result = result.filter((g) => favorites.has(g.id));
-    } else if (selectedFamily && selectedFamily !== "huggingface") {
+    } else if (
+      selectedFamily &&
+      selectedFamily !== "huggingface" &&
+      selectedFamily !== "recents"
+    ) {
       result = result.filter((g) => g.family === selectedFamily);
     }
 
@@ -425,13 +440,23 @@
       );
     }
 
-    // Sort: models that fit first, then by size (largest first)
+    // Sort: fits-now first, then fits-cluster-capacity, then too-large
     result.sort((a, b) => {
-      const aFits = a.variants.some((v) => canModelFit(v.id));
-      const bFits = b.variants.some((v) => canModelFit(v.id));
+      const getGroupFitRank = (group: ModelGroup): number => {
+        let hasClusterCapacityOnly = false;
+        for (const variant of group.variants) {
+          const fitStatus = getModelFitStatus(variant.id);
+          if (fitStatus === "fits_now") return 0;
+          if (fitStatus === "fits_cluster_capacity") {
+            hasClusterCapacityOnly = true;
+          }
+        }
+        return hasClusterCapacityOnly ? 1 : 2;
+      };
 
-      if (aFits && !bFits) return -1;
-      if (!aFits && bFits) return 1;
+      const aRank = getGroupFitRank(a);
+      const bRank = getGroupFitRank(b);
+      if (aRank !== bRank) return aRank - bRank;
 
       return (
         (b.smallestVariant.storage_size_megabytes || 0) -
@@ -445,6 +470,48 @@
   // Check if any favorites exist
   const hasFavorites = $derived(favorites.size > 0);
 
+  // Timestamp lookup for recent models
+  const recentTimestamps = $derived(
+    new Map(getRecentEntries().map((e) => [e.modelId, e.launchedAt])),
+  );
+
+  // Recent models: single-variant ModelGroups in launch order
+  const recentGroups = $derived.by((): ModelGroup[] => {
+    if (!recentModelIds || recentModelIds.length === 0) return [];
+    const result: ModelGroup[] = [];
+    for (const id of recentModelIds) {
+      const model = models.find((m) => m.id === id);
+      if (model) {
+        result.push({
+          id: model.base_model || model.id,
+          name: model.name || model.id,
+          capabilities: model.capabilities || ["text"],
+          family: model.family || "",
+          variants: [model],
+          smallestVariant: model,
+          hasMultipleVariants: false,
+        });
+      }
+    }
+    return result;
+  });
+
+  // Filtered recent groups (apply search query)
+  const filteredRecentGroups = $derived.by((): ModelGroup[] => {
+    if (!searchQuery.trim()) return recentGroups;
+    const query = searchQuery.toLowerCase().trim();
+    return recentGroups.filter(
+      (g) =>
+        g.name.toLowerCase().includes(query) ||
+        g.variants.some(
+          (v) =>
+            v.id.toLowerCase().includes(query) ||
+            (v.name || "").toLowerCase().includes(query) ||
+            (v.quantization || "").toLowerCase().includes(query),
+        ),
+    );
+  });
+
   function toggleGroupExpanded(groupId: string) {
     const next = new Set(expandedGroups);
     if (next.has(groupId)) {
@@ -602,11 +669,12 @@
         families={uniqueFamilies}
         {selectedFamily}
         {hasFavorites}
+        hasRecents={hasRecentsTab}
         onSelect={(family) => (selectedFamily = family)}
       />
 
       <!-- Model list -->
-      <div class="flex-1 overflow-y-auto flex flex-col">
+      <div class="flex-1 overflow-y-auto scrollbar-hide flex flex-col">
         {#if selectedFamily === "huggingface"}
           <!-- HuggingFace Hub view -->
           <div class="flex-1 flex flex-col min-h-0">
@@ -624,7 +692,7 @@
             </div>
 
             <!-- Results list -->
-            <div class="flex-1 overflow-y-auto">
+            <div class="flex-1 overflow-y-auto scrollbar-hide">
               {#if hfIsLoadingTrending && hfTrendingModels.length === 0}
                 <div
                   class="flex items-center justify-center py-12 text-white/40"
@@ -709,6 +777,45 @@
               </div>
             </div>
           </div>
+        {:else if selectedFamily === "recents"}
+          <!-- Recent models view -->
+          {#if filteredRecentGroups.length === 0}
+            <div
+              class="flex flex-col items-center justify-center h-full text-white/40 p-8"
+            >
+              <svg
+                class="w-12 h-12 mb-3"
+                viewBox="0 0 24 24"
+                fill="currentColor"
+              >
+                <path
+                  d="M13 3a9 9 0 0 0-9 9H1l3.89 3.89.07.14L9 12H6c0-3.87 3.13-7 7-7s7 3.13 7 7-3.13 7-7 7c-1.93 0-3.68-.79-4.94-2.06l-1.42 1.42A8.954 8.954 0 0 0 13 21a9 9 0 0 0 0-18zm-1 5v5l4.28 2.54.72-1.21-3.5-2.08V8H12z"
+                />
+              </svg>
+              <p class="font-mono text-sm">
+                {searchQuery
+                  ? "No matching recent models"
+                  : "No recently launched models"}
+              </p>
+            </div>
+          {:else}
+            {#each filteredRecentGroups as group}
+              <ModelPickerGroup
+                {group}
+                isExpanded={expandedGroups.has(group.id)}
+                isFavorite={favorites.has(group.id)}
+                {selectedModelId}
+                {canModelFit}
+                {getModelFitStatus}
+                onToggleExpand={() => toggleGroupExpanded(group.id)}
+                onSelectModel={handleSelect}
+                {onToggleFavorite}
+                onShowInfo={(g) => (infoGroup = g)}
+                downloadStatusMap={getVariantDownloadMap(group)}
+                launchedAt={recentTimestamps.get(group.variants[0]?.id ?? "")}
+              />
+            {/each}
+          {/if}
         {:else if filteredGroups.length === 0}
           <div
             class="flex flex-col items-center justify-center h-full text-white/40 p-8"
@@ -740,6 +847,7 @@
               isFavorite={favorites.has(group.id)}
               {selectedModelId}
               {canModelFit}
+              {getModelFitStatus}
               onToggleExpand={() => toggleGroupExpanded(group.id)}
               onSelectModel={handleSelect}
               {onToggleFavorite}

dashboard/src/lib/components/TopologyGraph.svelte

@@ -6,6 +6,8 @@
     isTopologyMinimized,
     debugMode,
     nodeThunderboltBridge,
+    nodeRdmaCtl,
+    nodeIdentities,
     type NodeInfo,
   } from "$lib/stores/app.svelte";
 
@@ -31,6 +33,8 @@
   const data = $derived(topologyData());
   const debugEnabled = $derived(debugMode());
   const tbBridgeData = $derived(nodeThunderboltBridge());
+  const rdmaCtlData = $derived(nodeRdmaCtl());
+  const identitiesData = $derived(nodeIdentities());
 
   function getNodeLabel(nodeId: string): string {
     const node = data?.nodes?.[nodeId];
@@ -333,14 +337,27 @@
       if (edge.source === a) entry.aToB = true;
       else entry.bToA = true;
 
-      const ip = edge.sendBackIp || "?";
-      const ifaceInfo = getInterfaceLabel(edge.source, ip);
+      let ip: string;
+      let ifaceLabel: string;
+      let missingIface: boolean;
+
+      if (edge.sourceRdmaIface || edge.sinkRdmaIface) {
+        ip = "RDMA";
+        ifaceLabel = `${edge.sourceRdmaIface || "?"} \u2192 ${edge.sinkRdmaIface || "?"}`;
+        missingIface = false;
+      } else {
+        ip = edge.sendBackIp || "?";
+        const ifaceInfo = getInterfaceLabel(edge.source, ip);
+        ifaceLabel = ifaceInfo.label;
+        missingIface = ifaceInfo.missing;
+      }
+
       entry.connections.push({
         from: edge.source,
         to: edge.target,
         ip,
-        ifaceLabel: ifaceInfo.label,
-        missingIface: ifaceInfo.missing,
+        ifaceLabel,
+        missingIface,
       });
       pairMap.set(key, entry);
     });
@@ -1120,15 +1137,17 @@
           .text(` (${ramUsagePercent.toFixed(0)}%)`);
       }
 
-      // Debug mode: Show TB bridge status
+      // Debug mode: Show TB bridge and RDMA status
       if (debugEnabled) {
+        let debugLabelY =
+          nodeInfo.y +
+          iconBaseHeight / 2 +
+          (showFullLabels ? 32 : showCompactLabels ? 26 : 22);
+        const debugFontSize = showFullLabels ? 9 : 7;
+        const debugLineHeight = showFullLabels ? 11 : 9;
+
         const tbStatus = tbBridgeData[nodeInfo.id];
         if (tbStatus) {
-          const tbY =
-            nodeInfo.y +
-            iconBaseHeight / 2 +
-            (showFullLabels ? 32 : showCompactLabels ? 26 : 22);
-          const tbFontSize = showFullLabels ? 9 : 7;
           const tbColor = tbStatus.enabled
             ? "rgba(234,179,8,0.9)"
             : "rgba(100,100,100,0.7)";
@@ -1136,12 +1155,46 @@
           nodeG
             .append("text")
             .attr("x", nodeInfo.x)
-            .attr("y", tbY)
+            .attr("y", debugLabelY)
             .attr("text-anchor", "middle")
             .attr("fill", tbColor)
-            .attr("font-size", tbFontSize)
+            .attr("font-size", debugFontSize)
             .attr("font-family", "SF Mono, Monaco, monospace")
             .text(tbText);
+          debugLabelY += debugLineHeight;
+        }
+
+        const rdmaStatus = rdmaCtlData[nodeInfo.id];
+        if (rdmaStatus !== undefined) {
+          const rdmaColor = rdmaStatus.enabled
+            ? "rgba(74,222,128,0.9)"
+            : "rgba(100,100,100,0.7)";
+          const rdmaText = rdmaStatus.enabled ? "RDMA:ON" : "RDMA:OFF";
+          nodeG
+            .append("text")
+            .attr("x", nodeInfo.x)
+            .attr("y", debugLabelY)
+            .attr("text-anchor", "middle")
+            .attr("fill", rdmaColor)
+            .attr("font-size", debugFontSize)
+            .attr("font-family", "SF Mono, Monaco, monospace")
+            .text(rdmaText);
+          debugLabelY += debugLineHeight;
+        }
+
+        const identity = identitiesData[nodeInfo.id];
+        if (identity?.osVersion) {
+          nodeG
+            .append("text")
+            .attr("x", nodeInfo.x)
+            .attr("y", debugLabelY)
+            .attr("text-anchor", "middle")
+            .attr("fill", "rgba(179,179,179,0.7)")
+            .attr("font-size", debugFontSize)
+            .attr("font-family", "SF Mono, Monaco, monospace")
+            .text(
+              `macOS ${identity.osVersion}${identity.osBuildVersion ? ` (${identity.osBuildVersion})` : ""}`,
+            );
         }
       }
     });

dashboard/src/lib/stores/app.svelte.ts

@@ -49,6 +49,7 @@ export interface NodeInfo {
   };
   last_macmon_update: number;
   friendly_name?: string;
+  os_version?: string;
 }
 
 export interface TopologyEdge {
@@ -56,6 +57,8 @@ export interface TopologyEdge {
   target: string;
   sendBackIp?: string;
   sendBackInterface?: string;
+  sourceRdmaIface?: string;
+  sinkRdmaIface?: string;
 }
 
 export interface TopologyData {
@@ -76,6 +79,8 @@ interface RawNodeIdentity {
   modelId?: string;
   chipId?: string;
   friendlyName?: string;
+  osVersion?: string;
+  osBuildVersion?: string;
 }
 
 interface RawMemoryUsage {
@@ -163,7 +168,7 @@ export interface ModelDownloadStatus {
 export interface PlacementPreview {
   model_id: string;
   sharding: "Pipeline" | "Tensor";
-  instance_meta: "MlxRing" | "MlxIbv" | "MlxJaccl";
+  instance_meta: "MlxRing" | "MlxJaccl";
   instance: unknown | null;
   memory_delta_by_node: Record<string, number> | null;
   error: string | null;
@@ -214,7 +219,6 @@ interface RawStateResponse {
     string,
     {
       MlxRingInstance?: Instance;
-      MlxIbvInstance?: Instance;
       MlxJacclInstance?: Instance;
     }
   >;
@@ -225,6 +229,19 @@ interface RawStateResponse {
   nodeMemory?: Record<string, RawMemoryUsage>;
   nodeSystem?: Record<string, RawSystemPerformanceProfile>;
   nodeNetwork?: Record<string, RawNodeNetworkInfo>;
+  // Thunderbolt identifiers per node
+  nodeThunderbolt?: Record<
+    string,
+    {
+      interfaces: Array<{
+        rdmaInterface: string;
+        domainUuid: string;
+        linkSpeed: string;
+      }>;
+    }
+  >;
+  // RDMA ctl status per node
+  nodeRdmaCtl?: Record<string, { enabled: boolean }>;
   // Thunderbolt bridge status per node
   nodeThunderboltBridge?: Record<
     string,
@@ -232,6 +249,20 @@ interface RawStateResponse {
   >;
   // Thunderbolt bridge cycles (nodes with bridge enabled forming loops)
   thunderboltBridgeCycles?: string[][];
+  // MetaInstances (declarative instance constraints)
+  metaInstances?: Record<string, MetaInstanceData>;
+}
+
+export interface MetaInstanceData {
+  metaInstanceId: string;
+  modelId: string;
+  sharding: string;
+  instanceMeta: string;
+  minNodes: number;
+  nodeIds: string[] | null;
+  placementError: string | null;
+  consecutiveFailures: number;
+  lastFailureError: string | null;
 }
 
 export interface MessageAttachment {
@@ -278,6 +309,7 @@ export interface Conversation {
   modelId: string | null;
   sharding: string | null;
   instanceType: string | null;
+  enableThinking: boolean | null;
 }
 
 const STORAGE_KEY = "exo-conversations";
@@ -425,6 +457,7 @@ function transformTopology(
       },
       last_macmon_update: Date.now() / 1000,
       friendly_name: identity?.friendlyName,
+      os_version: identity?.osVersion,
     };
   }
 
@@ -437,6 +470,8 @@ function transformTopology(
         if (!Array.isArray(edgeList)) continue;
         for (const edge of edgeList) {
           let sendBackIp: string | undefined;
+          let sourceRdmaIface: string | undefined;
+          let sinkRdmaIface: string | undefined;
           if (edge && typeof edge === "object" && "sinkMultiaddr" in edge) {
             const multiaddr = edge.sinkMultiaddr;
             if (multiaddr) {
@@ -444,10 +479,23 @@ function transformTopology(
                 multiaddr.ip_address ||
                 extractIpFromMultiaddr(multiaddr.address);
             }
+          } else if (
+            edge &&
+            typeof edge === "object" &&
+            "sourceRdmaIface" in edge
+          ) {
+            sourceRdmaIface = edge.sourceRdmaIface;
+            sinkRdmaIface = edge.sinkRdmaIface;
           }
 
           if (nodes[source] && nodes[sink] && source !== sink) {
-            edges.push({ source, target: sink, sendBackIp });
+            edges.push({
+              source,
+              target: sink,
+              sendBackIp,
+              sourceRdmaIface,
+              sinkRdmaIface,
+            });
           }
         }
       }
@@ -490,12 +538,33 @@ class AppStore {
   instances = $state<Record<string, unknown>>({});
   runners = $state<Record<string, unknown>>({});
   downloads = $state<Record<string, unknown[]>>({});
+  nodeDisk = $state<
+    Record<
+      string,
+      { total: { inBytes: number }; available: { inBytes: number } }
+    >
+  >({});
   placementPreviews = $state<PlacementPreview[]>([]);
   selectedPreviewModelId = $state<string | null>(null);
   isLoadingPreviews = $state(false);
   previewNodeFilter = $state<Set<string>>(new Set());
   lastUpdate = $state<number | null>(null);
+  metaInstances = $state<Record<string, MetaInstanceData>>({});
+  nodeIdentities = $state<Record<string, RawNodeIdentity>>({});
   thunderboltBridgeCycles = $state<string[][]>([]);
+  nodeThunderbolt = $state<
+    Record<
+      string,
+      {
+        interfaces: Array<{
+          rdmaInterface: string;
+          domainUuid: string;
+          linkSpeed: string;
+        }>;
+      }
+    >
+  >({});
+  nodeRdmaCtl = $state<Record<string, { enabled: boolean }>>({});
   nodeThunderboltBridge = $state<
     Record<
       string,
@@ -551,6 +620,7 @@ class AppStore {
           modelId: conversation.modelId ?? null,
           sharding: conversation.sharding ?? null,
           instanceType: conversation.instanceType ?? null,
+          enableThinking: conversation.enableThinking ?? null,
         }));
       }
     } catch (error) {
@@ -740,6 +810,7 @@ class AppStore {
       modelId: derivedModelId,
       sharding: derivedSharding,
       instanceType: derivedInstanceType,
+      enableThinking: null,
     };
 
     this.conversations.unshift(conversation);
@@ -765,6 +836,7 @@ class AppStore {
     this.hasStartedChat = true;
     this.isTopologyMinimized = true;
     this.isSidebarOpen = true; // Auto-open sidebar when chatting
+    this.thinkingEnabled = conversation.enableThinking ?? true;
     this.refreshConversationModelFromInstances();
 
     return true;
@@ -837,11 +909,7 @@ class AppStore {
 
     let instanceType: string | null = null;
     if (instanceTag === "MlxRingInstance") instanceType = "MLX Ring";
-    else if (
-      instanceTag === "MlxIbvInstance" ||
-      instanceTag === "MlxJacclInstance"
-    )
-      instanceType = "MLX RDMA";
+    else if (instanceTag === "MlxJacclInstance") instanceType = "MLX RDMA";
 
     let sharding: string | null = null;
     const inst = instance as {
@@ -1206,6 +1274,17 @@ class AppStore {
       if (data.downloads) {
         this.downloads = data.downloads;
       }
+      // MetaInstances
+      this.metaInstances = data.metaInstances ?? {};
+      if (data.nodeDisk) {
+        this.nodeDisk = data.nodeDisk;
+      }
+      // Node identities (for OS version mismatch detection)
+      this.nodeIdentities = data.nodeIdentities ?? {};
+      // Thunderbolt identifiers per node
+      this.nodeThunderbolt = data.nodeThunderbolt ?? {};
+      // RDMA ctl status per node
+      this.nodeRdmaCtl = data.nodeRdmaCtl ?? {};
       // Thunderbolt bridge cycles
       this.thunderboltBridgeCycles = data.thunderboltBridgeCycles ?? [];
       // Thunderbolt bridge status per node
@@ -1869,6 +1948,11 @@ class AppStore {
     }
   }
 
+  /**
+   * Whether thinking is enabled for the current conversation
+   */
+  thinkingEnabled = $state(true);
+
   /**
    * Selected model for chat (can be set by the UI)
    */
@@ -2047,6 +2131,7 @@ class AppStore {
       textContent?: string;
       preview?: string;
     }[],
+    enableThinking?: boolean | null,
   ): Promise<void> {
     if ((!content.trim() && (!files || files.length === 0)) || this.isLoading)
       return;
@@ -2194,6 +2279,9 @@ class AppStore {
           stream: true,
           logprobs: true,
           top_logprobs: 5,
+          ...(enableThinking != null && {
+            enable_thinking: enableThinking,
+          }),
         }),
       });
 
@@ -2852,6 +2940,18 @@ class AppStore {
     );
   }
 
+  /**
+   * Update the thinking preference for the active conversation
+   */
+  setConversationThinking(enabled: boolean) {
+    this.thinkingEnabled = enabled;
+    const conv = this.getActiveConversation();
+    if (conv) {
+      conv.enableThinking = enabled;
+      this.saveConversationsToStorage();
+    }
+  }
+
   /**
    * Start a download on a specific node
    */
@@ -2956,14 +3056,17 @@ export const tps = () => appStore.tps;
 export const totalTokens = () => appStore.totalTokens;
 export const topologyData = () => appStore.topologyData;
 export const instances = () => appStore.instances;
+export const metaInstances = () => appStore.metaInstances;
 export const runners = () => appStore.runners;
 export const downloads = () => appStore.downloads;
+export const nodeDisk = () => appStore.nodeDisk;
 export const placementPreviews = () => appStore.placementPreviews;
 export const selectedPreviewModelId = () => appStore.selectedPreviewModelId;
 export const isLoadingPreviews = () => appStore.isLoadingPreviews;
 export const lastUpdate = () => appStore.lastUpdate;
 export const isTopologyMinimized = () => appStore.isTopologyMinimized;
 export const selectedChatModel = () => appStore.selectedChatModel;
+export const thinkingEnabled = () => appStore.thinkingEnabled;
 export const debugMode = () => appStore.getDebugMode();
 export const topologyOnlyMode = () => appStore.getTopologyOnlyMode();
 export const chatSidebarVisible = () => appStore.getChatSidebarVisible();
@@ -2979,7 +3082,8 @@ export const sendMessage = (
     textContent?: string;
     preview?: string;
   }[],
-) => appStore.sendMessage(content, files);
+  enableThinking?: boolean | null,
+) => appStore.sendMessage(content, files, enableThinking);
 export const generateImage = (prompt: string, modelId?: string) =>
   appStore.generateImage(prompt, modelId);
 export const editImage = (
@@ -3022,6 +3126,8 @@ export const deleteAllConversations = () => appStore.deleteAllConversations();
 export const renameConversation = (id: string, name: string) =>
   appStore.renameConversation(id, name);
 export const getActiveConversation = () => appStore.getActiveConversation();
+export const setConversationThinking = (enabled: boolean) =>
+  appStore.setConversationThinking(enabled);
 
 // Sidebar actions
 export const isSidebarOpen = () => appStore.isSidebarOpen;
@@ -3038,7 +3144,12 @@ export const setChatSidebarVisible = (visible: boolean) =>
   appStore.setChatSidebarVisible(visible);
 export const refreshState = () => appStore.fetchState();
 
-// Thunderbolt bridge status
+// Node identities (for OS version mismatch detection)
+export const nodeIdentities = () => appStore.nodeIdentities;
+
+// Thunderbolt & RDMA status
+export const nodeThunderbolt = () => appStore.nodeThunderbolt;
+export const nodeRdmaCtl = () => appStore.nodeRdmaCtl;
 export const thunderboltBridgeCycles = () => appStore.thunderboltBridgeCycles;
 export const nodeThunderboltBridge = () => appStore.nodeThunderboltBridge;
 

dashboard/src/lib/stores/recents.svelte.ts

@@ -0,0 +1,75 @@
+/**
+ * RecentsStore - Manages recently launched models with localStorage persistence
+ */
+
+import { browser } from "$app/environment";
+
+const RECENTS_KEY = "exo-recent-models";
+const MAX_RECENT_MODELS = 20;
+
+interface RecentEntry {
+  modelId: string;
+  launchedAt: number;
+}
+
+class RecentsStore {
+  recents = $state<RecentEntry[]>([]);
+
+  constructor() {
+    if (browser) {
+      this.loadFromStorage();
+    }
+  }
+
+  private loadFromStorage() {
+    try {
+      const stored = localStorage.getItem(RECENTS_KEY);
+      if (stored) {
+        const parsed = JSON.parse(stored) as RecentEntry[];
+        this.recents = parsed;
+      }
+    } catch (error) {
+      console.error("Failed to load recent models:", error);
+    }
+  }
+
+  private saveToStorage() {
+    try {
+      localStorage.setItem(RECENTS_KEY, JSON.stringify(this.recents));
+    } catch (error) {
+      console.error("Failed to save recent models:", error);
+    }
+  }
+
+  recordLaunch(modelId: string) {
+    // Remove existing entry for this model (if any) to move it to top
+    const filtered = this.recents.filter((r) => r.modelId !== modelId);
+    // Prepend new entry
+    const next = [{ modelId, launchedAt: Date.now() }, ...filtered];
+    // Cap at max
+    this.recents = next.slice(0, MAX_RECENT_MODELS);
+    this.saveToStorage();
+  }
+
+  getRecentModelIds(): string[] {
+    return this.recents.map((r) => r.modelId);
+  }
+
+  hasAny(): boolean {
+    return this.recents.length > 0;
+  }
+
+  clearAll() {
+    this.recents = [];
+    this.saveToStorage();
+  }
+}
+
+export const recentsStore = new RecentsStore();
+
+export const hasRecents = () => recentsStore.hasAny();
+export const getRecentModelIds = () => recentsStore.getRecentModelIds();
+export const getRecentEntries = () => recentsStore.recents;
+export const recordRecentLaunch = (modelId: string) =>
+  recentsStore.recordLaunch(modelId);
+export const clearRecents = () => recentsStore.clearAll();

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

@@ -3,6 +3,7 @@
   import {
     topologyData,
     downloads,
+    nodeDisk,
     type DownloadProgress,
     refreshState,
     lastUpdate as lastUpdateStore,
@@ -37,10 +38,13 @@
     nodeId: string;
     nodeName: string;
     models: ModelEntry[];
+    diskAvailable?: number;
+    diskTotal?: number;
   };
 
   const data = $derived(topologyData());
   const downloadsData = $derived(downloads());
+  const nodeDiskData = $derived(nodeDisk());
 
   function getNodeLabel(nodeId: string): string {
     const node = data?.nodes?.[nodeId];
@@ -327,10 +331,17 @@
           ];
         }
 
+        // Get disk info for this node
+        const diskInfo = nodeDiskData?.[nodeId];
+        const diskAvailable = diskInfo?.available?.inBytes;
+        const diskTotal = diskInfo?.total?.inBytes;
+
         built.push({
           nodeId,
           nodeName: getNodeLabel(nodeId),
           models,
+          diskAvailable,
+          diskTotal,
         });
       }
 
@@ -417,6 +428,14 @@
                 <div class="text-xs text-exo-light-gray font-mono truncate">
                   {node.nodeId}
                 </div>
+                <div class="text-xs text-exo-light-gray font-mono mt-1">
+                  {formatBytes(
+                    node.models
+                      .filter((m) => m.status === "completed")
+                      .reduce((sum, m) => sum + m.totalBytes, 0),
+                  )} models{#if node.diskAvailable != null}
+                    - {formatBytes(node.diskAvailable)} free{/if}
+                </div>
               </div>
               <div
                 class="text-xs font-mono uppercase tracking-wider whitespace-nowrap shrink-0 text-right"
@@ -429,13 +448,6 @@
                     / {node.models.length} models</span
                   >
                 </div>
-                <div class="text-exo-light-gray normal-case tracking-normal">
-                  {formatBytes(
-                    node.models
-                      .filter((m) => m.status === "completed")
-                      .reduce((sum, m) => sum + m.totalBytes, 0),
-                  )} on disk
-                </div>
               </div>
             </div>
 

flake.nix

@@ -83,6 +83,9 @@
           _module.args.pkgs = import inputs.nixpkgs {
             inherit system;
             config.allowUnfreePredicate = pkg: (pkg.pname or "") == "metal-toolchain";
+            overlays = [
+              (import ./nix/apple-sdk-overlay.nix)
+            ];
           };
           treefmt = {
             projectRootFile = "flake.nix";

nix/apple-sdk-overlay.nix

@@ -0,0 +1,18 @@
+# Overlay that builds apple-sdk with a custom versions.json (for SDK 26.2).
+# The upstream nixpkgs package reads versions.json at eval time via a relative
+# path, so we can't override it through callPackage args. Instead, we copy
+# the upstream source and patch the one file.
+final: _prev:
+let
+  upstreamSrc = final.path + "/pkgs/by-name/ap/apple-sdk";
+  patchedSrc = final.runCommandLocal "apple-sdk-src-patched" { } ''
+    cp -r ${upstreamSrc} $out
+    chmod -R u+w $out
+    cp ${./apple-sdk/metadata/versions.json} $out/metadata/versions.json
+  '';
+in
+{
+  apple-sdk_26 = final.callPackage (patchedSrc + "/package.nix") {
+    darwinSdkMajorVersion = "26";
+  };
+}

nix/apple-sdk/metadata/versions.json

@@ -0,0 +1,26 @@
+{
+  "14": {
+    "urls": [
+      "https://swcdn.apple.com/content/downloads/14/48/052-59890-A_I0F5YGAY0Y/p9n40hio7892gou31o1v031ng6fnm9sb3c/CLTools_macOSNMOS_SDK.pkg",
+      "https://web.archive.org/web/20250211001355/https://swcdn.apple.com/content/downloads/14/48/052-59890-A_I0F5YGAY0Y/p9n40hio7892gou31o1v031ng6fnm9sb3c/CLTools_macOSNMOS_SDK.pkg"
+    ],
+    "version": "14.4",
+    "hash": "sha256-QozDiwY0Czc0g45vPD7G4v4Ra+3DujCJbSads3fJjjM="
+  },
+  "15": {
+    "urls": [
+      "https://swcdn.apple.com/content/downloads/52/01/082-41241-A_0747ZN8FHV/dectd075r63pppkkzsb75qk61s0lfee22j/CLTools_macOSNMOS_SDK.pkg",
+      "https://web.archive.org/web/20250530132510/https://swcdn.apple.com/content/downloads/52/01/082-41241-A_0747ZN8FHV/dectd075r63pppkkzsb75qk61s0lfee22j/CLTools_macOSNMOS_SDK.pkg"
+    ],
+    "version": "15.5",
+    "hash": "sha256-HBiSJuw1XBUK5R/8Sj65c3rftSEvQl/O9ZZVp/g1Amo="
+  },
+  "26": {
+    "urls": [
+      "https://swcdn.apple.com/content/downloads/60/22/089-71960-A_W8BL1RUJJ6/5zkyplomhk1cm7z6xja2ktgapnhhti6wwd/CLTools_macOSNMOS_SDK.pkg",
+      "https://web.archive.org/web/20250915230423/https://swcdn.apple.com/content/downloads/60/22/089-71960-A_W8BL1RUJJ6/5zkyplomhk1cm7z6xja2ktgapnhhti6wwd/CLTools_macOSNMOS_SDK.pkg"
+    ],
+    "version": "26.2",
+    "hash": "sha256-hXRlMieVv0smna5uiWRwq87IWOaPWtAjAldbi+wQXcw="
+  }
+}

nix/mlx.nix

@@ -41,7 +41,7 @@ let
 
   mlx = stdenv.mkDerivation rec {
     pname = "mlx";
-    version = let v = "0.30.5"; in
+    version = let v = "0.30.6"; in
       assert v == uvLockMlxVersion || throw "MLX version mismatch: nix/mlx.nix has ${v} but uv.lock has ${uvLockMlxVersion}. Update both the version and hash in nix/mlx.nix.";
       v;
     pyproject = true;
@@ -50,7 +50,7 @@ let
       owner = "ml-explore";
       repo = "mlx";
       tag = "v${version}";
-      hash = "sha256-OJk6jPlbaSlsUdk3ADz3tWcRzTWXRof3/q8Soe1AO6w=";
+      hash = "sha256-avD5EGhwgmPdXLAyQSqTO6AXk/W3ziH+f6AetjK3Sdo=";
     };
 
     patches = [
@@ -85,6 +85,7 @@ let
         (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_CPU" 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}")

pyproject.toml

@@ -17,8 +17,8 @@ dependencies = [
     "loguru>=0.7.3",
     "exo_pyo3_bindings", # rust bindings
     "anyio==4.11.0",
-    "mlx==0.30.5; sys_platform == 'darwin'",
-    "mlx[cpu]==0.30.5; sys_platform == 'linux'",
+    "mlx==0.30.6; sys_platform == 'darwin'",
+    "mlx[cpu]==0.30.6; sys_platform == 'linux'",
     "mlx-lm==0.30.6",
     "tiktoken>=0.12.0", # required for kimi k2 tokenizer
     "hypercorn>=0.18.0",
@@ -28,6 +28,8 @@ dependencies = [
     "pillow>=11.0,<12.0", # compatibility with mflux
     "mflux==0.15.5",
     "python-multipart>=0.0.21",
+    "msgspec>=0.19.0",
+    "zstandard>=0.23.0",
 ]
 
 [project.scripts]
@@ -57,6 +59,7 @@ dev = [
 [tool.uv.workspace]
 members = [
     "rust/exo_pyo3_bindings",
+    "bench",
 ]
 
 [tool.uv.sources]
@@ -129,7 +132,7 @@ markers = [
 env = [
   "EXO_TESTS=1"
 ]
-addopts = "-m 'not slow'"
+addopts = "-m 'not slow' --ignore=tests/start_distributed_test.py"
 filterwarnings = [
     "ignore:builtin type Swig:DeprecationWarning",
 ]

python/parts.nix

@@ -14,7 +14,9 @@
 
       # Override overlay to inject Nix-built components
       exoOverlay = final: prev: {
-        # Replace workspace exo_pyo3_bindings with Nix-built wheel
+        # Replace workspace exo_pyo3_bindings with Nix-built wheel.
+        # Preserve passthru so mkVirtualEnv can resolve dependency groups.
+        # Copy .pyi stub + py.typed marker so basedpyright can find the types.
         exo-pyo3-bindings = pkgs.stdenv.mkDerivation {
           pname = "exo-pyo3-bindings";
           version = "0.1.0";
@@ -22,6 +24,12 @@
           # Install from pre-built wheel
           nativeBuildInputs = [ final.pyprojectWheelHook ];
           dontStrip = true;
+          passthru = prev.exo-pyo3-bindings.passthru or { };
+          postInstall = ''
+            local siteDir=$out/${final.python.sitePackages}/exo_pyo3_bindings
+            cp ${inputs.self}/rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi $siteDir/
+            touch $siteDir/py.typed
+          '';
         };
       };
 
@@ -29,17 +37,32 @@
 
       # 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
           ];
         });
+      } // lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin {
+        # Use our pure Nix-built MLX with Metal support (macOS only)
+        mlx = self'.packages.mlx;
       };
 
+      # Additional overlay for Linux-specific fixes (type checking env).
+      # Native wheels have shared lib dependencies we don't need at type-check time.
+      linuxOverlay = final: prev:
+        let
+          ignoreMissing = drv: drv.overrideAttrs { autoPatchelfIgnoreMissingDeps = [ "*" ]; };
+          nvidiaPackages = lib.filterAttrs (name: _: lib.hasPrefix "nvidia-" name) prev;
+        in
+        lib.optionalAttrs pkgs.stdenv.hostPlatform.isLinux (
+          (lib.mapAttrs (_: ignoreMissing) nvidiaPackages) // {
+            mlx = ignoreMissing prev.mlx;
+            torch = ignoreMissing prev.torch;
+            triton = ignoreMissing prev.triton;
+          }
+        );
+
       pythonSet = (pkgs.callPackage inputs.pyproject-nix.build.packages {
         inherit python;
       }).overrideScope (
@@ -48,6 +71,7 @@
           overlay
           exoOverlay
           buildSystemsOverlay
+          linuxOverlay
         ]
       );
       exoVenv = pythonSet.mkVirtualEnv "exo-env" workspace.deps.default;
@@ -69,6 +93,16 @@
         text = ''exec python ${path} "$@"'';
       };
 
+      benchVenv = pythonSet.mkVirtualEnv "exo-bench-env" {
+        exo-bench = [ ];
+      };
+
+      mkBenchScript = name: path: pkgs.writeShellApplication {
+        inherit name;
+        runtimeInputs = [ benchVenv ];
+        text = ''exec python ${path} "$@"'';
+      };
+
       mkSimplePythonScript = name: path: pkgs.writeShellApplication {
         inherit name;
         runtimeInputs = [ pkgs.python313 ];
@@ -96,8 +130,8 @@
           exo = exoPackage;
           # Test environment for running pytest outside of Nix sandbox (needs GPU access)
           exo-test-env = testVenv;
-          exo-bench = mkPythonScript "exo-bench" (inputs.self + /bench/exo_bench.py);
         } // {
+        exo-bench = mkBenchScript "exo-bench" (inputs.self + /bench/exo_bench.py);
         exo-get-all-models-on-cluster = mkSimplePythonScript "exo-get-all-models-on-cluster" (inputs.self + /tests/get_all_models_on_cluster.py);
       };
 
@@ -108,6 +142,21 @@
           ${pkgs.ruff}/bin/ruff check ${inputs.self}
           touch $out
         '';
+
+        # Hermetic basedpyright type checking
+        typecheck = pkgs.runCommand "typecheck"
+          {
+            nativeBuildInputs = [
+              testVenv
+              pkgs.basedpyright
+            ];
+          }
+          ''
+            cd ${inputs.self}
+            export HOME=$TMPDIR
+            basedpyright --pythonpath ${testVenv}/bin/python
+            touch $out
+          '';
       };
     };
 }

resources/image_model_cards/exolabs--FLUX.1-Kontext-dev-4bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
+family = "flux"
+quantization = "4bit"
+base_model = "FLUX.1 Kontext"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 15475325472

resources/image_model_cards/exolabs--FLUX.1-Kontext-dev-8bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
+family = "flux"
+quantization = "8bit"
+base_model = "FLUX.1 Kontext"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 21426029632

resources/image_model_cards/exolabs--FLUX.1-Kontext-dev.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
+family = "flux"
+quantization = ""
+base_model = "FLUX.1 Kontext"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 33327437952

resources/image_model_cards/exolabs--FLUX.1-Krea-dev-4bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "4bit"
+base_model = "FLUX.1 Krea"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 15475325472

resources/image_model_cards/exolabs--FLUX.1-Krea-dev-8bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "8bit"
+base_model = "FLUX.1 Krea"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 21426029632

resources/image_model_cards/exolabs--FLUX.1-Krea-dev.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = ""
+base_model = "FLUX.1 Krea"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 33327437952

resources/image_model_cards/exolabs--FLUX.1-dev-4bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "4bit"
+base_model = "FLUX.1 Dev"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 15475325472

resources/image_model_cards/exolabs--FLUX.1-dev-8bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "8bit"
+base_model = "FLUX.1 Dev"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 21426029632

resources/image_model_cards/exolabs--FLUX.1-dev.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = ""
+base_model = "FLUX.1 Dev"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 33327437952

resources/image_model_cards/exolabs--FLUX.1-schnell-4bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "4bit"
+base_model = "FLUX.1 Schnell"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 15470210592

resources/image_model_cards/exolabs--FLUX.1-schnell-8bit.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = "8bit"
+base_model = "FLUX.1 Schnell"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 21415799872

resources/image_model_cards/exolabs--FLUX.1-schnell.toml

@@ -3,6 +3,10 @@ n_layers = 57
 hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
+family = "flux"
+quantization = ""
+base_model = "FLUX.1 Schnell"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 33306978432

resources/image_model_cards/exolabs--Qwen-Image-4bit.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = "4bit"
+base_model = "Qwen Image"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 26799533856

resources/image_model_cards/exolabs--Qwen-Image-8bit.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = "8bit"
+base_model = "Qwen Image"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 37014734400

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509-4bit.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = "4bit"
+base_model = "Qwen Image Edit"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 26799533856

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509-8bit.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = "8bit"
+base_model = "Qwen Image Edit"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 37014734400

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["ImageToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = ""
+base_model = "Qwen Image Edit"
+capabilities = ["image_edit"]
 
 [storage_size]
 in_bytes = 57445135488

resources/image_model_cards/exolabs--Qwen-Image.toml

@@ -4,6 +4,10 @@ hidden_size = 1
 supports_tensor = false
 tasks = ["TextToImage"]
 uses_cfg = true
+family = "qwen-image"
+quantization = ""
+base_model = "Qwen Image"
+capabilities = ["image_gen"]
 
 [storage_size]
 in_bytes = 57445135488

resources/inference_model_cards/mlx-community--Step-3.5-Flash-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Step-3.5-Flash-4bit"
+n_layers = 45
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "step"
+quantization = "4bit"
+base_model = "Step 3.5 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 114572190076

resources/inference_model_cards/mlx-community--Step-3.5-Flash-6bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Step-3.5-Flash-6bit"
+n_layers = 45
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "step"
+quantization = "6bit"
+base_model = "Step 3.5 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 159039627774

resources/inference_model_cards/mlx-community--Step-3.5-Flash-8Bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Step-3.5-Flash-8Bit"
+n_layers = 45
+hidden_size = 4096
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "step"
+quantization = "8bit"
+base_model = "Step 3.5 Flash"
+capabilities = ["text", "thinking"]
+
+[storage_size]
+in_bytes = 209082699847

rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi

@@ -130,7 +130,7 @@ class Multiaddr:
 
 @typing.final
 class NetworkingHandle:
-    def __new__(cls, identity: Keypair) -> NetworkingHandle: ...
+    def __new__(cls, identity: Keypair, bootstrap_peers: builtins.list[builtins.str] = ...) -> NetworkingHandle: ...
     async def connection_update_recv(self) -> ConnectionUpdate:
         r"""
         Receives the next `ConnectionUpdate` from networking.

rust/exo_pyo3_bindings/src/networking.rs

@@ -22,7 +22,6 @@ use pyo3::{Bound, Py, PyErr, PyResult, PyTraverseError, PyVisit, Python, pymetho
 use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pyclass_enum, gen_stub_pymethods};
 use std::net::IpAddr;
 use tokio::sync::{Mutex, mpsc, oneshot};
-use util::ext::VecExt as _;
 
 mod exception {
     use pyo3::types::PyTuple;
@@ -343,8 +342,10 @@ impl PyNetworkingHandle {
     // ---- Lifecycle management methods ----
 
     #[new]
-    fn py_new(identity: Bound<'_, PyKeypair>) -> PyResult<Self> {
+    #[pyo3(signature = (identity, bootstrap_peers=vec![]))]
+    fn py_new(identity: Bound<'_, PyKeypair>, bootstrap_peers: Vec<String>) -> PyResult<Self> {
         use pyo3_async_runtimes::tokio::get_runtime;
+        use std::str::FromStr;
 
         // create communication channels
         let (to_task_tx, to_task_rx) = mpsc::channel(MPSC_CHANNEL_SIZE);
@@ -354,9 +355,20 @@ impl PyNetworkingHandle {
         // get identity
         let identity = identity.borrow().0.clone();
 
+        // parse bootstrap peer multiaddrs
+        let parsed_peers: Vec<libp2p::Multiaddr> = bootstrap_peers
+            .into_iter()
+            .map(|s| libp2p::Multiaddr::from_str(&s))
+            .collect::<Result<_, _>>()
+            .pyerr()?;
+
+        if !parsed_peers.is_empty() {
+            log::info!("RUST: bootstrap peers: {:?}", parsed_peers);
+        }
+
         // create networking swarm (within tokio context!! or it crashes)
         let swarm = get_runtime()
-            .block_on(async { create_swarm(identity) })
+            .block_on(async { create_swarm(identity, parsed_peers) })
             .pyerr()?;
 
         // spawn tokio task running the networking logic
@@ -532,7 +544,9 @@ impl PyNetworkingHandle {
             .recv_many_py(limit)
             .allow_threads_py() // allow-threads-aware async call
             .await?
-            .map(|(t, d)| (t, d.pybytes())))
+            .into_iter()
+            .map(|(t, d)| (t, d.pybytes()))
+            .collect())
     }
 
     // TODO: rn this blocks main thread if anything else is awaiting the channel (bc its a mutex)

rust/networking/src/discovery.rs

@@ -106,6 +106,9 @@ pub struct Behaviour {
     managed: managed::Behaviour,
     mdns_discovered: HashMap<PeerId, BTreeSet<Multiaddr>>,
 
+    /// Addresses provided via --bootstrap-peer, re-dialed on the retry loop.
+    bootstrap_addrs: Vec<Multiaddr>,
+
     retry_delay: Delay, // retry interval
 
     // pending events to emmit => waker-backed Deque to control polling
@@ -113,13 +116,21 @@ pub struct Behaviour {
 }
 
 impl Behaviour {
-    pub fn new(keypair: &identity::Keypair) -> io::Result<Self> {
-        Ok(Self {
+    pub fn new(keypair: &identity::Keypair, bootstrap_peers: Vec<Multiaddr>) -> io::Result<Self> {
+        let mut behaviour = Self {
             managed: managed::Behaviour::new(keypair)?,
             mdns_discovered: HashMap::new(),
+            bootstrap_addrs: bootstrap_peers,
             retry_delay: Delay::new(RETRY_CONNECT_INTERVAL),
             pending_events: WakerDeque::new(),
-        })
+        };
+
+        // Immediately dial all bootstrap peers
+        for addr in &behaviour.bootstrap_addrs.clone() {
+            behaviour.dial_addr(addr.clone());
+        }
+
+        Ok(behaviour)
     }
 
     fn dial(&mut self, peer_id: PeerId, addr: Multiaddr) {
@@ -128,6 +139,14 @@ impl Behaviour {
         })
     }
 
+    /// Dial by address only — PeerId is resolved via Noise handshake on connect.
+    /// Used for bootstrap peers where only IP:port is known.
+    fn dial_addr(&mut self, addr: Multiaddr) {
+        self.pending_events.push_back(ToSwarm::Dial {
+            opts: DialOpts::unknown_peer_id().address(addr).build(),
+        })
+    }
+
     fn close_connection(&mut self, peer_id: PeerId, connection: ConnectionId) {
         // push front to make this IMMEDIATE
         self.pending_events.push_front(ToSwarm::CloseConnection {
@@ -362,13 +381,16 @@ impl NetworkBehaviour for Behaviour {
             Poll::Pending => {}
         }
 
-        // retry connecting to all mDNS peers periodically (fails safely if already connected)
+        // retry connecting to all mDNS + bootstrap peers periodically (fails safely if already connected)
         if self.retry_delay.poll_unpin(cx).is_ready() {
             for (p, mas) in self.mdns_discovered.clone() {
                 for ma in mas {
                     self.dial(p, ma)
                 }
             }
+            for addr in self.bootstrap_addrs.clone() {
+                self.dial_addr(addr)
+            }
             self.retry_delay.reset(RETRY_CONNECT_INTERVAL) // reset timeout
         }
 

rust/networking/src/swarm.rs

@@ -1,7 +1,7 @@
 use crate::alias;
 use crate::swarm::transport::tcp_transport;
 pub use behaviour::{Behaviour, BehaviourEvent};
-use libp2p::{SwarmBuilder, identity};
+use libp2p::{Multiaddr, SwarmBuilder, identity};
 
 pub type Swarm = libp2p::Swarm<Behaviour>;
 
@@ -16,11 +16,14 @@ pub const NETWORK_VERSION: &[u8] = b"v0.0.1";
 pub const OVERRIDE_VERSION_ENV_VAR: &str = "EXO_LIBP2P_NAMESPACE";
 
 /// Create and configure a swarm which listens to all ports on OS
-pub fn create_swarm(keypair: identity::Keypair) -> alias::AnyResult<Swarm> {
+pub fn create_swarm(
+    keypair: identity::Keypair,
+    bootstrap_peers: Vec<Multiaddr>,
+) -> alias::AnyResult<Swarm> {
     let mut swarm = SwarmBuilder::with_existing_identity(keypair)
         .with_tokio()
         .with_other_transport(tcp_transport)?
-        .with_behaviour(Behaviour::new)?
+        .with_behaviour(|kp| Behaviour::new(kp, bootstrap_peers))?
         .build();
 
     // Listen on all interfaces and whatever port the OS assigns
@@ -105,8 +108,7 @@ mod transport {
 mod behaviour {
     use crate::{alias, discovery};
     use libp2p::swarm::NetworkBehaviour;
-    use libp2p::{gossipsub, identity};
-    use std::time::Duration;
+    use libp2p::{Multiaddr, gossipsub, identity};
 
     /// Behavior of the Swarm which composes all desired behaviors:
     /// Right now its just [`discovery::Behaviour`] and [`gossipsub::Behaviour`].
@@ -117,9 +119,12 @@ mod behaviour {
     }
 
     impl Behaviour {
-        pub fn new(keypair: &identity::Keypair) -> alias::AnyResult<Self> {
+        pub fn new(
+            keypair: &identity::Keypair,
+            bootstrap_peers: Vec<Multiaddr>,
+        ) -> alias::AnyResult<Self> {
             Ok(Self {
-                discovery: discovery::Behaviour::new(keypair)?,
+                discovery: discovery::Behaviour::new(keypair, bootstrap_peers)?,
                 gossipsub: gossipsub_behaviour(keypair),
             })
         }
@@ -134,7 +139,6 @@ mod behaviour {
         gossipsub::Behaviour::new(
             MessageAuthenticity::Signed(keypair.clone()),
             ConfigBuilder::default()
-                .publish_queue_duration(Duration::from_secs(15))
                 .max_transmit_size(1024 * 1024)
                 .validation_mode(ValidationMode::Strict)
                 .build()

rust/util/Cargo.toml

@@ -13,13 +13,3 @@ path = "src/lib.rs"
 workspace = true
 
 [dependencies]
-# macro dependencies
-extend = { workspace = true }
-
-# utility dependencies
-thiserror = { workspace = true }
-once_cell = { workspace = true }
-internment = { workspace = true }
-derive_more = { workspace = true }
-bon = { workspace = true }
-recursion = { workspace = true }

rust/util/src/lib.rs

@@ -1,53 +1 @@
-//! TODO: crate documentation
-//!
-//! this is here as a placeholder documentation
-//!
-//!
-
-// enable Rust-unstable features for convenience
-#![feature(trait_alias)]
-#![feature(stmt_expr_attributes)]
-#![feature(type_alias_impl_trait)]
-#![feature(specialization)]
-#![feature(unboxed_closures)]
-#![feature(const_trait_impl)]
-#![feature(fn_traits)]
-
-pub mod nonempty;
 pub mod wakerdeque;
-
-pub(crate) mod private {
-    // sealed traits support
-    pub trait Sealed {}
-    impl<T: ?Sized> Sealed for T {}
-}
-
-/// Namespace for all the type/trait aliases used by this crate.
-pub(crate) mod alias {}
-
-/// Namespace for crate-wide extension traits/methods
-pub mod ext {
-    use extend::ext;
-
-    #[ext(pub, name = BoxedSliceExt)]
-    impl<T> Box<[T]> {
-        #[inline]
-        fn map<B, F>(self, f: F) -> Box<[B]>
-        where
-            F: FnMut(T) -> B,
-        {
-            self.into_iter().map(f).collect()
-        }
-    }
-
-    #[ext(pub, name = VecExt)]
-    impl<T> Vec<T> {
-        #[inline]
-        fn map<B, F>(self, f: F) -> Vec<B>
-        where
-            F: FnMut(T) -> B,
-        {
-            self.into_iter().map(f).collect()
-        }
-    }
-}

rust/util/src/nonempty.rs

@@ -1,138 +0,0 @@
-use std::slice::SliceIndex;
-use std::{ops, slice};
-use thiserror::Error;
-
-#[derive(Error, Debug)]
-#[error("Cannot create to `NonemptyArray` because the supplied slice is empty")]
-pub struct EmptySliceError;
-
-/// A pointer to a non-empty fixed-size slice allocated on the heap.
-#[derive(Clone, Debug, PartialEq, Eq, Hash, PartialOrd, Ord)]
-#[repr(transparent)]
-pub struct NonemptyArray<T>(Box<[T]>);
-
-#[allow(clippy::arbitrary_source_item_ordering)]
-impl<T> NonemptyArray<T> {
-    #[inline]
-    pub fn singleton(value: T) -> Self {
-        Self(Box::new([value]))
-    }
-
-    #[allow(clippy::missing_errors_doc)]
-    #[inline]
-    pub fn try_from_boxed_slice<S: Into<Box<[T]>>>(
-        boxed_slice: S,
-    ) -> Result<Self, EmptySliceError> {
-        let boxed_slice = boxed_slice.into();
-        if boxed_slice.is_empty() {
-            Err(EmptySliceError)
-        } else {
-            Ok(Self(boxed_slice))
-        }
-    }
-
-    #[must_use]
-    #[inline]
-    pub fn into_boxed_slice(self) -> Box<[T]> {
-        self.0
-    }
-
-    #[must_use]
-    #[inline]
-    pub fn to_vec(&self) -> Vec<T>
-    where
-        T: Clone,
-    {
-        self.0.to_vec()
-    }
-
-    #[must_use]
-    #[inline]
-    pub const fn as_slice(&self) -> &[T] {
-        &self.0
-    }
-
-    #[allow(clippy::indexing_slicing)]
-    #[must_use]
-    #[inline]
-    pub fn first(&self) -> &T {
-        &self.0[0]
-    }
-
-    #[allow(clippy::indexing_slicing, clippy::arithmetic_side_effects)]
-    #[must_use]
-    #[inline]
-    pub fn last(&self) -> &T {
-        &self.0[self.0.len() - 1]
-    }
-
-    #[must_use]
-    #[inline]
-    pub fn get<I>(&self, index: I) -> Option<&I::Output>
-    where
-        I: SliceIndex<[T]>,
-    {
-        self.0.get(index)
-    }
-
-    #[allow(clippy::len_without_is_empty)]
-    #[must_use]
-    #[inline]
-    pub const fn len(&self) -> usize {
-        self.0.len()
-    }
-
-    #[allow(clippy::iter_without_into_iter)]
-    #[inline]
-    pub fn iter(&self) -> slice::Iter<'_, T> {
-        self.0.iter()
-    }
-
-    #[allow(clippy::iter_without_into_iter)]
-    #[inline]
-    pub fn iter_mut(&mut self) -> slice::IterMut<'_, T> {
-        self.0.iter_mut()
-    }
-
-    #[inline]
-    #[must_use]
-    pub fn map<U, F: FnMut(T) -> U>(self, f: F) -> NonemptyArray<U> {
-        NonemptyArray(self.0.into_iter().map(f).collect())
-    }
-}
-
-impl<T> From<NonemptyArray<T>> for Box<[T]> {
-    #[inline]
-    fn from(value: NonemptyArray<T>) -> Self {
-        value.into_boxed_slice()
-    }
-}
-
-impl<T> ops::Index<usize> for NonemptyArray<T> {
-    type Output = T;
-
-    #[inline]
-    fn index(&self, index: usize) -> &Self::Output {
-        self.0.index(index)
-    }
-}
-
-impl<T> IntoIterator for NonemptyArray<T> {
-    type Item = T;
-    type IntoIter = std::vec::IntoIter<T>;
-
-    #[inline]
-    fn into_iter(self) -> Self::IntoIter {
-        self.into_boxed_slice().into_vec().into_iter()
-    }
-}
-
-impl<'a, T> IntoIterator for &'a NonemptyArray<T> {
-    type Item = &'a T;
-    type IntoIter = slice::Iter<'a, T>;
-
-    #[inline]
-    fn into_iter(self) -> Self::IntoIter {
-        self.iter()
-    }
-}

src/exo/download/coordinator.py

@@ -56,8 +56,49 @@ class DownloadCoordinator:
     event_receiver: Receiver[Event] = field(init=False)
     _tg: TaskGroup = field(init=False, default_factory=anyio.create_task_group)
 
+    # Per-model throttle for download progress events
+    _last_progress_time: dict[ModelId, float] = field(default_factory=dict)
+
     def __post_init__(self) -> None:
         self.event_sender, self.event_receiver = channel[Event]()
+        self.shard_downloader.on_progress(self._download_progress_callback)
+
+    async def _download_progress_callback(
+        self, callback_shard: ShardMetadata, progress: RepoDownloadProgress
+    ) -> None:
+        model_id = callback_shard.model_card.model_id
+        throttle_interval_secs = 1.0
+
+        if progress.status == "complete":
+            completed = DownloadCompleted(
+                shard_metadata=callback_shard,
+                node_id=self.node_id,
+                total_bytes=progress.total_bytes,
+            )
+            self.download_status[model_id] = completed
+            await self.event_sender.send(
+                NodeDownloadProgress(download_progress=completed)
+            )
+            if model_id in self.active_downloads:
+                del self.active_downloads[model_id]
+            self._last_progress_time.pop(model_id, None)
+        elif (
+            progress.status == "in_progress"
+            and current_time() - self._last_progress_time.get(model_id, 0.0)
+            > 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[model_id] = ongoing
+            await self.event_sender.send(
+                NodeDownloadProgress(download_progress=ongoing)
+            )
+            self._last_progress_time[model_id] = current_time()
 
     async def run(self) -> None:
         logger.info("Starting DownloadCoordinator")
@@ -119,12 +160,12 @@ class DownloadCoordinator:
     async def _start_download(self, shard: ShardMetadata) -> None:
         model_id = shard.model_card.model_id
 
-        # Check if already downloading or complete
+        # Check if already downloading, complete, or recently failed
         if model_id in self.download_status:
             status = self.download_status[model_id]
-            if isinstance(status, (DownloadOngoing, DownloadCompleted)):
+            if isinstance(status, (DownloadOngoing, DownloadCompleted, DownloadFailed)):
                 logger.debug(
-                    f"Download for {model_id} already in progress or complete, skipping"
+                    f"Download for {model_id} already in progress, complete, or failed, skipping"
                 )
                 return
 
@@ -169,46 +210,6 @@ class DownloadCoordinator:
         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)
@@ -283,6 +284,12 @@ class DownloadCoordinator:
                     _,
                     progress,
                 ) in self.shard_downloader.get_shard_download_status():
+                    model_id = progress.shard.model_card.model_id
+
+                    # Active downloads emit progress via the callback — don't overwrite
+                    if model_id in self.active_downloads:
+                        continue
+
                     if progress.status == "complete":
                         status: DownloadProgress = DownloadCompleted(
                             node_id=self.node_id,
@@ -290,7 +297,17 @@ class DownloadCoordinator:
                             total_bytes=progress.total_bytes,
                         )
                     elif progress.status in ["in_progress", "not_started"]:
-                        if progress.downloaded_bytes_this_session.in_bytes == 0:
+                        if (
+                            progress.downloaded_bytes.in_bytes
+                            >= progress.total_bytes.in_bytes
+                            > 0
+                        ):
+                            status = DownloadCompleted(
+                                node_id=self.node_id,
+                                shard_metadata=progress.shard,
+                                total_bytes=progress.total_bytes,
+                            )
+                        elif progress.downloaded_bytes_this_session.in_bytes == 0:
                             status = DownloadPending(
                                 node_id=self.node_id, shard_metadata=progress.shard
                             )

src/exo/download/download_utils.py

@@ -457,6 +457,8 @@ async def download_file_with_retry(
             )
         except HuggingFaceAuthenticationError:
             raise
+        except FileNotFoundError:
+            raise
         except HuggingFaceRateLimitError as e:
             if attempt == n_attempts - 1:
                 raise e

src/exo/main.py

@@ -46,7 +46,12 @@ class Node:
         keypair = get_node_id_keypair()
         node_id = NodeId(keypair.to_peer_id().to_base58())
         session_id = SessionId(master_node_id=node_id, election_clock=0)
-        router = Router.create(keypair)
+        bootstrap_peers: list[str] = []
+        if args.bootstrap_peer:
+            ip, port = args.bootstrap_peer.rsplit(":", 1)
+            bootstrap_peers.append(f"/ip4/{ip}/tcp/{port}")
+            logger.info(f"Bootstrap peer: {args.bootstrap_peer}")
+        router = Router.create(keypair, bootstrap_peers=bootstrap_peers)
         await router.register_topic(topics.GLOBAL_EVENTS)
         await router.register_topic(topics.LOCAL_EVENTS)
         await router.register_topic(topics.COMMANDS)
@@ -251,9 +256,10 @@ class Node:
 def main():
     args = Args.parse()
     soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
-    resource.setrlimit(resource.RLIMIT_NOFILE, (max(soft, 65535), hard))
+    target = min(max(soft, 65535), hard)
+    resource.setrlimit(resource.RLIMIT_NOFILE, (target, hard))
 
-    mp.set_start_method("spawn")
+    mp.set_start_method("spawn", force=True)
     # TODO: Refactor the current verbosity system
     logger_setup(EXO_LOG, args.verbosity)
     logger.info("Starting EXO")
@@ -282,6 +288,7 @@ class Args(CamelCaseModel):
     no_worker: bool = False
     no_downloads: bool = False
     fast_synch: bool | None = None  # None = auto, True = force on, False = force off
+    bootstrap_peer: str | None = None
 
     @classmethod
     def parse(cls) -> Self:
@@ -342,6 +349,13 @@ class Args(CamelCaseModel):
             dest="fast_synch",
             help="Force MLX FAST_SYNCH off",
         )
+        parser.add_argument(
+            "--bootstrap-peer",
+            type=str,
+            dest="bootstrap_peer",
+            default=None,
+            help="IP:PORT of an existing node to connect to (bypasses mDNS)",
+        )
 
         args = parser.parse_args()
         return cls(**vars(args))  # pyright: ignore[reportAny] - We are intentionally validating here, we can't do it statically

src/exo/master/adapters/chat_completions.py

@@ -3,7 +3,6 @@
 import time
 from collections.abc import AsyncGenerator
 from typing import Any
-from uuid import uuid4
 
 from exo.shared.types.api import (
     ChatCompletionChoice,
@@ -18,6 +17,7 @@ from exo.shared.types.api import (
     LogprobsContentItem,
     StreamingChoiceResponse,
     ToolCall,
+    Usage,
 )
 from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.common import CommandId
@@ -80,6 +80,7 @@ def chat_request_to_text_generation(
         seed=request.seed,
         stream=request.stream,
         tools=request.tools,
+        enable_thinking=request.enable_thinking,
         chat_template_messages=chat_template_messages
         if chat_template_messages
         else None,
@@ -125,6 +126,8 @@ async def generate_chat_stream(
     chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
 ) -> AsyncGenerator[str, None]:
     """Generate Chat Completions API streaming events from chunks."""
+    last_usage: Usage | None = None
+
     async for chunk in chunk_stream:
         if isinstance(chunk, ErrorChunk):
             error_response = ErrorResponse(
@@ -138,10 +141,12 @@ async def generate_chat_stream(
             yield "data: [DONE]\n\n"
             return
 
+        last_usage = chunk.usage or last_usage
+
         if isinstance(chunk, ToolCallChunk):
             tool_call_deltas = [
                 ToolCall(
-                    id=str(uuid4()),
+                    id=tool.id,
                     index=i,
                     function=tool,
                 )
@@ -161,12 +166,15 @@ async def generate_chat_stream(
                         finish_reason="tool_calls",
                     )
                 ],
+                usage=last_usage,
             )
             yield f"data: {tool_response.model_dump_json()}\n\n"
             yield "data: [DONE]\n\n"
             return
 
         chunk_response = chunk_to_response(chunk, command_id)
+        if chunk.finish_reason is not None:
+            chunk_response = chunk_response.model_copy(update={"usage": last_usage})
         yield f"data: {chunk_response.model_dump_json()}\n\n"
 
         if chunk.finish_reason is not None:
@@ -176,7 +184,9 @@ async def generate_chat_stream(
 async def collect_chat_response(
     command_id: CommandId,
     chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
-) -> ChatCompletionResponse:
+) -> AsyncGenerator[str]:
+    # This is an AsyncGenerator[str] rather than returning a ChatCompletionReponse because
+    # FastAPI handles the cancellation better but wouldn't auto-serialize for some reason
     """Collect all token chunks and return a single ChatCompletionResponse."""
     text_parts: list[str] = []
     tool_calls: list[ToolCall] = []
@@ -184,6 +194,7 @@ async def collect_chat_response(
     model: str | None = None
     finish_reason: FinishReason | None = None
     error_message: str | None = None
+    last_usage: Usage | None = None
 
     async for chunk in chunk_stream:
         if isinstance(chunk, ErrorChunk):
@@ -193,6 +204,8 @@ async def collect_chat_response(
         if model is None:
             model = chunk.model
 
+        last_usage = chunk.usage or last_usage
+
         if isinstance(chunk, TokenChunk):
             text_parts.append(chunk.text)
             if chunk.logprob is not None:
@@ -207,7 +220,7 @@ async def collect_chat_response(
         if isinstance(chunk, ToolCallChunk):
             tool_calls.extend(
                 ToolCall(
-                    id=str(uuid4()),
+                    id=tool.id,
                     index=i,
                     function=tool,
                 )
@@ -223,7 +236,7 @@ async def collect_chat_response(
     combined_text = "".join(text_parts)
     assert model is not None
 
-    return ChatCompletionResponse(
+    yield ChatCompletionResponse(
         id=command_id,
         created=int(time.time()),
         model=model,
@@ -241,4 +254,6 @@ async def collect_chat_response(
                 finish_reason=finish_reason,
             )
         ],
-    )
+        usage=last_usage,
+    ).model_dump_json()
+    return

src/exo/master/adapters/claude.py

@@ -3,9 +3,8 @@
 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.api import FinishReason, Usage
 from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.claude_api import (
     ClaudeContentBlock,
@@ -162,12 +161,14 @@ async def collect_claude_response(
     command_id: CommandId,
     model: str,
     chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
-) -> ClaudeMessagesResponse:
+) -> AsyncGenerator[str]:
+    # This is an AsyncGenerator[str] rather than returning a ChatCompletionReponse because
+    # FastAPI handles the cancellation better but wouldn't auto-serialize for some reason
     """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
+    last_usage: Usage | None = None
     error_message: str | None = None
 
     async for chunk in chunk_stream:
@@ -175,21 +176,21 @@ async def collect_claude_response(
             error_message = chunk.error_message or "Internal server error"
             break
 
+        last_usage = chunk.usage or last_usage
+
         if isinstance(chunk, ToolCallChunk):
             for tool in chunk.tool_calls:
                 tool_use_blocks.append(
                     ClaudeToolUseBlock(
-                        id=f"toolu_{uuid4().hex[:24]}",
+                        id=f"toolu_{tool.id}",
                         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)
@@ -209,11 +210,11 @@ async def collect_claude_response(
     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
+    # Use actual usage data if available
+    input_tokens = last_usage.prompt_tokens if last_usage else 0
+    output_tokens = last_usage.completion_tokens if last_usage else 0
 
-    return ClaudeMessagesResponse(
+    yield ClaudeMessagesResponse(
         id=f"msg_{command_id}",
         model=model,
         content=content,
@@ -222,7 +223,8 @@ async def collect_claude_response(
             input_tokens=input_tokens,
             output_tokens=output_tokens,
         ),
-    )
+    ).model_dump_json()
+    return
 
 
 async def generate_claude_stream(
@@ -250,7 +252,7 @@ async def generate_claude_stream(
 
     output_tokens = 0
     stop_reason: ClaudeStopReason | None = None
-    last_stats = None
+    last_usage: Usage | None = None
     next_block_index = 1  # text block is 0, tool blocks start at 1
 
     async for chunk in chunk_stream:
@@ -258,13 +260,14 @@ async def generate_claude_stream(
             # Close text block and bail
             break
 
+        last_usage = chunk.usage or last_usage
+
         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_id = f"toolu_{tool.id}"
                 tool_input_json = tool.arguments
 
                 # content_block_start for tool_use
@@ -291,7 +294,6 @@ async def generate_claude_stream(
             continue
 
         output_tokens += 1  # Count each chunk as one token
-        last_stats = chunk.stats or last_stats
 
         # content_block_delta
         delta_event = ClaudeContentBlockDeltaEvent(
@@ -303,9 +305,9 @@ async def generate_claude_stream(
         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
+    # Use actual token count from usage if available
+    if last_usage is not None:
+        output_tokens = last_usage.completion_tokens
 
     # content_block_stop for text block
     block_stop = ClaudeContentBlockStopEvent(index=0)

src/exo/master/adapters/responses.py

@@ -3,8 +3,8 @@
 from collections.abc import AsyncGenerator
 from itertools import count
 from typing import Any
-from uuid import uuid4
 
+from exo.shared.types.api import Usage
 from exo.shared.types.chunks import ErrorChunk, TokenChunk, ToolCallChunk
 from exo.shared.types.common import CommandId
 from exo.shared.types.openai_responses import (
@@ -122,13 +122,15 @@ async def collect_responses_response(
     command_id: CommandId,
     model: str,
     chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
-) -> ResponsesResponse:
+) -> AsyncGenerator[str]:
+    # This is an AsyncGenerator[str] rather than returning a ChatCompletionReponse because
+    # FastAPI handles the cancellation better but wouldn't auto-serialize for some reason
     """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
+    last_usage: Usage | None = None
     error_message: str | None = None
 
     async for chunk in chunk_stream:
@@ -136,32 +138,32 @@ async def collect_responses_response(
             error_message = chunk.error_message or "Internal server error"
             break
 
+        last_usage = chunk.usage or last_usage
+
         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]}",
+                        id=f"fc_{tool.id}",
+                        call_id=f"call_{tool.id}",
                         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
+    # Create usage from usage data if available
     usage = None
-    if last_stats is not None:
+    if last_usage 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,
+            input_tokens=last_usage.prompt_tokens,
+            output_tokens=last_usage.completion_tokens,
+            total_tokens=last_usage.total_tokens,
         )
 
     output: list[ResponseItem] = [
@@ -173,14 +175,15 @@ async def collect_responses_response(
     ]
     output.extend(function_call_items)
 
-    return ResponsesResponse(
+    yield ResponsesResponse(
         id=response_id,
         model=model,
         status="completed",
         output=output,
         output_text=accumulated_text,
         usage=usage,
-    )
+    ).model_dump_json()
+    return
 
 
 async def generate_responses_stream(
@@ -236,18 +239,19 @@ async def generate_responses_stream(
 
     accumulated_text = ""
     function_call_items: list[ResponseFunctionCallItem] = []
-    last_stats = None
+    last_usage: Usage | None = None
     next_output_index = 1  # message item is at 0
 
     async for chunk in chunk_stream:
         if isinstance(chunk, ErrorChunk):
             break
 
+        last_usage = chunk.usage or last_usage
+
         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]}"
+                fc_id = f"fc_{tool.id}"
+                call_id = f"call_{tool.id}"
 
                 # response.output_item.added for function_call
                 fc_item = ResponseFunctionCallItem(
@@ -303,7 +307,6 @@ async def generate_responses_stream(
             continue
 
         accumulated_text += chunk.text
-        last_stats = chunk.stats or last_stats
 
         # response.output_text.delta
         delta_event = ResponseTextDeltaEvent(
@@ -347,13 +350,13 @@ async def generate_responses_stream(
     )
     yield f"event: response.output_item.done\ndata: {item_done.model_dump_json()}\n\n"
 
-    # Create usage from stats if available
+    # Create usage from usage data if available
     usage = None
-    if last_stats is not None:
+    if last_usage 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,
+            input_tokens=last_usage.prompt_tokens,
+            output_tokens=last_usage.completion_tokens,
+            total_tokens=last_usage.total_tokens,
         )
 
     # response.completed

src/exo/master/api.py

@@ -3,7 +3,7 @@ import contextlib
 import json
 import random
 import time
-from collections.abc import AsyncGenerator, Awaitable, Callable
+from collections.abc import AsyncGenerator, Awaitable, Callable, Iterator
 from datetime import datetime, timezone
 from http import HTTPStatus
 from pathlib import Path
@@ -37,11 +37,13 @@ from exo.master.adapters.responses import (
     generate_responses_stream,
     responses_request_to_text_generation,
 )
+from exo.master.event_log import DiskEventLog
 from exo.master.image_store import ImageStore
 from exo.master.placement import place_instance as get_instance_placements
 from exo.shared.apply import apply
 from exo.shared.constants import (
     DASHBOARD_DIR,
+    EXO_EVENT_LOG_DIR,
     EXO_IMAGE_CACHE_DIR,
     EXO_MAX_CHUNK_SIZE,
     EXO_TRACING_CACHE_DIR,
@@ -69,8 +71,11 @@ from exo.shared.types.api import (
     ChatCompletionResponse,
     CreateInstanceParams,
     CreateInstanceResponse,
+    CreateMetaInstanceParams,
+    CreateMetaInstanceResponse,
     DeleteDownloadResponse,
     DeleteInstanceResponse,
+    DeleteMetaInstanceResponse,
     ErrorInfo,
     ErrorResponse,
     FinishReason,
@@ -113,8 +118,10 @@ from exo.shared.types.claude_api import (
 from exo.shared.types.commands import (
     Command,
     CreateInstance,
+    CreateMetaInstance,
     DeleteDownload,
     DeleteInstance,
+    DeleteMetaInstance,
     DownloadCommand,
     ForwarderCommand,
     ForwarderDownloadCommand,
@@ -123,10 +130,11 @@ from exo.shared.types.commands import (
     PlaceInstance,
     SendInputChunk,
     StartDownload,
+    TaskCancelled,
     TaskFinished,
     TextGeneration,
 )
-from exo.shared.types.common import CommandId, Id, NodeId, SessionId
+from exo.shared.types.common import CommandId, Id, MetaInstanceId, NodeId, SessionId
 from exo.shared.types.events import (
     ChunkGenerated,
     Event,
@@ -135,6 +143,7 @@ from exo.shared.types.events import (
     TracesMerged,
 )
 from exo.shared.types.memory import Memory
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.openai_responses import (
     ResponsesRequest,
     ResponsesResponse,
@@ -146,6 +155,8 @@ from exo.utils.banner import print_startup_banner
 from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.event_buffer import OrderedBuffer
 
+_API_EVENT_LOG_DIR = EXO_EVENT_LOG_DIR / "api"
+
 
 def _format_to_content_type(image_format: Literal["png", "jpeg", "webp"] | None) -> str:
     return f"image/{image_format or 'png'}"
@@ -175,7 +186,7 @@ class API:
         election_receiver: Receiver[ElectionMessage],
     ) -> None:
         self.state = State()
-        self._event_log: list[Event] = []
+        self._event_log = DiskEventLog(_API_EVENT_LOG_DIR)
         self.command_sender = command_sender
         self.download_command_sender = download_command_sender
         self.global_event_receiver = global_event_receiver
@@ -223,6 +234,8 @@ class API:
 
     def reset(self, new_session_id: SessionId, result_clock: int):
         logger.info("Resetting API State")
+        self._event_log.close()
+        self._event_log = DiskEventLog(_API_EVENT_LOG_DIR)
         self.state = State()
         self.session_id = new_session_id
         self.event_buffer = OrderedBuffer[Event]()
@@ -269,6 +282,9 @@ class API:
         self.app.get("/instance/previews")(self.get_placement_previews)
         self.app.get("/instance/{instance_id}")(self.get_instance)
         self.app.delete("/instance/{instance_id}")(self.delete_instance)
+        self.app.get("/meta_instances")(self.list_meta_instances)
+        self.app.post("/meta_instance")(self.create_meta_instance)
+        self.app.delete("/meta_instance/{meta_instance_id}")(self.delete_meta_instance)
         self.app.get("/models")(self.get_models)
         self.app.get("/v1/models")(self.get_models)
         self.app.post("/models/add")(self.add_custom_model)
@@ -289,7 +305,7 @@ class API:
         self.app.post("/v1/messages", response_model=None)(self.claude_messages)
         self.app.post("/v1/responses", response_model=None)(self.openai_responses)
         self.app.get("/state")(lambda: self.state)
-        self.app.get("/events")(lambda: self._event_log)
+        self.app.get("/events")(self.stream_events)
         self.app.post("/download/start")(self.start_download)
         self.app.delete("/download/{node_id}/{model_id:path}")(self.delete_download)
         self.app.get("/v1/traces")(self.list_traces)
@@ -298,12 +314,27 @@ class API:
         self.app.get("/v1/traces/{task_id}/raw")(self.get_trace_raw)
 
     async def place_instance(self, payload: PlaceInstanceParams):
+        model_card = await ModelCard.load(payload.model_id)
         command = PlaceInstance(
-            model_card=await ModelCard.load(payload.model_id),
+            model_card=model_card,
             sharding=payload.sharding,
             instance_meta=payload.instance_meta,
             min_nodes=payload.min_nodes,
         )
+
+        # Validate placement before sending — fail fast with a clear error
+        # instead of silently dropping the command in the master.
+        try:
+            get_instance_placements(
+                command,
+                topology=self.state.topology,
+                current_instances=self.state.instances,
+                node_memory=self.state.node_memory,
+                node_network=self.state.node_network,
+            )
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
         await self._send(command)
 
         return CreateInstanceResponse(
@@ -386,7 +417,12 @@ class API:
         if len(list(self.state.topology.list_nodes())) == 0:
             return PlacementPreviewResponse(previews=[])
 
-        model_card = await ModelCard.load(model_id)
+        try:
+            model_card = await ModelCard.load(model_id)
+        except Exception as exc:
+            raise HTTPException(
+                status_code=400, detail=f"Failed to load model card: {exc}"
+            ) from exc
         instance_combinations: list[tuple[Sharding, InstanceMeta, int]] = []
         for sharding in (Sharding.Pipeline, Sharding.Tensor):
             for instance_meta in (InstanceMeta.MlxRing, InstanceMeta.MlxJaccl):
@@ -510,6 +546,44 @@ class API:
             instance_id=instance_id,
         )
 
+    def list_meta_instances(self) -> dict[MetaInstanceId, MetaInstance]:
+        return dict(self.state.meta_instances)
+
+    async def create_meta_instance(
+        self, payload: CreateMetaInstanceParams
+    ) -> CreateMetaInstanceResponse:
+        meta_instance = MetaInstance(
+            model_id=payload.model_id,
+            sharding=payload.sharding,
+            instance_meta=payload.instance_meta,
+            min_nodes=payload.min_nodes,
+            node_ids=payload.node_ids,
+        )
+        command = CreateMetaInstance(meta_instance=meta_instance)
+        await self._send(command)
+        return CreateMetaInstanceResponse(
+            message="Command received.",
+            command_id=command.command_id,
+            meta_instance_id=meta_instance.meta_instance_id,
+        )
+
+    async def delete_meta_instance(
+        self, meta_instance_id: MetaInstanceId
+    ) -> DeleteMetaInstanceResponse:
+        meta = self.state.meta_instances.get(meta_instance_id)
+        if not meta:
+            raise HTTPException(status_code=404, detail="MetaInstance not found")
+
+        # Command processor handles cascade-deleting backing instances
+        command = DeleteMetaInstance(meta_instance_id=meta_instance_id)
+        await self._send(command)
+
+        return DeleteMetaInstanceResponse(
+            message="Command received.",
+            command_id=command.command_id,
+            meta_instance_id=meta_instance_id,
+        )
+
     async def _token_chunk_stream(
         self, command_id: CommandId
     ) -> AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None]:
@@ -529,16 +603,14 @@ class API:
                         break
 
         except anyio.get_cancelled_exc_class():
-            # TODO: TaskCancelled
-            """
-            self.command_sender.send_nowait(
-                ForwarderCommand(origin=self.node_id, command=command)
-            )
-            """
+            command = TaskCancelled(cancelled_command_id=command_id)
+            with anyio.CancelScope(shield=True):
+                await self.command_sender.send(
+                    ForwarderCommand(origin=self.node_id, command=command)
+                )
             raise
         finally:
-            command = TaskFinished(finished_command_id=command_id)
-            await self._send(command)
+            await self._send(TaskFinished(finished_command_id=command_id))
             if command_id in self._text_generation_queues:
                 del self._text_generation_queues[command_id]
 
@@ -633,11 +705,14 @@ class API:
                     "X-Accel-Buffering": "no",
                 },
             )
-
-        return await collect_chat_response(
-            command.command_id,
-            self._token_chunk_stream(command.command_id),
-        )
+        else:
+            return StreamingResponse(
+                collect_chat_response(
+                    command.command_id,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="application/json",
+            )
 
     async def bench_chat_completions(
         self, payload: BenchChatCompletionRequest
@@ -653,8 +728,7 @@ class API:
         command = TextGeneration(task_params=task_params)
         await self._send(command)
 
-        response = await self._collect_text_generation_with_stats(command.command_id)
-        return response
+        return await self._collect_text_generation_with_stats(command.command_id)
 
     async def _resolve_and_validate_text_model(self, model_id: ModelId) -> ModelId:
         """Validate a text model exists and return the resolved model ID.
@@ -689,6 +763,22 @@ class API:
             )
         return resolved_model
 
+    def stream_events(self) -> StreamingResponse:
+        def _generate_json_array(events: Iterator[Event]) -> Iterator[str]:
+            yield "["
+            first = True
+            for event in events:
+                if not first:
+                    yield ","
+                first = False
+                yield event.model_dump_json()
+            yield "]"
+
+        return StreamingResponse(
+            _generate_json_array(self._event_log.read_all()),
+            media_type="application/json",
+        )
+
     async def get_image(self, image_id: str) -> FileResponse:
         stored = self._image_store.get(Id(image_id))
         if stored is None:
@@ -856,6 +946,11 @@ class API:
                         del image_metadata[key]
 
         except anyio.get_cancelled_exc_class():
+            command = TaskCancelled(cancelled_command_id=command_id)
+            with anyio.CancelScope(shield=True):
+                await self.command_sender.send(
+                    ForwarderCommand(origin=self.node_id, command=command)
+                )
             raise
         finally:
             await self._send(TaskFinished(finished_command_id=command_id))
@@ -937,6 +1032,11 @@ class API:
 
             return (images, stats if capture_stats else None)
         except anyio.get_cancelled_exc_class():
+            command = TaskCancelled(cancelled_command_id=command_id)
+            with anyio.CancelScope(shield=True):
+                await self.command_sender.send(
+                    ForwarderCommand(origin=self.node_id, command=command)
+                )
             raise
         finally:
             await self._send(TaskFinished(finished_command_id=command_id))
@@ -1194,12 +1294,15 @@ class API:
                     "X-Accel-Buffering": "no",
                 },
             )
-
-        return await collect_claude_response(
-            command.command_id,
-            payload.model,
-            self._token_chunk_stream(command.command_id),
-        )
+        else:
+            return StreamingResponse(
+                collect_claude_response(
+                    command.command_id,
+                    payload.model,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="application/json",
+            )
 
     async def openai_responses(
         self, payload: ResponsesRequest
@@ -1227,11 +1330,15 @@ class API:
                 },
             )
 
-        return await collect_responses_response(
-            command.command_id,
-            payload.model,
-            self._token_chunk_stream(command.command_id),
-        )
+        else:
+            return StreamingResponse(
+                collect_responses_response(
+                    command.command_id,
+                    payload.model,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="application/json",
+            )
 
     def _calculate_total_available_memory(self) -> Memory:
         """Calculate total available memory across all nodes in bytes."""
@@ -1337,6 +1444,7 @@ class API:
                     with anyio.CancelScope(shield=True):
                         shutdown_ev.set()
         finally:
+            self._event_log.close()
             self.command_sender.close()
             self.global_event_receiver.close()
 

src/exo/master/event_log.py

@@ -0,0 +1,187 @@
+import contextlib
+import json
+from collections import OrderedDict
+from collections.abc import Iterator
+from datetime import datetime, timezone
+from io import BufferedRandom, BufferedReader
+from pathlib import Path
+
+import msgspec
+import zstandard
+from loguru import logger
+from pydantic import TypeAdapter
+
+from exo.shared.types.events import Event
+
+_EVENT_ADAPTER: TypeAdapter[Event] = TypeAdapter(Event)
+
+_HEADER_SIZE = 4  # uint32 big-endian
+_OFFSET_CACHE_SIZE = 128
+_MAX_ARCHIVES = 5
+
+
+def _serialize_event(event: Event) -> bytes:
+    return msgspec.msgpack.encode(event.model_dump(mode="json"))
+
+
+def _deserialize_event(raw: bytes) -> Event:
+    # Decode msgpack into a Python dict, then re-encode as JSON for Pydantic.
+    # Pydantic's validate_json() uses JSON-mode coercion (e.g. string -> enum)
+    # even under strict=True, whereas validate_python() does not. Going through
+    # JSON is the only way to get correct round-trip deserialization without
+    # disabling strict mode or adding casts everywhere.
+    as_json = json.dumps(msgspec.msgpack.decode(raw, type=dict))
+    return _EVENT_ADAPTER.validate_json(as_json)
+
+
+def _unpack_header(header: bytes) -> int:
+    return int.from_bytes(header, byteorder="big")
+
+
+def _skip_record(f: BufferedReader) -> bool:
+    """Skip one length-prefixed record. Returns False on EOF."""
+    header = f.read(_HEADER_SIZE)
+    if len(header) < _HEADER_SIZE:
+        return False
+    f.seek(_unpack_header(header), 1)
+    return True
+
+
+def _read_record(f: BufferedReader) -> Event | None:
+    """Read one length-prefixed record. Returns None on EOF."""
+    header = f.read(_HEADER_SIZE)
+    if len(header) < _HEADER_SIZE:
+        return None
+    length = _unpack_header(header)
+    payload = f.read(length)
+    if len(payload) < length:
+        return None
+    return _deserialize_event(payload)
+
+
+class DiskEventLog:
+    """Append-only event log backed by a file on disk.
+
+    On-disk format: sequence of length-prefixed msgpack records.
+    Each record is [4-byte big-endian uint32 length][msgpack payload].
+
+    Uses a bounded LRU cache of event index → byte offset for efficient
+    random access without storing an offset per event.
+    """
+
+    def __init__(self, directory: Path) -> None:
+        self._directory = directory
+        self._directory.mkdir(parents=True, exist_ok=True)
+        self._active_path = directory / "events.bin"
+        self._offset_cache: OrderedDict[int, int] = OrderedDict()
+        self._count: int = 0
+
+        # Rotate stale active file from a previous session/crash
+        if self._active_path.exists():
+            self._rotate(self._active_path, self._directory)
+
+        self._file: BufferedRandom = open(self._active_path, "w+b")  # noqa: SIM115
+
+    def _cache_offset(self, idx: int, offset: int) -> None:
+        self._offset_cache[idx] = offset
+        self._offset_cache.move_to_end(idx)
+        if len(self._offset_cache) > _OFFSET_CACHE_SIZE:
+            self._offset_cache.popitem(last=False)
+
+    def _seek_to(self, f: BufferedReader, target_idx: int) -> None:
+        """Seek f to the byte offset of event target_idx, using cache or scanning forward."""
+        if target_idx in self._offset_cache:
+            self._offset_cache.move_to_end(target_idx)
+            f.seek(self._offset_cache[target_idx])
+            return
+
+        # Find the highest cached index before target_idx
+        scan_from_idx = 0
+        scan_from_offset = 0
+        for cached_idx in self._offset_cache:
+            if cached_idx < target_idx:
+                scan_from_idx = cached_idx
+                scan_from_offset = self._offset_cache[cached_idx]
+
+        # Scan forward, skipping records
+        f.seek(scan_from_offset)
+        for _ in range(scan_from_idx, target_idx):
+            _skip_record(f)
+
+        self._cache_offset(target_idx, f.tell())
+
+    def append(self, event: Event) -> None:
+        packed = _serialize_event(event)
+        self._file.write(len(packed).to_bytes(_HEADER_SIZE, byteorder="big"))
+        self._file.write(packed)
+        self._count += 1
+
+    def read_range(self, start: int, end: int) -> Iterator[Event]:
+        """Yield events from index start (inclusive) to end (exclusive)."""
+        end = min(end, self._count)
+        if start < 0 or end < 0 or start >= end:
+            return
+
+        self._file.flush()
+        with open(self._active_path, "rb") as f:
+            self._seek_to(f, start)
+            for _ in range(end - start):
+                event = _read_record(f)
+                if event is None:
+                    break
+                yield event
+
+            # Cache where we ended up so the next sequential read is a hit
+            if end < self._count:
+                self._cache_offset(end, f.tell())
+
+    def read_all(self) -> Iterator[Event]:
+        """Yield all events from the log one at a time."""
+        if self._count == 0:
+            return
+        self._file.flush()
+        with open(self._active_path, "rb") as f:
+            for _ in range(self._count):
+                event = _read_record(f)
+                if event is None:
+                    break
+                yield event
+
+    def __len__(self) -> int:
+        return self._count
+
+    def close(self) -> None:
+        """Close the file and rotate active file to compressed archive."""
+        if self._file.closed:
+            return
+        self._file.close()
+        if self._active_path.exists() and self._count > 0:
+            self._rotate(self._active_path, self._directory)
+        elif self._active_path.exists():
+            self._active_path.unlink()
+
+    @staticmethod
+    def _rotate(source: Path, directory: Path) -> None:
+        """Compress source into a timestamped archive.
+
+        Keeps at most ``_MAX_ARCHIVES`` compressed copies.  Oldest beyond
+        the limit are deleted.
+        """
+        try:
+            stamp = datetime.now(timezone.utc).strftime("%Y-%m-%d_%H-%M-%S_%f")
+            dest = directory / f"events.{stamp}.bin.zst"
+            compressor = zstandard.ZstdCompressor()
+            with open(source, "rb") as f_in, open(dest, "wb") as f_out:
+                compressor.copy_stream(f_in, f_out)
+            source.unlink()
+            logger.info(f"Rotated event log: {source} -> {dest}")
+
+            # Prune oldest archives beyond the limit
+            archives = sorted(directory.glob("events.*.bin.zst"))
+            for old in archives[:-_MAX_ARCHIVES]:
+                old.unlink()
+        except Exception as e:
+            logger.opt(exception=e).warning(f"Failed to rotate event log {source}")
+            # Clean up the source even if compression fails
+            with contextlib.suppress(OSError):
+                source.unlink()

src/exo/master/main.py

@@ -1,9 +1,11 @@
-from datetime import datetime, timedelta, timezone
+from collections.abc import Sequence
+from datetime import datetime, timezone
 
 import anyio
 from anyio.abc import TaskGroup
 from loguru import logger
 
+from exo.master.event_log import DiskEventLog
 from exo.master.placement import (
     add_instance_to_placements,
     cancel_unnecessary_downloads,
@@ -11,11 +13,22 @@ from exo.master.placement import (
     get_transition_events,
     place_instance,
 )
+from exo.master.process_managers import ProcessManager
+from exo.master.process_managers.instance_health import InstanceHealthReconciler
+from exo.master.process_managers.meta_instance import MetaInstanceReconciler
+from exo.master.process_managers.node_timeout import NodeTimeoutReconciler
+from exo.master.reconcile import (
+    find_unsatisfied_meta_instances,
+    try_place_for_meta_instance,
+)
 from exo.shared.apply import apply
-from exo.shared.constants import EXO_TRACING_ENABLED
+from exo.shared.constants import EXO_EVENT_LOG_DIR, EXO_TRACING_ENABLED
+from exo.shared.models.model_cards import ModelCard
 from exo.shared.types.commands import (
     CreateInstance,
+    CreateMetaInstance,
     DeleteInstance,
+    DeleteMetaInstance,
     ForwarderCommand,
     ForwarderDownloadCommand,
     ImageEdits,
@@ -23,6 +36,7 @@ from exo.shared.types.commands import (
     PlaceInstance,
     RequestEventLog,
     SendInputChunk,
+    TaskCancelled,
     TaskFinished,
     TestCommand,
     TextGeneration,
@@ -34,10 +48,15 @@ from exo.shared.types.events import (
     IndexedEvent,
     InputChunkReceived,
     InstanceDeleted,
+    JacclSideChannelData,
+    JacclSideChannelGathered,
+    MetaInstanceCreated,
+    MetaInstanceDeleted,
+    MetaInstancePlacementFailed,
     NodeGatheredInfo,
-    NodeTimedOut,
     TaskCreated,
     TaskDeleted,
+    TaskStatusUpdated,
     TraceEventData,
     TracesCollected,
     TracesMerged,
@@ -57,7 +76,8 @@ 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.shared.types.worker.runners import RunnerId
+from exo.utils.channels import Receiver, Sender
 from exo.utils.event_buffer import MultiSourceBuffer
 
 
@@ -81,17 +101,16 @@ class Master:
         self.local_event_receiver = local_event_receiver
         self.global_event_sender = global_event_sender
         self.download_command_sender = download_command_sender
-        send, recv = channel[Event]()
-        self.event_sender: Sender[Event] = send
-        self._loopback_event_receiver: Receiver[Event] = recv
-        self._loopback_event_sender: Sender[ForwarderEvent] = (
-            local_event_receiver.clone_sender()
-        )
         self._multi_buffer = MultiSourceBuffer[NodeId, Event]()
-        # TODO: not have this
-        self._event_log: list[Event] = []
+        self._event_log = DiskEventLog(EXO_EVENT_LOG_DIR / "master")
         self._pending_traces: dict[TaskId, dict[int, list[TraceEventData]]] = {}
         self._expected_ranks: dict[TaskId, set[int]] = {}
+        self._jaccl_pending: dict[InstanceId, dict[int, dict[RunnerId, bytes]]] = {}
+        self._process_managers: Sequence[ProcessManager] = [
+            InstanceHealthReconciler(),
+            NodeTimeoutReconciler(),
+            MetaInstanceReconciler(),
+        ]
 
     async def run(self):
         logger.info("Starting Master")
@@ -100,14 +119,12 @@ class Master:
             async with self._tg as tg:
                 tg.start_soon(self._event_processor)
                 tg.start_soon(self._command_processor)
-                tg.start_soon(self._loopback_processor)
-                tg.start_soon(self._plan)
+                tg.start_soon(self._reconcile)
         finally:
+            self._event_log.close()
             self.global_event_sender.close()
             self.local_event_receiver.close()
             self.command_receiver.close()
-            self._loopback_event_sender.close()
-            self._loopback_event_receiver.close()
 
     async def shutdown(self):
         logger.info("Stopping Master")
@@ -278,7 +295,7 @@ class Master:
                         case DeleteInstance():
                             placement = delete_instance(command, self.state.instances)
                             transition_events = get_transition_events(
-                                self.state.instances, placement
+                                self.state.instances, placement, self.state.tasks
                             )
                             for cmd in cancel_unnecessary_downloads(
                                 placement, self.state.downloads
@@ -289,6 +306,85 @@ class Master:
                                     )
                                 )
                             generated_events.extend(transition_events)
+                        case CreateMetaInstance():
+                            logger.info(
+                                f"Creating MetaInstance for {command.meta_instance.model_id}"
+                                f" (min_nodes={command.meta_instance.min_nodes},"
+                                f" sharding={command.meta_instance.sharding})"
+                            )
+                            # Apply immediately so self.state is fresh across
+                            # the await below and the reconciler won't race.
+                            await self._apply_and_broadcast(
+                                MetaInstanceCreated(meta_instance=command.meta_instance)
+                            )
+                            # Immediate placement attempt for responsiveness
+                            model_card = await ModelCard.load(
+                                command.meta_instance.model_id
+                            )
+                            # Re-check: reconciler may have satisfied it during the await
+                            meta_id = command.meta_instance.meta_instance_id
+                            still_unsatisfied = any(
+                                m.meta_instance_id == meta_id
+                                for m in find_unsatisfied_meta_instances(
+                                    self.state.meta_instances,
+                                    self.state.instances,
+                                    self.state.topology,
+                                )
+                            )
+                            if still_unsatisfied:
+                                result = try_place_for_meta_instance(
+                                    command.meta_instance,
+                                    model_card,
+                                    self.state.topology,
+                                    self.state.instances,
+                                    self.state.node_memory,
+                                    self.state.node_network,
+                                )
+                                generated_events.extend(result.events)
+                                if result.error is not None:
+                                    generated_events.append(
+                                        MetaInstancePlacementFailed(
+                                            meta_instance_id=meta_id,
+                                            reason=result.error,
+                                        )
+                                    )
+                        case DeleteMetaInstance():
+                            backing_count = sum(
+                                1
+                                for inst in self.state.instances.values()
+                                if inst.meta_instance_id == command.meta_instance_id
+                            )
+                            logger.info(
+                                f"Deleting MetaInstance {command.meta_instance_id}"
+                                f" (cascade-deleting {backing_count} backing instance(s))"
+                            )
+                            generated_events.append(
+                                MetaInstanceDeleted(
+                                    meta_instance_id=command.meta_instance_id
+                                )
+                            )
+                            # Cascade-delete backing instances atomically,
+                            # cancelling any active tasks first.
+                            for iid, inst in self.state.instances.items():
+                                if inst.meta_instance_id == command.meta_instance_id:
+                                    for task in self.state.tasks.values():
+                                        if (
+                                            task.instance_id == iid
+                                            and task.task_status
+                                            in (
+                                                TaskStatus.Pending,
+                                                TaskStatus.Running,
+                                            )
+                                        ):
+                                            generated_events.append(
+                                                TaskStatusUpdated(
+                                                    task_status=TaskStatus.Cancelled,
+                                                    task_id=task.task_id,
+                                                )
+                                            )
+                                    generated_events.append(
+                                        InstanceDeleted(instance_id=iid)
+                                    )
                         case PlaceInstance():
                             placement = place_instance(
                                 command,
@@ -298,7 +394,7 @@ class Master:
                                 self.state.node_network,
                             )
                             transition_events = get_transition_events(
-                                self.state.instances, placement
+                                self.state.instances, placement, self.state.tasks
                             )
                             generated_events.extend(transition_events)
                         case CreateInstance():
@@ -308,7 +404,7 @@ class Master:
                                 self.state.instances,
                             )
                             transition_events = get_transition_events(
-                                self.state.instances, placement
+                                self.state.instances, placement, self.state.tasks
                             )
                             generated_events.extend(transition_events)
                         case SendInputChunk(chunk=chunk):
@@ -318,6 +414,18 @@ class Master:
                                     chunk=chunk,
                                 )
                             )
+                        case TaskCancelled():
+                            if (
+                                task_id := self.command_task_mapping.get(
+                                    command.cancelled_command_id
+                                )
+                            ) is not None:
+                                generated_events.append(
+                                    TaskStatusUpdated(
+                                        task_status=TaskStatus.Cancelled,
+                                        task_id=task_id,
+                                    )
+                                )
                         case TaskFinished():
                             generated_events.append(
                                 TaskDeleted(
@@ -326,42 +434,45 @@ class Master:
                                     ]
                                 )
                             )
-                            if command.finished_command_id in self.command_task_mapping:
-                                del self.command_task_mapping[
-                                    command.finished_command_id
-                                ]
+                            self.command_task_mapping.pop(
+                                command.finished_command_id, None
+                            )
                         case RequestEventLog():
                             # We should just be able to send everything, since other buffers will ignore old messages
-                            for i in range(command.since_idx, len(self._event_log)):
-                                await self._send_event(
-                                    IndexedEvent(idx=i, event=self._event_log[i])
-                                )
+                            # rate limit to 1000 at a time
+                            end = min(command.since_idx + 1000, len(self._event_log))
+                            for i, event in enumerate(
+                                self._event_log.read_range(command.since_idx, end),
+                                start=command.since_idx,
+                            ):
+                                await self._send_event(IndexedEvent(idx=i, event=event))
                     for event in generated_events:
-                        await self.event_sender.send(event)
+                        await self._apply_and_broadcast(event)
                 except ValueError as e:
                     logger.opt(exception=e).warning("Error in command processor")
 
-    # These plan loops are the cracks showing in our event sourcing architecture - more things could be commands
-    async def _plan(self) -> None:
+    async def _apply_and_broadcast(self, event: Event) -> None:
+        """Apply event to state, persist to disk, and broadcast to workers.
+
+        State is updated synchronously (before any await), so callers can
+        rely on ``self.state`` reflecting this event immediately after the
+        call.  Python's cooperative scheduling guarantees no interleaving
+        between the state read and write.
+        """
+        logger.debug(f"Master indexing event: {str(event)[:100]}")
+        indexed = IndexedEvent(event=event, idx=len(self._event_log))
+        self.state = apply(self.state, indexed)
+        event._master_time_stamp = datetime.now(tz=timezone.utc)  # pyright: ignore[reportPrivateUsage]
+        self._event_log.append(event)
+        await self._send_event(indexed)
+
+    async def _reconcile(self) -> None:
         while True:
-            # kill broken instances
-            connected_node_ids = set(self.state.topology.list_nodes())
-            for instance_id, instance in self.state.instances.items():
-                for node_id in instance.shard_assignments.node_to_runner:
-                    if node_id not in connected_node_ids:
-                        await self.event_sender.send(
-                            InstanceDeleted(instance_id=instance_id)
-                        )
-                        break
-
-            # time out dead nodes
-            for node_id, time in self.state.last_seen.items():
-                now = datetime.now(tz=timezone.utc)
-                if now - time > timedelta(seconds=30):
-                    logger.info(f"Manually removing node {node_id} due to inactivity")
-                    await self.event_sender.send(NodeTimedOut(node_id=node_id))
-
-            await anyio.sleep(10)
+            for pm in self._process_managers:
+                events = await pm.reconcile(self.state)
+                for event in events:
+                    await self._apply_and_broadcast(event)
+            await anyio.sleep(1)
 
     async def _event_processor(self) -> None:
         with self.local_event_receiver as local_events:
@@ -379,32 +490,15 @@ class Master:
                         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)
+                    if isinstance(event, JacclSideChannelData):
+                        await self._apply_and_broadcast(event)
+                        await self._handle_jaccl_side_channel(event)
+                        continue
 
-                    event._master_time_stamp = datetime.now(tz=timezone.utc)  # pyright: ignore[reportPrivateUsage]
                     if isinstance(event, NodeGatheredInfo):
                         event.when = str(datetime.now(tz=timezone.utc))
 
-                    self._event_log.append(event)
-                    await self._send_event(indexed)
-
-    async def _loopback_processor(self) -> None:
-        # this would ideally not be necessary.
-        # this is WAY less hacky than how I was working around this before
-        local_index = 0
-        with self._loopback_event_receiver as events:
-            async for event in events:
-                await self._loopback_event_sender.send(
-                    ForwarderEvent(
-                        origin=NodeId(f"master_{self.node_id}"),
-                        origin_idx=local_index,
-                        session=self.session_id,
-                        event=event,
-                    )
-                )
-                local_index += 1
+                    await self._apply_and_broadcast(event)
 
     # This function is re-entrant, take care!
     async def _send_event(self, event: IndexedEvent):
@@ -436,10 +530,49 @@ class Master:
         for trace_data in self._pending_traces[task_id].values():
             all_trace_data.extend(trace_data)
 
-        await self.event_sender.send(
+        await self._apply_and_broadcast(
             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]
+
+    async def _handle_jaccl_side_channel(self, event: JacclSideChannelData) -> None:
+        """Accumulate SideChannel contributions; when all runners for an instance
+        have submitted for the same sequence, emit JacclSideChannelGathered."""
+        iid = event.instance_id
+        seq = event.sequence
+
+        if iid not in self._jaccl_pending:
+            self._jaccl_pending[iid] = {}
+        if seq not in self._jaccl_pending[iid]:
+            self._jaccl_pending[iid][seq] = {}
+        self._jaccl_pending[iid][seq][event.runner_id] = event.data
+
+        instance = self.state.instances.get(iid)
+        if instance is None:
+            logger.warning(f"JacclSideChannelData for unknown instance {iid}")
+            return
+
+        expected_runners = set(instance.shard_assignments.runner_to_shard.keys())
+        submitted = set(self._jaccl_pending[iid][seq].keys())
+
+        logger.info(
+            f"JACCL side channel: instance={iid} seq={seq} "
+            f"submitted={len(submitted)}/{len(expected_runners)}"
+        )
+
+        if submitted >= expected_runners:
+            gathered = dict(self._jaccl_pending[iid][seq])
+            del self._jaccl_pending[iid][seq]
+            if not self._jaccl_pending[iid]:
+                del self._jaccl_pending[iid]
+
+            await self._apply_and_broadcast(
+                JacclSideChannelGathered(
+                    instance_id=iid,
+                    sequence=seq,
+                    gathered_data=gathered,
+                )
+            )

src/exo/master/placement.py

@@ -6,11 +6,11 @@ from typing import Sequence
 from exo.master.placement_utils import (
     Cycle,
     filter_cycles_by_memory,
+    get_largest_cycles,
     get_mlx_jaccl_coordinators,
     get_mlx_jaccl_devices_matrix,
     get_mlx_ring_hosts_by_node,
     get_shard_assignments,
-    get_smallest_cycles,
 )
 from exo.shared.models.model_cards import ModelId
 from exo.shared.topology import Topology
@@ -22,9 +22,15 @@ from exo.shared.types.commands import (
     PlaceInstance,
 )
 from exo.shared.types.common import NodeId
-from exo.shared.types.events import Event, InstanceCreated, InstanceDeleted
+from exo.shared.types.events import (
+    Event,
+    InstanceCreated,
+    InstanceDeleted,
+    TaskStatusUpdated,
+)
 from exo.shared.types.memory import Memory
 from exo.shared.types.profiling import MemoryUsage, NodeNetworkInfo
+from exo.shared.types.tasks import Task, TaskId, TaskStatus
 from exo.shared.types.worker.downloads import (
     DownloadOngoing,
     DownloadProgress,
@@ -100,23 +106,27 @@ def place_instance(
             "Pipeline parallelism is not supported for DeepSeek V3.1 (8-bit)"
         )
 
-    smallest_cycles = get_smallest_cycles(cycles_with_sufficient_memory)
+    largest_cycles = get_largest_cycles(cycles_with_sufficient_memory)
 
-    smallest_rdma_cycles = [
-        cycle for cycle in smallest_cycles if topology.is_rdma_cycle(cycle)
+    largest_rdma_cycles = [
+        cycle for cycle in largest_cycles if topology.is_rdma_cycle(cycle)
     ]
 
-    if command.instance_meta == InstanceMeta.MlxJaccl and smallest_rdma_cycles != []:
-        smallest_cycles = smallest_rdma_cycles
+    if command.instance_meta == InstanceMeta.MlxJaccl:
+        if not largest_rdma_cycles:
+            raise ValueError(
+                "Requested RDMA (MlxJaccl) but no RDMA-connected cycles available"
+            )
+        largest_cycles = largest_rdma_cycles
 
     cycles_with_leaf_nodes: list[Cycle] = [
         cycle
-        for cycle in smallest_cycles
+        for cycle in largest_cycles
         if any(topology.node_is_leaf(node_id) for node_id in cycle)
     ]
 
     selected_cycle = max(
-        cycles_with_leaf_nodes if cycles_with_leaf_nodes != [] else smallest_cycles,
+        cycles_with_leaf_nodes if cycles_with_leaf_nodes != [] else largest_cycles,
         key=lambda cycle: sum(
             (node_memory[node_id].ram_available for node_id in cycle),
             start=Memory(),
@@ -186,6 +196,7 @@ def delete_instance(
 def get_transition_events(
     current_instances: Mapping[InstanceId, Instance],
     target_instances: Mapping[InstanceId, Instance],
+    tasks: Mapping[TaskId, Task],
 ) -> Sequence[Event]:
     events: list[Event] = []
 
@@ -201,6 +212,18 @@ def get_transition_events(
     # find instances to delete
     for instance_id in current_instances:
         if instance_id not in target_instances:
+            for task in tasks.values():
+                if task.instance_id == instance_id and task.task_status in [
+                    TaskStatus.Pending,
+                    TaskStatus.Running,
+                ]:
+                    events.append(
+                        TaskStatusUpdated(
+                            task_status=TaskStatus.Cancelled,
+                            task_id=task.task_id,
+                        )
+                    )
+
             events.append(
                 InstanceDeleted(
                     instance_id=instance_id,

src/exo/master/placement_utils.py

@@ -37,11 +37,11 @@ def filter_cycles_by_memory(
     return filtered_cycles
 
 
-def get_smallest_cycles(
+def get_largest_cycles(
     cycles: list[Cycle],
 ) -> list[Cycle]:
-    min_nodes = min(len(cycle) for cycle in cycles)
-    return [cycle for cycle in cycles if len(cycle) == min_nodes]
+    max_nodes = max(len(cycle) for cycle in cycles)
+    return [cycle for cycle in cycles if len(cycle) == max_nodes]
 
 
 def allocate_layers_proportionally(

src/exo/master/process_managers/__init__.py

@@ -0,0 +1,12 @@
+from collections.abc import Sequence
+from typing import Protocol, runtime_checkable
+
+from exo.shared.types.events import Event
+from exo.shared.types.state import State
+
+
+@runtime_checkable
+class ProcessManager(Protocol):
+    """A reconciliation step that examines state and returns corrective events."""
+
+    async def reconcile(self, state: State) -> Sequence[Event]: ...

src/exo/master/process_managers/instance_health.py

@@ -0,0 +1,62 @@
+from collections.abc import Sequence
+from typing import final
+
+from loguru import logger
+
+from exo.master.reconcile import instance_connections_healthy, instance_runners_failed
+from exo.shared.types.events import Event, InstanceDeleted, InstanceRetrying
+from exo.shared.types.state import State
+
+MAX_INSTANCE_RETRIES = 3
+
+
+@final
+class InstanceHealthReconciler:
+    """Delete instances whose network connections are broken or whose runners have all failed."""
+
+    async def reconcile(self, state: State) -> Sequence[Event]:
+        events: list[Event] = []
+        for instance_id, instance in state.instances.items():
+            if not instance_connections_healthy(instance, state.topology):
+                events.append(
+                    InstanceDeleted(
+                        instance_id=instance_id,
+                        failure_error="Network connection lost",
+                    )
+                )
+                continue
+
+            is_failed, error_message = instance_runners_failed(
+                instance, state.runners, state.node_identities
+            )
+            if is_failed:
+                # Retry within the same instance if backed by a MetaInstance
+                mid = instance.meta_instance_id
+                mi = state.meta_instances.get(mid) if mid else None
+                if mid and mi and mi.consecutive_failures < MAX_INSTANCE_RETRIES:
+                    logger.info(
+                        f"Instance {instance_id} failed (attempt"
+                        f" {mi.consecutive_failures + 1}/{MAX_INSTANCE_RETRIES}),"
+                        f" retrying: {error_message}"
+                    )
+                    events.append(
+                        InstanceRetrying(
+                            instance_id=instance_id,
+                            meta_instance_id=mid,
+                            failure_error=error_message or "Runner failed",
+                        )
+                    )
+                else:
+                    if mid and mi:
+                        logger.warning(
+                            f"Instance {instance_id} exceeded retry limit"
+                            f" ({MAX_INSTANCE_RETRIES}), deleting:"
+                            f" {error_message}"
+                        )
+                    events.append(
+                        InstanceDeleted(
+                            instance_id=instance_id,
+                            failure_error=error_message,
+                        )
+                    )
+        return events

src/exo/master/process_managers/meta_instance.py

@@ -0,0 +1,91 @@
+from collections.abc import Sequence
+from typing import final
+
+import anyio
+from loguru import logger
+
+from exo.master.reconcile import (
+    find_unsatisfied_meta_instances,
+    try_place_for_meta_instance,
+)
+from exo.shared.models.model_cards import ModelCard
+from exo.shared.types.events import Event, InstanceCreated, MetaInstancePlacementFailed
+from exo.shared.types.state import State
+from exo.shared.types.worker.instances import Instance, InstanceId
+
+MODEL_CARD_LOAD_TIMEOUT_SECONDS = 10
+
+
+@final
+class MetaInstanceReconciler:
+    """Place instances for unsatisfied MetaInstances."""
+
+    async def reconcile(self, state: State) -> Sequence[Event]:
+        all_events: list[Event] = []
+        # Local copy for intermediate tracking — so placement of B
+        # sees A's instance and doesn't double-place on same resources.
+        current_instances: dict[InstanceId, Instance] = dict(state.instances)
+
+        unsatisfied = find_unsatisfied_meta_instances(
+            state.meta_instances,
+            current_instances,
+            state.topology,
+        )
+        for meta_instance in unsatisfied:
+            try:
+                with anyio.fail_after(MODEL_CARD_LOAD_TIMEOUT_SECONDS):
+                    model_card = await ModelCard.load(meta_instance.model_id)
+            except TimeoutError:
+                logger.warning(
+                    f"ModelCard.load timed out for {meta_instance.model_id}, skipping this cycle"
+                )
+                continue
+            except Exception as exc:
+                logger.warning(
+                    f"ModelCard.load failed for {meta_instance.model_id}: {exc}"
+                )
+                error = f"Failed to load model card: {exc}"
+                if meta_instance.placement_error != error:
+                    all_events.append(
+                        MetaInstancePlacementFailed(
+                            meta_instance_id=meta_instance.meta_instance_id,
+                            reason=error,
+                        )
+                    )
+                continue
+
+            result = try_place_for_meta_instance(
+                meta_instance,
+                model_card,
+                state.topology,
+                current_instances,
+                state.node_memory,
+                state.node_network,
+            )
+            # Update local instance map so next placement sees this one
+            for event in result.events:
+                if isinstance(event, InstanceCreated):
+                    logger.info(
+                        f"MetaInstance reconciler placed instance"
+                        f" {event.instance.instance_id} for"
+                        f" {meta_instance.model_id}"
+                    )
+                    current_instances[event.instance.instance_id] = event.instance
+            all_events.extend(result.events)
+
+            # Emit placement failure if error differs from what's already in state
+            if (
+                result.error is not None
+                and meta_instance.placement_error != result.error
+            ):
+                logger.warning(
+                    f"MetaInstance placement failed for"
+                    f" {meta_instance.model_id}: {result.error}"
+                )
+                all_events.append(
+                    MetaInstancePlacementFailed(
+                        meta_instance_id=meta_instance.meta_instance_id,
+                        reason=result.error,
+                    )
+                )
+        return all_events

src/exo/master/process_managers/node_timeout.py

@@ -0,0 +1,27 @@
+from collections.abc import Sequence
+from datetime import datetime, timedelta, timezone
+from typing import final
+
+from loguru import logger
+
+from exo.shared.types.events import Event, NodeTimedOut
+from exo.shared.types.state import State
+
+_DEFAULT_TIMEOUT = timedelta(seconds=30)
+
+
+@final
+class NodeTimeoutReconciler:
+    """Time out nodes that haven't been seen recently."""
+
+    def __init__(self, timeout: timedelta = _DEFAULT_TIMEOUT) -> None:
+        self.timeout = timeout
+
+    async def reconcile(self, state: State) -> Sequence[Event]:
+        now = datetime.now(tz=timezone.utc)
+        events: list[Event] = []
+        for node_id, last_seen in state.last_seen.items():
+            if now - last_seen > self.timeout:
+                logger.info(f"Removing node {node_id} due to inactivity")
+                events.append(NodeTimedOut(node_id=node_id))
+        return events

src/exo/master/reconcile.py

@@ -0,0 +1,240 @@
+from collections.abc import Mapping, Sequence
+from typing import NamedTuple
+
+from loguru import logger
+
+from exo.master.placement import get_transition_events, place_instance
+from exo.shared.models.model_cards import ModelCard
+from exo.shared.topology import Topology
+from exo.shared.types.commands import PlaceInstance
+from exo.shared.types.common import MetaInstanceId, NodeId
+from exo.shared.types.events import Event
+from exo.shared.types.meta_instance import MetaInstance
+from exo.shared.types.profiling import MemoryUsage, NodeIdentity, NodeNetworkInfo
+from exo.shared.types.topology import RDMAConnection, SocketConnection
+from exo.shared.types.worker.instances import (
+    BaseInstance,
+    Instance,
+    InstanceId,
+    MlxJacclInstance,
+    MlxRingInstance,
+)
+from exo.shared.types.worker.runners import (
+    RunnerFailed,
+    RunnerId,
+    RunnerShutdown,
+    RunnerStatus,
+)
+
+
+class PlacementResult(NamedTuple):
+    """Result of a placement attempt: events to apply and optional error reason."""
+
+    events: Sequence[Event]
+    error: str | None
+
+
+def _get_ring_order(instance: BaseInstance) -> list[NodeId]:
+    """Reconstruct ring order from shard device_rank."""
+    node_ranks: list[tuple[NodeId, int]] = []
+    for node_id, runner_id in instance.shard_assignments.node_to_runner.items():
+        shard = instance.shard_assignments.runner_to_shard[runner_id]
+        node_ranks.append((node_id, shard.device_rank))
+    node_ranks.sort(key=lambda x: x[1])
+    return [node_id for node_id, _ in node_ranks]
+
+
+def _ring_connections_healthy(instance: MlxRingInstance, topology: Topology) -> bool:
+    """Check that the specific IPs used by a ring instance still exist in the topology."""
+    ring = _get_ring_order(instance)
+    n = len(ring)
+    for node in ring:
+        hosts = instance.hosts_by_node[node]
+        for idx in range(n):
+            host = hosts[idx]
+            if host.ip in ("0.0.0.0", "198.51.100.1"):
+                continue  # self or placeholder
+            # Real connection: node → ring[idx]. Check specific IP.
+            connections = topology.get_all_connections_between(node, ring[idx])
+            if not any(
+                isinstance(c, SocketConnection)
+                and c.sink_multiaddr.ip_address == host.ip
+                for c in connections
+            ):
+                return False
+    return True
+
+
+def _jaccl_connections_healthy(instance: MlxJacclInstance, topology: Topology) -> bool:
+    """Check that the specific RDMA interfaces used by a JACCL instance still exist."""
+    ring = _get_ring_order(instance)
+    n = len(ring)
+    for i in range(n):
+        for j in range(n):
+            iface = instance.jaccl_devices[i][j]
+            if iface is None:
+                continue
+            connections = topology.get_all_connections_between(ring[i], ring[j])
+            if not any(
+                isinstance(c, RDMAConnection) and c.source_rdma_iface == iface
+                for c in connections
+            ):
+                return False
+    return True
+
+
+def instance_connections_healthy(instance: Instance, topology: Topology) -> bool:
+    """Check that an instance's nodes and specific connections are still in the topology."""
+    instance_nodes = set(instance.shard_assignments.node_to_runner.keys())
+    if not all(topology.contains_node(n) for n in instance_nodes):
+        return False
+    if len(instance_nodes) <= 1:
+        return True
+    match instance:
+        case MlxRingInstance():
+            return _ring_connections_healthy(instance, topology)
+        case MlxJacclInstance():
+            return _jaccl_connections_healthy(instance, topology)
+
+
+def instance_runners_failed(
+    instance: Instance,
+    runners: Mapping[RunnerId, RunnerStatus],
+    node_identities: Mapping[NodeId, NodeIdentity],
+) -> tuple[bool, str | None]:
+    """Check if an instance's runners have all reached terminal failure states.
+
+    Returns ``(True, error_message)`` when ALL runners are terminal
+    (``RunnerFailed`` or ``RunnerShutdown``) and at least one is ``RunnerFailed``.
+
+    Returns ``(False, None)`` when runners are still active, haven't reported
+    yet, or all gracefully shut down (no ``RunnerFailed``).
+    """
+    instance_runner_ids = set(instance.shard_assignments.node_to_runner.values())
+
+    if not instance_runner_ids:
+        return False, None
+
+    # Build reverse mapping: runner_id -> node_id
+    runner_to_node: dict[RunnerId, NodeId] = {
+        runner_id: node_id
+        for node_id, runner_id in instance.shard_assignments.node_to_runner.items()
+    }
+
+    has_any_failed = False
+    error_messages: list[str] = []
+
+    for runner_id in instance_runner_ids:
+        status = runners.get(runner_id)
+        if status is None:
+            # Runner hasn't reported yet — instance is still starting
+            return False, None
+        if isinstance(status, RunnerFailed):
+            has_any_failed = True
+            if status.error_message:
+                node_id = runner_to_node.get(runner_id)
+                name = (
+                    node_identities[node_id].friendly_name
+                    if node_id and node_id in node_identities
+                    else node_id or "unknown"
+                )
+                error_messages.append(f"{name}: {status.error_message}")
+        elif isinstance(status, RunnerShutdown):
+            pass  # Terminal but not a failure indicator on its own
+        else:
+            # Runner is still active (connecting, loading, running, etc.)
+            return False, None
+
+    if has_any_failed:
+        return True, "; ".join(error_messages) if error_messages else "Runner failed"
+
+    # All runners are Shutdown but none Failed — graceful shutdown, not a failure
+    return False, None
+
+
+def instance_satisfies_meta_instance(
+    meta_instance: MetaInstance,
+    instance: Instance,
+) -> bool:
+    """Check if a single instance satisfies a meta-instance's constraints.
+
+    This is a pure constraint check (model, min_nodes, node_ids).
+    Use ``instance_connections_healthy`` separately for topology health.
+    """
+    if instance.shard_assignments.model_id != meta_instance.model_id:
+        return False
+
+    instance_nodes = set(instance.shard_assignments.node_to_runner.keys())
+
+    if len(instance_nodes) < meta_instance.min_nodes:
+        return False
+
+    return meta_instance.node_ids is None or set(meta_instance.node_ids).issubset(
+        instance_nodes
+    )
+
+
+def find_unsatisfied_meta_instances(
+    meta_instances: Mapping[MetaInstanceId, MetaInstance],
+    instances: Mapping[InstanceId, Instance],
+    topology: Topology,
+) -> Sequence[MetaInstance]:
+    """Return meta-instances that have no healthy backing instance."""
+    unsatisfied: list[MetaInstance] = []
+    for meta_id, meta_instance in meta_instances.items():
+        has_healthy_backing = any(
+            instance.meta_instance_id == meta_id
+            and instance_connections_healthy(instance, topology)
+            for instance in instances.values()
+        )
+        if not has_healthy_backing:
+            unsatisfied.append(meta_instance)
+    return unsatisfied
+
+
+def try_place_for_meta_instance(
+    meta_instance: MetaInstance,
+    model_card: ModelCard,
+    topology: Topology,
+    current_instances: Mapping[InstanceId, Instance],
+    node_memory: Mapping[NodeId, MemoryUsage],
+    node_network: Mapping[NodeId, NodeNetworkInfo],
+) -> PlacementResult:
+    """Try to place an instance satisfying the meta-instance constraints.
+
+    Returns a :class:`PlacementResult` with events on success, or an error
+    reason on failure.
+    """
+    command = PlaceInstance(
+        model_card=model_card,
+        sharding=meta_instance.sharding,
+        instance_meta=meta_instance.instance_meta,
+        min_nodes=meta_instance.min_nodes,
+    )
+    try:
+        target_instances = place_instance(
+            command,
+            topology,
+            current_instances,
+            node_memory,
+            node_network,
+            required_nodes=(
+                set(meta_instance.node_ids) if meta_instance.node_ids else None
+            ),
+        )
+        # Tag the new instance with meta_instance_id
+        new_instance_ids = set(target_instances.keys()) - set(current_instances.keys())
+        if new_instance_ids:
+            new_id = next(iter(new_instance_ids))
+            target_instances[new_id] = target_instances[new_id].model_copy(
+                update={"meta_instance_id": meta_instance.meta_instance_id}
+            )
+        return PlacementResult(
+            events=list(get_transition_events(current_instances, target_instances, {})),
+            error=None,
+        )
+    except ValueError as e:
+        logger.debug(
+            f"MetaInstance placement not possible for {meta_instance.model_id}: {e}"
+        )
+        return PlacementResult(events=[], error=str(e))

src/exo/master/tests/test_claude_tool_use.py

@@ -4,7 +4,11 @@ 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.master.adapters.claude import (
+    ClaudeMessagesResponse,
+    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
@@ -17,6 +21,18 @@ async def _chunks_to_stream(
         yield chunk
 
 
+async def _collect_response(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None],
+) -> ClaudeMessagesResponse:
+    """Helper to consume the async generator and parse the JSON response."""
+    parts: list[str] = []
+    async for part in collect_claude_response(command_id, model, chunk_stream):
+        parts.append(part)
+    return ClaudeMessagesResponse.model_validate_json("".join(parts))
+
+
 MODEL = ModelId("test-model")
 COMMAND_ID = CommandId("cmd_test123")
 
@@ -47,7 +63,7 @@ class TestCollectClaudeResponseToolUse:
                 ],
             ),
         ]
-        response = await collect_claude_response(
+        response = await _collect_response(
             COMMAND_ID, "test-model", _chunks_to_stream(chunks)
         )
 
@@ -77,7 +93,7 @@ class TestCollectClaudeResponseToolUse:
                 ],
             ),
         ]
-        response = await collect_claude_response(
+        response = await _collect_response(
             COMMAND_ID, "test-model", _chunks_to_stream(chunks)
         )
 
@@ -102,7 +118,7 @@ class TestCollectClaudeResponseToolUse:
                 ],
             ),
         ]
-        response = await collect_claude_response(
+        response = await _collect_response(
             COMMAND_ID, "test-model", _chunks_to_stream(chunks)
         )
 
@@ -116,7 +132,7 @@ class TestCollectClaudeResponseToolUse:
 
     async def test_no_content_produces_empty_text_block(self):
         chunks: list[ErrorChunk | ToolCallChunk | TokenChunk] = []
-        response = await collect_claude_response(
+        response = await _collect_response(
             COMMAND_ID, "test-model", _chunks_to_stream(chunks)
         )
         assert len(response.content) == 1

src/exo/master/tests/test_event_log.py

@@ -0,0 +1,155 @@
+from pathlib import Path
+
+import pytest
+
+from exo.master.event_log import DiskEventLog
+from exo.shared.types.events import TestEvent
+
+
+@pytest.fixture
+def log_dir(tmp_path: Path) -> Path:
+    return tmp_path / "event_log"
+
+
+def test_append_and_read_back(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    events = [TestEvent() for _ in range(5)]
+    for e in events:
+        log.append(e)
+
+    assert len(log) == 5
+
+    result = list(log.read_all())
+    assert len(result) == 5
+    for original, restored in zip(events, result, strict=True):
+        assert original.event_id == restored.event_id
+
+    log.close()
+
+
+def test_read_range(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    events = [TestEvent() for _ in range(10)]
+    for e in events:
+        log.append(e)
+
+    result = list(log.read_range(3, 7))
+    assert len(result) == 4
+    for i, restored in enumerate(result):
+        assert events[3 + i].event_id == restored.event_id
+
+    log.close()
+
+
+def test_read_range_bounds(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    events = [TestEvent() for _ in range(3)]
+    for e in events:
+        log.append(e)
+
+    # Start beyond count
+    assert list(log.read_range(5, 10)) == []
+    # Negative start
+    assert list(log.read_range(-1, 2)) == []
+    # End beyond count is clamped
+    result = list(log.read_range(1, 100))
+    assert len(result) == 2
+
+    log.close()
+
+
+def test_empty_log(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    assert len(log) == 0
+    assert list(log.read_all()) == []
+    assert list(log.read_range(0, 10)) == []
+    log.close()
+
+
+def _archives(log_dir: Path) -> list[Path]:
+    return sorted(log_dir.glob("events.*.bin.zst"))
+
+
+def test_rotation_on_close(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    log.append(TestEvent())
+    log.close()
+
+    active = log_dir / "events.bin"
+    assert not active.exists()
+
+    archives = _archives(log_dir)
+    assert len(archives) == 1
+    assert archives[0].stat().st_size > 0
+
+
+def test_rotation_on_construction_with_stale_file(log_dir: Path):
+    log_dir.mkdir(parents=True, exist_ok=True)
+    (log_dir / "events.bin").write_bytes(b"stale data")
+
+    log = DiskEventLog(log_dir)
+    archives = _archives(log_dir)
+    assert len(archives) == 1
+    assert archives[0].exists()
+    assert len(log) == 0
+
+    log.close()
+
+
+def test_empty_log_no_archive(log_dir: Path):
+    """Closing an empty log should not leave an archive."""
+    log = DiskEventLog(log_dir)
+    log.close()
+
+    active = log_dir / "events.bin"
+
+    assert not active.exists()
+    assert _archives(log_dir) == []
+
+
+def test_close_is_idempotent(log_dir: Path):
+    log = DiskEventLog(log_dir)
+    log.append(TestEvent())
+    log.close()
+    archive = _archives(log_dir)
+    log.close()  # should not raise
+
+    assert _archives(log_dir) == archive
+
+
+def test_successive_sessions(log_dir: Path):
+    """Simulate two master sessions: both archives should be kept."""
+    log1 = DiskEventLog(log_dir)
+    log1.append(TestEvent())
+    log1.close()
+
+    first_archive = _archives(log_dir)[-1]
+
+    log2 = DiskEventLog(log_dir)
+    log2.append(TestEvent())
+    log2.append(TestEvent())
+    log2.close()
+
+    # Session 1 archive shifted to slot 2, session 2 in slot 1
+    second_archive = _archives(log_dir)[-1]
+    should_be_first_archive = _archives(log_dir)[-2]
+
+    assert first_archive.exists()
+    assert second_archive.exists()
+    assert first_archive != second_archive
+    assert should_be_first_archive == first_archive
+
+
+def test_rotation_keeps_at_most_5_archives(log_dir: Path):
+    """After 7 sessions, only the 5 most recent archives should remain."""
+    all_archives: list[Path] = []
+    for _ in range(7):
+        log = DiskEventLog(log_dir)
+        log.append(TestEvent())
+        log.close()
+        all_archives.append(_archives(log_dir)[-1])
+
+    for old in all_archives[:2]:
+        assert not old.exists()
+    for recent in all_archives[2:]:
+        assert recent.exists()

src/exo/master/tests/test_meta_instance_edge_cases.py

@@ -0,0 +1,778 @@
+"""Edge-case and regression tests for MetaInstance lifecycle, concurrent operations, and error handling."""
+
+import pytest
+
+from exo.master.process_managers.instance_health import (
+    MAX_INSTANCE_RETRIES,
+    InstanceHealthReconciler,
+)
+from exo.master.process_managers.meta_instance import MetaInstanceReconciler
+from exo.master.reconcile import (
+    find_unsatisfied_meta_instances,
+    instance_connections_healthy,
+    instance_runners_failed,
+    instance_satisfies_meta_instance,
+)
+from exo.shared.apply import apply
+from exo.shared.models.model_cards import ModelCard, ModelId, ModelTask
+from exo.shared.topology import Topology
+from exo.shared.types.common import Host, MetaInstanceId, NodeId
+from exo.shared.types.events import (
+    IndexedEvent,
+    InstanceCreated,
+    InstanceDeleted,
+    InstanceRetrying,
+    MetaInstanceCreated,
+    MetaInstanceDeleted,
+    MetaInstancePlacementFailed,
+    TaskStatusUpdated,
+)
+from exo.shared.types.memory import Memory
+from exo.shared.types.meta_instance import MetaInstance
+from exo.shared.types.multiaddr import Multiaddr
+from exo.shared.types.profiling import NodeIdentity
+from exo.shared.types.state import State
+from exo.shared.types.tasks import LoadModel, TaskId, TaskStatus
+from exo.shared.types.topology import Connection, SocketConnection
+from exo.shared.types.worker.instances import (
+    InstanceId,
+    MlxRingInstance,
+)
+from exo.shared.types.worker.runners import (
+    RunnerFailed,
+    RunnerId,
+    RunnerReady,
+    ShardAssignments,
+)
+from exo.shared.types.worker.shards import PipelineShardMetadata
+
+# --- Helpers (copied from test_reconcile.py for independence) ---
+
+
+def _model_card(model_id: str = "test-org/test-model") -> ModelCard:
+    return ModelCard(
+        model_id=ModelId(model_id),
+        storage_size=Memory.from_kb(1000),
+        n_layers=10,
+        hidden_size=30,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    )
+
+
+def _topology(*node_ids: str, connect: bool = True) -> Topology:
+    t = Topology()
+    nodes = [NodeId(n) for n in node_ids]
+    for n in nodes:
+        t.add_node(n)
+    if connect and len(nodes) > 1:
+        for i in range(len(nodes)):
+            j = (i + 1) % len(nodes)
+            t.add_connection(
+                Connection(
+                    source=nodes[i],
+                    sink=nodes[j],
+                    edge=SocketConnection(
+                        sink_multiaddr=Multiaddr(
+                            address=f"/ip4/10.0.0.{j + 1}/tcp/50000"
+                        )
+                    ),
+                )
+            )
+            t.add_connection(
+                Connection(
+                    source=nodes[j],
+                    sink=nodes[i],
+                    edge=SocketConnection(
+                        sink_multiaddr=Multiaddr(
+                            address=f"/ip4/10.0.0.{i + 1}/tcp/50000"
+                        )
+                    ),
+                )
+            )
+    return t
+
+
+def _meta_instance(
+    model_id: str = "test-org/test-model",
+    *,
+    min_nodes: int = 1,
+    node_ids: list[NodeId] | None = None,
+    meta_instance_id: MetaInstanceId | None = None,
+    consecutive_failures: int = 0,
+    last_failure_error: str | None = None,
+    placement_error: str | None = None,
+) -> MetaInstance:
+    return MetaInstance(
+        meta_instance_id=meta_instance_id or MetaInstanceId(),
+        model_id=ModelId(model_id),
+        min_nodes=min_nodes,
+        node_ids=node_ids,
+        consecutive_failures=consecutive_failures,
+        last_failure_error=last_failure_error,
+        placement_error=placement_error,
+    )
+
+
+def _instance(
+    model_id: str = "test-org/test-model",
+    node_ids: list[str] | None = None,
+    instance_id: InstanceId | None = None,
+    meta_instance_id: MetaInstanceId | None = None,
+) -> tuple[InstanceId, MlxRingInstance]:
+    iid = instance_id or InstanceId()
+    nodes = node_ids or ["node-a"]
+    n = len(nodes)
+    mc = _model_card(model_id)
+    ephemeral_port = 50000
+    node_to_runner = {NodeId(nd): RunnerId() for nd in nodes}
+    runner_to_shard = {
+        runner_id: PipelineShardMetadata(
+            model_card=mc,
+            device_rank=i,
+            world_size=n,
+            start_layer=0,
+            end_layer=mc.n_layers,
+            n_layers=mc.n_layers,
+        )
+        for i, runner_id in enumerate(node_to_runner.values())
+    }
+    hosts_by_node: dict[NodeId, list[Host]] = {}
+    for r, node_str in enumerate(nodes):
+        hosts: list[Host] = []
+        for idx in range(n):
+            if idx == r:
+                hosts.append(Host(ip="0.0.0.0", port=ephemeral_port))
+            elif n > 1 and idx in ((r - 1) % n, (r + 1) % n):
+                hosts.append(Host(ip=f"10.0.0.{idx + 1}", port=ephemeral_port))
+            else:
+                hosts.append(Host(ip="198.51.100.1", port=0))
+        hosts_by_node[NodeId(node_str)] = hosts
+    return iid, MlxRingInstance(
+        instance_id=iid,
+        shard_assignments=ShardAssignments(
+            model_id=ModelId(model_id),
+            runner_to_shard=runner_to_shard,
+            node_to_runner=node_to_runner,
+        ),
+        hosts_by_node=hosts_by_node,
+        ephemeral_port=ephemeral_port,
+        meta_instance_id=meta_instance_id,
+    )
+
+
+# =============================================================================
+# 1. MetaInstance lifecycle edge cases
+# =============================================================================
+
+
+def test_meta_instance_model_is_frozen():
+    """MetaInstance should be immutable (frozen model)."""
+    meta = _meta_instance()
+    try:
+        meta.model_id = ModelId("something-else")
+        raise AssertionError("Should have raised")
+    except Exception:
+        pass  # Expected — frozen model
+
+
+def test_meta_instance_created_then_deleted_roundtrip():
+    """Create and delete a MetaInstance through apply — state should be clean."""
+    state = State()
+    meta = _meta_instance()
+    state = apply(
+        state, IndexedEvent(idx=0, event=MetaInstanceCreated(meta_instance=meta))
+    )
+    assert meta.meta_instance_id in state.meta_instances
+    state = apply(
+        state,
+        IndexedEvent(
+            idx=1, event=MetaInstanceDeleted(meta_instance_id=meta.meta_instance_id)
+        ),
+    )
+    assert meta.meta_instance_id not in state.meta_instances
+    assert len(state.meta_instances) == 0
+
+
+def test_delete_nonexistent_meta_instance_is_safe():
+    """Deleting a MetaInstance that doesn't exist should not crash."""
+    state = State()
+    event = MetaInstanceDeleted(meta_instance_id=MetaInstanceId("nonexistent"))
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert len(new_state.meta_instances) == 0
+
+
+def test_placement_failed_for_nonexistent_meta_instance_is_safe():
+    """MetaInstancePlacementFailed for unknown ID should not crash."""
+    state = State()
+    event = MetaInstancePlacementFailed(
+        meta_instance_id=MetaInstanceId("nonexistent"),
+        reason="test",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert len(new_state.meta_instances) == 0
+
+
+def test_multiple_meta_instances_for_same_model():
+    """Multiple MetaInstances for the same model are tracked independently."""
+    state = State()
+    meta_a = _meta_instance("test-org/model-x")
+    meta_b = _meta_instance("test-org/model-x")
+    state = apply(
+        state, IndexedEvent(idx=0, event=MetaInstanceCreated(meta_instance=meta_a))
+    )
+    state = apply(
+        state, IndexedEvent(idx=1, event=MetaInstanceCreated(meta_instance=meta_b))
+    )
+    assert len(state.meta_instances) == 2
+    assert meta_a.meta_instance_id in state.meta_instances
+    assert meta_b.meta_instance_id in state.meta_instances
+
+
+# =============================================================================
+# 2. Retry logic edge cases
+# =============================================================================
+
+
+def test_retry_counter_resets_on_successful_instance_creation():
+    """When a new instance is created for a meta-instance, failures should reset."""
+    meta = _meta_instance(consecutive_failures=2, last_failure_error="old")
+    _, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    state = apply(state, IndexedEvent(idx=0, event=InstanceCreated(instance=inst)))
+    mi = state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+    # last_failure_error is preserved (for UI display)
+    assert mi.last_failure_error == "old"
+
+
+async def test_retry_count_increments_through_full_cycle():
+    """Walk through MAX_INSTANCE_RETRIES worth of retries, then verify delete."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    topology = _topology("node-a")
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        topology=topology,
+    )
+
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    for idx, i in enumerate(range(MAX_INSTANCE_RETRIES)):
+        # Simulate runners failing
+        state_with_runners = state.model_copy(
+            update={"runners": {runner_ids[0]: RunnerFailed(error_message=f"fail-{i}")}}
+        )
+        reconciler = InstanceHealthReconciler()
+        events = await reconciler.reconcile(state_with_runners)
+        assert len(events) == 1
+        assert isinstance(events[0], InstanceRetrying), f"iteration {i}"
+        state = apply(state, IndexedEvent(idx=idx, event=events[0]))
+
+    # After MAX_INSTANCE_RETRIES retries, failure counter should be at max
+    mi = state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == MAX_INSTANCE_RETRIES
+
+    # Next failure should result in deletion
+    state_with_runners = state.model_copy(
+        update={"runners": {runner_ids[0]: RunnerFailed(error_message="final")}}
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state_with_runners)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+
+
+async def test_health_reconciler_respects_exact_limit():
+    """At exactly MAX_INSTANCE_RETRIES, reconciler should delete, not retry."""
+    meta = _meta_instance(consecutive_failures=MAX_INSTANCE_RETRIES)
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="OOM")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+
+
+async def test_health_reconciler_at_limit_minus_one_retries():
+    """At MAX_INSTANCE_RETRIES - 1, reconciler should still retry."""
+    meta = _meta_instance(consecutive_failures=MAX_INSTANCE_RETRIES - 1)
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="OOM")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceRetrying)
+
+
+# =============================================================================
+# 3. Error handling edge cases
+# =============================================================================
+
+
+def test_runners_failed_with_empty_error_message():
+    """RunnerFailed with empty error_message should still report as failed."""
+    _, inst = _instance(node_ids=["node-a"])
+    runners = {
+        rid: RunnerFailed(error_message="")
+        for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    # Empty error message means we get the fallback
+    assert error == "Runner failed"
+
+
+def test_runners_failed_with_none_error_message():
+    """RunnerFailed with None error_message should still report as failed."""
+    _, inst = _instance(node_ids=["node-a"])
+    runners = {
+        rid: RunnerFailed(error_message=None)
+        for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    assert error == "Runner failed"
+
+
+def test_runners_failed_collects_all_error_messages():
+    """With multiple failed runners, all error messages should be collected."""
+    _, inst = _instance(node_ids=["node-a", "node-b", "node-c"])
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {
+        runner_ids[0]: RunnerFailed(error_message="OOM on GPU 0"),
+        runner_ids[1]: RunnerFailed(error_message="OOM on GPU 1"),
+        runner_ids[2]: RunnerFailed(error_message="OOM on GPU 2"),
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    assert error is not None
+    assert "OOM on GPU 0" in error
+    assert "OOM on GPU 1" in error
+    assert "OOM on GPU 2" in error
+
+
+def test_runners_failed_includes_friendly_name():
+    """Error messages should include node friendly names when available."""
+    _, inst = _instance(node_ids=["node-a"])
+    node_id = NodeId("node-a")
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {runner_ids[0]: RunnerFailed(error_message="OOM")}
+    identities = {node_id: NodeIdentity(friendly_name="My Mac Studio")}
+    is_failed, error = instance_runners_failed(inst, runners, identities)
+    assert is_failed is True
+    assert error is not None
+    assert "My Mac Studio" in error
+
+
+def test_instance_retrying_for_missing_instance_is_safe():
+    """InstanceRetrying for an instance not in state should not crash.
+
+    NOTE: When the instance is missing, the handler returns early WITHOUT
+    incrementing the MetaInstance failure counter. This means stale retry
+    events for already-deleted instances are silently dropped. This is
+    acceptable since the InstanceDeleted handler already increments failures.
+    """
+    meta = _meta_instance()
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = InstanceRetrying(
+        instance_id=InstanceId("nonexistent"),
+        meta_instance_id=meta.meta_instance_id,
+        failure_error="crash",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    # Does not crash, but failure count is NOT incremented (early return)
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+
+
+# =============================================================================
+# 4. Backward compatibility
+# =============================================================================
+
+
+def test_instance_without_meta_instance_id_works():
+    """Instances created without meta_instance_id should still function normally."""
+    _, inst = _instance(node_ids=["node-a"])
+    assert inst.meta_instance_id is None
+    topology = _topology("node-a")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+def test_instance_deleted_without_meta_does_not_affect_meta_instances():
+    """Deleting an instance without meta_instance_id should not affect meta_instances."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"])  # no meta_instance_id
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceDeleted(instance_id=iid, failure_error="crash")
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0  # unchanged
+
+
+def test_satisfies_ignores_meta_instance_id_binding():
+    """instance_satisfies_meta_instance checks constraints only, not binding."""
+    meta = _meta_instance()
+    _, inst = _instance(node_ids=["node-a"])  # no meta_instance_id set
+    # Should match on constraints (model, min_nodes) regardless of binding
+    assert instance_satisfies_meta_instance(meta, inst) is True
+
+
+def test_find_unsatisfied_uses_binding_not_constraints():
+    """find_unsatisfied checks meta_instance_id binding, not just constraint matching."""
+    meta = _meta_instance()
+    # Instance matches constraints but is NOT bound to this meta_instance
+    iid, inst = _instance(node_ids=["node-a"])
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta}, {iid: inst}, topology
+    )
+    # Should be unsatisfied because instance.meta_instance_id != meta.meta_instance_id
+    assert list(result) == [meta]
+
+
+# =============================================================================
+# 5. Concurrent / multi-instance scenarios
+# =============================================================================
+
+
+async def test_health_reconciler_handles_multiple_failing_instances():
+    """Multiple instances failing simultaneously should each get their own event."""
+    meta_a = _meta_instance()
+    meta_b = _meta_instance()
+    iid_a, inst_a = _instance(
+        node_ids=["node-a"], meta_instance_id=meta_a.meta_instance_id
+    )
+    iid_b, inst_b = _instance(
+        node_ids=["node-b"], meta_instance_id=meta_b.meta_instance_id
+    )
+    runner_ids_a = list(inst_a.shard_assignments.node_to_runner.values())
+    runner_ids_b = list(inst_b.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={
+            meta_a.meta_instance_id: meta_a,
+            meta_b.meta_instance_id: meta_b,
+        },
+        instances={iid_a: inst_a, iid_b: inst_b},
+        runners={
+            runner_ids_a[0]: RunnerFailed(error_message="OOM"),
+            runner_ids_b[0]: RunnerFailed(error_message="OOM"),
+        },
+        topology=_topology("node-a", "node-b"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 2
+    # Both should be InstanceRetrying since failures < MAX
+    assert all(isinstance(e, InstanceRetrying) for e in events)
+    instance_ids = {e.instance_id for e in events}  # type: ignore[union-attr]
+    assert instance_ids == {iid_a, iid_b}
+
+
+async def test_health_reconciler_mixed_healthy_and_failing():
+    """Only failing instances should produce events; healthy ones should not."""
+    meta_healthy = _meta_instance()
+    meta_failing = _meta_instance()
+    iid_h, inst_h = _instance(
+        node_ids=["node-a"], meta_instance_id=meta_healthy.meta_instance_id
+    )
+    iid_f, inst_f = _instance(
+        node_ids=["node-b"], meta_instance_id=meta_failing.meta_instance_id
+    )
+    runner_ids_h = list(inst_h.shard_assignments.node_to_runner.values())
+    runner_ids_f = list(inst_f.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={
+            meta_healthy.meta_instance_id: meta_healthy,
+            meta_failing.meta_instance_id: meta_failing,
+        },
+        instances={iid_h: inst_h, iid_f: inst_f},
+        runners={
+            runner_ids_h[0]: RunnerReady(),
+            runner_ids_f[0]: RunnerFailed(error_message="crash"),
+        },
+        topology=_topology("node-a", "node-b"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceRetrying)
+    assert events[0].instance_id == iid_f
+
+
+async def test_meta_instance_reconciler_empty_state():
+    """MetaInstanceReconciler with no meta_instances should produce no events."""
+    state = State()
+    reconciler = MetaInstanceReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 0
+
+
+# =============================================================================
+# 6. Placement error tracking
+# =============================================================================
+
+
+def test_placement_failed_sets_error():
+    """MetaInstancePlacementFailed should set placement_error on the MetaInstance."""
+    meta = _meta_instance()
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = MetaInstancePlacementFailed(
+        meta_instance_id=meta.meta_instance_id,
+        reason="Not enough memory",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.placement_error == "Not enough memory"
+
+
+def test_instance_created_clears_placement_error():
+    """InstanceCreated should clear placement_error on the MetaInstance."""
+    meta = _meta_instance(placement_error="Not enough memory")
+    _, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    state = apply(state, IndexedEvent(idx=0, event=InstanceCreated(instance=inst)))
+    mi = state.meta_instances[meta.meta_instance_id]
+    assert mi.placement_error is None
+
+
+def test_placement_error_does_not_increment_failures():
+    """Placement failures should only set placement_error, not increment consecutive_failures."""
+    meta = _meta_instance()
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = MetaInstancePlacementFailed(
+        meta_instance_id=meta.meta_instance_id,
+        reason="No resources",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+    assert mi.placement_error == "No resources"
+
+
+# =============================================================================
+# 7. State serialization roundtrip
+# =============================================================================
+
+
+def test_state_with_meta_instances_serializes():
+    """State with meta_instances should serialize and deserialize correctly."""
+    meta = _meta_instance(consecutive_failures=2, last_failure_error="test")
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    json_str = state.model_dump_json()
+    restored = State.model_validate_json(json_str)
+    assert meta.meta_instance_id in restored.meta_instances
+    mi = restored.meta_instances[meta.meta_instance_id]
+    assert mi.model_id == meta.model_id
+    assert mi.consecutive_failures == 2
+    assert mi.last_failure_error == "test"
+    assert iid in restored.instances
+    assert restored.instances[iid].meta_instance_id == meta.meta_instance_id
+
+
+# =============================================================================
+# 8. MetaInstanceReconciler error handling
+# =============================================================================
+
+
+async def test_meta_instance_reconciler_model_load_error_emits_placement_failed(
+    monkeypatch: "pytest.MonkeyPatch",
+):
+    """When ModelCard.load raises, reconciler emits MetaInstancePlacementFailed."""
+    import exo.master.process_managers.meta_instance as mi_mod
+
+    meta = _meta_instance()
+    topo = _topology("node-a")
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        topology=topo,
+    )
+
+    async def _failing_load(_model_id: ModelId) -> ModelCard:
+        raise RuntimeError("Network error")
+
+    monkeypatch.setattr(
+        mi_mod, "ModelCard", type("MC", (), {"load": staticmethod(_failing_load)})
+    )
+
+    reconciler = MetaInstanceReconciler()
+    events = await reconciler.reconcile(state)
+
+    placement_failed = [e for e in events if isinstance(e, MetaInstancePlacementFailed)]
+    assert len(placement_failed) == 1
+    assert "Failed to load model card" in placement_failed[0].reason
+    assert meta.meta_instance_id == placement_failed[0].meta_instance_id
+
+
+async def test_meta_instance_reconciler_model_load_error_skips_dedup(
+    monkeypatch: "pytest.MonkeyPatch",
+):
+    """When ModelCard.load error matches existing placement_error, no duplicate event."""
+    import exo.master.process_managers.meta_instance as mi_mod
+
+    meta = _meta_instance(placement_error="Failed to load model card: Network error")
+    topo = _topology("node-a")
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        topology=topo,
+    )
+
+    async def _failing_load(_model_id: ModelId) -> ModelCard:
+        raise RuntimeError("Network error")
+
+    monkeypatch.setattr(
+        mi_mod, "ModelCard", type("MC", (), {"load": staticmethod(_failing_load)})
+    )
+
+    reconciler = MetaInstanceReconciler()
+    events = await reconciler.reconcile(state)
+
+    # Error matches existing placement_error, so no duplicate event emitted
+    assert len(events) == 0
+
+
+async def test_meta_instance_reconciler_continues_after_error(
+    monkeypatch: "pytest.MonkeyPatch",
+):
+    """Reconciler should continue to next meta-instance after one fails to load."""
+    import exo.master.process_managers.meta_instance as mi_mod
+
+    meta_a = _meta_instance(model_id="org/model-a")
+    meta_b = _meta_instance(model_id="org/model-b")
+    topo = _topology("node-a")
+    state = State(
+        meta_instances={
+            meta_a.meta_instance_id: meta_a,
+            meta_b.meta_instance_id: meta_b,
+        },
+        topology=topo,
+    )
+
+    call_count = 0
+
+    async def _load_second_fails(model_id: ModelId) -> ModelCard:
+        nonlocal call_count
+        call_count += 1
+        raise RuntimeError(f"Cannot load {model_id}")
+
+    monkeypatch.setattr(
+        mi_mod, "ModelCard", type("MC", (), {"load": staticmethod(_load_second_fails)})
+    )
+
+    reconciler = MetaInstanceReconciler()
+    events = await reconciler.reconcile(state)
+
+    # Both meta-instances should have been attempted (not short-circuited)
+    assert call_count == 2
+    # Both should have placement failed events
+    placement_failed = [e for e in events if isinstance(e, MetaInstancePlacementFailed)]
+    assert len(placement_failed) == 2
+
+
+# =============================================================================
+# 8. Cascade delete with task cancellation
+# =============================================================================
+
+
+def test_cascade_delete_cancels_active_tasks():
+    """Deleting a MetaInstance should cancel tasks on backing instances.
+
+    Regression test: previously, cascade-deleting backing instances via
+    DeleteMetaInstance did not emit TaskStatusUpdated(Cancelled) for active
+    tasks, leaving orphaned task references in state.
+    """
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    task_id = TaskId()
+    task = LoadModel(task_id=task_id, instance_id=iid, task_status=TaskStatus.Running)
+
+    # Build state with meta-instance, backing instance, and active task
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        tasks={task_id: task},
+        topology=_topology("node-a"),
+    )
+
+    # Simulate the cascade-delete event sequence produced by main.py:
+    # 1. MetaInstanceDeleted
+    # 2. TaskStatusUpdated(Cancelled) for active tasks
+    # 3. InstanceDeleted
+    idx = 0
+    state = apply(
+        state,
+        IndexedEvent(
+            idx=idx,
+            event=MetaInstanceDeleted(meta_instance_id=meta.meta_instance_id),
+        ),
+    )
+    idx += 1
+    state = apply(
+        state,
+        IndexedEvent(
+            idx=idx,
+            event=TaskStatusUpdated(task_id=task_id, task_status=TaskStatus.Cancelled),
+        ),
+    )
+    idx += 1
+    state = apply(
+        state,
+        IndexedEvent(idx=idx, event=InstanceDeleted(instance_id=iid)),
+    )
+
+    # Verify everything is cleaned up
+    assert len(state.meta_instances) == 0
+    assert len(state.instances) == 0
+    assert state.tasks[task_id].task_status == TaskStatus.Cancelled
+
+
+def test_cascade_delete_skips_completed_tasks():
+    """Cascade delete should only cancel Pending/Running tasks, not completed ones."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+
+    running_task_id = TaskId()
+    completed_task_id = TaskId()
+    running_task = LoadModel(
+        task_id=running_task_id, instance_id=iid, task_status=TaskStatus.Running
+    )
+    completed_task = LoadModel(
+        task_id=completed_task_id, instance_id=iid, task_status=TaskStatus.Complete
+    )
+
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        tasks={running_task_id: running_task, completed_task_id: completed_task},
+        topology=_topology("node-a"),
+    )
+
+    # Only the running task should be cancelled — we verify the logic pattern
+    # by checking which tasks are Pending or Running
+    active_tasks = [
+        t
+        for t in state.tasks.values()
+        if t.instance_id == iid
+        and t.task_status in (TaskStatus.Pending, TaskStatus.Running)
+    ]
+    assert len(active_tasks) == 1
+    assert active_tasks[0].task_id == running_task_id

src/exo/master/tests/test_placement.py

@@ -239,7 +239,7 @@ def test_get_transition_events_no_change(instance: Instance):
     target_instances = {instance_id: instance}
 
     # act
-    events = get_transition_events(current_instances, target_instances)
+    events = get_transition_events(current_instances, target_instances, {})
 
     # assert
     assert len(events) == 0
@@ -252,7 +252,7 @@ def test_get_transition_events_create_instance(instance: Instance):
     target_instances: dict[InstanceId, Instance] = {instance_id: instance}
 
     # act
-    events = get_transition_events(current_instances, target_instances)
+    events = get_transition_events(current_instances, target_instances, {})
 
     # assert
     assert len(events) == 1
@@ -266,7 +266,7 @@ def test_get_transition_events_delete_instance(instance: Instance):
     target_instances: dict[InstanceId, Instance] = {}
 
     # act
-    events = get_transition_events(current_instances, target_instances)
+    events = get_transition_events(current_instances, target_instances, {})
 
     # assert
     assert len(events) == 1

src/exo/master/tests/test_placement_utils.py

@@ -3,10 +3,10 @@ import pytest
 from exo.master.placement_utils import (
     allocate_layers_proportionally,
     filter_cycles_by_memory,
+    get_largest_cycles,
     get_mlx_jaccl_coordinators,
     get_shard_assignments,
     get_shard_assignments_for_pipeline_parallel,
-    get_smallest_cycles,
 )
 from exo.master.tests.conftest import (
     create_node_memory,
@@ -143,7 +143,7 @@ def test_filter_multiple_cycles_by_memory():
     }
 
 
-def test_get_smallest_cycles():
+def test_get_largest_cycles():
     # arrange
     node_a_id = NodeId()
     node_b_id = NodeId()
@@ -175,12 +175,12 @@ def test_get_smallest_cycles():
     cycles = [c for c in topology.get_cycles() if len(c) != 1]  # ignore singletons
 
     # act
-    smallest_cycles = get_smallest_cycles(cycles)
+    largest_cycles = get_largest_cycles(cycles)
 
     # assert
-    assert len(smallest_cycles) == 1
-    assert len(smallest_cycles[0]) == 2
-    assert set(n for n in smallest_cycles[0]) == {node_a_id, node_b_id}
+    assert len(largest_cycles) == 1
+    assert len(largest_cycles[0]) == 3
+    assert set(n for n in largest_cycles[0]) == {node_a_id, node_b_id, node_c_id}
 
 
 @pytest.mark.parametrize(

src/exo/master/tests/test_reconcile.py

@@ -0,0 +1,742 @@
+from exo.master.process_managers.instance_health import InstanceHealthReconciler
+from exo.master.reconcile import (
+    find_unsatisfied_meta_instances,
+    instance_connections_healthy,
+    instance_runners_failed,
+    instance_satisfies_meta_instance,
+)
+from exo.shared.apply import apply
+from exo.shared.models.model_cards import ModelCard, ModelId, ModelTask
+from exo.shared.topology import Topology
+from exo.shared.types.common import Host, MetaInstanceId, NodeId
+from exo.shared.types.events import (
+    IndexedEvent,
+    InstanceCreated,
+    InstanceDeleted,
+    InstanceRetrying,
+    MetaInstanceCreated,
+    MetaInstanceDeleted,
+)
+from exo.shared.types.memory import Memory
+from exo.shared.types.meta_instance import MetaInstance
+from exo.shared.types.multiaddr import Multiaddr
+from exo.shared.types.state import State
+from exo.shared.types.topology import Connection, SocketConnection
+from exo.shared.types.worker.instances import (
+    InstanceId,
+    MlxRingInstance,
+)
+from exo.shared.types.worker.runners import (
+    RunnerFailed,
+    RunnerId,
+    RunnerLoading,
+    RunnerReady,
+    RunnerShutdown,
+    ShardAssignments,
+)
+from exo.shared.types.worker.shards import PipelineShardMetadata
+
+
+def _model_card(model_id: str = "test-org/test-model") -> ModelCard:
+    return ModelCard(
+        model_id=ModelId(model_id),
+        storage_size=Memory.from_kb(1000),
+        n_layers=10,
+        hidden_size=30,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    )
+
+
+def _topology(*node_ids: str, connect: bool = True) -> Topology:
+    """Build a topology with nodes connected in a bidirectional ring with unique IPs.
+
+    Node at index ``i`` gets IP ``10.0.0.{i+1}``. Edges go in both directions
+    between consecutive nodes (including wrap-around).
+    """
+    t = Topology()
+    nodes = [NodeId(n) for n in node_ids]
+    for n in nodes:
+        t.add_node(n)
+    if connect and len(nodes) > 1:
+        for i in range(len(nodes)):
+            j = (i + 1) % len(nodes)
+            t.add_connection(
+                Connection(
+                    source=nodes[i],
+                    sink=nodes[j],
+                    edge=SocketConnection(
+                        sink_multiaddr=Multiaddr(
+                            address=f"/ip4/10.0.0.{j + 1}/tcp/50000"
+                        )
+                    ),
+                )
+            )
+            t.add_connection(
+                Connection(
+                    source=nodes[j],
+                    sink=nodes[i],
+                    edge=SocketConnection(
+                        sink_multiaddr=Multiaddr(
+                            address=f"/ip4/10.0.0.{i + 1}/tcp/50000"
+                        )
+                    ),
+                )
+            )
+    return t
+
+
+def _meta_instance(
+    model_id: str = "test-org/test-model",
+    *,
+    min_nodes: int = 1,
+    node_ids: list[NodeId] | None = None,
+    meta_instance_id: MetaInstanceId | None = None,
+) -> MetaInstance:
+    return MetaInstance(
+        meta_instance_id=meta_instance_id or MetaInstanceId(),
+        model_id=ModelId(model_id),
+        min_nodes=min_nodes,
+        node_ids=node_ids,
+    )
+
+
+def _instance(
+    model_id: str = "test-org/test-model",
+    node_ids: list[str] | None = None,
+    instance_id: InstanceId | None = None,
+    meta_instance_id: MetaInstanceId | None = None,
+) -> tuple[InstanceId, MlxRingInstance]:
+    """Create a test instance with hosts_by_node matching ``_topology()`` IPs."""
+    iid = instance_id or InstanceId()
+    nodes = node_ids or ["node-a"]
+    n = len(nodes)
+    mc = _model_card(model_id)
+    ephemeral_port = 50000
+    node_to_runner = {NodeId(nd): RunnerId() for nd in nodes}
+    runner_to_shard = {
+        runner_id: PipelineShardMetadata(
+            model_card=mc,
+            device_rank=i,
+            world_size=n,
+            start_layer=0,
+            end_layer=mc.n_layers,
+            n_layers=mc.n_layers,
+        )
+        for i, runner_id in enumerate(node_to_runner.values())
+    }
+    # Build hosts_by_node with IPs matching _topology() convention:
+    # node at index idx has IP 10.0.0.{idx+1}
+    hosts_by_node: dict[NodeId, list[Host]] = {}
+    for r, node_str in enumerate(nodes):
+        hosts: list[Host] = []
+        for idx in range(n):
+            if idx == r:
+                hosts.append(Host(ip="0.0.0.0", port=ephemeral_port))
+            elif n > 1 and idx in ((r - 1) % n, (r + 1) % n):
+                hosts.append(Host(ip=f"10.0.0.{idx + 1}", port=ephemeral_port))
+            else:
+                hosts.append(Host(ip="198.51.100.1", port=0))
+        hosts_by_node[NodeId(node_str)] = hosts
+    return iid, MlxRingInstance(
+        instance_id=iid,
+        shard_assignments=ShardAssignments(
+            model_id=ModelId(model_id),
+            runner_to_shard=runner_to_shard,
+            node_to_runner=node_to_runner,
+        ),
+        hosts_by_node=hosts_by_node,
+        ephemeral_port=ephemeral_port,
+        meta_instance_id=meta_instance_id,
+    )
+
+
+# --- instance_satisfies_meta_instance (pure constraint matching) ---
+
+
+def test_satisfies_matching_model():
+    meta = _meta_instance()
+    _, inst = _instance(node_ids=["node-a"])
+    assert instance_satisfies_meta_instance(meta, inst) is True
+
+
+def test_not_satisfies_wrong_model():
+    meta = _meta_instance("test-org/model-a")
+    _, inst = _instance("test-org/model-b")
+    assert instance_satisfies_meta_instance(meta, inst) is False
+
+
+def test_not_satisfies_missing_required_node():
+    meta = _meta_instance(node_ids=[NodeId("node-c")])
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    assert instance_satisfies_meta_instance(meta, inst) is False
+
+
+def test_not_satisfies_fewer_than_min_nodes():
+    meta = _meta_instance(min_nodes=3)
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    assert instance_satisfies_meta_instance(meta, inst) is False
+
+
+def test_satisfies_with_node_ids_specified():
+    meta = _meta_instance(node_ids=[NodeId("node-a"), NodeId("node-b")], min_nodes=2)
+    _, inst = _instance(node_ids=["node-a", "node-b", "node-c"])
+    assert instance_satisfies_meta_instance(meta, inst) is True
+
+
+# --- instance_connections_healthy ---
+
+
+def test_healthy_single_node_present():
+    _, inst = _instance(node_ids=["node-a"])
+    topology = _topology("node-a")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+def test_unhealthy_single_node_missing():
+    _, inst = _instance(node_ids=["node-a"])
+    topology = Topology()  # empty
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_healthy_two_node_ring():
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    topology = _topology("node-a", "node-b")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+def test_unhealthy_two_node_edge_removed():
+    """Nodes present but edge removed — ring broken."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    topology = _topology("node-a", "node-b", connect=False)
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_unhealthy_two_node_ip_changed():
+    """Edge exists but with a different IP than instance was configured with."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    # Build topology with different IPs than _instance() expects
+    topology = Topology()
+    topology.add_node(NodeId("node-a"))
+    topology.add_node(NodeId("node-b"))
+    topology.add_connection(
+        Connection(
+            source=NodeId("node-a"),
+            sink=NodeId("node-b"),
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/192.168.99.99/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=NodeId("node-b"),
+            sink=NodeId("node-a"),
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/192.168.99.98/tcp/50000")
+            ),
+        )
+    )
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_healthy_three_node_ring():
+    _, inst = _instance(node_ids=["node-a", "node-b", "node-c"])
+    topology = _topology("node-a", "node-b", "node-c")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+def test_unhealthy_three_node_one_edge_removed():
+    """Remove one edge from a three-node ring — instance unhealthy."""
+    _, inst = _instance(node_ids=["node-a", "node-b", "node-c"])
+    # Build topology with one direction of one edge missing
+    topology = Topology()
+    nodes = [NodeId("node-a"), NodeId("node-b"), NodeId("node-c")]
+    for n in nodes:
+        topology.add_node(n)
+    # Add all edges except node-a → node-b
+    topology.add_connection(
+        Connection(
+            source=nodes[1],
+            sink=nodes[0],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.1/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=nodes[1],
+            sink=nodes[2],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.3/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=nodes[2],
+            sink=nodes[1],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.2/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=nodes[2],
+            sink=nodes[0],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.1/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=nodes[0],
+            sink=nodes[2],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.3/tcp/50000")
+            ),
+        )
+    )
+    # Missing: node-a → node-b (ip 10.0.0.2)
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_unhealthy_node_missing_from_topology():
+    """Instance has a node that's not in the topology at all."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    topology = _topology("node-a")  # node-b not present
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_healthy_extra_nodes_in_topology():
+    """Extra nodes in topology don't affect instance health."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    topology = _topology("node-a", "node-b", "node-c")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+# --- find_unsatisfied_meta_instances ---
+
+
+def test_unsatisfied_no_meta_instances():
+    result = find_unsatisfied_meta_instances({}, {}, Topology())
+    assert list(result) == []
+
+
+def test_unsatisfied_one_satisfied():
+    meta = _meta_instance()
+    id_a, inst_a = _instance(meta_instance_id=meta.meta_instance_id)
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta},
+        {id_a: inst_a},
+        topology,
+    )
+    assert list(result) == []
+
+
+def test_unsatisfied_one_not_satisfied():
+    meta = _meta_instance("test-org/model-x")
+    id_a, inst_a = _instance("test-org/model-y")
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta}, {id_a: inst_a}, topology
+    )
+    assert list(result) == [meta]
+
+
+def test_unsatisfied_mix():
+    meta_satisfied = _meta_instance("test-org/model-a")
+    meta_unsatisfied = _meta_instance("test-org/model-b")
+    id_a, inst_a = _instance(
+        "test-org/model-a", meta_instance_id=meta_satisfied.meta_instance_id
+    )
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {
+            meta_satisfied.meta_instance_id: meta_satisfied,
+            meta_unsatisfied.meta_instance_id: meta_unsatisfied,
+        },
+        {id_a: inst_a},
+        topology,
+    )
+    assert list(result) == [meta_unsatisfied]
+
+
+def test_unsatisfied_node_disconnect():
+    meta = _meta_instance()
+    id_a, inst_a = _instance(
+        node_ids=["node-a", "node-b"], meta_instance_id=meta.meta_instance_id
+    )
+    topology = _topology("node-a")  # node-b disconnected
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta},
+        {id_a: inst_a},
+        topology,
+    )
+    assert list(result) == [meta]
+
+
+def test_unsatisfied_edge_break():
+    """Instance exists but its connections broke — meta-instance becomes unsatisfied."""
+    meta = _meta_instance()
+    id_a, inst_a = _instance(
+        node_ids=["node-a", "node-b"], meta_instance_id=meta.meta_instance_id
+    )
+    topology = _topology("node-a", "node-b", connect=False)  # nodes present, no edges
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta},
+        {id_a: inst_a},
+        topology,
+    )
+    assert list(result) == [meta]
+
+
+def test_unsatisfied_idempotent():
+    meta = _meta_instance("test-org/model-x")
+    topology = _topology("node-a")
+    meta_instances = {meta.meta_instance_id: meta}
+    instances: dict[InstanceId, MlxRingInstance] = {}
+    result_1 = list(
+        find_unsatisfied_meta_instances(meta_instances, instances, topology)
+    )
+    result_2 = list(
+        find_unsatisfied_meta_instances(meta_instances, instances, topology)
+    )
+    assert result_1 == result_2
+
+
+def test_unsatisfied_exclusive_binding():
+    """Two MetaInstances for the same model: one is bound via meta_instance_id, the other is unsatisfied."""
+    meta_a = _meta_instance("test-org/model-x")
+    meta_b = _meta_instance("test-org/model-x")
+    id_inst, inst = _instance(
+        "test-org/model-x", meta_instance_id=meta_a.meta_instance_id
+    )
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {
+            meta_a.meta_instance_id: meta_a,
+            meta_b.meta_instance_id: meta_b,
+        },
+        {id_inst: inst},
+        topology,
+    )
+    assert list(result) == [meta_b]
+
+
+# --- apply handlers ---
+
+
+def test_apply_meta_instance_created():
+    state = State()
+    meta = _meta_instance()
+    event = MetaInstanceCreated(meta_instance=meta)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert meta.meta_instance_id in new_state.meta_instances
+    assert new_state.meta_instances[meta.meta_instance_id] == meta
+
+
+def test_apply_meta_instance_deleted():
+    meta = _meta_instance()
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = MetaInstanceDeleted(meta_instance_id=meta.meta_instance_id)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert meta.meta_instance_id not in new_state.meta_instances
+
+
+def test_apply_meta_instance_deleted_clears_failure_info():
+    meta = _meta_instance().model_copy(
+        update={"consecutive_failures": 2, "last_failure_error": "OOM"}
+    )
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = MetaInstanceDeleted(meta_instance_id=meta.meta_instance_id)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert meta.meta_instance_id not in new_state.meta_instances
+
+
+# --- instance_runners_failed ---
+
+
+def test_runners_failed_all_failed():
+    """All runners in RunnerFailed -> instance is failed."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    runners = {
+        rid: RunnerFailed(error_message="OOM")
+        for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    assert error is not None
+    assert "OOM" in error
+
+
+def test_runners_failed_mixed_failed_shutdown():
+    """One Failed + one Shutdown = failed."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {
+        runner_ids[0]: RunnerFailed(error_message="crash"),
+        runner_ids[1]: RunnerShutdown(),
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    assert error is not None
+    assert "crash" in error
+
+
+def test_runners_not_failed_all_shutdown():
+    """All Shutdown (graceful) = not a failure."""
+    _, inst = _instance(node_ids=["node-a"])
+    runners = {
+        rid: RunnerShutdown() for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, _ = instance_runners_failed(inst, runners, {})
+    assert is_failed is False
+
+
+def test_runners_not_failed_still_active():
+    """Some runners still active = not failed yet."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {
+        runner_ids[0]: RunnerFailed(error_message="OOM"),
+        runner_ids[1]: RunnerLoading(),
+    }
+    is_failed, _ = instance_runners_failed(inst, runners, {})
+    assert is_failed is False
+
+
+def test_runners_not_failed_no_status():
+    """Runner not yet reported = not failed."""
+    _, inst = _instance(node_ids=["node-a"])
+    is_failed, _ = instance_runners_failed(inst, {}, {})
+    assert is_failed is False
+
+
+def test_runners_not_failed_healthy():
+    """Runners in Ready state = not failed."""
+    _, inst = _instance(node_ids=["node-a"])
+    runners = {
+        rid: RunnerReady() for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, _ = instance_runners_failed(inst, runners, {})
+    assert is_failed is False
+
+
+# --- failure tracking in apply_instance_deleted ---
+
+
+def test_apply_instance_deleted_tracks_failure():
+    """InstanceDeleted with failure_error increments meta instance failure count."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceDeleted(instance_id=iid, failure_error="Runner OOM")
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 1
+    assert mi.last_failure_error == "Runner OOM"
+
+
+def test_apply_instance_deleted_increments_failure():
+    """Subsequent failures increment the counter."""
+    meta = _meta_instance().model_copy(
+        update={"consecutive_failures": 2, "last_failure_error": "previous error"}
+    )
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceDeleted(instance_id=iid, failure_error="new error")
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 3
+    assert mi.last_failure_error == "new error"
+
+
+def test_apply_instance_deleted_no_failure_no_tracking():
+    """InstanceDeleted without failure_error does not track."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceDeleted(instance_id=iid)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+
+
+def test_apply_instance_deleted_orphan_no_tracking():
+    """InstanceDeleted for orphan instance (no meta_instance_id) does not track."""
+    iid, inst = _instance(node_ids=["node-a"])
+    state = State(instances={iid: inst})
+    event = InstanceDeleted(instance_id=iid, failure_error="crash")
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert len(new_state.meta_instances) == 0
+
+
+# --- InstanceRetrying ---
+
+
+def test_apply_instance_retrying_removes_runners():
+    """InstanceRetrying removes the instance's runners from state but keeps the instance."""
+    meta = _meta_instance()
+    iid, inst = _instance(
+        node_ids=["node-a", "node-b"], meta_instance_id=meta.meta_instance_id
+    )
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {
+        runner_ids[0]: RunnerFailed(error_message="OOM"),
+        runner_ids[1]: RunnerShutdown(),
+    }
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners=runners,
+    )
+    event = InstanceRetrying(
+        instance_id=iid,
+        meta_instance_id=meta.meta_instance_id,
+        failure_error="OOM",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    # Instance still exists
+    assert iid in new_state.instances
+    # Runners removed
+    assert runner_ids[0] not in new_state.runners
+    assert runner_ids[1] not in new_state.runners
+
+
+def test_apply_instance_retrying_increments_failure():
+    """InstanceRetrying increments consecutive_failures on the MetaInstance."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceRetrying(
+        instance_id=iid,
+        meta_instance_id=meta.meta_instance_id,
+        failure_error="crash",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 1
+    assert mi.last_failure_error == "crash"
+
+
+def test_apply_instance_retrying_skips_missing_runners():
+    """InstanceRetrying doesn't assert if runners haven't reported yet."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    # No runners in state at all
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceRetrying(
+        instance_id=iid,
+        meta_instance_id=meta.meta_instance_id,
+        failure_error="crash",
+    )
+    # Should not raise
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert iid in new_state.instances
+
+
+def test_apply_instance_created_resets_failure_counter():
+    """InstanceCreated resets consecutive_failures but preserves last_failure_error."""
+    meta = _meta_instance().model_copy(
+        update={"consecutive_failures": 3, "last_failure_error": "old error"}
+    )
+    _, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = InstanceCreated(instance=inst)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+    assert mi.last_failure_error == "old error"
+    assert mi.placement_error is None
+
+
+# --- InstanceHealthReconciler retry-vs-delete ---
+
+
+async def test_health_reconciler_retries_when_under_limit():
+    """InstanceHealthReconciler emits InstanceRetrying when consecutive_failures < 3."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="OOM")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceRetrying)
+    assert events[0].instance_id == iid
+    assert events[0].meta_instance_id == meta.meta_instance_id
+
+
+async def test_health_reconciler_deletes_when_limit_reached():
+    """InstanceHealthReconciler emits InstanceDeleted when consecutive_failures >= 3."""
+    meta = _meta_instance().model_copy(update={"consecutive_failures": 3})
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="OOM")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+
+
+async def test_health_reconciler_deletes_without_meta_instance():
+    """Instances without a MetaInstance are deleted immediately on runner failure."""
+    iid, inst = _instance(node_ids=["node-a"])
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="crash")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+
+
+async def test_health_reconciler_network_failure_always_deletes():
+    """Network failure always triggers InstanceDeleted regardless of retry count."""
+    meta = _meta_instance()
+    iid, inst = _instance(
+        node_ids=["node-a", "node-b"], meta_instance_id=meta.meta_instance_id
+    )
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        topology=_topology("node-a"),  # node-b missing
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+    assert events[0].failure_error == "Network connection lost"

src/exo/routing/router.py

@@ -101,8 +101,12 @@ class TopicRouter[T: CamelCaseModel]:
 
 class Router:
     @classmethod
-    def create(cls, identity: Keypair) -> "Router":
-        return cls(handle=NetworkingHandle(identity))
+    def create(
+        cls,
+        identity: Keypair,
+        bootstrap_peers: list[str] | None = None,
+    ) -> "Router":
+        return cls(handle=NetworkingHandle(identity, bootstrap_peers or []))
 
     def __init__(self, handle: NetworkingHandle):
         self.topic_routers: dict[str, TopicRouter[CamelCaseModel]] = {}
@@ -207,10 +211,10 @@ class Router:
                 try:
                     logger.trace(f"Sending message on {topic} with payload {data}")
                     await self._net.gossipsub_publish(topic, data)
-                # As a hack, this also catches AllQueuesFull
-                # Need to fix that ASAP.
-                except (NoPeersSubscribedToTopicError, AllQueuesFullError):
+                except NoPeersSubscribedToTopicError:
                     pass
+                except AllQueuesFullError:
+                    logger.warning(f"All peer queues full, dropping message on {topic}")
 
 
 def get_node_id_keypair(

src/exo/shared/apply.py

@@ -4,7 +4,7 @@ from datetime import datetime
 
 from loguru import logger
 
-from exo.shared.types.common import NodeId
+from exo.shared.types.common import MetaInstanceId, NodeId
 from exo.shared.types.events import (
     ChunkGenerated,
     Event,
@@ -12,6 +12,12 @@ from exo.shared.types.events import (
     InputChunkReceived,
     InstanceCreated,
     InstanceDeleted,
+    InstanceRetrying,
+    JacclSideChannelData,
+    JacclSideChannelGathered,
+    MetaInstanceCreated,
+    MetaInstanceDeleted,
+    MetaInstancePlacementFailed,
     NodeDownloadProgress,
     NodeGatheredInfo,
     NodeTimedOut,
@@ -28,9 +34,11 @@ from exo.shared.types.events import (
     TracesCollected,
     TracesMerged,
 )
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.profiling import (
     NodeIdentity,
     NodeNetworkInfo,
+    NodeRdmaCtlStatus,
     NodeThunderboltInfo,
     ThunderboltBridgeStatus,
 )
@@ -47,7 +55,9 @@ from exo.utils.info_gatherer.info_gatherer import (
     MemoryUsage,
     MiscData,
     NodeConfig,
+    NodeDiskUsage,
     NodeNetworkInterfaces,
+    RdmaCtlStatus,
     StaticNodeInformation,
     ThunderboltBridgeInfo,
 )
@@ -63,12 +73,22 @@ def event_apply(event: Event, state: State) -> State:
             | InputChunkReceived()
             | TracesCollected()
             | TracesMerged()
+            | JacclSideChannelData()
+            | JacclSideChannelGathered()
         ):  # Pass-through events that don't modify state
             return state
         case InstanceCreated():
             return apply_instance_created(event, state)
         case InstanceDeleted():
             return apply_instance_deleted(event, state)
+        case InstanceRetrying():
+            return apply_instance_retrying(event, state)
+        case MetaInstanceCreated():
+            return apply_meta_instance_created(event, state)
+        case MetaInstanceDeleted():
+            return apply_meta_instance_deleted(event, state)
+        case MetaInstancePlacementFailed():
+            return apply_meta_instance_placement_failed(event, state)
         case NodeTimedOut():
             return apply_node_timed_out(event, state)
         case NodeDownloadProgress():
@@ -171,20 +191,123 @@ def apply_task_failed(event: TaskFailed, state: State) -> State:
     return state.model_copy(update={"tasks": new_tasks})
 
 
+def _update_meta_instance(
+    state: State, mid: MetaInstanceId, **fields: object
+) -> Mapping[MetaInstanceId, MetaInstance]:
+    mi = state.meta_instances[mid]
+    return {**state.meta_instances, mid: mi.model_copy(update=fields)}
+
+
 def apply_instance_created(event: InstanceCreated, state: State) -> State:
     instance = event.instance
     new_instances: Mapping[InstanceId, Instance] = {
         **state.instances,
         instance.instance_id: instance,
     }
-    return state.model_copy(update={"instances": new_instances})
+    update: dict[str, object] = {"instances": new_instances}
+    # Reset failure tracking when a new instance is created for a meta-instance
+    if instance.meta_instance_id and instance.meta_instance_id in state.meta_instances:
+        mi = state.meta_instances[instance.meta_instance_id]
+        if mi.placement_error is not None or mi.consecutive_failures > 0:
+            update["meta_instances"] = _update_meta_instance(
+                state,
+                instance.meta_instance_id,
+                placement_error=None,
+                consecutive_failures=0,
+            )
+    return state.model_copy(update=update)
 
 
 def apply_instance_deleted(event: InstanceDeleted, state: State) -> State:
+    deleted_instance = state.instances.get(event.instance_id)
     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})
+    update: dict[str, object] = {"instances": new_instances}
+
+    # Track failure on the MetaInstance itself
+    if (
+        event.failure_error
+        and deleted_instance
+        and deleted_instance.meta_instance_id
+        and deleted_instance.meta_instance_id in state.meta_instances
+    ):
+        mid = deleted_instance.meta_instance_id
+        mi = state.meta_instances[mid]
+        update["meta_instances"] = {
+            **state.meta_instances,
+            mid: mi.model_copy(
+                update={
+                    "consecutive_failures": mi.consecutive_failures + 1,
+                    "last_failure_error": event.failure_error,
+                }
+            ),
+        }
+
+    return state.model_copy(update=update)
+
+
+def apply_instance_retrying(event: InstanceRetrying, state: State) -> State:
+    """Runners failed but retry limit not reached — remove runners, keep instance."""
+    instance = state.instances.get(event.instance_id)
+    if instance is None:
+        # Instance was already deleted (e.g. cascade from DeleteMetaInstance).
+        # The InstanceDeleted handler already incremented consecutive_failures
+        # on the MetaInstance, so skipping here avoids double-counting.
+        return state
+
+    # Remove all runners belonging to this instance from state
+    runner_ids_to_remove = set(instance.shard_assignments.node_to_runner.values())
+    new_runners: Mapping[RunnerId, RunnerStatus] = {
+        rid: rs for rid, rs in state.runners.items() if rid not in runner_ids_to_remove
+    }
+
+    update: dict[str, object] = {"runners": new_runners}
+
+    # Increment failure count on the MetaInstance
+    if event.meta_instance_id in state.meta_instances:
+        update["meta_instances"] = _update_meta_instance(
+            state,
+            event.meta_instance_id,
+            consecutive_failures=state.meta_instances[
+                event.meta_instance_id
+            ].consecutive_failures
+            + 1,
+            last_failure_error=event.failure_error,
+        )
+
+    return state.model_copy(update=update)
+
+
+def apply_meta_instance_created(event: MetaInstanceCreated, state: State) -> State:
+    new_meta: Mapping[MetaInstanceId, MetaInstance] = {
+        **state.meta_instances,
+        event.meta_instance.meta_instance_id: event.meta_instance,
+    }
+    return state.model_copy(update={"meta_instances": new_meta})
+
+
+def apply_meta_instance_deleted(event: MetaInstanceDeleted, state: State) -> State:
+    new_meta: Mapping[MetaInstanceId, MetaInstance] = {
+        mid: mi
+        for mid, mi in state.meta_instances.items()
+        if mid != event.meta_instance_id
+    }
+    return state.model_copy(update={"meta_instances": new_meta})
+
+
+def apply_meta_instance_placement_failed(
+    event: MetaInstancePlacementFailed, state: State
+) -> State:
+    if event.meta_instance_id not in state.meta_instances:
+        return state
+    return state.model_copy(
+        update={
+            "meta_instances": _update_meta_instance(
+                state, event.meta_instance_id, placement_error=event.reason
+            )
+        }
+    )
 
 
 def apply_runner_status_updated(event: RunnerStatusUpdated, state: State) -> State:
@@ -223,6 +346,9 @@ def apply_node_timed_out(event: NodeTimedOut, state: State) -> State:
     node_memory = {
         key: value for key, value in state.node_memory.items() if key != event.node_id
     }
+    node_disk = {
+        key: value for key, value in state.node_disk.items() if key != event.node_id
+    }
     node_system = {
         key: value for key, value in state.node_system.items() if key != event.node_id
     }
@@ -239,6 +365,9 @@ def apply_node_timed_out(event: NodeTimedOut, state: State) -> State:
         for key, value in state.node_thunderbolt_bridge.items()
         if key != event.node_id
     }
+    node_rdma_ctl = {
+        key: value for key, value in state.node_rdma_ctl.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 = (
@@ -256,10 +385,12 @@ def apply_node_timed_out(event: NodeTimedOut, state: State) -> State:
             "last_seen": last_seen,
             "node_identities": node_identities,
             "node_memory": node_memory,
+            "node_disk": node_disk,
             "node_system": node_system,
             "node_network": node_network,
             "node_thunderbolt": node_thunderbolt,
             "node_thunderbolt_bridge": node_thunderbolt_bridge,
+            "node_rdma_ctl": node_rdma_ctl,
             "thunderbolt_bridge_cycles": thunderbolt_bridge_cycles,
         }
     )
@@ -288,6 +419,8 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:
             update["node_memory"] = {**state.node_memory, event.node_id: info.memory}
         case MemoryUsage():
             update["node_memory"] = {**state.node_memory, event.node_id: info}
+        case NodeDiskUsage():
+            update["node_disk"] = {**state.node_disk, event.node_id: info.disk_usage}
         case NodeConfig():
             pass
         case MiscData():
@@ -302,7 +435,12 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:
         case StaticNodeInformation():
             current_identity = state.node_identities.get(event.node_id, NodeIdentity())
             new_identity = current_identity.model_copy(
-                update={"model_id": info.model, "chip_id": info.chip}
+                update={
+                    "model_id": info.model,
+                    "chip_id": info.chip,
+                    "os_version": info.os_version,
+                    "os_build_version": info.os_build_version,
+                }
             )
             update["node_identities"] = {
                 **state.node_identities,
@@ -354,6 +492,11 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:
                         new_tb_bridge, state.node_network
                     )
                 )
+        case RdmaCtlStatus():
+            update["node_rdma_ctl"] = {
+                **state.node_rdma_ctl,
+                event.node_id: NodeRdmaCtlStatus(enabled=info.enabled),
+            }
 
     return state.model_copy(update=update)
 

src/exo/shared/constants.py

@@ -43,7 +43,8 @@ DASHBOARD_DIR = (
 )
 
 # Log files (data/logs or cache)
-EXO_LOG = EXO_CACHE_HOME / "exo.log"
+EXO_LOG_DIR = EXO_CACHE_HOME / "exo_log"
+EXO_LOG = EXO_LOG_DIR / "exo.log"
 EXO_TEST_LOG = EXO_CACHE_HOME / "exo_test.log"
 
 # Identity (config)
@@ -60,6 +61,7 @@ EXO_MAX_CHUNK_SIZE = 512 * 1024
 
 EXO_CUSTOM_MODEL_CARDS_DIR = EXO_DATA_HOME / "custom_model_cards"
 
+EXO_EVENT_LOG_DIR = EXO_DATA_HOME / "event_log"
 EXO_IMAGE_CACHE_DIR = EXO_CACHE_HOME / "images"
 EXO_TRACING_CACHE_DIR = EXO_CACHE_HOME / "traces"
 

src/exo/shared/logging.py

@@ -1,11 +1,30 @@
 import logging
 import sys
+from collections.abc import Iterator
 from pathlib import Path
 
+import zstandard
 from hypercorn import Config
 from hypercorn.logging import Logger as HypercornLogger
 from loguru import logger
 
+_MAX_LOG_ARCHIVES = 5
+
+
+def _zstd_compress(filepath: str) -> None:
+    source = Path(filepath)
+    dest = source.with_suffix(source.suffix + ".zst")
+    cctx = zstandard.ZstdCompressor()
+    with open(source, "rb") as f_in, open(dest, "wb") as f_out:
+        cctx.copy_stream(f_in, f_out)
+    source.unlink()
+
+
+def _once_then_never() -> Iterator[bool]:
+    yield True
+    while True:
+        yield False
+
 
 class InterceptLogger(HypercornLogger):
     def __init__(self, config: Config):
@@ -53,13 +72,16 @@ def logger_setup(log_file: Path | None, verbosity: int = 0):
             enqueue=True,
         )
     if log_file:
+        rotate_once = _once_then_never()
         logger.add(
             log_file,
             format="[ {time:YYYY-MM-DD HH:mm:ss.SSS} | {level: <8} | {name}:{function}:{line} ] {message}",
             level="INFO",
             colorize=False,
             enqueue=True,
-            rotation="1 week",
+            rotation=lambda _, __: next(rotate_once),
+            retention=_MAX_LOG_ARCHIVES,
+            compression=_zstd_compress,
         )
 
 

src/exo/shared/models/model_cards.py

@@ -30,11 +30,11 @@ from exo.utils.pydantic_ext import CamelCaseModel
 # kinda ugly...
 # TODO: load search path from config.toml
 _custom_cards_dir = Path(str(EXO_CUSTOM_MODEL_CARDS_DIR))
-_csp = [Path(RESOURCES_DIR) / "inference_model_cards", _custom_cards_dir]
-if EXO_ENABLE_IMAGE_MODELS:
-    _csp.append(Path(RESOURCES_DIR) / "image_model_cards")
-
-CARD_SEARCH_PATH = _csp
+CARD_SEARCH_PATH = [
+    Path(RESOURCES_DIR) / "inference_model_cards",
+    Path(RESOURCES_DIR) / "image_model_cards",
+    _custom_cards_dir,
+]
 
 _card_cache: dict[ModelId, "ModelCard"] = {}
 
@@ -49,10 +49,16 @@ async def _refresh_card_cache():
                 pass
 
 
+def _is_image_card(card: "ModelCard") -> bool:
+    return any(t in (ModelTask.TextToImage, ModelTask.ImageToImage) for t in card.tasks)
+
+
 async def get_model_cards() -> list["ModelCard"]:
     if len(_card_cache) == 0:
         await _refresh_card_cache()
-    return list(_card_cache.values())
+    if EXO_ENABLE_IMAGE_MODELS:
+        return list(_card_cache.values())
+    return [c for c in _card_cache.values() if not _is_image_card(c)]
 
 
 class ModelTask(str, Enum):
@@ -183,6 +189,7 @@ class ConfigData(BaseModel):
             ["MiniMaxM2ForCausalLM"],
             ["LlamaForCausalLM"],
             ["GptOssForCausalLM"],
+            ["Step3p5ForCausalLM"],
         ]
 
     @model_validator(mode="before")

src/exo/shared/types/api.py

@@ -1,12 +1,12 @@
 import time
 from collections.abc import Generator
 from typing import Annotated, Any, Literal
+from uuid import uuid4
 
-from pydantic import BaseModel, Field, field_validator
-from pydantic_core import PydanticUseDefault
+from pydantic import BaseModel, Field
 
 from exo.shared.models.model_cards import ModelCard, ModelId
-from exo.shared.types.common import CommandId, NodeId
+from exo.shared.types.common import CommandId, MetaInstanceId, NodeId
 from exo.shared.types.memory import Memory
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding, ShardMetadata
@@ -60,6 +60,7 @@ class ChatCompletionMessageText(BaseModel):
 
 
 class ToolCallItem(BaseModel):
+    id: str = Field(default_factory=lambda: str(uuid4()))
     name: str
     arguments: str
 
@@ -197,6 +198,7 @@ class ChatCompletionRequest(BaseModel):
     top_p: float | None = None
     top_k: int | None = None
     tools: list[dict[str, Any]] | None = None
+    enable_thinking: bool | None = None
     tool_choice: str | dict[str, Any] | None = None
     parallel_tool_calls: bool | None = None
     user: str | None = None
@@ -225,13 +227,6 @@ class PlaceInstanceParams(BaseModel):
     instance_meta: InstanceMeta = InstanceMeta.MlxRing
     min_nodes: int = 1
 
-    @field_validator("sharding", "instance_meta", mode="plain")
-    @classmethod
-    def use_default(cls, v: object):
-        if not v or not isinstance(v, (Sharding, InstanceMeta)):
-            raise PydanticUseDefault()
-        return v
-
 
 class CreateInstanceParams(BaseModel):
     instance: Instance
@@ -267,6 +262,26 @@ class DeleteInstanceResponse(BaseModel):
     instance_id: InstanceId
 
 
+class CreateMetaInstanceParams(BaseModel):
+    model_id: ModelId
+    sharding: Sharding = Sharding.Pipeline
+    instance_meta: InstanceMeta = InstanceMeta.MlxRing
+    min_nodes: int = 1
+    node_ids: list[NodeId] | None = None
+
+
+class CreateMetaInstanceResponse(BaseModel):
+    message: str
+    command_id: CommandId
+    meta_instance_id: MetaInstanceId
+
+
+class DeleteMetaInstanceResponse(BaseModel):
+    message: str
+    command_id: CommandId
+    meta_instance_id: MetaInstanceId
+
+
 class AdvancedImageParams(BaseModel):
     seed: Annotated[int, Field(ge=0)] | None = None
     num_inference_steps: Annotated[int, Field(ge=1, le=100)] | None = None

src/exo/shared/types/commands.py

@@ -6,7 +6,8 @@ from exo.shared.types.api import (
     ImageGenerationTaskParams,
 )
 from exo.shared.types.chunks import InputImageChunk
-from exo.shared.types.common import CommandId, NodeId
+from exo.shared.types.common import CommandId, MetaInstanceId, NodeId
+from exo.shared.types.meta_instance import MetaInstance
 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, ShardMetadata
@@ -48,6 +49,18 @@ class DeleteInstance(BaseCommand):
     instance_id: InstanceId
 
 
+class CreateMetaInstance(BaseCommand):
+    meta_instance: MetaInstance
+
+
+class DeleteMetaInstance(BaseCommand):
+    meta_instance_id: MetaInstanceId
+
+
+class TaskCancelled(BaseCommand):
+    cancelled_command_id: CommandId
+
+
 class TaskFinished(BaseCommand):
     finished_command_id: CommandId
 
@@ -89,6 +102,9 @@ Command = (
     | PlaceInstance
     | CreateInstance
     | DeleteInstance
+    | CreateMetaInstance
+    | DeleteMetaInstance
+    | TaskCancelled
     | TaskFinished
     | SendInputChunk
 )

src/exo/shared/types/common.py

@@ -42,6 +42,10 @@ class CommandId(Id):
     pass
 
 
+class MetaInstanceId(Id):
+    """Identifier for a MetaInstance."""
+
+
 class Host(CamelCaseModel):
     ip: str
     port: int

src/exo/shared/types/events.py

@@ -1,11 +1,14 @@
+import base64
+from collections.abc import Mapping
 from datetime import datetime
-from typing import final
+from typing import Annotated, final
 
-from pydantic import Field
+from pydantic import BeforeValidator, Field, PlainSerializer
 
 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.common import CommandId, Id, MetaInstanceId, NodeId, SessionId
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.tasks import Task, TaskId, TaskStatus
 from exo.shared.types.worker.downloads import DownloadProgress
 from exo.shared.types.worker.instances import Instance, InstanceId
@@ -14,6 +17,28 @@ from exo.utils.info_gatherer.info_gatherer import GatheredInfo
 from exo.utils.pydantic_ext import CamelCaseModel, FrozenModel, TaggedModel
 
 
+def _decode_base64_bytes(v: bytes | str) -> bytes:
+    if isinstance(v, bytes):
+        return v
+    return base64.b64decode(v)
+
+
+def _encode_base64_bytes(v: bytes) -> str:
+    return base64.b64encode(v).decode("ascii")
+
+
+Base64Bytes = Annotated[
+    bytes,
+    BeforeValidator(_decode_base64_bytes),
+    PlainSerializer(_encode_base64_bytes, return_type=str),
+]
+"""bytes that serialize to/from base64 strings in JSON.
+
+Needed because TaggedModel's wrap validator converts JSON→Python validation
+context, which breaks strict-mode bytes deserialization from JSON strings.
+"""
+
+
 class EventId(Id):
     """
     Newtype around `ID`
@@ -66,6 +91,30 @@ class InstanceCreated(BaseEvent):
 
 class InstanceDeleted(BaseEvent):
     instance_id: InstanceId
+    failure_error: str | None = None
+
+
+class MetaInstanceCreated(BaseEvent):
+    meta_instance: MetaInstance
+
+
+class MetaInstanceDeleted(BaseEvent):
+    meta_instance_id: MetaInstanceId
+
+
+@final
+class MetaInstancePlacementFailed(BaseEvent):
+    meta_instance_id: MetaInstanceId
+    reason: str
+
+
+@final
+class InstanceRetrying(BaseEvent):
+    """Runners failed but retry count is below the limit — restart runners, keep instance."""
+
+    instance_id: InstanceId
+    meta_instance_id: MetaInstanceId
+    failure_error: str
 
 
 class RunnerStatusUpdated(BaseEvent):
@@ -132,6 +181,25 @@ class TracesMerged(BaseEvent):
     traces: list[TraceEventData]
 
 
+@final
+class JacclSideChannelData(BaseEvent):
+    """A runner's local contribution to a JACCL SideChannel all_gather round."""
+
+    instance_id: InstanceId
+    runner_id: RunnerId
+    sequence: int
+    data: Base64Bytes
+
+
+@final
+class JacclSideChannelGathered(BaseEvent):
+    """Gathered result of a JACCL SideChannel all_gather round."""
+
+    instance_id: InstanceId
+    sequence: int
+    gathered_data: Mapping[RunnerId, Base64Bytes]
+
+
 Event = (
     TestEvent
     | TaskCreated
@@ -141,6 +209,10 @@ Event = (
     | TaskAcknowledged
     | InstanceCreated
     | InstanceDeleted
+    | InstanceRetrying
+    | MetaInstanceCreated
+    | MetaInstanceDeleted
+    | MetaInstancePlacementFailed
     | RunnerStatusUpdated
     | RunnerDeleted
     | NodeTimedOut
@@ -152,6 +224,8 @@ Event = (
     | TopologyEdgeDeleted
     | TracesCollected
     | TracesMerged
+    | JacclSideChannelData
+    | JacclSideChannelGathered
 )
 
 

src/exo/shared/types/meta_instance.py

@@ -0,0 +1,25 @@
+from typing import final
+
+from pydantic import Field
+
+from exo.shared.models.model_cards import ModelId
+from exo.shared.types.common import MetaInstanceId, NodeId
+from exo.shared.types.worker.instances import InstanceMeta
+from exo.shared.types.worker.shards import Sharding
+from exo.utils.pydantic_ext import FrozenModel
+
+
+@final
+class MetaInstance(FrozenModel):
+    """Declarative constraint: ensure an instance matching these parameters always exists."""
+
+    meta_instance_id: MetaInstanceId = Field(default_factory=MetaInstanceId)
+    model_id: ModelId
+    sharding: Sharding = Sharding.Pipeline
+    instance_meta: InstanceMeta = InstanceMeta.MlxRing
+    min_nodes: int = 1
+    node_ids: list[NodeId] | None = None
+    # Failure tracking
+    placement_error: str | None = None
+    consecutive_failures: int = 0
+    last_failure_error: str | None = None

src/exo/shared/types/profiling.py

@@ -1,4 +1,6 @@
+import shutil
 from collections.abc import Sequence
+from pathlib import Path
 from typing import Literal, Self
 
 import psutil
@@ -38,6 +40,22 @@ class MemoryUsage(CamelCaseModel):
         )
 
 
+class DiskUsage(CamelCaseModel):
+    """Disk space usage for the models directory."""
+
+    total: Memory
+    available: Memory
+
+    @classmethod
+    def from_path(cls, path: Path) -> Self:
+        """Get disk usage stats for the partition containing path."""
+        total, _used, free = shutil.disk_usage(path)
+        return cls(
+            total=Memory.from_bytes(total),
+            available=Memory.from_bytes(free),
+        )
+
+
 class SystemPerformanceProfile(CamelCaseModel):
     # TODO: flops_fp16: float
 
@@ -63,6 +81,8 @@ class NodeIdentity(CamelCaseModel):
     model_id: str = "Unknown"
     chip_id: str = "Unknown"
     friendly_name: str = "Unknown"
+    os_version: str = "Unknown"
+    os_build_version: str = "Unknown"
 
 
 class NodeNetworkInfo(CamelCaseModel):
@@ -77,6 +97,12 @@ class NodeThunderboltInfo(CamelCaseModel):
     interfaces: Sequence[ThunderboltIdentifier] = []
 
 
+class NodeRdmaCtlStatus(CamelCaseModel):
+    """Whether RDMA is enabled on this node (via rdma_ctl)."""
+
+    enabled: bool
+
+
 class ThunderboltBridgeStatus(CamelCaseModel):
     """Whether the Thunderbolt Bridge network service is enabled on this node."""
 

src/exo/shared/types/state.py

@@ -6,11 +6,14 @@ from pydantic import ConfigDict, Field, field_serializer, field_validator
 from pydantic.alias_generators import to_camel
 
 from exo.shared.topology import Topology, TopologySnapshot
-from exo.shared.types.common import NodeId
+from exo.shared.types.common import MetaInstanceId, NodeId
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.profiling import (
+    DiskUsage,
     MemoryUsage,
     NodeIdentity,
     NodeNetworkInfo,
+    NodeRdmaCtlStatus,
     NodeThunderboltInfo,
     SystemPerformanceProfile,
     ThunderboltBridgeStatus,
@@ -39,6 +42,7 @@ class State(CamelCaseModel):
         arbitrary_types_allowed=True,
     )
     instances: Mapping[InstanceId, Instance] = {}
+    meta_instances: Mapping[MetaInstanceId, MetaInstance] = {}
     runners: Mapping[RunnerId, RunnerStatus] = {}
     downloads: Mapping[NodeId, Sequence[DownloadProgress]] = {}
     tasks: Mapping[TaskId, Task] = {}
@@ -49,10 +53,12 @@ class State(CamelCaseModel):
     # Granular node state mappings (update independently at different frequencies)
     node_identities: Mapping[NodeId, NodeIdentity] = {}
     node_memory: Mapping[NodeId, MemoryUsage] = {}
+    node_disk: Mapping[NodeId, DiskUsage] = {}
     node_system: Mapping[NodeId, SystemPerformanceProfile] = {}
     node_network: Mapping[NodeId, NodeNetworkInfo] = {}
     node_thunderbolt: Mapping[NodeId, NodeThunderboltInfo] = {}
     node_thunderbolt_bridge: Mapping[NodeId, ThunderboltBridgeStatus] = {}
+    node_rdma_ctl: Mapping[NodeId, NodeRdmaCtlStatus] = {}
 
     # Detected cycles where all nodes have Thunderbolt bridge enabled (>2 nodes)
     thunderbolt_bridge_cycles: Sequence[Sequence[NodeId]] = []

src/exo/shared/types/tasks.py

@@ -24,6 +24,7 @@ class TaskStatus(str, Enum):
     Complete = "Complete"
     TimedOut = "TimedOut"
     Failed = "Failed"
+    Cancelled = "Cancelled"
 
 
 class BaseTask(TaggedModel):
@@ -60,6 +61,11 @@ class TextGeneration(BaseTask):  # emitted by Master
     error_message: str | None = Field(default=None)
 
 
+class CancelTask(BaseTask):
+    cancelled_task_id: TaskId
+    runner_id: RunnerId
+
+
 class ImageGeneration(BaseTask):  # emitted by Master
     command_id: CommandId
     task_params: ImageGenerationTaskParams
@@ -87,6 +93,7 @@ Task = (
     | LoadModel
     | StartWarmup
     | TextGeneration
+    | CancelTask
     | ImageGeneration
     | ImageEdits
     | Shutdown