fix: pass _cancel_sender in RunnerSupervisor test helper

After merging main (api cancellation #1276), the RunnerSupervisor dataclass requires a _cancel_sender field. Update the test helper to create and pass this channel. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ci: retrigger CI (darwin runner back)
2026-02-16 10:25:22 -08:00 · 2026-02-16 10:01:27 -08:00 · 2026-02-16 10:01:27 -08:00 · 2026-02-16 10:01:27 -08:00 · 2026-02-16 10:01:27 -08:00 · 2026-02-16 10:01:27 -08:00
98 changed files with 4384 additions and 997 deletions
@@ -0,0 +1,15 @@
+.venv/
+.direnv/
+target/
+.git/
+.idea/
+.pytest_cache/
+.ruff_cache/
+dashboard/node_modules/
+dashboard/.svelte-kit/
+dashboard/build/
+dist/
+*.pdb
+**/__pycache__
+**/.DS_Store
+.mlx_typings/
@@ -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"
@@ -0,0 +1,44 @@
+name: e2e-tests
+
+on:
+  push:
+    branches:
+      - e2e-tests
+  pull_request:
+    branches:
+      - staging
+      - main
+
+jobs:
+  e2e:
+    runs-on: ubuntu-latest
+    timeout-minutes: 45
+    steps:
+      - name: Free up disk space
+        run: |
+          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc \
+            /opt/hostedtoolcache /usr/local/share/boost /usr/share/swift \
+            /opt/microsoft /opt/az
+          docker system prune -af
+          df -h /
+
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          lfs: false
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Build E2E image with cache
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: e2e/Dockerfile
+          tags: exo-e2e:latest
+          load: true
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
+
+      - name: Run E2E tests
+        run: python3 e2e/run_all.py
@@ -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 }}
@@ -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: ...
@@ -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.
@@ -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"
@@ -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).
@@ -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:
@@ -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 {
@@ -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
@@ -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?
@@ -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 {
@@ -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 [] }
@@ -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:
@@ -0,0 +1 @@
+collect_ignore = ["tests/start_distributed_test.py"]
@@ -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>

@@ -225,6 +225,7 @@
  }

  function handleDeleteClick(messageId: string) {
+    if (loading) return;
    deleteConfirmId = messageId;
  }

@@ -255,7 +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'
@@ -317,9 +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
@@ -751,8 +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"
@@ -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
@@ -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",
@@ -89,6 +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"
@@ -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 });
@@ -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(
@@ -244,7 +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}
@@ -255,13 +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
@@ -360,6 +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
@@ -384,7 +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>

@@ -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,
@@ -383,7 +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);
    }

@@ -427,13 +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) -
@@ -447,6 +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)) {
@@ -604,6 +669,7 @@
        families={uniqueFamilies}
        {selectedFamily}
        {hasFavorites}
+        hasRecents={hasRecentsTab}
        onSelect={(family) => (selectedFamily = family)}
      />

@@ -711,6 +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"
@@ -742,6 +847,7 @@
              isFavorite={favorites.has(group.id)}
              {selectedModelId}
              {canModelFit}
+              {getModelFitStatus}
              onToggleExpand={() => toggleGroupExpanded(group.id)}
              onSelectModel={handleSelect}
              {onToggleFavorite}
@@ -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})` : ""}`,
+            );
        }
      }
    });
@@ -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 {
@@ -225,6 +230,19 @@ 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,
@@ -278,6 +296,7 @@ export interface Conversation {
  modelId: string | null;
  sharding: string | null;
  instanceType: string | null;
+  enableThinking: boolean | null;
 }

 const STORAGE_KEY = "exo-conversations";
@@ -425,6 +444,7 @@ function transformTopology(
      },
      last_macmon_update: Date.now() / 1000,
      friendly_name: identity?.friendlyName,
+      os_version: identity?.osVersion,
    };
  }

@@ -437,6 +457,8 @@ 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 +466,23 @@ 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 +525,32 @@ 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);
+  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 +606,7 @@ class AppStore {
          modelId: conversation.modelId ?? null,
          sharding: conversation.sharding ?? null,
          instanceType: conversation.instanceType ?? null,
+          enableThinking: conversation.enableThinking ?? null,
        }));
      }
    } catch (error) {
@@ -740,6 +796,7 @@ class AppStore {
      modelId: derivedModelId,
      sharding: derivedSharding,
      instanceType: derivedInstanceType,
+      enableThinking: null,
    };

    this.conversations.unshift(conversation);
@@ -765,6 +822,7 @@ 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;
@@ -1206,6 +1264,15 @@ class AppStore {
      if (data.downloads) {
        this.downloads = data.downloads;
      }
+      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 +1936,11 @@ 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 +2119,7 @@ class AppStore {
      textContent?: string;
      preview?: string;
    }[],
+    enableThinking?: boolean | null,
  ): Promise<void> {
    if ((!content.trim() && (!files || files.length === 0)) || this.isLoading)
      return;
@@ -2194,6 +2267,9 @@ class AppStore {
          stream: true,
          logprobs: true,
          top_logprobs: 5,
+          ...(enableThinking != null && {
+            enable_thinking: enableThinking,
+          }),
        }),
      });

@@ -2852,6 +2928,18 @@ 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
   */
@@ -2958,12 +3046,14 @@ export const topologyData = () => appStore.topologyData;
 export const instances = () => appStore.instances;
 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 +3069,8 @@ 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 +3113,8 @@ 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 +3131,12 @@ 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;

@@ -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 = $state<RecentEntry[]>([]);
+
+  constructor() {
+    if (browser) {
+      this.loadFromStorage();
+    }
+  }
+
+  private loadFromStorage() {
+    try {
+      const stored = localStorage.getItem(RECENTS_KEY);
+      if (stored) {
+        const parsed = JSON.parse(stored) as RecentEntry[];
+        this.recents = parsed;
+      }
+    } catch (error) {
+      console.error("Failed to load recent models:", error);
+    }
+  }
+
+  private saveToStorage() {
+    try {
+      localStorage.setItem(RECENTS_KEY, JSON.stringify(this.recents));
+    } catch (error) {
+      console.error("Failed to save recent models:", error);
+    }
+  }
+
+  recordLaunch(modelId: string) {
+    // Remove existing entry for this model (if any) to move it to top
+    const filtered = this.recents.filter((r) => r.modelId !== modelId);
+    // Prepend new entry
+    const next = [{ modelId, launchedAt: Date.now() }, ...filtered];
+    // Cap at max
+    this.recents = next.slice(0, MAX_RECENT_MODELS);
+    this.saveToStorage();
+  }
+
+  getRecentModelIds(): string[] {
+    return this.recents.map((r) => r.modelId);
+  }
+
+  hasAny(): boolean {
+    return this.recents.length > 0;
+  }
+
+  clearAll() {
+    this.recents = [];
+    this.saveToStorage();
+  }
+}
+
+export const recentsStore = new RecentsStore();
+
+export const hasRecents = () => recentsStore.hasAny();
+export const getRecentModelIds = () => recentsStore.getRecentModelIds();
+export const getRecentEntries = () => recentsStore.recents;
+export const recordRecentLaunch = (modelId: string) =>
+  recentsStore.recordLaunch(modelId);
+export const clearRecents = () => recentsStore.clearAll();
@@ -12,6 +12,11 @@
    toggleFavorite,
    getFavoritesSet,
  } from "$lib/stores/favorites.svelte";
+  import {
+    hasRecents,
+    getRecentModelIds,
+    recordRecentLaunch,
+  } from "$lib/stores/recents.svelte";
  import {
    hasStartedChat,
    isTopologyMinimized,
@@ -37,8 +42,11 @@
    toggleTopologyOnlyMode,
    chatSidebarVisible,
    toggleChatSidebarVisible,
+    nodeThunderbolt,
+    nodeRdmaCtl,
    thunderboltBridgeCycles,
    nodeThunderboltBridge,
+    nodeIdentities,
    type DownloadProgress,
    type PlacementPreview,
  } from "$lib/stores/app.svelte";
@@ -62,8 +70,50 @@
  const sidebarVisible = $derived(chatSidebarVisible());
  const tbBridgeCycles = $derived(thunderboltBridgeCycles());
  const tbBridgeData = $derived(nodeThunderboltBridge());
+  const identitiesData = $derived(nodeIdentities());
+  const tbIdentifiers = $derived(nodeThunderbolt());
+  const rdmaCtlData = $derived(nodeRdmaCtl());
  const nodeFilter = $derived(previewNodeFilter());

+  // Detect macOS version mismatches across cluster nodes
+  const macosVersionMismatch = $derived.by(() => {
+    if (!identitiesData) return null;
+    const entries = Object.entries(identitiesData);
+    // Filter to macOS nodes (version starts with a digit, e.g. "15.3")
+    const macosNodes = entries.filter(([_, id]) => {
+      const v = id.osVersion;
+      return v && v !== "Unknown" && /^\d/.test(v);
+    });
+    if (macosNodes.length < 2) return null;
+    // Compare on buildVersion for precise mismatch detection
+    const buildVersions = new Set(
+      macosNodes.map(([_, id]) => id.osBuildVersion ?? id.osVersion),
+    );
+    if (buildVersions.size <= 1) return null;
+    return macosNodes.map(([nodeId, id]) => ({
+      nodeId,
+      friendlyName: getNodeName(nodeId),
+      version: id.osVersion!,
+      buildVersion: id.osBuildVersion ?? "Unknown",
+    }));
+  });
+
+  // Detect TB5 nodes where RDMA is not enabled
+  const tb5WithoutRdma = $derived.by(() => {
+    const rdmaCtl = rdmaCtlData;
+    if (!rdmaCtl) return false;
+    const ids = tbIdentifiers;
+    if (!ids) return false;
+    // Find nodes with TB5 hardware (any TB interface)
+    const tb5NodeIds = Object.entries(ids)
+      .filter(([_, node]) => node.interfaces.length > 0)
+      .map(([id]) => id);
+    if (tb5NodeIds.length < 2) return false;
+    // At least one TB5 node has RDMA disabled
+    return tb5NodeIds.some((id) => rdmaCtl[id]?.enabled !== true);
+  });
+  let tb5InfoDismissed = $state(false);
+
  // Helper to get friendly node name from node ID
  function getNodeName(nodeId: string): string {
    const node = data?.nodes?.[nodeId];
@@ -96,6 +146,12 @@
    }
  }

+  // Warning icon SVG path (reused across warning snippets)
+  const warningIconPath =
+    "M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z";
+  const infoIconPath =
+    "M13 16h-1v-4h-1m1-4h.01M21 12a9 9 0 11-18 0 9 9 0 0118 0z";
+
  let mounted = $state(false);

  // Instance launch state
@@ -113,6 +169,10 @@
      capabilities?: string[];
    }>
  >([]);
+  type ModelMemoryFitStatus =
+    | "fits_now"
+    | "fits_cluster_capacity"
+    | "too_large";

  // Model tasks lookup for ChatForm - maps both short IDs and full HuggingFace IDs
  const modelTasks = $derived(() => {
@@ -130,6 +190,19 @@
    return tasks;
  });

+  const modelCapabilities = $derived(() => {
+    const caps: Record<string, string[]> = {};
+    for (const model of models) {
+      if (model.capabilities && model.capabilities.length > 0) {
+        caps[model.id] = model.capabilities;
+        if (model.hugging_face_id) {
+          caps[model.hugging_face_id] = model.capabilities;
+        }
+      }
+    }
+    return caps;
+  });
+
  // Helper to check if a model supports image generation
  function modelSupportsImageGeneration(modelId: string): boolean {
    const model = models.find(
@@ -228,6 +301,10 @@
  // Favorites state (reactive)
  const favoritesSet = $derived(getFavoritesSet());

+  // Recent models state (reactive)
+  const recentModelIds = $derived(getRecentModelIds());
+  const showRecentsTab = $derived(hasRecents());
+
  // Slider dragging state
  let isDraggingSlider = $state(false);
  let sliderTrackElement: HTMLDivElement | null = $state(null);
@@ -450,14 +527,41 @@
    );
  });

+  // Calculate total memory in the cluster (in GB)
+  const clusterTotalMemoryGB = $derived(() => {
+    if (!data) return 0;
+    return (
+      Object.values(data.nodes).reduce((acc, n) => {
+        const total =
+          n.macmon_info?.memory?.ram_total ?? n.system_info?.memory ?? 0;
+        return acc + total;
+      }, 0) /
+      (1024 * 1024 * 1024)
+    );
+  });
+
+  function getModelMemoryFitStatus(model: {
+    id: string;
+    name?: string;
+    storage_size_megabytes?: number;
+  }): ModelMemoryFitStatus {
+    const modelSizeGB = getModelSizeGB(model);
+    if (modelSizeGB <= availableMemoryGB()) {
+      return "fits_now";
+    }
+    if (modelSizeGB <= clusterTotalMemoryGB()) {
+      return "fits_cluster_capacity";
+    }
+    return "too_large";
+  }
+
  // Check if a model has enough memory to run
  function hasEnoughMemory(model: {
    id: string;
    name?: string;
    storage_size_megabytes?: number;
  }): boolean {
-    const modelSizeGB = getModelSizeGB(model);
-    return modelSizeGB <= availableMemoryGB();
+    return getModelMemoryFitStatus(model) === "fits_now";
  }

  // Sorted models for dropdown - biggest first, unrunnable at the end
@@ -630,6 +734,9 @@
        // Always auto-select the newly launched model so the user chats to what they just launched
        setSelectedChatModel(modelId);

+        // Record the launch in recent models history
+        recordRecentLaunch(modelId);
+
        // Scroll to the bottom of instances container to show the new instance
        // Use multiple attempts to ensure DOM has updated with the new instance
        const scrollToBottom = () => {
@@ -1657,6 +1764,249 @@
  });
 </script>

+{#snippet clusterWarnings()}
+  {#if tbBridgeCycles.length > 0 || macosVersionMismatch || (tb5WithoutRdma && !tb5InfoDismissed)}
+    <div class="absolute top-4 left-4 flex flex-col gap-2 z-40">
+      {#if tbBridgeCycles.length > 0}
+        {@const cycle = tbBridgeCycles[0]}
+        {@const serviceName = getTbBridgeServiceName(cycle)}
+        {@const disableCmd = `sudo networksetup -setnetworkserviceenabled "${serviceName}" off`}
+        <div class="group relative" role="alert">
+          <div
+            class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+          >
+            <svg
+              class="w-5 h-5 text-yellow-400 flex-shrink-0"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+              stroke-width="2"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                d={warningIconPath}
+              />
+            </svg>
+            <span class="text-sm font-mono text-yellow-200">
+              THUNDERBOLT BRIDGE CYCLE DETECTED
+            </span>
+          </div>
+
+          <!-- Tooltip on hover -->
+          <div
+            class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
+          >
+            <p class="text-xs text-white/80 mb-2">
+              A network routing cycle was detected between nodes connected via
+              Thunderbolt Bridge. This can cause connectivity issues.
+            </p>
+            <p class="text-xs text-white/60 mb-2">
+              <span class="text-yellow-300">Affected nodes:</span>
+              {cycle.map(getNodeName).join(" → ")}
+            </p>
+            <p class="text-xs text-white/60 mb-1">
+              <span class="text-yellow-300">To fix:</span> Disable the Thunderbolt
+              Bridge on one of the affected nodes:
+            </p>
+            <button
+              type="button"
+              onclick={() => copyToClipboard(disableCmd)}
+              class="w-full flex items-center gap-2 text-[10px] font-mono bg-exo-black/60 px-2 py-1.5 rounded text-exo-yellow break-all text-left hover:bg-exo-black/80 transition-colors cursor-pointer group/copy"
+              title="Click to copy"
+            >
+              <span class="flex-1">{disableCmd}</span>
+              <svg
+                class="w-3.5 h-3.5 flex-shrink-0 text-white/40 group-hover/copy:text-exo-yellow transition-colors"
+                fill="none"
+                viewBox="0 0 24 24"
+                stroke="currentColor"
+                stroke-width="2"
+              >
+                {#if copiedCommand}
+                  <path
+                    stroke-linecap="round"
+                    stroke-linejoin="round"
+                    d="M5 13l4 4L19 7"
+                  />
+                {:else}
+                  <path
+                    stroke-linecap="round"
+                    stroke-linejoin="round"
+                    d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z"
+                  />
+                {/if}
+              </svg>
+            </button>
+          </div>
+        </div>
+      {/if}
+
+      {#if macosVersionMismatch}
+        <div class="group relative" role="alert">
+          <div
+            class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+          >
+            <svg
+              class="w-5 h-5 text-yellow-400 flex-shrink-0"
+              fill="none"
+              viewBox="0 0 24 24"
+              stroke="currentColor"
+              stroke-width="2"
+            >
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                d={warningIconPath}
+              />
+            </svg>
+            <span class="text-sm font-mono text-yellow-200">
+              INCOMPATIBLE macOS VERSIONS
+            </span>
+          </div>
+
+          <!-- Tooltip on hover -->
+          <div
+            class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
+          >
+            <p class="text-xs text-white/80 mb-2">
+              Nodes in this cluster are running different macOS versions. This
+              may cause inference compatibility issues.
+            </p>
+            <div class="text-xs text-white/60 mb-2">
+              <span class="text-yellow-300">Node versions:</span>
+              {#each macosVersionMismatch as node}
+                <div class="ml-2">
+                  {node.friendlyName} — macOS {node.version} ({node.buildVersion})
+                </div>
+              {/each}
+            </div>
+            <p class="text-xs text-white/60">
+              <span class="text-yellow-300">Suggested action:</span> Update all nodes
+              to the same macOS version for best compatibility.
+            </p>
+          </div>
+        </div>
+      {/if}
+
+      {#if tb5WithoutRdma && !tb5InfoDismissed}
+        <div
+          class="flex items-center gap-2 px-3 py-2 rounded border border-blue-400/50 bg-blue-400/10 backdrop-blur-sm"
+          role="status"
+        >
+          <svg
+            class="w-5 h-5 text-blue-400 flex-shrink-0"
+            fill="none"
+            viewBox="0 0 24 24"
+            stroke="currentColor"
+            stroke-width="2"
+          >
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d={infoIconPath}
+            />
+          </svg>
+          <span class="text-sm font-mono text-blue-200"> RDMA AVAILABLE </span>
+          <button
+            type="button"
+            onclick={() => (tb5InfoDismissed = true)}
+            class="ml-1 text-blue-300/60 hover:text-blue-200 transition-colors cursor-pointer"
+            title="Dismiss"
+          >
+            <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="M6 18L18 6M6 6l12 12"
+              />
+            </svg>
+          </button>
+        </div>
+      {/if}
+    </div>
+  {/if}
+{/snippet}
+
+{#snippet clusterWarningsCompact()}
+  {#if tbBridgeCycles.length > 0 || macosVersionMismatch || (tb5WithoutRdma && !tb5InfoDismissed)}
+    <div class="absolute top-2 left-2 flex flex-col gap-1">
+      {#if tbBridgeCycles.length > 0}
+        <div
+          class="flex items-center gap-1.5 px-2 py-1 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm"
+          title="Thunderbolt Bridge cycle detected"
+        >
+          <svg
+            class="w-3.5 h-3.5 text-yellow-400"
+            fill="none"
+            viewBox="0 0 24 24"
+            stroke="currentColor"
+            stroke-width="2"
+          >
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d={warningIconPath}
+            />
+          </svg>
+          <span class="text-[10px] font-mono text-yellow-200">TB CYCLE</span>
+        </div>
+      {/if}
+      {#if macosVersionMismatch}
+        <div
+          class="flex items-center gap-1.5 px-2 py-1 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm"
+          title="Incompatible macOS versions detected"
+        >
+          <svg
+            class="w-3.5 h-3.5 text-yellow-400"
+            fill="none"
+            viewBox="0 0 24 24"
+            stroke="currentColor"
+            stroke-width="2"
+          >
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d={warningIconPath}
+            />
+          </svg>
+          <span class="text-[10px] font-mono text-yellow-200"
+            >macOS MISMATCH</span
+          >
+        </div>
+      {/if}
+      {#if tb5WithoutRdma && !tb5InfoDismissed}
+        <div
+          class="flex items-center gap-1.5 px-2 py-1 rounded border border-blue-400/50 bg-blue-400/10 backdrop-blur-sm"
+          title="Thunderbolt 5 detected — RDMA can be enabled for better performance"
+        >
+          <svg
+            class="w-3.5 h-3.5 text-blue-400"
+            fill="none"
+            viewBox="0 0 24 24"
+            stroke="currentColor"
+            stroke-width="2"
+          >
+            <path
+              stroke-linecap="round"
+              stroke-linejoin="round"
+              d={infoIconPath}
+            />
+          </svg>
+          <span class="text-[10px] font-mono text-blue-200">RDMA AVAILABLE</span
+          >
+        </div>
+      {/if}
+    </div>
+  {/if}
+{/snippet}
+
 <!-- Global event listeners for slider dragging -->
 <svelte:window
  onmousemove={handleSliderMouseMove}
@@ -1724,17 +2074,40 @@
            onNodeClick={togglePreviewNodeFilter}
          />

-          <!-- Thunderbolt Bridge Cycle Warning -->
-          {#if tbBridgeCycles.length > 0}
-            {@const cycle = tbBridgeCycles[0]}
-            {@const serviceName = getTbBridgeServiceName(cycle)}
-            {@const disableCmd = `sudo networksetup -setnetworkserviceenabled "${serviceName}" off`}
-            <div class="absolute top-4 left-4 group" role="alert">
-              <div
-                class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+          {@render clusterWarnings()}
+
+          <!-- TB5 RDMA Available Info -->
+          {#if tb5WithoutRdma && !tb5InfoDismissed}
+            <div
+              class="absolute left-4 flex items-center gap-2 px-3 py-2 rounded border border-blue-400/50 bg-blue-400/10 backdrop-blur-sm"
+              class:top-16={tbBridgeCycles.length > 0}
+              class:top-4={tbBridgeCycles.length === 0}
+              role="status"
+            >
+              <svg
+                class="w-5 h-5 text-blue-400 flex-shrink-0"
+                fill="none"
+                viewBox="0 0 24 24"
+                stroke="currentColor"
+                stroke-width="2"
+              >
+                <path
+                  stroke-linecap="round"
+                  stroke-linejoin="round"
+                  d="M13 16h-1v-4h-1m1-4h.01M21 12a9 9 0 11-18 0 9 9 0 0118 0z"
+                />
+              </svg>
+              <span class="text-sm font-mono text-blue-200">
+                RDMA AVAILABLE
+              </span>
+              <button
+                type="button"
+                onclick={() => (tb5InfoDismissed = true)}
+                class="ml-1 text-blue-300/60 hover:text-blue-200 transition-colors cursor-pointer"
+                title="Dismiss"
              >
                <svg
-                  class="w-5 h-5 text-yellow-400 flex-shrink-0"
+                  class="w-4 h-4"
                  fill="none"
                  viewBox="0 0 24 24"
                  stroke="currentColor"
@@ -1743,60 +2116,10 @@
                  <path
                    stroke-linecap="round"
                    stroke-linejoin="round"
-                    d="M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z"
+                    d="M6 18L18 6M6 6l12 12"
                  />
                </svg>
-                <span class="text-sm font-mono text-yellow-200">
-                  THUNDERBOLT BRIDGE CYCLE DETECTED
-                </span>
-              </div>
-
-              <!-- Tooltip on hover -->
-              <div
-                class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
-              >
-                <p class="text-xs text-white/80 mb-2">
-                  A network routing cycle was detected between nodes connected
-                  via Thunderbolt Bridge. This can cause connectivity issues.
-                </p>
-                <p class="text-xs text-white/60 mb-2">
-                  <span class="text-yellow-300">Affected nodes:</span>
-                  {cycle.map(getNodeName).join(" → ")}
-                </p>
-                <p class="text-xs text-white/60 mb-1">
-                  <span class="text-yellow-300">To fix:</span> Disable the Thunderbolt
-                  Bridge on one of the affected nodes:
-                </p>
-                <button
-                  type="button"
-                  onclick={() => copyToClipboard(disableCmd)}
-                  class="w-full flex items-center gap-2 text-[10px] font-mono bg-exo-black/60 px-2 py-1.5 rounded text-exo-yellow break-all text-left hover:bg-exo-black/80 transition-colors cursor-pointer group/copy"
-                  title="Click to copy"
-                >
-                  <span class="flex-1">{disableCmd}</span>
-                  <svg
-                    class="w-3.5 h-3.5 flex-shrink-0 text-white/40 group-hover/copy:text-exo-yellow transition-colors"
-                    fill="none"
-                    viewBox="0 0 24 24"
-                    stroke="currentColor"
-                    stroke-width="2"
-                  >
-                    {#if copiedCommand}
-                      <path
-                        stroke-linecap="round"
-                        stroke-linejoin="round"
-                        d="M5 13l4 4L19 7"
-                      />
-                    {:else}
-                      <path
-                        stroke-linecap="round"
-                        stroke-linejoin="round"
-                        d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z"
-                      />
-                    {/if}
-                  </svg>
-                </button>
-              </div>
+              </button>
            </div>
          {/if}

@@ -1843,17 +2166,21 @@
              onNodeClick={togglePreviewNodeFilter}
            />

-            <!-- Thunderbolt Bridge Cycle Warning -->
-            {#if tbBridgeCycles.length > 0}
-              {@const cycle = tbBridgeCycles[0]}
-              {@const serviceName = getTbBridgeServiceName(cycle)}
-              {@const disableCmd = `sudo networksetup -setnetworkserviceenabled "${serviceName}" off`}
-              <div class="absolute top-4 left-4 group" role="alert">
+            {@render clusterWarnings()}
+
+            <!-- TB5 RDMA Available Info -->
+            {#if tb5WithoutRdma && !tb5InfoDismissed}
+              <div
+                class="absolute left-4 group"
+                class:top-16={tbBridgeCycles.length > 0}
+                class:top-4={tbBridgeCycles.length === 0}
+                role="status"
+              >
                <div
-                  class="flex items-center gap-2 px-3 py-2 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm cursor-help"
+                  class="flex items-center gap-2 px-3 py-2 rounded border border-blue-400/50 bg-blue-400/10 backdrop-blur-sm"
                >
                  <svg
-                    class="w-5 h-5 text-yellow-400 flex-shrink-0"
+                    class="w-5 h-5 text-blue-400 flex-shrink-0"
                    fill="none"
                    viewBox="0 0 24 24"
                    stroke="currentColor"
@@ -1862,60 +2189,62 @@
                    <path
                      stroke-linecap="round"
                      stroke-linejoin="round"
-                      d="M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z"
+                      d="M13 16h-1v-4h-1m1-4h.01M21 12a9 9 0 11-18 0 9 9 0 0118 0z"
                    />
                  </svg>
-                  <span class="text-sm font-mono text-yellow-200">
-                    THUNDERBOLT BRIDGE CYCLE DETECTED
+                  <span class="text-sm font-mono text-blue-200">
+                    RDMA AVAILABLE
                  </span>
-                </div>
-
-                <!-- Tooltip on hover -->
-                <div
-                  class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-yellow-500/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
-                >
-                  <p class="text-xs text-white/80 mb-2">
-                    A network routing cycle was detected between nodes connected
-                    via Thunderbolt Bridge. This can cause connectivity issues.
-                  </p>
-                  <p class="text-xs text-white/60 mb-2">
-                    <span class="text-yellow-300">Affected nodes:</span>
-                    {cycle.map(getNodeName).join(" → ")}
-                  </p>
-                  <p class="text-xs text-white/60 mb-1">
-                    <span class="text-yellow-300">To fix:</span> Disable the Thunderbolt
-                    Bridge on one of the affected nodes:
-                  </p>
                  <button
                    type="button"
-                    onclick={() => copyToClipboard(disableCmd)}
-                    class="w-full flex items-center gap-2 text-[10px] font-mono bg-exo-black/60 px-2 py-1.5 rounded text-exo-yellow break-all text-left hover:bg-exo-black/80 transition-colors cursor-pointer group/copy"
-                    title="Click to copy"
+                    onclick={() => (tb5InfoDismissed = true)}
+                    class="ml-1 text-blue-300/60 hover:text-blue-200 transition-colors cursor-pointer"
+                    title="Dismiss"
                  >
-                    <span class="flex-1">{disableCmd}</span>
                    <svg
-                      class="w-3.5 h-3.5 flex-shrink-0 text-white/40 group-hover/copy:text-exo-yellow transition-colors"
+                      class="w-4 h-4"
                      fill="none"
                      viewBox="0 0 24 24"
                      stroke="currentColor"
                      stroke-width="2"
                    >
-                      {#if copiedCommand}
-                        <path
-                          stroke-linecap="round"
-                          stroke-linejoin="round"
-                          d="M5 13l4 4L19 7"
-                        />
-                      {:else}
-                        <path
-                          stroke-linecap="round"
-                          stroke-linejoin="round"
-                          d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z"
-                        />
-                      {/if}
+                      <path
+                        stroke-linecap="round"
+                        stroke-linejoin="round"
+                        d="M6 18L18 6M6 6l12 12"
+                      />
                    </svg>
                  </button>
                </div>
+
+                <!-- Tooltip on hover -->
+                <div
+                  class="absolute top-full left-0 mt-2 w-80 p-3 rounded border border-blue-400/30 bg-exo-dark-gray/95 backdrop-blur-sm opacity-0 invisible group-hover:opacity-100 group-hover:visible transition-all duration-200 z-50 shadow-lg"
+                >
+                  <p class="text-xs text-white/80 mb-2">
+                    Thunderbolt 5 hardware detected on multiple nodes. Enable
+                    RDMA for significantly faster inter-node communication.
+                  </p>
+                  <p class="text-xs text-white/60 mb-1.5">
+                    <span class="text-blue-300">To enable:</span>
+                  </p>
+                  <ol
+                    class="text-xs text-white/60 list-decimal list-inside space-y-0.5 mb-1.5"
+                  >
+                    <li>Connect nodes with TB5 cables</li>
+                    <li>Boot to Recovery (hold power 10s → Options)</li>
+                    <li>
+                      Run
+                      <code class="text-blue-300 bg-blue-400/10 px-1 rounded"
+                        >rdma_ctl enable</code
+                      >
+                    </li>
+                    <li>Reboot</li>
+                  </ol>
+                  <p class="text-xs text-white/40">
+                    Requires macOS 26.2+, TB5 cables, and matching OS versions.
+                  </p>
+                </div>
              </div>
            {/if}

@@ -1954,6 +2283,7 @@
                showHelperText={false}
                showModelSelector={true}
                modelTasks={modelTasks()}
+                modelCapabilities={modelCapabilities()}
              />
            </div>
          </div>
@@ -2733,6 +3063,7 @@
                placeholder="Ask anything"
                showModelSelector={true}
                modelTasks={modelTasks()}
+                modelCapabilities={modelCapabilities()}
              />
            </div>
          </div>
@@ -2773,30 +3104,7 @@
                  onNodeClick={togglePreviewNodeFilter}
                />

-                <!-- Thunderbolt Bridge Cycle Warning (compact) -->
-                {#if tbBridgeCycles.length > 0}
-                  <div
-                    class="absolute top-2 left-2 flex items-center gap-1.5 px-2 py-1 rounded border border-yellow-500/50 bg-yellow-500/10 backdrop-blur-sm"
-                    title="Thunderbolt Bridge cycle detected - click to view details"
-                  >
-                    <svg
-                      class="w-3.5 h-3.5 text-yellow-400"
-                      fill="none"
-                      viewBox="0 0 24 24"
-                      stroke="currentColor"
-                      stroke-width="2"
-                    >
-                      <path
-                        stroke-linecap="round"
-                        stroke-linejoin="round"
-                        d="M12 9v2m0 4h.01m-6.938 4h13.856c1.54 0 2.502-1.667 1.732-3L13.732 4c-.77-1.333-2.694-1.333-3.464 0L3.34 16c-.77 1.333.192 3 1.732 3z"
-                      />
-                    </svg>
-                    <span class="text-[10px] font-mono text-yellow-200"
-                      >TB CYCLE</span
-                    >
-                  </div>
-                {/if}
+                {@render clusterWarningsCompact()}
              </div>
            </button>

@@ -3252,11 +3560,17 @@
  {models}
  {selectedModelId}
  favorites={favoritesSet}
+  {recentModelIds}
+  hasRecents={showRecentsTab}
  existingModelIds={new Set(models.map((m) => m.id))}
  canModelFit={(modelId) => {
    const model = models.find((m) => m.id === modelId);
    return model ? hasEnoughMemory(model) : false;
  }}
+  getModelFitStatus={(modelId): ModelMemoryFitStatus => {
+    const model = models.find((m) => m.id === modelId);
+    return model ? getModelMemoryFitStatus(model) : "too_large";
+  }}
  onSelect={handleModelPickerSelect}
  onClose={() => (isModelPickerOpen = false)}
  onToggleFavorite={toggleFavorite}
@@ -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>

@@ -0,0 +1,58 @@
+# Stage 1: Build the dashboard
+FROM node:22-slim AS dashboard
+WORKDIR /app/dashboard
+COPY dashboard/package.json dashboard/package-lock.json ./
+RUN npm ci
+COPY dashboard/ .
+RUN npm run build
+
+# Stage 2: Build and run exo
+FROM python:3.13-slim
+
+# Install system dependencies
+# libblas-dev/liblapack-dev/liblapacke-dev are required by MLX CPU backend on Linux
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    pkg-config \
+    libssl-dev \
+    libblas-dev \
+    liblapack-dev \
+    liblapacke-dev \
+    curl \
+    protobuf-compiler \
+    iptables \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install Rust nightly
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain nightly
+ENV PATH="/root/.cargo/bin:${PATH}"
+
+# Wrap g++ with -fpermissive to fix MLX CPU JIT compilation with GCC 14
+# (GCC 14 treats _Float128/_Float32/_Float64 as built-in types, conflicting with MLX-generated code)
+# Must be done BEFORE uv sync so any source builds also get the fix
+RUN mv /usr/bin/g++ /usr/bin/g++.real && \
+    printf '#!/bin/sh\nexec /usr/bin/g++.real -fpermissive "$@"\n' > /usr/bin/g++ && \
+    chmod +x /usr/bin/g++
+
+# Install uv
+COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
+
+WORKDIR /app
+
+# Copy dependency files first for better layer caching
+COPY pyproject.toml Cargo.toml uv.lock README.md ./
+COPY rust/ ./rust/
+COPY bench/pyproject.toml ./bench/pyproject.toml
+
+# Copy source and resources
+COPY src/ ./src/
+COPY resources/ ./resources/
+
+# Copy built dashboard from stage 1
+COPY --from=dashboard /app/dashboard/build ./dashboard/build/
+
+# Install Python deps and build Rust bindings, then clean up build artifacts
+# to keep the layer small (Rust target/ and cargo registry can be 1-2 GB)
+RUN uv sync && rm -rf /app/rust/target /root/.cargo/registry /root/.cargo/git
+
+CMD [".venv/bin/exo", "-v"]
@@ -0,0 +1,195 @@
+"""Shared E2E test infrastructure for exo cluster tests."""
+
+import asyncio
+import json
+import os
+import sys
+from pathlib import Path
+from urllib.error import URLError
+from urllib.request import Request, urlopen
+
+E2E_DIR = Path(__file__).parent.resolve()
+TIMEOUT = int(os.environ.get("E2E_TIMEOUT", "120"))
+
+
+class Cluster:
+    """Async wrapper around a docker compose exo cluster."""
+
+    def __init__(self, name: str, overrides: list[str] | None = None):
+        self.name = name
+        self.project = f"e2e-{name}"
+        compose_files = [str(E2E_DIR / "docker-compose.yml")]
+        for path in overrides or []:
+            compose_files.append(str(E2E_DIR / path))
+        self._compose_base = [
+            "docker",
+            "compose",
+            "-p",
+            self.project,
+            *[arg for f in compose_files for arg in ("-f", f)],
+        ]
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *exc):
+        await self.stop()
+
+    async def _run(self, *args: str, check: bool = True) -> str:
+        proc = await asyncio.create_subprocess_exec(
+            *self._compose_base,
+            *args,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.STDOUT,
+        )
+        stdout, _ = await proc.communicate()
+        output = stdout.decode()
+        if check and proc.returncode != 0:
+            print(output, file=sys.stderr)
+            raise RuntimeError(
+                f"docker compose {' '.join(args)} failed (rc={proc.returncode})"
+            )
+        return output
+
+    async def build(self):
+        # Skip build if the image was pre-built (e.g. in CI with buildx cache)
+        proc = await asyncio.create_subprocess_exec(
+            "docker",
+            "image",
+            "inspect",
+            "exo-e2e:latest",
+            stdout=asyncio.subprocess.DEVNULL,
+            stderr=asyncio.subprocess.DEVNULL,
+        )
+        await proc.wait()
+        if proc.returncode == 0:
+            print("  Using pre-built image (exo-e2e:latest)")
+            return
+        print("  Building images...")
+        await self._run("build", "--quiet")
+
+    async def start(self):
+        print("  Starting cluster...")
+        await self._run("up", "-d")
+
+    async def stop(self):
+        print("  Cleaning up...")
+        await self._run("down", "--timeout", "5", check=False)
+
+    async def logs(self) -> str:
+        return await self._run("logs", check=False)
+
+    async def exec(
+        self, service: str, *cmd: str, check: bool = True
+    ) -> tuple[int, str]:
+        """Run a command inside a running container. Returns (returncode, output)."""
+        proc = await asyncio.create_subprocess_exec(
+            *self._compose_base,
+            "exec",
+            "-T",
+            service,
+            *cmd,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.STDOUT,
+        )
+        stdout, _ = await proc.communicate()
+        output = stdout.decode()
+        if check and proc.returncode != 0:
+            raise RuntimeError(
+                f"exec {' '.join(cmd)} in {service} failed (rc={proc.returncode})"
+            )
+        return proc.returncode, output
+
+    async def wait_for(self, description: str, check_fn, timeout: int = TIMEOUT):
+        """Poll check_fn every 2s until it returns True or timeout expires."""
+        print(f"  Waiting for {description}...")
+        deadline = asyncio.get_event_loop().time() + timeout
+        while asyncio.get_event_loop().time() < deadline:
+            if await check_fn():
+                print(f"  {description}")
+                return
+            await asyncio.sleep(2)
+        output = await self.logs()
+        print(f"--- cluster logs ---\n{output}\n---", file=sys.stderr)
+        raise TimeoutError(f"Timed out waiting for {description}")
+
+    async def assert_healthy(self):
+        """Verify the cluster formed correctly: nodes started, discovered each other, elected a master, API responds."""
+
+        async def both_nodes_started():
+            log = await self.logs()
+            return log.count("Starting node") >= 2
+
+        async def nodes_discovered():
+            log = await self.logs()
+            return log.count("ConnectionMessageType.Connected") >= 2
+
+        async def master_elected():
+            log = await self.logs()
+            return "demoting self" in log
+
+        async def api_responding():
+            try:
+                with urlopen("http://localhost:52415/v1/models", timeout=3) as resp:
+                    return resp.status == 200
+            except (URLError, OSError):
+                return False
+
+        await self.wait_for("Both nodes started", both_nodes_started)
+        await self.wait_for("Nodes discovered each other", nodes_discovered)
+        await self.wait_for("Master election resolved", master_elected)
+        await self.wait_for("API responding", api_responding)
+
+    async def _api(
+        self, method: str, path: str, body: dict | None = None, timeout: int = 30
+    ) -> dict:
+        """Make an API request to the cluster. Returns parsed JSON."""
+        url = f"http://localhost:52415{path}"
+        data = json.dumps(body).encode() if body else None
+        req = Request(
+            url, data=data, headers={"Content-Type": "application/json"}, method=method
+        )
+        loop = asyncio.get_event_loop()
+        resp_bytes = await loop.run_in_executor(
+            None, lambda: urlopen(req, timeout=timeout).read()
+        )
+        return json.loads(resp_bytes)
+
+    async def place_model(self, model: str, timeout: int = 600):
+        """Place a model instance on the cluster (triggers download) and wait until it's ready."""
+        await self._api("POST", "/place_instance", {"model_id": model})
+
+        async def model_ready():
+            try:
+                resp = await self._api("GET", "/v1/models")
+                return any(m.get("id") == model for m in resp.get("data", []))
+            except Exception:
+                return False
+
+        await self.wait_for(f"Model {model} ready", model_ready, timeout=timeout)
+
+    async def chat(
+        self, model: str, messages: list[dict], timeout: int = 600, **kwargs
+    ) -> dict:
+        """Send a chat completion request. Retries until model is downloaded and inference completes."""
+        body = json.dumps({"model": model, "messages": messages, **kwargs}).encode()
+        deadline = asyncio.get_event_loop().time() + timeout
+        last_error = None
+
+        while asyncio.get_event_loop().time() < deadline:
+            try:
+                req = Request(
+                    "http://localhost:52415/v1/chat/completions",
+                    data=body,
+                    headers={"Content-Type": "application/json"},
+                )
+                loop = asyncio.get_event_loop()
+                resp_bytes = await loop.run_in_executor(
+                    None, lambda r=req: urlopen(r, timeout=300).read()
+                )
+                return json.loads(resp_bytes)
+            except Exception as e:
+                last_error = e
+                await asyncio.sleep(5)
+
+        raise TimeoutError(f"Chat request failed after {timeout}s: {last_error}")
@@ -0,0 +1,20 @@
+services:
+  exo-node-1:
+    image: exo-e2e:latest
+    build:
+      context: ..
+      dockerfile: e2e/Dockerfile
+    environment:
+      - EXO_LIBP2P_NAMESPACE=docker-e2e
+    command: [".venv/bin/exo", "-v"]
+    ports:
+      - "52415:52415"
+
+  exo-node-2:
+    image: exo-e2e:latest
+    build:
+      context: ..
+      dockerfile: e2e/Dockerfile
+    environment:
+      - EXO_LIBP2P_NAMESPACE=docker-e2e
+    command: [".venv/bin/exo", "-v"]
@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+"""Discovers and runs all E2E tests in e2e/test_*.py.
+
+Tests with '# slow' on the first line of their docstring are skipped
+unless --slow is passed or E2E_SLOW=1 is set.
+"""
+
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+E2E_DIR = Path(__file__).parent.resolve()
+
+
+def is_slow(test_file: Path) -> bool:
+    """Check if the test file is marked as slow (has '# slow' in first 3 lines)."""
+    with open(test_file) as f:
+        for line in f:
+            if line.strip().startswith("#"):
+                continue
+            if line.strip().startswith('"""') or line.strip().startswith("'''"):
+                # Read into the docstring
+                for doc_line in f:
+                    if "slow" in doc_line.lower() and doc_line.strip().startswith(
+                        "slow"
+                    ):
+                        return True
+                    if '"""' in doc_line or "'''" in doc_line:
+                        break
+            break
+    return False
+
+
+def main():
+    run_slow = "--slow" in sys.argv or os.environ.get("E2E_SLOW") == "1"
+    if "--update-snapshots" in sys.argv:
+        os.environ["UPDATE_SNAPSHOTS"] = "1"
+    test_files = sorted(E2E_DIR.glob("test_*.py"))
+    if not test_files:
+        print("No test files found")
+        sys.exit(1)
+
+    passed = 0
+    failed = 0
+    skipped = 0
+    failures = []
+
+    for test_file in test_files:
+        name = test_file.stem
+        if is_slow(test_file) and not run_slow:
+            print(f"=== {name} === SKIPPED (slow, use --slow to run)")
+            skipped += 1
+            continue
+
+        print(f"=== {name} ===")
+        result = subprocess.run([sys.executable, str(test_file)])
+        if result.returncode == 0:
+            passed += 1
+        else:
+            # Retry once — Docker networking (mDNS) can be slow on first boot
+            print(f"\n=== {name} === RETRYING (attempt 2/2)")
+            result = subprocess.run([sys.executable, str(test_file)])
+            if result.returncode == 0:
+                passed += 1
+            else:
+                failed += 1
+                failures.append(name)
+        print()
+
+    total = passed + failed + skipped
+    print("================================")
+    print(
+        f"{passed}/{total} tests passed" + (f", {skipped} skipped" if skipped else "")
+    )
+
+    if failed:
+        print(f"Failed: {' '.join(failures)}")
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,78 @@
+"""Snapshot testing infrastructure for E2E tests.
+
+Provides deterministic regression testing by comparing inference output
+against committed baseline snapshots. Tests FAIL if no baseline exists —
+baselines must be explicitly generated and committed.
+
+Generate baselines: UPDATE_SNAPSHOTS=1 python3 e2e/run_all.py --slow
+Update after intentional changes: UPDATE_SNAPSHOTS=1 python3 e2e/run_all.py --slow
+
+Snapshots are stored per-architecture (e.g. snapshots/x86_64/, snapshots/arm64/)
+since floating-point results differ between CPU architectures.
+"""
+
+import difflib
+import json
+import os
+import platform
+from pathlib import Path
+
+ARCH = platform.machine()
+SNAPSHOTS_DIR = Path(__file__).parent / "snapshots" / ARCH
+
+
+def assert_snapshot(
+    name: str,
+    content: str,
+    metadata: dict,
+) -> None:
+    """Compare content against a saved snapshot, or create one if missing.
+
+    Args:
+        name: Snapshot identifier (used as filename: snapshots/{arch}/{name}.json).
+        content: The actual inference output to compare.
+        metadata: Additional context stored alongside content (model, seed, etc.).
+                  Not used for comparison -- purely documentary.
+
+    Raises:
+        AssertionError: If content doesn't match the saved snapshot.
+
+    Environment:
+        UPDATE_SNAPSHOTS=1: Overwrite existing snapshot with actual content.
+    """
+    snapshot_file = SNAPSHOTS_DIR / f"{name}.json"
+    update = os.environ.get("UPDATE_SNAPSHOTS") == "1"
+
+    if update:
+        # Explicitly regenerate snapshot
+        SNAPSHOTS_DIR.mkdir(parents=True, exist_ok=True)
+        snapshot_data = {**metadata, "arch": ARCH, "content": content}
+        snapshot_file.write_text(json.dumps(snapshot_data, indent=2) + "\n")
+        print(f"  Updated snapshot: {ARCH}/{snapshot_file.name}")
+    elif not snapshot_file.exists():
+        raise AssertionError(
+            f"No baseline snapshot for '{name}' on {ARCH}.\n"
+            f"Expected file: {snapshot_file}\n\n"
+            f"Generate baselines with: UPDATE_SNAPSHOTS=1 python3 e2e/run_all.py --slow"
+        )
+    else:
+        snapshot = json.loads(snapshot_file.read_text())
+        expected = snapshot["content"]
+        if content != expected:
+            diff = "\n".join(
+                difflib.unified_diff(
+                    expected.splitlines(),
+                    content.splitlines(),
+                    fromfile=f"expected ({snapshot_file.relative_to(SNAPSHOTS_DIR.parent.parent)})",
+                    tofile="actual",
+                    lineterm="",
+                )
+            )
+            raise AssertionError(
+                f"Snapshot mismatch for '{name}' on {ARCH}!\n\n"
+                f"{diff}\n\n"
+                f"Expected: {expected!r}\n"
+                f"Actual:   {content!r}\n\n"
+                f"To update: UPDATE_SNAPSHOTS=1 python3 e2e/run_all.py --slow"
+            )
+        print(f"  Output matches snapshot ({ARCH}/{snapshot_file.name})")
@@ -0,0 +1,22 @@
+"""Test: Basic cluster formation.
+
+Verifies two nodes discover each other, elect a master, and the API responds.
+"""
+
+import asyncio
+import sys
+
+sys.path.insert(0, str(__import__("pathlib").Path(__file__).parent))
+from conftest import Cluster
+
+
+async def main():
+    async with Cluster("cluster_formation") as cluster:
+        await cluster.build()
+        await cluster.start()
+        await cluster.assert_healthy()
+        print("PASSED: cluster_formation")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,61 @@
+"""Test: Deterministic inference output (snapshot test).
+
+Sends a chat completion request with a fixed seed,
+then verifies the output matches a known-good snapshot. This ensures
+inference produces consistent results across runs.
+
+Uses MLX CPU backend in Docker on x86 Linux.
+"""
+
+import asyncio
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent))
+from snapshot import assert_snapshot
+
+from conftest import Cluster
+
+MODEL = "mlx-community/Qwen3-0.6B-4bit"
+SEED = 42
+PROMPT = "What is 2+2? Reply with just the number."
+MAX_TOKENS = 32
+
+
+async def main():
+    async with Cluster("inference_snapshot") as cluster:
+        await cluster.build()
+        await cluster.start()
+        await cluster.assert_healthy()
+
+        print(f"  Launching model {MODEL}...")
+        await cluster.place_model(MODEL)
+
+        print(f"  Sending chat completion (seed={SEED})...")
+        resp = await cluster.chat(
+            model=MODEL,
+            messages=[{"role": "user", "content": PROMPT}],
+            seed=SEED,
+            temperature=0,
+            max_tokens=MAX_TOKENS,
+        )
+
+        content = resp["choices"][0]["message"]["content"]
+        print(f"  Response: {content!r}")
+
+        assert_snapshot(
+            name="inference_snapshot",
+            content=content,
+            metadata={
+                "model": MODEL,
+                "seed": SEED,
+                "prompt": PROMPT,
+                "max_tokens": MAX_TOKENS,
+            },
+        )
+
+        print("PASSED: inference_snapshot")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,47 @@
+"""Test: Cluster works without internet access.
+
+Verifies exo functions correctly when containers can talk to each other
+but cannot reach the internet. Uses iptables to block all outbound traffic
+except private subnets and multicast (for mDNS discovery).
+"""
+
+import asyncio
+import sys
+
+sys.path.insert(0, str(__import__("pathlib").Path(__file__).parent))
+from conftest import Cluster
+
+
+async def main():
+    async with Cluster(
+        "no_internet",
+        overrides=["tests/no_internet/docker-compose.override.yml"],
+    ) as cluster:
+        await cluster.build()
+        await cluster.start()
+        await cluster.assert_healthy()
+
+        # Verify internet is actually blocked from inside the containers
+        for node in ["exo-node-1", "exo-node-2"]:
+            rc, _ = await cluster.exec(
+                node,
+                "curl",
+                "-sf",
+                "--max-time",
+                "3",
+                "https://huggingface.co",
+                check=False,
+            )
+            assert rc != 0, f"{node} should not be able to reach the internet"
+            print(f"  {node}: internet correctly blocked")
+
+        # Verify exo detected no internet connectivity
+        log = await cluster.logs()
+        assert "Internet connectivity: False" in log, "exo should detect no internet"
+        print("  exo correctly detected no internet connectivity")
+
+        print("PASSED: no_internet")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,65 @@
+"""Test: Runner chaos — abrupt runner death detection.
+slow
+
+Sends a chat completion with the EXO_RUNNER_MUST_DIE trigger, which causes
+the runner process to call os._exit(1) (simulating an OOM kill). Verifies that
+the RunnerSupervisor health check detects the death and the system doesn't hang.
+
+Requires a machine that can run MLX inference at reasonable speed (Apple Silicon).
+Run with: python3 e2e/run_all.py --slow  or  E2E_SLOW=1 python3 e2e/run_all.py
+"""
+
+import asyncio
+import contextlib
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent))
+from conftest import Cluster
+
+MODEL = "mlx-community/Qwen3-0.6B-4bit"
+
+
+async def main():
+    async with Cluster("runner_chaos") as cluster:
+        await cluster.build()
+        await cluster.start()
+        await cluster.assert_healthy()
+
+        # Place the model so a runner is loaded and ready
+        print(f"  Launching model {MODEL}...")
+        await cluster.place_model(MODEL)
+
+        # Send a chat request with the die trigger.
+        # The runner will call os._exit(1) mid-inference, simulating OOM kill.
+        # The chat request itself will fail — that's expected.
+        print("  Sending EXO_RUNNER_MUST_DIE trigger...")
+        with contextlib.suppress(Exception):
+            await cluster.chat(
+                model=MODEL,
+                messages=[{"role": "user", "content": "EXO RUNNER MUST DIE"}],
+                timeout=60,
+            )
+
+        # Wait for the health check to detect the death and emit RunnerFailed
+        async def health_check_detected():
+            log = await cluster.logs()
+            return "runner process died unexpectedly" in log
+
+        await cluster.wait_for(
+            "Health check detected runner death",
+            health_check_detected,
+            timeout=30,
+        )
+
+        # Verify RunnerFailed was emitted (visible in logs)
+        log = await cluster.logs()
+        assert "runner process died unexpectedly" in log, (
+            f"Expected health check to detect runner death but it didn't.\nLogs:\n{log}"
+        )
+
+        print("PASSED: runner_chaos")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,60 @@
+"""Test: Code generation snapshot.
+slow
+
+Verifies deterministic output for a code generation prompt.
+"""
+
+import asyncio
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent))
+from snapshot import assert_snapshot
+
+from conftest import Cluster
+
+MODEL = "mlx-community/Qwen3-0.6B-4bit"
+SEED = 42
+PROMPT = (
+    "Write a Python function to reverse a string. Only output the code, no explanation."
+)
+MAX_TOKENS = 64
+
+
+async def main():
+    async with Cluster("snapshot_code_gen") as cluster:
+        await cluster.build()
+        await cluster.start()
+        await cluster.assert_healthy()
+
+        print(f"  Launching model {MODEL}...")
+        await cluster.place_model(MODEL)
+
+        print(f"  Sending chat completion (seed={SEED})...")
+        resp = await cluster.chat(
+            model=MODEL,
+            messages=[{"role": "user", "content": PROMPT}],
+            seed=SEED,
+            temperature=0,
+            max_tokens=MAX_TOKENS,
+        )
+
+        content = resp["choices"][0]["message"]["content"]
+        print(f"  Response: {content!r}")
+
+        assert_snapshot(
+            name="snapshot_code_gen",
+            content=content,
+            metadata={
+                "model": MODEL,
+                "seed": SEED,
+                "prompt": PROMPT,
+                "max_tokens": MAX_TOKENS,
+            },
+        )
+
+        print("PASSED: snapshot_code_gen")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,65 @@
+"""Test: Edge case snapshots.
+slow
+
+Verifies deterministic output for edge-case prompts: single word input,
+special characters, and unicode.
+"""
+
+import asyncio
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent))
+from snapshot import assert_snapshot
+
+from conftest import Cluster
+
+MODEL = "mlx-community/Qwen3-0.6B-4bit"
+SEED = 42
+MAX_TOKENS = 32
+
+CASES = [
+    ("edge_single_word", "Hi"),
+    ("edge_special_chars", "What does 2 * (3 + 4) / 7 - 1 equal? Use <math> tags."),
+    ("edge_unicode", "Translate 'hello' to Japanese, Chinese, and Korean."),
+]
+
+
+async def main():
+    async with Cluster("snapshot_edge") as cluster:
+        await cluster.build()
+        await cluster.start()
+        await cluster.assert_healthy()
+
+        print(f"  Launching model {MODEL}...")
+        await cluster.place_model(MODEL)
+
+        for snapshot_name, prompt in CASES:
+            print(f"  [{snapshot_name}] Sending: {prompt!r}")
+            resp = await cluster.chat(
+                model=MODEL,
+                messages=[{"role": "user", "content": prompt}],
+                seed=SEED,
+                temperature=0,
+                max_tokens=MAX_TOKENS,
+            )
+
+            content = resp["choices"][0]["message"]["content"]
+            print(f"  [{snapshot_name}] Response: {content!r}")
+
+            assert_snapshot(
+                name=snapshot_name,
+                content=content,
+                metadata={
+                    "model": MODEL,
+                    "seed": SEED,
+                    "prompt": prompt,
+                    "max_tokens": MAX_TOKENS,
+                },
+            )
+
+        print("PASSED: snapshot_edge")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,58 @@
+"""Test: Longer output snapshot.
+slow
+
+Verifies deterministic output with a higher max_tokens (128).
+"""
+
+import asyncio
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent))
+from snapshot import assert_snapshot
+
+from conftest import Cluster
+
+MODEL = "mlx-community/Qwen3-0.6B-4bit"
+SEED = 42
+PROMPT = "Explain how a binary search algorithm works."
+MAX_TOKENS = 128
+
+
+async def main():
+    async with Cluster("snapshot_long_output") as cluster:
+        await cluster.build()
+        await cluster.start()
+        await cluster.assert_healthy()
+
+        print(f"  Launching model {MODEL}...")
+        await cluster.place_model(MODEL)
+
+        print(f"  Sending chat completion (seed={SEED}, max_tokens={MAX_TOKENS})...")
+        resp = await cluster.chat(
+            model=MODEL,
+            messages=[{"role": "user", "content": PROMPT}],
+            seed=SEED,
+            temperature=0,
+            max_tokens=MAX_TOKENS,
+        )
+
+        content = resp["choices"][0]["message"]["content"]
+        print(f"  Response: {content!r}")
+
+        assert_snapshot(
+            name="snapshot_long_output",
+            content=content,
+            metadata={
+                "model": MODEL,
+                "seed": SEED,
+                "prompt": PROMPT,
+                "max_tokens": MAX_TOKENS,
+            },
+        )
+
+        print("PASSED: snapshot_long_output")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,73 @@
+"""Test: Multi-model snapshot tests.
+slow
+
+Verifies deterministic output across different model architectures to catch
+model-specific regressions. Each model uses its own snapshot file.
+Run with: python3 e2e/run_all.py --slow  or  E2E_SLOW=1 python3 e2e/run_all.py
+"""
+
+import asyncio
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent))
+from snapshot import assert_snapshot
+
+from conftest import Cluster
+
+SEED = 42
+PROMPT = "What is the capital of France?"
+MAX_TOKENS = 32
+
+MODELS = [
+    "mlx-community/SmolLM2-135M-Instruct",
+    "mlx-community/Llama-3.2-1B-Instruct-4bit",
+    "mlx-community/gemma-2-2b-it-4bit",
+]
+
+
+async def main():
+    async with Cluster("snapshot_multi_model") as cluster:
+        await cluster.build()
+        await cluster.start()
+        await cluster.assert_healthy()
+
+        for model in MODELS:
+            short_name = (
+                model.split("/")[-1].lower().replace("-", "_").replace(".", "_")
+            )
+            snapshot_name = f"snapshot_multi_{short_name}"
+
+            print(f"  Launching model {model}...")
+            await cluster.place_model(model)
+
+            print(f"  Sending chat completion (seed={SEED})...")
+            resp = await cluster.chat(
+                model=model,
+                messages=[{"role": "user", "content": PROMPT}],
+                seed=SEED,
+                temperature=0,
+                max_tokens=MAX_TOKENS,
+            )
+
+            content = resp["choices"][0]["message"]["content"]
+            print(f"  [{short_name}] Response: {content!r}")
+
+            assert_snapshot(
+                name=snapshot_name,
+                content=content,
+                metadata={
+                    "model": model,
+                    "seed": SEED,
+                    "prompt": PROMPT,
+                    "max_tokens": MAX_TOKENS,
+                },
+            )
+
+            print(f"  [{short_name}] PASSED")
+
+        print("PASSED: snapshot_multi_model")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,58 @@
+"""Test: Reasoning/math snapshot.
+slow
+
+Verifies deterministic output for a simple reasoning prompt.
+"""
+
+import asyncio
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent))
+from snapshot import assert_snapshot
+
+from conftest import Cluster
+
+MODEL = "mlx-community/Qwen3-0.6B-4bit"
+SEED = 42
+PROMPT = "If I have 3 apples and give away 1, how many do I have? Think step by step."
+MAX_TOKENS = 64
+
+
+async def main():
+    async with Cluster("snapshot_reasoning") as cluster:
+        await cluster.build()
+        await cluster.start()
+        await cluster.assert_healthy()
+
+        print(f"  Launching model {MODEL}...")
+        await cluster.place_model(MODEL)
+
+        print(f"  Sending chat completion (seed={SEED})...")
+        resp = await cluster.chat(
+            model=MODEL,
+            messages=[{"role": "user", "content": PROMPT}],
+            seed=SEED,
+            temperature=0,
+            max_tokens=MAX_TOKENS,
+        )
+
+        content = resp["choices"][0]["message"]["content"]
+        print(f"  Response: {content!r}")
+
+        assert_snapshot(
+            name="snapshot_reasoning",
+            content=content,
+            metadata={
+                "model": MODEL,
+                "seed": SEED,
+                "prompt": PROMPT,
+                "max_tokens": MAX_TOKENS,
+            },
+        )
+
+        print("PASSED: snapshot_reasoning")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,32 @@
+# Block all outbound internet traffic using iptables while preserving:
+#   - Multicast (224.0.0.0/4) for mDNS peer discovery
+#   - Private subnets (10/8, 172.16/12, 192.168/16) for inter-container communication
+#   - Loopback (127/8)
+# Requires NET_ADMIN capability for iptables.
+services:
+  exo-node-1:
+    cap_add:
+      - NET_ADMIN
+    entrypoint: ["/bin/sh", "-c"]
+    command:
+      - |
+        iptables -A OUTPUT -d 127.0.0.0/8 -j ACCEPT
+        iptables -A OUTPUT -d 10.0.0.0/8 -j ACCEPT
+        iptables -A OUTPUT -d 172.16.0.0/12 -j ACCEPT
+        iptables -A OUTPUT -d 192.168.0.0/16 -j ACCEPT
+        iptables -A OUTPUT -d 224.0.0.0/4 -j ACCEPT
+        iptables -A OUTPUT -j REJECT
+        exec .venv/bin/exo -v
+  exo-node-2:
+    cap_add:
+      - NET_ADMIN
+    entrypoint: ["/bin/sh", "-c"]
+    command:
+      - |
+        iptables -A OUTPUT -d 127.0.0.0/8 -j ACCEPT
+        iptables -A OUTPUT -d 10.0.0.0/8 -j ACCEPT
+        iptables -A OUTPUT -d 172.16.0.0/12 -j ACCEPT
+        iptables -A OUTPUT -d 192.168.0.0/16 -j ACCEPT
+        iptables -A OUTPUT -d 224.0.0.0/4 -j ACCEPT
+        iptables -A OUTPUT -j REJECT
+        exec .venv/bin/exo -v
@@ -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";
@@ -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";
+  };
+}
@@ -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="
+  }
+}
@@ -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-SV/3MXt+SuJ69XfLfXycold6KgtXSM7OE0KwMSNw+eE=";
+      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}")
@@ -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]
@@ -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;
@@ -118,6 +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
+          '';
      };
    };
 }
@@ -0,0 +1,12 @@
+model_id = "mlx-community/SmolLM2-135M-Instruct"
+n_layers = 30
+hidden_size = 576
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "llama"
+quantization = "bf16"
+base_model = "SmolLM2 135M"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 269060381
@@ -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
@@ -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
@@ -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
@@ -0,0 +1,12 @@
+model_id = "mlx-community/gemma-2-2b-it-4bit"
+n_layers = 26
+hidden_size = 2304
+supports_tensor = false
+tasks = ["TextGeneration"]
+family = "gemma2"
+quantization = "4bit"
+base_model = "Gemma 2 2B"
+capabilities = ["text"]
+
+[storage_size]
+in_bytes = 1492755242
@@ -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;
@@ -532,7 +531,9 @@ 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)
@@ -106,7 +106,6 @@ mod behaviour {
    use crate::{alias, discovery};
    use libp2p::swarm::NetworkBehaviour;
    use libp2p::{gossipsub, identity};
-    use std::time::Duration;

    /// Behavior of the Swarm which composes all desired behaviors:
    /// Right now its just [`discovery::Behaviour`] and [`gossipsub::Behaviour`].
@@ -134,7 +133,6 @@ 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()
@@ -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 }
@@ -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()
-        }
-    }
-}
@@ -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()
-    }
-}
@@ -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,
@@ -251,7 +251,8 @@ 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")
    # TODO: Refactor the current verbosity system
@@ -17,6 +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
@@ -79,6 +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,
@@ -124,6 +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(
@@ -137,6 +141,8 @@ 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(
@@ -160,12 +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:
@@ -175,7 +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] = []
@@ -183,6 +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):
@@ -192,6 +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:
@@ -222,7 +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,
@@ -240,4 +254,6 @@ async def collect_chat_response(
                finish_reason=finish_reason,
            )
        ],
-    )
+        usage=last_usage,
+    ).model_dump_json()
+    return
@@ -4,7 +4,7 @@ import json
 from collections.abc import AsyncGenerator
 from typing import Any

-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,
@@ -161,12 +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:
@@ -174,6 +176,8 @@ 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(
@@ -183,12 +187,10 @@ async def collect_claude_response(
                        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)
@@ -208,11 +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,
@@ -221,7 +223,8 @@ async def collect_claude_response(
            input_tokens=input_tokens,
            output_tokens=output_tokens,
        ),
-    )
+    ).model_dump_json()
+    return


 async def generate_claude_stream(
@@ -249,7 +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:
@@ -257,8 +260,9 @@ 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
@@ -290,7 +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(
@@ -302,9 +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)
@@ -4,6 +4,7 @@ from collections.abc import AsyncGenerator
 from itertools import count
 from typing import Any

+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 (
@@ -121,13 +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:
@@ -135,6 +138,8 @@ 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(
@@ -145,22 +150,20 @@ async def collect_responses_response(
                        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] = [
@@ -172,14 +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(
@@ -235,15 +239,16 @@ 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_{tool.id}"
                call_id = f"call_{tool.id}"
@@ -302,7 +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(
@@ -346,13 +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
@@ -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,
@@ -123,6 +125,7 @@ from exo.shared.types.commands import (
    PlaceInstance,
    SendInputChunk,
    StartDownload,
+    TaskCancelled,
    TaskFinished,
    TextGeneration,
 )
@@ -146,6 +149,8 @@ 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 +180,7 @@ 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 +228,8 @@ 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]()
@@ -289,7 +296,7 @@ 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)
@@ -534,16 +541,14 @@ 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]

@@ -638,11 +643,14 @@ 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
@@ -658,8 +666,7 @@ 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.
@@ -694,6 +701,22 @@ 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:
@@ -861,6 +884,11 @@ 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))
@@ -942,6 +970,11 @@ 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))
@@ -1199,12 +1232,15 @@ 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
@@ -1232,11 +1268,15 @@ 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."""
@@ -1342,6 +1382,7 @@ class API:
                    with anyio.CancelScope(shield=True):
                        shutdown_ev.set()
        finally:
+            self._event_log.close()
            self.command_sender.close()
            self.global_event_receiver.close()

@@ -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()
@@ -4,6 +4,7 @@ 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,
@@ -12,7 +13,7 @@ from exo.master.placement import (
    place_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.types.commands import (
    CreateInstance,
    DeleteInstance,
@@ -23,6 +24,7 @@ from exo.shared.types.commands import (
    PlaceInstance,
    RequestEventLog,
    SendInputChunk,
+    TaskCancelled,
    TaskFinished,
    TestCommand,
    TextGeneration,
@@ -38,6 +40,7 @@ from exo.shared.types.events import (
    NodeTimedOut,
    TaskCreated,
    TaskDeleted,
+    TaskStatusUpdated,
    TraceEventData,
    TracesCollected,
    TracesMerged,
@@ -88,8 +91,7 @@ class Master:
            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]] = {}

@@ -103,6 +105,7 @@ class Master:
                tg.start_soon(self._loopback_processor)
                tg.start_soon(self._plan)
        finally:
+            self._event_log.close()
            self.global_event_sender.close()
            self.local_event_receiver.close()
            self.command_receiver.close()
@@ -278,7 +281,7 @@ 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
@@ -298,7 +301,7 @@ 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 +311,7 @@ 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 +321,18 @@ 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(
@@ -326,20 +341,18 @@ 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
                            # rate limit to 1000 at a time
-                            for i in range(
-                                command.since_idx,
-                                min(command.since_idx + 1000, len(self._event_log)),
+                            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=self._event_log[i])
-                                )
+                                await self._send_event(IndexedEvent(idx=i, event=event))
                    for event in generated_events:
                        await self.event_sender.send(event)
                except ValueError as e:
@@ -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,
@@ -186,6 +192,7 @@ 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 +208,18 @@ 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,
@@ -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
@@ -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()
@@ -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
@@ -207,10 +207,10 @@ 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(
@@ -31,6 +31,7 @@ from exo.shared.types.events import (
 from exo.shared.types.profiling import (
    NodeIdentity,
    NodeNetworkInfo,
+    NodeRdmaCtlStatus,
    NodeThunderboltInfo,
    ThunderboltBridgeStatus,
 )
@@ -47,7 +48,9 @@ from exo.utils.info_gatherer.info_gatherer import (
    MemoryUsage,
    MiscData,
    NodeConfig,
+    NodeDiskUsage,
    NodeNetworkInterfaces,
+    RdmaCtlStatus,
    StaticNodeInformation,
    ThunderboltBridgeInfo,
 )
@@ -223,6 +226,9 @@ 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 +245,9 @@ 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 +265,12 @@ 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 +299,8 @@ 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 +315,12 @@ 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 +372,11 @@ 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)

@@ -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"

@@ -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,
        )


@@ -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")
@@ -3,8 +3,7 @@ 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
@@ -199,6 +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
@@ -227,13 +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
@@ -48,6 +48,10 @@ class DeleteInstance(BaseCommand):
    instance_id: InstanceId


+class TaskCancelled(BaseCommand):
+    cancelled_command_id: CommandId
+
+
 class TaskFinished(BaseCommand):
    finished_command_id: CommandId

@@ -89,6 +93,7 @@ Command = (
    | PlaceInstance
    | CreateInstance
    | DeleteInstance
+    | TaskCancelled
    | TaskFinished
    | SendInputChunk
 )
@@ -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."""

@@ -8,9 +8,11 @@ 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.profiling import (
+    DiskUsage,
    MemoryUsage,
    NodeIdentity,
    NodeNetworkInfo,
+    NodeRdmaCtlStatus,
    NodeThunderboltInfo,
    SystemPerformanceProfile,
    ThunderboltBridgeStatus,
@@ -49,10 +51,12 @@ 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]] = []
@@ -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
@@ -40,5 +40,6 @@ class TextGenerationTaskParams(BaseModel, frozen=True):
    stop: str | list[str] | None = None
    seed: int | None = None
    chat_template_messages: list[dict[str, Any]] | None = None
+    enable_thinking: bool | None = None
    logprobs: bool = False
    top_logprobs: int | None = None
@@ -12,6 +12,7 @@ class ThunderboltConnection(CamelCaseModel):
 class ThunderboltIdentifier(CamelCaseModel):
    rdma_interface: str
    domain_uuid: str
+    link_speed: str = ""


 ## Intentionally minimal, only collecting data we care about - there's a lot more
@@ -19,6 +20,7 @@ class ThunderboltIdentifier(CamelCaseModel):

 class _ReceptacleTag(BaseModel, extra="ignore"):
    receptacle_id_key: str | None = None
+    current_speed_key: str | None = None


 class _ConnectivityItem(BaseModel, extra="ignore"):
@@ -42,7 +44,9 @@ class ThunderboltConnectivityData(BaseModel, extra="ignore"):
        # if tag not in ifaces: return None
        iface = f"rdma_{ifaces[tag]}"
        return ThunderboltIdentifier(
-            rdma_interface=iface, domain_uuid=self.domain_uuid_key
+            rdma_interface=iface,
+            domain_uuid=self.domain_uuid_key,
+            link_speed=self.receptacle_1_tag.current_speed_key or "",
        )

    def conn(self) -> ThunderboltConnection | None:
@@ -62,6 +62,7 @@ class PartialImageResponse(BaseRunnerResponse):
 class ToolCallResponse(BaseRunnerResponse):
    tool_calls: list[ToolCallItem]
    usage: Usage | None
+    stats: GenerationStats | None = None


 class FinishedResponse(BaseRunnerResponse):
@@ -125,7 +125,9 @@ class MpSender[T]:
            self._state.buffer.put(item, block=True)

    async def send_async(self, item: T) -> None:
-        await to_thread.run_sync(self.send, item, limiter=CapacityLimiter(1))
+        await to_thread.run_sync(
+            self.send, item, limiter=CapacityLimiter(1), abandon_on_cancel=True
+        )

    def close(self) -> None:
        if not self._state.closed.is_set():
@@ -8,16 +8,17 @@ from subprocess import CalledProcessError
 from typing import Self, cast

 import anyio
-from anyio import create_task_group, open_process
+from anyio import create_task_group, fail_after, open_process, to_thread
 from anyio.abc import TaskGroup
 from anyio.streams.buffered import BufferedByteReceiveStream
 from anyio.streams.text import TextReceiveStream
 from loguru import logger
 from pydantic import ValidationError

-from exo.shared.constants import EXO_CONFIG_FILE
+from exo.shared.constants import EXO_CONFIG_FILE, EXO_MODELS_DIR
 from exo.shared.types.memory import Memory
 from exo.shared.types.profiling import (
+    DiskUsage,
    MemoryUsage,
    NetworkInterfaceInfo,
    ThunderboltBridgeStatus,
@@ -31,7 +32,13 @@ from exo.utils.channels import Sender
 from exo.utils.pydantic_ext import TaggedModel

 from .macmon import MacmonMetrics
-from .system_info import get_friendly_name, get_model_and_chip, get_network_interfaces
+from .system_info import (
+    get_friendly_name,
+    get_model_and_chip,
+    get_network_interfaces,
+    get_os_build_version,
+    get_os_version,
+)

 IS_DARWIN = sys.platform == "darwin"

@@ -177,11 +184,18 @@ class StaticNodeInformation(TaggedModel):

    model: str
    chip: str
+    os_version: str
+    os_build_version: str

    @classmethod
    async def gather(cls) -> Self:
        model, chip = await get_model_and_chip()
-        return cls(model=model, chip=chip)
+        return cls(
+            model=model,
+            chip=chip,
+            os_version=get_os_version(),
+            os_build_version=await get_os_build_version(),
+        )


 class NodeNetworkInterfaces(TaggedModel):
@@ -196,6 +210,28 @@ class MacThunderboltConnections(TaggedModel):
    conns: Sequence[ThunderboltConnection]


+class RdmaCtlStatus(TaggedModel):
+    enabled: bool
+
+    @classmethod
+    async def gather(cls) -> Self | None:
+        if not IS_DARWIN or shutil.which("rdma_ctl") is None:
+            return None
+        try:
+            with anyio.fail_after(5):
+                proc = await anyio.run_process(["rdma_ctl", "status"], check=False)
+        except (TimeoutError, OSError):
+            return None
+        if proc.returncode != 0:
+            return None
+        output = proc.stdout.decode("utf-8").lower().strip()
+        if "enabled" in output:
+            return cls(enabled=True)
+        if "disabled" in output:
+            return cls(enabled=False)
+        return None
+
+
 class ThunderboltBridgeInfo(TaggedModel):
    status: ThunderboltBridgeStatus

@@ -284,6 +320,20 @@ class MiscData(TaggedModel):
        return cls(friendly_name=await get_friendly_name())


+class NodeDiskUsage(TaggedModel):
+    """Disk space information for the models directory."""
+
+    disk_usage: DiskUsage
+
+    @classmethod
+    async def gather(cls) -> Self:
+        return cls(
+            disk_usage=await to_thread.run_sync(
+                lambda: DiskUsage.from_path(EXO_MODELS_DIR)
+            )
+        )
+
+
 async def _gather_iface_map() -> dict[str, str] | None:
    proc = await anyio.run_process(
        ["networksetup", "-listallhardwareports"], check=False
@@ -310,10 +360,12 @@ GatheredInfo = (
    | NodeNetworkInterfaces
    | MacThunderboltIdentifiers
    | MacThunderboltConnections
+    | RdmaCtlStatus
    | ThunderboltBridgeInfo
    | NodeConfig
    | MiscData
    | StaticNodeInformation
+    | NodeDiskUsage
 )


@@ -326,6 +378,9 @@ class InfoGatherer:
    memory_poll_rate: float | None = None if IS_DARWIN else 1
    macmon_interval: float | None = 1 if IS_DARWIN else None
    thunderbolt_bridge_poll_interval: float | None = 10 if IS_DARWIN else None
+    static_info_poll_interval: float | None = 60
+    rdma_ctl_poll_interval: float | None = 10 if IS_DARWIN else None
+    disk_poll_interval: float | None = 30
    _tg: TaskGroup = field(init=False, default_factory=create_task_group)

    async def run(self):
@@ -335,25 +390,38 @@ class InfoGatherer:
                    tg.start_soon(self._monitor_macmon, macmon_path)
                tg.start_soon(self._monitor_system_profiler_thunderbolt_data)
                tg.start_soon(self._monitor_thunderbolt_bridge_status)
+                tg.start_soon(self._monitor_rdma_ctl_status)
            tg.start_soon(self._watch_system_info)
            tg.start_soon(self._monitor_memory_usage)
            tg.start_soon(self._monitor_misc)
+            tg.start_soon(self._monitor_static_info)
+            tg.start_soon(self._monitor_disk_usage)

            nc = await NodeConfig.gather()
            if nc is not None:
                await self.info_sender.send(nc)
-            sni = await StaticNodeInformation.gather()
-            await self.info_sender.send(sni)

    def shutdown(self):
        self._tg.cancel_scope.cancel()

+    async def _monitor_static_info(self):
+        if self.static_info_poll_interval is None:
+            return
+        while True:
+            try:
+                with fail_after(30):
+                    await self.info_sender.send(await StaticNodeInformation.gather())
+            except Exception as e:
+                logger.warning(f"Error gathering static node info: {e}")
+            await anyio.sleep(self.static_info_poll_interval)
+
    async def _monitor_misc(self):
        if self.misc_poll_interval is None:
            return
        while True:
            try:
-                await self.info_sender.send(await MiscData.gather())
+                with fail_after(10):
+                    await self.info_sender.send(await MiscData.gather())
            except Exception as e:
                logger.warning(f"Error gathering misc data: {e}")
            await anyio.sleep(self.misc_poll_interval)
@@ -361,20 +429,26 @@ class InfoGatherer:
    async def _monitor_system_profiler_thunderbolt_data(self):
        if self.system_profiler_interval is None:
            return
-        iface_map = await _gather_iface_map()
-        if iface_map is None:
-            return

        while True:
            try:
-                data = await ThunderboltConnectivity.gather()
-                assert data is not None
+                with fail_after(30):
+                    iface_map = await _gather_iface_map()
+                    if iface_map is None:
+                        raise ValueError("Failed to gather interface map")

-                idents = [it for i in data if (it := i.ident(iface_map)) is not None]
-                await self.info_sender.send(MacThunderboltIdentifiers(idents=idents))
+                    data = await ThunderboltConnectivity.gather()
+                    assert data is not None

-                conns = [it for i in data if (it := i.conn()) is not None]
-                await self.info_sender.send(MacThunderboltConnections(conns=conns))
+                    idents = [
+                        it for i in data if (it := i.ident(iface_map)) is not None
+                    ]
+                    await self.info_sender.send(
+                        MacThunderboltIdentifiers(idents=idents)
+                    )
+
+                    conns = [it for i in data if (it := i.conn()) is not None]
+                    await self.info_sender.send(MacThunderboltConnections(conns=conns))
            except Exception as e:
                logger.warning(f"Error gathering Thunderbolt data: {e}")
            await anyio.sleep(self.system_profiler_interval)
@@ -402,8 +476,9 @@ class InfoGatherer:
            return
        while True:
            try:
-                nics = await get_network_interfaces()
-                await self.info_sender.send(NodeNetworkInterfaces(ifaces=nics))
+                with fail_after(10):
+                    nics = await get_network_interfaces()
+                    await self.info_sender.send(NodeNetworkInterfaces(ifaces=nics))
            except Exception as e:
                logger.warning(f"Error gathering network interfaces: {e}")
            await anyio.sleep(self.interface_watcher_interval)
@@ -413,37 +488,70 @@ class InfoGatherer:
            return
        while True:
            try:
-                curr = await ThunderboltBridgeInfo.gather()
-                if curr is not None:
-                    await self.info_sender.send(curr)
+                with fail_after(30):
+                    curr = await ThunderboltBridgeInfo.gather()
+                    if curr is not None:
+                        await self.info_sender.send(curr)
            except Exception as e:
                logger.warning(f"Error gathering Thunderbolt Bridge status: {e}")
            await anyio.sleep(self.thunderbolt_bridge_poll_interval)

+    async def _monitor_rdma_ctl_status(self):
+        if self.rdma_ctl_poll_interval is None:
+            return
+        while True:
+            try:
+                curr = await RdmaCtlStatus.gather()
+                if curr is not None:
+                    await self.info_sender.send(curr)
+            except Exception as e:
+                logger.warning(f"Error gathering RDMA ctl status: {e}")
+            await anyio.sleep(self.rdma_ctl_poll_interval)
+
+    async def _monitor_disk_usage(self):
+        if self.disk_poll_interval is None:
+            return
+        while True:
+            try:
+                with fail_after(5):
+                    await self.info_sender.send(await NodeDiskUsage.gather())
+            except Exception as e:
+                logger.warning(f"Error gathering disk usage: {e}")
+            await anyio.sleep(self.disk_poll_interval)
+
    async def _monitor_macmon(self, macmon_path: str):
        if self.macmon_interval is None:
            return
        # macmon pipe --interval [interval in ms]
-        try:
-            async with await open_process(
-                [macmon_path, "pipe", "--interval", str(self.macmon_interval * 1000)]
-            ) as p:
-                if not p.stdout:
-                    logger.critical("MacMon closed stdout")
-                    return
-                async for text in TextReceiveStream(
-                    BufferedByteReceiveStream(p.stdout)
-                ):
-                    await self.info_sender.send(MacmonMetrics.from_raw_json(text))
-        except CalledProcessError as e:
-            stderr_msg = "no stderr"
-            stderr_output = cast(bytes | str | None, e.stderr)
-            if stderr_output is not None:
-                stderr_msg = (
-                    stderr_output.decode()
-                    if isinstance(stderr_output, bytes)
-                    else str(stderr_output)
+        while True:
+            try:
+                async with await open_process(
+                    [
+                        macmon_path,
+                        "pipe",
+                        "--interval",
+                        str(self.macmon_interval * 1000),
+                    ]
+                ) as p:
+                    if not p.stdout:
+                        logger.critical("MacMon closed stdout")
+                        return
+                    async for text in TextReceiveStream(
+                        BufferedByteReceiveStream(p.stdout)
+                    ):
+                        await self.info_sender.send(MacmonMetrics.from_raw_json(text))
+            except CalledProcessError as e:
+                stderr_msg = "no stderr"
+                stderr_output = cast(bytes | str | None, e.stderr)
+                if stderr_output is not None:
+                    stderr_msg = (
+                        stderr_output.decode()
+                        if isinstance(stderr_output, bytes)
+                        else str(stderr_output)
+                    )
+                logger.warning(
+                    f"MacMon failed with return code {e.returncode}: {stderr_msg}"
                )
-            logger.warning(
-                f"MacMon failed with return code {e.returncode}: {stderr_msg}"
-            )
+            except Exception as e:
+                logger.warning(f"Error in macmon monitor: {e}")
+            await anyio.sleep(self.macmon_interval)
@@ -1,3 +1,4 @@
+import platform
 import socket
 import sys
 from subprocess import CalledProcessError
@@ -8,6 +9,34 @@ from anyio import run_process
 from exo.shared.types.profiling import InterfaceType, NetworkInterfaceInfo


+def get_os_version() -> str:
+    """Return the OS version string for this node.
+
+    On macOS this is the macOS version (e.g. ``"15.3"``).
+    On other platforms it falls back to the platform name (e.g. ``"Linux"``).
+    """
+    if sys.platform == "darwin":
+        version = platform.mac_ver()[0]
+        return version if version else "Unknown"
+    return platform.system() or "Unknown"
+
+
+async def get_os_build_version() -> str:
+    """Return the macOS build version string (e.g. ``"24D5055b"``).
+
+    On non-macOS platforms, returns ``"Unknown"``.
+    """
+    if sys.platform != "darwin":
+        return "Unknown"
+
+    try:
+        process = await run_process(["sw_vers", "-buildVersion"])
+    except CalledProcessError:
+        return "Unknown"
+
+    return process.stdout.decode("utf-8", errors="replace").strip() or "Unknown"
+
+
 async def get_friendly_name() -> str:
    """
    Asynchronously gets the 'Computer Name' (friendly name) of a Mac.
@@ -35,6 +35,9 @@ from mlx_lm.models.qwen3_moe import Model as Qwen3MoeModel
 from mlx_lm.models.qwen3_moe import Qwen3MoeSparseMoeBlock
 from mlx_lm.models.qwen3_next import Model as Qwen3NextModel
 from mlx_lm.models.qwen3_next import Qwen3NextDecoderLayer, Qwen3NextSparseMoeBlock
+from mlx_lm.models.step3p5 import Model as Step35Model
+from mlx_lm.models.step3p5 import Step3p5MLP as Step35MLP
+from mlx_lm.models.step3p5 import Step3p5Model as Step35InnerModel
 from transformers.models.qwen3.modeling_qwen3 import Qwen3DecoderLayer

 from exo.shared.logging import logger
@@ -264,6 +267,19 @@ def pipeline_auto_parallel(
            )
        )

+    if isinstance(inner_model_instance, Step35InnerModel):
+        inner_model_instance.num_layers = len(layers)
+        sliding_layers = [
+            i for i, layer in enumerate(layers) if getattr(layer, "is_sliding", False)
+        ]
+        full_layers = [
+            i
+            for i, layer in enumerate(layers)
+            if not getattr(layer, "is_sliding", True)
+        ]
+        inner_model_instance._swa_idx = 0 if not sliding_layers else sliding_layers[0]
+        inner_model_instance._full_idx = 0 if not full_layers else full_layers[0]
+
    _set_layers(model, layers)

    assert isinstance(layers, list), (
@@ -427,6 +443,14 @@ def tensor_auto_parallel(
            all_to_sharded_linear_in_place,
            sharded_to_all_linear_in_place,
        )
+    elif isinstance(model, Step35Model):
+        tensor_parallel_sharding_strategy = Step35ShardingStrategy(
+            group,
+            all_to_sharded_linear,
+            sharded_to_all_linear,
+            all_to_sharded_linear_in_place,
+            sharded_to_all_linear_in_place,
+        )
    else:
        raise ValueError(f"Unsupported model type: {type(model)}")

@@ -981,3 +1005,46 @@ class GptOssShardingStrategy(TensorParallelShardingStrategy):
            layer.mlp.sharding_group = self.group  # pyright: ignore[reportAttributeAccessIssue]
            mx.eval(layer)
        return model
+
+
+class Step35ShardingStrategy(TensorParallelShardingStrategy):
+    def shard_model(
+        self,
+        model: nn.Module,
+        timeout_seconds: float,
+        on_timeout: TimeoutCallback | None,
+    ) -> nn.Module:
+        model = cast(Step35Model, model)
+
+        for layer in model.layers:
+            eval_with_timeout(
+                layer.parameters(), timeout_seconds / len(model.layers), on_timeout
+            )
+            layer.self_attn.q_proj = self.all_to_sharded_linear(layer.self_attn.q_proj)
+            layer.self_attn.k_proj = self.all_to_sharded_linear(layer.self_attn.k_proj)
+            layer.self_attn.v_proj = self.all_to_sharded_linear(layer.self_attn.v_proj)
+            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
+
+            layer.self_attn.num_heads //= self.N
+            layer.self_attn.num_kv_heads //= self.N
+
+            if getattr(layer.self_attn, "use_head_wise_attn_gate", False):
+                layer.self_attn.g_proj = self.all_to_sharded_linear(
+                    layer.self_attn.g_proj
+                )
+
+            if isinstance(layer.mlp, Step35MLP):
+                layer.mlp.gate_proj = self.all_to_sharded_linear(layer.mlp.gate_proj)
+                layer.mlp.up_proj = self.all_to_sharded_linear(layer.mlp.up_proj)
+                layer.mlp.down_proj = self.sharded_to_all_linear(layer.mlp.down_proj)
+            else:
+                layer.mlp.sharding_group = self.group
+                self.all_to_sharded_linear_in_place(layer.mlp.share_expert.gate_proj)
+                self.all_to_sharded_linear_in_place(layer.mlp.share_expert.up_proj)
+                self.sharded_to_all_linear_in_place(layer.mlp.share_expert.down_proj)
+                self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.gate_proj)
+                self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.up_proj)
+                self.sharded_to_all_linear_in_place(layer.mlp.switch_mlp.down_proj)
+
+            mx.eval(layer)
+        return model
@@ -393,10 +393,11 @@ def mlx_generate(
                    f"Model generated unexpected finish_reason: {out.finish_reason}"
                )

+            total_prompt_tokens = len(all_prompt_tokens)
            usage = Usage(
-                prompt_tokens=int(out.prompt_tokens),
+                prompt_tokens=total_prompt_tokens,
                completion_tokens=completion_tokens,
-                total_tokens=int(out.prompt_tokens) + completion_tokens,
+                total_tokens=total_prompt_tokens + completion_tokens,
                prompt_tokens_details=PromptTokensDetails(
                    cached_tokens=prefix_hit_length
                ),
@@ -1,6 +1,5 @@
 import json
 import os
-import resource
 import sys
 import time
 from pathlib import Path
@@ -63,12 +62,8 @@ from exo.worker.engines.mlx.auto_parallel import (
 from exo.worker.runner.bootstrap import logger

 Group = mx.distributed.Group
-# Needed for 8 bit model
-resource.setrlimit(resource.RLIMIT_NOFILE, (2048, 4096))


-# TODO: Test this
-#  ALSO https://github.com/exo-explore/exo/pull/233#discussion_r2549683673
 def get_weights_size(model_shard_meta: ShardMetadata) -> Memory:
    return Memory.from_float_kb(
        (model_shard_meta.end_layer - model_shard_meta.start_layer)
@@ -86,30 +81,6 @@ class ModelLoadingTimeoutError(Exception):
    pass


-def mx_barrier(group: Group | None = None):
-    mx.eval(
-        mx.distributed.all_sum(
-            mx.array(1.0),
-            stream=mx.default_stream(mx.Device(mx.cpu)),
-            group=group,
-        )
-    )
-
-
-def broadcast_from_zero(value: int, group: Group | None = None):
-    if group is None:
-        return value
-
-    if group.rank() == 0:
-        a = mx.array([value], dtype=mx.int32)
-    else:
-        a = mx.array([0], dtype=mx.int32)
-
-    m = mx.distributed.all_sum(a, stream=mx.Device(mx.DeviceType.cpu), group=group)
-    mx.eval(m)
-    return int(m.item())
-
-
 class HostList(RootModel[list[str]]):
    @classmethod
    def from_hosts(cls, hosts: list[Host]) -> "HostList":
@@ -465,11 +436,19 @@ def apply_chat_template(
        partial_assistant_content = cast(str, formatted_messages[-1].get("content", ""))
        formatted_messages = formatted_messages[:-1]

+    extra_kwargs: dict[str, Any] = {}
+    if task_params.enable_thinking is not None:
+        # Qwen3 and GLM use "enable_thinking"; DeepSeek uses "thinking".
+        # Jinja ignores unknown variables, so passing both is safe.
+        extra_kwargs["enable_thinking"] = task_params.enable_thinking
+        extra_kwargs["thinking"] = task_params.enable_thinking
+
    prompt: str = tokenizer.apply_chat_template(
        formatted_messages,
        tokenize=False,
        add_generation_prompt=True,
        tools=task_params.tools,
+        **extra_kwargs,
    )

    if partial_assistant_content:
@@ -586,3 +565,23 @@ def mlx_cleanup(
    import gc

    gc.collect()
+
+
+def mx_any(bool_: bool, group: Group | None) -> bool:
+    if group is None:
+        return bool_
+    num_true = mx.distributed.all_sum(
+        mx.array(bool_), group=group, stream=mx.default_stream(mx.Device(mx.cpu))
+    )
+    mx.eval(num_true)
+    return num_true.item() > 0
+
+
+def mx_barrier(group: Group | None):
+    if group is None:
+        return
+    mx.eval(
+        mx.distributed.all_sum(
+            mx.array(1.0), group=group, stream=mx.default_stream(mx.Device(mx.cpu))
+        )
+    )
@@ -33,6 +33,7 @@ from exo.shared.types.events import (
 from exo.shared.types.multiaddr import Multiaddr
 from exo.shared.types.state import State
 from exo.shared.types.tasks import (
+    CancelTask,
    CreateRunner,
    DownloadModel,
    ImageEdits,
@@ -184,6 +185,14 @@ class Worker:
            )
            if task is None:
                continue
+
+            # Gate DownloadModel on backoff BEFORE emitting TaskCreated
+            # to prevent flooding the event log with useless events
+            if isinstance(task, DownloadModel):
+                model_id = task.shard_metadata.model_card.model_id
+                if not self._download_backoff.should_proceed(model_id):
+                    continue
+
            logger.info(f"Worker plan: {task.__class__.__name__}")
            assert task.task_status
            await self.event_sender.send(TaskCreated(task_id=task.task_id, task=task))
@@ -199,9 +208,6 @@ class Worker:
                    )
                case DownloadModel(shard_metadata=shard):
                    model_id = shard.model_card.model_id
-                    if not self._download_backoff.should_proceed(model_id):
-                        continue
-
                    self._download_backoff.record_attempt(model_id)

                    await self.download_command_sender.send(
@@ -219,15 +225,22 @@ class Worker:
                        )
                    )
                case Shutdown(runner_id=runner_id):
+                    runner = self.runners.pop(runner_id)
                    try:
                        with fail_after(3):
-                            await self.runners.pop(runner_id).start_task(task)
+                            await runner.start_task(task)
                    except TimeoutError:
                        await self.event_sender.send(
                            TaskStatusUpdated(
                                task_id=task.task_id, task_status=TaskStatus.TimedOut
                            )
                        )
+                    finally:
+                        runner.shutdown()
+                case CancelTask(
+                    cancelled_task_id=cancelled_task_id, runner_id=runner_id
+                ):
+                    await self.runners[runner_id].cancel_task(cancelled_task_id)
                case ImageEdits() if task.task_params.total_input_chunks > 0:
                    # Assemble image from chunks and inject into task
                    cmd_id = task.command_id
@@ -265,18 +278,18 @@ class Worker:
                        del self.input_chunk_buffer[cmd_id]
                    if cmd_id in self.input_chunk_counts:
                        del self.input_chunk_counts[cmd_id]
-                    await self.runners[self._task_to_runner_id(task)].start_task(
-                        modified_task
-                    )
+                    await self._start_runner_task(modified_task)
                case task:
-                    await self.runners[self._task_to_runner_id(task)].start_task(task)
+                    await self._start_runner_task(task)

    def shutdown(self):
        self._tg.cancel_scope.cancel()

-    def _task_to_runner_id(self, task: Task):
-        instance = self.state.instances[task.instance_id]
-        return instance.shard_assignments.node_to_runner[self.node_id]
+    async def _start_runner_task(self, task: Task):
+        if (instance := self.state.instances.get(task.instance_id)) is not None:
+            await self.runners[
+                instance.shard_assignments.node_to_runner[self.node_id]
+            ].start_task(task)

    async def _nack_request(self, since_idx: int) -> None:
        # We request all events after (and including) the missing index.
@@ -315,8 +328,6 @@ class Worker:
            for event in self.out_for_delivery.copy().values():
                await self.local_event_sender.send(event)

-    ## Op Executors
-
    def _create_supervisor(self, task: CreateRunner) -> RunnerSupervisor:
        """Creates and stores a new AssignedRunner with initial downloading status."""
        runner = RunnerSupervisor.create(
@@ -4,6 +4,7 @@ from collections.abc import Mapping, Sequence

 from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.tasks import (
+    CancelTask,
    ConnectToGroup,
    CreateRunner,
    DownloadModel,
@@ -53,13 +54,14 @@ def plan(
 ) -> Task | None:
    # Python short circuiting OR logic should evaluate these sequentially.
    return (
-        _kill_runner(runners, all_runners, instances)
+        _cancel_tasks(runners, tasks)
+        or _kill_runner(runners, all_runners, instances)
        or _create_runner(node_id, runners, instances)
        or _model_needs_download(node_id, runners, global_download_status)
        or _init_distributed_backend(runners, all_runners)
        or _load_model(runners, all_runners, global_download_status)
        or _ready_to_warmup(runners, all_runners)
-        or _pending_tasks(runners, tasks, all_runners, input_chunk_buffer)
+        or _pending_tasks(runners, tasks, all_runners, input_chunk_buffer or {})
    )


@@ -270,7 +272,7 @@ def _pending_tasks(
    runners: Mapping[RunnerId, RunnerSupervisor],
    tasks: Mapping[TaskId, Task],
    all_runners: Mapping[RunnerId, RunnerStatus],
-    input_chunk_buffer: Mapping[CommandId, dict[int, str]] | None = None,
+    input_chunk_buffer: Mapping[CommandId, dict[int, str]],
 ) -> Task | None:
    for task in tasks.values():
        # for now, just forward chat completions
@@ -284,7 +286,7 @@ def _pending_tasks(
        if isinstance(task, ImageEdits) and task.task_params.total_input_chunks > 0:
            cmd_id = task.command_id
            expected = task.task_params.total_input_chunks
-            received = len((input_chunk_buffer or {}).get(cmd_id, {}))
+            received = len(input_chunk_buffer.get(cmd_id, {}))
            if received < expected:
                continue  # Wait for all chunks to arrive

@@ -292,16 +294,33 @@ def _pending_tasks(
            if task.instance_id != runner.bound_instance.instance.instance_id:
                continue

-            # I have a design point here; this is a state race in disguise as the task status doesn't get updated to completed fast enough
-            # however, realistically the task status should be set to completed by the LAST runner, so this is a true race
-            # the actual solution is somewhat deeper than this bypass - TODO!
+            # the task status _should_ be set to completed by the LAST runner
+            # it is currently set by the first
+            # this is definitely a hack
            if task.task_id in runner.completed:
                continue

-            # TODO: Check ordering aligns with MLX distributeds expectations.
-
            if isinstance(runner.status, RunnerReady) and all(
                isinstance(all_runners[global_runner_id], (RunnerReady, RunnerRunning))
                for global_runner_id in runner.bound_instance.instance.shard_assignments.runner_to_shard
            ):
                return task
+
+
+def _cancel_tasks(
+    runners: Mapping[RunnerId, RunnerSupervisor],
+    tasks: Mapping[TaskId, Task],
+) -> Task | None:
+    for task in tasks.values():
+        if task.task_status != TaskStatus.Cancelled:
+            continue
+        for runner_id, runner in runners.items():
+            if task.instance_id != runner.bound_instance.instance.instance_id:
+                continue
+            if task.task_id in runner.cancelled:
+                continue
+            return CancelTask(
+                instance_id=task.instance_id,
+                cancelled_task_id=task.task_id,
+                runner_id=runner_id,
+            )
@@ -1,21 +1,36 @@
+from __future__ import annotations
+
 import os
+import threading
+from multiprocessing.sharedctypes import Synchronized

 import loguru

 from exo.shared.types.events import Event, RunnerStatusUpdated
-from exo.shared.types.tasks import Task
+from exo.shared.types.tasks import Task, TaskId
 from exo.shared.types.worker.instances import BoundInstance, MlxJacclInstance
 from exo.shared.types.worker.runners import RunnerFailed
 from exo.utils.channels import ClosedResourceError, MpReceiver, MpSender

 logger: "loguru.Logger" = loguru.logger

+HEARTBEAT_INTERVAL_SECONDS = 0.5
+
+
+def _heartbeat_loop(heartbeat: Synchronized[int], stop: threading.Event) -> None:
+    """Daemon thread that periodically increments the heartbeat counter."""
+    while not stop.is_set():
+        heartbeat.value += 1
+        stop.wait(HEARTBEAT_INTERVAL_SECONDS)
+

 def entrypoint(
    bound_instance: BoundInstance,
    event_sender: MpSender[Event],
    task_receiver: MpReceiver[Task],
+    cancel_receiver: MpReceiver[TaskId],
    _logger: "loguru.Logger",
+    heartbeat: Synchronized[int] | None = None,
 ) -> None:
    fast_synch_override = os.environ.get("EXO_FAST_SYNCH")
    if fast_synch_override == "on" or (
@@ -34,11 +49,22 @@ def entrypoint(

    logger.info(f"Fast synch flag: {os.environ['MLX_METAL_FAST_SYNCH']}")

+    # Start heartbeat thread so the supervisor can detect if we freeze.
+    stop_heartbeat = threading.Event()
+    heartbeat_thread: threading.Thread | None = None
+    if heartbeat is not None:
+        heartbeat_thread = threading.Thread(
+            target=_heartbeat_loop,
+            args=(heartbeat, stop_heartbeat),
+            daemon=True,
+        )
+        heartbeat_thread.start()
+
    # Import main after setting global logger - this lets us just import logger from this module
    try:
        from exo.worker.runner.runner import main

-        main(bound_instance, event_sender, task_receiver)
+        main(bound_instance, event_sender, task_receiver, cancel_receiver)
    except ClosedResourceError:
        logger.warning("Runner communication closed unexpectedly")
    except Exception as e:
@@ -52,6 +78,9 @@ def entrypoint(
            )
        )
    finally:
+        stop_heartbeat.set()
+        if heartbeat_thread is not None:
+            heartbeat_thread.join(timeout=1)
        try:
            event_sender.close()
            task_receiver.close()
@@ -1,5 +1,8 @@
 import base64
 import json
+import math
+import os
+import resource
 import time
 from collections.abc import Generator
 from functools import cache
@@ -87,6 +90,7 @@ from exo.worker.engines.mlx.utils_mlx import (
    initialize_mlx,
    load_mlx_items,
    mlx_force_oom,
+    mx_any,
 )
 from exo.worker.runner.bootstrap import logger

@@ -111,7 +115,11 @@ def main(
    bound_instance: BoundInstance,
    event_sender: MpSender[Event],
    task_receiver: MpReceiver[Task],
+    cancel_receiver: MpReceiver[TaskId],
 ):
+    soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
+    resource.setrlimit(resource.RLIMIT_NOFILE, (min(max(soft, 2048), hard), hard))
+
    instance, runner_id, shard_metadata = (
        bound_instance.instance,
        bound_instance.bound_runner_id,
@@ -125,11 +133,15 @@ def main(
        time.sleep(timeout)

    setup_start_time = time.time()
+    cancelled_tasks = set[TaskId]()

-    model: Model | DistributedImageModel | None = None
+    # type checker was unhappy with me - splitting these fixed it
+    inference_model: Model | None = None
+    image_model: DistributedImageModel | None = None
    tokenizer = None
    group = None
    kv_prefix_cache: KVPrefixCache | None = None
+    check_for_cancel_every: int | None = None

    current_status: RunnerStatus = RunnerIdle()
    logger.info("runner created")
@@ -142,6 +154,7 @@ def main(
            if task.task_id in seen:
                logger.warning("repeat task - potential error")
            seen.add(task.task_id)
+            cancelled_tasks.discard(TaskId("CANCEL_CURRENT_TASK"))
            event_sender.send(
                TaskStatusUpdated(task_id=task.task_id, task_status=TaskStatus.Running)
            )
@@ -187,7 +200,7 @@ def main(
                        time.sleep(0.5)

                    if ModelTask.TextGeneration in shard_metadata.model_card.tasks:
-                        model, tokenizer = load_mlx_items(
+                        inference_model, tokenizer = load_mlx_items(
                            bound_instance, group, on_timeout=on_model_load_timeout
                        )
                        logger.info(
@@ -199,7 +212,7 @@ def main(
                        ModelTask.TextToImage in shard_metadata.model_card.tasks
                        or ModelTask.ImageToImage in shard_metadata.model_card.tasks
                    ):
-                        model = initialize_image_model(bound_instance)
+                        image_model = initialize_image_model(bound_instance)
                    else:
                        raise ValueError(
                            f"Unknown model task(s): {shard_metadata.model_card.tasks}"
@@ -207,8 +220,6 @@ def main(
                    current_status = RunnerLoaded()
                    logger.info("runner loaded")
                case StartWarmup() if isinstance(current_status, RunnerLoaded):
-                    assert model
-
                    current_status = RunnerWarmingUp()
                    logger.info("runner warming up")
                    event_sender.send(
@@ -220,16 +231,31 @@ def main(

                    logger.info(f"warming up inference for instance: {instance}")
                    if ModelTask.TextGeneration in shard_metadata.model_card.tasks:
-                        assert not isinstance(model, DistributedImageModel)
+                        assert inference_model
                        assert tokenizer

+                        t = time.monotonic()
                        toks = warmup_inference(
-                            model=model,
+                            model=inference_model,
                            tokenizer=tokenizer,
                            group=group,
-                            # kv_prefix_cache=kv_prefix_cache,  # supply for warmup-time prefix caching
                        )
                        logger.info(f"warmed up by generating {toks} tokens")
+                        check_for_cancel_every = min(
+                            math.ceil(toks / min(time.monotonic() - t, 0.001)), 100
+                        )
+                        if group is not None:
+                            check_for_cancel_every = int(
+                                mx.max(
+                                    mx.distributed.all_gather(
+                                        mx.array([check_for_cancel_every]), group=group
+                                    )
+                                ).item()
+                            )
+
+                        logger.info(
+                            f"runner checking for cancellation every {check_for_cancel_every} tokens"
+                        )
                        logger.info(
                            f"runner initialized in {time.time() - setup_start_time} seconds"
                        )
@@ -237,8 +263,8 @@ def main(
                        ModelTask.TextToImage in shard_metadata.model_card.tasks
                        or ModelTask.ImageToImage in shard_metadata.model_card.tasks
                    ):
-                        assert isinstance(model, DistributedImageModel)
-                        image = warmup_image_generator(model=model)
+                        assert image_model
+                        image = warmup_image_generator(model=image_model)
                        if image is not None:
                            logger.info(f"warmed up by generating {image.size} image")
                        else:
@@ -258,9 +284,9 @@ def main(
                        )
                    )
                    event_sender.send(TaskAcknowledged(task_id=task.task_id))
-
-                    assert model and not isinstance(model, DistributedImageModel)
+                    assert inference_model
                    assert tokenizer
+                    assert check_for_cancel_every

                    try:
                        _check_for_debug_prompts(task_params)
@@ -270,7 +296,7 @@ def main(

                        # Generate responses using the actual MLX generation
                        mlx_generator = mlx_generate(
-                            model=model,
+                            model=inference_model,
                            tokenizer=tokenizer,
                            task=task_params,
                            prompt=prompt,
@@ -295,11 +321,11 @@ def main(
                            patch_glm_tokenizer(tokenizer)

                        # GPT-OSS specific parsing to match other model formats.
-                        elif isinstance(model, GptOssModel):
+                        elif isinstance(inference_model, GptOssModel):
                            mlx_generator = parse_gpt_oss(mlx_generator)

                        if tokenizer.has_tool_calling and not isinstance(
-                            model, GptOssModel
+                            inference_model, GptOssModel
                        ):
                            assert tokenizer.tool_call_start
                            assert tokenizer.tool_call_end
@@ -312,7 +338,18 @@ def main(
                            )

                        completion_tokens = 0
+                        tokens_since_last_cancel_check = 0
                        for response in mlx_generator:
+                            tokens_since_last_cancel_check += 1
+                            if tokens_since_last_cancel_check >= check_for_cancel_every:
+                                tokens_since_last_cancel_check = 0
+                                cancelled_tasks.update(cancel_receiver.collect())
+                                want_to_cancel = (task.task_id in cancelled_tasks) or (
+                                    TaskId("CANCEL_CURRENT_TASK") in cancelled_tasks
+                                )
+                                if mx_any(want_to_cancel, group):
+                                    break
+
                            match response:
                                case GenerationResponse():
                                    completion_tokens += 1
@@ -360,6 +397,7 @@ def main(
                                                    tool_calls=response.tool_calls,
                                                    model=shard_metadata.model_card.model_id,
                                                    usage=response.usage,
+                                                    stats=response.stats,
                                                ),
                                            )
                                        )
@@ -384,7 +422,7 @@ def main(
                case ImageGeneration(
                    task_params=task_params, command_id=command_id
                ) if isinstance(current_status, RunnerReady):
-                    assert isinstance(model, DistributedImageModel)
+                    assert image_model
                    logger.info(f"received image generation request: {str(task)[:500]}")
                    current_status = RunnerRunning()
                    logger.info("runner running")
@@ -397,7 +435,9 @@ def main(

                    try:
                        image_index = 0
-                        for response in generate_image(model=model, task=task_params):
+                        for response in generate_image(
+                            model=image_model, task=task_params
+                        ):
                            is_primary_output = _is_primary_output_node(shard_metadata)

                            if is_primary_output:
@@ -447,7 +487,7 @@ def main(
                case ImageEdits(task_params=task_params, command_id=command_id) if (
                    isinstance(current_status, RunnerReady)
                ):
-                    assert isinstance(model, DistributedImageModel)
+                    assert image_model
                    logger.info(f"received image edits request: {str(task)[:500]}")
                    current_status = RunnerRunning()
                    logger.info("runner running")
@@ -460,7 +500,9 @@ def main(

                    try:
                        image_index = 0
-                        for response in generate_image(model=model, task=task_params):
+                        for response in generate_image(
+                            model=image_model, task=task_params
+                        ):
                            if _is_primary_output_node(shard_metadata):
                                match response:
                                    case PartialImageResponse():
@@ -519,14 +561,20 @@ def main(
                    raise ValueError(
                        f"Received {task.__class__.__name__} outside of state machine in {current_status=}"
                    )
-            event_sender.send(
-                TaskStatusUpdated(task_id=task.task_id, task_status=TaskStatus.Complete)
+            was_cancelled = (task.task_id in cancelled_tasks) or (
+                TaskId("CANCEL_CURRENT_TASK") in cancelled_tasks
            )
+            if not was_cancelled:
+                event_sender.send(
+                    TaskStatusUpdated(
+                        task_id=task.task_id, task_status=TaskStatus.Complete
+                    )
+                )
            event_sender.send(
                RunnerStatusUpdated(runner_id=runner_id, runner_status=current_status)
            )
            if isinstance(current_status, RunnerShutdown):
-                del model, tokenizer, group
+                del inference_model, image_model, tokenizer, group
                mx.clear_cache()
                import gc

@@ -760,7 +808,9 @@ def parse_tool_calls(
                    tools = [_validate_single_tool(tool) for tool in parsed]
                else:
                    tools = [_validate_single_tool(parsed)]
-                yield ToolCallResponse(tool_calls=tools, usage=response.usage)
+                yield ToolCallResponse(
+                    tool_calls=tools, usage=response.usage, stats=response.stats
+                )

            except (
                json.JSONDecodeError,
@@ -791,7 +841,8 @@ def parse_tool_calls(
                    text=tool_call_start + "".join(tool_call_text_parts),
                    token=0,
                    finish_reason=response.finish_reason,
-                    usage=None,
+                    usage=response.usage,
+                    stats=response.stats,
                )
            continue
        # fallthrough
@@ -949,6 +1000,7 @@ def _validate_single_tool(obj: dict[str, Any]) -> ToolCallItem:
 EXO_RUNNER_MUST_FAIL = "EXO RUNNER MUST FAIL"
 EXO_RUNNER_MUST_OOM = "EXO RUNNER MUST OOM"
 EXO_RUNNER_MUST_TIMEOUT = "EXO RUNNER MUST TIMEOUT"
+EXO_RUNNER_MUST_DIE = "EXO RUNNER MUST DIE"


 def _check_for_debug_prompts(task_params: TextGenerationTaskParams) -> None:
@@ -964,6 +1016,9 @@ def _check_for_debug_prompts(task_params: TextGenerationTaskParams) -> None:
    if not prompt:
        return

+    if EXO_RUNNER_MUST_DIE in prompt:
+        logger.info("Abrupt process death triggered (simulates OOM kill)")
+        os._exit(1)
    if EXO_RUNNER_MUST_FAIL in prompt:
        logger.info("raising exception")
        raise Exception("Artificial runner exception - for testing purposes only.")
@@ -1,12 +1,17 @@
+from __future__ import annotations
+
 import contextlib
+import multiprocessing
 import signal
 from dataclasses import dataclass, field
 from multiprocessing import Process
+from multiprocessing.sharedctypes import Synchronized
 from typing import Self

 import anyio
 from anyio import (
    BrokenResourceError,
+    CancelScope,
    ClosedResourceError,
    to_thread,
 )
@@ -26,6 +31,7 @@ from exo.shared.types.worker.runners import (
    RunnerIdle,
    RunnerLoading,
    RunnerRunning,
+    RunnerShutdown,
    RunnerShuttingDown,
    RunnerStatus,
    RunnerWarmingUp,
@@ -36,6 +42,8 @@ from exo.worker.runner.bootstrap import entrypoint

 PREFILL_TIMEOUT_SECONDS = 60
 DECODE_TIMEOUT_SECONDS = 5
+HEALTH_CHECK_INTERVAL_SECONDS = 1
+HEARTBEAT_STALE_CHECKS = 10


@dataclass(eq=False)
@@ -47,9 +55,15 @@ class RunnerSupervisor:
    _ev_recv: MpReceiver[Event]
    _task_sender: MpSender[Task]
    _event_sender: Sender[Event]
+    _cancel_sender: MpSender[TaskId]
+    _heartbeat: Synchronized[int]
    status: RunnerStatus = field(default_factory=RunnerIdle, init=False)
    pending: dict[TaskId, anyio.Event] = field(default_factory=dict, init=False)
    completed: set[TaskId] = field(default_factory=set, init=False)
+    cancelled: set[TaskId] = field(default_factory=set, init=False)
+    _death_handled: bool = field(default=False, init=False)
+    _last_heartbeat_value: int = field(default=0, init=False)
+    _heartbeat_stale_count: int = field(default=0, init=False)

    @classmethod
    def create(
@@ -60,8 +74,10 @@ class RunnerSupervisor:
        initialize_timeout: float = 400,
    ) -> Self:
        ev_send, ev_recv = mp_channel[Event]()
-        # A task is kind of a runner command
        task_sender, task_recv = mp_channel[Task]()
+        cancel_sender, cancel_recv = mp_channel[TaskId]()
+
+        heartbeat: Synchronized[int] = multiprocessing.Value("Q", 0)

        runner_process = Process(
            target=entrypoint,
@@ -69,7 +85,9 @@ class RunnerSupervisor:
                bound_instance,
                ev_send,
                task_recv,
+                cancel_recv,
                logger,
+                heartbeat,
            ),
            daemon=True,
        )
@@ -83,20 +101,26 @@ class RunnerSupervisor:
            initialize_timeout=initialize_timeout,
            _ev_recv=ev_recv,
            _task_sender=task_sender,
+            _cancel_sender=cancel_sender,
            _event_sender=event_sender,
+            _heartbeat=heartbeat,
        )

        return self

    async def run(self):
        self.runner_process.start()
-        await self._forward_events()
+        async with anyio.create_task_group() as tg:
+            tg.start_soon(self._forward_events)
+            tg.start_soon(self._health_check, tg.cancel_scope)

    def shutdown(self):
        logger.info("Runner supervisor shutting down")
        self._ev_recv.close()
        self._task_sender.close()
        self._event_sender.close()
+        self._cancel_sender.send(TaskId("CANCEL_CURRENT_TASK"))
+        self._cancel_sender.close()
        self.runner_process.join(1)
        if not self.runner_process.is_alive():
            logger.info("Runner process succesfully terminated")
@@ -112,14 +136,6 @@ class RunnerSupervisor:
        logger.critical("Runner process didn't respond to SIGTERM, killing")
        self.runner_process.kill()

-        self.runner_process.join(1)
-        if not self.runner_process.is_alive():
-            return
-
-        logger.critical(
-            "Runner process didn't respond to SIGKILL. System resources may have leaked"
-        )
-
    async def start_task(self, task: Task):
        if task.task_id in self.pending:
            logger.warning(
@@ -141,6 +157,17 @@ class RunnerSupervisor:
            return
        await event.wait()

+    async def cancel_task(self, task_id: TaskId):
+        if task_id in self.completed:
+            logger.info(f"Unable to cancel {task_id} as it has been completed")
+            return
+        self.cancelled.add(task_id)
+        with anyio.move_on_after(0.5) as scope:
+            await self._cancel_sender.send_async(task_id)
+        if scope.cancel_called:
+            logger.error("RunnerSupervisor cancel pipe blocked")
+            await self._check_runner(TimeoutError("cancel pipe blocked"))
+
    async def _forward_events(self):
        with self._ev_recv as events:
            try:
@@ -168,9 +195,99 @@ class RunnerSupervisor:
                        self.completed.add(event.task_id)
                    await self._event_sender.send(event)
            except (ClosedResourceError, BrokenResourceError) as e:
-                await self._check_runner(e)
-                for tid in self.pending:
-                    self.pending[tid].set()
+                if not self._death_handled:
+                    self._death_handled = True
+                    await self._check_runner(e)
+                    for tid in self.pending:
+                        self.pending[tid].set()
+
+    async def _health_check(self, cancel_scope: CancelScope) -> None:
+        """Periodically check if the runner process is alive and responsive.
+
+        Detects two failure modes:
+        1. Process death (e.g. OOM kill) without cleanly closing the event
+           channel, which would leave _forward_events blocked on queue.get().
+        2. Unresponsive process (e.g. frozen by OS memory pressure, deadlock)
+           detected via a stale heartbeat counter.
+        """
+        while True:
+            await anyio.sleep(HEALTH_CHECK_INTERVAL_SECONDS)
+
+            if not self.runner_process.is_alive():
+                self._handle_process_exit(cancel_scope)
+                return
+
+            # Check heartbeat counter — if it hasn't changed between
+            # consecutive checks, the subprocess may be frozen.
+            current = self._heartbeat.value
+            if current > 0:
+                if current == self._last_heartbeat_value:
+                    self._heartbeat_stale_count += 1
+                    if self._heartbeat_stale_count >= HEARTBEAT_STALE_CHECKS:
+                        logger.error(
+                            f"Health check: runner process unresponsive "
+                            f"(heartbeat stale for {self._heartbeat_stale_count} checks), killing"
+                        )
+                        self._handle_unresponsive(cancel_scope)
+                        return
+                else:
+                    self._heartbeat_stale_count = 0
+                self._last_heartbeat_value = current
+
+    def _handle_process_exit(self, cancel_scope: CancelScope) -> None:
+        """Handle runner process that has exited."""
+        if not self._death_handled:
+            self._death_handled = True
+            if isinstance(
+                self.status, (RunnerShutdown, RunnerShuttingDown, RunnerFailed)
+            ):
+                logger.info("Health check: runner process exited (expected)")
+            else:
+                rc = self.runner_process.exitcode
+                if isinstance(rc, int) and rc < 0:
+                    sig = -rc
+                    try:
+                        cause = f"signal={sig} ({signal.strsignal(sig)})"
+                    except Exception:
+                        cause = f"signal={sig}"
+                else:
+                    cause = f"exitcode={rc}"
+
+                logger.error(
+                    f"Health check: runner process died unexpectedly ({cause})"
+                )
+                self._event_sender.send_nowait(
+                    RunnerStatusUpdated(
+                        runner_id=self.bound_instance.bound_runner_id,
+                        runner_status=RunnerFailed(
+                            error_message=f"Terminated ({cause})"
+                        ),
+                    )
+                )
+                self.shutdown()
+
+            for tid in self.pending:
+                self.pending[tid].set()
+
+        cancel_scope.cancel()
+
+    def _handle_unresponsive(self, cancel_scope: CancelScope) -> None:
+        """Handle runner process that is alive but unresponsive."""
+        if not self._death_handled:
+            self._death_handled = True
+            self._event_sender.send_nowait(
+                RunnerStatusUpdated(
+                    runner_id=self.bound_instance.bound_runner_id,
+                    runner_status=RunnerFailed(
+                        error_message="Runner process unresponsive (heartbeat timeout)"
+                    ),
+                )
+            )
+            for tid in self.pending:
+                self.pending[tid].set()
+            self.shutdown()
+
+        cancel_scope.cancel()

    def __del__(self) -> None:
        if self.runner_process.is_alive():
@@ -1,7 +1,9 @@
 # Check tasks are complete before runner is ever ready.
+import unittest.mock
 from collections.abc import Iterable
 from typing import Callable

+import mlx.core as mx
 import pytest

 import exo.worker.runner.runner as mlx_runner
@@ -19,6 +21,7 @@ from exo.shared.types.tasks import (
    Shutdown,
    StartWarmup,
    Task,
+    TaskId,
    TaskStatus,
    TextGeneration,
 )
@@ -113,6 +116,7 @@ def patch_out_mlx(monkeypatch: pytest.MonkeyPatch):
    monkeypatch.setattr(mlx_runner, "load_mlx_items", make_nothin((1, MockTokenizer)))
    monkeypatch.setattr(mlx_runner, "warmup_inference", make_nothin(1))
    monkeypatch.setattr(mlx_runner, "_check_for_debug_prompts", nothin)
+    monkeypatch.setattr(mlx_runner, "mx_any", make_nothin(False))
    # Mock apply_chat_template since we're using a fake tokenizer (integer 1).
    # Returns a prompt without thinking tag so detect_thinking_prompt_suffix returns None.
    monkeypatch.setattr(mlx_runner, "apply_chat_template", make_nothin("test prompt"))
@@ -163,6 +167,7 @@ def _run(tasks: Iterable[Task]):
    )

    task_sender, task_receiver = mp_channel[Task]()
+    _cancel_sender, cancel_receiver = mp_channel[TaskId]()
    event_sender = EventCollector()

    with task_sender:
@@ -173,8 +178,16 @@ def _run(tasks: Iterable[Task]):
        # this is some c++ nonsense
        task_receiver.close = nothin
        task_receiver.join = nothin
-
-        mlx_runner.main(bound_instance, event_sender, task_receiver)  # type: ignore[arg-type]
+        with unittest.mock.patch(
+            "exo.worker.runner.runner.mx.distributed.all_gather",
+            make_nothin(mx.array([1])),
+        ):
+            mlx_runner.main(
+                bound_instance,
+                event_sender,  # pyright: ignore[reportArgumentType]
+                task_receiver,
+                cancel_receiver,
+            )

        return event_sender.events

@@ -1 +1,204 @@
-# TODO:
+from __future__ import annotations
+
+import multiprocessing
+import os
+import signal as signal_module
+from collections.abc import Callable
+from multiprocessing.sharedctypes import Synchronized
+from typing import Any
+
+import anyio
+
+from exo.shared.types.events import Event, RunnerStatusUpdated
+from exo.shared.types.tasks import Task, TaskId
+from exo.shared.types.worker.runners import RunnerFailed, RunnerIdle, RunnerShutdown
+from exo.utils.channels import Receiver, Sender, channel, mp_channel
+from exo.worker.runner.runner_supervisor import (
+    HEALTH_CHECK_INTERVAL_SECONDS,
+    HEARTBEAT_STALE_CHECKS,
+    RunnerSupervisor,
+)
+
+from ...constants import (
+    INSTANCE_1_ID,
+    MODEL_A_ID,
+    NODE_A,
+    RUNNER_1_ID,
+)
+from ..conftest import get_bound_mlx_ring_instance
+
+
+def _die_immediately() -> None:
+    """Subprocess target that exits with a non-zero code."""
+    os._exit(1)
+
+
+def _die_with_signal() -> None:
+    """Subprocess target that kills itself with SIGKILL (simulates OOM)."""
+    os.kill(os.getpid(), signal_module.SIGKILL)
+
+
+def _exit_cleanly() -> None:
+    """Subprocess target that exits with code 0."""
+    os._exit(0)
+
+
+def _hang_forever() -> None:
+    """Subprocess target that hangs without updating heartbeat (simulates freeze)."""
+    import time
+
+    # Write one heartbeat so the supervisor starts tracking, then stop.
+    time.sleep(100000)
+
+
+def _build_supervisor(
+    event_sender: Sender[Event],
+    target: Callable[..., Any],
+) -> RunnerSupervisor:
+    """Build a RunnerSupervisor with a custom subprocess target.
+
+    Uses a clone of event_sender (matching real Worker behavior) so that
+    closing the supervisor's copy doesn't close the test's receiver.
+    """
+    bound_instance = get_bound_mlx_ring_instance(
+        instance_id=INSTANCE_1_ID,
+        model_id=MODEL_A_ID,
+        runner_id=RUNNER_1_ID,
+        node_id=NODE_A,
+    )
+
+    _ev_send, ev_recv = mp_channel[Event]()
+    task_sender, _task_recv = mp_channel[Task]()
+    cancel_sender, _cancel_recv = mp_channel[TaskId]()
+    runner_process = multiprocessing.Process(target=target, daemon=True)
+    heartbeat: Synchronized[int] = multiprocessing.Value("Q", 0)
+
+    return RunnerSupervisor(
+        bound_instance=bound_instance,
+        shard_metadata=bound_instance.bound_shard,
+        runner_process=runner_process,
+        initialize_timeout=10,
+        _ev_recv=ev_recv,
+        _task_sender=task_sender,
+        _cancel_sender=cancel_sender,
+        _event_sender=event_sender.clone(),
+        _heartbeat=heartbeat,
+    )
+
+
+def _collect_failed_events(
+    event_receiver: Receiver[Event],
+) -> list[RunnerFailed]:
+    """Drain the receiver and return all RunnerFailed statuses."""
+    out: list[RunnerFailed] = []
+    while True:
+        try:
+            event = event_receiver.receive_nowait()
+        except Exception:
+            break
+        if isinstance(event, RunnerStatusUpdated) and isinstance(
+            event.runner_status, RunnerFailed
+        ):
+            out.append(event.runner_status)
+    return out
+
+
+async def test_health_check_detects_dead_process():
+    """When the runner process dies with a non-zero exit code, the health check
+    should emit a RunnerFailed event and run() should return."""
+    event_sender, event_receiver = channel[Event]()
+    supervisor = _build_supervisor(event_sender, _die_immediately)
+
+    with anyio.fail_after(HEALTH_CHECK_INTERVAL_SECONDS + 5):
+        await supervisor.run()
+
+    failures = _collect_failed_events(event_receiver)
+    assert len(failures) == 1
+    assert failures[0].error_message is not None
+    assert "exitcode=1" in failures[0].error_message
+
+
+async def test_health_check_detects_signal_death():
+    """When the runner process is killed by a signal (e.g. OOM -> SIGKILL),
+    the health check should report the signal in the failure message."""
+    event_sender, event_receiver = channel[Event]()
+    supervisor = _build_supervisor(event_sender, _die_with_signal)
+
+    with anyio.fail_after(HEALTH_CHECK_INTERVAL_SECONDS + 5):
+        await supervisor.run()
+
+    failures = _collect_failed_events(event_receiver)
+    assert len(failures) == 1
+    assert failures[0].error_message is not None
+    assert "signal=9" in failures[0].error_message
+
+
+async def test_health_check_releases_pending_tasks():
+    """When the runner dies, any pending start_task() waiters should be unblocked."""
+    event_sender, _event_receiver = channel[Event]()
+    supervisor = _build_supervisor(event_sender, _die_immediately)
+
+    # Register a pending waiter as if start_task() was waiting for acknowledgement
+    task_event = anyio.Event()
+    tid = TaskId("pending-task")
+    supervisor.pending[tid] = task_event
+
+    with anyio.fail_after(HEALTH_CHECK_INTERVAL_SECONDS + 5):
+        await supervisor.run()
+
+    assert task_event.is_set()
+
+
+async def test_clean_exit_no_failure_when_shutdown_status():
+    """When the runner was in RunnerShutdown status and exits with code 0,
+    no RunnerFailed event should be emitted."""
+    event_sender, event_receiver = channel[Event]()
+    supervisor = _build_supervisor(event_sender, _exit_cleanly)
+
+    # Simulate that the runner had already reported shutdown via events
+    supervisor.status = RunnerShutdown()
+
+    with anyio.fail_after(HEALTH_CHECK_INTERVAL_SECONDS + 5):
+        await supervisor.run()
+
+    failures = _collect_failed_events(event_receiver)
+    assert len(failures) == 0
+
+
+async def test_unexpected_exit_code_zero_emits_failure():
+    """When the runner exits with code 0 but was NOT in a shutdown state,
+    this is unexpected and should still emit RunnerFailed."""
+    event_sender, event_receiver = channel[Event]()
+    supervisor = _build_supervisor(event_sender, _exit_cleanly)
+
+    assert isinstance(supervisor.status, RunnerIdle)
+
+    with anyio.fail_after(HEALTH_CHECK_INTERVAL_SECONDS + 5):
+        await supervisor.run()
+
+    failures = _collect_failed_events(event_receiver)
+    assert len(failures) == 1
+    assert failures[0].error_message is not None
+    assert "exitcode=0" in failures[0].error_message
+
+
+async def test_heartbeat_timeout_detects_unresponsive_process():
+    """When the runner process is alive but its heartbeat goes stale,
+    the health check should kill it and emit RunnerFailed."""
+    event_sender, event_receiver = channel[Event]()
+    supervisor = _build_supervisor(event_sender, _hang_forever)
+
+    # Pre-seed the heartbeat counter with a non-zero value and set the
+    # supervisor's last-seen value to match so it appears stale immediately.
+    # Set stale count to HEARTBEAT_STALE_CHECKS - 1 so a single check triggers.
+    supervisor._heartbeat.value = 42  # pyright: ignore[reportPrivateUsage]
+    supervisor._last_heartbeat_value = 42  # pyright: ignore[reportPrivateUsage]
+    supervisor._heartbeat_stale_count = HEARTBEAT_STALE_CHECKS - 1  # pyright: ignore[reportPrivateUsage]
+
+    with anyio.fail_after(HEALTH_CHECK_INTERVAL_SECONDS + 5):
+        await supervisor.run()
+
+    failures = _collect_failed_events(event_receiver)
+    assert len(failures) == 1
+    assert failures[0].error_message is not None
+    assert "unresponsive" in failures[0].error_message.lower()
@@ -380,6 +380,7 @@ dependencies = [
    { name = "mlx", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "mlx", extra = ["cpu"], marker = "sys_platform == 'linux'" },
    { name = "mlx-lm", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "msgspec", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "openai-harmony", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "pillow", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "psutil", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -389,6 +390,7 @@ dependencies = [
    { name = "tiktoken", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "tomlkit", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "types-aiofiles", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "zstandard", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]

 [package.dev-dependencies]
@@ -414,9 +416,10 @@ requires-dist = [
    { name = "hypercorn", specifier = ">=0.18.0" },
    { name = "loguru", specifier = ">=0.7.3" },
    { name = "mflux", specifier = "==0.15.5" },
-    { name = "mlx", marker = "sys_platform == 'darwin'", specifier = "==0.30.5" },
-    { name = "mlx", extras = ["cpu"], marker = "sys_platform == 'linux'", specifier = "==0.30.5" },
+    { name = "mlx", marker = "sys_platform == 'darwin'", specifier = "==0.30.6" },
+    { name = "mlx", extras = ["cpu"], marker = "sys_platform == 'linux'", specifier = "==0.30.6" },
    { name = "mlx-lm", specifier = "==0.30.6" },
+    { name = "msgspec", specifier = ">=0.19.0" },
    { name = "openai-harmony", specifier = ">=0.0.8" },
    { name = "pillow", specifier = ">=11.0,<12.0" },
    { name = "psutil", specifier = ">=7.0.0" },
@@ -426,6 +429,7 @@ requires-dist = [
    { name = "tiktoken", specifier = ">=0.12.0" },
    { name = "tomlkit", specifier = ">=0.14.0" },
    { name = "types-aiofiles", specifier = ">=24.1.0.20250708" },
+    { name = "zstandard", specifier = ">=0.23.0" },
 ]

 [package.metadata.requires-dev]
@@ -1042,22 +1046,22 @@ wheels = [

 [[package]]
 name = "mlx"
-version = "0.30.5"
+version = "0.30.6"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
    { name = "mlx-metal", marker = "sys_platform == 'darwin'" },
 ]
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/99/bf/e399eac4371a9d80a4bff41e1500a914a852d026ef84c30c8e427964752f/mlx-0.30.5-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:1fd93ce3de56b01699dc18a9673c89510f1772125fc48f6385dca66be34500fc", size = 572736, upload-time = "2026-02-03T04:27:58.063Z" },
-    { url = "https://files.pythonhosted.org/packages/56/e3/ad176bc95791eb09dd4414811c8fc27741c3e4832824bf1f7881f8103928/mlx-0.30.5-cp313-cp313-macosx_15_0_arm64.whl", hash = "sha256:8e24d40e8c5827a242d7b8ba2379c8ece5776b7163e05ceb7a110ec161582f91", size = 572738, upload-time = "2026-02-03T04:27:59.4Z" },
-    { url = "https://files.pythonhosted.org/packages/34/16/a1b4da06eec1efe8bb84ab450c790e7291a8610d619517965d71d31aa094/mlx-0.30.5-cp313-cp313-macosx_26_0_arm64.whl", hash = "sha256:481555f8632a54a36405a80348d88631564cecf1c16c3a9a9bf971b11c353887", size = 572813, upload-time = "2026-02-03T04:28:03.41Z" },
-    { url = "https://files.pythonhosted.org/packages/d6/7c/97c1539a593393c65cf19a0ae89005fafa13321cfb185dbb409fe9b0a955/mlx-0.30.5-cp313-cp313-manylinux_2_35_aarch64.whl", hash = "sha256:e89df90121910595529d37711a8a14f345e468b6e6caa453f40578acc90b7297", size = 621230, upload-time = "2026-02-03T04:28:05.632Z" },
-    { url = "https://files.pythonhosted.org/packages/47/ef/a266e0584ab5e301568eecf1b1e83ba92e87a395de6a7539404386ffd555/mlx-0.30.5-cp313-cp313-manylinux_2_35_x86_64.whl", hash = "sha256:027412cf65319c6d09d726a4910795811c995adeecd4d19b859a5c4eab0b0cf2", size = 662868, upload-time = "2026-02-03T04:28:07.034Z" },
-    { url = "https://files.pythonhosted.org/packages/a7/7f/2089917d98a12528a43f50e0cae675432322a36683252451d06819dbf8b1/mlx-0.30.5-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:ce54786846716f43b18733af1ccbd7f4801d659e6191da9d8297e88950533533", size = 574246, upload-time = "2026-02-03T04:28:08.368Z" },
-    { url = "https://files.pythonhosted.org/packages/d7/09/7950b1587fab3fb4922371628d57b643297ec7e336dcfb44acf2738b087f/mlx-0.30.5-cp314-cp314-macosx_15_0_arm64.whl", hash = "sha256:da405026dadbf7e5c838bdbf941484c307ad4629b8f137e52782c3093ef00fc9", size = 574245, upload-time = "2026-02-03T04:28:10.23Z" },
-    { url = "https://files.pythonhosted.org/packages/37/86/8592f35b2cb25077bb5e8bf0566b9bb2061b36baa774db3e5990e59848a4/mlx-0.30.5-cp314-cp314-macosx_26_0_arm64.whl", hash = "sha256:d2cc2d131f65f83141ffeb858eda3ae4b694030b2e1d7ed393b3a8f5fa6944f8", size = 574262, upload-time = "2026-02-03T04:28:11.548Z" },
-    { url = "https://files.pythonhosted.org/packages/3d/c7/9393ae41bff83a69321aea8d8f481faf530f3b3f5912e364094ef2245e2b/mlx-0.30.5-cp314-cp314-manylinux_2_35_aarch64.whl", hash = "sha256:3a58a108ed4cea1f644fcf0cf2a7fbaabcc43f3022da46b61cf50a7d80fe77f8", size = 623221, upload-time = "2026-02-03T04:28:13.406Z" },
-    { url = "https://files.pythonhosted.org/packages/74/e9/ee0fa04cb294382f778c7800367e8bcceac1832d0e17554ab8cabd2fc48a/mlx-0.30.5-cp314-cp314-manylinux_2_35_x86_64.whl", hash = "sha256:e623031b72e49b3c2ffbca203cc6a853f0c6c11afe271a9b410598d3fd9479f8", size = 662788, upload-time = "2026-02-03T04:28:14.764Z" },
+    { url = "https://files.pythonhosted.org/packages/ae/5b/e460e144a34d5529e010056cccf50b538d56ed001473bc6b246018fd58cb/mlx-0.30.6-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:ed86f8bffc174c2f259ca589ea25464c96cf69d1bb457074a2bf2ef53737e54f", size = 573515, upload-time = "2026-02-06T03:45:23.405Z" },
+    { url = "https://files.pythonhosted.org/packages/60/25/69833fefb9a3fef30b56792b1bcd022496c4fea83e45411d289b77ef7546/mlx-0.30.6-cp313-cp313-macosx_15_0_arm64.whl", hash = "sha256:c52294958269e20f300639a17c1900ca8fc737d859ddda737f9811e94bd040e5", size = 573516, upload-time = "2026-02-06T03:45:24.618Z" },
+    { url = "https://files.pythonhosted.org/packages/9c/6a/7e7fbeebc5cb51b6a5eba96b263a6298707bcbdc059f4b0b73e088bc3dea/mlx-0.30.6-cp313-cp313-macosx_26_0_arm64.whl", hash = "sha256:b5b6636f7c49a4d86d8ec82643b972f45a144a7a9f3a967b27b2e6e22cf71e6a", size = 573592, upload-time = "2026-02-06T03:45:25.928Z" },
+    { url = "https://files.pythonhosted.org/packages/93/06/280f6f2ba80520a7109730425eda0d966658793aa0d02d8be8d351f75253/mlx-0.30.6-cp313-cp313-manylinux_2_35_aarch64.whl", hash = "sha256:67e6c9e30a9faeacc209917ef5523177cf9b086914b6b5d83ff886e4294b727d", size = 622011, upload-time = "2026-02-06T03:45:28.165Z" },
+    { url = "https://files.pythonhosted.org/packages/fe/35/f872afbee9c079cc69924d9e9c46f5663adb7da58cba3511db082dd307c1/mlx-0.30.6-cp313-cp313-manylinux_2_35_x86_64.whl", hash = "sha256:47db8b16fcb6f6c5a47c0bdb24ed377b41237017ac93aa6cb6aa206c9bdf82e4", size = 663650, upload-time = "2026-02-06T03:45:30.315Z" },
+    { url = "https://files.pythonhosted.org/packages/60/23/361dc7a5797634e4d7e9bdd6564c6b28f9b1246672632def2f91bf066b18/mlx-0.30.6-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:78804a89dcff4a838f7c2da72392fe87a523e95122a3c840e53df019122aad45", size = 575028, upload-time = "2026-02-06T03:45:31.549Z" },
+    { url = "https://files.pythonhosted.org/packages/a8/69/1854484d414171586814dfbe8def95f75c4ea2c7341ba13ba8ee675f7c62/mlx-0.30.6-cp314-cp314-macosx_15_0_arm64.whl", hash = "sha256:ec13584ab069665cc7ad34a05494d9291cd623aef6ae96be48875fc87cfc25d6", size = 575026, upload-time = "2026-02-06T03:45:33.072Z" },
+    { url = "https://files.pythonhosted.org/packages/6b/b8/3adbc441924209a7e4c568308b2a0b54bd09aee6a68db5bae85304791e54/mlx-0.30.6-cp314-cp314-macosx_26_0_arm64.whl", hash = "sha256:b2c5e8a090a753ef99a1380a4d059c983083f36198864f6df9faaf1223d083df", size = 575041, upload-time = "2026-02-06T03:45:34.814Z" },
+    { url = "https://files.pythonhosted.org/packages/3f/54/9d9e06804fb2088202a2cdf60458e00b221f71420bea285720b60f9e82b5/mlx-0.30.6-cp314-cp314-manylinux_2_35_aarch64.whl", hash = "sha256:9ceddede4af0de31d1f6b3099f70e5469d60cd7c546975dedbdbeab3519cab3f", size = 624002, upload-time = "2026-02-06T03:45:36Z" },
+    { url = "https://files.pythonhosted.org/packages/42/92/3140a15a50cb1f9267a6552171e1dfa577861de53e093124bc43707f2a0e/mlx-0.30.6-cp314-cp314-manylinux_2_35_x86_64.whl", hash = "sha256:4a6ffd2d16728cf95f63a1b555d7c2eaeea686a0e6b73228bd265411cb5d77a4", size = 663569, upload-time = "2026-02-06T03:45:37.242Z" },
 ]

 [package.optional-dependencies]
@@ -1070,16 +1074,16 @@ cuda13 = [

 [[package]]
 name = "mlx-cpu"
-version = "0.30.5"
+version = "0.30.6"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/d7/23/5837effd6be5ed322a456623a4ad51d7c569ee0234284d96c081f73f8bda/mlx_cpu-0.30.5-py3-none-manylinux_2_35_aarch64.whl", hash = "sha256:5e848db0830f70be3e706e465badf0570ebee3379470de9fb84c8d5b45b9e2e5", size = 8685370, upload-time = "2026-02-03T03:08:32.512Z" },
-    { url = "https://files.pythonhosted.org/packages/9f/0e/b5c98e9a2d767a967b5f0e464375bb4bbc4da2190a51586ef3bdca4e7875/mlx_cpu-0.30.5-py3-none-manylinux_2_35_x86_64.whl", hash = "sha256:3d32fef8958c3e1abff39e77309d715f131609ad63c8c3be0f54e2778fbd36ba", size = 10256264, upload-time = "2026-02-03T03:08:34.872Z" },
+    { url = "https://files.pythonhosted.org/packages/c1/fd/b52c37d8514c321675bd9c8dacf1a26f493bb1004adbb3986feda76f2776/mlx_cpu-0.30.6-py3-none-manylinux_2_35_aarch64.whl", hash = "sha256:d4ca39400e06a762ab8c43c1da7e4ada87d9880b864d10eff8023dc910c76985", size = 8689310, upload-time = "2026-02-06T01:48:34.598Z" },
+    { url = "https://files.pythonhosted.org/packages/67/a1/f097b2042c3373abeb0365b52a5f5fc9176fe5c14682483b53012f714697/mlx_cpu-0.30.6-py3-none-manylinux_2_35_x86_64.whl", hash = "sha256:573baa804539b64fc3e1740d7055e80e7b47eaa4139bbace69d6f9896d2081e4", size = 10262929, upload-time = "2026-02-06T01:48:37.157Z" },
 ]

 [[package]]
 name = "mlx-cuda-13"
-version = "0.30.5"
+version = "0.30.6"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
    { name = "nvidia-cublas", marker = "sys_platform == 'linux'" },
@@ -1088,8 +1092,8 @@ dependencies = [
    { name = "nvidia-nccl-cu13", marker = "sys_platform == 'linux'" },
 ]
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/bf/cb/10622469930c6f4024cc9853be387158d9e39e0e19d97ee07120265e9ddc/mlx_cuda_13-0.30.5-py3-none-manylinux_2_35_aarch64.whl", hash = "sha256:6cd61592b4624ced451b312f1ce0cb6c951c438d9639c1b56dc43abfebb0fbaf", size = 66899934, upload-time = "2026-02-03T04:10:55.162Z" },
-    { url = "https://files.pythonhosted.org/packages/f8/9e/144f35c6b9ceaec2ddda184868e58e2a09188e2f82547dd1eabc74adddd4/mlx_cuda_13-0.30.5-py3-none-manylinux_2_35_x86_64.whl", hash = "sha256:b88db12c6eec02667902edc5a7ef3febac0822412a1477259f06abb564d09724", size = 69738714, upload-time = "2026-02-03T04:10:58.941Z" },
+    { url = "https://files.pythonhosted.org/packages/08/6b/27cd4abed63ca9e4086536acde4486ea7a762dda5b36bb867ed954efe93a/mlx_cuda_13-0.30.6-py3-none-manylinux_2_35_aarch64.whl", hash = "sha256:1a5e58eaf4b1a4eb6cb0111ecb4099f6f1380403599f8992ec93dfbad8d9c6f1", size = 66904875, upload-time = "2026-02-06T02:49:19.183Z" },
+    { url = "https://files.pythonhosted.org/packages/b1/ae/8864834f1c4637bd2f5da7fd8a6d2fb554a0bad16fbe53bb5ecdfcf1611f/mlx_cuda_13-0.30.6-py3-none-manylinux_2_35_x86_64.whl", hash = "sha256:35792e799bbf1498f3e0792c5c34c64bd13075de6d34d88e223de4b1000bab82", size = 69745930, upload-time = "2026-02-06T02:49:22.844Z" },
 ]

 [[package]]
@@ -1112,12 +1116,12 @@ wheels = [

 [[package]]
 name = "mlx-metal"
-version = "0.30.5"
+version = "0.30.6"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/5f/60/1bb7019c10645a3a7f3c5decea9241f5ff18133da96d396dabf27dd6f6c4/mlx_metal-0.30.5-py3-none-macosx_14_0_arm64.whl", hash = "sha256:f8287d0204170231aefda0b03acbb3b54336848be83e7916d7ef18d44e23593a", size = 38259369, upload-time = "2026-02-03T04:27:19.468Z" },
-    { url = "https://files.pythonhosted.org/packages/31/4a/0b6a59629f9ee4de76c1b018a0c4d5a38844048f6ebccc2c72d2dba89bf0/mlx_metal-0.30.5-py3-none-macosx_15_0_arm64.whl", hash = "sha256:9ce0b169ec7043d46d63a7b04e048fdd51246a134c8ee7eccb9dfb8db9f3f468", size = 38254849, upload-time = "2026-02-03T04:27:23.08Z" },
-    { url = "https://files.pythonhosted.org/packages/8e/d4/038a7281a44366d963b30b3867423f59f65e3dae479cffe0e63855c600f3/mlx_metal-0.30.5-py3-none-macosx_26_0_arm64.whl", hash = "sha256:04cfe0bf2b95a294c5072e7084a708ca3e1ba9454775af17b1be2877a4606a9f", size = 47536555, upload-time = "2026-02-03T04:27:26.385Z" },
+    { url = "https://files.pythonhosted.org/packages/f3/85/44406b521f920248fad621334d4dc15e77660a494edf890e7cbee33bf38d/mlx_metal-0.30.6-py3-none-macosx_14_0_arm64.whl", hash = "sha256:ea6d0c973def9a5b4f652cc77036237db3f88c9d0af63701d76b5fddde99b820", size = 38437818, upload-time = "2026-02-06T03:44:56.19Z" },
+    { url = "https://files.pythonhosted.org/packages/d0/cb/10a516995f7d0c154b0d7e633c54b51e96977a86a355105b6474cfcbe0d0/mlx_metal-0.30.6-py3-none-macosx_15_0_arm64.whl", hash = "sha256:0f8cb94634d07e06a372d6ad9a090f38a18bab1ff19a140aede60eacf707bb94", size = 38433701, upload-time = "2026-02-06T03:44:59.678Z" },
+    { url = "https://files.pythonhosted.org/packages/4c/7d/70cb272f7373c334709f210ed8420511fc9d64d05a7a646c0b3b94c29c04/mlx_metal-0.30.6-py3-none-macosx_26_0_arm64.whl", hash = "sha256:d761ae26304f2c4b454eeea7f612a56919d9e5e57dbb1dc0788f8e34aa6f41c2", size = 47718448, upload-time = "2026-02-06T03:45:03.133Z" },
 ]

 [[package]]
@@ -1138,6 +1142,32 @@ wheels = [
    { url = "https://files.pythonhosted.org/packages/43/e3/7d92a15f894aa0c9c4b49b8ee9ac9850d6e63b03c9c32c0367a13ae62209/mpmath-1.3.0-py3-none-any.whl", hash = "sha256:a0b2b9fe80bbcd81a6647ff13108738cfb482d481d826cc0e02f5b35e5c88d2c", size = 536198, upload-time = "2023-03-07T16:47:09.197Z" },
 ]

+[[package]]
+name = "msgspec"
+version = "0.20.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/ea/9c/bfbd12955a49180cbd234c5d29ec6f74fe641698f0cd9df154a854fc8a15/msgspec-0.20.0.tar.gz", hash = "sha256:692349e588fde322875f8d3025ac01689fead5901e7fb18d6870a44519d62a29", size = 317862, upload-time = "2025-11-24T03:56:28.934Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8a/d1/b902d38b6e5ba3bdddbec469bba388d647f960aeed7b5b3623a8debe8a76/msgspec-0.20.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:9c1ff8db03be7598b50dd4b4a478d6fe93faae3bd54f4f17aa004d0e46c14c46", size = 196463, upload-time = "2025-11-24T03:55:43.405Z" },
+    { url = "https://files.pythonhosted.org/packages/57/b6/eff0305961a1d9447ec2b02f8c73c8946f22564d302a504185b730c9a761/msgspec-0.20.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:f6532369ece217fd37c5ebcfd7e981f2615628c21121b7b2df9d3adcf2fd69b8", size = 188650, upload-time = "2025-11-24T03:55:44.761Z" },
+    { url = "https://files.pythonhosted.org/packages/99/93/f2ec1ae1de51d3fdee998a1ede6b2c089453a2ee82b5c1b361ed9095064a/msgspec-0.20.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f9a1697da2f85a751ac3cc6a97fceb8e937fc670947183fb2268edaf4016d1ee", size = 218834, upload-time = "2025-11-24T03:55:46.441Z" },
+    { url = "https://files.pythonhosted.org/packages/28/83/36557b04cfdc317ed8a525c4993b23e43a8fbcddaddd78619112ca07138c/msgspec-0.20.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7fac7e9c92eddcd24c19d9e5f6249760941485dff97802461ae7c995a2450111", size = 224917, upload-time = "2025-11-24T03:55:48.06Z" },
+    { url = "https://files.pythonhosted.org/packages/8f/56/362037a1ed5be0b88aced59272442c4b40065c659700f4b195a7f4d0ac88/msgspec-0.20.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:f953a66f2a3eb8d5ea64768445e2bb301d97609db052628c3e1bcb7d87192a9f", size = 222821, upload-time = "2025-11-24T03:55:49.388Z" },
+    { url = "https://files.pythonhosted.org/packages/92/75/fa2370ec341cedf663731ab7042e177b3742645c5dd4f64dc96bd9f18a6b/msgspec-0.20.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:247af0313ae64a066d3aea7ba98840f6681ccbf5c90ba9c7d17f3e39dbba679c", size = 227227, upload-time = "2025-11-24T03:55:51.125Z" },
+    { url = "https://files.pythonhosted.org/packages/bb/18/62dc13ab0260c7d741dda8dc7f481495b93ac9168cd887dda5929880eef8/msgspec-0.20.0-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:eead16538db1b3f7ec6e3ed1f6f7c5dec67e90f76e76b610e1ffb5671815633a", size = 196407, upload-time = "2025-11-24T03:55:55.001Z" },
+    { url = "https://files.pythonhosted.org/packages/dd/1d/b9949e4ad6953e9f9a142c7997b2f7390c81e03e93570c7c33caf65d27e1/msgspec-0.20.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:703c3bb47bf47801627fb1438f106adbfa2998fe586696d1324586a375fca238", size = 188889, upload-time = "2025-11-24T03:55:56.311Z" },
+    { url = "https://files.pythonhosted.org/packages/1e/19/f8bb2dc0f1bfe46cc7d2b6b61c5e9b5a46c62298e8f4d03bbe499c926180/msgspec-0.20.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6cdb227dc585fb109305cee0fd304c2896f02af93ecf50a9c84ee54ee67dbb42", size = 219691, upload-time = "2025-11-24T03:55:57.908Z" },
+    { url = "https://files.pythonhosted.org/packages/b8/8e/6b17e43f6eb9369d9858ee32c97959fcd515628a1df376af96c11606cf70/msgspec-0.20.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:27d35044dd8818ac1bd0fedb2feb4fbdff4e3508dd7c5d14316a12a2d96a0de0", size = 224918, upload-time = "2025-11-24T03:55:59.322Z" },
+    { url = "https://files.pythonhosted.org/packages/1c/db/0e833a177db1a4484797adba7f429d4242585980b90882cc38709e1b62df/msgspec-0.20.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:b4296393a29ee42dd25947981c65506fd4ad39beaf816f614146fa0c5a6c91ae", size = 223436, upload-time = "2025-11-24T03:56:00.716Z" },
+    { url = "https://files.pythonhosted.org/packages/c3/30/d2ee787f4c918fd2b123441d49a7707ae9015e0e8e1ab51aa7967a97b90e/msgspec-0.20.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:205fbdadd0d8d861d71c8f3399fe1a82a2caf4467bc8ff9a626df34c12176980", size = 227190, upload-time = "2025-11-24T03:56:02.371Z" },
+    { url = "https://files.pythonhosted.org/packages/4c/06/3da3fc9aaa55618a8f43eb9052453cfe01f82930bca3af8cea63a89f3a11/msgspec-0.20.0-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:f84703e0e6ef025663dd1de828ca028774797b8155e070e795c548f76dde65d5", size = 200389, upload-time = "2025-11-24T03:56:06.375Z" },
+    { url = "https://files.pythonhosted.org/packages/83/3b/cc4270a5ceab40dfe1d1745856951b0a24fd16ac8539a66ed3004a60c91e/msgspec-0.20.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:7c83fc24dd09cf1275934ff300e3951b3adc5573f0657a643515cc16c7dee131", size = 193198, upload-time = "2025-11-24T03:56:07.742Z" },
+    { url = "https://files.pythonhosted.org/packages/cd/ae/4c7905ac53830c8e3c06fdd60e3cdcfedc0bbc993872d1549b84ea21a1bd/msgspec-0.20.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5f13ccb1c335a124e80c4562573b9b90f01ea9521a1a87f7576c2e281d547f56", size = 225973, upload-time = "2025-11-24T03:56:09.18Z" },
+    { url = "https://files.pythonhosted.org/packages/d9/da/032abac1de4d0678d99eaeadb1323bd9d247f4711c012404ba77ed6f15ca/msgspec-0.20.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:17c2b5ca19f19306fc83c96d85e606d2cc107e0caeea85066b5389f664e04846", size = 229509, upload-time = "2025-11-24T03:56:10.898Z" },
+    { url = "https://files.pythonhosted.org/packages/69/52/fdc7bdb7057a166f309e0b44929e584319e625aaba4771b60912a9321ccd/msgspec-0.20.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:d931709355edabf66c2dd1a756b2d658593e79882bc81aae5964969d5a291b63", size = 230434, upload-time = "2025-11-24T03:56:12.48Z" },
+    { url = "https://files.pythonhosted.org/packages/cb/fe/1dfd5f512b26b53043884e4f34710c73e294e7cc54278c3fe28380e42c37/msgspec-0.20.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:565f915d2e540e8a0c93a01ff67f50aebe1f7e22798c6a25873f9fda8d1325f8", size = 231758, upload-time = "2025-11-24T03:56:13.765Z" },
+]
+
 [[package]]
 name = "multidict"
 version = "6.7.0"
@@ -2486,3 +2516,37 @@ wheels = [
    { url = "https://files.pythonhosted.org/packages/cf/72/6a85bba382f22cf78add705d8c3731748397d986e197e53ecc7835e76de7/yarl-1.22.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:c0232bce2170103ec23c454e54a57008a9a72b5d1c3105dc2496750da8cfa47c", size = 341473, upload-time = "2025-10-06T14:12:10.994Z" },
    { url = "https://files.pythonhosted.org/packages/73/ae/b48f95715333080afb75a4504487cbe142cae1268afc482d06692d605ae6/yarl-1.22.0-py3-none-any.whl", hash = "sha256:1380560bdba02b6b6c90de54133c81c9f2a453dee9912fe58c1dcced1edb7cff", size = 46814, upload-time = "2025-10-06T14:12:53.872Z" },
 ]
+
+[[package]]
+name = "zstandard"
+version = "0.25.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/fd/aa/3e0508d5a5dd96529cdc5a97011299056e14c6505b678fd58938792794b1/zstandard-0.25.0.tar.gz", hash = "sha256:7713e1179d162cf5c7906da876ec2ccb9c3a9dcbdffef0cc7f70c3667a205f0b", size = 711513, upload-time = "2025-09-14T22:15:54.002Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/35/0b/8df9c4ad06af91d39e94fa96cc010a24ac4ef1378d3efab9223cc8593d40/zstandard-0.25.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:ec996f12524f88e151c339688c3897194821d7f03081ab35d31d1e12ec975e94", size = 795735, upload-time = "2025-09-14T22:17:26.042Z" },
+    { url = "https://files.pythonhosted.org/packages/3f/06/9ae96a3e5dcfd119377ba33d4c42a7d89da1efabd5cb3e366b156c45ff4d/zstandard-0.25.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:a1a4ae2dec3993a32247995bdfe367fc3266da832d82f8438c8570f989753de1", size = 640440, upload-time = "2025-09-14T22:17:27.366Z" },
+    { url = "https://files.pythonhosted.org/packages/d9/14/933d27204c2bd404229c69f445862454dcc101cd69ef8c6068f15aaec12c/zstandard-0.25.0-cp313-cp313-manylinux2010_i686.manylinux2014_i686.manylinux_2_12_i686.manylinux_2_17_i686.whl", hash = "sha256:e96594a5537722fdfb79951672a2a63aec5ebfb823e7560586f7484819f2a08f", size = 5343070, upload-time = "2025-09-14T22:17:28.896Z" },
+    { url = "https://files.pythonhosted.org/packages/6d/db/ddb11011826ed7db9d0e485d13df79b58586bfdec56e5c84a928a9a78c1c/zstandard-0.25.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:bfc4e20784722098822e3eee42b8e576b379ed72cca4a7cb856ae733e62192ea", size = 5063001, upload-time = "2025-09-14T22:17:31.044Z" },
+    { url = "https://files.pythonhosted.org/packages/db/00/87466ea3f99599d02a5238498b87bf84a6348290c19571051839ca943777/zstandard-0.25.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:457ed498fc58cdc12fc48f7950e02740d4f7ae9493dd4ab2168a47c93c31298e", size = 5394120, upload-time = "2025-09-14T22:17:32.711Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/95/fc5531d9c618a679a20ff6c29e2b3ef1d1f4ad66c5e161ae6ff847d102a9/zstandard-0.25.0-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:fd7a5004eb1980d3cefe26b2685bcb0b17989901a70a1040d1ac86f1d898c551", size = 5451230, upload-time = "2025-09-14T22:17:34.41Z" },
+    { url = "https://files.pythonhosted.org/packages/63/4b/e3678b4e776db00f9f7b2fe58e547e8928ef32727d7a1ff01dea010f3f13/zstandard-0.25.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:8e735494da3db08694d26480f1493ad2cf86e99bdd53e8e9771b2752a5c0246a", size = 5547173, upload-time = "2025-09-14T22:17:36.084Z" },
+    { url = "https://files.pythonhosted.org/packages/4e/d5/ba05ed95c6b8ec30bd468dfeab20589f2cf709b5c940483e31d991f2ca58/zstandard-0.25.0-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:3a39c94ad7866160a4a46d772e43311a743c316942037671beb264e395bdd611", size = 5046736, upload-time = "2025-09-14T22:17:37.891Z" },
+    { url = "https://files.pythonhosted.org/packages/50/d5/870aa06b3a76c73eced65c044b92286a3c4e00554005ff51962deef28e28/zstandard-0.25.0-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:172de1f06947577d3a3005416977cce6168f2261284c02080e7ad0185faeced3", size = 5576368, upload-time = "2025-09-14T22:17:40.206Z" },
+    { url = "https://files.pythonhosted.org/packages/5d/35/398dc2ffc89d304d59bc12f0fdd931b4ce455bddf7038a0a67733a25f550/zstandard-0.25.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:3c83b0188c852a47cd13ef3bf9209fb0a77fa5374958b8c53aaa699398c6bd7b", size = 4954022, upload-time = "2025-09-14T22:17:41.879Z" },
+    { url = "https://files.pythonhosted.org/packages/9a/5c/36ba1e5507d56d2213202ec2b05e8541734af5f2ce378c5d1ceaf4d88dc4/zstandard-0.25.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:1673b7199bbe763365b81a4f3252b8e80f44c9e323fc42940dc8843bfeaf9851", size = 5267889, upload-time = "2025-09-14T22:17:43.577Z" },
+    { url = "https://files.pythonhosted.org/packages/70/e8/2ec6b6fb7358b2ec0113ae202647ca7c0e9d15b61c005ae5225ad0995df5/zstandard-0.25.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:0be7622c37c183406f3dbf0cba104118eb16a4ea7359eeb5752f0794882fc250", size = 5433952, upload-time = "2025-09-14T22:17:45.271Z" },
+    { url = "https://files.pythonhosted.org/packages/7b/01/b5f4d4dbc59ef193e870495c6f1275f5b2928e01ff5a81fecb22a06e22fb/zstandard-0.25.0-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:5f5e4c2a23ca271c218ac025bd7d635597048b366d6f31f420aaeb715239fc98", size = 5814054, upload-time = "2025-09-14T22:17:47.08Z" },
+    { url = "https://files.pythonhosted.org/packages/b2/e5/fbd822d5c6f427cf158316d012c5a12f233473c2f9c5fe5ab1ae5d21f3d8/zstandard-0.25.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:4f187a0bb61b35119d1926aee039524d1f93aaf38a9916b8c4b78ac8514a0aaf", size = 5360113, upload-time = "2025-09-14T22:17:48.893Z" },
+    { url = "https://files.pythonhosted.org/packages/3d/5c/f8923b595b55fe49e30612987ad8bf053aef555c14f05bb659dd5dbe3e8a/zstandard-0.25.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:e29f0cf06974c899b2c188ef7f783607dbef36da4c242eb6c82dcd8b512855e3", size = 795887, upload-time = "2025-09-14T22:17:54.198Z" },
+    { url = "https://files.pythonhosted.org/packages/8d/09/d0a2a14fc3439c5f874042dca72a79c70a532090b7ba0003be73fee37ae2/zstandard-0.25.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:05df5136bc5a011f33cd25bc9f506e7426c0c9b3f9954f056831ce68f3b6689f", size = 640658, upload-time = "2025-09-14T22:17:55.423Z" },
+    { url = "https://files.pythonhosted.org/packages/5d/7c/8b6b71b1ddd517f68ffb55e10834388d4f793c49c6b83effaaa05785b0b4/zstandard-0.25.0-cp314-cp314-manylinux2010_i686.manylinux_2_12_i686.manylinux_2_28_i686.whl", hash = "sha256:f604efd28f239cc21b3adb53eb061e2a205dc164be408e553b41ba2ffe0ca15c", size = 5379849, upload-time = "2025-09-14T22:17:57.372Z" },
+    { url = "https://files.pythonhosted.org/packages/a4/86/a48e56320d0a17189ab7a42645387334fba2200e904ee47fc5a26c1fd8ca/zstandard-0.25.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:223415140608d0f0da010499eaa8ccdb9af210a543fac54bce15babbcfc78439", size = 5058095, upload-time = "2025-09-14T22:17:59.498Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/ad/eb659984ee2c0a779f9d06dbfe45e2dc39d99ff40a319895df2d3d9a48e5/zstandard-0.25.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:2e54296a283f3ab5a26fc9b8b5d4978ea0532f37b231644f367aa588930aa043", size = 5551751, upload-time = "2025-09-14T22:18:01.618Z" },
+    { url = "https://files.pythonhosted.org/packages/61/b3/b637faea43677eb7bd42ab204dfb7053bd5c4582bfe6b1baefa80ac0c47b/zstandard-0.25.0-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:ca54090275939dc8ec5dea2d2afb400e0f83444b2fc24e07df7fdef677110859", size = 6364818, upload-time = "2025-09-14T22:18:03.769Z" },
+    { url = "https://files.pythonhosted.org/packages/31/dc/cc50210e11e465c975462439a492516a73300ab8caa8f5e0902544fd748b/zstandard-0.25.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e09bb6252b6476d8d56100e8147b803befa9a12cea144bbe629dd508800d1ad0", size = 5560402, upload-time = "2025-09-14T22:18:05.954Z" },
+    { url = "https://files.pythonhosted.org/packages/c9/ae/56523ae9c142f0c08efd5e868a6da613ae76614eca1305259c3bf6a0ed43/zstandard-0.25.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:a9ec8c642d1ec73287ae3e726792dd86c96f5681eb8df274a757bf62b750eae7", size = 4955108, upload-time = "2025-09-14T22:18:07.68Z" },
+    { url = "https://files.pythonhosted.org/packages/98/cf/c899f2d6df0840d5e384cf4c4121458c72802e8bda19691f3b16619f51e9/zstandard-0.25.0-cp314-cp314-musllinux_1_2_i686.whl", hash = "sha256:a4089a10e598eae6393756b036e0f419e8c1d60f44a831520f9af41c14216cf2", size = 5269248, upload-time = "2025-09-14T22:18:09.753Z" },
+    { url = "https://files.pythonhosted.org/packages/1b/c0/59e912a531d91e1c192d3085fc0f6fb2852753c301a812d856d857ea03c6/zstandard-0.25.0-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:f67e8f1a324a900e75b5e28ffb152bcac9fbed1cc7b43f99cd90f395c4375344", size = 5430330, upload-time = "2025-09-14T22:18:11.966Z" },
+    { url = "https://files.pythonhosted.org/packages/a0/1d/7e31db1240de2df22a58e2ea9a93fc6e38cc29353e660c0272b6735d6669/zstandard-0.25.0-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:9654dbc012d8b06fc3d19cc825af3f7bf8ae242226df5f83936cb39f5fdc846c", size = 5811123, upload-time = "2025-09-14T22:18:13.907Z" },
+    { url = "https://files.pythonhosted.org/packages/f6/49/fac46df5ad353d50535e118d6983069df68ca5908d4d65b8c466150a4ff1/zstandard-0.25.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:4203ce3b31aec23012d3a4cf4a2ed64d12fea5269c49aed5e4c3611b938e4088", size = 5359591, upload-time = "2025-09-14T22:18:16.465Z" },
+]
				`@@ -0,0 +1 @@`
				`collect_ignore = ["tests/start_distributed_test.py"]`